A fault detection method, system, device, and medium

Abstract

The application discloses a fault detection method, system, device and medium, and relates to the field of industrial fault analysis.The method comprises the following steps: obtaining operation data of a target device; dividing the operation data into a plurality of target window data by using a set sliding window; inputting the plurality of target window data into a fault diagnosis classification model to obtain fault information of the target device; wherein a multilayer convolutional neural network extracts features of the plurality of target window data to obtain target feature information data corresponding to each target window data; a multilayer bidirectional long short-term memory network performs time sequence calculation on each target feature information data to obtain a plurality of target time sequence feature data; a trained classifier respectively performs attention mechanism calculation on each target time sequence feature data to obtain fusion feature data of the target, and outputs the fault information of the target device according to the fusion feature data of the target; and the application can realize accurate fault detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial fault analysis, and in particular to a fault detection method, system, device and medium. BACKGROUND

[0002] In recent years, with the development of industrial productivity, various industrial production lines have made important contributions to the development of social productivity. In the actual production environment, industrial production equipment will be subjected to continuous vibration and impact, which leads to the wear and aging of equipment materials and parts, thereby causing the industrial equipment to be prone to failure. When the producer realizes the failure, a large number of defective products may have been produced, or even the entire industrial equipment has collapsed and stopped, thereby causing huge losses. If failure prediction can be made before failure occurs, and the parts that will have problems are replaced in advance, the life of the industrial equipment can be improved, and the sudden failure of a device can be avoided to have a serious impact on the entire industrial production. Nowadays, with the popularization of the concept of "smart factory", various industrial equipment is equipped with corresponding sensors, and it is easy to collect data such as vibration, temperature, current, voltage, etc. of the equipment. Through the analysis of these real-time sensor data, the failure prediction of the industrial equipment is a key method to effectively avoid production accidents and improve industrial production efficiency.

[0003] Traditional fault diagnosis methods, such as fault physics (Li et al.; Yang et al.; Zhu et al.) and fault tree analysis (Kabir et al.), generally focus on the operational mechanism or theoretical analysis of the system. They have been widely used in the field of fault diagnosis with high reliability requirements such as aerospace and electronics. However, when faced with complex systems like chemical processes, they are usually not feasible because it is difficult to mathematically model or analyze these complex systems.

[0004] With the development of sensor technology, a lot of data related to system operation can be easily collected and obtained. These heterogeneous and multi-source data can involve rich fault information. Therefore, in recent years, many data-driven methods (Cai et al.; El Koujok et al.; Li 2018 et al.; Serdio et al.; Zhang et al.) have been developed for fault diagnosis using these collected data, and they have been widely applied in various industrial sectors. In particular, data-driven methods based on deep learning (Lei et al.; Li et al.; Rodríguez Ramos et al.; Zhang et al.; Wang et al.) have achieved remarkable results in complex system fault diagnosis due to their advantages in feature extraction.

[0005] However, the diagnosis of the fault through the collected data does not consider that the performance of the electronic product will degrade over time, and therefore the time delay of the fault occurrence also has a certain influence on the occurrence of the fault. Therefore, how to realize accurate detection of the fault is crucial. SUMMARY

[0006] The purpose of the present application is to provide a fault detection method, system, device and medium, which can realize accurate detection of the fault.

[0007] To achieve the above purpose, the present application provides the following scheme:

[0008] A fault detection method, the method comprising:

[0009] obtaining running data of a target device;

[0010] dividing the running data into a plurality of target window data by using a set sliding window;

[0011] inputting the plurality of target window data into a fault diagnosis classification model to obtain fault information of the target device; the fault information comprising: fault attribute and fault occurrence time;

[0012] The fault diagnosis classification model comprises: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network and a trained classifier.

[0013] The multi-layer convolutional neural network is used for feature extraction of the plurality of target window data to obtain target feature information data corresponding to each target window data.

[0014] The multi-level bidirectional long short-term memory network is used for time sequence calculation of each target feature information data to obtain a plurality of target time sequence feature data.

[0015] The trained classifier is used for attention mechanism calculation of each target time sequence feature data to obtain target fusion feature data, and outputs the fault information of the target device according to the target fusion feature data.

[0016] Optionally, the determination method of the fault diagnosis classification model comprises:

[0017] obtaining training data of the target device; the training data comprising: running data of known fault information;

[0018] dividing the running data in the training data into a plurality of training window data by using a set sliding window;

[0019] construct a classification network; the classification network comprises: a plurality of layers of convolutional neural networks, a plurality of layers of bidirectional long short-term memory networks and a classifier connected in sequence;

[0020] input each of the training window data into the plurality of layers of convolutional neural networks in the classification network for feature extraction to obtain training feature information data corresponding to each of the training window data;

[0021] input each of the training feature information data into the plurality of layers of bidirectional long short-term memory networks as input, perform time sequence calculation on each of the training feature information data to obtain a plurality of training time sequence feature data;

[0022] input each of the training time sequence feature data and the corresponding fault information into the classifier as input, perform attention mechanism calculation on each of the training time sequence feature data to obtain training fusion feature data, and train parameters of the classifier with the minimum classification loss function as the target to obtain the trained classifier; the parameters include: weight matrix and bias matrix;

[0023] the fault diagnosis classification model comprises: the plurality of layers of convolutional neural networks, the plurality of layers of bidirectional long short-term memory networks and the trained classifier.

[0024] Optionally, input each of the training window data into the plurality of layers of convolutional neural networks in the classification network for feature extraction to obtain training feature information data corresponding to each of the training window data, specifically comprising:

[0025] input each of the training window data into the convolution kernel of the plurality of layers of convolutional neural networks for convolution operation to obtain a plurality of output convolution results; wherein the convolution kernel comprises a plurality of channels; each of the channels corresponds to one of the training window data;

[0026] stack the output convolution results at corresponding positions to obtain a plurality of training feature information data.

[0027] Optionally, input each of the training feature information data into the plurality of layers of bidirectional long short-term memory networks as input, perform time sequence calculation on each of the training feature information data to obtain a plurality of training time sequence feature data, specifically comprising:

[0028] input each of the training feature information data into the reset layer of the plurality of layers of bidirectional long short-term memory networks for nonlinear activation operation to obtain a plurality of nonlinear results; the reset layer comprises a plurality of forward long short-term memory network units and a plurality of reverse long short-term memory network units; the nonlinear results include forward nonlinear results and reverse nonlinear results;

[0029] For any of the nonlinear results, the nonlinear results are input to an update layer in the multi-level bidirectional long short-term memory network, the forward nonlinear results and the backward nonlinear results are spliced to obtain output results;

[0030] The output results are input to an output layer in the multi-level bidirectional long short-term memory network for multiplication operation to obtain a plurality of training time sequence feature data.

[0031] Optionally, according to the fusion feature data of the target, the trained classifier is specific to output the fault information of the target device according to the fusion feature data of the target.

[0032] According to the fusion feature data of the target, the trained classifier is specific to output the fault information of the target device according to the fusion feature data of the target. f y att +b f ) output the fault information of the target device; wherein y is the fault information of the target device; softmax is a normalized exponential function; W f is a weight matrix in the trained classifier; b f is a bias matrix in the trained classifier; y att is the fusion feature data of the target.

[0033] A fault detection system, the system comprises:

[0034] A data acquisition module for acquiring running data of a target device;

[0035] A division module for dividing the running data into a plurality of target window data by using a set sliding window;

[0036] An output module for inputting a plurality of the target window data into a fault diagnosis classification model to obtain fault information of the target device; the fault information includes fault attributes and fault occurrence time;

[0037] The fault diagnosis classification model comprises a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network and a trained classifier.

[0038] The multi-layer convolutional neural network is used for feature extraction of a plurality of the target window data to obtain target feature information data corresponding to each of the target window data.

[0039] The multi-level bidirectional long short-term memory network is used for time sequence calculation of each of the target feature information data to obtain a plurality of target time sequence feature data.

[0040] The trained classifier is used for attention mechanism calculation on each target time sequence feature data respectively, to obtain fusion feature data of the target, and outputs the fault information of the target device according to the fusion feature data of the target.

[0041] An electronic device comprising a memory for storing a computer program and a processor for running the computer program to make the electronic device execute the fault detection method.

[0042] A computer readable storage medium storing a computer program, which is executed by a processor to implement the fault detection method.

[0043] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0044] The present application provides a fault detection method, system, device and medium, through a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network and a trained classifier, feature extraction, time sequence calculation and attention mechanism operation are performed on target window data respectively, and finally the detection of fault information is completed, the present application realizes the acquisition of time sequence data by adopting a multi-level bidirectional long short-term memory network, so as to consider the time delay of the fault, thereby improving the accuracy of fault detection, thereby, the present application can realize accurate fault detection. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below, and obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0046] Figure 1 The flowchart of the fault detection method provided by the embodiment of the present application;

[0047] Figure 2 The flowchart of the fault detection algorithm in actual application provided by the embodiment of the present application;

[0048] Figure 3 The multi-level bidirectional long short-term memory network structure diagram provided by the embodiment of the present application;

[0049] Figure 4 The structure diagram of the fault detection system provided by the embodiment of the present application.

[0050] Symbol explanation:

[0051] Data acquisition module-1, division module-2, output module-3. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0053] Generally, in practical applications, data-driven fault diagnosis can be regarded as a classification task. Fault diagnosis using data-driven methods usually includes the following four steps: data preprocessing, feature extraction, classifier construction and fault diagnosis. For various systems such as power, pneumatic, hydraulic network, chemical process, long transmission line, robot, etc., the occurrence of faults may have a certain time delay. In the fault diagnosis of these systems, the existence of the time delay of fault occurrence must be considered, otherwise the performance of fault diagnosis will be greatly reduced. Delayed fault diagnosis has been widely studied in circuit systems. Recently, a delayed circuit system fault prediction method based on FFT, PCA and CNN is proposed (Khalil et al.). In the fault diagnosis of complex electronic systems, Cai et al. consider that the performance of electronic products will degrade over time, and use dynamic Bayesian networks to simulate the dynamic degradation process of electronic products. In the fault diagnosis of rolling bearings, Liu et al. use RNN to process the time dependence of signals. From the above analysis, it can be seen that the fault diagnosis method based on deep learning is more suitable for complex systems. In recent years, although some deep learning-based fault diagnosis methods have been proposed, they rarely consider the time delay of fault occurrence.

[0054] Therefore, it is meaningful to propose a fault diagnosis method for complex systems based on deep learning, which needs to consider both the feature extraction method based on deep learning and the time delay of fault occurrence.

[0055] The purpose of the present application is to provide a fault detection method, system, device and medium, which can realize accurate detection of faults.

[0056] In order to make the above-mentioned purposes, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0057] Embodiment 1

[0058] As shown in Figure 1 The embodiments of the present application provide a fault detection method, which comprises:

[0059] Step 100: acquiring running data of a target device.

[0060] Step 200: dividing the running data into multiple target window data by using a set sliding window.

[0061] Step 300: inputting the multiple target window data into a fault diagnosis classification model to obtain fault information of the target equipment; the fault information includes: fault attribute and fault occurrence time.

[0062] The fault diagnosis classification model includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a trained classifier.

[0063] The multi-layer convolutional neural network is used for feature extraction of the multiple target window data to obtain target feature information data corresponding to each target window data.

[0064] The multi-level bidirectional long short-term memory network is used for time sequence calculation of each target feature information data to obtain multiple target time sequence feature data.

[0065] The trained classifier is used for attention mechanism calculation of each target time sequence feature data respectively to obtain target fusion feature data, and outputs the fault information of the target equipment according to the target fusion feature data.

[0066] Specifically, according to the target fusion feature data, the trained classifier is specifically used for:

[0067] outputting the fault information of the target equipment according to the target fusion feature data by using a formula y = softmax(W f y att +b f ); wherein y is the fault information of the target equipment; softmax is a normalized exponential function; W f is a weight matrix in the trained classifier; b f is a bias matrix in the trained classifier; and y att is the target fusion feature data.

[0068] Specifically, the determination method of the fault diagnosis classification model specifically includes:

[0069] obtaining training data of the target equipment; the training data includes running data with known fault information.

[0070] Dividing the running data in the training data into multiple training window data by using a set sliding window.

[0071] In addition, in actual application, the actual operation process is shown in Figure 2First, the multivariate time series data needs to be processed by a sliding window to obtain fault data samples considering time delay.

[0072] The sliding window processing of the fault data obtains two-dimensional data of the time sequence window, which is as follows:

[0073] Assume is the original multivariate time series data, x t is the observation vector of the system at time t, m represents the number of observation attributes, and n represents the length of the observation time. Y = {y i |T=1,2,3,...,n} is the label of X, then the original data sample can be represented as D = {(X, Y)}. The sliding window is represented by a rectangular box, the length of which is the number of attributes of X, and the width of which is d (0 < d ≤ n). The data sample for training is obtained continuously from X by the sliding window. The step size of the sliding window movement is an integer λ (0 < λ ≤ n-d). During the i-th movement of the sliding window, assume that the subsequence of X is represented as X[i:i+d], then the label of X[i:i+d] is Y[i+d], where [p:q] represents the operation of extracting elements between p and q from an ordered set. Thus, the sample (X[i:i+d], Y[i+d]) will be used as input.

[0074] The original one-dimensional sample (the data at a certain time in the multivariate time series data) only contains feature information. After sliding window processing, the one-dimensional sample is converted into a two-dimensional sample, which integrates feature information and time delay information. Using these two-dimensional samples to train the fault diagnosis classification model can enable it to learn both feature information and time delay information. In addition, the size of the feature information and the time delay information contained in these two-dimensional samples can be adjusted according to actual conditions: the size of the feature information can be adjusted by adjusting the size of m, and the size of the time delay information can be adjusted by adjusting the width of the sliding window. By integrating more attributes and longer time delays in the sliding window processing, two-dimensional samples more suitable for complex system fault diagnosis can be obtained.

[0075] Construct a classification network; the classification network comprises: a plurality of layers of convolutional neural networks, a plurality of layers of bidirectional long short-term memory networks, and a classifier connected in sequence.

[0076] Each training window data is input into the plurality of layers of convolutional neural networks in the classification network for feature extraction to obtain training feature information data corresponding to each training window data.

[0077] Each training window data is input into the plurality of layers of convolutional neural networks in the classification network for feature extraction to obtain training feature information data corresponding to each training window data, which specifically comprises:

[0078] The respective training window data are input into convolution kernels of the multi-layer convolutional neural network, convolution operation is performed, and a plurality of output convolution results are obtained; wherein the convolution kernel comprises a plurality of channels; each channel corresponds to a training window data.

[0079] The plurality of output convolution results are accumulated and stacked at corresponding positions to obtain a plurality of training feature information data.

[0080] In other words, the multi-layer convolutional neural network is designed to perform serialized feature extraction, and zero padding operation is used to maintain the dimensionality of the output and the input equal. The convolution operation process can be described as:

[0081]

[0082] wherein W c , b c and f represent the weight of the convolution kernel, the bias term and the activation function, respectively. The symbol represents the convolution operation. c i is the convolution result corresponding to the i-th obtained window data of the sliding window.

[0083] It can be defined as the local region of the input x t：t+d = [x t , x t+1 , x t+2 ..., x t+d ,] d represents the multiplication of elements between the window width and the convolution kernel W c x t：t+d represents the input data sequence at time t.

[0084] In the algorithm implementation process, a plurality of convolution layers are stacked to extract more descriptive feature expressions. Finally, the features extracted by the multi-layer convolution can be represented as:

[0085]

[0086] The respective training feature information data are input into the multi-level bidirectional long short-term memory network, and time series calculation is performed on the respective training feature information data to obtain a plurality of training time series feature data.

[0087] The structure diagram of the multi-level bidirectional long short-term memory network is shown in Figure 3 . Wherein the respective training feature information data are input into the multi-level bidirectional long short-term memory network, and time series calculation is performed on the respective training feature information data to obtain a plurality of training time series feature data, which specifically includes:

[0088] Each training feature information data is input into the reset layer of a multi-level bidirectional long short-term memory network, and nonlinear activation operations are performed to obtain multiple nonlinear results. The reset layer includes multiple forward long short-term memory network units and multiple backward long short-term memory network units. The nonlinear results include forward nonlinear results and backward nonlinear results.

[0089] For any of the aforementioned nonlinear results, the nonlinear result is input into the update layer of a multi-level bidirectional long short-term memory network, and the forward nonlinear result and the reverse nonlinear result are concatenated to obtain the output result.

[0090] Multiple output results are input into the output layer of a multi-level bidirectional long short-term memory network for multiplication to obtain multiple training time-series feature data.

[0091] In other words, to better process time-series data, this invention designs a multi-level bidirectional long short-term memory network based on a gated recurrent (GRU) structure to achieve feature updates based on temporal correlations. Its main idea is to introduce an adaptive gating mechanism that selectively adds new information and forgets previous information.

[0092] The calculation and update process of GRU can be described as follows:

[0093] r t =σ(W r ·[h t-1 x t ]+b r );

[0094] u t =σ(W u ·[h t-1 x t ]+b u );

[0095]

[0096]

[0097] Where r, u, and h represent the output values ​​of the reset layer, update layer, and output layer, respectively; r and These represent the input and new memory of the cell, respectively. The subscripts t and t-1 represent the current and previous times of the state. W and b represent the weight matrix and bias term, respectively. σ and tanh represent the nonlinear activation function, and ⊙ represents element-wise multiplication.

[0098] A unidirectional GRU can only predict the output of the current time based on information from the previous time step. However, in practical industrial applications of fault diagnosis, the output of the current state depends not only on previous states but also on future states. Therefore, applying a bidirectional GRU can more effectively model temporal information. Specifically, a bidirectional GRU consists of two GRUs superimposed on each other, with the output determined by the states of these two GRUs. The first layer is forward-looking, and the second layer is backward-looking, with the hidden state h... t The calculation process is as follows:

[0099]

[0100]

[0101]

[0102] Where U and W represent the weight matrices of the hidden state and the input, respectively. Furthermore, b represents the bias term. The arrows indicate the direction of time transition. In this invention, to more effectively model temporal information, multiple bidirectional GRUs are stacked to form a multi-level bidirectional long short-term memory network, enabling temporal feature extraction and updating within the model.

[0103] Each training time-series feature data and the corresponding fault information are used as input to the classifier. Attention mechanism is applied to each training time-series feature data to obtain the fused feature data for training. The parameters of the classifier are trained with the goal of minimizing the classification loss function to obtain the trained classifier. The parameters include: weight matrix and bias matrix.

[0104] The fault diagnosis classification model includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a trained classifier.

[0105] To measure the impact of different temporal feature data on the detection results, this invention employs an attention mechanism to quantify the importance of the temporal feature data corresponding to each sliding window. Specifically, it utilizes the attention mechanism to explore the importance of temporal window features, thereby achieving effective fusion of temporal features.

[0106] The feature vectors for each time series window, i.e., the feature vectors of the time series feature data, are:

[0107] H = [h0, h1, h2, ..., h t ];

[0108] The computational process of the attention mechanism is as follows:

[0109]

[0110] In the actual operation process, the values of the input vector matrices Q, K and V are H. The attention mechanism assigns a corresponding importance weight to each state vector, and the weight matrix is defined as W F and Attention(Q, K, V) represents the updated time sequence feature vector. The present application realizes the progressive distribution of the importance of the time sequence window features by stacking multiple layers of attention mechanisms, which is beneficial to the subsequent time sequence feature fusion operation:

[0111]

[0112] w n represents the normalized weight of each sliding window feature. y att The fusion feature data obtained after the multi-layer attention mechanism represents the overall feature description of the fault sample.

[0113] Then, a softmax classifier is used for the final fault diagnosis, i.e., fault condition classification. The classification loss function is:

[0114]

[0115] wherein, represents the probability that the ith sample is predicted as the ground truth, i.e., the probability that the fusion feature data corresponding to the ith movement at this time is predicted as fault data.

[0116] represents the sum of the probabilities that the fusion feature data corresponding to the ith movement at this time (sample i) is predicted as other categories; M represents the number of other categories, N represents the number of fusion feature data in the batch, j is the serial number of other categories, and L is the classification loss function. The classification loss function aims to improve the classification effect of the classifier by constraining the predicted value y based on the loss value between the fault prediction result and the actual value, and guides the parameter optimization.

[0117] Embodiment 2

[0118] As shown in Figure 4 , the present application provides a fault detection system, which comprises a data acquisition module 1, a division module 2 and an output module 3.

[0119] The data acquisition module 1 is used to acquire the running data of the target device.

[0120] The division module 2 is used to divide the running data into a plurality of target window data by using a set sliding window.

[0121] The output module 3 is used to input the plurality of target window data into a fault diagnosis classification model to obtain fault information of the target device; the fault information comprises a fault attribute and a fault occurrence time.

[0122] The fault diagnosis classification model includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a trained classifier.

[0123] Multilayer convolutional neural networks are used to extract features from multiple target window data to obtain target feature information data corresponding to each target window data.

[0124] Multi-level bidirectional long short-term memory networks are used to perform temporal calculations on the feature information data of various targets to obtain multiple target temporal feature data.

[0125] The trained classifier is used to perform attention mechanism calculations on the temporal feature data of each target to obtain the fused feature data of the target, and outputs the fault information of the target device based on the fused feature data of the target.

[0126] Example 3

[0127] This invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the fault detection method in Embodiment 1.

[0128] In one embodiment, the aforementioned electronic device may be a server.

[0129] As an optional implementation, this embodiment of the invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the fault detection method in Embodiment 1.

[0130] Advantages of this invention:

[0131] This invention proposes a fault detection algorithm based on a multi-level bidirectional long short-term memory network and an attention mechanism. During the feature extraction process of fault samples, a sliding window approach is used to consider the time delay of the fault. The multi-level bidirectional long short-term memory network is used to model the temporal dependencies of the data. The attention mechanism is introduced to measure the importance of temporal features, thereby obtaining descriptive sample feature representations that can significantly improve fault diagnosis accuracy, such as in industrial scenarios.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0133] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A fault detection method, characterized in that, The method includes: Obtain the operating data of the target device; The running data is divided into multiple target window data by setting a sliding window; Multiple target window data are input into a fault diagnosis classification model to obtain fault information of the target device; the fault information includes: fault attributes and fault occurrence time; The fault diagnosis classification model includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a trained classifier; The multi-layer convolutional neural network is used to extract features from multiple target window data to obtain target feature information data corresponding to each target window data. The multi-level bidirectional long short-term memory network is used to perform temporal calculations on each of the target feature information data to obtain multiple target temporal feature data. The trained classifier is used to perform attention mechanism calculations on the temporal feature data of each target to obtain the fused feature data of the target, and outputs the fault information of the target device based on the fused feature data of the target. Each of the target window data is input into the convolution kernel of the multi-layer convolutional neural network to perform convolution operations, resulting in multiple corresponding output convolution results; wherein, the convolution kernel includes multiple channels; each channel corresponds to one of the training window data; The corresponding output convolution results are accumulated and stacked at their respective positions to obtain the target feature information data corresponding to each target window data.

2. The fault detection method according to claim 1, characterized in that, The method for determining the fault diagnosis classification model specifically includes: Acquire training data for the target device; the training data includes: operational data with known fault information; The running data in the training data is divided into multiple training window data by setting a sliding window; Construct a classification network; the classification network includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a classifier connected in sequence; Each training window data is input into a multi-layer convolutional neural network in the classification network for feature extraction, thereby obtaining training feature information data corresponding to each training window data. Each of the training feature information data is used as the input of the multi-level bidirectional long short-term memory network, and time-series calculations are performed on each of the training feature information data to obtain multiple training time-series feature data. Each training time-series feature data and the corresponding fault information are used as input to the classifier. Attention mechanism is calculated on each training time-series feature data to obtain the fused feature data for training. The parameters of the classifier are trained with the goal of minimizing the classification loss function to obtain the trained classifier. The parameters include: weight matrix and bias matrix. The fault diagnosis classification model includes: the multi-layer convolutional neural network, the multi-level bidirectional long short-term memory network, and the trained classifier.

3. The fault detection method according to claim 2, characterized in that, Each training window data is input into a multi-layer convolutional neural network in the classification network for feature extraction, resulting in training feature information data corresponding to each training window data, specifically including: Each training window data is input into the convolution kernel of the multi-layer convolutional neural network to perform convolution operations and obtain multiple output convolution results; wherein, the convolution kernel includes multiple channels; each channel corresponds to one training window data; The multiple output convolution results are accumulated and stacked at corresponding positions to obtain multiple training feature information data.

4. The fault detection method according to claim 2, characterized in that, Each of the training feature information data is used as input to the multi-level bidirectional long short-term memory network. Temporal calculations are performed on each of the training feature information data to obtain multiple training temporal feature data, specifically including: Each of the training feature information data is input into the reset layer of the multi-level bidirectional long short-term memory network, and nonlinear activation operation is performed to obtain multiple nonlinear results; the reset layer includes multiple forward long short-term memory network units and multiple backward long short-term memory network units; the nonlinear results include forward nonlinear results and backward nonlinear results; For any of the aforementioned nonlinear results, the nonlinear result is input into the update layer of the multi-level bidirectional long short-term memory network, and the forward nonlinear result and the reverse nonlinear result are concatenated to obtain the output result; Multiple output results are input into the output layer of the multi-level bidirectional long short-term memory network for multiplication to obtain multiple training time-series feature data.

5. The fault detection method according to claim 1, characterized in that, Based on the target's fused feature data, the trained classifier is specifically used for outputting fault information of the target device based on the target's fused feature data, specifically for: Formulas are used based on the fusion feature data of the target. Output the fault information of the target device; wherein... y For fault information of the target device; softmax It is a normalized exponential function; W f This is the weight matrix in the trained classifier; b f This is the bias matrix in the trained classifier; y att The goal is to fuse feature data.

6. A fault detection system, characterized in that, The system includes: The data acquisition module is used to acquire the operating data of the target device; The partitioning module is used to divide the running data into multiple target window data using a set sliding window; The output module is used to input multiple target window data into the fault diagnosis classification model to obtain the fault information of the target device; the fault information includes: fault attributes and fault occurrence time; The fault diagnosis classification model includes: a multi-layer convolutional neural network, a multi-level bidirectional long short-term memory network, and a trained classifier; The multi-layer convolutional neural network is used to extract features from multiple target window data to obtain target feature information data corresponding to each target window data. The multi-level bidirectional long short-term memory network is used to perform temporal calculations on each of the target feature information data to obtain multiple target temporal feature data. The trained classifier is used to perform attention mechanism calculations on the temporal feature data of each target to obtain the fused feature data of the target, and outputs the fault information of the target device based on the fused feature data of the target. Each of the target window data is input into the convolution kernel of the multi-layer convolutional neural network to perform convolution operations, resulting in multiple corresponding output convolution results; wherein, the convolution kernel includes multiple channels; each channel corresponds to one of the training window data; The corresponding output convolution results are accumulated and stacked at their respective positions to obtain the target feature information data corresponding to each target window data.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the fault detection method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the fault detection method as described in any one of claims 1 to 5.

Images