A fault diagnosis method for electrical equipment based on fault transmission and spatial mining
Patent Information
- Application Number
- CN202610919201.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-24
- Publication Date
- 2026-09-11
AI Technical Summary
[0003]在电气设备故障诊断领域，现有方法多聚焦于单一测点信号的时频特征分析或浅层次的多测点信息叠加，而忽略了复杂设备中故障激励与多测点振动响应之间的时间因果传递关系以及空间多测点信号的深层关联性
本发明通过符号相位转移熵量化各测点间故障信息传递规律并构建关系图，有效表征故障动态传播路径与时序因果依赖；结合多测点时频域特征搭建特征图，精准刻画测点之间的空间拓扑关联，摒弃传统数据简单拼接的处理方式。依托多通道自适应图卷积网络对两类图结构特征进行联合提取与融合，实现故障时序传递与空间关联的协同挖掘，充分捕捉故障传播的全局模式。同时引入一致性约束与独立性准则优化融合特征，降低特征共线性干扰、强化有效故障特征表达。
Smart Images
Figure CN122734705A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical equipment fault diagnosis technology, and specifically to an electrical equipment fault diagnosis method based on fault transmission and spatial mining. Background Technology
[0002] Electrical equipment fault diagnosis is a key technology for ensuring the safe and stable operation of power systems and preventing significant economic losses caused by sudden failures. By accurately identifying early-stage potential faults, equipment maintenance costs can be significantly reduced and service life extended, contributing to the sustainable development of industrial production.
[0003] In the field of electrical equipment fault diagnosis, existing methods mostly focus on the time-frequency characteristic analysis of signals from a single measuring point or the superficial superposition of information from multiple measuring points, neglecting the temporal causal transmission relationship between fault excitation and vibration response at multiple measuring points in complex equipment, as well as the deep correlation between signals from multiple measuring points in space. On the one hand, when a fault propagates within the equipment, the vibration signals from different measuring points often exhibit temporal causal dependencies. For example, an early fault in a component may trigger abnormal vibrations at other measuring points through mechanical coupling, but traditional methods struggle to quantify this dynamic fault information transmission path. On the other hand, signals from measuring points at different locations within the equipment exhibit topological correlations in space, but existing models often treat multi-measuring-point data with simple splicing or independent processing, resulting in insufficient spatial feature fusion and difficulty in capturing the global pattern of fault propagation. This lack of joint modeling of temporal causality and spatial topology limits the refined extraction of fault features and the improvement of diagnostic accuracy under complex operating conditions.
[0004] Therefore, there is an urgent need for an intelligent diagnostic method that can simultaneously characterize the spatiotemporal causal network relationship of signals from multiple measurement points and achieve the synergistic fusion of dynamic information transmission and spatial features. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a method for diagnosing electrical equipment faults based on fault transmission and spatial mining. This invention discloses a fault diagnosis method for electrical equipment based on fault propagation and spatial mining, comprising: Step S1: Collect vibration signals from multiple measuring points of electrical equipment, preprocess the vibration signals, calculate the symbol phase transfer entropy between each measuring point based on the preprocessed signals, and construct a relationship diagram representing the transmission relationship of fault information. Step S2: Extract the time-frequency domain features of the vibration signals at each measuring point, and construct a feature map representing the correlation of the measuring point features based on the time-frequency domain features; Step S3: Input the relationship graph and feature graph into a multi-channel adaptive graph convolutional network to jointly extract and fuse the structural features of the two types of graphs to obtain the fused global features; Step S4: Optimize the global features based on consistency constraints and independence criteria, and complete the fault identification and diagnosis of electrical equipment through a classification model.
[0006] As a further improvement of the present invention, in step S1, the vibration signal is preprocessed, specifically including: The original vibration signal is converted into an instantaneous phase time sequence using Hilbert transform; The instantaneous phase time series is sequentially reconstructed in phase space and symbolized, and finally the symbolic phase transfer entropy is calculated.
[0007] As a further improvement of the present invention, the phase space reconstruction uses the pseudo-nearest neighbor method to determine the embedding dimension m, wherein the delay time τ is 3-4, the symbolic permutation order is 1, and the time step δ is 1. As a further improvement of the present invention, in step S1, constructing a relationship diagram representing the fault information transmission relationship specifically includes: A relational graph is constructed using the symbolic phase transfer entropy calculated between each measurement point as elements of the adjacency matrix. As a further improvement of the present invention, in step S2, the time-frequency domain features include time-domain statistical features and frequency-domain features; the time-domain statistical features include mean, variance, kurtosis, and peak factor, and the frequency-domain features include centroid frequency, mean square frequency, spectral entropy, and cepstral peak. As a further improvement of the present invention, in step S2, constructing a feature map representing the correlation of measurement point features specifically includes: calculating a similarity matrix using cosine similarity based on the time-frequency domain feature vectors of each measurement point, and establishing an adjacency matrix using the k-nearest neighbor method, thereby constructing a feature map. As a further improvement of the present invention, in step S3, the joint extraction and fusion of the two types of graph structure features specifically includes: A multi-channel adaptive graph convolutional network is used to extract unique embedding features and shared common embedding features from the relationship graph and the feature graph, respectively. Global average pooling is performed on each unique embedding feature and the common embedding feature to obtain the corresponding global feature vector; An attention mechanism is introduced to calculate the attention weight coefficients of each feature vector. Based on the weight coefficients, the features of multiple classes are adaptively weighted and fused to obtain the fused global features. As a further improvement of the present invention, the calculation process of the attention mechanism includes: Set up three sets of attention weight matrices and corresponding bias terms, and then perform operations on each pooled global feature vector with the corresponding weight matrix and bias term to obtain the initial attention score; The Softmax function is used to normalize all initial attention scores to obtain standardized attention weight coefficients; By combining the weight coefficients with the corresponding embedded features, a weighted summation is performed to complete the fusion of multi-source features. As a further improvement of the present invention, in step S4, optimizing the global features based on consistency constraints and independence criteria includes: The embedded features output from the relation graph and feature graph are subjected to L2 normalization, and a consistency constraint loss function is constructed to enhance the similarity between the feature outputs of the two types of graphs. The Hilbert-Schmidt independence criterion is used to construct an independence-constrained loss function, which weakens the collinearity interference between different embedded features and preserves differentiated fault information. As a further improvement of the present invention, in step S4, the classification model is a fully connected classification layer equipped with a Softmax activation function, and the cross-entropy loss update model is adopted to finally output the fault type and status diagnosis results of the electrical equipment. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention quantifies the fault information transmission patterns between measurement points and constructs a relationship graph through symbolic phase transfer entropy, effectively characterizing the dynamic propagation path and temporal causal dependence of faults. It combines time-frequency domain features from multiple measurement points to build a feature graph, accurately depicting the spatial topological relationships between measurement points, abandoning the traditional method of simply splicing data. Relying on a multi-channel adaptive graph convolutional network, it jointly extracts and fuses the two types of graph structure features, achieving collaborative mining of fault temporal transmission and spatial relationships, fully capturing the global pattern of fault propagation. Simultaneously, it introduces consistency constraints and independence criteria to optimize the fused features, reducing feature collinearity interference and enhancing the effective expression of fault features.
[0008] This invention enables accurate identification of potential electrical equipment faults, effectively improves the accuracy of fault diagnosis under complex operating conditions, reduces equipment maintenance costs, extends equipment lifespan, ensures the safe and stable operation of the power system, and helps industrial production to continue reliably. Attached Figure Description
[0009] Figure 1 This is a flowchart of an electrical equipment fault diagnosis method based on fault propagation and spatial mining, as disclosed in one embodiment of the present invention. Figure 2 The image shows the vibration data waveforms of bearing housing A in the X direction under different faults in an embodiment of the electrical equipment fault diagnosis method based on fault transmission and spatial mining disclosed in this invention. Figure 3 The image shows the vibration data waveforms of bearing housing A in the Y direction under different faults in an embodiment of the electrical equipment fault diagnosis method based on fault transmission and spatial mining disclosed in this invention. Figure 4 The image shows the vibration data waveforms of bearing housing B in the X direction under different faults in an embodiment of the electrical equipment fault diagnosis method based on fault transmission and spatial mining disclosed in this invention. Figure 5 The image shows the vibration waveforms of bearing housing B in the Y direction under different faults in an embodiment of the electrical equipment fault diagnosis method based on fault transmission and spatial mining disclosed in this invention. Figure 6 This is a schematic diagram of the diagnostic results of the proposed multi-channel adaptive graph convolutional neural model under different signal-to-noise ratios in an embodiment of the electrical equipment fault diagnosis method based on fault propagation and spatial mining disclosed in this invention. Detailed Implementation
[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0011] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, a fault diagnosis method for electrical equipment based on fault transmission and spatial mining according to the present invention includes: Step S1: Signal Acquisition, Preprocessing, and Relationship Diagram Construction Raw vibration signals from multiple measuring points of electrical equipment were collected and preprocessed: the raw vibration signals were transformed into instantaneous phase time series using Hilbert transform; the phase space of the instantaneous phase time series was reconstructed using the pseudo-nearest neighbor method, where the embedding dimension m ranged from 3 to 7, the delay time τ was set to 3 to 4, the symbolic permutation order was set to 1, and the time step δ was set to 1. The reconstructed sequence was then symbolized, and the symbolic phase transfer entropy between different measuring points was calculated. This value was used as an element of the adjacency matrix At to construct a relational graph representing fault information transmission and temporal causal relationships. .
[0012] Step S2: Time-frequency domain feature extraction and feature map construction The time-frequency domain features of the vibration signals at each measuring point are extracted, including time-domain statistical features and frequency-domain features. The time-domain statistical features include mean, variance, kurtosis, and peak factor, while the frequency-domain features include centroid frequency, mean square frequency, spectral entropy, and cepstral peak value. Based on the time-frequency domain feature vectors of each measuring point, a cosine similarity matrix is calculated between features, and an adjacency matrix A is generated using the k-nearest neighbor method. f Finally, a feature map representing the spatial feature association of measurement points is constructed. .
[0013] Step S3: Joint extraction and fusion of features from multi-channel adaptive graph convolutional networks The relationship graph and feature map are jointly input into a multi-channel adaptive graph convolutional network to jointly extract and fuse the structural features of the two types of graphs, resulting in fused global features. Specifically, unique embedding features and shared embedding features between the two types of graphs are extracted. Global average pooling is then performed on each set of embedding features to obtain the corresponding global feature vectors. An attention mechanism is introduced to calculate the attention weight coefficients of each feature vector. Based on these weight coefficients, the multi-class features are adaptively weighted and fused to obtain the fused global features. This includes setting three sets of attention weight matrices and bias terms, calculating initial attention scores based on the feature vectors, normalizing the scores using the Softmax function to obtain standardized attention weight coefficients, and then adaptively weighting and fusing the multi-class features based on these weight coefficients to obtain the fused global features.
[0014] Step S4: Feature Optimization and Fault Classification Diagnosis The embedded features output from the relationship graph and feature graph are subjected to L2 normalization, and a consistency constraint loss function is constructed to enhance the similarity between the feature outputs of the two types of graphs. Simultaneously, the Hilbert-Schmidt independence criterion (HSIC) is used to construct an independence constraint loss function to weaken collinearity interference between different embedded features, preserve differentiated fault information, and complete global feature optimization. The optimized features are then input into a classification model, which is a fully connected classification layer with a Softmax activation function. Cross-entropy loss is used to update the model, ultimately identifying the fault type of electrical equipment and completing fault diagnosis.
[0015] Example 1 This embodiment uses a rotating machinery test bench as the test object to simulate typical electrical / rotating equipment faults in industrial settings, setting four common fault conditions: bearing inner ring fault, bearing outer ring fault, shaft misalignment, and rotor imbalance.
[0016] Measurement points were set up on bearing housings A and B of the test bench. Vibration data in both the X and Y directions were collected simultaneously for each bearing housing, for a total of 4 sets of measurement point signals. The sampling frequency of the equipment was set to 10kHz, and the length of a single sampling data set was set to 2048. The raw vibration data of all measurement points under the above four fault conditions were collected, and the vibration waveform data corresponding to each fault was saved (e.g., Figure 2-5 The figure shows the vibration waveforms of bearing housings A and B in the X and Y directions under different fault conditions.
[0017] Step S1: Data preprocessing, symbol phase transfer entropy calculation, and relational graph construction For a given time series and Its instantaneous phase time series θx(t) and θy(t) are respectively derived from and Obtained through Hilbert transform. For a given time series... x Its analysis sequence ζ Defined as: ; In the formula, A ( t () represents the instantaneous amplitude; θ ( t )for x ( t The instantaneous phase of ); for x ( t Perform Hilbert transform to obtain Specifically: ; In the formula, PV is the integral of Cauchy's principal value.
[0018] right θx(t) and θy(t) Phase space reconstruction is performed by using the pseudo-nearest neighbor method to reconstruct the phase space of the instantaneous phase time series: starting from the low dimension, the proportion of pseudo-nearest neighbor points in the phase space trajectory is calculated. When the proportion of pseudo-nearest neighbor points is less than 1% and no longer decreases with the increase of dimension, the current dimension is determined to be the optimal embedding dimension m.
[0019] After completing the phase space reconstruction, the sequence is symbolized. Using the relative frequencies of the symbols, the joint probability and conditional probability of the permutation index sequence are estimated, and the corresponding probabilities are calculated. Y arrive X The sign phase transition entropy is: ; ; ; In the formula: Let x be the sign instantaneous phase of the time series at time i; Let y be the sign of the instantaneous phase of the time series y at time i; Time series x In time i + δ The instantaneous phase of the symbol at a given point; δ represents a time step. Table 1 shows the symbol phase transfer entropy values obtained between experimental vibration signals.
[0020] Table 1. Symbol phase transition entropy values between experimental vibration signals Step S2: Time-frequency domain feature extraction and feature map construction For the vibration signals of the four measurement points, eight-dimensional features were extracted one by one: time-domain features (mean, variance, kurtosis, peak factor) and frequency-domain features (centroid frequency, mean square frequency, spectral entropy, cepstral peak). The features of the individual measurement points were combined into a feature vector.
[0021] Based on feature vectors, cosine similarity is used to calculate the similarity matrix between n nodes. By combining the k-nearest neighbor rule to filter edge relationships, an adjacency matrix is generated. Combine the feature vector matrix X to construct the k-nearest neighbor feature map. The feature maps are extracted using a graph convolutional network. Layer node embedding for: ; In the formula, It is the first in GCN l Layer weight matrix; ReLU It is the ReLU activation function, initially ; , yes The angle matrix; n represents the number of nodes. This indicates the final layer output embedding. A high-dimensional information transfer matrix, constructed based on symbolic phase transfer entropy, is used as an adjacency matrix to build a relational graph. This graph emphasizes the correlation between sensor signals under different states, reflecting the causal relationships and information flow intensity between fault excitation and response, and helps describe the fault propagation path, thus aiding the model in more accurately understanding and predicting faults. Combining the advantages and disadvantages of relational graphs and topology graphs allows for a flexible balance of information between different graphs, helping to capture potential dependencies and information transfer paths between components, and improving the robustness of the fault diagnosis model to different types of faults.
[0022] Step S3: Feature extraction and attention fusion using multi-channel adaptive graph convolutional network To extract common information from the two graphs, a weight matrix is constructed from the feature maps. Relationship diagram Extracting node embeddings Z CT and Z CF The common embedding Z of the two graph networks C The sum is: ; To ensure the output features of the two graph networks and To enhance the accuracy of information fusion between relational graphs and feature graphs, consistency constraints were designed based on the similarity between the graphs. Firstly, for... and The embedding matrix is normalized using L2 normalization. , Then, two normalized matrices can be used to capture the similarity of n nodes, with the consistency constraint £C as follows: and as follows: ; Embedded and From the same relation graph G t =(A t In the learning process of the two embeddings (X), to ensure that they can capture different information, the Hilbert-Schmidt independence criterion (HSIC) is adopted to enhance the difference between the two embeddings, reduce collinearity and mutual interference of relational graph and topology graph network embeddings, and improve the robustness and accuracy of electrical equipment fault diagnosis model. and The HSIC constraint is: ; In the formula, and They are respectively and The Gram matrix; I is the identity matrix; e is a column vector of all ones; similarly... and The parallax is also enhanced through HSIC, with a parallax constraint of £. d ,in: ; Utilizing attention mechanisms right and and a common embedding By performing weighted fusion, a more comprehensive feature representation is obtained: ; To highlight key fault information, this embodiment introduces a three-branch attention mechanism to adaptively weight and fuse three types of embedded features. The specific calculation process is as follows: Let the node embedding dimension be d, and the attention weight matrices of the three branches be WI, WF, and WC ∈ R. d ×1, with bias terms bI, bF, bC ∈ R. Global average pooling is performed on the node embeddings NI, NF, and NC of the three branches to obtain global feature vectors VI, VF, and VC. Attention scores for each branch are calculated using fully connected layers. ; The weight coefficients of each branch are obtained by Softmax normalization: ; The final fusion features are: This fusion feature integrates features unique to relational graphs, features unique to feature graphs, and common features of both types of graphs, achieving adaptive weighted fusion of multi-source information.
[0023] Step S4: Feature Optimization and Fault Classification Diagnosis Feature optimization: Based on the consistency constraint loss £C and independence constraint loss £ in step S3 d The network is jointly optimized to enhance the consistency of features between the two types of graphs while retaining the differentiated fault information of each branch and reducing feature collinearity interference.
[0024] Global pooling and classification: The optimized global features are subjected to global average pooling and input into a fully connected classification layer with a Softmax activation function. The model parameters are updated using cross-entropy loss to complete the fault identification and diagnosis of five states.
[0025] Experimental results description: like Figure 6 As shown, in order to verify the effectiveness of the method proposed in this embodiment, the diagnostic results of different signal-to-noise ratio models were compared. As the experimental results show, under the condition of no noise original signal, the diagnostic accuracy of this method for the five categories of electrical equipment status reached a high level, demonstrating excellent basic fault identification capability. After adding noise interference of different intensities, the overall accuracy of the model remained stable, with only a slight decrease in the diagnosis of category 1 and category 5. The accuracy of categories 2 to 4 remained at a high level under signal-to-noise ratio conditions of 10dB, 5dB and even 1dB.
[0026] The above experimental results verify the anti-interference ability and robustness advantages of the method of the present invention, indicating that by modeling the fault transmission relationship, mining the spatial topology features and using the multi-source feature fusion strategy, the noise interference problem in complex industrial sites can be effectively overcome. Even in a strong noise environment, the method can still stably achieve accurate fault diagnosis of electrical equipment, providing reliable technical support for equipment operation and maintenance.
[0027] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for fault diagnosis of electrical equipment based on fault propagation and spatial mining, characterized in that, include: Step S1: Collect vibration signals from multiple measuring points of electrical equipment, preprocess the vibration signals, calculate the symbol phase transfer entropy between each measuring point based on the preprocessed signals, and construct a relationship diagram representing the transmission relationship of fault information. Step S2: Extract the time-frequency domain features of the vibration signals at each measuring point, and construct a feature map representing the correlation of the measuring point features based on the time-frequency domain features; Step S3: Input the relationship graph and feature graph into a multi-channel adaptive graph convolutional network to jointly extract and fuse the structural features of the two types of graphs to obtain the fused global features; Step S4: Optimize the global features based on consistency constraints and independence criteria, and complete the fault identification and diagnosis of electrical equipment through a classification model.
2. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S1, the vibration signal is preprocessed, specifically including: The original vibration signal is converted into an instantaneous phase time sequence using Hilbert transform; The instantaneous phase time series is sequentially reconstructed in phase space and symbolized, and finally the symbolic phase transfer entropy is calculated.
3. The electrical equipment fault diagnosis method according to claim 2, characterized in that, The phase space reconstruction uses the pseudo-nearest neighbor method to determine the embedding dimension m, where the delay time τ is 3-4, the symbolic permutation order is 1, and the time step δ is 1.
4. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S1, constructing a relational graph representing the transmission relationship of fault information specifically includes: using the symbol phase transfer entropy calculated between each measurement point as the element of the adjacency matrix to construct the relational graph.
5. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S2, the time-frequency domain features include time-domain statistical features and frequency-domain features; The time-domain statistical features include mean, variance, kurtosis, and peak factor, while the frequency-domain features include centroid frequency, mean square frequency, spectral entropy, and cepstral peak value.
6. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S2, a feature map representing the correlation of measurement point features is constructed, which specifically includes: calculating the similarity matrix using cosine similarity based on the time-frequency domain feature vectors of each measurement point, and establishing an adjacency matrix using the k-nearest neighbor method, thereby constructing the feature map.
7. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S3, the two types of graph structure features are jointly extracted and fused, specifically including: A multi-channel adaptive graph convolutional network is used to extract unique embedding features and shared common embedding features from the relationship graph and the feature graph, respectively. Global average pooling is performed on each unique embedding feature and the common embedding feature to obtain the corresponding global feature vector; An attention mechanism is introduced to calculate the attention weight coefficients of each feature vector. Based on the weight coefficients, the features of multiple classes are adaptively weighted and fused to obtain the fused global features.
8. The electrical equipment fault diagnosis method according to claim 7, characterized in that, The computational process of the attention mechanism includes: Set up three sets of attention weight matrices and corresponding bias terms, and then perform operations on each pooled global feature vector with the corresponding weight matrix and bias term to obtain the initial attention score; The Softmax function is used to normalize all initial attention scores to obtain standardized attention weight coefficients; By combining the weight coefficients with the corresponding embedded features, a weighted summation is performed to complete the fusion of multi-source features.
9. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S4, optimizing the global features based on consistency constraints and independence criteria includes: The embedded features output from the relation graph and feature graph are subjected to L2 normalization, and a consistency constraint loss function is constructed to enhance the similarity between the feature outputs of the two types of graphs. The Hilbert-Schmidt independence criterion is used to construct an independence-constrained loss function, which weakens the collinearity interference between different embedded features and preserves differentiated fault information.
10. The electrical equipment fault diagnosis method according to claim 1, characterized in that, In step S4, the classification model is a fully connected classification layer equipped with a Softmax activation function, and the model is updated using cross-entropy loss. Finally, the fault type and status diagnosis results of the electrical equipment are output.