Open set fault diagnosis method and system based on graph small wavelength short-term memory network
By embedding discrete graph wavelet convolution and graph wavelet energy pooling into LSTM units, and combining the graph wavelet short-term memory network (GWLSTM) method with a fully connected classifier, the shortcomings of existing fault diagnosis methods in feature extraction and unknown fault identification in complex multivariable signals are solved, and multi-scale spatiotemporal feature adaptive extraction and reliable identification of unknown faults are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-12
AI Technical Summary
Existing deep learning-based fault diagnosis methods are insufficient in extracting robust features from complex multivariate signals and identifying unknown faults, making it difficult to meet the needs of open-set diagnosis.
A method based on graph wavelet short-term memory network (GWLSTM) is adopted. Multi-scale spatiotemporal feature extraction is performed in LSTM units by embedding discrete graph wavelet convolution, and a learnable soft contraction function is introduced. Combined with graph wavelet energy pooling module and fully connected classifier, reliable identification of unknown faults is achieved.
It effectively models the complex correlations of multivariate signals, improves the robustness of feature representation, enables the classification of known fault categories and the reliable identification of unknown faults, and enhances the accuracy and robustness of open set diagnosis.
Smart Images
Figure CN122020246A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an open-set fault diagnosis method and system based on graph-based short-term memory networks, belonging to the field of mechanical fault diagnosis. Background Technology
[0002] In recent years, deep learning-based intelligent fault diagnosis methods have been widely used in prediction and health management due to their powerful data-driven feature learning capabilities. However, existing methods still face two key challenges: first, how to extract robust and condition-independent features from complex multivariate signals to adapt to different operating conditions and environmental changes; and second, the recognition performance of traditional deep models significantly decreases when unknown or unseen fault types appear, making it difficult to meet the open-set diagnostic requirements under complex real-world conditions. Although graph neural networks can be used to model the spatial correlation of multi-source signals in recent years, existing methods still fall short in simultaneously capturing spatiotemporal dependencies and achieving robust open-set fault identification.
[0003] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art in this country. Summary of the Invention
[0004] This invention provides an open-set fault diagnosis method based on graph wavelet short-term memory networks. This method embeds discrete graph wavelet convolutions into LSTM units to achieve adaptive extraction of multi-scale spatiotemporal features. Simultaneously, a learnable soft contraction function is introduced to suppress noise-sensitive components and enhance feature robustness. Furthermore, a GWPool operation is designed to adaptively aggregate the node feature matrices into a global graph representation, which is then tested to calculate graph energy, enabling reliable identification of unknown faults.
[0005] The technical solution of this invention is:
[0006] According to a first aspect of the present invention, an open-set fault diagnosis method based on graph small-wavelength short-term memory networks is provided, comprising:
[0007] Step S1: Acquire multi-source sensor data under different states of the transmission; normalize the acquired multi-source sensor data and treat each sensor as a graph node to construct an adjacency matrix between nodes; at the same time, use a sliding window to generate a node feature matrix for each state mode; where the state includes normal state and fault state.
[0008] Step S2: Construct GWLSTM memory network units, and encode the node feature matrix at multiple scales by stacking GWLSTM memory network units, and fuse time series information and graph structure dependencies to obtain node representations;
[0009] Step S3: Construct a graph wavelet energy pooling module to adaptively aggregate node representations into a global graph representation based on graph wavelet energy pooling;
[0010] Step S4: Input the global graph representation into a two-layer fully connected classifier, and combine it with the energy function for open set discrimination to achieve the classification of known fault categories and reliable identification of unknown faults.
[0011] Furthermore, the stacked GWLSTM memory network units are represented as follows:
[0012] ;
[0013] ;
[0014] In the formula, It is the first layer of GWLSTM memory network unit The node representation learned after encoding X. It is a second-layer GWLSTM memory network unit based on The node representations learned after further encoding Represents the node feature matrix, It is an adjacency matrix.
[0015] Furthermore, the GWLSTM memory network unit organically combines discrete graph wavelet convolution with the long short-term memory network gating mechanism, and its update process is formally represented as follows:
[0016] ;
[0017] ;
[0018] ;
[0019] ;
[0020] in, , , , These represent the forgetting gate, input gate, cell input gate, and output gate, respectively. , , and Let represent the node feature matrices applied to the input time t. The frequency domain filter matrix is used to adjust the operation of the forget gate, input gate, cell input, and output gate respectively; , , and Let represent the frequency domain filter matrices applied to the hidden state, respectively, and let represent the hidden state at time t-1. Connect to the forget gate, input gate, cell input, and output gate; These are trainable parameters that control the filtering behavior; , , and These represent the bias terms of the forgetting gate, input gate, cell input, and output gate, respectively. Represents the discrete graph wavelet convolution operator. Represents the discrete graph wavelet inverse transform operator; It is a learnable soft contraction function.
[0021] Furthermore, the graph wavelet energy pooling module is represented as follows:
[0022] ;
[0023] in, The global graph representation obtained by the graph wavelet energy pooling module represents the graph representation. This represents the vector composed of the soft-thresholded wavelet coefficients of all nodes at the s-th scale and the v-th feature channel.
[0024] Furthermore, the process of achieving the classification of known fault categories and the reliable identification of unknown faults is divided into a training phase and a testing phase:
[0025] During the training phase, the proposed GWLSTM memory network model is trained using labeled known states with the standard cross-entropy loss function; the GWLSTM memory network model includes GWLSTM memory network units, graph wavelet energy pooling modules, and a fully connected classifier;
[0026] During the testing phase, an open-set classifier based on graph energy was introduced. The open-set classifier constructs an energy function that maps the graph representation to a scale energy, and this energy function is used as the scoring function of the open-set classifier for a given energy threshold.
[0027] ;
[0028] in, For an open-set classifier, if the energy value of the fault... Above the energy threshold If the probability is -1, it is considered an open-set fault and marked as -1; otherwise, the label corresponding to the highest probability will be used. Mark it.
[0029] According to a second aspect of the present invention, an open set fault diagnosis system based on graph small wavelength short-term memory network is provided, comprising a module of the open set fault diagnosis method based on graph small wavelength short-term memory network as described above.
[0030] According to a third aspect of the present invention, a terminal is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the open set fault diagnosis method based on graph small wavelength short-term memory network as described above.
[0031] The beneficial effects of this invention are:
[0032] This invention achieves adaptive extraction of multi-scale spatial-temporal features by embedding discrete graph wavelet convolutions in LSTM units, effectively modeling complex correlations between multivariate signals. It significantly improves the robustness of feature representation by introducing a learnable soft contraction function to suppress noise-sensitive components. Furthermore, it designs a GWPool operation to adaptively aggregate node-level features into a global graph representation, thereby compressing redundant structures and strengthening key patterns. Finally, the global graph representation is input into a two-layer fully connected classifier, and an energy function is used for open-set discrimination during the testing phase to achieve reliable classification of known fault categories and identification of unknown faults. Attached Figure Description
[0033] Figure 1 This is a flowchart of the method of the present invention.
[0034] Figure 2 Based on the schematic diagram of the gear transmission system fault simulation test bench in this embodiment.
[0035] Figure 3 This is a schematic diagram of the open set fault diagnosis process of GWLSTM in this embodiment.
[0036] Figure 4 This is a schematic diagram illustrating the score distribution of each method in implementing open set fault diagnosis on task M3 in this embodiment.
[0037] Figure 5 This is a schematic diagram of the classification confusion matrix for feature extraction using GWLSTM proposed on task M3 in this embodiment.
[0038] Figure 6 This is a schematic diagram of t-SNE for feature extraction using GWLSTM proposed on task M3 in this embodiment. Detailed Implementation
[0039] 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. 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. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.
[0040] Example 1: As Figures 1-6 As shown, according to a first aspect of the present invention, an open-set fault diagnosis method based on graph small-wavelength short-term memory networks is provided, comprising the following steps:
[0041] Step S1: Acquire multi-source sensor data under different states of the transmission; normalize the acquired multi-source sensor data, and treat each sensor as a graph node to construct an adjacency matrix between nodes. Simultaneously, a sliding window is used to generate the node feature matrix for each state mode. Among them, the status includes normal status and fault status.
[0042] Step S2: Construct GWLSTM memory network units. Multi-scale encoding of the node feature matrix is performed using stacked GWLSTM memory network units, fusing time-series information and graph structure dependencies to learn robust and discriminative node representations. The stacked GWLSTM memory network units are represented as follows:
[0043] ;
[0044] ;
[0045] In the formula, It is the first layer of GWLSTM memory network unit right The node representation learned after encoding. It is a second-layer GWLSTM memory network unit based on The node representations learned after further encoding Represents the node feature matrix, It is an adjacency matrix.
[0046] Step S3: Construct a graph wavelet pooling (GWPool) module to represent nodes based on graph wavelet pooling. Adaptive aggregation is used for global graph representation, thereby compressing redundant structures and enhancing key patterns. The graph wavelet energy pooling module is represented as follows:
[0047] ;
[0048] In the formula, The global graph representation obtained by the graph wavelet energy pooling module represents the graph representation. This represents the discrete graph wavelet convolution operator.
[0049] Step S4: Input the global graph representation into a two-layer fully connected classifier, and combine it with the energy function for open set discrimination to achieve the classification of known fault categories and reliable identification of unknown faults.
[0050] Furthermore, the GWLSTM memory network unit organically combines discrete graph wavelet convolution with the Long Short-Term Memory (LSTM) network gating mechanism, and its update process is formally represented as follows:
[0051] ;
[0052] ;
[0053] ;
[0054] ;
[0055] in, , , , These represent the forgetting gate, input gate, cell input gate, and output gate, respectively. , , and Let represent the node feature matrices applied to the input time t. The frequency domain filter matrix is used to adjust the operation of the forget gate, input gate, cell input and output gate respectively. The value of t is 1,...,k, where k is the number of data segments generated by the sliding window. , , and Let represent the frequency domain filter matrices applied to the hidden state, respectively, and let represent the hidden state at time t-1. Connect to the forget gate, input gate, cell input, and output gate; These are trainable parameters that control the filtering behavior; , , and These represent the bias terms for the forget gate, input gate, cell input, and output gate, respectively; Ω represents the discrete graph wavelet convolution operator. Represents the discrete graph wavelet inverse transform operator; It is a learnable soft contraction function.
[0056] Furthermore, the learnable soft contraction function Defined as:
[0057] ;
[0058] in, It is a learnable threshold for feature denoising. Initialize to e -4 .
[0059] Furthermore, the graph wavelet energy pooling module is represented as follows:
[0060] ;
[0061] in, The global graph representation obtained by the graph wavelet energy pooling module represents the graph representation. This represents the vector (containing low-pass and high-pass components) composed of the soft-thresholded wavelet coefficients of all nodes at the s-th scale and the v-th feature channel, where s and v represent the number of wavelet scales and the number of feature channels, respectively.
[0062] Furthermore, the process of achieving the classification of known fault categories and the reliable identification of unknown faults is divided into a training phase and a testing phase:
[0063] During the training phase, the proposed GWLSTM memory network model is trained using labeled known states and the standard cross-entropy loss function. The GWLSTM memory network model comprises stacked GWLSTM memory network units, a graph wavelet energy pooling module, and a fully connected classifier. The cross-entropy loss function... Specifically:
[0064] ;
[0065] ;
[0066] in, For the sample Predict the probability distribution of the labels. It is a sample The probability distribution of the true labels, where M represents the sample size. Number the sample; Represents an i-dimensional logical function for predicting the i-th class of states. It is a non-linear activation function. Indicates a fully connected layer;
[0067] After training, open set fault testing was performed. During the testing phase, an open set classifier based on graph energy was introduced, which constructed an energy function. This maps the graph representation to a scale energy; if the fully connected classifier during the training phase is represented as... Then the relationship between the energy function and the prediction logistic vector is: The energy function for any global graph representation can be defined as:
[0068] ;
[0069] in, Let f represent the energy value of H in the global graph, f represent the fully connected classifier during the training phase, y represent the true label, C represent the total number of known classes, and f(H) represent the total number of classes. [c] Let f represent the output score of the fully connected classifier f during the training phase on the global graph representation H in class c, where c represents the class number.
[0070] Using the energy function, the classification loss function in the testing phase becomes:
[0071] ;
[0072] in, This is the set of indices for the test set samples.
[0073] The classification loss function described above pushes down known categories and pulls up unknown categories. Therefore, the energy function can be used as the score function for an open-set classifier given an energy threshold.
[0074] ;
[0075] in, For an open-set classifier, if the energy value of the fault is higher than the energy threshold... If the probability is -1, it is considered an open set fault and marked as -1; otherwise, it is marked with the label corresponding to the highest probability. In this invention, the open set threshold is set to -5.
[0076] According to a second aspect of the present invention, an open-set fault diagnosis system based on graph small-wavelength short-term memory networks is provided, comprising modules of the open-set fault diagnosis method based on graph small-wavelength short-term memory networks described above. It should be noted that the functions of each module in the open-set fault diagnosis system based on graph small-wavelength short-term memory networks are similar to the functions of each step in the open-set fault diagnosis method based on graph small-wavelength short-term memory networks; therefore, to avoid repetition, these functions will not be elaborated here. For details, please refer to the descriptions of each step in the aforementioned method.
[0077] According to a third aspect of the present invention, a terminal is provided, including a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the open set fault diagnosis method based on graph small wavelength short-term memory network as described above.
[0078] Example 2: The following describes an optional embodiment of the present invention in conjunction with simulation:
[0079] This embodiment uses the MCC5-THU gearbox dataset collected by Tsinghua University, which fully considers the speed and load variations of gearboxes in industrial scenarios. The experiment was conducted in... Figure 2 The test was conducted on the test bench shown, which recorded gearbox fault data. The motor speed range was 0 to 3000 rpm, the load range was 0 to 20 Nm, and the sampling frequency was 12.8 kHz. The collected data came from eight sensors: acceleration sensors, speed sensors, and torque sensors for the motor and gearbox in the X, Y, and Z directions. A total of 20 states were collected in this experiment, including healthy state (normal state), missing tooth fault state, three different degrees of gear pitting fault state, three different degrees of gear wear fault state, three different degrees of gear tooth breakage fault state, three different degrees of gear tooth cracking fault state, and two combined faults under three different fault degrees, as shown in Table 1.
[0080] Table 1
[0081]
[0082] Based on the MCC5-THU transmission dataset, the diagnostic method of this invention is used, and the specific steps are as follows:
[0083] Step S1: First, perform Min-Max normalization on the multi-source sensor data collected in different states, mapping the data to the [0,1] interval to eliminate the influence of amplitude differences between different sensors, and obtain the normalized data. The states include normal states and fault states, and the specific normalization formula is as follows:
[0084] ;
[0085] in, This represents the maximum value in the corresponding sensor data under the current state. This represents the minimum value in the sensor data corresponding to the current state. This represents a specific sampling point of the sensor data in the current state.
[0086] Then, each sensor is treated as a graph node, and a sensor graph structure is constructed for each state mode. RadiusGraph is used to mine the relationships between the sensor nodes. A sliding window of size 1024 is used to segment the acquired sensor data, and the segmented graph segments are used as node features in a predefined sensor graph. Therefore, for each state, a set of graph data with the same graph structure but different node features can be obtained.
[0087] The adjacency matrix constructed above It can be represented as:
[0088] ;
[0089] In the formula, Representing the adjacency matrix The element in the p-th row and q-th column, As the radius threshold, Represents a node and The cosine similarity between them; N represents the total number of research objects, i.e., the number of sensors.
[0090] In this embodiment, data was collected under the condition of a fixed gear output shaft torque of 10 Nm, with the rotational speed changing from 0 to 2000 rpm within one cycle and then back to 0. Acceleration sensor data from the motor and gearbox in the X, Y, and Z directions were selected to generate graph-structured data (six sensors), and three open-set fault diagnosis tasks were formulated based on this data—labeled M1, M2, and M3 in Table 2. Task M1 involves one unknown fault category, M2 contains two, and M3 is the most challenging, containing three unknown faults. For each known fault type, 700 graphs were generated for each state through the above process, with 80% used for training and the remaining 20% for testing. Each unknown category contributed 140 graphs solely for testing purposes. Taking M1 as an example, a total of 11,200 graphs were used for training, and 2,940 graphs were used for evaluation, as shown in Table 2.
[0091] Table 2
[0092]
[0093] Step S2: Based on the graph data constructed in Step S1, discrete graph wavelet convolution is introduced to perform multi-scale spatial modeling of node features under the constraints of the graph topology. Specifically, based on the discrete graph wavelet transform operator, low-pass and high-pass filtering is applied to node features at different scales to achieve multi-scale graph convolutional encoding, i.e.:
[0094] ;
[0095] in, Represents a learnable frequency domain filter. These are trainable parameters that control the filtering behavior; G represents the graph. Represents the discrete graph wavelet convolution operator. This represents the discrete graph wavelet inverse transform operator.
[0096] After multi-scale encoding of node features, the data is input into the gating unit, enabling the generation of node-level spatiotemporal fusion feature representations during the temporal recursion process. Specifically, to achieve spatial feature extraction of graph data during the temporal recursion process, the aforementioned discrete graph wavelet convolution is embedded into the gating mechanism of the traditional LSTM unit. This performs multi-scale modeling of the topological dependencies of multidimensional degenerate features, thereby obtaining node-level spatiotemporal fusion degenerate features and constructing the GWLSTM memory network unit. Its update process can be formally represented as follows:
[0097] ;
[0098] ;
[0099] ;
[0100] ;
[0101] in, , , , These represent the forgetting gate, input gate, cell input gate, and output gate, respectively. , , and Let represent the node feature matrix applied to the input time t, respectively. The frequency domain filter matrix is used to adjust the operation of the forget gate, input gate, cell input and output gate respectively. The value of t is 1,...,k, where k is the number of data segments generated by the sliding window. , , and represents the frequency domain filter matrix applied to the hidden state, respectively, connecting the hidden state at time t-1 to the forget gate, input gate, cell input, and output gate; are trainable parameters that control the filtering behavior; These are trainable parameters that control the filtering behavior; , , and These represent the bias terms for the forget gate, input gate, cell input, and output gate, respectively; Ω represents the discrete graph wavelet convolution operator. Represents the discrete graph wavelet inverse transform operator; A learnable soft contraction function is introduced after discrete graph wavelet decomposition to improve the robustness of the characteristics under noise and operating condition disturbances. Defined as:
[0102] ;
[0103] in, It is a learnable threshold for feature denoising. Initialize to e -4 .
[0104] Next, calculate the cell state c at time t. t and hidden state h t The details are as follows:
[0105] ;
[0106] ;
[0107] in, Represents the dot product. It is a non-linear activation function.
[0108] The GWLSTM memory network unit constructed by the above formula enables the model to dynamically model the degradation process in the time dimension, thereby realizing the joint extraction and fusion of spatiotemporal features.
[0109] Step S3, as follows Figure 3 As shown, the node-level hidden state h output by the above GWLSTM memory network unit at time t is obtained by using a pooling module based on graph wavelet energy. t The nodes in the graph are adaptively aggregated into a global graph representation, thereby compressing redundant structures and strengthening key patterns. Specifically, firstly, the node-level hidden states h are... t Discrete graph wavelet decomposition is performed to obtain low-pass and high-pass coefficients at different scales. Then, aggregation weights are adaptively determined based on the wavelet energy distribution corresponding to each scale, and the coefficients at each scale are weighted and fused to highlight the main degradation modes and reduce noise and redundant information. Its basic form can be expressed as:
[0110] ;
[0111] in, This represents the vector (containing low-pass and high-pass components) composed of the soft-thresholded wavelet coefficients of all nodes at the s-th scale and the v-th feature channel, where s and v represent the number of wavelet scales and the number of feature channels, respectively. It is a global graph representation obtained from the multi-scale features of the decomposition.
[0112] In step S4, such as Figure 3 As shown, the global graph represents... Input a two-layer fully connected classifier and combine it with an energy function. Open set discrimination is performed to achieve reliable classification of known fault categories and identification of unknown faults. This process can be divided into a training phase and a testing phase.
[0113] First, during the training phase, the labeled known states, namely the training samples in Table 2, are used, and the proposed GWLSTM model is trained with the standard cross-entropy loss function. The predicted logarithm of the graph representation is defined as:
[0114] ;
[0115] in, Represents an i-dimensional logic function for predicting the i-th type of fault. It is a non-linear activation function. This indicates a fully connected layer.
[0116] Subsequently, the cross-entropy loss function for model training can be expressed as:
[0117] ;
[0118] Among them, P r Predict the probability distribution of the label for sample r. This represents the probability distribution of the true labels of sample r, where M represents the number of samples and r is the sample ID. Open set fault testing is performed after training is complete.
[0119] During the testing phase, an open-set classifier based on graph energy was introduced, which constructs a function. This maps the graph representation to a scale energy. If we represent the classifier during the training phase as... The relationship between energy and the prediction logistic vector is: The energy function represented by any graph can be defined as:
[0120] ;
[0121] in, Let H represent the energy value of the global graph, f represent the classifier during the training phase, y represent the true label, and C represent the total number of faults of known categories. This represents the output score of classifier f for sample H in class c during the training phase, where c represents the class number.
[0122] Using the energy function described above, the classification loss function for the testing phase becomes:
[0123] ;
[0124] in, This is the set of indices for the test set samples.
[0125] The classification loss function described above pushes down known categories and pulls up unknown categories. Therefore, the energy function can be used as the score function for an open-set classifier given a threshold.
[0126] ;
[0127] in, For an open-set classifier, if the energy of the fault is higher than the energy threshold... If the probability is -1, it is considered an open set fault and marked as -1; otherwise, it is marked with the label corresponding to the highest probability. In this invention, the open set threshold is set to -5.
[0128] The test samples in Table 2 were tested using the open set fault classifier described above.
[0129] In this example, this disclosure selects two advanced feature extraction networks as comparison models: Multi-Resolution Feature Graph Convolutional Network (MRFGCN) and Graph Sample and Aggregation Network (GraphSage). The metrics used are Acc, RAcc, and OSR. Acc represents accuracy, RAcc represents the model's ability to correctly identify known classes on a test set containing unknown classes, and OSR represents the proportion of unknown classes correctly identified by the model. During model training, each model is trained for 100 epochs with a batch size of 100. The Adam optimizer with an initial learning rate of 0.001 is selected for model optimization. Each method is evaluated after five independent runs, and the final performance is reported as the average result. Furthermore, only the GWLSTM memory network model of this invention uses an open-set classifier based on graph energy for open-set classification, while the other methods rely on confidence scores based on softmax. The comparative experimental results are shown in Table 3 below:
[0130] Table 3
[0131]
[0132] The results show that the proposed GWLSTM achieves the best open-set fault diagnosis performance across all tasks. In the most complex open-set fault diagnosis task M3, GWLSTM improves the OSR metric by 10.1% and 11.54% compared to MRFGCN and GraphSage, respectively. Furthermore, it can be observed that the proposed method achieves the smallest standard deviation compared to the comparative methods, thus demonstrating its effectiveness. Figure 4 The score distribution of GWLSTM and two contrasting methods on the M3 task is shown. Figure 5 and Figure 6 The corresponding confusion matrix and t-SNE visualization of features extracted by GWLSTM are shown. As mentioned earlier, from Figures 4-6 As can be seen, the GWLSTM of this invention can effectively distinguish the scores of known and unknown samples, and clearly differentiate between them. In contrast, the baseline method exhibits significant overlap, indicating that the known and unknown categories are more easily confused. Furthermore, the confusion matrix and t-SNE results confirm that GWLSTM achieves highly accurate classification of both known and unknown faults, with only a small number of misclassifications. These findings further validate the superiority of the proposed method.
[0133] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.
Claims
1. A method for open-set fault diagnosis based on graph-based short-term memory networks, characterized in that, include: Step S1: Acquire multi-source sensor data under different conditions of the transmission; The acquired multi-source sensor data is normalized, and each sensor is treated as a graph node to construct an adjacency matrix between nodes. At the same time, a sliding window is used to generate a node feature matrix for each state mode, where the state includes normal state and fault state. Step S2: Construct GWLSTM memory network units, and encode the node feature matrix at multiple scales by stacking GWLSTM memory network units, and fuse time series information and graph structure dependencies to obtain node representations; Step S3: Construct a graph wavelet energy pooling module to adaptively aggregate node representations into a global graph representation based on graph wavelet energy pooling; Step S4: Input the global graph representation into a two-layer fully connected classifier, and combine it with the energy function for open set discrimination to achieve the classification of known fault categories and reliable identification of unknown faults.
2. The open-set fault diagnosis method based on graph small-wavelength short-term memory network according to claim 1, characterized in that, The stacked GWLSTM memory network units are represented as follows: ； ； In the formula, It is the first layer of GWLSTM memory network unit right The node representation learned after encoding. It is a second-layer GWLSTM memory network unit. based on The node representations learned after further encoding, Represents the node feature matrix, It is an adjacency matrix.
3. The open-set fault diagnosis method based on graph small-wavelength short-term memory network according to claim 1, characterized in that, The GWLSTM memory network unit organically combines discrete graph wavelet convolution with the long short-term memory network gating mechanism, and its update process is formally represented as follows: ； ； ； ； in, , , , These represent the forgetting gate, input gate, cell input gate, and output gate, respectively. , , and They respectively represent the values applied to the input time. Node feature matrix The frequency domain filter matrix is used to adjust the operation of the forget gate, input gate, cell input, and output gate respectively; , , and Let each represent a frequency domain filter matrix applied to the hidden state, and let each represent a frequency domain filter matrix applied to the hidden state. Hidden state of time Connect to the forget gate, input gate, cell input, and output gate; These are trainable parameters that control the filtering behavior; , , and These represent the bias terms of the forgetting gate, input gate, cell input, and output gate, respectively. Represents the discrete graph wavelet convolution operator. Represents the discrete graph wavelet inverse transform operator; It is a learnable soft contraction function.
4. The open-set fault diagnosis method based on graph small-wavelength short-term memory network according to claim 1, characterized in that, The graph wavelet energy pooling module is represented as follows: ； in, The global graph representation obtained by the graph wavelet energy pooling module; Indicates the first Scale, First The vector consisting of the soft-thresholded wavelet coefficients of all nodes on each feature channel.
5. The open-set fault diagnosis method based on graph small-wavelength short-term memory network according to claim 1, characterized in that, The process of achieving reliable classification of known fault categories and reliable identification of unknown faults is divided into a training phase and a testing phase: During the training phase, the proposed GWLSTM memory network model is trained using labeled known states with the standard cross-entropy loss function; the GWLSTM memory network model includes GWLSTM memory network units, graph wavelet energy pooling modules, and a fully connected classifier; During the testing phase, an open-set classifier based on graph energy was introduced. The open-set classifier constructs an energy function that maps the graph representation to a scale energy, and this energy function is used as the scoring function of the open-set classifier for a given energy threshold. ； in, For an open-set classifier, if the energy value of the fault... Above the energy threshold If the probability is -1, it is considered an open-set fault and marked as -1; otherwise, the label corresponding to the highest probability will be used. Mark it.
6. An open-set fault diagnosis system based on graph-based short-term memory networks, characterized in that, The module includes the open set fault diagnosis method based on graph small wavelength short-term memory network as described in any one of claims 1-5.
7. A terminal, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, the processor being configured to perform the steps of the open set fault diagnosis method based on graph small wavelength short-term memory networks as described in any one of claims 1-5.