A method for industrial process fault diagnosis based on impulse graph neural network

Abstract

The present invention discloses an industrial process fault diagnosis method based on a pulse graph neural network, and belongs to the technical field of industrial process fault diagnosis. The method includes: collecting process variables under different fault conditions in the industrial production process; performing data preprocessing and graph construction on the process variables, including data standardization, sliding window sampling to enhance time series correlation, and constructing graph data according to the K-nearest neighbor method; inputting the graph data into the pulse graph neural network for training, including using the graph convolution layer to complete graph feature extraction and using the Bernoulli function for pulse encoding and using the pulse neuron and the fully connected layer to learn the node features and train the model; using the trained model to predict and warn faults, detect real-time unknown fault category data, and when a fault occurs, provide alarm information to facilitate maintenance personnel to efficiently locate and handle it.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process fault diagnosis, and in particular to a fault diagnosis method based on a pulse graph neural network in an industrial production process. Background Art

[0002] Industrial processes play a vital role in modern society, covering multiple fields such as production, manufacturing, and energy. These processes often involve complex chemical, mechanical, and electrical systems, and their normal operation is crucial to production efficiency and product quality. Failures are inevitable in industrial processes, which may lead to production interruptions, resource waste, and equipment damage. Therefore, early detection and resolution of failures are essential to maintaining the reliability and stability of industrial processes.

[0003] Industrial process fault diagnosis technology can be generally divided into three categories: model analysis-based methods, knowledge-driven methods, and data-driven methods. Model analysis-based methods rely on in-depth understanding and modeling of process objects. They detect and diagnose faults by analyzing and comparing the dynamic characteristics of the system. Establishing a system model requires a clear understanding of the mechanism and structure of the detection object, which is challenging in actual industrial production processes. Knowledge-driven methods simulate the expert reasoning process by acquiring prior expert experience and knowledge to make judgments on fault models. This method is suitable for situations where the relationship between faults is clear and the decision logic is strong. For process objects, the coupling between variables is strong, and it is difficult to construct universal expert experience. Data-driven methods use intelligent optimization algorithms and machine learning techniques based on process data collected in real time by field sensors in industrial processes to effectively identify process fault patterns. Data-driven methods are widely used in the field of fault diagnosis due to their strong applicability, high feasibility, and excellent data mining capabilities. However, traditional data-driven methods have problems such as high data quality requirements, high computational complexity, and poor real-time performance.

[0004] Among the advanced data-driven methods, graph neural networks represent data as graph structures and perform operations and reasoning on the graphs, which can better capture the relationships and contextual information between data. They have been successfully applied to fields such as social network analysis, recommendation systems, and bioinformatics, and have also achieved important results in fault diagnosis. However, the computational complexity of graph neural networks is relatively high, especially for large-scale complex systems such as chemical processes, which require a lot of computing resources and time. In recent years, the third-generation artificial neural network, the spiking neural network, has received extensive attention and research. Inspired by the way the human brain transmits information, the spiking neural network is a bio-simulated neuron model that simulates the dynamic behavior of brain neurons. It has low energy consumption and high-efficiency information transmission characteristics, and has good application prospects in the field of fault diagnosis. Summary of the invention

[0005] The purpose of the present invention is to provide a fault diagnosis method based on pulse graph neural network to overcome the above technical challenges, so as to timely and accurately detect and diagnose faults occurring in industrial processes, while having the characteristics of low energy consumption and high efficiency, ensuring the stability of the chemical production process, reducing safety accidents and improving the economic benefits of the factory.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A fault diagnosis method based on impulse graph neural network comprises the following steps:

[0008] S1: Industrial production process data collection. Data collection is performed on different process variables that need to be monitored in normal and fault conditions in the chemical production process. In the case of multiple fault conditions, data under different fault conditions need to be distinguished. The selection of process variables needs to be comprehensive, covering all equipment involved in the process technology.

[0009] S2: Data preprocessing and graph data construction. The collected industrial production process variable data is fully preprocessed. The industrial process technology contains many process variables. All variables are standardized on their own scale and converted into a normal distribution between 0 and 1, so as to achieve the uniform scale of the variables. The standardized data is processed by sliding window sampling and data enhancement. Finally, the window data is used as the original data for graph construction. The KNN graph construction method is used to construct all nodes into several graphs.

[0010] S3: Input the graph data into the pulse graph neural network for learning. The core method of the present invention relies on a pulse graph neural network algorithm, which is mainly a neural network model with a high energy consumption advantage. It mainly uses a graph convolution layer to extract graph features and a Bernoulli function for pulse encoding, converting floating-point operations into binary addition operations to reduce computing energy consumption, and further learning through pulse neurons and fully connected layers. The pulse graph neural network is trained using data with different fault labels in the industrial production process to learn the fault characteristics of the data. After the training is completed, a high-precision multi-classification fault diagnosis model is obtained.

[0011] S4: Use the learned model to predict and warn about faults. Input the data of unknown faults into the trained neural network model, predict the fault type, and send out an alarm through the system to prompt maintenance personnel to deal with the fault in time.

[0012] Furthermore, industrial processes are large and small industrial processes that can collect process variable data of different equipment under different working conditions, not limited to a certain production process. Different working conditions include: normal working state, different fault conditions such as reactor cooling water temperature failure, condenser failure, feed abnormality, etc.

[0013] Furthermore, to pre-process the data under different working conditions, we first need to distinguish these different working condition data, that is, assign "labels" to the data. Then standardize the data with different labels. There are different numbers of process variables in industrial processes. Small process variables are within ten, and large processes have dozens or even hundreds of process variables. These variables have different scales. Standardize each variable to eliminate the impact of scale on training.

[0014] The standardized variables are sampled using sliding windows, and sliding window operations are performed using the set step size and window length rules to enhance the connection between variables at previous and next moments. The sampled windows are then used to construct graph data according to the node similarity rules. A window is regarded as a node, and the connection relationship between nodes is determined according to the graph construction rules. Data with different labels are constructed into many nodes, and then all nodes of all labels are used to construct graphs, thereby obtaining several graphs with node connections according to certain rules. The nodes in the graphs are classified to determine the fault category.

[0015] Furthermore, the industrial production process data is constructed into a graph and then input into the pulse graph neural network model to train the fault diagnosis model, which includes an input layer, a graph convolution layer, an activation function, a pulse neuron, a fully connected layer and an output layer.

[0016] Before training, the data set is divided into training, validation, and test sets in a 6:2:2 ratio.

[0017] Multiple samples are randomly sampled from the nodes in the training set to form a minimum batch input model.

[0018] Furthermore, the function of the graph convolution layer is to aggregate the neighbor information of the node, capture the data features, and learn the direct correlation of the fault. The industrial production process data after preprocessing exists in the form of a sliding window. A window is regarded as a node to construct the graph data. The graph data can be represented as G(V, A, X, E). The calculation formula of the graph convolution is:

[0019]

[0020] Formula (1) describes the calculation between layers, A represents the adjacency matrix, D represents the degree matrix, H (l) represents the node feature matrix of the lth layer, W is the learnable weight matrix, I is the identity matrix.

[0021] The activation function σ uses the ReLu activation function.

[0022] Furthermore, the convolutional industrial production process data node features will pass through the spiking neuron. The spiking neuron model is also called the leaky integral-flush (LIF) model. The LIF model is derived by analogizing the spiking neuron to a circuit model containing inductors and capacitors. The derivation process is as follows:

[0023] Assume that any time-varying current I(t) is input into the spiking neuron, which may be a pulse signal transmitted by electrical stimulation or neuron. Since the current in the circuit is conserved, we have:

[0024] I(t)=I R +I C (2)

[0025] Among them I R =u R / R refers to the current passing through the membrane resistance. From the perspective of biological theory, it represents the current generated by the movement of ions inside and outside the membrane, while u R =uu reset .u reset is the membrane potential threshold of the spiking neuron, u represents the input voltage of the neuron, and u R Indicates the pulse emission voltage.

[0026] I C is the current obtained by charging the capacitor, and its expression is:

[0027]

[0028] The total input I(t) of the circuit can be expressed as I(t) = (u(t) - u reset ) / R+Cdu(t) / d(t), introducing the time constant τ m =RC, the potential update mode of the LIF neuron of the industrial production process data can be expressed as:

[0029]

[0030] Furthermore, the industrial production process data is simply linearly transformed through the fully connected layer to further learn the data features.

[0031] Furthermore, in order to enable the node feature matrix obtained after graph convolution to be input into the pulse neuron in a suitable manner, the node features need to be pulse encoded. The present invention uses the Bernoulli function to encode the node features.

[0032] The pulse node features after pulse neuron processing enter the fully connected layer, and the pulse emission intensity of different nodes is obtained through linear change. The value of the corresponding category is determined by statistically analyzing the pulse emission intensity.

[0033] The output layer is encoded in one-hot mode (unique hot encoding), and the output is a vector. The vector dimension is the total number of fault classification categories.

[0034] The present invention is beneficial in that:

[0035] 1. The fault diagnosis method proposed in the present invention does not require the construction of an accurate mathematical model for the industrial process, does not involve system model identification, and does not require prior knowledge provided by expert experience as a fault diagnosis standard. The data-driven fault diagnosis task can be completed through the process variables collected at the chemical site, which has strong versatility and feasibility.

[0036] 2. The fault diagnosis method based on pulse neural network proposed in the present invention is a neural network with high bionic characteristics. Since the data processed by the pulse neural network is converted into binary data and the signal is transmitted in the form of pulses, a large amount of computing resource consumption can be saved. Implementing the model on a neuromorphic chip can save energy consumption and improve the efficiency of fault diagnosis, and it has advancedness and high energy efficiency.

[0037] 3. The fault diagnosis method proposed in the present invention does not require additional processing of industrial production process data or the addition of other processes. It only needs to directly collect the process variables collected by the field sensors under different fault conditions and normal conditions to complete the fault diagnosis of the industrial process.

[0038] 4. The fault diagnosis method based on pulse graph neural network proposed in the present invention is a universal industrial process fault diagnosis process, not only applicable to a specific process. It is applicable to process technologies that can collect process variable data under different working conditions and has good versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the fault diagnosis method of the present invention;

[0040] Figure 2 It is a visual schematic diagram of the fault diagnosis method process of the present invention;

[0041] Figure 3 It is a schematic diagram of the impulse graph neural network model described in the present invention. DETAILED DESCRIPTION

[0042] To make the purpose, technical solution and advantages of this application clearer. The technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Aiming at the fault risks in the chemical production process, the present invention proposes a fault diagnosis method based on pulse graph neural network. The main implementation process of the method is shown in the attached figure. Figure 1 As shown, the visualization steps of the process are as follows Figure 2 As shown below, according to Figure 1 Describe the entire process systematically; it is mainly divided into four steps. The process framework of this fault diagnosis can be described as follows:

[0044] First, the data set of the industrial process to be processed is collected, and then the collected original data set is preprocessed and graph data is constructed. According to the data label and the proposed graph construction method, the data is represented as a graph. The graph data is input into the neural network model, and the information of the graph data is aggregated through the convolution layer to obtain a new node feature matrix that aggregates the neighbor information. The node feature matrix is ​​converted into a binary pulse of 0 or 1 by pulse coding. After further learning through the fully connected layer and pulse neurons, the number of output pulses is decoded to achieve fault classification and prediction. Finally, a fault alarm is issued based on the fault prediction and diagnosis results.

[0045] In the process framework of the fault diagnosis method proposed in the present invention, it is necessary to first perform data preprocessing and graph data construction on the original data set, and represent the data as a graph according to the data label and the graph construction method proposed by us. The graph data is input into the neural network model, and the information of the graph data is aggregated through the convolution layer to obtain new features about the nodes in the graph data that aggregate the neighbor information. The core of this process is to extract the abstract features of the relevant data set layer by layer through the stacking of network layers. Subsequently, we convert the feature vector into a pulse input between 0 and 1 through the Bernoulli function, which acts as a pulse encoder. The next series of fully connected layers and LIF neurons act as pulse decoders and finally generate pulse outputs. By counting the number of pulses emitted, we can classify and predict different types of faults. By pulse encoding and decoding the feature information of the node, node classification can be achieved, so as to achieve the purpose of fault diagnosis in industrial processes.

[0046] Specifically, S1 collects process variable data under different working conditions of industrial processes, which mainly depends on the coverage of field equipment and various sensors. At the same time, according to the requirements of fault diagnosis tasks, it collects data under different fault conditions that need to be identified.

[0047] S2 is data preprocessing and graph data construction, which is divided into data standardization, sliding window sampling and graph data construction.

[0048] S21 is data standardization. Different sample sets X of industrial production process data that have been assigned labels nm ={x nm1 ,x nm2 ,...,x nmj}. Where n represents the label category, m represents the total number of samples in a label, j represents the number of dimensions of a sample, and a sample point represents the data at a sampling moment, which contains multiple process variables. For example, a small process may have less than ten variables, while a large process may have dozens or even hundreds of process variables. Different variables are standardized on this variable scale, and the calculation formula is as follows:

[0049]

[0050] The mean of the j-th dimension variable is It represents the standard deviation of the variable, and normalizes variables of different scales to the range of 0 to 1.

[0051] S22 is sliding window sampling. A standardized 1×j-dimensional industrial process data is regarded as a data sample, that is, a step-length data in the sliding window sampling. The original sequence of length L is processed using a sliding window setting with a window length of W and a step length of constant S. The data sample dimension after sliding window sampling becomes 1×j×W, and the number of windows N after sampling can be expressed as:

[0052]

[0053] Sliding window sampling can observe and sample different parts of the industrial production process data set, thereby extracting features more comprehensively. This global feature extraction captures the commonality and correlation of the entire data set and enhances the feature expression capability. Secondly, there is data context association between the overlapping windows generated by sliding window sampling, which helps to transfer and share related information between windows and enhance the modeling of global context.

[0054] S23 is the construction of graph data. The standardized industrial production process data is sampled by sliding windows, and N windows are regarded as N nodes. The K-nearest neighbor (KNN) method is used to construct the graph. For each node, the KNN method selects the K nearest neighbor nodes, and then connects these neighbor nodes with the original node to form a graph data set. It can be expressed as: Ne(v i )=KNN(k',v i ,Ψ). KNN(·) means returning the set Ψ=[v i+1 ,v i+2 ,...,v i+k ]Node v i The k' nearest neighbors. The edge weight between each node can be estimated by a Gaussian and weight function, and its calculation formula is:

[0055]

[0056] S3 is to input the industrial production process graph data into the impulse graph neural network model for training, such as Figure 3 As shown, the main features of the structure of the spike graph neural network model are graph feature extraction and spike coding.

[0057] The model represents and extracts graph data features through stacking graph convolution layers, and learns fault features through pulse coding and propagation, ultimately achieving classification and prediction functions. The graph convolution layer is responsible for integrating the information of the constructed graph data set, aggregating the feature information of neighboring nodes, and outputting node features in the form of pulse sequences through the Bernoulli coding rule. After the fully connected layer, the pulse sequences of different nodes are further learned and processed. When passing through the pulse neuron, the pulse intensity of the neuron is obtained according to the set threshold and pulse emission mechanism, and the classification prediction result is obtained according to the pulse size. The different components of the network work together to organically combine graph feature extraction and coding with the final classification judgment, realizing end-to-end processing of graph data.

[0058] S31 is graph feature extraction. The data combination of graphs is different from that of Euclidean space data. The graph structure is relatively complex. Different graphs contain different numbers of nodes, and different nodes have different neighbors. For graphs, the correlation between samples is usually represented by the relationship between nodes and edges. G(V, A, X, E) is used to represent industrial production process graph data. V represents the nodes in the graph, A is the neighbor matrix, which represents the connection relationship between nodes, X is the attribute matrix of the node itself, and E is the edge between nodes. The calculation formula for graph convolution of industrial process graph data is:

[0059]

[0060] H(l) represents the feature vector of the node at layer l, H (l+1) represents the feature vector of the node at layer l+1 after convolution, W l represents the learnable parameters of the lth layer of convolution, σ represents the activation function, Represents a symmetric normalized matrix. Formula (8) expresses the calculation between graph convolution layers, and the information of the adjacency matrix is ​​shared between layers.

[0061] S32 is pulse coding. According to the potential update formula of the pulse neuron of the industrial production process diagram data of the present invention, the membrane potential is Indicates that ΔV m represents the changing membrane potential, then the potential calculation formula of the LIF neuron is:

[0062]

[0063] The node feature matrix obtained after graph convolution is a continuous floating point number. The node feature output after the pulse encoder is defined as represents the total pulse value of the fully connected layer neuron before the tth time step, representing a presynaptic pulse with a value of 0 or 1. The premise of pulse coding is to assume that the size of the encoded pulse probability is positively correlated with the importance of the node feature, that is, each feature triggers a pulse s ij The relationship between the probability p and the node characteristics is encoded using the Bernoulli function as follows:

[0064] p(s ij )~Bernoulli(r ij ),r ij =min(r ij ,1) (10)

[0065] r ij Represents the jth feature of the i-th node. The node features obtained after graph convolution can significantly reflect the information of the data, so the larger the node feature value, the greater the probability that the encoder generates a pulse.

[0066] S33 is to train the pulse graph neural network model. The industrial production process data samples marked with different labels are divided into training sets and test sets or training, validation and test sets. The pulse graph neural network is input in batches, the batch sampling is set to 64, and the learning rate, pulse neuron threshold, number of graph convolution layers and other parameters and hyperparameters are set to train the model. The loss function is calculated by calculating the difference between the predicted fault category and the true category, which is used for back propagation to update the network weights.

[0067] The pulse graph neural network consists of a graph convolution layer, a fully connected layer, and a LIF neuron layer, which realizes the spatiotemporal feature extraction, pulse encoding, and decoding of graph data. The probability function is used for pulse encoding, so that the input feature matrix is ​​converted into a pulse signal, and then the features represented by the pulse are further learned through the fully connected layer.

[0068] S4 uses a trained model for fault prediction and alarm. Specifically, it uses a trained pulse graph neural network model to classify and predict faults in industrial process data of unknown working conditions. The input is real-time industrial production process data to judge normal and different fault conditions. When a fault is detected, the system issues an alarm message, and the model inputs the fault label. The label determines the specific type of fault that occurred in the factory, allowing staff to accurately and efficiently handle faults that occur in production.

[0069] The fault diagnosis method based on pulse graph neural network proposed in the present invention establishes a fault classification model with strong feature extraction capability and low energy consumption, completes the extraction of graph data features through graph convolution layer, uses pulse coding and pulse neurons to simplify complex floating point numbers into 0 and 1 values, and simplifies complex multiplication and addition operations into addition operations, thus achieving efficient fault prediction and classification, and finally performing fault alarms to improve the efficiency of maintenance personnel. The fault diagnosis method proposed in the present invention has advantages in data with strong time series characteristics, especially industrial processes. Traditional deep learning models convert data into feature vectors of fixed size, and it is easy to ignore the time series information between nodes in subsequent processing. The model proposed in the present invention simulates the pulse transmission process through pulse neurons, converts the feature information of graph data into sparse pulse signals, captures the time series characteristics and dynamic evolution relationships in the data, and successfully improves the accuracy and efficiency of fault diagnosis.

[0070] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0071] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

[0072] Finally, it should be noted that the above describes the basic principles, main features and advantages of the present invention. The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features can be replaced by equivalents. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected.

Claims

1. An industrial process fault diagnosis method based on impulse graph neural network, characterized in that: The following steps are involved: S1: Industrial production process data collection: collect data on different process variables that need to be monitored in normal and fault conditions in the chemical production process. In the case of multiple fault conditions, the data under different fault conditions need to be distinguished. The selection of process variables needs to be comprehensive, covering all equipment involved in the process technology. S2: Data preprocessing and graph data construction: fully preprocess the collected industrial production process variable data, standardize all variables on their own scale, and transform them into a normal distribution between 0 and 1, so as to achieve the unification of the scale of the variables; The standardized data is sampled using a sliding window and data enhancement is performed. Finally, the window data is used as the original data for graph construction. The KNN graph construction method is used to construct all nodes into several graphs. S3: Input graph data into the spiking graph neural network for learning; Relying on the pulse graph neural network algorithm, the graph convolution layer is used to extract graph features, and the Bernoulli function is used for pulse encoding, which converts floating-point operations into binary addition operations to reduce computing energy consumption, and further learns through pulse neurons and fully connected layers; the pulse graph neural network is trained using data with different fault labels in the industrial production process to learn the fault characteristics of the data. After training, a high-precision multi-classification fault diagnosis model is obtained; S4: Use the learned model to predict faults and issue alarms; input unknown fault data into the trained neural network model, predict the fault type, issue an alarm, and prompt maintenance personnel to handle the fault in a timely manner.

2. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: Industrial processes are large and small industrial processes that can collect process variable data of different equipment under different working conditions, and are not limited to a certain production process; different working conditions include: normal working state, different fault conditions.

3. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: To pre-process the data under different working conditions, it is necessary to distinguish the data under different working conditions, that is, assign labels to the data and standardize the data with different labels; The standardized variables are sampled using a sliding window, and the sliding window operation is performed using the set step size and window length rules to enhance the connection between the variables at the previous and next moments; Afterwards, the sampled windows are used to construct graph data according to the rules of node similarity. A window is regarded as a node, and the connection relationship between nodes is determined according to the graph construction rules. Data with different labels are constructed into many nodes, and then the nodes of all labels are used to construct a graph, thereby obtaining several graphs with nodes connected according to certain rules. The nodes in the graph are classified and the fault category is determined.

4. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: After the industrial production process data is constructed into a graph, it is input into the pulse graph neural network model to train the fault diagnosis model, including input layer, graph convolution layer, activation function, pulse neuron, fully connected layer and output layer; Before training, the data set is divided into training, validation, and test sets according to the ratio of 6:2:

2. Multiple samples are randomly sampled from the nodes in the training set to form a minimum batch input model.

5. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: The function of the graph convolution layer is to aggregate the neighbor information of the node, capture the data features, and learn the direct correlation of the fault; After preprocessing, the industrial production process data exists in the form of sliding windows. A window is regarded as a node to construct graph data. The graph data can be expressed as G(V, A, X, E). The calculation formula of graph convolution is: Formula (1) describes the calculation between layers, A represents the adjacency matrix, D represents the degree matrix, and H (l) represents the node feature matrix of the lth layer, W is the learnable weight matrix, I is the identity matrix; The activation function σ uses the ReLu activation function.

6. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: The node features of the industrial production process data after convolution will pass through the spiking neuron. The spiking neuron model is also called the leaky integral release (LIF) model. The LIF model is derived by analogizing the spiking neuron to a circuit model containing inductors and capacitors. The derivation process is as follows: Any time-varying current I(t) is input into the pulse neuron, which is a pulse signal transmitted by electrical stimulation or neurons. The current conservation in the circuit is: I(t)=I R +I C (2) Among them I R =u R / R refers to the current passing through the membrane resistance. From the perspective of biological theory, it represents the current generated by the movement of ions inside and outside the membrane, while u R =uu reset ;u reset is the membrane potential threshold of the spiking neuron, u represents the input voltage of the neuron, and u R Indicates the pulse emission voltage; I C is the current obtained by charging the capacitor, and its expression is: The total input I(t) of the circuit is expressed as I(t) = (u(t) - u reset ) / R+Cdu(t) / d(t), introducing the time constant τ m =RC, the potential update mode of LIF neurons of industrial production process data is expressed as:

7. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1, characterized in that: The industrial production process data is simply linearly transformed through the fully connected layer to further learn the data features.

8. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1 is characterized in that: In order to make the node feature matrix obtained after graph convolution be input into the pulse neuron in a suitable way, the node features are pulse-encoded and the Bernoulli function is used to encode the node features; The pulse node features processed by the pulse neurons enter the fully connected layer, and the pulse emission intensity of different nodes is obtained through linear changes. The value of the corresponding category is determined by statistically analyzing the pulse emission intensity. The output layer is encoded in a one-hot manner and the output is a vector whose dimension is the total number of fault classification categories.

9. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1, characterized in that: S2 includes the following steps: S21 is data standardization; for different sample sets X of industrial production process data that have been assigned labels nm ={x nm1 ,x nm2 ,...,x nmj }; n represents the label category, m represents the total number of samples in a label, j represents the number of dimensions of a sample, and a sample point represents the data at a sampling moment; standardize different variables, and the calculation formula is as follows: The mean of the j-th dimension variable is It represents the standard deviation of the variable, which unifies variables of different scales into the range of 0 to 1 through standardization; S22 is sliding window sampling; a 1×j-dimensional industrial process data after standardization is regarded as a data sample, that is, a step-length data in the sliding window sampling; a sliding window setting with a window length of W and a step length of constant S is used to process the original sequence of length L. The data sample dimension after sliding window sampling becomes 1×j×W, and the number of windows N after sampling is expressed as: S23 is the construction of graph data. The standardized industrial production process data is sampled by sliding windows, and N windows are regarded as N nodes. The K-nearest neighbor KNN method is used to construct the graph. For each node, the KNN method selects the K nearest neighbor nodes, and then establishes connections between the neighbor nodes and the original node to form a graph data set; it is expressed as: Ne(v i )=KNN(k',v i ,Ψ); KNN(·) means returning the set Ψ=[v i+1 ,v i+2 ,...,v i+k ]Node v i The k' nearest neighbors of each node; the edge weight between each node is estimated by a Gaussian and weight function, and the calculation formula is:

10. The method for industrial process fault diagnosis based on impulse graph neural network according to claim 1, characterized in that: Each sample point after sliding window sampling is regarded as a node, and the K nearest neighbor (KNN) method is used to construct a graph. All label data are preprocessed in the same way to form nodes, and the nodes are connected by edges to construct several graphs, and the categories of the nodes are learned, trained and classified. G(V,A,X,E) is used to represent industrial process graph data. V represents the nodes in the graph, A is the neighbor matrix, which represents the connection relationship between nodes, X is the attribute matrix of the node itself, and E is the edge between nodes. The calculation formula of graph convolution is: H (l) represents the feature vector of the node at layer l, H (l+1) represents the feature vector of the node at layer l+1 after convolution, W l represents the learnable parameters of the lth layer of convolution, σ represents the activation function, represents a symmetric normalized matrix; The spiking neuron model used is the leaky integral-fire (LIF) model, and its membrane potential is defined as Indicates that ΔV m Represents the changing membrane potential. The potential calculation formula of LIF neurons is: The node feature matrix obtained after graph convolution is a continuous floating point number, and pulse coding is performed using a probability function, which is the Bernoulli function.

Images