AIoT-enhanced health management system using soft and stretchable triboelectric sensors for human behavior monitoring

Sedentary, inadequate sleep and exercise can affect human health. Artificial intelligence (AI) and Internet of Things (IoT) create the Artificial Intelligence of Things (AIoT), providing the possibility to solve these problems. This paper presents a novel


| INTRODUCTION
In recent years, there has been a growing public concern regarding suboptimal health conditions resulting from unhealthy lifestyles.2][3] For instance, insufficient exercise can lead to fatigue and dizziness, while increasing the risk of obesity and cardiovascular diseases. 4Additionally, sitting for extended periods not only impedes blood circulation but also reduces metabolism, resulting in cardiovascular disease, cervical and lumbar diseases. 5Prolonged poor sleep can lead to serious consequences such as memory loss, high blood pressure, and coronary heart disease. 6Therefore, balancing the relationship between exercise, work and sleep and developing reasonable lifestyle habits is crucial for health management.8][9] While smartphones and watches are popular for tracking sleep and exercise, they may not be ideal due to issues like discomfort, low detection sensitivity, radiation, and reliance on external power sources. 10The long-term use of the sitting posture monitoring system based on the passive restraint structure brings about comfort and dependence problems. 11Traditional sleep monitors are expensive, complex, and inconvenient to carry, often requiring hospital guidance. 12It should be noted that these health monitoring devices are typically designed to operate independently, with specific functions and complex structures, which can create obstacles to effective information exchange and hinder the establishment of a comprehensive and efficient monitoring system.While vision-based monitoring systems are popular for their high efficiency, they require good lighting and raise privacy concerns.Given these challenges, it is crucial to develop a self-powered, comfortable, non-invasive, high-sensitivity, and simple-structured overall health management system for all-day work, sleep, and exercise behavior monitoring.
4][15][16][17][18] Therefore, wearable electronic devices have emerged as a possible solution to monitor daily human behavior while avoiding privacy issues associated with visual monitoring systems.However, a major challenge for wearable devices is the constant need for an energy supply, limiting their long-term usage. 19,202][23][24][25][26][27] However, piezoelectric and pyroelectric effects have some limitations in material selection and temperature requirements, respectively.Meanwhile, the electromagnetic effect can be unstable due to magnetic field fluctuations, impacting sensing accuracy. 286][37][38][39][40] Notably, TENG sensors based on flexible substrates and electrodes allow for enhanced collection of dynamic signals from the human body, enabling the recognition of different behaviors.Liquid metal, as a novel electrode material, possesses high conductivity, flexibility, and good controllability, enabling the fabrication of fully flexible TENG sensors for richer usage scenarios.Therefore, using TENG combined with flexible materials offers a viable solution for designing self-powered sensors to monitor human activity in diverse scenarios and develop overall health management systems.
][43][44][45] Among these fields, health management systems have gained substantial interest in both research and commercialization.1][52] Note that combining these technologies can further transform realtime monitoring of human physiology to the prevention and management of chronic diseases.For instance, Tao et al. proposed an intelligent artificial larynx that incorporates machine learning (ML) technology to restore the vocalization ability of voice-impaired individuals, 53 while An et al. developed an intelligent neck motion detector that enhances neck monitoring and rehabilitation performance. 54Such fusion of AI and IoT technologies has given rise to a new research field known as artificial Intelligence of Things (AIoT), which can greatly improve the ability of multi-modal tactile perception, thereby improving the effectiveness of health management. 55,56herefore, the utilization of flexible TENG sensors for capturing behavioral data in various human motions, including gait, sitting posture, and sleep, and so on, while integrating an AIoT-based health management system for analyzing these signals, holds significant research value.Meanwhile, the integration of this flexible TENG with AIoT enables these health management systems to have notable features such as cost-effectiveness, comfort, privacy protection, and all-day functionality.
Till now, there has been a growing interest in using TENG and AIoT technology for daily health monitoring and human motion analysis.One significant research direction in this field is gait analysis, which involves attaching sensors to capture human gait information. 47,57,58By utilizing data analysis methods such as ML, signals from TENG-based socks and insoles can be analyzed for gait monitoring, 13,42,[59][60][61][62][63][64][65][66][67][68][69][70] particularly for recognizing abnormal gaits (as shown in Table S1).There remain continuous challenges to balance the relationship between sensor amount, monitoring accuracy, and comfort for gait detection, that is, designing a minimum system with better sensory performance.Similarly, although sitting posture correction systems based on wearable TENG patches or cushions have also been developed (Table S2), [71][72][73][74][75][76] how to reduce the number of sensors while improving recognition accuracy with usage comfort and privacy protection is also required.Moreover, HMI to correct sitting posture is necessary for health management.3][84][85][86] Developing an intelligent pillow to capture head motions shows its advantages, but a TENG-based pillow with ML-enhanced intelligence has not been reported yet.Such design may reduce sensor amounts required to build pillows while improving the sensing accuracy for more head movement situations, which will contribute to analyzing sleeping status.Notably, all the aforementioned scenarios can utilize modularly designed TENG sensors individually or in an array, decreasing the complexity of the system.Thereby, this monitoring integrated with working, exercise, and sleeping contains full human activity information, and it brings new insight for health management with the assistance of AIoT.
Herein, this paper presents a soft and stretchable TENG sensor designed for human behavior monitoring in AIoT-based health management during daily activities (Figure 1A).The sensor consists of Galinstan liquid metal and silicone rubber, with excellent stretchability and flexibility.One TENG sensor is placed on the right insole to create an intelligent insole, enabling the capture of diverse walking patterns through triboelectric signals.
Additionally, the CNN model algorithm is introduced to recognize the identities of different individuals and classify exercise states, enabling personalized exercise monitoring and evaluation to enhance the effectiveness of workouts (see Figure 1B).Similarly, the smart cushion is composed of 12 TENG sensor arrays, which provide identification of different persons and discrimination of different sitting postures by applying the CNN model for signal analysis (Figure 1C).This system is further integrated with sedentary reminding and lamp control to promote healthy work habits.Moreover, a smart pillow equipped with 15 TENG sensors is utilized to record real-time head position and motion during sleep, offering emergency alerts for adverse conditions to improve sleep safety and quality (see Figure 1D).Eventually, based on the data collected from these intelligent devices, a health management strategy is proposed, and it demonstrates that the user's health status can be displayed in real-time through the HMI.In conclusion, this flexible triboelectric sensorbased intelligent system covers a wide range of daily behaviors, providing an affordable and versatile solution for AIoT-based health management (Figure 1E).

| Structure, working principle and characterization of SS-TENG sensor
The advancement of wearable sensors for human behavior monitoring necessitates a balance between sensor comfort and performance.Leveraging TENG technology's material versatility, we opt for Galinstan as the TENG electrode, chosen for its room-temperature liquid state and exceptional conductivity.To ensure flexibility, we encase the Galinstan electrode in silicone rubber, serving as the triboelectric material. 87To establish dependable electrical connections, we utilize external wires that extend deep into the TENG cavity.Simultaneously, we meticulously consider the length of these internal wires to increase exposure to liquid metal.This fusion results in the SS-TENG, a soft and stretchable sensor (Figure S1a,b), tailored for wearable human behavior monitoring devices.
First, we design a 3D-printable mold for crafting the silicone rubber chamber of the SS-TENG sensor.This chamber consists of a cover plate and a lower housing with unique groove features.Using a Zortrax M200 Plus 3D printer, we construct the mold with Z-ABS material.We generate two types of silicone rubber films from Ecoflex 00-30 (Smooth-on) by combining parts A and B in a 1:1 weight ratio.After stirring for 2 min and vacuum treatment for 3 min to remove bubbles, we pour the mixture into the mold, solidifying it at 70 C for 30 min.Galinstan Alloy (as shown in Figure S1a), a composition of 68.5% Gallium, 21.5% Indium, and 10.0% Tin, is selected for its room-temperature liquid state.We bond two silicone films using liquid silicone, and inject liquid metal into the silicone rubber cavity through a syringe, connecting a wire for electrical linkage.The complete fabrication process of the sensor can be seen in Figure S2.
In this study, the SS-TENG sensor (4 cm Â 4 cm Â 3 mm) functions in a single-electrode mode, illustrated in Figure S1c.To show the structure of the SS-TENG sensor more clearly, no external object is placed in the picture shown in Figure S1b.When compressed, it generates charges on both the silicone rubber and an external object (like Ni-fabric) through triboelectrification. Initial contact surface charges are balanced, resulting in no electron flow in the external circuit.Upon Ni-fabric separation from silicone, this balance is disrupted.Electrons move from liquid metal to ground, compensating for silicone's negative charges.Conversely, as Ni-fabric reconnects with silicone, electrons flow from ground to electrode, yielding a reverse output pulse.This mechanism adeptly transforms mechanical energy into electrical energy, facilitating self-powered sensing and human behavior monitoring applications.To address the potential problem of internal liquid metal leakage under significant impact loads, we design a support structure within the cavity, as illustrated in Figure S3.Through stress and strain analysis using the Solidworks Simulation module, we find that the SS-TENG sensor with the support structure exhibits enhanced load-bearing capacity when subjected to the same longitudinal deformation as the SS-TENG sensor without the support structure under external loads (Figure S4a,c).Notably, the use of the support structure prevents extensive deformation within the sensor cavity (Figure S4b), reducing the likelihood of liquid metal overflow to some extent.Conversely, significant deformation occurs inside the sensor cavity without the support structure (Figure S4d).This approach not only safeguards against potential leakage but also conserves liquid metal, effectively reducing the manufacturing cost of SS-TENG sensors.
Next, we utilize a linear motor to assess the electrical performance of the SS-TENG (depicted in Figure S5).The SS-TENG sensor is positioned opposite the linear motor, and the motor's pressing plate is coated with Ni-fabric.We vary compressive force from 5 to 150 N at a 1 Hz frequency to analyze the open-circuit voltage, short-circuit current and transferred charge of the SS-TENG sensors under diverse normal forces (as shown in Figure S6a-c).The findings reveal that as force increases, the opencircuit voltage rises from 3 to around 75 V, confirming its exceptional sensing range and sensitivity.Beyond 80 N, voltage remains essentially unchanged, indicating good voltage stability under higher pressures.Simultaneously, short-circuit current goes from 0.15 μA at 1 N to approximately 3 μA at 150 N and transferred charge grows from 45 to 67 nC due to increased contact area.Despite the significant voltage levels generated by the proposed SS-TENG sensor, it appears to yield a relatively low output current ranging from 0.15 to 3 uA.The limited amount of small, short-lived current may prove insufficient for the autonomous operation of the TENG sensor.9][90] Frequency testing from 0.5 to 5 Hz (Figure S6d-f) demonstrates voltage and current correlation with frequency rise, while transferred charges remain stable.The proposed SS-TENG sensor exhibits the potential for self-powered human behavior monitoring during cyclic compression.To further assess the sensitivity of the SS-TENG sensor, we establish a minimum compression force of 0.5 N. As depicted in Figure S6g, the sensor exhibits remarkable sensitivity at forces below 20 N, registering 2.47 V/N.However, as the force applied surpasses 40 N and reaches up to 150 N, the sensor demonstrates a stable output sensitivity of 0.14 V/N.This underscores its robust performance even under substantial compression forces.Furthermore, flexibility serves as a key criterion for assessing the sensor when considering it as a wearable electronic device.Illustrated in Figure S6h, the sensor effortlessly undergoes stretching, twisting, and bending due to the incorporation of flexible packaging materials and liquid metal electrodes.
We further conduct various tests to evaluate the energy harvesting capabilities and durability of our sensor.Under 1 Hz and 20 N excitation, we vary load resistances (from 0.1 to 1000 MΩ) to measure the SS-TENG sensor's output performance.In Figure S7, results show that output voltage gradually climbs to a peak of 50 V with increasing load resistance, following Ohm's law as current decreases.At 200 MΩ load resistance, the sensor achieves around 19 μW peak output power.Durability tests involve cyclic loading (Figure S8a), revealing the SS-TENG sensor's anti-fatigue ability and stable output over 5000 cycles.Liquids like coffee, orange juice, and water are also tested, with the sensor's signal remaining steady even after 15 min of soaking (Figure S8b).Notably, the sensor remains robust even after 10 h of water exposure, showcasing its washability.Overall, the SS-TENG sensor's wide sensing range, sensitivity, and flexibility render it fitting for versatile human behavior monitoring applications.

| Demonstration of gait analysis via smart insole
Exercise is crucial for maintaining good health, while long-term work pressure and extended working hours will reduce the available time for physical activity.Walking, as a fundamental form of exercise in daily life, has garnered increased attention due to its ability to promote metabolism, alleviate fatigue, and its convenience, as it can be performed anytime and anywhere.Human gait during walking encompasses valuable sensory information that can be monitored using various sensors to acquire different kinds of physical parameters such as step count and walking speed.Currently, insoles or socks based on different principles are commonly applied for this purpose. 13,42n this work, the SS-TENG sensor is integrated into the heel of an insole to create a smart insole (Figure 2B), that is, insole-based TENG (I-TENG).This integration is chosen because the heel area experiences higher and continuous pressure shocks during walking compared to other regions.Here, we conduct experiments on an 80-kg volunteer, using a commercially available pressure sensor to measure and record the force exerted by the heel while walking.As shown in Figure S9a, when the heel is not in contact with the ground, the numerical change of the pressure sensor is negligible.Although there are some fluctuations when the heel is in contact with the ground, the force range is about 150 N during states (2) and (3) and does not show an obvious force peak when the heel leaves the ground again.This observation suggests that the pressure produced by the heel of an 80 kg person during normal walking is approximately 150 N, although this result may be affected by the shape of the sole of the foot.Considering that heavier users may cause liquid metal leakage when stepping on the SS-TENG sensor, we use a pressure testing machine to test the voltage response of the SS-TENG under a pressing force of 250 N.
As shown in Figure S9b, the sensor can stably output voltage signals, and without leakage of liquid metal.
According to the above pressure test on the heel during walking.As the human walks, the I-TENG is continuously released and compressed to generate triboelectric signals.To verify this phenomenon, experiments are conducted with a volunteer who wears shoes for normal walking, where only the right shoe is equipped with the I-TENG (Figure S10).Meanwhile, a portable signal processing circuit (Figure S11) is utilized to collect and process the signals generated by the sensor, improving signal acquisition accuracy while maintaining portability.During the gait cycle, when the user steps forward, the right heel lifts off the ground, generating a negative triboelectric peak.Then, when the right heel contacts the ground, it results in a positive triboelectric peak (Figure 2C).
The key to assessing exercise intensity and ensuring effective health management lies in obtaining personal information and evaluating exercise status.Notably, gait information always encapsulates various human features including body weight and walking habits, and this information is inherently encoded into the triboelectric signals generated by the smart insole.In addition to traditional analysis strategies based on signal frequency and peak values, machine learning techniques enable the extraction of subtle differences from triboelectric signals for revealing implicit information.Thus, combined with a CNN predictive model, the SS-TENG-based insole is possible to identify different people, help derive personalized information, and combine exercise state judgment for exercise amount calculation.The fundamental layout and workflow of the smart insole are depicted in Figure 2A.
In  Note that although larger datasets with more sensors may offer more features, potentially leading to higher recognition accuracy, we choose to collect sensory information from a single sensor under the right foot to reduce data volume and simplify the system.By evaluating the cross-entropy loss function, we optimize the structure of the CNN model and set two convolutional layers and two pooling layers ultimately as shown in Figures 2H and S12a.Eventually, the CNN-based model demonstrates high positive predictive value and true positive rate for people recognition, resulting in an overall recognition accuracy of 94.86% (Figure 2F).In our quest for more favorable test outcomes, we amass an additional 60 sets of data, thereby enlarging the training set to a total of 160 sets.However, contrary to our expectations, the test results do not show a significant enhancement despite the increased data volume (as shown in Figure S13a).In order to find a more ideal algorithm to solve this problem.We also apply the initial 100 group datasets to predict classification accuracy using an artificial neural network (ANN) and support vector machine (SVM).The accuracy achieved is only 67.43% and 69.14%, respectively, as illustrated in Figure S13b.This observation indirectly underscores the capability of the CNN model we employ to extract features with subtle differences.
Building upon the smart insole people recognition, we apply the same algorithm model to assess the activities of two volunteers.This attempt is guided by the understanding that exercise intensity correlates with activity actions, thereby enhancing the insole's adaptability and monitoring capabilities.Employing the I-TENG data acquisition circuit and the CNN algorithm model from the earlier identification, we record 150 sets of walking, jogging, and jumping actions performed by the two participants.Each set of data samples is standardized to consist of 600 data points, creating a dataset.Simultaneously, a monitoring window ensures uniform activity action steps within each sample.Figure 2E illustrates typical sensor signal outputs for the walking, jogging, and jumping states of the two participants.The 150 action samples for each participant are randomly divided into three groups following a 4:1:1 ratio (train: 100 samples, test: 25 samples, and validation: 25 samples).Subsequently, the selected 125 samples are input into the CNN framework to construct the predictive model.Ultimately, the CNN-based model demonstrates a high positive predictive value and true positive rate in recognizing a total of 6 actions performed by the two participants, yielding an overall recognition accuracy of 97.33% (see Figure 2G).It is noteworthy that the CNN classifier we employ exhibits robust performance in gait recognition when confronted with variations in weight (for the same volunteer), step duration, and sensor misalignment.Refer to Note S1 for detailed information.Furthermore, by analyzing the time series of voltage peak outputs corresponding to different gait states, it becomes possible to calculate the walking speed by combining it with the individual stride distance.After experiments, it was observed that the time intervals between peak voltages generated by the insole during fast walking, normal walking, and slow walking are approximately 0.83, 1.15, and 1.53 s, respectively (Figure S14).Thereby, real-time walking speed can be calculated by utilizing the formula Vs = L/T, where L represents the stride length in a walking cycle and T represents the time taken to complete the cycle.For an adult male with a height of 180 cm and a weight of 70 kg, walking and jogging consume about 0.05 and 0.15 calories per step.Combined with the step count and speed information provided by the smart insole, the activity energy consumed can be calculated based on the user's gait and body shape information.In Figure S15, we provide estimates of activity energy expenditure after half an hour of exercise at different gaits for the man.Thereafter, a visual interface has been developed to enable real-time monitoring of insole signals, facilitating the detection and analysis of walking speed, gait states and active energy (Video S1).
In short, using ML and a visual interface, we are able to apply identity recognition and activity recognition to the health management system.This enables us to provide customized exercise plans and monitoring for different individuals through an IoT-based remote setup.Notably, with only 100 training samples from one SS-TENG sensor, we are able to achieve remarkably accurate recognition.In order to show its function, we further conduct a demonstration including the identification of 7 individuals and the recognition of a total of 6 motions for two individuals (see Video S2), and further used a visual interface to show how the system accesses personal accounts to complete personalized exercise monitoring (see Video S1).The implementation of such an intelligent system allows for the monitoring and management of daily exercise status and volume, providing valuable assistance for health management.

| Demonstration of sitting monitoring via smart cushion
Maintaining correct sitting posture is crucial for daily work and study, contributing to enhanced physical and mental well-being. 11Currently, researchers are exploring two main methods for real-time monitoring of sitting posture.One approach involves using cameras to record the user's posture from video streams and then applying various algorithms to assess the sitting posture. 91Another solution is to use passive restraint belts or brackets to maintain users in a correct sitting posture during work or study. 92These approaches may pose challenges for privacy protection or lead to passive disengagement of body muscles and ineffective muscle memory formation.
Herein, we propose a novel approach that addresses comfort, privacy, and effective posture correction by utilizing 12 SS-TENG sensors to create a smart cushion, that is, cushion-based TENG (C-TENGs), comprising a base and a backrest (see Figures 3B and S16).EPE foam and Velcro ensure both comfort and easy removability of the sensor array (Figure S17), allowing the smart cushion to be adaptable to various chairs for collecting sitting posture information.Notably, since sitting posture information doesn't involve human faces or other privacysensitive information, machine learning technologies can be safely used for complex classification tasks.CNNs have shown strong performance in identity recognition and gait classification in the previous section.Hence, the combination of a smart cushion with a CNN-based prediction model can provide a comprehensive and reliable sitting posture monitoring solution for different users.
In this study, a CNN-based model utilizes signals acquired from a smart cushion to evaluate user status, including user identification and sitting posture monitoring.The entire process is illustrated in Figure 3A different users, as depicted in Figure 3E.Similarly, we continue to augment the collection of training set data but do not achieve a more desirable accuracy (refer to Figure S13a); instead, it affects the convergence speed.This observation also indicates that a dataset of 140 sets is sufficient for distinguishing the identities of the five participants.Moreover, the real-time user identity recognition based on C-TENGs is demonstrated in Video S3.
After obtaining the user's identity using the smart cushion, monitoring their sitting posture becomes crucial.To assess the sitting posture, we recruited a volunteer to collect signals while adopting various sitting postures, including normal sitting, leaning back, leaning forward, and crossing legs (left/right).Utilizing the 12 SS-TENG sensors, distinct feature differences can be observed among the five different sitting postures.During the signal acquisition process, 150 sets of data are collected for each of the five different sitting postures, following a ratio of 4:1:1, we selected 100 sets of data as the training set and fed them into the CNN framework to establish a prediction model.Notably, in comparison to identity recognition, different sitting postures exhibit more pronounced feature differences (Figure 3D).Consequently, by using only 100 sets of data to construct a prediction model through the CNN network, we achieve a recognition accuracy of 97.6% for the five different sitting postures (Figure 3F).Real-time sitting posture monitoring, based on C-TENGs, is effectively demonstrated in Video S4.Simultaneously, any improper sitting posture is promptly displayed on the interface and receives timely reminders.
Furthermore, extended periods of sitting during work or study can greatly increase the risk of various health issues, including eye strain, neck strain, and spine problems.Therefore, it is crucial to have effective reminders for sedentary sitting.We utilized two SS-TENG sensors positioned in the middle of the C-TENGs to sense the sitting time and provide prolonged sitting reminders as shown in Figure 3G.The MCU collects the dual-channel data from the cushion and processes it through the processing circuit and Bluetooth module.When the C-TENGs generates the sitting signal, the MCU sends a corresponding control command through Bluetooth, and the timer starts to count (Figure 3F).If the user doesn't leave the cushion entirely within the set time, the timer continues to run until reaching the time threshold, even if the user changes their sitting position (Video S5).On the other hand, if the user leaves the cushion before the set reminder time, the timer is automatically reset and waits for the next sitting signal.This method enables fastresponse statistics of sitting time (Video S6).Additionally, we have developed a wireless desk lamp control system, as shown in Figure S18.The MCU detects the sitting and standing signals, and sends control commands via the Bluetooth to turn on the desk lamp.When the MCU detects the leaving signal from the cushion, it triggers the system to turn off the lamp (Video S7).This demonstration not only brings convenience to work and study but also helps to reduce energy waste caused by forgetting to turn off the lights when leaving the work area.
Overall, our proposed smart cushion enabled by ML offers a comfortable and fast-response option for realtime sitting posture monitoring and excessive sedentary time reminding.Its user-friendly design allows easy assembly and disassembly, making it adaptable to various seats for monitoring different users.By integrating user identification, sitting posture monitoring, and sedentary reminders into a comprehensive health management system, it effectively mitigates health risks associated with poor posture and prolonged sitting.In future, we will integrate the haptic feedback technology in the AIoTbased health management system. 93Hence, we can remind the users on inappropriate poses during long sitting situation.

| Demonstration of sleeping monitoring via smart pillow
Monitoring postural changes during sleep is significant for evaluating sleep quality, and real-time head movement is an important indicator of the overall body state during sleep. 79Although numerous wearable sleep monitoring devices are available in the market that utilizes a three-dimensional acceleration sensor to collect body motion during sleep, wearing these devices may cause discomfort and potentially affect sleep quality. 94Therefore, there requires alternative approaches to avoid such limitations.
To address this issue, we develop a smart pillow measuring 700 mm Â 350 mm Â 70 mm with 15 flexible SS-TENG sensors arranged in arrays, that is, pillow-based TENG (P-TENG).EPE foam is used as the upper and lower interlayers to form a pillow, and Velcro is introduced in the interlayer to fix the SS-TENG sensor (Figure S19).A multi-channel signal acquisition system is employed to collect P-TENGs signals generated by head motion during sleep in real-time (Figure 4B).The collected P-TENGs signals are then sent to a built-in machine learning model for head position and head motion analysis (Figure 4A).To achieve this function, we divided the surface of the smart pillow into seven zones labeled as A-G along the length direction (Figure S20), and set five sleep scenarios, including lying down, getting up, turning left, turning right, and falling out of pillow.As the head touches different zones of the pillow during the lying-down process, the P-TENG signals provide realtime feedback on head motion and position (Figure S21a).Similarly, the motion of getting up can also generate output from the corresponding sensor (Figure S21b).Notably, to prevent users from being too close to the edge of the pillow during sleep, which could result in a stiff neck or falling the smart pillow provides reminders when the head falls on areas A and G (Video S8).
It is important to consider the motions during sleep, particularly when turning over.To define these motions, the shifting of the head between different positions is used to simulate turning over.For instance, when the head shifts from zone F to zone D, this motion is considered as turning to the right side.Thereby, nine signal patterns are identified for turning to the right (Figure S21d pillow, as demonstrated in Figure S21c and Video S8.It is noteworthy that even if the head turns from the center of the pillow to the edge, such as from zone C to zone A or from zone E to zone G, the system can provide a warning to avoid any harmful effects. Additionally, we further apply the CNN algorithm to distinguish different signal features corresponding to head motion and position as shown in Figure 4C.Using the signal output of lying down, getting up, turning head to the right, and turning the head to the left, the accuracy of judging head motion and position reached 96.43%, 95.71%, 93.89%, and 95.56%, respectively (Figure S16), reflecting our designed CNN model has good robustness and performance.The feasibility of using smart pillows (P-TENGs) for sleep motion monitoring is further verified as shown in Video S9.Generally, our SS-TENG-based smart pillow provides a comfortable and non-invasive way to monitor head motion during sleep.

| Intelligent health management system based on wearable SS-TENG sensor
In modern society, many people gradually realize that lack of exercise, sedentary lifestyle, and poor sleep quality will damage their health.Health management is becoming an essential part of people's daily lives, this study proposes a smart insole, cushion and pillow to monitor human behavior in these scenarios and establish a health management system based on information collected by SS-TENG sensors.This system's execution can be categorized into three primary states: sleep scene monitoring, working scene monitoring, and continuous motion monitoring throughout (Figure 5A).Commencing each morning, when the user wears shoes embedded with smart insoles, the I-TENG sensor is activated.It identifies the user's identity through simple gait information and establishes a connection with the system, initiating motion monitoring.During this phase, movement is monitored based on three pivotal factors: gait analysis, step count, and calories burned.The data recording engine evaluates the daily level of physical activity based on these influential factors.When the user sits on a chair furnished with the smart cushion, the data response from the C-TENGs is linked to the system after verifying the user's identity.Subsequently, the system monitors the user's sitting posture.Smart cushion can assess the duration of sitting, instances of poor posture, and the frequency of reminders.This information aids individuals in adjusting their sitting postures, thereby mitigating physical discomfort and reducing the risk of diseases associated with prolonged sitting.The integration of an identity recognition module widens the monitoring scope of smart cushions, ensuring independent and uninterrupted monitoring information between different users.Upon the disappearance of signals from both C-TENGs and I-TENG, the signal generated by P-TENGs will be determined whether the user is going to bed or not.At this point, the smart pillow is activated for sleep monitoring.It evaluates sleep quality by recording sleep duration, the number of warnings (indicating potential sleep disturbances), and the number of head movements.This data empowers individuals to gain a better understanding of their sleep patterns and facilitates early detection of sleep-related issues.Based on the collected data, the system provides personalized exercise and improvement recommendations, such as increasing physical activity or increasing sleep duration (Figure 5C).All monitoring data is presented to users through an intuitive and concise visual interface (as illustrated in Figure 5Bii-iv).Users can access their personal data at any time, enabling them to effectively manage their health.Through analysis and integration of this data, the system can provide personalized health advice and guidance, assisting individuals in rectifying detrimental habits and enhancing their overall wellbeing.In summary, this intelligent health management system utilizes smart insoles, cushions, and pillows equipped with SS-TENG sensors to continuously monitor individuals' behavior in different scenarios.It provides personalized health advice and guidance, empowering individuals to make positive adjustments to their habits and enhance their overall health.

| CONCLUSION
This paper introduces the SS-TENG, a soft and stretchable sensor designed to monitor human health during various daily activities, including exercise, work, and sleep.The SS-TENG sensor utilizing liquid metal electrodes and silicone material is highly flexible and adaptable.The design incorporates a support structure with the liquid metal cavity, minimizing liquid metal usage and boosting strain capacity.When integrated into an insole, it monitors the gait by decoupling triboelectric signals and provides real-time information on step count, speed, and activity consumption, enabling people to better understand their exercise amount.By combining a CNN-based algorithm, such a smart insole is demonstrated to recognize seven people and distinguish three movement modes with solely one sensor with a high accuracy rate of 94.86% and 97.33%, respectively.This significantly reduces the complexity of personalized exercise assessment systems.To monitor sitting postures, a smart cushion is constructed with 12 SS-TENGs, pearl foam and Velcro, providing sufficient flexibility for sitting comfort.Moreover, the CNN-based model achieves excellent real-time recognition accuracy of 98.86% for identification of five participants and 98.4% for recognizing five sitting postures.This cushion is further proven to achieve sitting management by providing sedentary reminders and lamp control based on the posture recognition result.In addition, the SS-TENG sensors are further used to create a smart pillow, which could record real-time head motion and position during sleep.Our result indicates that it achieves 96.25% accuracy in identifying eight typical sleep patterns (can be extended to monitor 34 head motions), providing valuable insights into improving sleep quality and emergency warnings.By analyzing data collected from these smart devices, a visual intelligent health management strategy is proposed that quantitatively analyzes daily health status and provides personalized suggestions.In summary, this soft and stretchable triboelectric sensor-based intelligent system covers a wide range of daily behaviors.Establish linkage between exercise, work and sleep through flexible and wearable triboelectric sensors, offering a lowcost and general-purpose solution for AIoT-based health management.METHODS

| Characterization and electrical measurement
We utilize the electrometer (Model 6514B, Keithley) to measure various parameters including open-circuit voltage, short-circuit current, and short-circuit charge of the SS-TENG.The collected data was recorded by a highquality oscilloscope (DSOX3034T, Keysight).To measure the sensor output under different external stimuli, we employed a linear motor (C1100, LinMot) to apply 0.5-150 N forces to the SS-TENG sensor.This was achieved by adjusting the working displacement, speed, and acceleration of the electromagnetic motor at different frequencies.A multi-material test system (ZQ-990, ZHIQU) was used to test the voltage output of the SS-TENG sensor under a pressing force of 250 N.

| Sensory system and applications
To collect real-time signals, a signal processing circuit was connected to a microcontroller unit (Mega2560, Arduino).The Arduino is equipped with the ATmega2560 chip (Arduino Mega2560, Zejie), and has 16 channels of analog and digital input channels.The collected signals were then transmitted to a computer via a Universal Serial Bus (USB) communication cable.In the smart insole system, the TENG sensor transmits data to the computer through the serial port at a baud rate of 9600.In this configuration, the ADC's sampling rate is approximately 150 Hz, allowing a single channel to collect 150 data points per second.Conversely, in the smart cushion and pillow systems, where the serial port transmission baud rate is 57 600, the ADC operates at a sampling rate of about 120 Hz and 100 Hz, respectively.In these scenarios, a single channel can collect 120 and 100 data points per second.For applications such as sitto-stand time and lamp control, a commercial Bluetooth module (HC-05) was used to transmit data.The Bluetooth module transmission rate is 9600.The machine learning system utilized for analyzing the collected signals was based on a convolutional neural network (CNN), which was developed in Python using the Tensor-Flow framework.

| Study participation
Prior to participation in the experiments, informed consent was obtained from the volunteer in all experiments.

F I G U R E 1
Schematics of artificial intelligence of things-based intelligent health management system.(A) All-day human activity monitoring system for intelligent health management using triboelectric nanogenerator (TENG) sensors.(B) Schematic of TENG-based smart insole for gait analysis.(C) Schematic of TENG-based smart cushion for sitting monitoring.(D) Schematics of TENG-based smart pillow for sleep monitoring.(E) Future applications of intelligent health management system using TENG sensors.

F
I G U R E 2 Gait analysis based on SS-TENG-based smart insole and machine learning.(A) Schematic diagram of the gait analysis by smart insole and machine learning.(B) Optical photo of the smart insole with SS-TENG sensor.(C) Schematic of two states of the I-TENG during the normal walking cycle.(D) I-TENG outputs corresponding to seven participants.(E) Real-time activity signals for different participants.(F) Confusion map of the prediction with the gait patterns of seven participants.(G) Confusion map of different activity patterns for two participants.(H) The CNN model for people identification and activity pattern recognition.
order to establish the CNN-based model, seven participants with 50-80 kg weights are invited to perform identical stepping motions after wearing shoes with the smart insole.During the experiment, the signals generated by the I-TENG are first collected and processed by an analog-to-digital converter (ADC) and microcontroller (MCU)-based hardware circuit.Since the smart insole only uses one SS-TENG sensor, we only need to use one channel of the MCU during the signal collection process and transmit it to the computer for storage through the serial port.Subsequently, each participant records 150 sets of the I-TENG sensor signals (3 steps per set), resulting in a dataset containing 600 data points for a single channel per sample.Meanwhile, a monitoring window that can accommodate 600 data points is employed to visualize the signals, ensuring that each sample contains three consecutive steps.The single-channel raw data is used directly as the training feature.Each feature represents a data point in the time series during stepping, encompassing information such as contact force, stepping speed, and contact time.
Figure 2D shows the typical sensor outputs from different participants, and each participant's 150 samples are randomly divided into three groups in a 4:1:1 ratio (training: 100 samples, testing: 25 samples, validation: 25 samples).The selected 125 samples are fed into the CNN framework to build the prediction model.
. To begin with, the signal collection process involves stipulating that each C-TENG signal sample represents a complete sitting posture.In order to ensure the consistency of each signal collection, we collect the signal brought by each normal sitting movement.Here, we uniformly window each sitting movement into a single SS-TENG channel to generate 350 feature points.By utilizing the 12 analog input ports of the MCU, 12 Â 350 data points can be acquired from the 12 SS-TENG sensors of the smart cushion each time, forming a data sample.Each data sample comprises the voltage amplitude of different SS-TENG sensors every time an individual sits down on the seat cushion, along with the changing trend of the voltage in the time domain.Figure 3C illustrates the utilization of an ADC to collect 210 sets of data during normal sitting postures involving five different users.Furthermore, the collected multidimensional data from the 12 channels are converted into one-dimensional vectors.Subsequently, the collected data are randomly divided into training, testing, and verification sets in a ratio of 4:1:1.Specifically, 140 groups are used for training, 35 groups for testing, and another 35 groups for verification.The CNN framework is employed to construct the prediction model by inputting the training set, which includes five different identity labels.By training the CNN model and optimizing it using the loss function (Figure S12b), we can achieve a recognition accuracy of 98.40% for distinguishing the identities of the five F I G U R E 3 Sit analysis based on SS-TENG-based smart cushion and machine learning.(A) Schematic diagram of the sit analysis by smart cushion and machine learning.(B) Optical photo of the smart cushion system.(C) 3D plots of the C-TENGs sensors outputs with the normal sitting posture corresponding to seven participants.(D) 3D plots of the C-TENGs sensors outputs corresponding to five sitting postures.(E) Confusion map of the prediction of five participants.(F) Confusion map of five different sitting postures.(G) Hardware and flow chart for sitting time statistical and signal mapping based on two-channel C-TENGs.
), while the other nine signal patterns for turning to the left are shown in Figure S21e.To address the risk of falling from the bed for infants, old people, and unconscious patients when they roll over to the side during sleep, an early warning function based on the smart pillow is designed to detect if the head falls off the edge of the F I G U R E 4 Sleep monitoring based on SS-TENG-based smart pillow and machine learning.(A) Schematic diagrams of the head motion and position monitoring by smart pillow and machine learning.(B) Optical photo of the smart pillow system.(C) 3D plots of the P-TENGs sensors outputs corresponding to different motions and positions.(D) Overview of the recognition system.(E) Confusion map of machine learning training results, where the front part of the label corresponds to the head motion and the back part corresponds to the head position.
Here, the data acquisition and machine learning training process are similar to the smart insole and smart cushion.Each data sample has 300 features, and 100 samples are utilized to train the model.Specifically, we selected 8 typical signals from 7 lying down signals, 7 getting up signals, 9 turning right signals, 9 turning left signals, and 2 fall out of pillow signals (Figure S22).After training, the CNN-based model achieves a judgment accuracy of 96.25% for eight different head motions (Figure 4D).

F
I G U R E 5 SS-TENG-based intelligent health management system.(A) Overview of the operation of the SS-TENG-based intelligent health management system.(B) Visualization program connected to SS-TENGs for gait, sit and sleep monitoring.(C) Feasible stay healthy recommendations based on the state of the day.