An industrial equipment fault diagnosis method

Abstract

The application discloses an industrial equipment fault diagnosis method, and belongs to the technical field of intelligent fault diagnosis, and comprises the following steps: constructing a sensing signal matrix based on a sensing monitoring signal; performing feature extraction and iterative training to obtain a signal relationship strength matrix; clustering the relationship strength between the sensing monitoring signals, and solving a bidirectional influence evaluation target function based on a gradient descent algorithm to obtain a signal optimization sorting result; rearranging the signal relationship strength matrix, and fitting to obtain the causal strength relationship between the sensing monitoring signals, forming a plurality of causal correlation signal pairs, and assigning weight coefficients to the corresponding pairs to construct a causal diagram; and according to the causal diagram, performing fault classification through a path-sensitive attention mechanism, an adaptive information gating mechanism, a multi-channel feature integration mechanism and a hierarchical information processing mechanism to obtain an industrial equipment fault diagnosis result. The application solves the problem of difficult accurate diagnosis of industrial equipment faults under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis technology, and in particular relates to a fault diagnosis method for industrial equipment. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, various mechanical, electrical, and automated equipment are playing an increasingly important role in industrial production. However, these devices inevitably face the influence of various complex environmental factors during long-term operation, such as temperature fluctuations, humidity changes, mechanical vibrations, and electrical interference, making them prone to multi-dimensional failures. These failures not only significantly shorten the service life of the equipment but also pose a serious challenge to the safe and stable operation of the production system. Therefore, accurately monitoring the real-time operating status of equipment and promptly identifying potential failure modes is of great scientific significance and practical value for ensuring the efficient operation and maintenance and economic benefits of the production system.

[0003] Existing methods for fault diagnosis in industrial equipment are mainly divided into two categories: those based on manual feature engineering and those based on data-driven approaches. However, these methods primarily focus on the statistical correlation between data points, neglecting the underlying causal mechanisms. This severely limits their reliable application in critical infrastructure. Faults in industrial equipment typically follow specific physical laws and causal chains, rather than simple statistical correlations. For example, bearing wear can lead to localized temperature increases, which in turn cause changes in vibration frequency and increased energy consumption, forming a clear causal relationship. Traditional correlation analysis methods struggle to distinguish such causal relationships, especially when multiple parameters change at roughly the same time. They cannot accurately determine the causal order between parameters, thus failing to uncover the mechanism of fault occurrence and potentially leading to insufficient diagnostic accuracy. This is particularly true when facing complex operating conditions and novel faults, where the diagnostic accuracy error is even greater. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides an industrial equipment fault diagnosis method that solves the problem of difficulty in accurately diagnosing industrial equipment faults under complex working conditions.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] The present invention provides a method for diagnosing faults in industrial equipment, comprising the following steps:

[0007] S1. Deploy several sensors of different types onto industrial equipment for sensing and monitoring, and construct a sensing signal matrix based on the sensing and monitoring signals;

[0008] S2. Based on the dual-network feature difference loss function and dual-network training loss function, the sensing signal matrix is ​​input into the dual network composed of deep autoencoder and multilayer sensing mechanism for feature extraction and iterative training to obtain the signal relationship intensity matrix.

[0009] S3. Cluster the relationship strengths between the sensing and monitoring signals in the signal relationship strength matrix, and solve the objective function for bidirectional impact assessment based on the gradient descent algorithm to obtain the signal optimization ranking results;

[0010] S4. Rearrange the signal relationship strength matrix according to the optimization sorting results, and obtain the causal strength relationship between the sensing and monitoring signals through triangulation and addition model fitting to form several causal related signal pairs. Assign weight coefficients to each causal related signal pair and construct a causal graph.

[0011] S5. Based on the cause-effect graph, fault classification is performed through path-sensitive attention mechanism, adaptive information gating mechanism, multi-channel feature integration mechanism and hierarchical information processing mechanism to obtain the fault diagnosis results of industrial equipment.

[0012] Further, step S1 includes the following steps:

[0013] S11. Deploy several different types of sensors onto industrial equipment and continuously collect sensing and monitoring signals during the operation of the industrial equipment to construct an original signal dataset.

[0014] S12. Transform the original signal dataset into a sensing signal matrix;

[0015] The calculation expression for the sensing signal matrix is ​​as follows:

[0016] ,

[0017] in, Represents the sensing signal matrix, This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 1. This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 2. This indicates that the sample originates from a Class 1 sensor and the sample size is [number missing]. The sensor monitoring signal, This indicates a sensor monitoring signal originating from two types of sensors with a sample size of 1. This indicates sensor monitoring signals originating from two types of sensors with a sample size of 2. This indicates that the samples originated from two types of sensors and the sample size was [number missing]. The sensor monitoring signal, Indicates origin Sensor-like signals with a sample size of 1; Indicates origin Sensing signals of type sensor with a sample size of 2, Indicates origin Sensor type and sample size The sensor monitoring signal.

[0018] Further, step S2 includes the following steps:

[0019] S21. Input the sensor signal matrix into the deep autoencoder for pre-training to obtain the first signal relationship feature representation;

[0020] S22. Input the first signal relationship feature representation into the multilayer perceptron for structural simplification, and output the second signal relationship feature representation;

[0021] S23. Construct a dual-network feature difference loss function based on the first signal relationship feature representation and the second signal relationship feature representation;

[0022] The calculation expression for the dual-network feature difference loss function is as follows:

[0023] ,

[0024] in, This represents the feature difference loss function of the two networks. This represents the characteristic representation of the first signal relationship. This represents the characteristic representation of the second signal relationship. The first signal relationship characteristic represents the value of the corresponding determinant. The second signal relationship characteristic represents the value of the corresponding determinant;

[0025] S24. Construct the energy function;

[0026] The energy function is calculated as follows:

[0027] ,

[0028] in, This indicates the energy value of the sensor monitoring signal relationship. Indicates the first The sensor monitoring signal and the first Energy weights among individual sensor monitoring signals Indicates the first One sensor monitoring signal, Indicates the first One sensor monitoring signal, Indicates the first The bias of a sensor monitoring signal;

[0029] S25. Construct a dual-network training loss function based on the first signal relationship feature representation, the second signal relationship feature representation, and the energy function;

[0030] The calculation expression for the dual-network training loss function is as follows:

[0031] ,

[0032] ,

[0033] ,

[0034]

[0035] in, This represents the training loss function for the dual networks. Indicates the weights of the transfer learning loss. This represents the transfer learning loss function. Indicates the loss weights of the energy model. Represents the energy model loss function. Represents the weight of physical constraint loss. Represents the physical constraint loss function. Represents the square of the 2-norm. Represents the expectation function, Represents the sensing signal matrix Regarding time gradient, Represents the energy weight matrix. Represents the 1-norm;

[0036] S26. With the goal of minimizing the feature difference loss function of the dual network, according to the training loss function of the dual network, the dual network composed of deep autoencoder and multilayer sensor is iteratively trained using stochastic gradient descent or adaptive momentum optimization algorithm until the dual network stops iterating, and the signal relationship intensity matrix is ​​obtained.

[0037] The expression for calculating the signal relationship intensity matrix is ​​as follows:

[0038] , ,

[0039] in, Represents the signal strength matrix. Indicates belonging to, This indicates the strength of the relationship between the first sensing signal and the second sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the first and second sensor monitoring signals in the sensor signal matrix. This indicates that the first sensing monitoring signal in the sensing signal matrix is ​​related to the second sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals This indicates the strength of the relationship between the second sensing signal and the first sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the second sensing signal and the second sensing signal in the sensing signal matrix. This indicates that the second sensing monitoring signal in the sensing signal matrix is ​​related to the first sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal. Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal Represents the first element in the sensing signal matrix. The sensor monitoring signal and the first The strength of the relationship between individual sensor monitoring signals.

[0040] Further, step S3 includes the following steps:

[0041] S31. Based on the density-based dynamic clustering algorithm, set the neighborhood radius and minimum number of samples, and cluster the relationship strength between the sensing and monitoring signals in the signal relationship strength matrix to obtain a multi-level signal relationship strength matrix.

[0042] S32. Define causal ordering variables and a two-way impact assessment objective function, where, Refers to the first The sensor monitoring signal is ranked first The probability of a sensor monitoring signal preceding a given signal;

[0043] The calculation expression for the objective function of the two-way impact assessment is as follows:

[0044] ,

[0045] in, This represents the objective function for two-way impact assessment. Indicates the influence of the balance coefficient;

[0046] S33. Based on the multi-level signal relationship intensity matrix, solve the bidirectional impact assessment objective function using the gradient descent algorithm to iteratively optimize the causal order of the sensing and monitoring signals, and obtain the signal optimization ranking result.

[0047] The calculation expression for the signal optimization sorting result is as follows:

[0048] ,

[0049] in, This indicates the result of the optimized signal sorting. This indicates the sensor monitoring signal that is the first in the causal transmission sequence. This indicates the sensor monitoring signal that is the second in the causal transmission order. This indicates that the causal transmission order is the first. The position of the sensor monitoring signal.

[0050] Further, step S4 includes the following steps:

[0051] S41. Based on the signal optimization and sorting results, rearrange the signal relationship strength matrix according to the row-first-column arrangement to obtain the preliminary causal structure matrix;

[0052] S42. Perform triangulation on the initial causal structure matrix to obtain the triangulated causal structure matrix;

[0053] S43. An additive model is used to fit the causal relationship strength between each sensing and monitoring signal and its causal parent node in the triangularized causal structure matrix, wherein the sensing and monitoring signal and its causal parent node constitute a causal correlation signal pair.

[0054] The calculation expression for the addition model is as follows:

[0055] ,

[0056] in, Indicates the first The strength of the causal relationship between a sensor monitoring signal and its causal parent node Indicates the first The set of causal parent node indices for each sensor monitoring signal. Indicates the first The sensor monitoring signal and its first The causal coefficients between causal parent nodes Indicates the first The first sensor monitoring signal One causal parent node, Indicates a bias in causality;

[0057] S44. Based on the strength of the causal relationship between each sensing and monitoring signal and its causal parent node, assign weight coefficients to each causal related signal pair in the triangulated causal structure matrix to obtain the target causal structure matrix.

[0058] The calculation expression for the target causal structure matrix is ​​as follows:

[0059] ,

[0060] in, Represents the target causal structure matrix. This indicates the strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. This indicates the strength of the direct causal influence between the first and second sensor monitoring signals. Indicates the first sensor monitoring signal and the second... The intensity of the direct causal influence of each sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the first sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the second sensor monitoring signal. This indicates that the second sensor monitoring signal is related to the first... The intensity of the direct causal influence of each sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The sensor monitoring signal and the first The intensity of the direct causal influence of each sensor monitoring signal;

[0061] S45. Construct a causal graph based on the target causal structure matrix.

[0062] Further, step S5 includes the following steps:

[0063] S51. Based on the path-sensitive attention mechanism, the attention coefficients of each causal related signal to the corresponding node are calculated based on the causal graph, and the enhanced node representation is generated by aggregating the information of neighboring nodes.

[0064] S52. Generate message vectors for each node in the cause-effect graph according to the adaptive information gating mechanism, calculate the control value of the node, update the node representation of each node, and obtain the current node representation.

[0065] S53. Based on the multi-channel feature integration mechanism, extract the structural information of causal graphs of different scales and types, and perform feature connection, dimensionality compression and residual preservation to obtain the structural features of the multi-channel fused graph.

[0066] S54. According to the hierarchical information processing mechanism, calculate the importance score of each node in the causal graph, select key information nodes based on the importance score, and aggregate and compress the node representations of key information nodes to obtain the important node representations.

[0067] S55. Weighted fusion of the enhanced node representation, the current node representation, the multi-channel fusion graph structural features, and the important node representation is performed to obtain the fusion sensing information representation matrix;

[0068] S56. Based on the fused sensor information representation matrix, the probability distribution of various faults corresponding to the causal graph is output through the fully connected layer and the probability mapping function to obtain the fault diagnosis results of industrial equipment.

[0069] Further, S51 includes the following steps:

[0070] S511. Based on the path-sensitive attention mechanism, the attention coefficients of each causal related signal to the corresponding node are calculated based on the causal graph.

[0071] The expression for calculating the attention coefficient is as follows:

[0072] ,

[0073] in, This indicates that the causal correlation signal corresponds to the first... The node and the first Attention coefficients between nodes Represents the attention vector. Indicates the first The eigenvector transformation matrix corresponding to each node This represents a vector concatenation operation. Indicates the first The eigenvector transformation matrix corresponding to each node;

[0074] S512. Normalize the attention coefficients and aggregate neighbor node information based on the normalized attention coefficients to generate reinforced node representations;

[0075] The computational expression for the enhanced node is as follows:

[0076] ,

[0077] in, Indicates the first A reinforced node representation of a node. This represents the sigmoid function. Indicates the first The node and the first Attention coefficients after normalization between nodes Indicates the relationship with the first Each node has a set of nodes that are directly connected to it.

[0078] Further, S52 includes the following steps:

[0079] S521. Based on the adaptive information gating mechanism, generate message vectors for each node in the causal graph.

[0080] S522. Based on the message vector, the control value of the node is calculated through the gating network;

[0081] The expression for calculating the control value of the node is as follows:

[0082] ,

[0083] in, Indicates the first Control values ​​for each node This represents the learnable weight matrix of the gated network. Indicates the first The node representation of a node. Indicates the first The message vector of each node;

[0084] S523. Update the node representation of each node based on the node's control value to obtain the current node representation;

[0085] The computational expression represented by the current node is as follows:

[0086] ,

[0087] in, Indicates the first The current node representation of each node.

[0088] Further, S53 includes the following steps:

[0089] S531. Based on the multi-channel feature integration mechanism, multiple parallel feature extraction channels are set up to extract structural information of causal graphs of different scales and types, and the graph structure features output by each channel are obtained.

[0090] S532. Connect the graph structure features output from each channel, and compress the dimension and retain the information through the projection layer to obtain the multi-channel fused graph structure features.

[0091] The calculation expression for the multi-channel fusion graph structural features is as follows:

[0092] ,

[0093] in, This indicates the structural features of the multi-channel fusion graph. Indicates the weight of channel fusion. express The result of splicing the graph structure features output from each channel along the feature dimension. This indicates the bias for channel fusion. Represents the residual mapping matrix. This represents the structural information of the cause-effect graph.

[0094] Further, S54 includes the following steps:

[0095] S541. Calculate the importance score of each node in the cause-effect graph according to the hierarchical information processing mechanism.

[0096] The formula for calculating the importance score of each node in the causal graph is as follows:

[0097] ,

[0098] ,

[0099] in, In a causal graph, the first... The importance score of each node The importance score learning parameter matrix, Indicates the first The node characteristics of each node. This represents vector concatenation. Indicating the relationship between the cause-and-effect graph and the first Each node contains the total information of its parent node and byte points with a direct causal relationship. Indicates the relationship with the first S542. Based on the importance scores of each node, select several key information nodes using a node selection model.

[0100] The calculation expression for the node selection model is as follows:

[0101] ,

[0102] in, This represents the index of the key information node in the cause-effect graph. This indicates selecting the highest-rated option. 1 node The importance score represents the node. Indicates the pooling ratio. Indicates the total number of nodes. Indicates rounding down;

[0103] S543. Aggregate and compress the node representations of the selected key information nodes to obtain the important node representations.

[0104] The beneficial effects of the present invention are as follows: The industrial equipment fault diagnosis method provided by the present invention, by introducing causal theory into industrial equipment fault diagnosis, can accurately determine the causal order between parameters, explore the mechanism of fault generation, and improve the diagnostic accuracy.

[0105] This invention extracts features from the sensor signal matrix using a dual network composed of a deep autoencoder and a multilayer sensing mechanism, and performs deep feature transformation through quantized energy modeling to capture the correlation between sensor monitoring signals. The scheme divides the feature space into multiple dynamic categories representing causal patterns through adaptive dynamic causal ordering, and constructs a multi-level signal relationship strength matrix by iteratively updating cluster centers and calculating feature membership. Furthermore, it searches for the optimal feature sequence through a bidirectional influence assessment objective function to determine the causal ordering relationship between variables, achieving causal ordering optimization and causal graph construction, thus providing a foundation for accurate diagnosis of industrial equipment faults. This invention analyzes and extracts features from the causal graph through a parallel path-sensitive attention mechanism, an adaptive information gating mechanism, a multi-channel feature integration mechanism, and a hierarchical information processing mechanism, and weightedly fuses features for industrial equipment fault classification. This significantly improves the representation capability in complex fault scenarios, enhances the perception capability of key causal paths and the detection capability of abnormal patterns, improves robustness, and achieves global perception of the overall industrial equipment fault state.

[0106] Other advantages of the present invention will be analyzed in more detail in the following embodiments. Attached Figure Description

[0107] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0108] Figure 1 This is a flowchart illustrating the steps of an industrial equipment fault diagnosis method according to an embodiment of the present invention. Detailed Implementation

[0109] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0110] like Figure 1 As shown, in one embodiment of the present invention, the present invention provides a method for diagnosing faults in industrial equipment, comprising the following steps:

[0111] S1. Deploy several sensors of different types onto industrial equipment for sensing and monitoring, and construct a sensing signal matrix based on the sensing and monitoring signals;

[0112] S1 includes the following steps:

[0113] S11. Deploy several different types of sensors onto industrial equipment and continuously collect sensing and monitoring signals during the operation of the industrial equipment to construct an original signal dataset.

[0114] In this embodiment, the sensors include vibration sensors, temperature sensors, current sensors, etc.; industrial equipment includes machining equipment, automated production line equipment, etc.

[0115] In this solution, the sensor deployment is based on the type of industrial equipment and monitoring needs, and the sensor locations are reasonably selected to ensure that the signal covers key operating parts, the acquisition frequency meets the requirements for capturing fault characteristics, and the sensor signals under stable operation and different working conditions of industrial equipment are continuously collected.

[0116] S12. Transform the original signal dataset into a sensing signal matrix;

[0117] In this scheme, the sensor detection signals in the original signal dataset are organized according to the corresponding sensor type and quantity to form a sensor signal matrix;

[0118] The calculation expression for the sensing signal matrix is ​​as follows:

[0119] ,

[0120] in, Represents the sensing signal matrix, This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 1. This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 2. This indicates that the sample originates from a Class 1 sensor and the sample size is [number missing]. The sensor monitoring signal, This indicates a sensor monitoring signal originating from two types of sensors with a sample size of 1. This indicates sensor monitoring signals originating from two types of sensors with a sample size of 2. This indicates that the samples originated from two types of sensors and the sample size was [number missing]. The sensor monitoring signal, Indicates origin Sensor-like signals with a sample size of 1; Indicates origin Sensing signals of type sensor with a sample size of 2, Indicates origin Sensor type and sample size The sensor monitoring signal.

[0121] S2. Based on the dual-network feature difference loss function and dual-network training loss function, the sensing signal matrix is ​​input into the dual network composed of deep autoencoder and multilayer sensing mechanism for feature extraction and iterative training to obtain the signal relationship intensity matrix.

[0122] In this scheme, deep feature transformation is performed by using high-order feature representation technology of deep autoencoder and quantization energy modeling method of multilayer perceptron. This enables the extraction of potential relationships between sensing and monitoring signals from the sensing signal matrix. Domain knowledge constraints are incorporated to adjust the feature representation, ensuring that the strength of the learned relationships between sensing and monitoring signals can fully meet the inherent characteristics of the physical system. Finally, a high-dimensional representation matrix representing the strength of the relationships between sensing and monitoring signals, namely the signal relationship strength matrix, is output.

[0123] S2 includes the following steps:

[0124] S21. Input the sensing signal matrix into the deep autoencoder for pre-training to obtain the first signal relationship feature representation; in this scheme, the first signal relationship feature representation is a matrix form of the relationship feature representation between sensing and monitoring signals, wherein each matrix element represents the relationship between sensing and monitoring signals.

[0125] S22. Input the first signal relationship feature representation into the multilayer perceptron for structural simplification, and output the second signal relationship feature representation; in this scheme, the second signal relationship feature representation is a matrix-form relationship feature representation between sensing and monitoring signals that is simplified based on the first signal relationship feature representation.

[0126] S23. Construct a dual-network feature difference loss function based on the first signal relationship feature representation and the second signal relationship feature representation;

[0127] The calculation expression for the dual-network feature difference loss function is as follows:

[0128] ,

[0129] in, This represents the feature difference loss function of the two networks. This represents the characteristic representation of the first signal relationship. This represents the characteristic representation of the second signal relationship. The first signal relationship characteristic represents the value of the corresponding determinant. The second signal relationship characteristic represents the value of the corresponding determinant;

[0130] S24. Construct the energy function;

[0131] The energy function is calculated as follows:

[0132] ,

[0133] in, This indicates the energy value of the sensor monitoring signal relationship. Indicates the first The sensor monitoring signal and the first Energy weights among individual sensor monitoring signals Indicates the first One sensor monitoring signal, Indicates the first One sensor monitoring signal, Indicates the first The bias of each sensor monitoring signal; In this scheme, the complex correlation between each sensor monitoring signal is captured by introducing a quantitative energy modeling method, the data configuration quality is characterized by an energy function, and physical constraint terms are incorporated to construct a dual-network training loss function;

[0134] S25. Construct a dual-network training loss function based on the first signal relationship feature representation, the second signal relationship feature representation, and the energy function;

[0135] The calculation expression for the dual-network training loss function is as follows:

[0136] ,

[0137] ,

[0138] ,

[0139]

[0140] in, This represents the training loss function for the dual networks. Indicates the weights of the transfer learning loss. This represents the transfer learning loss function. Indicates the loss weights of the energy model. Represents the energy model loss function. Represents the weight of physical constraint loss. Represents the physical constraint loss function. Represents the square of the 2-norm. Represents the expectation function, Represents the sensing signal matrix Regarding time gradient, Represents the energy weight matrix. Represents the 1-norm; in this embodiment, The value is 0.5. The value is 0.3. The value is 0.2; in this scheme, the energy weight matrix is ​​obtained by capturing the complex correlation between sensing and monitoring signals in the energy function. constitute;

[0141] S26. With the goal of minimizing the feature difference loss function of the dual network, based on the training loss function of the dual network, the dual network composed of the deep autoencoder and the multilayer sensor is iteratively trained using stochastic gradient descent or adaptive momentum optimization algorithm until the dual network stops iterating, and the signal relationship intensity matrix is ​​obtained. In this scheme, the preset dual network stopping iteration condition is that the number of repeated iterations reaches the preset dual network iteration number threshold, or the value of the feature difference loss function of the dual network is less than the preset difference loss threshold for a consecutive preset difference detection round.

[0142] The expression for calculating the signal relationship intensity matrix is ​​as follows:

[0143] , ,

[0144] in, Represents the signal strength matrix. Indicates belonging to, This indicates the strength of the relationship between the first sensing signal and the second sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the first and second sensor monitoring signals in the sensor signal matrix. This indicates that the first sensing monitoring signal in the sensing signal matrix is ​​related to the second sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals This indicates the strength of the relationship between the second sensing signal and the first sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the second sensing signal and the second sensing signal in the sensing signal matrix. This indicates that the second sensing monitoring signal in the sensing signal matrix is ​​related to the first sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal. Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal Represents the first element in the sensing signal matrix. The sensor monitoring signal and the first The strength of the relationship between individual sensor monitoring signals.

[0145] In this embodiment, when iteratively training the dual network composed of a deep autoencoder and a multilayer perceptron using stochastic gradient descent or adaptive momentum optimization algorithms, the learning rate is set to 0.001.

[0146] S3. Cluster the relationship strengths between the sensing and monitoring signals in the signal relationship strength matrix, and solve the objective function for bidirectional impact assessment based on the gradient descent algorithm to obtain the signal optimization ranking results;

[0147] S3 includes the following steps:

[0148] S31. Based on the density-based dynamic clustering algorithm, set the neighborhood radius and minimum number of samples, and cluster the relationship strength between the sensing and monitoring signals in the signal relationship strength matrix to obtain a multi-level signal relationship strength matrix.

[0149] In this embodiment, the neighborhood radius is set to 0.4~0.6, and the minimum number of samples is set to 3~5. During clustering, the cluster centers are initialized, the membership degree of the relationship strength between each sensing and monitoring signal relative to the cluster center is calculated, and the signals are assigned to the cluster center with the highest membership degree. The cluster centers are iteratively updated to obtain a multi-level signal relationship strength matrix after clustering. In this scheme, the multi-level signal relationship strength matrix has the same matrix form as the signal relationship strength matrix, and the corresponding elements in the matrix represent the same objects. However, the matrix elements in the multi-level signal relationship strength matrix comprehensively consider the cluster membership degree, pattern similarity, and original association strength of the relationship strength between sensing and monitoring signals, providing a more comprehensive and accurate measurement basis for causal ranking.

[0150] In this scheme, adaptive clustering is achieved through a density-based dynamic clustering algorithm, which divides the feature space in the signal relationship intensity matrix into multiple dynamic categories representing causal patterns. That is, the cluster centers will be adjusted with iterative updates to adapt to the dynamic changes in feature distribution and reflect the causal patterns under different working conditions. Adaptive learning is achieved by adjusting the cluster centers.

[0151] S32. Define causal ordering variables and a two-way impact assessment objective function, where, Refers to the first The sensor monitoring signal is ranked first The probability of a sensor monitoring signal preceding a given signal;

[0152] The calculation expression for the objective function of the two-way impact assessment is as follows:

[0153] ,

[0154] in, This represents the objective function for two-way impact assessment. Indicates the influence of the balance coefficient;

[0155] In this embodiment, The value is in units of 0.1 and belongs to the range {0, 0.1, 0.2, ..., 1}.

[0156] In this scheme, by setting a causal ranking variable, the discrete causal ranking problem is transformed into a continuous optimization problem, effectively avoiding the risk of combinatorial explosion. The two-way influence evaluation objective function can encourage the maximization of forward causal influence while suppressing backward causal influence. Forward causal influence refers to the causal influence between the top-ranked sensor monitoring signals and the bottom-ranked sensor monitoring signals, while backward causal influence refers to the causal influence between the bottom-ranked sensor monitoring signals and the top-ranked sensor monitoring signals.

[0157] S33. Based on the multi-level signal relationship intensity matrix, solve the bidirectional impact assessment objective function using the gradient descent algorithm to iteratively optimize the causal order of the sensing and monitoring signals, and obtain the signal optimization ranking result.

[0158] The calculation expression for the signal optimization sorting result is as follows:

[0159] ,

[0160] in, This indicates the result of the optimized signal sorting. This indicates the sensor monitoring signal that is the first in the causal transmission sequence. This indicates the sensor monitoring signal that is the second in the causal transmission order. This indicates that the causal transmission order is the first. The position of the sensor monitoring signal.

[0161] S4. Rearrange the signal relationship strength matrix according to the optimization sorting results, and obtain the causal strength relationship between the sensing and monitoring signals through triangulation and addition model fitting to form several causal related signal pairs. Assign weight coefficients to each causal related signal pair and construct a causal graph.

[0162] S4 includes the following steps:

[0163] S41. Based on the signal optimization and sorting results, the signal relationship strength matrix is ​​rearranged according to the row-first, column-second arrangement to obtain the preliminary causal structure matrix. In this scheme, by rearranging the signal relationship strength matrix according to the causal order, it can be ensured that the preliminary causal matrix has potential directed acyclicity.

[0164] S42. Perform triangulation on the initial causal structure matrix to obtain the triangulated causal structure matrix;

[0165] In this scheme, the triangulation operation of the initial causal structure matrix is ​​achieved through row or column permutation, thereby strictly ensuring the directed acyclicity of the causal graph;

[0166] S43. An additive model is used to fit the causal relationship strength between each sensing and monitoring signal and its causal parent node in the triangularized causal structure matrix, wherein the sensing and monitoring signal and its causal parent node constitute a causal correlation signal pair.

[0167] The calculation expression for the addition model is as follows:

[0168] ,

[0169] in, Indicates the first The strength of the causal relationship between a sensor monitoring signal and its causal parent node Indicates the first The set of causal parent node indices for each sensor monitoring signal. Indicates the first The sensor monitoring signal and its first The causal coefficients between causal parent nodes Indicates the first The first sensor monitoring signal One causal parent node, This indicates a causal bias. In this embodiment, the causal parent node of the sensing and monitoring signal refers to the sensing and monitoring signal whose causal order in the triangulated causal structure matrix precedes that of the sensing and monitoring signal. When the sensing and monitoring signal is the effect in the causal relationship, its corresponding causal parent node is the cause in the causal relationship.

[0170] In this plan, These are the direct influence coefficients obtained through regression modeling under a defined causal structure constraint. In this embodiment, the least squares method is used for regression modeling estimation. .

[0171] S44. Based on the strength of the causal relationship between each sensing and monitoring signal and its causal parent node, assign weight coefficients to each causal related signal pair in the triangulated causal structure matrix to obtain the target causal structure matrix.

[0172] The calculation expression for the target causal structure matrix is ​​as follows:

[0173] ,

[0174] in, Represents the target causal structure matrix. This indicates the strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. This indicates the strength of the direct causal influence between the first and second sensor monitoring signals. Indicates the first sensor monitoring signal and the second... The intensity of the direct causal influence of each sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the first sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the second sensor monitoring signal. This indicates that the second sensor monitoring signal is related to the first... The intensity of the direct causal influence of each sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The sensor monitoring signal and the first The intensity of the direct causal influence of each sensor monitoring signal;

[0175] S45. Construct a causal graph based on the target causal structure matrix.

[0176] In this solution, a causal graph is constructed using graph visualization tools, such as the NetworkX library. Nodes in the causal graph represent sensor monitoring signals to reflect the corresponding monitoring characteristics of industrial equipment. Directed edges represent the causal relationships between the various sensor monitoring signals in the target causal structure matrix, and the weights of the directed edges correspond to the strength of each direct causal influence in the target causal structure matrix. In this embodiment, the node positions in the causal graph can also be optimized using an edge filtering layout method to clearly demonstrate the causal transmission paths between variables, thereby forming a more intuitive causal graph and aiding in understanding the causal influence logic between the operating parameters of industrial equipment.

[0177] S5. Based on the cause-effect graph, fault classification is performed through path-sensitive attention mechanism, adaptive information gating mechanism, multi-channel feature integration mechanism and hierarchical information processing mechanism to obtain the fault diagnosis results of industrial equipment.

[0178] S5 includes the following steps:

[0179] S51. Based on the path-sensitive attention mechanism, the attention coefficients of each causal related signal to the corresponding node are calculated based on the causal graph, and the enhanced node representation is generated by aggregating the information of neighboring nodes.

[0180] S51 includes the following steps:

[0181] S511. Based on the path-sensitive attention mechanism, the attention coefficients of each causal signal pair corresponding to the node are calculated based on the causal graph. In this scheme, the attention coefficients of each causal signal pair can measure the importance of the association between nodes in the causal graph, thereby capturing complex node dependencies and highlighting key causal paths.

[0182] The expression for calculating the attention coefficient is as follows:

[0183] ,

[0184] in, This indicates that the causal correlation signal corresponds to the first... The node and the first Attention coefficients between nodes Represents the attention vector. Indicates the first The eigenvector transformation matrix corresponding to each node This represents a vector concatenation operation. Indicates the first The eigenvector transformation matrix corresponding to each node;

[0185] S512. Normalize the attention coefficients and aggregate neighbor node information based on the normalized attention coefficients to generate reinforced node representations;

[0186] The computational expression for the enhanced node is as follows:

[0187] ,

[0188] in, Indicates the first A reinforced node representation of a node. This represents the sigmoid function. Indicates the first The node and the first Attention coefficients after normalization between nodes Indicates the relationship with the first Each node has a set of nodes that are directly connected to it.

[0189] In this scheme, the nodes in the causal graph are enhanced through a path-sensitive attention mechanism to highlight the impact of key causal paths on fault diagnosis, while effectively suppressing the influence of noise or weak causal connections, thus significantly improving the accuracy of fault diagnosis for industrial equipment.

[0190] S52. Generate message vectors for each node in the cause-effect graph according to the adaptive information gating mechanism, calculate the control value of the node, update the node representation of each node, and obtain the current node representation.

[0191] S52 includes the following steps:

[0192] S521. Based on the adaptive information gating mechanism, a message vector is generated for each node in the causal graph. In this scheme, the message vector is obtained by aggregating the features of the node's neighboring nodes.

[0193] S522. Based on the message vector, the control value of the node is calculated through the gating network;

[0194] The expression for calculating the control value of the node is as follows:

[0195] ,

[0196] in, Indicates the first Control values ​​for each node This represents the learnable weight matrix of the gated network. Indicates the first The node representation of a node. Indicates the first The message vector of each node; in this scheme, the learnable weight matrix of the gating network is initialized at the beginning of the gating network training and optimized through backpropagation, and finally learns the best strategy for dynamically adjusting the information flow of nodes.

[0197] S523. Update the node representation of each node based on the node's control value to obtain the current node representation;

[0198] The computational expression represented by the current node is as follows:

[0199] ,

[0200] in, Indicates the first The current node representation of each node.

[0201] In this scheme, the adaptive information gating mechanism filters and adjusts the information flow through a fine control mechanism. By combining the node representation of the node at the time of input of the causal graph with the current node representation, it achieves fine information control. It can dynamically adjust the information flow according to the current state and context of the node, effectively handle abnormal states and mode transitions, and significantly enhance the ability to detect abnormal patterns.

[0202] S53. Based on the multi-channel feature integration mechanism, extract the structural information of causal graphs of different scales and types, and perform feature connection, dimensionality compression and residual preservation to obtain the structural features of the multi-channel fused graph.

[0203] S53 includes the following steps:

[0204] S531. Based on the multi-channel feature integration mechanism, multiple parallel feature extraction channels are set up to extract structural information of causal graphs of different scales and types, and the graph structure features output by each channel are obtained.

[0205] S532. Connect the graph structure features output from each channel, and compress the dimension and retain the information through the projection layer to obtain the multi-channel fused graph structure features.

[0206] The calculation expression for the multi-channel fusion graph structural features is as follows:

[0207] ,

[0208] in, This indicates the structural features of the multi-channel fusion graph. Indicates the weight of channel fusion. express The result of splicing the graph structure features output from each channel along the feature dimension. This indicates the bias for channel fusion. Represents the residual mapping matrix. This represents the structural information of the cause-effect graph.

[0209] In this scheme, the multi-channel feature integration mechanism not only connects the outputs of multiple channels, but also adds residual connections to maintain smooth information flow, enabling the graph structure of the causal graph to be understood from multiple perspectives, thereby capturing complex failure modes and propagation paths to improve robustness.

[0210] S54. According to the hierarchical information processing mechanism, calculate the importance score of each node in the causal graph, select key information nodes based on the importance score, and aggregate and compress the node representations of key information nodes to obtain the important node representations.

[0211] S54 includes the following steps:

[0212] S541. Calculate the importance score of each node in the cause-effect graph according to the hierarchical information processing mechanism.

[0213] The formula for calculating the importance score of each node in the causal graph is as follows:

[0214] ,

[0215] ,

[0216] in, In a causal graph, the first... The importance score of each node The importance score learning parameter matrix, Indicates the first The node characteristics of each node. This represents vector concatenation. Indicating the relationship between the cause-and-effect graph and the first Each node contains the total information of its parent node and byte points with a direct causal relationship. Indicates the relationship with the first A set of nodes that are directly connected to each other in a directed manner;

[0217] In this scheme, node features refer to the local sensing information carried by the node after processing by attention and gating mechanisms; this scheme combines the node's own node features with its environmental features through vector concatenation to achieve a comprehensive evaluation of both global and local features; in this scheme, the importance score learning parameter matrix is ​​more important for identifying specific industrial equipment faults by automatically adjusting and learning the causal structure of the corresponding type. Able to reflect the first The contextual position of each node in the causal chain; It is based on the target causal structure matrix Determined and the first Each node has a set of nodes that are directly connected to it in a directed manner;

[0218] In this scheme, the importance score of a node can reflect the degree of contribution of the node to fault diagnosis.

[0219] S542. Based on the importance scores of each node, select several key information nodes using the node selection model;

[0220] The calculation expression for the node selection model is as follows:

[0221] ,

[0222] in, This represents the index of the key information node in the cause-effect graph. This indicates selecting the highest-rated option. 1 node The importance score represents the node. Indicates the pooling ratio. Indicates the total number of nodes. Indicates rounding down;

[0223] S543. Aggregate and compress the node representations of the selected key information nodes to obtain the important node representations.

[0224] In this scheme, node selection allows for the retention of the most important nodes, effectively reducing computational complexity while preserving key information and enhancing the understanding of the global state.

[0225] S55. Weighted fusion of the enhanced node representation, the current node representation, the multi-channel fusion graph structural features, and the important node representation is performed to obtain the fusion sensing information representation matrix;

[0226] S56. Based on the fused sensor information representation matrix, the probability distribution of various faults corresponding to the causal graph is output through the fully connected layer and the probability mapping function to obtain the fault diagnosis results of industrial equipment.

[0227] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults in industrial equipment, characterized in that, Includes the following steps: S1. Deploy several sensors of different types onto industrial equipment for sensing and monitoring, and construct a sensing signal matrix based on the sensing and monitoring signals; S2. Based on the dual-network feature difference loss function and dual-network training loss function, the sensing signal matrix is ​​input into the dual network composed of deep autoencoder and multilayer sensing mechanism for feature extraction and iterative training to obtain the signal relationship intensity matrix. S3. Cluster the relationship strengths between the sensing and monitoring signals in the signal relationship strength matrix, and solve the objective function for bidirectional impact assessment based on the gradient descent algorithm to obtain the signal optimization ranking results; S3 includes the following steps: S31. Based on the density-based dynamic clustering algorithm, set the neighborhood radius and minimum number of samples, and cluster the relationship strength between the sensing and monitoring signals in the signal relationship strength matrix to obtain a multi-level signal relationship strength matrix. S32. Define causal ordering variables and a two-way impact assessment objective function, where, Refers to the first The sensor monitoring signal is ranked first The probability of a sensor monitoring signal preceding a given signal; The calculation expression for the objective function of the two-way impact assessment is as follows: ， in, This represents the objective function for two-way impact assessment. Indicates the influence of the balance coefficient. Represents the first element in the sensing signal matrix. The sensor monitoring signal and the first The strength of the relationship between individual sensor monitoring signals; S33. Based on the multi-level signal relationship intensity matrix, solve the bidirectional impact assessment objective function using the gradient descent algorithm to iteratively optimize the causal order of the sensing and monitoring signals, and obtain the signal optimization ranking result. The calculation expression for the signal optimization sorting result is as follows: ， in, This indicates the result of the optimized signal sorting. This indicates the sensor monitoring signal that is the first in the causal transmission sequence. This indicates the sensor monitoring signal that is the second in the causal transmission order. This indicates that the causal transmission order is the first. Position sensing and monitoring signals; S4. Rearrange the signal relationship strength matrix according to the optimization sorting results, and obtain the causal strength relationship between the sensing and monitoring signals through triangulation and addition model fitting to form several causal related signal pairs. Assign weight coefficients to each causal related signal pair and construct a causal graph. S5. Based on the cause-effect graph, fault classification is performed through path-sensitive attention mechanism, adaptive information gating mechanism, multi-channel feature integration mechanism and hierarchical information processing mechanism to obtain the fault diagnosis results of industrial equipment.

2. The industrial equipment fault diagnosis method according to claim 1, characterized in that, S1 includes the following steps: S11. Deploy several different types of sensors onto industrial equipment and continuously collect sensing and monitoring signals during the operation of the industrial equipment to construct an original signal dataset. S12. Transform the original signal dataset into a sensing signal matrix; The calculation expression for the sensing signal matrix is ​​as follows: ， in, Represents the sensing signal matrix, This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 1. This indicates a sensor monitoring signal originating from a Class 1 sensor and having a sample size of 2. This indicates that the sample originates from a Class 1 sensor and the sample size is [number missing]. The sensor monitoring signal, This indicates a sensor monitoring signal originating from two types of sensors with a sample size of 1. This indicates sensor monitoring signals originating from two types of sensors with a sample size of 2. This indicates that the samples originated from two types of sensors and the sample size was [number missing]. The sensor monitoring signal, Indicates origin Sensor-like signals with a sample size of 1; Indicates origin Sensing signals of type sensor with a sample size of 2, Indicates origin Sensor type and sample size The sensor monitoring signal.

3. The industrial equipment fault diagnosis method according to claim 2, characterized in that, S2 includes the following steps: S21. Input the sensor signal matrix into the deep autoencoder for pre-training to obtain the first signal relationship feature representation; S22. Input the first signal relationship feature representation into the multilayer perceptron for structural simplification, and output the second signal relationship feature representation; S23. Construct a dual-network feature difference loss function based on the first signal relationship feature representation and the second signal relationship feature representation; The calculation expression for the dual-network feature difference loss function is as follows: ， in, This represents the feature difference loss function of the two networks. This represents the characteristic representation of the first signal relationship. This represents the characteristic representation of the second signal relationship. The first signal relationship characteristic represents the value of the corresponding determinant. The second signal relationship characteristic represents the value of the corresponding determinant; S24. Construct the energy function; The energy function is calculated as follows: ， in, This indicates the energy value of the sensor monitoring signal relationship. Indicates the first The sensor monitoring signal and the first Energy weights among individual sensor monitoring signals Indicates the first One sensor monitoring signal, Indicates the first One sensor monitoring signal, Indicates the first The bias of a sensor monitoring signal; S25. Construct a dual-network training loss function based on the first signal relationship feature representation, the second signal relationship feature representation, and the energy function; The calculation expression for the dual-network training loss function is as follows: ， ， ， in, This represents the training loss function for the dual-network system. Indicates the weights of the transfer learning loss. This represents the transfer learning loss function. Indicates the loss weights of the energy model. Represents the energy model loss function. Represents the physical constraint loss weights. Represents the physical constraint loss function. Represents the square of the 2-norm. Represents the expectation function, Represents the sensing signal matrix Regarding time gradient, Represents the energy weight matrix. Represents the 1-norm; S26. With the goal of minimizing the feature difference loss function of the dual network, according to the training loss function of the dual network, the dual network composed of deep autoencoder and multilayer sensor is iteratively trained using stochastic gradient descent or adaptive momentum optimization algorithm until the dual network stops iterating, and the signal relationship intensity matrix is ​​obtained. The expression for calculating the signal relationship intensity matrix is ​​as follows: ， ， in, Represents the signal strength matrix. Indicates belonging to, This indicates the strength of the relationship between the first sensing signal and the second sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the first and second sensor monitoring signals in the sensor signal matrix. This indicates that the first sensing monitoring signal in the sensing signal matrix is ​​related to the second sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals This indicates the strength of the relationship between the second sensing signal and the first sensing signal in the sensing signal matrix. This indicates the strength of the relationship between the second sensing signal and the second sensing signal in the sensing signal matrix. This indicates that the second sensing monitoring signal in the sensing signal matrix is ​​related to the first sensing monitoring signal. The strength of the relationship between individual sensor monitoring signals Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal. Represents the first element in the sensing signal matrix. The strength of the relationship between the first sensor monitoring signal and the second sensor monitoring signal.

4. The industrial equipment fault diagnosis method according to claim 3, characterized in that, S4 includes the following steps: S41. Based on the signal optimization and sorting results, rearrange the signal relationship strength matrix according to the row-first-column arrangement to obtain the preliminary causal structure matrix; S42. Perform triangulation on the initial causal structure matrix to obtain the triangulated causal structure matrix; S43. An additive model is used to fit the causal relationship strength between each sensing and monitoring signal and its causal parent node in the triangularized causal structure matrix, wherein the sensing and monitoring signal and its causal parent node constitute a causal correlation signal pair. The calculation expression for the addition model is as follows: ， in, Indicates the first The strength of the causal relationship between a sensor monitoring signal and its causal parent node Indicates the first The set of causal parent node indices for each sensor monitoring signal. Indicates the first The sensor monitoring signal and its first The causal coefficients between causal parent nodes Indicates the first The first sensor monitoring signal One causal parent node, Indicates a bias in causality; S44. Based on the strength of the causal relationship between each sensing and monitoring signal and its causal parent node, assign weight coefficients to each causal related signal pair in the triangulated causal structure matrix to obtain the target causal structure matrix. The calculation expression for the target causal structure matrix is ​​as follows: ， in, Represents the target causal structure matrix. This indicates the strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. This indicates the strength of the direct causal influence between the first and second sensor monitoring signals. Indicates the first sensor monitoring signal and the second... The intensity of the direct causal influence of each sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the first sensor monitoring signal. This indicates the strength of the direct causal influence between the second sensor monitoring signal and the second sensor monitoring signal. This indicates that the second sensor monitoring signal is related to the first... The intensity of the direct causal influence of each sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The strength of the direct causal influence between the first sensor monitoring signal and the second sensor monitoring signal. Indicates the first The sensor monitoring signal and the first The intensity of the direct causal influence of each sensor monitoring signal; S45. Construct a causal graph based on the target causal structure matrix.

5. The industrial equipment fault diagnosis method according to claim 4, characterized in that, S5 includes the following steps: S51. Based on the path-sensitive attention mechanism, the attention coefficients of each causal related signal to the corresponding node are calculated based on the causal graph, and the enhanced node representation is generated by aggregating the information of neighboring nodes. S52. Generate message vectors for each node in the cause-effect graph according to the adaptive information gating mechanism, calculate the control value of the node, update the node representation of each node, and obtain the current node representation. S53. Based on the multi-channel feature integration mechanism, extract the structural information of causal graphs of different scales and types, and perform feature connection, dimensionality compression and residual preservation to obtain the structural features of the multi-channel fused graph. S54. According to the hierarchical information processing mechanism, calculate the importance score of each node in the causal graph, select key information nodes based on the importance score, and aggregate and compress the node representations of key information nodes to obtain the important node representations. S55. Weighted fusion of the enhanced node representation, the current node representation, the multi-channel fusion graph structural features, and the important node representation is performed to obtain the fusion sensing information representation matrix; S56. Based on the fused sensor information representation matrix, the probability distribution of various faults corresponding to the causal graph is output through the fully connected layer and the probability mapping function to obtain the fault diagnosis results of industrial equipment.

6. The industrial equipment fault diagnosis method according to claim 5, characterized in that, S51 includes the following steps: S511. Based on the path-sensitive attention mechanism, the attention coefficients of each causal related signal to the corresponding node are calculated based on the causal graph. The expression for calculating the attention coefficient is as follows: ， in, Indicates the causal correlation signal for the corresponding first... The node and the first Attention coefficients between nodes Represents the attention vector. Indicates the first The eigenvector transformation matrix corresponding to each node This represents a vector concatenation operation. Indicates the first The eigenvector transformation matrix corresponding to each node; S512. Normalize the attention coefficients and aggregate neighbor node information based on the normalized attention coefficients to generate reinforced node representations; The computational expression for the enhanced node is as follows: ， in, Indicates the first A reinforced node representation of a node. This represents the sigmoid function. Indicates the first The node and the first Attention coefficients after normalization between nodes Indicates the relationship with the first Each node has a set of nodes that are directly connected to it.

7. The industrial equipment fault diagnosis method according to claim 6, characterized in that, S52 includes the following steps: S521. Based on the adaptive information gating mechanism, generate message vectors for each node in the causal graph. S522. Based on the message vector, the control value of the node is calculated through the gating network; The expression for calculating the control value of the node is as follows: ， in, Indicates the first Control values ​​for each node This represents the learnable weight matrix of the gated network. Indicates the first The node representation of a node. Indicates the first The message vector of each node; S523. Update the node representation of each node based on the node's control value to obtain the current node representation; The computational expression represented by the current node is as follows: ， in, Indicates the first The current node representation of each node.

8. The industrial equipment fault diagnosis method according to claim 7, characterized in that, S53 includes the following steps: S531. Based on the multi-channel feature integration mechanism, multiple parallel feature extraction channels are set up to extract structural information of causal graphs of different scales and types, and the graph structure features output by each channel are obtained. S532. Connect the graph structure features output from each channel, and compress the dimension and retain the information through the projection layer to obtain the multi-channel fused graph structure features. The calculation expression for the multi-channel fusion graph structural features is as follows: ， in, This indicates the structural features of the multi-channel fusion graph. Indicates the weight of channel fusion. express The result of splicing the graph structure features output from each channel along the feature dimension. This indicates the bias for channel fusion. Represents the residual mapping matrix. This represents the structural information of the cause-effect graph.

9. The industrial equipment fault diagnosis method according to claim 8, characterized in that, S54 includes the following steps: S541. Calculate the importance score of each node in the cause-effect graph according to the hierarchical information processing mechanism. The formula for calculating the importance score of each node in the causal graph is as follows: ， ， in, In a causal graph, the first... The importance score of each node The importance score learning parameter matrix, Indicates the first The node characteristics of each node. This represents vector concatenation. Indicates the relationship between the cause-and-effect graph and the first Each node contains the total information of its parent node and byte points with a direct causal relationship. Indicates the relationship with the first A set of nodes that are directly connected to each other in a directed manner; S542. Based on the importance scores of each node, select several key information nodes using the node selection model; The calculation expression for the node selection model is as follows: ， in, This represents the index of the key information node in the cause-effect graph. This indicates selecting the highest-rated option. 1 node The importance score represents the node. Indicates the pooling ratio. Indicates the total number of nodes. Indicates rounding down; S543. Aggregate and compress the node representations of the selected key information nodes to obtain the important node representations.

Images