Sensor data fusion for the industrial artificial intelligence of things

The emergence of smart sensors, artificial intelligence, and deep learning technologies yield artificial intelligence of things, also known as the AIoT. Sophisticated cooperation of these technologies is vital for the effective processing of industrial sensor data. This paper introduces a new framework for addressing the different challenges of the AIoT applications. The proposed framework is an intelligent combination of multi‐agent systems, knowledge graphs and deep learning. Deep learning architectures are used to create models from different sensor‐based data. Multi‐agent systems can be used for simulating the collective behaviours of the smart sensors using IoT settings. The communication among different agents is realized by integrating knowledge graphs. Different optimizations based on constraint satisfaction as well as evolutionary computation are also investigated. Experimental analysis is undertaken to compare the methodology presented to state‐of‐the‐art AIoT technologies. We show through experimentation that our designed framework achieves good performance compared to baseline solutions.

sensor data became a necessity. One of the main reasons for this is that companies do not analyse their data but learn from the data to extract valuable knowledge and decision in the future.
Although IoT technologies provide tools to deal with smart sensors analysis, the exponentially growing network environment in IoT also brings to light complex and heterogeneous information, which will be a challenge in future industrial settings (Da et al., 2014;Baza et al., 2020;Li et al., 2018;Zhang & Tao, 2021). Without strong intelligent techniques for learning from events as well as data, serious degradation in terms of systems performances may be lead.

| Motivation
Various artificial intelligence (AI) is widely applied to solve complex problems in IoT applications (Tan et al., 2021). Deep learning, multi-agent systems, knowledge graphs, are just examples of such methods that prove their success in the recent decade. Deep learning (Dai et al., 2019;Dai et al., 2020;Dwivedi et al., 2019;Liu et al., 2019;Luo et al., 2020;Ma et al., 2018;Weng et al., 2019) are used to learn from the previous examples of the industrial environment. For instance, the work in Weng et al. (2019) incorporated both blockchain and reinforcement learning to address the federated learning challenges such as security and privacy. Multi-agents systems (Alqahtani et al., 2020;Casado-Vara et al., 2018;Chen, Srivastava, et al., 2020;Ciatto et al., 2019;Cicirelli et al., 2019;Zhou et al., 2019) are used to address the challenges of the internet of things applications. For instance, the works in Alqahtani et al. (2020) and Casado-Vara et al. (2018) use intelligent agents to improve the security and privacy in the internet of things environments. This is ensured by integrating blockchain technology and certificate authentication mechanisms for data sharing among users of the platform. Besides, the works (Ciatto et al., 2019;Cicirelli et al., 2019) attempt to enhance cognition of the internet of things applications by integrating the main features of the intelligent agents' behaviours in the whole internet of things system. In the context of knowledge graphs, several semantic-based solutions have been proposed in the AIoT (Bellini et al., 2014;Iwendi et al., 2020;Le-Phuoc et al., 2016;Li et al., 2019;Qiu et al., 2018;Qiu et al., 2019). For instance, Qui et al. (2019) combines both word co-occurrences and community detection to build the semantic graph of heterogeneous information of the smart sensors data.
Previous AIoT solutions do not address the industrial challenges where large-scale data are usually collected from distributed and heterogeneous smart sensors. Therefore, new methods are needed to improve the existing AIoT systems in both accuracy and runtime performances.

| Contributions
The novelty of the present research is to develop an end-to-end solution to overcome the limitations of the AIoT systems, by proving an efficient and intelligent collaborative framework using a multi-agent system, the knowledge graph, and deep learning. The main contributions are fourfold as follows: 1. We propose a new intelligent collaborative framework to handle smart sensor data in the AIoT. The proposed framework uses deep learning and a multi-agent system for creating automated models representing large-scale data collected from heterogeneous and distributed smart sensors.
2. We use both the convolution neural network and the recurrent neural network in the learning phase. The convolution neural network deals with matrix representation such as images, where the recurrent neural network deals with time series representation such as trajectories.
3. We offer a novel intelligent method for enhancing communication among the various actors by combining the knowledge graph process. We also investigate different optimizations such as constraint satisfaction programming, and evolutionary computation to make the whole framework robust and efficient. 4. We evaluate the proposed framework using two industrial use cases, including smart building, and intelligent transportation. The results show that our novel framework can be useful when directly compared to two baseline methods for object detection and two other baseline methods for outlier detection.
The remainder of this work is structured as follows: Related work is studied in Section 2. The proposed framework and designed algorithm are discussed in Section 3. We report our experimental results in Section 4. Section 5 presents the main finding of the application of hybrid collaborative intelligent framework (jHCIFj) in industrial settings. Section 6 concludes the paper.

| LITERATURE REVIEW
Different deep learning methods have been developed in the context of artificial intelligence internet of things applications. Dai et al. (2019) created an architecture based on reinforcement learning that can be used to secure next-generation networks in IIoT. The designed system was shown to maximize utility as well as accurately cache data sharing throughout the network. Weng et al. (2019) implemented a framework called DeepChain, which is a deep learning-based framework that is also distributed able to navigate federated learning problems. In DeepChain learners could in essence behave incorrectly where updating parameters. Liu et al. (2019) tackle IIoT problems and were able to adopt a learning approachable to give a mechanism used for the evaluation of IIoT systems using in terms of decentralization, scalability, security, throughput, and latency. Dai et al. (2020) again investigate the problem of online offloading but this time made us of a Markov-based decision process. The authors integrated deep learning as well as the genetic algorithm (GA). The process was shown to be able to maximize offloading performance in the long term. Luo et al. (2020) administered a distributed IoT technology that could synchronize many continuous local views using independent SDN (software-defined networking) controllers. The technology could reach a consensus even on global views. The approach itself was shown to reduce computational resources significantly even when considering hidden features of particular controllers as well as resource constraints in the environment used.
To improve cognition, and autonomy several solutions based on multi-agent systems have been proposed. Cicirelli et al. (2019) suggested a multi-agents platform to establish cognition in decentralized smart city based algorithms. It also exploits the intelligent agents for developing internet of things applications and makes it reactive, proactive, and cognitive behaviours. Roberto et al. (2018) developed a multi-agent system to analyse and manage the supply chain process in the internet of things settings. It also provides an efficient mechanism for ensuring security using blockchain technology. Ciatto et al. (2019) developed agentify system which enhances the autonomy of the intelligent agents by exploring sociality and situatedness features of each agent in the system. Alqahtani et al. (2020) developed a multi-agent system for enhancing monitoring and improving the security scheme in the cloud internet of things environment. It creates a middleware layer to manage the communication among users in the platform. Intelligent agents provide security by sharing trust and certificates for authentication. Alsboui et al. (2020) suggested the use of distributed ledger platform based on mobile agents to improve the scalability of data sharing in tremendous P2P networks. It also provides an intelligence mechanism, where the mobile agents are integrated for node-level communications and data collections.
Communication is a crucial issue in multi-agent systems. A knowledge graph is a powerful artificial intelligence representation that targets the approach for reliably extracting semantic knowledge from heterogeneous data from smart sensors. Because of similarities, data from smart cities are initially calculated using the term co-occurrence. Following that, a semantic network is created based on similarities between smart city data.
Finally, a community identification method is used, which may partition a given smart city into independent, disjoint communities, each of which can be defined as a separate concept.
Through this short review of the literature, one can see that there is a lot of exciting research that has been implemented with connections to AIoT. Many research items look into learning mechanisms for the creation of automated models, while some methods also explore multi-agent systems that are capable of dealing with smart sensor data from distributed environments. Moreover, other methods dive into semantic modelling for helping data sharing in IoT environments. Nonetheless, these methods have not been shown to be successful in industrial settings, where there has not been an abundance of recent research. In a different direction altogether, in this research work, we investigate to our knowledge the first dedicated complete framework that combines knowledge graphs, multi-agent systems, and deep learning to envision well-used solutions to solve smart sensor data for applicability in AIoT.

| Principle
Let us begin by describing the key elements of the HCIF. As shown in Figure 1. The input data is collected from smart sensors in the internet of things settings. The data is entered into the HCIF framework, for analysis and pattern learning. The output of the HCIF will be a set of guidelines to the user for helping the decision-making process. Note that the proposed HCIF framework is a generic one, which may be applied for solving any decision-making tasks including, prediction, outlier analysis, regression, and so on. Our framework builds upon several intelligent components such as deep learning, multi-agent systems, knowledge graphs, constraint satisfaction, and evolutionary computation for dealing with smart sensor data. In particular, we use deep learning to create automated models from the smart sensors data. As HCIF is a generic framework, different deep learning architectures are used, the convolution neural network is used for handling images and videos, where the recurrent neural network is used for handling time series data. The multi-agents system is then used to deal with distributed and heterogeneous smart sensor data, where the communication is realized using the knowledge graph techniques. Different optimization has been integrated, the first one is the constraint satisfaction programming, where the second one is the evolutionary computation process. Each agent first performs the deep learning architectures to build intelligent models. In the inference step, each agent uses ontology matching to interpret its output to the other agents in the system. The security privacy issues are modelled by the set of constraints, and solved by two strategies; the first one is an exact strategy, where the complete instantiation is generated, and the second one is the approximate-based strategy, where the aim is to find an instantiation converged to the optimal one. PSO is also made use of for the optimization of processing runtime for the solver as developed. HCIF components are detailed and described in the rest of this section.

| Deep learning
Here, we made use of two DL architectures for use in the training of smart sensor data.
1. Convolution neural network: Convolution Neural Network (CNN) (Mishkin et al., 2017) is a group of deep frameworks that can be applied often into computer vision (CV) problems like visual recognition as well as object detection. CNN has also recently been applied to many other data representation types like time series as well as textual data. The primary purpose of CNN is in feature extraction directly from matrix data making use of convolutional filters. Here, we can define convolutional filters as the set of weights applied directly to each and every element within the matrix data pixels. Weights are adjusted as well as learned making use of traditional backpropagation methods. In our research, we used VGG16, the well-known CNNs, which is known for being able to achieve 92.8% top six test accuracy on Image Net, with close to 16 million image examples in close to 2000 classes.() 2. Recurrent neural network: Recurrent neural network (RNN) architecture is used (De Mulder et al., 2015). The input consists of heterogeneous smart sensor data. A multilayer-feed-forward network is applied made up of many neurons that are formed into layers. For any given layer, a neuron gets linked to all other neurons using a combination of different weight values. Input layer neurons get associated with every input data. Output layer neurons get connected to each output. The overall aim is to be able to minimize the error between the data that is output from the network and its corresponding ground truth data.
The mth neuron output from layer l, defined as s m l is shown in Equation 1. We note here that the output sums for all neurons for a given layer set to 1, as illustrated in Equation 2. with F I G U R E 1 Hybrid collaborative intelligent framework where b m l is the bias value for neuron s m l , j l j is the number of neurons in layer l, s j lÀ1 is the output of jth neuron in l À 1 layer, ω mj lÀ1 is the weight value connects neurons s m l to s j lÀ1 , and σ : ð Þ: activation function.
The updating weight rule is given in each iteration i by multiplying the previous weights with the ith smart sensor data D i , the learning rate μ, and the error rate E i , as illustrated in Equation 3: where the error rate E i is described in Equation 4 as follows: The algorithm starts by initializing weight values. At iteration i, neurons of the input layer get input data that consists of smart-sensor data. Input gets explored through neurons in the network making use of Equation 1. Network output gets compared to ground truth, where any error is found using Equations 4. Error-values are propagated throughout the network for updating weight value as shown using Equation 3. The process is repeated for all input.

| Multi-agents systems
Collaboration is an important factor in the AIoT. To simulate the interaction among the different entities in the AIoT, a multi-agent system is integrated into the HCIF framework. To increase the smart behaviour of the agents in HCIF, reinforcement learning is developed. Therefore, we define a multi-agents system by a tuple < A, S,U,R > . A is the set of agents, each can be shown to be a decision process based on Markov, where S is defined as a finite set of environment states, where U is defined as a set of actions and also where R can be defined as the reward function.
Each agent's behaviour as it pertains to A can be represented by a policy, which specifies how every agent can choose actions for a given state.
Each agent's purpose is to maximize a given objective function based on its policy. For example, in the case of prediction, a policy of each agent is to maximize the number of predicted objects. The following describes the different components of the multi-agents system of HCIF: 1. State: Each agent's future action is determined by the preceding states' decisions. As a result, each agent's state is made up of two parts: the set of prior actions and the current data to be handled. The size of S is determined by observations in a particular database.
2. Action: It is the assignment of each observation's decision-making behaviour in the database. For instance, in a prediction job, it is the assignment of each object's class.
3. Reward: Determining a suitable reward function is critical. It enables each agent in A to learn more effectively. We used data that contained ground truth to provide rewards to the agent's activities. The following is the definition of the reward function: is the decision of the agent A i , whether the observation O j has correct action or not. L O j À Á is the ground-truth of the observation O j .

Environment:
The environment is a collection of databases that store a huge amount of data from smart sensor devices. This enables the environment to produce specific situations for the agent's training and to estimate the optimal actions to perform.
Each and every agent A i begins by scanning all of the observations in the ith smart sensor, agent A i then calculates the 1st observation, as well as remaining observations for the ith sensor. The given reward function R A i , U j À Á is calculated on each decision using ground truth of 1st observation. As detailed, the process given is repeated then for the remainder of observations of ith sensor. The resulting local decision set LD i can be derived for each and every agent A i .

| Knowledge graphs
Intelligent communication may be used to integrate the local choices of the agents in A by utilizing a knowledge graph. Ontology matching is one of the robust techniques, which allows representing the knowledge graphs. In the following, we present an approach that allows matching between the different ontologies of the agents in A. . The problem statement's purpose relating to the ontologymatching problem among agents that are A can be said to be for the determination of common properties that exist between ontologies. In other words, for the determination of the function M as is given in Equation 5. where The naïve approach to the ontology-matching problem is to compare all values of instances across ontologies. The outcome of the alignment is determined by the matching procedure. Each match may result in a unique alignment instance. Each alignment result is then reviewed and compared to the reference alignment. The reference alignment is one that has been recommended by a user or domain expert. The reference alignment comprises all instances of the common ontology. Any sort of naive approach can be highly costly from a time perspective especially for smart sensors that are known to usually collect vast amounts of data. As a solution to this problem, in this paper we suggest and propose a novel algorithm that can be used directly with ontology matching in any smart sensor environment. The algorithm first breaks up the entire set of instances of each and every ontology into many clusters that still have dependencies. Through this process, each cluster still has some correlated instances that will need to be processed later. Second, the algorithm explores all of the instances in a given cluster to try and locate common fea- properties are shared between cluster instances to some threshold value. This threshold value as such allows each cluster to remain highly correlated. Through a matching process, the algorithm investigates all of the instances in a given cluster to look for alignments. Alignments between instances of two ontologies and their clusters are performed instead of looking to perform alignments between instances of ontologies one at a time.

| Optimization
To improve privacy, we employ an optimization solution that is based on a constraint satisfaction problem. The limitations of the whole smart sensor data set in IoT settings are divided into various groups. The clusters are investigated in order to identify the optimal configuration that secures privacy. We design two techniques for investigating constraint clusters: Approximation-based strategy. Clusters are processed separately not considering shared data between sites. Local instantiation of data is carried out on each cluster using a search process. A merging function is used to create a global instantiation. This function combines all local instantiations of data. Such an approach returns a subset of all data from the whole set of constraints. This is because the shared data were not taken into account in the search process.
Proposition 1. An upper/lower bound of constraints satisfied by the approximation-based strategy, noted j A j, is j C j (j C j À P k i¼1 S G i ð Þ , where C is the clusters of the constraints, k is the number of clusters, S G i ð Þ is the number of conflicts constraints of the cluster G i ), and we note j C j À Proof. In the worst case, the number of non-satisfied constraints of the given approximation-based strategy is equal to the number of constraints from the shared variables between all clusters. This is maybe realized, where the shared variables appeared in the constraints of the clusters. In this case, the number of satisfied constraints of the approximation-based strategy is equal to j C j minus all the number of constraints, containing the shared variables of all clusters, equal to P k i¼1 S G i ð Þ. In the case of the conflict, constraints do not use the shared variables during the search process, the number of satisfied constraints of the approximation-based strategy is j C j. □ According to this proposition, the quality of the approximation-based method is largely dependent on the quantity of common data between all clusters. The approximation-based method can meet all requirements if the number of shared data is kept to a minimum. This will be fixed by carefully selecting the number of clusters.
Exact strategy: In the search process, this method takes into account both shared data and clusters. This enables us to identify the best potential instantiation of smart sensor data to meet all restrictions. To generate the local instantiation of data, the search method is first performed to each cluster of constraints. The local instantiation of data is concatenated with the instantiation of shared data to get the global instantiation of all constraints.
In our previous works (Belhadi, Djenouri, Diaz, et al., 2021;Djenouri et al., 2020), we tried GA (Genetic Algorithm) and PSO (Particle Swarm Optimization) in solving the hyper-parameters issues, however, we find PSO provides better intensification and diversification compared to the GA. PSO (Darwish et al., 2020) can be used for solver optimization in terms of runtime. PSO was chosen for this assignment because of its wellknown balance of intensification and diversity, both of which are crucial in this context (constraint satisfaction). Next, the key idea of PSO is given for solving our constraint problem.

Initialization of Population:
We randomly generate the initial population for particles making use of the solution space.
Particle Updating: Let us consider the set of particles P, then there exists a vector X for position such that X ti ¼ x i1 x i2 x i3 …x in ð Þ T and we say that the velocity vector is V ti ¼ v i1 v i2 v i3 …v in ð Þ T at iteration t for each particle i that it is composed of. The particles themselves are able to update their own positions using velocity formula given in Equations 7 and 8. and where i ¼ 1, 2,…, j P j.
Here we can say that individual particles may be able to update positions making use of Equation 7, where two factors are given, c 1 as well as c 2 , that can respectively be shown to have contribution in particle movement for each and every iteration. Here, let p t be the position of the best particle iteration t. Furthermore, let us define p Ã as the position of the best particle overall iteration. Equation 8 may be used in the creation of updates in particle position. Let the parameter w be positive as well as constant, and say that w can be used for balancing the global search, which is known as exploration, as well as the local search, known simply as exploitation.
Fitness computing: It is used to evaluate the solution S, where S is the number of constraints met. If a constraint does not contradict the data exchanged between the sensors, it is met. The objective is to maximize this value as much as possible. The following is the fitness function: where satisfied C i ð Þ) = 1 if C i is satisfied, 0, otherwise.

| PERFORMANCE EVALUATION
In this section, we compare the proposed HCIF and compare it with state-of-the-art solutions. In particular, the following two case studies have been analysed: 1. Smart building: We used the NESTA162-bus Data (Thams et al., 2019) which is a set of data that contains N À 1 possible contingencies that represent all plausible operating points for any given energy demand profile. This set contains over 1 million points. In this dataset, topology changes are also included for N À 1 investigations. The aim is to identify fault diagnosis in electric power systems for smart building applications using object detection. For evaluating the objects detected, accuracy represented by mean average precision (mAP) and computational runtime are used. mAP can be considered a well-known metric for the evaluation of object detection systems, which is the sum of precision of all objects detected over the number of all objects detected, as shown in Equation 10: where Precision i ð Þ is defined as precision at rank i, and n is defined as all detected objects. In other words, the first i ranked objects can be considered while all other remaining objects can be ignored.We compare HCIF with the clustering particle swarm optimization for object detection (CPOD) , and fast region convolution neural network (FastRCNN) (Girshick, 2015) which are state of the art detection solutions.
2. Intelligent transportation: For ITS, we make use of data from ECML PKDD 2015 available at Alqahtani et al., 2020. This database is time-series based containing actual trajectories from over 400 taxis in Porto, Portugal in the span of 2013-2014. The CSV file contains over 3 GB. Each row defined a singular trip, which includes TripID, TaxiID, CallType, as well as timestamps and GPS information. The CSV file also contains a singular coordinates pair for every trip for every 15 s of that particular trip. The last item in each row is the logged destination for the journey, and the first item logged is the beginning of the journey. The aim of this dataset was originally to find local as well as global taxi trajectory outliers. As a generality, a well-known common issue in anomaly detection is in the actual evaluation procedure itself. This will be the case as well in new IoT environments such as AIoT, where the baseline is usually unknown due to the newness of the environment. To help with quantitative evaluation, in this paper, we make use of the methodology implemented with (Belhadi et al., 2020) where synthetic anomalous patterns are injected.
Injecting local outliers: local outliers are generated by adding noise several times with a certain probability p $ U 0:8, 1:0 ð Þand a given threshold μ; Injecting global outliers: From the local outliers, we again add noise but now only a few times with a certain probability p $ U 0:0, 1:0 ð Þand a given μ.
For both injections, each data d i in each dataset is changed as follows: The ratio of corrected returned outliers to all outliers is used to do the evaluation. This number is between 0 and 1, with a larger value indicating greater accuracy. We compare HCIF with the collective abnormal of human behaviours (CaHB) (Belhadi, Djenouri, Srivastava, et al., 2021), and group trajectory outliers (GTO) (Belhadi et al., 2020) which are the state-of-the-art anomaly detection solutions.
Our experimental evaluation has been performed on a personal computer with a 64-bit core i7 processor running Windows 10 and 16 GB of RAM. The host CPU is a 64-bit quad-core Intel Xeon E5520 running at 2.27 GHz. The GPU is an Nvidia Tesla C2075 with 448 CUDA cores for 448 (14 multiprocessors with 32 cores each) and a clock speed of 1.15 GHz. It has a global memory capacity of 2.8 GB, a shared memory capacity of 49.15 KB, and a warp size of 32. Single precision is utilized on both the CPU and GPU. In our implementation scenario, we employed GPU blocks to replicate the environment of a multi-agent system. Each agent is assigned to a single GPU block, with the associated agent sharing the block's shared memory. The communication between agents is accomplished through the use of the GPU host's global and constant memory.

| Smart building
Our experimental evaluation in this subsection aims to validate the usability of HCIF in smart building applications. Figure 2 presents both runtime and accuracy of HCIF compared to the two baseline algorithms FastRCNN (Girshick, 2015), and CPOD  using smart building dataset while varying the number of points and images. As shown in the first and third plots of Figure 1, the accuracy of the HCIF on NESTA162-bus database is compared with FastRCNN (Girshick, 2015), and CPOD . Making adjustments to the points as well as images for input, HCIF can be seen to outperform both baseline algorithms all three comparatives (accuracy, runtime, and mAP). This is explained by the fact that HCIF uses efficient strategies to identify objects by incorporating both the convolution neural network and the multiagents system, the learning is highly optimized using particle swarm optimization and constraint satisfaction programming. The second, and the fourth plots of the Figure 2 the runtime of the HCIF on NESTA162-bus database is compared with FastRCNN (Girshick, 2015), and CPOD . Causing variations in image numbers from 2000 to 7000, HCIF can outperform both baseline algorithms in just runtime.
We can give meaning to this by stating that HCIF prunes the bounding boxes by using intelligent communication among the different agents in the framework efficiently.

| Intelligent transportation
Our experimental evaluation in this subsection aims to validate the usability of HCIF in intelligent transportation applications. Figure 3 presents both runtime and accuracy of HCIF compared to the two baseline algorithms CaHB (Belhadi, Djenouri, Srivastava, et al., 2021), and GTO (Belhadi et al., 2020) using intelligent transportation dataset while varying the number of points and trajectories. As shown in the first and the third plots of Figure 3, the accuracy of the HCIF on ECML PKDD database, is compared with CaHB (Belhadi, Djenouri, Srivastava, et al., 2021), and GTO (Belhadi et al., 2020). By varying with the number of points and trajectories used as input, HCIF can be seen to outperform both baseline algorithms in just accuracy, shown using the ratio of the corrected outliers. This is explained by the fact that HCIF uses efficient strategies to detect both local and global anomalous by incorporating both the recurrent neural network and the multi-agents system. The second, and the fourth plots of the Figure 3, the runtime of HCIF on ECML PKDD database is compared with CaHB (Belhadi, Djenouri, Srivastava, et al., 2021), and GTO (Belhadi et al., 2020). Giving variations in trajectory numbers from 1500 to 5000, HCIF can be seen to outperform both baseline algorithms in just runtime. We can give meaning to this by stating that HCIF efficiently explores the solution space of both local and global outliers. Indeed, the local outlier is determined on each agent, where robust communication among agents allows the identification of global outliers.

| DISCUSSIONS AND FUTURE PERSPECTIVES
The main results of using a hybrid framework to handle smart sensor data from AIoT settings are discussed in this section. F I G U R E 2 Performance comparison of the hybrid collaborative intelligent framework and the state-of-the-art object detection solutions in the context of smart building application by varying the number of points, and the number of images computation meet the artificial intelligence internet of things for handling complex smart sensors data. Further findings of this work are that the multi-agents system benefits from the ontology matching process. Thus, each agent can communicate with the other agents through the set of common instances and properties derived by the ontology matching process. Finally, our framework is designed generically and as such can easily be applied to differing AIoT applications, which requires intelligent collaboration among different complex entities. This scenario reflects well the current needs in the industrial environments. Instead of the positive outcomes of this paper, future work may include: 1. Learning step improvements. Convolution neural networks and recurrent neural networks have been used as deep learning architectures. Additional deep learning models may be investigated for reducing the error-learning rate. Therefore, a topic of interest is the integrating of other deep learning models into the proposed framework in the future, such as transfer learning , active learning (Deng et al., 2020), and reinforcement learning (Chen, Cui, et al., 2020). We also could potentially look at mechanisms for the automation of the tuning procedure for hyper parameters in deep-learning models. Forcing the framework to take several iterations to find the best hyper-parameters seems wasteful. Using some sort of knowledge base that can then be studied based on correlative effects would be an interesting future direction.
2. Runtime processing improvements. For a fast framework performance boost as well as allow its application to larger problems through the exploitation of HP-computing (high-performance computing) tools like supercomputers (Chen et al., 2018) and techniques like cluster computing (Zhang et al., 2017). With this in mind, our aim as researchers is to identify an independent job in every cluster of sensor data by keeping in mind the HP computing challenges that may occur which include communication, load balancing, synchronization, as well as memory management. Furthermore, making use of methodologies to handle load balancing becomes very important. A known way to handle this well-known F I G U R E 3 Performance comparison of the hybrid collaborative intelligent framework and the state-of-the-art outlier detection solutions in the context of intelligent transportation application by varying the number of points, and the number of trajectories issue is through the development of decomposition strategies that can allow the location of equitable clusters concerning the number of smart sensors identified per cluster.
3. Use cases. This paper already presented four use cases of the proposed framework, namely smart building, and intelligent transportation. Taking into account the positive results, further work in domain-specific problems that present big data may be of interest. Some examples include business intelligence applications or the studying of business-specific financial data. For automated trading, runtime performance would be critical since trading can be very time-dependent.

| CONCLUSION
This paper proposes a novel intelligent framework for the collaborative IoT by combining different components. Deep learning architectures such as convolution neural networks and recurrent neural networks are used to learn from the different smart sensors' data. A multi-agent system is used to simulate the collective behaviours of the smart sensors in the internet of things environment. The communication among the different agents is assured by developing a new strategy for the ontology matching process. Different optimizations based on constraint satisfaction and meta-heuristics are also investigated. Our proposed framework is directly compared to existing artificial intelligence internet of things technologies. Our results are strong and show clearly that our proposed framework has good performance compared to the baseline solutions. With respect to future perspectives, we plan to improve the learning process by adding additional deep learning models by using transfer learning, and active learning. Another perspective of this research work is the ability to handle big data using high-performance computing. Exploring other use cases such as business intelligence is also on our future agenda.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are openly available in NESTA162-bus Data at https://arxiv.org/abs/1806.01074, reference number 42.