Internet of Things sensing device reliability assessment method based on graph neural network and multi-task learning

Abstract

The invention discloses an Internet of Things sensing device reliability assessment method based on a graph neural network and multi-task learning. The method specifically comprises the following steps: data collection and preprocessing; multi-level model design is carried out, the reliability of the multi-level model is evaluated, and the multi-level model design comprises sensor level design, component level design, board card level design and complete machine level design; carrying out multi-level network fusion; and through particle swarm optimization of multi-level model parameters, the reliability evaluation performance is improved. According to the method, by constructing the multi-level full-quantity mapping relation, the accuracy of reliability evaluation between the levels is improved, the overall fault trend of the equipment is accurately predicted, then the stability of the overall system is enhanced, and the reliability of the equipment is more accurately evaluated especially under the complex electromagnetic environment and environmental stress; the method is very important for operation and maintenance management of the digital converter station, can recognize potential faults in advance, optimizes a maintenance plan, and avoids shutdown and economic loss caused by sudden faults.

Description

Technical Field

[0001] The present application relates to the technical field of the Internet of Things in the power system, and particularly to a reliability evaluation method for an Internet of Things sensing device based on a graph neural network and multi-task learning. Background Art

[0002] The application of reliability evaluation of Internet of Things sensing devices in digital converter stations is of great significance. As the hub node of a new type of power system, the stability and reliability of the Internet of Things sensing network in a digital converter station directly affect the intelligent management of the power system and the efficiency of equipment fault detection. Due to the complex electromagnetic environment where the converter station is located and being affected by working conditions such as high and low temperatures, salt spray corrosion, and alternating damp heat, Internet of Things sensing devices often face interference from many uncertain factors, which will affect the communication efficiency and data accuracy of the devices, and thus reduce the operating efficiency and service quality of the system. Therefore, relying on advanced Internet of Things, edge computing, and artificial intelligence technologies to establish a set of equipment status evaluation methods for complex environments can not only discover potential device failures in advance and improve the accuracy of fault prediction, but also ensure the stable operation of digital converter stations through real-time monitoring and intelligent analysis of device status, and promote the digital transformation and intelligent upgrade of device management.

[0003] Existing reliability evaluation methods for Internet of Things sensing devices mostly focus on single-level evaluation or analysis based on expert experience and rules, and it is difficult to comprehensively consider the interdependent relationships between different levels of devices and the influence of complex environmental factors. Single-level reliability evaluation methods usually only focus on the reliability of a certain level of the device (such as the whole machine, components, modules, etc.), while ignoring the mutual influence and dependence between different levels. For example, the degradation of sensing materials may cause a decline in the performance of components, ultimately affecting the reliability of the whole machine. Empirical and rule-driven methods, such as fault tree analysis (FTA) and failure mode and effects analysis (FMEA), although they can provide an intuitive analysis of device fault modes, they rely heavily on expert experience and assumptions and cannot handle new or unknown fault modes; moreover, with the development of device technology and the change of the use environment, the original empirical rules often cannot adapt to new fault situations, thus reducing the accuracy and flexibility of evaluation. In addition, fault tree analysis and failure mode and effects analysis usually lack in-depth consideration of the multi-level coupling relationship of devices and are difficult to provide accurate predictions in complex working environments. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a reliability evaluation method for Internet of Things perception devices based on graph neural network and multi-task learning. This evaluation method comprehensively evaluates the reliability of devices by establishing a multi-level full-scale mapping relationship, combining graph neural network and multi-task learning, and accurately models the mutual dependence and influence between different levels. Through the fusion and optimization of multi-source heterogeneous data, it can effectively cope with the interference brought by complex environments, provide a more accurate and comprehensive reliability evaluation, and at the same time have the ability of adaptive optimization, showing significant advantages in device management and fault prediction, greatly improving the stability and intelligence level of the system.

[0005] To solve the above technical problems, the present invention is implemented in the following ways:

[0006] The reliability evaluation method for Internet of Things perception devices based on graph neural network and multi-task learning specifically includes the following steps:

[0007] S1. Data collection and preprocessing;

[0008] S2. Multi-level model design, evaluating the reliability of the multi-level model, and the multi-level model design includes sensor level design, component level design, board level design and whole machine level design;

[0009] S3. Multi-level network fusion;

[0010] S4. Optimize the parameters of the multi-level model through particle swarm optimization to improve the reliability evaluation performance.

[0011] Further, the specific method of step S1 is as follows:

[0012] Collect relevant data from the sensors, components, boards and whole machines of the Internet of Things perception devices. The data includes fault data, operating status data and environmental data, and clean and preprocess the data to remove outliers and fill in missing data to ensure the consistency and integrity of the data.

[0013] Further, the specific method of sensor level design in step S2 is as follows:

[0014] The reliability evaluation of sensor materials adopts a model of convolutional neural network (CNN). The model target is to output the reliability D1 of sensor materials. The input is the multi-dimensional features of sensor materials including physical properties and environmental data, and this feature constitutes the input data N represents the number of sample data, M represents the feature dimension of each sample data, and the reliability of sensor materials is obtained through convolutional neural network. The specific expression is as follows:

[0015] D1 = f1(X1)

[0016] Among them, f1(·) represents the convolutional neural network model, and D1 represents the reliability of the sensor material, which is a value between [0,1];

[0017] The calculation process of the convolutional neural network model includes the following steps:

[0018] 1) Convolutional layer

[0019] The convolutional layer uses convolutional weights K represents the size of the convolutional kernel. The expression of the output of the convolutional layer is as follows:

[0020] Z1 = X1 · W1 + b1

[0021] Among them, Z1 represents the output of the convolutional layer, and b1 represents the bias term;

[0022] 2) Activation function

[0023] The output of the convolutional layer passes through the ReLU activation function, and its expression is as follows:

[0024] A1 = ReLU(Z1) = max(0, Z1)

[0025] The ReLU activation function changes all negative values to 0 to increase the nonlinearity of the network;

[0026] 3) Pooling layer

[0027] The pooling layer selects the maximum value in the local area. The pooling operation uses a 2×2 window and a stride of 2. The expression of the output of the pooling layer is as follows:

[0028] P1 = MaxPool(A1)

[0029] Among them, P1 represents the output of the pooling layer. The pooling layer reduces the data dimension and improves the calculation efficiency;

[0030] 4) Fully connected layer

[0031] The pooled data is flattened and input into the fully connected layer. The weight of the fully connected layer is W2. The expression of the output of the fully connected layer is as follows:

[0032] F1 = P1 · W2 + b2

[0033] Among them, F1 represents the output of the fully connected layer, and b2 represents the bias term;

[0034] 5) Output layer

[0035] The output layer of the convolutional neural network model is processed through the Sigmoid activation function to obtain the reliability D1. Its specific expression is as follows:

[0036]

[0037] Among them, σ represents the Sigmoid activation function.

[0038] Furthermore, the specific method for component-level design in step S2 is as follows:

[0039] The component level is modeled using a graph neural network (GNN). The model objective is to output the reliability D2 at the component level, and the input is the input data X2. The reliability at the component level is obtained through the graph neural network, and the specific expression is as follows:

[0040] D2 = g2(X2)

[0041] Among them, g2 represents the function obtained through the GNN model;

[0042] The nodes in the graph are components, and the edges are the interdependent relationships between components. The graph is represented as G = (V, E), where V = {v1, v2,..., v n} represents the set of nodes, and E = {e1, e2,..., e m} represents the set of edges. The feature vector i of each node v represents the feature vector of the l-th layer node; The graph convolutional layer updates the feature vector of each node through the weighted average of neighbor node information, and the specific expression is as follows:

[0043]

[0044] Among them, represents the set of neighbor nodes of node v i , W( l ) represents the learning parameter of the l-th layer, and b( l ) represents the bias term; σ represents the activation function;

[0045] The update process is iterated, and finally the node feature vector is used to predict the component reliability D2, and its specific expression is as follows:

[0046]

[0047] Among them, φ(·) represents a subsequent neural network layer for mapping the node features to the reliability score.

[0048] Furthermore, the specific method for board-level design in step S2 is as follows:

[0049] The board level is jointly trained using a multi-task learning (MTL) model. The model objective is to output the reliability D3 at the board level, and the input is the input data X3. The reliability at the board level is obtained through the multi-task learning model, and the specific expression is as follows:

[0050]

[0051] Among them, f k (·) represents the k-th task model of the MTL model, and θ k represents the model parameters of the k-th task;

[0052] The multi-task set T at the board level = {T1, T2,..., T K}}, and each task T k corresponds to an objective function. The output of the model is the multi-task prediction result, and the specific expression is as follows:

[0053]

[0054] In the MTL model, the parameters are divided into shared parameters and task-specific parameters. The shared parameters are used to capture the commonalities of all tasks, and the task-specific parameters are used to capture the unique features of each task. The output expression of each task is as follows:

[0055] f k = W shared X3 + W k X3

[0056] Among them, W shared represents the shared parameter, and W k represents the task-specific parameter;

[0057] The goal of the MTL model is to minimize the weighted sum of the multi-task loss functions, and the expression is as follows:

[0058]

[0059] Among them, L k represents the loss function of the k-th task, λ k represents the weight of task T k , and Y k represents the true label.

[0060] Furthermore, the specific method for the design at the whole machine level in step S2 is as follows:

[0061] The whole machine level adopts a reinforcement learning (RL) model. Considering the reliability D4 of the whole machine under different operating conditions and environments, the model is trained through deep Q-learning to explore the optimal reliability evaluation under various operating conditions. The goal of the model is to output the reliability D4 at the whole machine level. Among them, the reinforcement learning model is a Markov decision process, denoted as a five-tuple (S, A, P, R, γ).

[0062] The state space S represents the operating state of the whole machine, and s t ∈ S is the operating state at time t;

[0063] The action space A represents the actions at the whole machine level, including adjusting the workload, modifying the maintenance interval, changing the working mode, etc. t a ∈ A represents the action at time t;

[0064] The state transition probability P(s'|s,a) represents the probability distribution of the change of the whole machine operation state from s to s' when the action a is executed;

[0065] The reward function R(s t ,a t ) represents the reward function obtained by executing the action, which is related to the device health state:

[0066] R(s t ,a t ) = -C failure -α r ·C maintenance +βr·D4

[0067] Among them, C failure represents the cost (high value) brought by device failure, C maintenance represents the cost of maintenance actions, α r , β r respectively represent the weight parameters of the reward function, which are used to balance the failure, maintenance cost and reliability;

[0068] The discount factor γ represents the influence weight of controlling future rewards, and the range is 0 < γ < 1;

[0069] The goal of the reinforcement learning model is to find an optimal policy π * (a|s), so that the long-term cumulative reward is maximized, and the specific expression is as follows:

[0070]

[0071] The expression of the state-action value function Q(s,a) is as follows:

[0072]

[0073] Among them, represents the expectation function, Q(s,a) represents the expected reliability after executing the action a in the state s, a' represents the next action that can be taken, and s' represents the system state after executing the current action a;

[0074] The reliability D4 at the whole machine level is optimized through deep Q learning, and the state-action value function Q(s,a) is updated as:

[0075]

[0076] Among them, α represents the learning rate;

[0077] After training is completed, select the optimal action a in the given state s * , and the expression is as follows:

[0078]

[0079] Execute the optimal action a * Obtain the corresponding reliability of the whole machine level, and the expression is as follows:

[0080] D4 = Q(s, a * ).

[0081] Furthermore, the specific method of step S3 is as follows:

[0082] Fuse the reliability models of the sensor level, component level, board level, and whole machine level in step S2 to construct a multi-level network. The multi-level network fusion establishes a reliability mapping relationship between levels, enabling the overall reliability of the device to not only be based on the individual evaluation results of each level but also consider the interactive effects between levels, thereby improving the overall evaluation effect;

[0083] Use a multi-layer perceptron to fuse the reliabilities D1, D2, D3, and D4 of different levels. The process expression is as follows:

[0084] h′1 = σ(W1[D1, D2, D3, D4] + b1)

[0085] h′2 = σ(W2h′1 + b2)

[0086] D′ = σ(W3h′2 + b3)

[0087] Among them, D' represents the overall reliability and is used as the final evaluation result. W1, W2, and W3 represent weight matrices, and b1, b2, and b3 represent bias terms.

[0088] Furthermore, the specific method of step S4 is as follows:

[0089] Use the particle swarm optimization method (PSO, Particle Swarm Optimization) to optimize the parameters of the multi-level network. PSO can help search for the optimal network weights and biases to minimize the final evaluation error. The specific expression is as follows:

[0090] min L(D, D actual )

[0091] Among them, L represents the loss function using the mean square error, and D actual represents the actual reliability;

[0092] The particle swarm optimization method searches for the optimal solution by adjusting the particle positions and velocities. The particle swarm optimization process is as follows:

[0093] Set the number of particle swarms P. Each particle represents the parameter set Θ = {W1, W2, W3, b1, b2, b3} of the model; calculate the loss value L under the current parameters of each particle, and update the particle velocity and position expressions as follows:

[0094] v i,t+1 = ωv i，t + c1r1(p best,i - Θ i ) + c2r2(g best - Θ i )

[0095] Θ i，t+1 = Θ i，t + v it+1

[0096] Among them, v i,t represents the velocity of particle i in the t-th generation, ω represents the inertia weight, which is used to balance local search and global search, c1 and c2 represent learning factors, which control the influence of individual and global optimal positions, r1 and r2 represent random numbers to ensure the particle search ability, p best,i represents the best solution of particle i so far, and g best represents the global optimal solution;

[0097] If the loss function L converges or reaches the maximum number of iterations, stop the optimization to obtain the optimal model parameters.

[0098] Compared with the prior art, the beneficial effects of the present invention are:

[0099] The present invention improves the accuracy of reliability evaluation between levels by constructing a multi-level full-scale mapping relationship, accurately predicts the overall fault trend of the device, and then enhances the stability of the overall system. Especially in complex electromagnetic environments and environmental stresses, the reliability of the device can be more accurately evaluated; it is crucial for the operation and maintenance management of digital converter stations, can identify potential faults in advance, optimize maintenance plans, and avoid outages and economic losses caused by sudden faults.

[0100] By using technologies such as deep learning, graph neural networks, multi-task learning, and reinforcement learning for modeling, the system can extract effective features from a large amount of data, perform accurate prediction and analysis, and has stronger learning ability and generalization ability; by combining particle swarm optimization and reinforcement learning, not only the parameters of each level model are optimized, but also the evaluation strategy is dynamically adjusted according to the changes in the device state, improving the adaptive ability of the system.

[0101] Meanwhile, it can effectively process various heterogeneous data from different levels and different sources. Through comprehensive analysis, it improves the utilization efficiency of data and reduces the complexity of data preprocessing. Brief Description of the Drawings

[0102] Figure 1 It is a schematic flow chart of the reliability evaluation method of the present invention. Detailed Embodiment

[0103] The following further elaborates on the detailed embodiment of the present invention in conjunction with the drawings and specific embodiments.

[0104] The present invention evaluates the reliability of the IoT sensing device under complex environmental conditions through a multi-level model. The core is to divide the reliability of the device into four levels: sensing material, component, board, and whole machine. Through technologies such as graph neural network and multi-task learning, a reliability mapping relationship between each level is established to achieve accurate evaluation of the overall reliability of the device; at the same time, particle swarm optimization is combined to optimize the model parameters to improve the accuracy and robustness of the evaluation model.

[0105] As Figure 1 shown, the reliability evaluation method of the IoT sensing device based on graph neural network and multi-task learning specifically includes the following steps:

[0106] S1. Data collection and preprocessing, the specific method is as follows:

[0107] Collect relevant data from the sensors, components, boards, and whole machines of the IoT sensing device. The data includes fault data, operating status data, and environmental data, and clean and preprocess the data to remove outliers and fill in missing data to ensure the consistency and integrity of the data.

[0108] S2. Multi-level model design, evaluate the reliability of the multi-level model. The multi-level model design includes sensor level design, component level design, board level design, and whole machine level design. The specific method of the sensor level design is as follows:

[0109] The reliability evaluation of the sensor material adopts a model of convolutional neural network (CNN). The model objective is to output the reliability D1 of the sensor material. The input is the multi-dimensional features of the sensor material including physical properties (such as hardness, conductivity, temperature, pressure, etc.) and environmental data (such as humidity, temperature, radiation, etc.). This feature constitutes the input data N represents the number of sample data, and M represents the feature dimension of each sample data. The reliability of the sensor material is obtained through the convolutional neural network. The specific expression is as follows:

[0110] D1 = f1(X1)

[0111] Among them, f1(·) represents the convolutional neural network model, D1 represents the reliability of the sensor material, which is a numerical value between [0,1], indicating the reliability of the material in the current environment;

[0112] The calculation process of the convolutional neural network model includes the following steps:

[0113] 1) Convolutional layer

[0114] The convolutional layer uses convolutional weights K represents the size of the convolutional kernel. The expression of the output of the convolutional layer is as follows:

[0115] Z1 = X1 · W1 + b1

[0116] Among them, Z1 represents the output of the convolutional layer, and b1 represents the bias term;

[0117] 2) Activation function

[0118] The output of the convolutional layer passes through the ReLU activation function, and its expression is as follows:

[0119] A1 = ReLU(Z1) = max(0, Z1)

[0120] The ReLU activation function changes all negative values to 0 to increase the non-linearity of the network;

[0121] 3) Pooling layer

[0122] The pooling layer selects the maximum value in the local area. The pooling operation uses a 2×2 window with a stride of 2. The expression of the output of the pooling layer is as follows:

[0123] P1 = MaxPool(A1)

[0124] Among them, P1 represents the output of the pooling layer. The pooling layer reduces the data dimension and improves the calculation efficiency;

[0125] 4) Fully connected layer

[0126] The pooled data is flattened and input into the fully connected layer. The weight of the fully connected layer is W2. The expression of the output of the fully connected layer is as follows:

[0127] F1 = P1 · W2 + b2

[0128] Among them, F1 represents the output of the fully connected layer, and b2 represents the bias term;

[0129] 5) Output layer

[0130] The output layer of the convolutional neural network model is processed through the Sigmoid activation function to obtain the reliability D1. Its specific expression is as follows:

[0131]

[0132] Among them, σ represents the Sigmoid activation function.

[0133] The specific method for component-level design is as follows:

[0134] The graph neural network (GNN) is a deep learning model that can effectively process graph-structured data and is suitable for describing the interdependent relationships between component levels. In the component level of the Internet of Things sensing device, components usually form a graph structure through relationships such as electrical connections and signal dependencies, and these relationships have an important impact on the reliability of components. The component level is modeled using the graph neural network (GNN). The model objective is to output the reliability D2 of the component level, and the input is the input data X2. The reliability of the component level is obtained through the graph neural network, and the specific expression is as follows:

[0135] D2 = g2(X2)

[0136] Among them, g2 represents the function obtained through the GNN model;

[0137] The core of the graph neural network is to update the representation of each node through information propagation of nodes and edges. The update of a node is a process of weighted summation based on the information of its neighbor nodes. The graph nodes in the graph structure are components, and the edges are the interdependent relationships between components. The graph is represented as G = (V, E), where V = {v1, v2,..., v n} represents the set of nodes, and E = {e1, e2,..., e m} represents the set of edges. The feature vector i of each node v represents the feature vector of the node at the l-th layer; the graph convolutional layer updates the feature vector of each node through weighted averaging of neighbor node information, and the specific expression is as follows:

[0138]

[0139] Among them, represents the set of neighbor nodes of node v i , W (l) represents the learning parameter of the l-th layer, and b (l) represents the bias term; σ represents the activation function;

[0140] The update process is iterated, and finally the node feature vector is used to predict the component reliability D2, and its specific expression is as follows:

[0141]

[0142] Among them, φ(·) represents a subsequent neural network layer used to map the node features to the reliability score.

[0143] The specific method for the board-level design is as follows:

[0144] The board-level uses a multi-task learning (MTL) model for joint training. The model objective is to output the reliability D3 at the board level, and the input is the input data X3. The board-level reliability is obtained through the multi-task learning model, and the specific expression is as follows:

[0145]

[0146] where f k (·) represents the k-th task model of the MTL model, and θ k represents the model parameters of the k-th task;

[0147] The multi-task set T at the board level = {T1, T2, …, T K} Each task T k corresponds to an objective function, and the output of the model is the multi-task prediction result. The specific expression is as follows:

[0148]

[0149] In the MTL model, the parameters are divided into shared parameters and task-specific parameters. The shared parameters are used to capture the commonalities of all tasks, and the task-specific parameters are used to capture the unique features of each task. The output expression of each task is as follows:

[0150] f k = W shared X3 + W k X3

[0151] where W shared represents the shared parameters, and W k represents the task-specific parameters;

[0152] The objective of the MTL model is to minimize the weighted sum of the multi-task loss functions. The expression is as follows:

[0153]

[0154] where L k represents the loss function of the k-th task, λ k represents the weight of task T k (set to 1 / k), and Y k represents the true label. By jointly optimizing the loss functions of multiple tasks, the MTL effectively improves the accuracy of board-level reliability prediction.

[0155] The specific method for the whole-machine level design is as follows:

[0156] At the whole-machine level, a reinforcement learning (RL) model is adopted to consider the reliability D4 of the whole machine under different operating conditions and environments. The model is trained through deep Q-learning to explore the optimal reliability evaluation under various operating conditions. The goal of the model is to output the reliability D4 at the whole-machine level. The reinforcement learning model is a Markov decision process, denoted as a five-tuple (S, A, P, R, γ).

[0157] The state space S represents the operating state of the whole machine, and s t ∈S is the operating state at time t;

[0158] The action space A represents the actions at the whole-machine level, including adjusting the workload, modifying the maintenance interval, changing the working mode, etc. And a t ∈A is the action at time t;

[0159] The state transition probability P(s'|s, a) represents the probability distribution of the operating state of the whole machine changing from s to s' when the action a is executed;

[0160] The reward function R(s t , a t ) represents the reward function obtained by executing the action, which is related to the device health state:

[0161] R(s t , a t ) = -C failure -α r ·C maintenance +β r ·D4

[0162] Where C failure represents the cost (high value) brought by device failure, C maintenance represents the cost of maintenance actions, and α r , β r represent the weight parameters of the reward function respectively, used to balance the failure, maintenance cost and reliability;

[0163] The discount factor γ represents the influence weight of controlling future rewards, and the range is 0 < γ < 1;

[0164] The goal of the reinforcement learning model is to find an optimal policy π * (a|s), so as to maximize the long-term cumulative reward. The specific expression is as follows:

[0165]

[0166] The expression of the state-action value function Q(s, a) is as follows:

[0167]

[0168] Where Let \(Q(s,a)\) denote the expected reliability after executing action \(a\) in state \(s\), \(a'\) denote the next available action, and \(s'\) denote the system state after executing the current action \(a\);

[0169] The reliability \(D4\) at the whole - machine level is optimized through deep Q - learning, and the state - action value function \(Q(s,a)\) is updated as follows:

[0170]

[0171] where \(\alpha\) represents the learning rate;

[0172] After training, select the optimal action \(a\) in the given state \(s\) * , and the expression is as follows:

[0173]

[0174] Execute the optimal action \(a\) * Obtain the corresponding reliability at the whole - machine level, and the expression is as follows:

[0175] D4 = Q(s,a * ).

[0176] S3. Multilevel network fusion, and the specific method is as follows:

[0177] Fuse the reliability models of the sensor level, component level, board - card level, and whole - machine level in step S2 to construct a multilevel network. The multilevel network fusion establishes a reliability mapping relationship between levels, enabling the overall reliability of the device to not only be based on the individual evaluation results of each level but also consider the interactive effects between levels, thereby improving the overall evaluation effect;

[0178] Adopt a multi - layer perceptron to fuse the reliabilities \(D1\), \(D2\), \(D3\), and \(D4\) at different levels. The process expression is as follows:

[0179] h′1 = σ(W1[D1,D2,D3,D4]+b1)

[0180] h′2 = σ(W2h′1+b2)

[0181] D′ = σ(W3h′2+b3)

[0182] where \(D'\) represents the overall reliability and is used as the final evaluation result, \(W1\), \(W2\), \(W3\) represent weight matrices, and \(b1\), \(b2\), \(b3\) represent bias terms.

[0183] S4. Optimize the multilevel model parameters through particle swarm optimization to improve the reliability evaluation performance, and the specific method is as follows:

[0184] Optimize the parameters of the multi-level network using the Particle Swarm Optimization (PSO) method. PSO can help search for the optimal network weights and biases to minimize the final evaluation error, and the specific expression is as follows:

[0185] min L(D′,D actual )

[0186] where L represents the loss function using the mean squared error, and D actual represents the actual reliability;

[0187] The Particle Swarm Optimization method searches for the optimal solution by adjusting the particle positions and velocities. The particle swarm optimization process is as follows:

[0188] Set the number of particle swarms P. Each particle represents the model parameter set Θ = {W1, W2, W3, b1, b2, b3}; calculate the loss value L under the current parameters of each particle, and update the particle velocity and position expressions as follows:

[0189] v i,t+1 = ωv i,t + c1r1(p best,i - Θ i ) + c2r2(g best - Θ i )

[0190] Θ i,t+1 = Θ i,t + v i,t+1

[0191] where v i,t represents the velocity of particle i in the t-th generation, ω represents the inertia weight used to balance local search and global search, c1 and c2 represent the learning factors that control the influence of the individual and global optimal positions, r1 and r2 represent random numbers to ensure the search ability of the particle, p best,i represents the best solution found by particle i so far, and g best represents the global optimal solution;

[0192] If the loss function L converges or reaches the maximum number of iterations, stop the optimization to obtain the optimal model parameters.

[0193] The above description is only the implementation manner of the present invention. Once again, for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements can still be made to the present invention, and these improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A reliability evaluation method for an IoT perception device based on a graph neural network and multi-task learning, characterized in that, Specifically, it includes the following steps: S1. Data collection and preprocessing; S2. Multilevel model design, and evaluate the reliability of the multilevel model. The multilevel model design includes sensor level design, component level design, board level design, and whole machine level design; S3. Multilevel network fusion; S4. Optimize the multilevel model parameters through particle swarm optimization to improve the reliability evaluation performance.

2. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of step S1 is as follows: Collect relevant data from the sensors, components, boards, and whole machines of the Internet of Things perception device. The data includes fault data, operating status data, and environmental data, and clean and preprocess the data to remove outliers and fill in missing data to ensure the consistency and integrity of the data.

3. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of sensor level design in step S2 is as follows: The reliability evaluation of the sensor material adopts a convolutional neural network model. The model aims to output the reliability D1 of the sensor material, and the input is the multi-dimensional features of the sensor material including physical attributes and environmental data, which constitute the input data N represents the number of sample data, and M represents the feature dimension of each sample data. The reliability of the sensor material is obtained through a convolutional neural network, and the specific expression is as follows: D1 = f1(X1) where f1(·) represents a convolutional neural network model, and D1 represents the reliability of the sensor material; The calculation process of the convolutional neural network model includes the following steps: 1) Convolutional layer The convolutional layer uses convolutional weights K represents the size of the convolutional kernel, and the expression for the output of the convolutional layer is as follows: Z1 = X1·W1 + b1 where Z1 represents the output of the convolutional layer, and b1 represents the bias term; 2) Activation function The output of the convolutional layer passes through the ReLU activation function, and its expression is as follows: A1 = ReLU(Z1) = max(0, Z1) The ReLU activation function changes all negative values to 0 to increase the non-linearity of the network; 3) Pooling layer The pooling layer selects the maximum value of the local area. The pooling operation uses a 2×2 window and a stride of 2. The expression of the pooling layer output is as follows: P1 = MaxPool(A1) where P1 represents the output of the pooling layer, and the pooling layer reduces the data dimension; 4) Fully connected layer The pooled data is flattened and input into the fully connected layer. The weight of the fully connected layer is W2, and the expression of the fully connected layer output is as follows: F1 = P1·W2 + b2 where F1 represents the output of the fully connected layer, and b2 represents the bias term; 5) Output layer The output layer of the convolutional neural network model is processed through the Sigmoid activation function to obtain the reliability D1, and its specific expression is as follows: where σ represents the Sigmoid activation function.

4. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of component level design in step S2 is as follows: The component level is modeled by a graph neural network. The model target is to output the reliability D2 of the component level. The input is the input data X2. The reliability of the component level is obtained through the graph neural network, and the specific expression is as follows: D2 = g2(X2) where g2 represents the function obtained through the GNN model; The nodes in the graph are components, and the edges are the interdependencies between components. The graph is represented by G = (V, E), where V = {v1, v2, …, v n } represents a node set, E = {e1, e2, …, e m } represents an edge set, each node v i The eigenvector of Represents the feature vector of the l-th layer node; the graph convolution layer updates the feature vector of each node through the weighted average of the neighbor node information. The specific expression is as follows: Among them, represents the set of neighbor nodes of node v i , W (l) represents the learning parameters of the l-th layer, and b (l) represents the bias term; σ represents the activation function; The update process is iterated, and the final node feature vector is used to predict the component reliability D2, and its specific expression is as follows: where φ(·) represents a subsequent neural network layer for mapping the node features to the reliability score.

5. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of the board level design in step S2 is as follows: The board level uses a multi-task learning model for joint training. The model target is to output the reliability D3 of the board level, and the input is the input data X3. The board level reliability is obtained through the multi-task learning model. The specific expression is as follows: Among them, f k (·) represents the k-th task model of the MTL model, and θ k represents the model parameters of the k-th task; The multi-task set T at the board level = {T1, T2, …, T K}, and each task T k corresponds to an objective function. The output of the model is the multi-task prediction result, and the specific expression is as follows: The parameters in the MTL model are divided into shared parameters and task-specific parameters. The shared parameters are used to capture the commonalities of all tasks, and the task-specific parameters are used to capture the unique features of each task. The output expression of each task is as follows: f k = W shared X3 + W k X3 Among them, W shared represents a shared parameter, and W k represents a task-specific parameter; The goal of the MTL model is to minimize the weighted sum of the multi-task loss functions. The expression is as follows: Among them, L k represents the loss function of the k-th task, and λ k represents the weight of task T k , and Y k represents the true label.

6. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of the whole machine level design in step S2 is as follows: The whole machine level uses a reinforcement learning model. The model target is to output the reliability D4 of the whole machine level. The reinforcement learning model is a Markov decision process, denoted as a five-tuple (S, A, P, R, γ). The state space S represents the operating state of the whole machine, and s t ∈ S is the operating state at time t; The action space A represents the actions at the whole machine level, including adjusting the workload, modifying the maintenance interval, and changing the working mode, where a t ∈ A represents the action at time t; The state transition probability P(s'|s,a) represents the probability distribution of the operating state of the whole machine changing from s to s' when the action a is executed. Reward function R(s t ,a t ) represents the reward function obtained by performing an action: R(s t ,a t )=-C failure -α r ·D maintenance +β r ·D4 Among them, C failure represents the cost brought by equipment failure, C maintenance represents the cost of maintenance actions, α r , β r respectively represent the weight parameters of the reward function, which are used to balance the failure, maintenance cost and reliability; The discount factor γ represents the influence weight of controlling future rewards, and the range is 0 < γ < 1. The goal of the reinforcement learning model is to find an optimal policy π * (a|s) that maximizes the long-term cumulative reward, and the specific expression is as follows: The state-action value function Q(s,a) is expressed as follows: Among them, represents the expected function, Q(s,a) represents the expected reliability after performing action a in state s, a' represents the action that can be taken next, and s' represents the system state after performing the current action a; The reliability D4 of the whole machine level is optimized through deep Q learning, and the state-action value function Q(s,a) is updated as: where α represents the learning rate. After training is completed, select the optimal action a in the given state s * , and the expression is as follows: Execute the optimal action a * Obtain the corresponding reliability at the whole-machine level, and the expression is as follows: D4 = Q(s,a * )。 7. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of step S3 is as follows: Fuse the reliability models of the sensor level, component level, board level and whole machine level in step S2 to construct a multi-level network, and establish a reliability mapping relationship between levels through the fusion of the multi-level network. Use a multi-layer perceptron to fuse the reliabilities D1, D2, D3 and D4 of different levels. The process expression is as follows: h′1 = σ(W1[D1,D2,D3,D4]+b1) h′2 = σ(W2h′1+b2) D′ = σ(W3h′2+b3) where D' represents the overall reliability and is used as the final evaluation result. W1, W2, W3 represent weight matrices, and b1, b2, b3 represent bias terms.

8. The reliability evaluation method of the Internet of Things perception device based on graph neural network and multi-task learning according to claim 1, characterized in that The specific method of step S4 is as follows: Use the particle swarm optimization method to optimize the parameters of the multi-level network to minimize the final evaluation error. The specific expression is as follows: min L(D′,D actual ) Among them, L represents the loss function using mean square error, and D actual represents the actual reliability; The particle swarm optimization method searches for the optimal solution by adjusting the particle positions and velocities. The particle swarm optimization process is as follows: Set the number of particle swarms \(P\), and each particle represents the parameter set of the model \(\Theta=\{W1, W2, W3, b1, b2, b3\}\); calculate the loss value \(L\) under the current parameters of each particle, and update the expressions of particle velocity and position as follows: v i,t+1 = ωv i,t + c1r1(p best,i - Θ i ) + c2r2(g best - Θ i ) Θ i,t+1 = Θ i,t + v i,t+1 Among them, v i,t represents the velocity of particle i in the t-th generation, ω represents the inertia weight, which is used to balance local search and global search, c1 and c2 represent the learning factors, which control the influence of the individual and global optimal positions, r1 and r2 represent random numbers that ensure the search ability of the particle, p best,i represents the best solution of particle i so far, and g best represents the global optimal solution; If the loss function \(L\) converges or reaches the maximum number of iterations, stop the optimization to obtain the optimal model parameters.

Images