CN-122020596-A - Fault detection method based on converged network

Abstract

The invention discloses a fault detection method based on a fusion network, which relates to the technical field of fault detection of industrial equipment, wherein a convolutional neural network model is used for capturing local modes and variation trends through local feature extraction and sliding window operation on time sequence data of the equipment through a convolutional layer, a transducer module is used for helping a CNN model to capture global dependency relationship features in the time sequence data through a self-attention mechanism, a bidirectional gating recursion module is used for simultaneously considering forward and reverse information of the time sequence data and capturing front and rear dependency relationship features of the time sequence data, a cross attention mechanism module is introduced for feature fusion and then compressing the feature into features with fixed length through self-adaptive average pooling mapping, and the processed features are classified through a full connection layer to output a prediction result of fault categories.

Inventors

• CHANG YAOHUI
• JIANG YUAN
• WU KAIWEN
• LU JINHUAN
• LIU TANG
• LI JIN
• GENG RUNHUA
• YANG YINGJIAN
• ZHOU YUFEI
• LI ZIHENG
• JIANG HUI
• LI NAN

Assignees

• 石河子大学
• 新疆天业（集团）有限公司

Dates

Publication Date

20260512

Application Date

20260212

Claims (7)

1. 1.A fault detection method based on a converged network is characterized by comprising the following steps: s1, a convolutional neural network CNN model captures a local mode and a change trend by carrying out local feature extraction and sliding window operation on time sequence data of equipment through a convolutional layer; S2, a self-attention mechanism module converter helps the CNN model in S1 to capture the global dependency characteristics in the time sequence data; s3, a bidirectional gating recursion module BiGRU simultaneously considers forward and reverse information of time sequence data through a gating cyclic neural network GRU and captures the front-back dependency characteristics of the time sequence data; S4, introducing a cross attention mechanism module to combine the output information of the S2 and the S3 for feature fusion; s5, compressing the fusion characteristics based on the S4 output into characteristics with fixed length through self-adaptive average pooling mapping; and S6, classifying the features processed in the step S5 through a full connection layer, and finally outputting a prediction result of the fault class.
2. 2. The method for detecting faults based on the converged network of claim 1, wherein the time sequence data of the equipment in S1 is cut and normalized through a self-defined make_ datasets function to generate data meeting the input requirements of the converged network, and the data is converted into a DataLoader format of PyTorch and supports batch processing.
3. 3. A method for detecting faults based on a converged network as claimed in claim 2, wherein S1 a plurality of convolution layers are created by a make_layers method, each layer comprises convolution, a ReLU activation function and a pooling operation, each convolution layer extracts local features in data, the pooling operation reduces the spatial size of the data, the ReLU activation function increases the nonlinear expression capability after each convolution, and the maximum value in a local window is selected to compress the time axis length while retaining the strongest signal.
4. 4. The method for detecting faults based on the converged network of claim 3, wherein the specific process of S1 is as follows: S11, extracting the characteristics of the local mode from the time sequence data, scanning the time sequence in a sliding window mode through a plurality of one-dimensional convolution layers, and learning short-term information such as the change trend, the edge, the mutation and the like of signals in each local area; S12, adding a ReLU activation function after each group of convolution layers is operated, accessing a maximum pooling layer, and compressing the time axis length and reserving the strongest signal by selecting the maximum value in a window; s13, carrying out superposition optimization capturing of the local mode and the capability of the change trend on the basis of the S12.
5. 5. The method for detecting faults based on the converged network of claim 4, wherein: the specific process in S2 is as follows: s21, converting the local features extracted by the CNN model in the S1 into time steps in a sequence form as input of a transducer; S22, calculating attention weights according to the self-attention mechanism and carrying out weighted summation on the characteristics of all time steps to update the representation of the current time step; s23, merging a multi-head self-attention mechanism to learn different types of time dependency relationships in a plurality of subspaces simultaneously; s24, adding a feedforward neural network, residual connection and normalization processing means after the self-attention mechanism; s25, setting a two-layer stacked transducer encoder after the pooling layer of the S1 to improve the global representation capability.
6. 6. The method for detecting faults based on the converged network of claim 5, wherein the specific process of S3 is as follows: s31, forward GRU capturing history information by processing a time sequence according to a time step sequence through a gate control cyclic neural network GRU; S32, reverse GRU capturing future information by processing the time sequence according to a time step reverse sequence through a gate control cyclic neural network GRU; S33, carrying out BiGRU module double-layer stacking, wherein the first layer captures shallow layer time sequence dependence in the time dimension and outputs a context enhanced sequence, and the second layer further digs high-level time structure change on the basis.
7. 7. The method for detecting faults based on the fusion network of claim 6, wherein the specific process of S4 is that a cross attention mechanism module integrates two types of information of global dependency characteristics of a transducer and front-back dependency characteristics extracted by BiGRU, a Query-Key Value mechanism is used, biGRU output is used as a Query, the transducer output is used as a Key Key and a Value, and attention weights between the Key Key and the Value are calculated.

Description

Fault detection method based on converged network Technical Field The invention relates to the technical field of fault detection of industrial equipment, in particular to a fault detection method based on a converged network. Background With the continuous improvement of the automation and intelligence level of industrial equipment, equipment fault prediction and health monitoring become more and more important, the method can help us to discover potential problems of equipment in advance, avoid sudden faults, reduce equipment downtime and improve production efficiency, and the traditional fault diagnosis method generally relies on manual experience and rules to manually check and formulate complex rules to judge faults, so that the method is low in efficiency, has limited processing capacity for complex and high-dimensional time sequence data, and is difficult to deal with massive and diverse data in modern industry. In recent years, deep learning technology is gradually applied to fault detection due to the fact that the deep learning technology can automatically extract important features in data and identify rules of the data, particularly a Convolutional Neural Network (CNN) and a cyclic neural network (RNN) are successful in many scenes, however, single CNN or RNN still has limitations when processing time series data, and complex time series dependency relations in the data cannot be completely mined. Disclosure of Invention The invention aims to provide a fault detection method based on a fusion network, which combines a convolutional neural network CNN, transformer, biGRU and a Cross-Attention mechanism Cross-Attention to construct a new multi-module joint model for efficiently processing complex time sequence data and realizing accurate fault prediction. In order to achieve the above object, the present invention provides a method for detecting faults based on a converged network, comprising the following steps: s1, a convolutional neural network CNN model captures a local mode and a change trend by carrying out local feature extraction and sliding window operation on time sequence data of equipment through a convolutional layer; S2, a self-attention mechanism module converter helps the CNN model in S1 to capture the global dependency characteristics in the time sequence data; s3, a bidirectional gating recursion module BiGRU simultaneously considers forward and reverse information of time sequence data through a gating cyclic neural network GRU and captures the front-back dependency characteristics of the time sequence data; S4, introducing a cross attention mechanism module to combine the output information of the S2 and the S3 for feature fusion; s5, compressing the fusion characteristics based on the S4 output into characteristics with fixed length through self-adaptive average pooling mapping; and S6, classifying the features processed in the step S5 through a full connection layer, and finally outputting a prediction result of the fault class. Preferably, the time sequence data of the device in S1 is cut and normalized by a custom make_ datasets function to generate data meeting the input requirement of the fusion network, and the data is converted into a DataLoader format of PyTorch and supports batch processing. Preferably, in S1, a plurality of convolution layers are created by a make_layers method, each layer including convolution, a ReLU activation function, and a pooling operation, each convolution layer extracting local features in data, the pooling operation reducing the spatial size of the data, the ReLU activation function increasing the nonlinear expression capability after each convolution, compressing the time axis length by selecting the maximum value in the local window, while retaining the strongest signal. Preferably, the specific process of S1 is as follows: S11, extracting the characteristics of the local mode from the time sequence data, scanning the time sequence in a sliding window mode through a plurality of one-dimensional convolution layers, and learning short-term information such as the change trend, the edge, the mutation and the like of signals in each local area; S12, adding a ReLU activation function after each group of convolution layers is operated, accessing a maximum pooling layer, and compressing the time axis length and reserving the strongest signal by selecting the maximum value in a window; s13, carrying out superposition optimization capturing of the local mode and the capability of the change trend on the basis of the S12. Preferably, the specific process in S2 is as follows: s21, converting the local features extracted by the CNN model in the S1 into time steps in a sequence form as input of a transducer; S22, calculating attention weights according to the self-attention mechanism and carrying out weighted summation on the characteristics of all time steps to update the representation of the current time step; s23, merging a multi-head self-attenti