An early weak fault diagnosis method based on self-adjusting joint distribution adaptation

Abstract

The present invention discloses an early weak fault diagnosis method based on self-adjusting joint distribution adaptation, and relates to the field of mechanical engineering technology. The method comprises the following steps: step 1: collecting one-dimensional vibration signals of the source domain and early weak fault signals of the target domain. Step 2: building a Gabor-CNN backbone network. Step 3: re-dividing the two-dimensional image samples of the source domain after preprocessing. Step 4: statistically analyzing the distribution of target domain samples with the same pseudo-label, and calculating the average value to use as the threshold of the target domain samples for classification. Step 5: calculating the distribution average value corresponding to the target domain sample data before each round of backbone network classification, and calculating the ratio with the threshold in step 4, and using the ratio as the coefficient of the conditional distribution loss and the marginal distribution loss in the total loss function; and thus performing successive iterations to achieve the final classification. This method can realize the effective classification of early weak faults in the degradation stage of mechanical equipment, and prevent the high losses caused by serious faults in the later stage.

Description

Technical Field

[0001] The present invention relates to a fault diagnosis method based on deep transfer learning, to the field of mechanical engineering technology, and in particular to an early weak fault diagnosis method based on self-adjusting joint distribution adaptation. Background Art

[0002] In a long-term complex and harsh working environment, key components of mechanical equipment are prone to various faults, which will directly affect the safe operation of mechanical equipment. Diagnosing early weak faults of various components of the machinery will effectively prevent serious consequences of the equipment in the later stage. At present, the diagnosis methods for early mechanical faults mainly include signal processing-based methods and data-driven methods.

[0003] Signal processing methods used in the field of early mechanical fault diagnosis include variational mode extraction, spectral wavelet adaptive threshold denoising, improved EMD-Kurtogram method, etc. The core idea is to fully extract fault frequency information and identify and classify it through filtering and denoising methods.

[0004] Data-driven early fault diagnosis methods can be divided into deep learning-based methods and transfer learning-based methods. These methods include capsule networks combined with attention mechanisms, nonlinear sparse blind deconvolution, and support vector machines based on genetic algorithms. The core of deep learning-based methods is to extract sample fault features so that the model can classify fault types based on significant differences between samples. For early weak faults of bearings, deep learning methods usually require feature enhancement to achieve accurate identification and classification. Transfer learning methods, on the basis of deep learning, use the existing knowledge in the source domain to classify the target domain data by narrowing the correlation between the source domain and the target domain, which can effectively solve the problem of unlabeled target domain samples in practical applications. Joint distribution adaptation is an adaptive method that combines cross-domain marginal distribution and conditional distribution. It can effectively adapt to the distribution difference between the source domain and the target domain and is widely used in the transfer learning process. It includes joint distribution adaptation methods combined with neighborhood-preserving embedding and manifold embedding, joint distribution offset difference methods based on AL-MMD, and transfer learning methods with adaptive regularization learning. Transfer learning-based methods effectively improve the classification performance of the model by aligning the difference features between the source domain and the target domain and using source domain label data.

[0005] The shortcomings of the above technical solution are:

[0006] (1) During the normal operation of the machine, it is impossible to collect labeled data of various early weak fault type samples, resulting in a lack of training set samples.

[0007] (2) The early weak fault vibrations of mechanical equipment components are affected by environmental noise, and the fault characteristics presented are not obvious enough. Converting the one-dimensional signal into a two-dimensional grayscale image and using 2D-CNN for feature extraction cannot effectively separate the fault characteristics.

[0008] (3) Although the joint distribution adaptation combines the marginal distribution and the conditional distribution, the conditional distribution error function has a large correlation with the pseudo-label samples. When the confidence of the pseudo-label samples is low, serious misclassification will occur, resulting in a decrease in classification accuracy. Summary of the invention

[0009] The technical purpose of the present invention is to propose an early weak fault diagnosis method based on self-adjusting joint distribution adaptation. This method is not obvious enough for weak fault characteristics caused by degradation of mechanical equipment during operation. It is necessary to preprocess the weak fault signal to enhance the characteristics, and a new network is needed to effectively extract the enhanced characteristics.

[0010] At present, most of the faults collected in public data sets are artificially created, which have large distribution differences from the degradation faults existing in actual industrial scenarios. The task of migrating artificial faults to degradation faults is more in line with actual application situations.

[0011] Conditional distribution has high requirements on the confidence of pseudo labels of target domain samples after preliminary classification. When there are too many noisy pseudo labels, the conditional distribution loss will be too large, resulting in a sharp decrease in classification accuracy. A purification method for pseudo-label samples is needed to improve the confidence of pseudo-label samples.

[0012] In the early stages of training, since the model has not fully converged, the preliminary pre-classification joint distribution loss function will have a negative transfer effect on the classification results. An adaptive adjustment parameter is needed to change the weight of the joint distribution loss in the total loss function to prevent serious misclassification results.

[0013] In view of the above technical deficiencies, the present invention provides an early weak fault diagnosis method based on deep transfer learning to solve the problems existing in the existing fault diagnosis method based on transfer learning. The method comprises the following steps:

[0014] Step 1: Collect the one-dimensional vibration signal in the source domain and the early weak fault signal in the target domain. The early weak fault signal usually refers to the abnormal vibration signal of the bearing between the normal and fault states. And perform data preprocessing, including noise reduction processing and time-frequency analysis, in order to reduce the impact of environmental noise on weak fault signals and enhance the weak fault features in the signal. Convert the one-dimensional vibration signal into a two-dimensional grayscale image.

[0015] Step 2: Build a Gabor-CNN backbone network and use the source domain data preprocessed in step 1 to train the network model. The backbone network extracts the texture of the two-dimensional image samples at various angles through Gabor filters of multiple scales, thereby effectively learning different types of weak fault features. The network model is then used to perform preliminary classification of weak fault samples in the target domain.

[0016] Step 3: Re-divide the source domain two-dimensional image samples after preprocessing n times, each time dividing them into samples with label n and samples with label non-n, use the samples to train a binary classifier, and perform a two-stage screening on the pseudo-label target domain samples.

[0017] Step 4: Count the distribution of target domain samples with the same pseudo-label after the second-stage screening, and calculate the average value, which is used as the threshold for classifying target domain samples.

[0018] Step 5: Calculate the distribution mean of the target domain sample data before each round of classification by the backbone network, and compare it with the threshold in step 4. Use the ratio as the coefficient of the conditional distribution loss and marginal distribution loss in the total loss function; perform successive iterations to achieve the final classification.

[0019] Furthermore, in step 1, a sliding window with a length of 1024 is used to intercept a one-dimensional vibration signal, and then an envelope spectrum analysis based on Hilbert transform is used for noise reduction, and the cutoff frequency is set to 300 to filter out some high-frequency noise. The time domain signal is Fourier transformed to obtain a frequency domain signal, and the frequency domain signal is converted into a two-dimensional image using the Gram angle field. The position of each pixel corresponds to the frequency of the original signal, so the position information in the image reflects the sequence relationship of the weak fault signal in the frequency domain. The Gram angle field encodes the nonlinear dynamic characteristics of the early weak fault signal into the image by calculating the inner product of the sine and cosine functions. This conversion method can capture the phase relationship of the weak anomaly in the signal and can better reflect the nonlinear dynamic behavior therein. The samples are scaled and grayed to generate a grayscale image with a final size of 64*64.

[0020] Furthermore, in step 2: first, build a Gabor-CNN backbone network. In each convolution layer, the convolution filter is modulated by Gabor filters of different directions and scales to generate convolution Gabor directional filters GoFs, and the Gabor filter is used to replace the convolution kernel in the CNN network, thereby realizing the convolution process. Gabor-CNN is trained using source domain samples of various fault types in step 1. The Gabor filter can decompose the source domain image in frequency and direction at different scales, and learn the relevant features of specific faults at different scales and directions. Then, the trained model is used to preliminarily classify the early weak fault samples in the target domain. The filter response map obtained after the convolution operation of the Gabor filter can reflect the frequency components of the image at different scales and directions, corresponding to the characteristic information of the weak fault samples.

[0021] Furthermore, in step 3: the pseudo-label noise filtering module filters out the noisy pseudo-labels of the target domain samples by using the following method: first, source domain samples with label n and source domain samples with label non-n are used as training sets to train n networks for secondary screening of target domain samples with pseudo-label n. Binary classification is performed on target domain samples containing noisy pseudo-labels.

[0022] Furthermore, in step 4: the root mean square error between every two samples of the target domain sample set after the secondary screening is calculated and the average value is taken as the distribution density benchmark value of the target domain sample set.

[0023] Furthermore, in step 5: the sample distribution average of the target domain data after classification by the backbone network is compared with the distribution density reference value in step 4, and the ratio of the two is used as the weight coefficient for controlling the marginal distribution and the conditional distribution. The joint distribution function uses two coefficients to control the weights of the marginal distribution and the conditional distribution, and then in each round of training iteration, the joint distribution error is gradually introduced into the loss function to complete the adaptive migration from the source domain data to the target domain data. The loss function of the marginal distribution is CORAL, and the loss function of the conditional distribution is MMD.

[0024] Compared with the prior art, the present invention has the following technical advantages.

[0025] After the one-dimensional vibration signal is subjected to envelope spectrum denoising in the present invention, the influence of environmental noise on the early weak fault vibration signal is first reduced. The signal converted to the frequency domain is transformed into a two-dimensional image using the Gram angular field, and the weak fault features are converted into textures of different brightness in the image, effectively achieving feature enhancement. The Gabor directional filter can capture the direction and scale information of weak fault features on multiple channels, so that Gabor-CNN can effectively extract features from two-dimensional image samples.

[0026] The present invention improves the problem of noisy pseudo-labels in target domain data after preliminary classification, and effectively filters the noise labels of the target domain samples of preliminary classification through a method of re-dividing the source domain data and training a secondary screening model. This is more conducive to the normal back propagation of conditional distribution loss and marginal distribution loss in the subsequent loss function and the model optimization process.

[0027] The present invention proposes an improved joint distribution method, which can controllably adjust the coefficients of the loss function to avoid serious negative transfer in the training process caused by excessive conditional distribution and marginal distribution losses when the difference between source domain data and target domain data is too large.

[0028] Aiming at the situation that most existing data sets are artificial fault samples, while the faults in actual industrial scenarios are degradation faults, the present invention proposes a method which can realize the migration task from artificial faults to degradation faults, and has higher practical application value.

[0029] The present invention can achieve effective classification of early weak faults in the degradation stage of mechanical equipment, can achieve effective prevention and repair of large mechanical faults, and prevent high losses caused by later serious faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 These are the normal, inner ring fault, rolling element fault, and outer ring fault sample images after preprocessing of the CRWU bearing dataset.

[0031] Figure 2 These are the normal, inner race fault, rolling element fault, and outer race fault sample images after preprocessing of the IMS bearing dataset.

[0032] Figure 3 It is the overall flow chart of the proposed method.

[0033] Figure 4 It is a diagram of the process of re-dividing source domain samples and outputting the training results of the two-stage screening network.

[0034] Figure 5 It is the accuracy curve of four groups of migration tasks.

[0035] Figure 6 It is the loss curve of four groups of migration tasks. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0037] Bearings are the most common parts in mechanical equipment and are prone to various faults that cause equipment downtime. The effectiveness of the present invention is demonstrated by using bearing data as an example. The data selected are part of the bearing vibration signals of the Western Reserve University bearing data set and the IMS bearing data set of the University of Cincinnati. The CWRU bearing data set is used as the source domain data, and there are four states: normal, inner ring fault, rolling element fault, and outer ring fault, where the size of each fault type is 0.1778mm. At the same time, four working conditions are selected, and the bearing speeds are 1730rpm, 1750rpm, 1772rpm, and 1797rpm respectively. The early degradation part of the IMS bearing data set is intercepted as the target domain data, and there are four states: normal, inner ring fault, rolling element fault, and outer ring fault, and the bearing speed is 2000rpm. First, signal preprocessing is performed, and the one-dimensional signals of different fault types are intercepted with 1024 as the sliding window length, and envelope spectrum denoising is performed, and the cutoff frequency is set to 300Hz. After denoising, it is converted into a two-dimensional grayscale image, such as Figure 1 , Figure 2 shown.

[0038] The source domain data is input into the Gabor-CNN backbone network. The network parameters are shown in Table 1:

[0039] Table 1 Gabor-CNN network parameters

[0040]

[0041]

[0042] The network trained with source domain data is used to preliminarily classify the target domain data. At this time, the target domain data is preliminarily divided into 4 different states. Then the source domain data is re-divided n times, and a set of secondary screening network models is trained for each division. The network model is based on 2D-CNN, and the parameters are shown in the following table:

[0043] Table 2 2D-CNN network parameters

[0044]

[0045] After the target domain data has been screened twice by the above model, the sample distribution average of each category of the target domain data is calculated, and the sample distribution of the target domain data before screening is calculated at the same time. The ratio of the two is used as the coefficient of the joint distribution loss in the loss function (the default is 0), as shown in formula (8).

[0046] The backbone network extracts the features of the source domain data and the target domain data, and uses MMD to measure the conditional distribution difference between the two domain features, that is, the distribution distance between samples with the same label. CORAL is used to measure the marginal distribution difference between the two domain features, that is, the distribution distance between the source domain and the target domain samples. The obtained conditional distribution and marginal distribution losses are multiplied by the corresponding coefficients and added to the total loss function. The backbone network is back-propagated and optimized, and the final classification results of the target domain samples are output after iteration.

[0047]

[0048] Formula (1) shows that the Gram angle field defines polar coordinates and maintains the numerical relationship while introducing the time relationship. represents the frequency domain data of the normalized vibration signal, In this method, it represents the frequency sequence of the vibration signal. i With r i Then it represents the angle and radius of the polar coordinates of the inner product in the transformed two-dimensional image.

[0049]

[0050] Formula (2) defines the Gabor kernel, where z = (x, y), The scale v = 0, ..., V, and the direction u = 0, ..., U. In the experiment, U is set to π, V is set to π / 2, and k is used to define kernels of different scales and directions.

[0051]

[0052] Formula (3) and Formula (4) define the calculation process of MMD and CORAL, H represents the reproducing kernel Hilbert space (RKHS), and Φ represents the mapping function of RKHS. S With Cov T represents the covariance measure between the source domain and the target domain, X S ,X T Represents the source domain and target domain samples, and d represents the output of the network.

[0053]

[0054] Formula (5) represents the calculation process of conditional probability distribution, X S ,Y S ,X T ,Y T They represent source domain samples, source domain labels, target domain samples, and target domain labels respectively. C represents the total number of label types, T represents feature extraction, and E represents spatial mapping.

[0055]

[0056] Formula (6) represents the calculation process of marginal distribution.

[0057]

[0058] Formula (7) represents the calculation process of the sample distribution mean value of each category of target domain data. RSMSE is the root mean square value of the distance between all samples, where i′ represents the data except i in the set {1,2,3,…,n}.

[0059]

[0060] In formula (8) represents the weight coefficient in the loss function, represents the target domain sample with label i after filtering, It indicates the target domain samples with label i before filtering.

[0061]

[0062] Formula (9) represents the calculation formula of the final loss function, L Train Represents the cross entropy loss during training.

[0063] The four working condition samples of the CWRU dataset are used as the source domain, and the early degradation samples of the IMS dataset are used as the target domain. Four groups of migration experiments are set up. The results are as follows: Figure 5 , Figure 6 shown.

[0064] Table 3 Final accuracy of four groups of migration experiments

[0065]

[0066] The present invention pre-processes the early weak vibration signal to reduce the noise impact and show the fault characteristics more clearly. Gabor-CNN is used as the backbone network model to improve the model's ability to extract different types of weak fault characteristics.

[0067] By performing a secondary screening on the target domain samples, the target domain sample distribution with filtered label noise is obtained, and the ratio of the target domain sample distribution before and after screening is used as the weight coefficient to control the joint distribution loss in the total loss function, thereby achieving adaptive migration from the source domain to the target domain.

Claims

1. An early weak fault diagnosis method based on self-adjusting joint distribution adaptation, characterized in that: The method comprises the following steps, Step 1: Collect the one-dimensional vibration signal in the source domain and the early weak fault signal in the target domain. The early weak fault signal usually refers to the abnormal vibration signal of the bearing between the normal and fault states; and perform data preprocessing, including noise reduction processing and time-frequency analysis, in order to reduce the impact of environmental noise on the weak fault signal and enhance the weak fault characteristics in the signal; Convert one-dimensional vibration signal into two-dimensional grayscale image; Step 2: Build a Gabor-CNN backbone network and use the source domain data preprocessed in step 1 to train the network model. The Gabor-CNN backbone network extracts the texture of the two-dimensional image samples at various angles through Gabor filters of multiple scales and learns different types of weak fault features. The network model is then used to preliminarily classify weak fault samples in the target domain; Step 3: Re-divide the source domain two-dimensional image samples after preprocessing n times, each time dividing them into samples with label n and samples with label non-n, use the samples to train a binary classifier, and perform a two-stage screening of pseudo-label target domain samples; Step 4: Count the distribution of target domain samples with the same pseudo-label after the second-stage screening, and calculate the average value, which is used as the threshold of the target domain samples for classification; Step 5: Calculate the distribution mean of the target domain sample data after each round of backbone network classification, and compare it with the threshold in step 4. Use the ratio as the coefficient of the conditional distribution loss and marginal distribution loss in the total loss function; perform successive iterations to achieve the final classification.

2. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 1 is characterized in that: In the step 1, a sliding window with a length of 1024 is used to intercept a one-dimensional vibration signal, and then an envelope spectrum analysis based on Hilbert transform is used for noise reduction, and the cutoff frequency is set to 300 to filter out some high-frequency noise; the time domain signal is Fourier transformed to obtain a frequency domain signal, and the frequency domain signal is converted into a two-dimensional image using the Gram angle field. The position of each pixel corresponds to the frequency of the original signal, and the position information in the image reflects the sequence relationship of the weak fault signal in the frequency domain; the Gram angle field encodes the nonlinear dynamic characteristics of the early weak fault signal into the image by calculating the inner product of the sine and cosine functions; The samples are scaled and grayscaled to generate a grayscale image with a final size of 64*64.

3. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 1 is characterized in that: In the step 2: a Gabor-CNN backbone network is constructed; in each convolution layer, the convolution filter is modulated by Gabor filters of different directions and scales to generate convolution Gabor directional filters GoFs, and the convolution kernel in the CNN network is replaced by the Gabor filter, thereby realizing the convolution process; the Gabor-CNN is trained using source domain samples of various fault types in step 1, and the Gabor filter decomposes the source domain image in frequency and direction at different scales, and learns relevant features of specific faults at different scales and directions; The trained model is then used to perform preliminary classification of early weak fault samples in the target domain. The filter response map obtained after the Gabor filter convolution operation reflects the frequency components of the image at different scales and directions, corresponding to the characteristic information of the weak fault samples.

4. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 1 is characterized in that: In step 3: the following method is used to filter out the noisy pseudo-labels of the target domain samples. First, source domain samples with a label of n and source domain samples with a label other than n are used as training sets to train n networks for secondary screening of target domain samples with a pseudo-label of n; and target domain samples containing noisy pseudo-labels are classified into two categories.

5. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 1 is characterized in that: In step 4: the root mean square error between every two samples of the target domain sample set after the two-stage screening is calculated and the average value is taken as the distribution density benchmark value of the target domain sample set.

6. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 5 is characterized in that: The sample distribution average of the target domain data after classification by the backbone network is compared with the distribution density benchmark value in step 4, and the ratio of the two is used as the weight coefficient for controlling the marginal distribution and the conditional distribution; the joint distribution function uses two coefficients to control the weights of the marginal distribution and the conditional distribution, and then in each round of training iteration, the joint distribution error is gradually introduced into the loss function to complete the adaptive migration from the source domain data to the target domain data.

7. The early weak fault diagnosis method based on self-adjusting joint distribution adaptation according to claim 6 is characterized in that: The loss function of the marginal distribution is CORAL, and the loss function of the conditional distribution is MMD.

Images