Intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data

Abstract

The present invention discloses an intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data. The first step is to prepare a large amount of unlabeled data as self-supervised training data, prepare labeled data, and divide it into supervised training data and test data; the second step is to enhance the unlabeled data based on three enhancement methods; the third step is to build a self-supervised network including feature extraction and feature mapping; the fourth step is to select the corresponding features of sparse category data through inverse attention; the fifth step is to design the noise contrast estimation loss function based on the selected features, and optimize the self-supervised network; the sixth step is to adjust the diagnostic network based on the feature extraction network after self-supervised pre-training through a small amount of labeled fault data; the seventh step is to input the monitoring signal into the above-mentioned trained network. The present invention is more in line with the problem of no health status label and type imbalance of actual monitoring data, and improves the diagnostic ability of the existing fault diagnosis network.

Description

Technical Field

[0001] The present invention belongs to the technical field of processing and analyzing vibration signals of rotating machinery, and in particular relates to an intelligent fault diagnosis method based on self-supervision of unbalanced label-free monitoring data. Background Art

[0002] With the increasing requirements for machine operation, the monitoring of mechanical working conditions is becoming more and more important. Among them, intelligent and rapid diagnosis of health status is one of the important contents. As one of the signal sources containing rich information about mechanical faults, vibration signals are widely used in fault diagnosis of machinery, especially rotating machinery. Mechanical fault diagnosis relies on the extraction of fault signal features. Traditional fault diagnosis methods require a lot of manual intervention feature extraction methods, which are no longer sufficient to process industrial big data. With the development of artificial intelligence, artificial neural networks have achieved remarkable results in many fields, especially in the development of pattern recognition fields such as image recognition and speech recognition.

[0003] Due to the powerful nonlinear feature extraction capability of artificial neural networks, researchers have introduced them into the field of mechanical fault diagnosis and conducted extensive research on them. Artificial neural networks have a high degree of automation, which greatly reduces the need for human participation in the feature extraction process. Among various neural network models, unsupervised learning methods have attracted particular attention because they do not require label participation. These include autoencoders, restricted Boltzmann machines, sparse filtering, etc. Although many powerful unsupervised networks have emerged, there are still some common problems among them.

[0004] Existing unsupervised pre-training methods are difficult to effectively extract real features, and the analysis of features is too simple. Self-supervised network pre-training methods based on contrastive learning used in other fields can form positive and negative samples through data enhancement, encourage the learning of essential patterns in the data, and further strengthen the mining of essential patterns rather than background noise patterns during pre-training. It can recognize and analyze the deep structural commonality of samples through complex surface data information. This also inspires the further effective use of massive unlabeled monitoring data in the field of fault diagnosis.

[0005] Self-supervised methods can learn feature representations under signal contrast, which is very important for further mining pattern features in samples and data sets. However, existing diagnostic methods have not done enough research on this, and effective data enhancement forms and suitable network structures that match vibration signals also need to be further explored. At the same time, in the unlabeled data sets obtained through monitoring, the quantitative balance of different types of samples is difficult to guarantee and unknown. The imbalance of the anonymity of such sample types also greatly affects the effectiveness of existing self-supervised pre-training methods. Summary of the invention

[0006] Purpose of the invention: To address the problem that massive unlabeled unbalanced monitoring data cannot be effectively used for auxiliary training of fault diagnosis models, the present invention proposes an intelligent fault diagnosis method based on self-supervision of unbalanced unbalanced monitoring data, which can make full use of unlabeled unbalanced monitoring data in real scenarios for self-supervised pre-training, and can obtain high diagnostic accuracy only based on a very small amount of labeled fault data, greatly expanding the application scope of the intelligent diagnosis method.

[0007] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0008] The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data includes the following steps:

[0009] The first step is to obtain the time domain vibration signal, use the unlabeled and data type unbalanced monitoring data as the data set as the self-supervised pre-training data, and divide the labeled time domain vibration signal into a training set and a test set;

[0010] In the second step, the unlabeled data are enhanced and expanded using multiplicative noise enhancement, reverse shift enhancement, and additive noise enhancement.

[0011] The third step is to build a self-supervised fault diagnosis network, which includes feature extraction and mapping. After feature extraction, inverse attention selection of features is performed to optimize features to achieve balanced learning and global maturity of features.

[0012] The fifth step is to design the noise contrast estimation loss function based on the selected features and use the enhanced vibration signal unlabeled dataset to pre-train the self-supervised network;

[0013] Step 6: Global adjustment training of the fault diagnosis network: Use the labeled training data set to adjust the parameters of the self-supervised network trained in step 5;

[0014] In the seventh step, the vibration signal test set is input into the trained network and its health status type is output.

[0015] Furthermore, the multiplicative noise enhancement formula in the second step is as follows:

[0016] ,

[0017] in represents the first samples, The vibration signal that needs to be enhanced samples, Indicates signal The average amplitude, enhancement amplitude and The amplitude is proportional to Indicates signal The standard deviation of Indicates that from is the mean, Get the same as the Gaussian distribution constructed for the variance Random number sequences of the same length.

[0018] Furthermore, the reverse offset enhancement process in the second step is as follows:

[0019] Set sample The length is , flip the vibration signal in the time dimension, so that the first The data points in the new sample are located at , then connect the signal end to end, so that all data points move backward in a cycle Positions, get the reverse offset enhanced samples .

[0020] Furthermore, the additive noise enhancement process in the second step is as follows:

[0021] According to the vibration signal Samples The average amplitude , standard deviation , enhanced by formula (2), where Indicates that from is the mean, The Gaussian distribution constructed for the standard deviation is obtained with Random number sequences of the same length; is a further mixed salt and pepper noise addition method, which indicates middle Data points are randomly selected using the rand function And set the selected data points to 0 through mask operation, while the values ​​of other data points remain unchanged. represents the first The samples are expressed as follows:

[0022] .

[0023] Furthermore, the self-supervised network constructed in the third step includes feature extraction and mapping, specifically including:

[0024] The first Multiplicative noise enhanced samples , reverse order offset enhanced samples and additive noise enhanced samples Splicing as Passed to the feature encoder network in formula (3) Perform feature extraction, Adopt 1DCNN network structure;

[0025] ,

[0026] in It is the output representation of the last average pooling layer of the 1DCNN network. Each enhanced sample obtains its corresponding output through 1DCNN.

[0027] Then the feature nonlinear mapping is performed using the projection head The representation obtained by encoding Mapped to the latent space of the corresponding contrast loss, the mapping relationship is shown in formula (4);

[0028] ,

[0029] in It is a representation Mapping output, is the ReLu activation function, and is the parameter matrix, It is used to linearly transform the features after nonlinear activation to obtain the final output. Used to perform linear transformation and dimensionality reduction on input features.

[0030] Furthermore, the balanced learning and global maturation process of features is as follows:

[0031] The mapping output obtained in the third step is recorded as the feature , first calculate the feature Maturity ; The kurtosis shown in formula (5) Calculate the activation value of the same feature in all samples and obtain its distinguishability to evaluate its maturity;

[0032] ,

[0033] In terms of maturity After sorting, the proportion of feature dimensions that need to be suppressed is , for maturity Sort from large to small and get the characteristics The corresponding sorting of each dimension feature , according to formula (6), lock the position mask of the neuron to be inhibited;

[0034] ,

[0035] Finally, the position With features Multiply to get new features.

[0036] Furthermore, the fifth step is to design the loss function and perform pre-training, which includes:

[0037] No. Multiplicative noise enhanced samples and reverse offset enhanced samples The positive sample pairs are transformed by the feature extraction network to obtain feature representation and ; The cosine similarity in formula (7) is used to represent the distance between features. The matching method of other enhanced samples is the same;

[0038] ,

[0039] Positive sample pair The contrast loss function The definition of is shown in formula (8), , Indicates The original samples are reversed and shifted to enhance the samples;

[0040] ,

[0041] in is the temperature parameter, is the total number of original samples, and self-supervised network training based on unlabeled data is achieved by minimizing the contrast loss function.

[0042] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0043] The present invention designs a specific effective data enhancement mode for vibration signals, which forms a driving force for the effective learning of the self-supervised network. It can fully explore the fault mode, assist in removing the working conditions and background noise modes therein, provide an excellent optimization starting point for subsequent supervised training, and effectively reduce the demand for labeled monitoring data in the supervised training process.

[0044] The three enhancement methods for vibration signals in the present invention greatly enrich the sample forms and can effectively assist self-supervised learning; the subsequent feature mapping is followed by inverse attention selection and focused optimization of features corresponding to sparse category data, which achieves balanced learning and global maturity of features and handles anonymous category imbalance in unlabeled data; the noise contrast estimation loss function designed based on the selected features can fully stimulate the self-supervised learning process and achieve full utilization of unlabeled imbalanced monitoring data.

[0045] The feature extraction network based on self-supervised pre-training of the present invention only needs a very small amount of labeled fault data to perform global adjustment of the fault diagnosis network, so as to realize direct and accurate fault diagnosis, which greatly reduces the training difficulty of the fault diagnosis network in real scenarios. The present invention is used in the fault diagnosis of rotating machinery to realize effective intelligent fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of an intelligent fault diagnosis network based on self-supervised pre-training of unbalanced unlabeled monitoring data according to the present invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0048] The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data described in the present invention first designs three data enhancement methods for vibration signals, so as to facilitate the subsequent self-supervised pre-training process to perform positive and negative example comparison training through enhanced samples, and then constructs a feature extraction and feature mapping module, performs inverse attention selection after feature extraction and mapping, and focuses on optimizing the features of rare category data with weaker optimization effects in the network, thereby achieving balanced learning of global features and effectively dealing with the anonymous imbalance problem of unlabeled data categories; then, the noise contrast estimation loss function is designed based on the selected features, and the self-supervised network is optimized through enhanced data to fully utilize the unbalanced unlabeled monitoring data; finally, based on the feature extraction network after self-supervised pre-training, a small amount of labeled fault data is used to perform global fine-tuning of the fault diagnosis network; in the use stage, the monitoring signal is input into the above-mentioned trained network to achieve direct diagnosis of the fault, which will greatly improve the utilization effect of massive unlabeled monitoring data. The process is as follows Figure 1 As shown, the specific implementation process is as follows:

[0049] The first step is to obtain training and test vibration signals. A large amount of unlabeled and unbalanced monitoring data constitutes the dataset. , as self-supervised pre-training data, is the original vibration signal, express Data points. A small number of labeled time-domain vibration signals As input, It is a type of fault signal, containing data points; is the total number of categories of different health status; is the fault type label for each category of fault. After segmenting the signal, 10% of the data is taken as the labeled training sample set, and the remaining 90% of the data is taken as the test sample set.

[0050] The second step is unlabeled data enhancement. Three data enhancement methods designed for vibration signals are used to expand the unlabeled data: multiplicative noise enhancement, reverse shift enhancement, and additive noise enhancement.

[0051] Multiplicative noise enhancement: Enhancement is performed using formula (1), where is the vibration signal sample that needs to be enhanced, and its enhancement amplitude is The amplitude of is proportional to and can be adjusted adaptively. Indicates signal The average amplitude of Indicates signal The standard deviation of . Indicates that from is the mean, Get the same as the Gaussian distribution constructed for the variance Random number sequences of the same length. Represents samples after multiplicative noise enhancement.

[0052] .

[0053] Reverse frequency shift enhancement: The reverse frequency shift mainly flips the vibration signal in the time dimension, making the original The data points in the new sample are located at ,in For sample The length of the signal is then connected end to end so that all data points move backwards in a cycle Positions, get reverse offset enhanced data .

[0054] Additive noise enhancement: According to The average amplitude , standard deviation , enhanced by formula (2), where Likewise, from is the mean, The Gaussian distribution constructed for the standard deviation is obtained with Random number sequences of the same length. is a further mixed salt and pepper noise addition method, which indicates middle Data points are randomly selected using the rand function And set the selected data points to 0 through mask operation, while the values ​​of other data points remain unchanged. represents the first samples.

[0055] .

[0056] Original sample Three new enhanced data samples are available , and , so that after data augmentation the dataset will be tripled, , and These three samples are a group of positive pairs with each other, and the new samples obtained by random data augmentation with any other samples are negative pairs.

[0057] The third step is to build a self-supervised network, which includes two modules: feature extraction and mapping. , and Splice to Passed to the feature encoder network in formula (3) Perform feature extraction, A 1DCNN network structure can be used, which can be adjusted according to different input data formats.

[0058] ,

[0059] in It is the output representation of the last average pooling layer of the 1DCNN network. Each enhanced sample obtains its corresponding output through 1DCNN.

[0060] This is followed by a feature nonlinear mapping step, in which the projection head The representation obtained by encoding Mapped to the latent space of the corresponding contrast loss, the mapping relationship is shown in formula (4).

[0061] ,

[0062] in It is a representation Mapping output, is the ReLu activation function, and is the parameter matrix, It is used to linearly transform the features after nonlinear activation to obtain the final output. Used to perform linear transformation and dimensionality reduction on input features. Compare More conducive to constructing contrast loss.

[0063] In the fourth step, after the feature extraction part, the inverse attention selection of features corresponding to the rare category data is performed. The features with weaker optimization effects in the network are optimized, so as to achieve balanced learning and global maturity of the features. Rare category data refers to the types of faults that appear less frequently in the training data set. It is usually difficult to obtain enough samples for these categories, resulting in insufficient recognition ability of the model. Weak features refer to those that contribute less to the loss function during model training or have limited improvement in overall prediction performance.

[0064] The mapping output obtained in the third step is recorded as the feature , first calculate the feature Maturity Because the learned features will be used for classification tasks, the sparse activation characteristics of the features are very important. Therefore, it is necessary to optimize the distinguishability of all features equally, that is, as shown in formula (5): (Kurtosis) Calculate the activation value of the same feature in all samples and obtain its distinguishability to evaluate its maturity.

[0065] ,

[0066] In terms of maturity After sorting, the proportion of feature dimensions that need to be suppressed is , for maturity Sort from large to small and get the characteristics The corresponding sorting of each dimension feature , according to formula (6), the position mask of the neurons to be inhibited is locked.

[0067] ,

[0068] Finally, the position With features Multiply to get new features.

[0069] Step 5: Loss function design and self-supervised network training. The noise contrast estimation loss function is designed based on the selected features, and the self-supervised network is optimized based on the enhanced data. , The positive sample pairs are transformed by the feature extraction network to obtain feature representation and , the cosine similarity in formula (7) is used to represent the distance between features. The matching method of other enhanced samples is the same.

[0070] ,

[0071] Therefore, the positive sample pair The contrast loss function The definition of is shown in formula (8), , Indicates The original samples are reversed and shifted to the enhanced samples.

[0072] ,

[0073] in is the temperature parameter, is the total number of original samples. Self-supervised network training based on unlabeled data is achieved by minimizing the contrast loss function.

[0074] Step 6: Global adjustment training of the fault diagnosis network. The trained self-supervised network enters the classification task adjustment process, at which time the projection head is discarded. , the feature encoder network And the representation obtained by training the network Put it into the classification task, and use the split 10% labeled training data set to adjust the parameters of the model. The previous stage of the model has learned effective features from a large amount of unlabeled data, so continue to train the model with a small amount of labeled training data set.

[0075] In the seventh step, 90% of the monitoring signals used for testing are input into the trained network model and their health status types are output. Compared with the existing technology, the above network is more suitable for the problem of unbalanced categories of monitoring data without health status labels and anonymity that must be faced in the actual fault diagnosis network training. It realizes the use of unbalanced monitoring data to assist the diagnosis network training and effectively improves the generalization ability of the existing fault diagnosis network.

[0076] The intelligent fault diagnosis model can make full use of unlabeled unbalanced monitoring data in real scenarios for effective self-supervised pre-training, and can obtain high diagnostic accuracy by adjusting only a small amount of labeled fault data, which greatly expands the application scope of intelligent diagnosis methods. An effective hybrid enhancement method for vibration signals, and the hybrid enhancement method does not require data labels and does not require parameter adjustment. The enhancement process is adaptively adjusted according to the characteristics of the vibration signal itself, and the operation is simple and easy to implement. The original unlabeled monitoring data collected by the sensor can be continuously and directly used for training, achieving the role of effective auxiliary diagnosis network training. The feature selection method directly focuses on the poorly optimized features in network optimization, especially the features corresponding to the rare sample category, which is better suitable for unlabeled monitoring data with unbalanced anonymous data types. When constructing a self-supervised network, it includes two modules: feature extraction and feature mapping. The newly added feature mapping part is used in the self-supervision process and discarded during the global adjustment, which realizes the partial isolation of downstream tasks from the self-supervised pre-training process, making the self-supervised training network more suitable for downstream fault diagnosis tasks.

Claims

1. An intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data, comprising the following steps: The first step is to obtain the time domain vibration signal, use the unlabeled and data type unbalanced monitoring data as the data set as the self-supervised pre-training data, and divide the labeled time domain vibration signal into a training set and a test set; In the second step, the unlabeled data are enhanced and expanded using multiplicative noise enhancement, reverse shift enhancement, and additive noise enhancement. The third step is to build a self-supervised fault diagnosis network, which includes feature extraction and mapping, including: The i-th multiplicative noise enhanced sample Reverse offset enhanced sample and additive noise enhanced samples Splice as X i Pass it to the feature encoder network f(·) in formula (3) for feature extraction. f(·) adopts 1DCNN network structure; h i =f(X i )=1DCNN(X i ) (3), where h i It is the output representation of the last average pooling layer of the 1DCNN network. Each enhanced sample obtains its corresponding output through 1DCNN. Then, feature nonlinear mapping is performed, and the representation h obtained by encoding is transformed into i Mapped to the latent space of the corresponding contrast loss, the mapping relationship is shown in formula (4); q i =g(h i )=W 1 σ(W 2 h i ) (4), where q i It is the representation of h i Mapping output, σ(·) is the ReLu activation function, W 1 and W 2 is the parameter matrix, W 1 It is used to linearly transform the features after nonlinear activation to obtain the final output, W 2 Used to perform linear transformation and dimensionality reduction on input features; The fourth step is to perform inverse attention selection of features after feature extraction to optimize features to achieve balanced learning and global maturity of features. The process of balanced learning and global maturity of features is as follows: The mapping output obtained in the third step is recorded as feature z. First, the maturity g of feature z is calculated. z ; The activation value of the same feature in all samples is calculated by the kurtosis shown in formula (5), and its distinguishability is obtained to evaluate its maturity; g z =jurtosis(z) (5), In terms of maturity z After sorting, the maturity g is z Sort from large to small to get the ranking corresponding to the features of each dimension in feature z g , according to formula (6), lock the position mask of the neuron to be inhibited; Finally, multiply the location (i) by the feature Z to obtain a new feature; The fifth step is to design the noise contrast estimation loss function based on the selected features and use the enhanced vibration signal unlabeled dataset to pre-train the self-supervised network; Step 6: Global adjustment training of the fault diagnosis network: Use the labeled training data set to adjust the parameters of the self-supervised network trained in step 5; In the seventh step, the vibration signal test set is input into the trained network and its health status type is output.

2. The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data according to claim 1 is characterized in that: The multiplicative noise enhancement formula in the second step is as follows: in represents the i-th sample after multiplicative noise enhancement, x i is the i-th sample of the vibration signal to be enhanced, z i Represents signal x i The average amplitude, enhancement amplitude and x i The amplitude is proportional to s i Represents signal x i The standard deviation of random(z i ,s i ) indicates that from z i is the mean, s i Get the Gaussian distribution constructed for the variance with x i Random number sequences of the same length.

3. The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data according to claim 1 is characterized in that: The reverse offset enhancement process in the second step is as follows: Let sample x i The length of the vibration signal is L, and the time dimension of the vibration signal is flipped, and the position of the t-th data point in the original sample is Lt in the new sample. Then the signal is connected end to end, and all data points are moved back r positions in a cycle to obtain a reverse shift enhanced sample.

4. The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data according to claim 1 is characterized in that: The additive noise enhancement process in the second step is as follows: According to the vibration signal i-th sample x i The average amplitude z i , standard deviation s i , enhanced by formula (2), where random(z i ,s i ) indicates that from z i is the mean, s i The Gaussian distribution constructed with the standard deviation of x i Random number sequence of the same length; mask(rand(L, 0.1×L)) is a further mixed salt and pepper noise addition method, indicating that x i The rand function randomly selects 0.1×L data points from the L data points, and sets the selected data points to 0 through the mask operation, while the values ​​of other data points remain unchanged. represents the i-th sample after additive noise enhancement, which is expressed as follows:

5. The intelligent fault diagnosis method based on self-supervision of unbalanced unlabeled monitoring data according to claim 1 is characterized in that: The fifth step is to design the loss function and perform pre-training, which includes: The i-th multiplicative noise enhanced sample and reverse offset enhanced samples The positive sample pairs are transformed by the feature extraction network to obtain feature representation and The cosine similarity in formula (7) is used to represent the distance between features. The matching method of other enhanced samples is the same; Positive sample pair Contrastive loss function Loss i，k The definition of is shown in formula (8), k≠i, represents the sample after the kth original sample is enhanced by reverse shift; Where τ is the temperature parameter, N is the total number of original samples, and self-supervised network training based on unlabeled data is achieved by minimizing the contrastive loss function.

Images