Fault diagnosis model, method, system, device and readable medium based on small sample learning

Abstract

The present application relates to the technical field of fault diagnosis, and in particular to a fault diagnosis model, method, system, device and readable medium based on small sample learning. The construction method of the fault diagnosis model based on small sample learning comprises the following steps: collecting vibration signal data on a target moving component, converting the vibration signal data into feature images, and constructing a data set storing feature images labeled with fault categories; then making the constructed fault diagnosis model perform machine learning on the data set; the fault diagnosis model comprises a feature extraction module and a classification module, the feature extraction module learns deep local descriptors of the feature images based on a residual element multiplication network to output local features, and the classification module queries the similarity between the feature images and each fault category based on the output of the feature extraction module using a neighbor classification algorithm. The method can significantly improve the fault diagnosis effect and improve the accuracy of the diagnosis result when the training samples are unbalanced under small samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, and in particular to a fault diagnosis model, method, system, device and readable medium based on small sample learning. BACKGROUND

[0002] Traditional bearing fault diagnosis methods usually rely on large-scale historical data and accurate models to predict and identify possible fault patterns. However, in some cases, data may be limited, for example, industrial data often has a high class bias problem, most of the time the equipment is in normal operation state, and the number of fault samples is much smaller than that of healthy samples. Only limited sample data can be used for analysis. In this case, traditional machine learning methods often have difficulty in accurately identifying potential fault patterns and predicting possible problems in bearings.

[0003] Small sample learning-based technology proposes a new solution to this problem. The key advantage of small sample learning is that it can build an efficient model in a data-scarce situation. This method uses a small amount of data samples for model training, which can extract key features from limited data and use special algorithms and techniques to effectively identify potential fault patterns in bearings. Further identify the potential fault signals of the bearing. In addition, this method also has good generalization ability, which can adapt to new data and continuously improve the accuracy of the model to learn and infer more extensive patterns and features from a small number of samples, thereby improving the reliability and accuracy of bearing fault diagnosis. SUMMARY

[0004] In order to overcome the defect that the bearing diagnosis is difficult to use machine learning due to the lack of samples in the prior art, the present application proposes a construction method of a fault diagnosis model based on small sample learning, which can construct a high-precision fault diagnosis model based on a small number of healthy samples. This method is suitable for mechanical moving parts such as bearings.

[0005] The construction method of the fault diagnosis model based on small sample learning proposed by the present application first collects vibration signal data on the target moving part, converts the vibration signal data into feature images, and constructs a data set storing feature images with labeled fault categories. Then let the constructed fault diagnosis model perform machine learning on the data set.

[0006] The fault diagnosis model includes a feature extraction module and a classification module. The feature extraction module learns the deep local descriptors of the feature images based on the residual element multiplication network to output local features. The classification module queries the similarity of the feature images to each fault category based on the output of the feature extraction module using a nearest neighbor classification algorithm.

[0007] Preferably, the feature extraction module comprises a plurality of residual blocks connected in sequence; the residual block comprises: a first activation function, a batch normalization layer, a convolution layer, a first full connection layer, a second full connection layer, a third full connection layer, a dimension stacking unit and a second activation function; the first activation function, the batch normalization layer and the convolution layer are connected in sequence, the input of the first full connection layer and the input of the second full connection layer are both connected to the output of the convolution layer, the output of the first full connection layer and the output of the second full connection layer are both connected to the input of the third full connection layer; the input of the dimension stacking unit is connected to the input of the fault diagnosis model and the output of the third full connection layer respectively, the output of the dimension stacking unit is connected to the input of the second activation function, the dimension stacking unit sends the feature image of the fault diagnosis model input and the output data of the third full connection layer after dimension stacking to the second activation function for activation, and the output of the second activation function is the output of the residual block.

[0008] Preferably, the classification module embeds the local feature of the feature extraction module into a vector space, and calculates the similarity of each feature image in the vector space, the classification module finds k nearest neighbors of the feature image of the to-be-detected sample from each fault category according to the similarity, calculates the probability that the to-be-detected sample belongs to each fault category according to the similarity of the nearest neighbors, and k is a set value.

[0009] Preferably, the calculation formula of the probability p(i) that the to-be-detected sample belongs to the i-th category is as follows:

[0010]

[0011] Wherein, M is the dimension of the embedded vector space, x m represents the m-th array of the local feature ψ(X) of the to-be-detected sample after the feature extraction module, 1≤m≤M; represents the m-th array of the local feature of the j-th sample in the i-th category after the feature extraction module , and the superscript T represents transposition; ||.|| represents the two-norm.

[0012] Preferably, in the iteration process of the fault diagnosis model, the loss function Loss is calculated according to the following formula:

[0013]

[0014] Wherein, y(i) is the parameter after smoothing processing of the real category label probability distribution, ε is a set smoothing parameter, ε is taken between 0 and 1; n is the number of fault categories; p(i) is the probability that the to-be-detected sample belongs to the i-th fault category predicted by the fault category diagnosis model.

[0015] Preferably, let the converted two-dimensional image of the vibration signal data be denoted as an original image, and the feature image in the data set comprises the original image and an image obtained after image enhancement is performed on the original image.

[0016] The method comprises the following steps: firstly, constructing a fault diagnosis model of the object to be diagnosed by using the method for constructing a fault diagnosis model based on small sample learning; then, collecting vibration signal data of the object to be diagnosed, and converting the vibration signal data into a two-dimensional image; inputting the two-dimensional image as a sample to be detected into the fault diagnosis model; and obtaining a fault category corresponding to maximum probability output by the fault diagnosis model as a diagnosis result.

[0017] The device comprises a signal collection module, a signal processing module and a fault diagnosis model.

[0018] The system comprises a memory and a processor, the memory stores a computer program, the processor is connected to the memory, and the processor is used for executing the computer program to realize the fault diagnosis method.

[0019] The readable medium stores a computer program, and the computer program is used for realizing the fault diagnosis method when being executed.

[0020] The method has the following advantages:

[0021] The method can significantly improve the fault diagnosis effect, and improve the accuracy of the diagnosis result when the training sample is unbalanced under a small sample. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a training flowchart of a bearing operation fault diagnosis model based on small sample learning

[0023] Figure 2 is a simulation diagram for comparing the training accuracy obtained by the method and a comparative method in the embodiment

[0024] Figure 3 A comparison simulation diagram of loss values obtained by the method of the present application and the comparative method in the embodiment is shown.

[0025] Figure 4 A comparison simulation diagram of test accuracy obtained by the method of the present application and the comparative method in the embodiment is shown.

[0026] Figure 5 A comparison simulation diagram of validation accuracy obtained by the method of the present application and the comparative method in the embodiment is shown.

[0027] ResEMNet-LabelSmooth, ResEMNet--CrossEntropy, Conv64F-CrossEntropy, Conv64F-LabelSmooth. DETAILED DESCRIPTION

[0028] 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. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0029] The data sample in the embodiment is denoted as {feature image, fault category}. The feature image includes two categories: one is a two-dimensional image obtained by converting the sampled vibration data signal through signal processing, and the other is an image obtained by image enhancement on the two-dimensional image of the first category.

[0030] The vibration data signal can be converted into a two-dimensional image after being processed by the matplotlib library.

[0031] The image enhancement includes random photometric transformation, random cropping, etc.

[0032] In the embodiment, the fault diagnosis model performs fault diagnosis on the feature image obtained based on the vibration data signal processing, and outputs the fault category.

[0033] The fault diagnosis model includes a feature extraction module and a classification module. The feature extraction module includes sequentially connected first, second, third and fourth residual blocks.

[0034] The first, second, third and fourth residual blocks have the same structure and are collectively referred to as residual blocks. The input of the residual block is used as the input of the fault diagnosis model.

[0035] The residual block comprises a first activation function, a batch normalization layer, a convolution layer, a first full connection layer, a second full connection layer, a third full connection layer, a dimension stacking unit and a second activation function. The first activation function, the batch normalization layer and the convolution layer are sequentially connected, and the feature image is activated by the first activation function, and then output after batch normalization and convolution processing.

[0036] The input of the first full connection layer and the input of the second full connection layer are connected to the output of the convolution layer, and the output of the first full connection layer and the output of the second full connection layer are connected to the input of the third full connection layer. In this way, the output of the convolution layer is processed by the first full connection layer and the second full connection layer respectively, and then the output of the second full connection layer and the output of the second full connection layer are point multiplied and input to the third full connection layer for processing. The input of the dimension stacking unit is connected to the input of the fault diagnosis model and the output of the third full connection layer respectively, and the output of the dimension stacking unit is connected to the input of the second activation function. The dimension stacking unit performs dimension stacking on the feature image input by the fault diagnosis model and the output data of the third full connection layer, and sends the result to the second activation function for activation. The output of the second activation function is the output of the residual block.

[0037] In this way, let the input of the residual block be input and the output be output, and the processing process of the residual block is represented by the following formula:

[0038] DW = Conv (BN (LR1 (input)))

[0039] output = LR2 (input + FC3 (FC1 (DW) * FC2 (DW)))

[0040] Where DW represents the output of the convolution layer, Conv represents the convolution layer, BN represents the batch normalization layer, LR1 represents the first activation function, LR2 represents the second activation function, FC1, FC2 and FC3 represent the first full connection layer, the second full connection layer and the third full connection layer respectively.

[0041] The first residual block, the second residual block, the third residual block and the fourth residual block are sequentially connected, and the output of the fourth residual module is the output of the feature extraction module. Let the first residual block, the second residual block, the third residual block and the fourth residual block be B1, B2, B3 and B4 respectively, and the feature image input by the fault diagnosis model be X', then the local feature ψ(X') output by the feature extraction module is represented by the following formula:

[0042] ψ(X') = B4(B3(B2(B1(X')))

[0043] Classification module

[0044] The classification module first embeds the local feature output by the feature extraction module into a vector space Rd×M , denoted as:

[0045] ψ(X') = [x'1, x'2, …, x' m ,…, x' M ]

[0046] x' m denotes the mth array of the local feature ψ(X'), 1≤m≤M; each number contains d numerical values; that is, x'1, x'2, x' M denote the 1st array, the 2nd array and the Mth array of the local feature ψ(X') respectively.

[0047] The classification module finds the k nearest neighbors of the to-be-detected sample from each fault category in combination with the similarity, and calculates the probability of the to-be-detected sample belonging to each fault category according to the similarity of the nearest neighbors.

[0048] Taking the i-th fault category as an example, the classification module first calculates the similarity of the to-be-detected sample and each sample in the i-th fault category in the data set, and then selects the k samples with the largest similarity as the nearest neighbors of the to-be-detected sample, k being a set value, which can be any integer greater than 1 in specific implementation, such as 3, 5, 10, etc.

[0049] Let the to-be-detected sample be denoted as X, and after processing by the feature extraction module, we have:

[0050] ψ(X) = [x1, x2, …, x m ,…, x M ]

[0051] x m denotes the mth array of the local feature ψ(X), 1≤m≤M; each number contains d numerical values; that is, x1, x2, x M denote the 1st array, the 2nd array and the Mth array of the local feature ψ(X) respectively.

[0052] Let the j-th sample of the i-th fault category in the data set be denoted as After processing by the feature extraction module, we have:

[0053]

[0054] denotes the mth array of the local feature , 1≤m≤M; each number contains d numerical values; that is, denote the 1st array, the 2nd array and the Mth array of the local feature respectively.

[0055] Let the similarity between the to-be-detected sample X and the sample be represented by the cosine similarity, then we have:

[0056]

[0057] According to the order from large to small of the cosine similarity, the samples in the i-th fault category are sorted, and the probability p(i) that the to-be-detected sample belongs to the i-th fault category is:

[0058]

[0059] Wherein, Φ represents the classification module, c i represents the i-th fault category;

[0060] The output of the classification module is the probability {p(i)} 1≤i≤n that the to-be-detected sample belongs to each fault category.

[0061] In this embodiment, the training of the fault diagnosis model includes the following steps:

[0062] S1, extracting data samples {feature images, fault categories} from the data set to construct a training set and a validation set respectively; and constructing a fault diagnosis model;

[0063] S2, extracting training samples from the training set, and allowing the fault diagnosis model to learn the training samples;

[0064] S3, extracting validation samples from the validation set, allowing the fault diagnosis model to predict the fault category probability of the validation samples, and calculating the loss function Loss;

[0065]

[0066] Wherein, y(i) is the parameter after smoothing processing of the real category label probability distribution, ε is a set smoothing parameter, ε is taken between 0 and 1; n is the number of fault categories; p(i) is the probability that the fault category diagnosis model predicts that the validation sample belongs to the i-th fault category;

[0067] S4, judging whether the fault diagnosis model converges; if not, returning to step S2; if yes, fixing the fault diagnosis model and outputting.

[0068] The bearing fault diagnosis method proposed in this embodiment includes the following steps:

[0069] St1, collecting vibration signal data of the bearing in continuous time, performing two-dimensional transformation on the vibration signal data, and obtaining a two-dimensional image as a to-be-detected sample.

[0070] It is worth noting that the length of the vibration signal data is consistent in the model training process and the application process.

[0071] St2, input the sample to be detected into the trained fault diagnosis model, and obtain the fault category corresponding to the maximum probability output by the fault diagnosis model as the diagnosis result.

[0072] The above fault prediction model is verified on bearing diagnosis in combination with specific embodiments. In this embodiment, vibration signal sampling is performed under the working state of the bearing, so that one vibration signal data contains 6003 columns, wherein the first column is an id field representing a sample number, the 2nd to 6001st columns are vibration signal values sampled in time sequence, and the last column label field is label data, i.e., the fault category of the bearing. In this embodiment, the fault category is represented by numbers 0 to 9, wherein "0" represents normal, and other labels are shown in Table 1.

[0073] Table 1: Bearing fault state label

[0074]

[0075] In Table 1, the number "1" represents that the bearing adopts a diameter A1 and is in an outer ring fault state when sampling; and the like.

[0076] In this embodiment, a total of 792 vibration signal data are collected when the bearing is working, containing all labels 0-9. The vibration signal data is converted into a two-dimensional image after processing by the matplotlib library, and then a data set is constructed, the data sample in the data set is recorded as {feature image, fault category}, the feature image is the two-dimensional image converted from the vibration signal data, or the image after image enhancement of the two-dimensional image; the fault category in the data sample is the working state when the vibration signal data corresponding to the feature image is sampled, which is represented by numbers 0-9 in this embodiment. In this embodiment, the data set is divided into a training set, a validation set and a test set in proportions of 60%, 20% and 20% respectively.

[0077] In this embodiment, the fault diagnosis model proposed by the application is recorded as ResEMNet-LabelSmooth; in addition, three comparative models are selected, which are:

[0078] ResEMNet--CrossEntropy: using the fault diagnosis model architecture proposed by the application, in the model training process, the loss function uses a cross-entropy loss function;

[0079] Conv64F-CrossEntropy: compared with the fault diagnosis model architecture proposed by the application, the feature extraction module is replaced by a standard convolutional layer architecture, and in the model training process, the loss function uses a cross-entropy loss function;

[0080] Conv64F-LabelSmooth: compared with the fault diagnosis model architecture proposed in the application, the feature extraction module is replaced by a standard convolutional layer architecture, and in the model training process, the loss function adopts a traditional label smoothing loss.

[0081] The four models are trained on the training set and the validation set, and the model performance is tested on the test set after the model converges.

[0082] In this embodiment, when the four models learn on the training set, the training parameters are set to 5-way one-shot, that is, there are 5 categories, each category has only one sample, the batchsize (each training batch) size is set to 16, the total epoch (total training rounds) is 50, and the episode (model iteration number, that is, the loss transmission number) is 100; that is, in each iteration, first train 16 batches of 80 samples, each batch uses 5 samples, and they come from 5 different categories; then extract the validation set samples to calculate the loss, and update the model parameters according to the loss; in this embodiment, the model is calculated as a round of training every 100 iterations, and the model convergence condition is: complete 50 rounds of training.

[0083] In this embodiment, the prediction accuracy of the four models on the training set is as shown in Figure 2 It can be seen that the feature extraction module used in the model ResEMNet-LabelSmooth proposed in the application combined with the label smoothing regularization method has a much higher prediction accuracy on the training set than the traditional convolutional network architecture of the model Conv64F-CrossEntropy and the original cross-entropy loss, and the model ResEMNet-LabelSmooth proposed in the application is compared with the model Conv64F-LabelSmooth, which proves that the feature extraction module can reach convergence faster than the traditional convolutional network architecture in the training process.

[0084] Figure 3 is the comparison of the loss values of the four models in the training process, it can be seen that even after using the label smoothing regularization method, the loss in the application can still quickly converge.

[0085] Figure 4is the comparison of the prediction accuracy of the four models on the test set. It can be seen that after the convergence of the four models, the prediction accuracy can reach more than 90%, but the ResEMNet-LabelSmooth and the comparative model ResEMNet--CrossEntropy of the model of the application using the feature extraction module can converge faster, and the prediction accuracy of the ResEMNet-LabelSmooth of the application is more smooth and stable than that of the ResEMNet--CrossEntropy. It can be seen that the model constructed by the application is more stable, and the test accuracy will not decrease with the increase of the iteration round, avoiding the overfitting phenomenon. It can be seen that the application has better convergence stability and good robustness.

[0086] Figure 5 is the comparison of the prediction accuracy of the four models on the test set. It can be seen that after the convergence of the four models, the prediction accuracy can reach more than 90%, but the ResEMNet-LabelSmooth and the comparative model ResEMNet--CrossEntropy of the model of the application using the feature extraction module can converge faster, and the prediction accuracy of the ResEMNet-LabelSmooth of the application is more smooth and stable than that of the ResEMNet--CrossEntropy. It can be seen that the model constructed by the application is more stable, and the test accuracy will not decrease with the increase of the iteration round, avoiding the overfitting phenomenon. It can be seen that the application has better convergence stability and good robustness. Figure 2 Figure 3 Figure 4 Figure 5 It can be seen that the model of the application has the highest accuracy on the training set, the validation set and the test set, and the model converges faster and has more stable performance than the other comparative models.

[0087] Of course, for those skilled in the art, the application is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures that can be realized in other specific forms without departing from the spirit or essential characteristics of the application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0088] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be properly combined to form other embodiments that those skilled in the art can understand.

[0089] The technologies, shapes and configurations not described in detail in the application are well-known technologies.​​​

Claims

1. A method for constructing a fault diagnosis model based on small sample learning, characterized in that, First, vibration signal data is collected on a target moving component, and the vibration signal data is converted into feature images to construct a data set storing feature images labeled with fault categories; then the constructed fault diagnosis model is allowed to perform machine learning on the data set; The fault diagnosis model comprises a feature extraction module and a classification module, the feature extraction module learns a deep local descriptor of the feature images based on a residual element multiplication network to output local features, and the classification module queries the similarity of the feature images to each fault category based on the output of the feature extraction module using a nearest neighbor classification algorithm; The feature extraction module comprises a plurality of residual blocks connected in sequence; each residual block comprises a first activation function, a batch normalization layer, a convolution layer, a first full connection layer, a second full connection layer, a third full connection layer, a dimension stacking unit, and a second activation function; the first activation function, the batch normalization layer, and the convolution layer are connected in sequence, the input of the first full connection layer and the input of the second full connection layer are both connected to the output of the convolution layer, and the output of the first full connection layer and the output of the second full connection layer are both connected to the input of the third full connection layer; the input of the dimension stacking unit is connected to the input of the fault diagnosis model and the output of the third full connection layer respectively, the output of the dimension stacking unit is connected to the input of the second activation function, the dimension stacking unit is used to stack the feature images input by the fault diagnosis model and the output data of the third full connection layer in dimension and then send them to the second activation function for activation, and the output of the second activation function is the output of the residual block; The classification module embeds the local features output by the feature extraction module into a vector space and calculates the similarity of each feature image in the vector space, the classification module finds k nearest neighbors of the feature image of the to-be-detected sample from each fault category according to the similarity, calculates the probability that the to-be-detected sample belongs to each fault category according to the similarity of the nearest neighbors, and k is a set value; The calculation formula of the probability p(i) that the to-be-detected sample belongs to the i-th category is as follows: where M is the dimension of the embedding vector space, x m represents the mth array of the local feature ψ(X) of the sample to be detected after being processed by the feature extraction module, 1≤m≤M; represents the mth array of the local feature ψ(X) of the sample to be detected after being processed by the feature extraction module, 1≤m≤M; represents the mth array of the local feature ψ(X) of the sample to be detected after being processed by the feature extraction module, 1≤m≤M; During the iteration process of the fault diagnosis model, the loss function Loss is calculated according to the following formula: where y(i) is a parameter obtained by smoothing the real category label probability distribution, ε is a set smoothing parameter, ε is taken as a value between 0 and 1; n is the number of fault categories; and p(i) is the probability that the fault diagnosis model predicts that the verification sample belongs to the i-th fault category. 2.The method of claim 1, wherein, Let the two-dimensional image converted from the vibration signal data be an original image, and the feature images in the data set include the original image and an image obtained by performing image enhancement on the original image.

3. A failure diagnosis method employing a construction method of a failure diagnosis model based on small sample learning as claimed in claim 1 or 2, characterized by, First, the fault diagnosis model of the to-be-diagnosed object is constructed by using the construction method of the fault diagnosis model based on small sample learning according to claim 1 or 2; then the vibration signal data of the to-be-diagnosed object is collected and converted into a two-dimensional image, the two-dimensional image is input into the fault diagnosis model as a to-be-detected sample, and the fault category corresponding to the maximum probability output by the fault diagnosis model is obtained as a diagnosis result.

4. A failure diagnosing apparatus employing the failure diagnosing method according to claim 3, characterized by It comprises: The signal acquisition module is used for collecting vibration signal data of an object to be diagnosed; the signal processing module is connected with the signal acquisition module and the fault diagnosis model respectively; the signal acquisition module converts the vibration signal data into a two-dimensional image and inputs the two-dimensional image into the fault diagnosis model; and the fault diagnosis model outputs a diagnosis result.

5. A failure diagnosis system characterized by comprising: The computer program is stored in the memory and is executed by the processor to implement the fault diagnosis method of claim 3.

6. A readable medium characterized by The computer program is stored in the memory and is executed by the processor to implement the fault diagnosis method of claim 3.

Images