Fault re-diagnosis method based on diagnosis confusion and computer readable storage medium

Abstract

The present application relates to a fault re-diagnosis method based on diagnostic confusion and a computer readable storage medium, the fault re-diagnosis method comprising: constructing a training set and a test set; training a fault re-diagnosis model to obtain a trained fault re-diagnosis model; the fault re-diagnosis model comprising a basic fault diagnosis module, a fault re-diagnosis module and a feature fusion module, and each module comprising a feature extraction layer and a feature classification layer connected in sequence, training all samples through the basic fault diagnosis module, transforming high-confusion fault signals into high-confusion fault images through spatial domain transformation, and training the fault re-diagnosis module through the high-confusion fault images; inputting test samples into the trained fault re-diagnosis model to obtain the final fault diagnosis result. The advantage is that the fault re-diagnosis method can accurately diagnose faults, effectively avoids fault confusion, and improves the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a fault re-diagnosis method based on diagnostic confusion and a computer-readable storage medium. Background Art

[0002] The advancement and continuous improvement of industrial technology and equipment have led to increasingly automated and complex industrial processes. Industrial production systems are highly coupled and complex, and even the slightest deviation within them can paralyze the entire system, leading to safety incidents. Fault detection and diagnosis technologies can predict the changing trends of industrial equipment and play a key role in preventing industrial system failures. Therefore, in order to prevent serious industrial accidents caused by faults in industrial production processes, fault diagnosis technology has attracted widespread attention from researchers.

[0003] In actual industrial processes, due to the different impacts of personnel, equipment, materials, and the environment in different processes, the difficulty of diagnosing different fault signals is also different. In the process of industrial fault diagnosis, especially when diagnosing multiple faults, there is often a certain type of fault or several types of faults whose diagnosis rate is significantly lower than that of other faults, and the probability of misdiagnosis is higher than that of other faults. For certain minor and difficult-to-diagnose faults, since their fault manifestations are very similar to those of one or several other fault signals, the diagnosis rate of the fault is very low. When using deep learning intelligent methods to diagnose the fault, it is often easy to misdiagnose it, that is, it is confused with other faults. Therefore, the fault diagnosis accuracy in the existing technology is limited, and therefore there is a need to improve the fault diagnosis method in the existing technology. Summary of the Invention

[0004] The first technical problem to be solved by the present invention is to provide a fault re-diagnosis method based on diagnostic confusion that can improve the accuracy of fault diagnosis in response to the above-mentioned prior art.

[0005] The second technical problem to be solved by the present invention is to provide a computer-readable storage medium capable of executing the above-mentioned fault re-diagnosis method based on diagnostic confusion.

[0006] The technical solution adopted by the present invention to solve the first technical problem is: a fault re-diagnosis method based on diagnostic confusion, characterized by comprising the following steps:

[0007] S1. Acquire multiple fault signals and preprocess all the fault signals to obtain preprocessed fault signals;

[0008] S2. The preprocessed fault signals are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a preprocessed fault signal and a fault label corresponding to the fault signal;

[0009] S3. Construct a fault re-diagnosis model, and train the constructed fault re-diagnosis model in batches using the training set to obtain a trained fault re-diagnosis model;

[0010] The fault re-diagnosis model constructed includes a basic fault diagnosis module, a fault re-diagnosis module and a feature fusion module. The basic fault diagnosis module includes a first feature extraction layer and a first feature classification layer connected in sequence. The fault re-diagnosis module includes a second feature extraction layer and a second feature classification layer connected in sequence. The feature fusion module includes a feature fusion layer and a third feature classification layer connected in sequence.

[0011] The specific process of training the constructed fault re-diagnosis model using a batch of training samples is as follows:

[0012] S3-1. Using all training samples in the current batch to train a basic fault diagnosis module, obtaining a trained basic fault diagnosis module; screening out high-confusion fault signals from all training samples in the current batch, performing spatial domain transformation on each high-confusion fault signal to obtain a high-confusion fault image; using the high-confusion fault image to train a fault re-diagnosis module, obtaining a trained fault re-diagnosis module;

[0013] S3-2. Input the output results of the first feature extraction layer and the output results of the second feature extraction layer into the feature fusion layer to obtain fused fault features. The fused fault features are then input into the third feature classification layer for classification to obtain the final fault diagnosis results. The loss function is calculated using the fault label in each training sample and the corresponding fault diagnosis result to train the feature fusion module, thereby obtaining a trained feature fusion module.

[0014] S4. Select a test sample from the test set. The test sample is a pre-processed fault signal. The test sample is input into the fault re-diagnosis model trained in S3 to obtain the final fault diagnosis result of the test sample.

[0015] Preferably, the method for screening the high-confusion fault signal in S3-1 is:

[0016] Input all training samples of the current batch into the basic fault diagnosis module after training to obtain the fault classification results corresponding to each training sample. The fault classification results correspond to the fault diagnosis rate or false positive rate.

[0017] And judge each training sample in the following way:

[0018] Determine whether the fault diagnosis rate or false positive rate corresponding to the current training sample is within a preset range. If so, the current training sample is a non-high-confusion fault signal; if not, the current training sample is a high-confusion fault signal.

[0019] Preferably, the specific structure of the first feature extraction layer is:

[0020] The first feature extraction layer includes a first feature extraction unit, a receptive field coverage unit, a first convolutional layer and a second convolutional layer connected in sequence, and the output of the second convolutional layer corresponds to the output end of the first feature extraction layer.

[0021] Preferably, the specific structure of the second feature extraction layer is:

[0022] The second feature extraction layer includes a second feature extraction unit, a feature fusion unit, a first 2D convolutional layer and a second 2D convolutional layer connected in sequence. The output of the second 2D convolutional layer is further flattened and then serves as the output end of the second feature extraction layer.

[0023] Preferably, the specific structure of the feature fusion layer is:

[0024] The feature fusion layer includes a bottleneck convolution layer and N network layers connected in sequence; the bottleneck convolution layer is used to reduce the dimension of the merged features; the structure of each network layer is the same, including M convolution layers and 1 maximum pooling layer connected in sequence, and N and M are both positive integers.

[0025] Preferably, the first feature classification layer, the second feature classification layer and the third feature classification layer have the same structure, and all include a global average pooling layer and a fully connected layer connected to the global average pooling layer.

[0026] The technical solution adopted by the present invention to solve the above-mentioned second technical problem is: a computer-readable storage medium, which stores a computer program and can be read and executed by a processor, and is characterized in that: when the computer program is executed by the processor, it implements the above-mentioned fault re-diagnosis method based on diagnostic confusion.

[0027] Compared with existing technologies, the advantages of the present invention are: All samples are trained using a basic fault diagnosis module, thereby achieving a certain level of diagnostic capability for all faults; High-confusion fault signals are transformed into high-confusion fault images, which are then used to train the fault re-diagnosis module, thereby achieving specific diagnostic capability for high-confusion faults; and finally, the output results of the first feature extraction layer and the output results of the second feature extraction layer are fused to improve the overall diagnostic effectiveness of the fault re-diagnosis model. Therefore, this fault re-diagnosis method can accurately diagnose faults, effectively avoid fault confusion, and improve the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 1 is a principle block diagram of a fault re-diagnosis model according to an embodiment of the present invention;

[0029] Figure 2 4 is a network structure diagram of the fault re-diagnosis model in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

[0031] The fault re-diagnosis method based on diagnostic confusion in this embodiment includes the following steps:

[0032] S1. Acquire multiple fault signals and preprocess all the fault signals to obtain preprocessed fault signals;

[0033] In this embodiment, the device to be fault diagnosed is a motor or other device prone to faults. The device's historical operating data (for example, one or more of the following: total current, total voltage, total power, vibration frequency of the device housing, device switch status, current of each device component, voltage of each device component, power of each device component, rotation speed of a rotating device component, and vibration frequency of a vibrating device component) is obtained, and a fault signal is obtained for the device experiencing the fault. The above-mentioned method for obtaining the fault signal is conventional and will not be elaborated on here.

[0034] The above-mentioned preprocessing can be filtering, normalization and other operations;

[0035] S2. The preprocessed fault signals are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a preprocessed fault signal and a fault label corresponding to the fault signal;

[0036] S3. Construct a fault re-diagnosis model, and train the constructed fault re-diagnosis model in batches using the training set to obtain a trained fault re-diagnosis model;

[0037] The fault re-diagnosis model constructed includes a basic fault diagnosis module, a fault re-diagnosis module and a feature fusion module. The basic fault diagnosis module includes a first feature extraction layer and a first feature classification layer connected in sequence. The fault re-diagnosis module includes a second feature extraction layer and a second feature classification layer connected in sequence. The feature fusion module includes a feature fusion layer and a third feature classification layer connected in sequence.

[0038] like Figure 1 As shown in Figure 2, the specific process of training the constructed fault re-diagnosis model using a batch of training samples is as follows:

[0039] S3-1. Using all training samples in the current batch to train a basic fault diagnosis module, obtaining a trained basic fault diagnosis module; screening out high-confusion fault signals from all training samples in the current batch, performing spatial domain transformation on each high-confusion fault signal to obtain a high-confusion fault image; using the high-confusion fault image to train a fault re-diagnosis module, obtaining a trained fault re-diagnosis module;

[0040] In this embodiment, the method for screening the high-confusion fault signal in S3-1 is:

[0041] Input all training samples of the current batch into the basic fault diagnosis module after training to obtain the fault classification results corresponding to each training sample. The fault classification results correspond to the fault diagnosis rate or false positive rate.

[0042] And judge each training sample in the following way:

[0043] Determine whether the fault diagnosis rate or false positive rate corresponding to the current training sample is within a preset range. If so, the current training sample is a non-high-confusion fault signal; if not, the current training sample is a high-confusion fault signal;

[0044] In this embodiment, the preset range is selected as follows:

[0045] A box plot is used to represent all fault diagnosis classification results (i.e., fault diagnosis rate or false positive rate). Q1 is taken as the first quartile (smaller quartile) of the sample data, which is equal to the 25% number after the sample data is sorted from small to large. Q3 is taken as the third quartile (larger quartile) of the sample data, which is equal to the 75% number after the sample data is sorted from small to large. Then, the interquartile range (IQR) is:

[0046] IQR=Q3-Q1

[0047] The data upper limit (maximum) and lower limit (minimum) are expressed as:

[0048] maximum=Q3+1.5IQR

[0049] minimum=Q1-1.5IQR

[0050] If the fault diagnosis rate or false positive rate corresponding to the current training sample is outside [minimum, maximum], the current training sample is a high confusion fault signal;

[0051] S3-2. Input the output results of the first feature extraction layer and the output results of the second feature extraction layer into the feature fusion layer to obtain fused fault features. The fused fault features are then input into the third feature classification layer for classification to obtain the final fault diagnosis results. The loss function is calculated using the fault label in each training sample and the corresponding fault diagnosis result to train the feature fusion module, thereby obtaining a trained feature fusion module.

[0052] S4. Select a test sample from the test set. The test sample is a pre-processed fault signal. The test sample is input into the fault re-diagnosis model trained in S3 to obtain the final fault diagnosis result of the test sample.

[0053] like Figure 2 As shown, the specific structure of the first feature extraction layer in this embodiment is: the first feature extraction layer includes a first feature extraction unit, a receptive field coverage unit, a first convolutional layer and a second convolutional layer connected in sequence, and the output of the second convolutional layer corresponds to the output end of the first feature extraction layer.

[0054] The specific structure of the second feature extraction layer is as follows: the second feature extraction layer includes a second feature extraction unit, a feature fusion unit, a first 2D convolutional layer and a second 2D convolutional layer connected in sequence. The output of the second 2D convolutional layer is also flattened and then serves as the output end of the second feature extraction layer.

[0055] The above-mentioned first feature extraction unit adopts two convolutional layers connected in sequence, the second feature extraction unit adopts two 2D convolutional layers connected in sequence, and the feature fusion unit and the receptive field coverage unit both adopt the maximum pooling layer. The working processes of the above-mentioned units are all existing technologies and will not be elaborated here.

[0056] The specific structure of the feature fusion layer is as follows: the feature fusion layer includes a bottleneck convolution layer and N network layers connected in sequence; the bottleneck convolution layer is used to reduce the dimensionality of the merged features; each network layer has the same structure, including M convolution layers connected in sequence and one maximum pooling layer, where N and M are both positive integers. In this embodiment, N = 2 and M = 2.

[0057] Because the output feature dimensions of the first and second feature extraction layers are different, the output of the second feature extraction layer needs to be flattened. This converts the two-dimensional fault features into one dimension, making the output dimensions of the first and second feature extraction layers consistent. The bottleneck convolution layer reduces the dimensionality of the combined features of the first and second feature extraction layers to optimize the network's computational efficiency. To further deepen feature fusion and extraction, two consecutive operations are performed, each passing the features through two convolutional layers and a maximum pooling layer. This allows the network to capture the fused feature representation and enhance its ability to identify key information. Finally, a global average pooling layer is used to aggregate the features, and a fully connected layer classifies these features for accurate fault diagnosis.

[0058] In addition, the structures of the first feature classification layer, the second feature classification layer and the third feature classification layer are the same, and all include a global average pooling layer and a fully connected layer connected to the global average pooling layer.

[0059] Diagnostic confusion in this embodiment does not refer to a specific fault, but rather to the unevenness of the model's fault diagnosis performance. This can be caused by a few fault features having varying degrees of significance, or by similarities in some fault features. Furthermore, diagnostic confusion is based on the model's diagnostic results and cannot be separated from the specific fault diagnosis model. Ideally, if the model can accurately diagnose any fault, there would be no diagnostic confusion. Therefore, diagnostic confusion is related to both the inherent characteristics of the data and the performance of the model.

[0060] Generally speaking, when a fault with a low diagnostic probability is misdiagnosed, the probability of it being misdiagnosed as another fault is nearly average; the fault is randomly misdiagnosed as any other fault. However, in the case of high-confusion fault misdiagnosis, the majority of samples of a high-confusion fault are non-randomly misdiagnosed as a subset of the faults, resulting in a high false positive rate for these faults. Table 1.1 below shows an example of the misdiagnosis probability for a high-confusion fault. While non-high-confusion fault A has a potential for misdiagnosis, it is randomly misdiagnosed as one of the other four faults. However, the misdiagnosis rate for high-confusion fault B is uneven, with it being more likely to be misdiagnosed as C and D. This misdiagnosis relationship is bidirectional; that is, when other high-confusion faults occur, they also face the risk of being misdiagnosed as the former. This high probability of misdiagnosis reflects, to some extent, the similarities between these fault types and the challenges faced by diagnostic models in diagnosing these high-confusion faults.

[0061] Table 1.1 Examples of misjudgment of high confusion faults

[0062]

[0063] If we group several high-confusion faults together and all remaining non-high-confusion faults together, and then perform binary classification, the classification accuracy should be very high—high-confusion faults are rarely misclassified as non-high-confusion faults, while a small number of non-high-confusion faults may be misclassified as high-confusion faults. This reveals two characteristics of diagnostic confusion: first, some faults (high-confusion faults) are easily misclassified; second, the accuracy of group classification is high, with the high-confusion fault group outperforming the non-high-confusion fault group.

[0064] Therefore, in this embodiment, the basic fault diagnosis module is trained on all samples, thereby achieving a certain level of diagnostic capability for all faults. Furthermore, the high-confusion fault signal is transformed into a high-confusion fault image through spatial domain transformation, and the fault re-diagnosis module is trained on this high-confusion fault image, thereby achieving specific diagnostic capability for high-confusion faults. Finally, the output of the first feature extraction layer and the output of the second feature extraction layer are fused to improve the overall diagnostic effectiveness of the fault re-diagnosis model. Therefore, this fault re-diagnosis method can accurately diagnose faults and improve the accuracy of fault diagnosis.

[0065] This embodiment also relates to a computer-readable storage medium that stores a computer program and can be read and executed by a processor. When the computer program is executed by the processor, the above-mentioned fault re-diagnosis method based on diagnostic confusion is implemented.

[0066] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A fault re-diagnosis method based on diagnostic confusion, characterized in that The steps include: S1. Acquire multiple fault signals and preprocess all the fault signals to obtain preprocessed fault signals; S2. The preprocessed fault signals are combined into a data set, and the data set is constructed into a training set and a test set; wherein each training sample in the training set includes a preprocessed fault signal and a fault label corresponding to the fault signal; S3. Construct a fault re-diagnosis model, and train the constructed fault re-diagnosis model in batches using the training set to obtain a trained fault re-diagnosis model; The fault re-diagnosis model constructed includes a basic fault diagnosis module, a fault re-diagnosis module and a feature fusion module. The basic fault diagnosis module includes a first feature extraction layer and a first feature classification layer connected in sequence. The fault re-diagnosis module includes a second feature extraction layer and a second feature classification layer connected in sequence. The feature fusion module includes a feature fusion layer and a third feature classification layer connected in sequence; The specific process of training the constructed fault re-diagnosis model using a batch of training samples is as follows: S3-1. Using all training samples in the current batch to train a basic fault diagnosis module, obtaining a trained basic fault diagnosis module; screening out high-confusion fault signals from all training samples in the current batch, performing spatial domain transformation on each high-confusion fault signal to obtain a high-confusion fault image; using the high-confusion fault image to train a fault re-diagnosis module, obtaining a trained fault re-diagnosis module; S3-2. Input the output results of the first feature extraction layer and the output results of the second feature extraction layer into the feature fusion layer to obtain fused fault features. The fused fault features are then input into the third feature classification layer for classification to obtain the final fault diagnosis results. The loss function is calculated using the fault label in each training sample and the corresponding fault diagnosis result to train the feature fusion module, thereby obtaining a trained feature fusion module. S4. Select a test sample from the test set. The test sample is a pre-processed fault signal. The test sample is input into the fault re-diagnosis model trained in S3 to obtain the final fault diagnosis result of the test sample.

2. The fault re-diagnosis method according to claim 1, characterized in that: The method for screening the high-confusion fault signal in S3-1 is: Input all training samples of the current batch into the basic fault diagnosis module after training to obtain the fault classification results corresponding to each training sample. The fault classification results correspond to the fault diagnosis rate or false positive rate. And judge each training sample in the following way: Determine whether the fault diagnosis rate or false positive rate corresponding to the current training sample is within a preset range. If so, the current training sample is a non-high-confusion fault signal; if not, the current training sample is a high-confusion fault signal.

3. The fault re-diagnosis method according to claim 1 or 2, characterized in that: The specific structure of the first feature extraction layer is: The first feature extraction layer includes a first feature extraction unit, a receptive field coverage unit, a first convolutional layer and a second convolutional layer connected in sequence, and the output of the second convolutional layer corresponds to the output end of the first feature extraction layer.

4. The fault re-diagnosis method according to claim 3, characterized in that: The specific structure of the second feature extraction layer is: The second feature extraction layer includes a second feature extraction unit, a feature fusion unit, a first 2D convolutional layer and a second 2D convolutional layer connected in sequence. The output of the second 2D convolutional layer is further flattened and then serves as the output end of the second feature extraction layer.

5. The fault re-diagnosis method according to claim 4, characterized in that: The specific structure of the feature fusion layer is: The feature fusion layer includes a bottleneck convolution layer and N network layers connected in sequence; the bottleneck convolution layer is used to reduce the dimension of the merged features; the structure of each network layer is the same, including M convolution layers and 1 maximum pooling layer connected in sequence, and N and M are both positive integers.

6. The fault re-diagnosis method according to claim 5, characterized in that: The first feature classification layer, the second feature classification layer and the third feature classification layer have the same structure, and all include a global average pooling layer and a fully connected layer connected to the global average pooling layer.

7. A computer-readable storage medium storing a computer program and capable of being read and executed by a processor, characterized in that: When the computer program is executed by a processor, the fault re-diagnosis method based on diagnostic confusion according to any one of claims 1 to 6 is implemented.

Images