Rotating machine fault diagnosis method based on artificial intelligence

Abstract

The invention discloses a rotating machine fault diagnosis method based on artificial intelligence. The method comprises the following steps: step 1, collecting a vibration signal from a rotating machine to be detected; step 2, preprocessing the collected vibration signals; 3, dividing the preprocessed signals into a 80% training data set and a 20% test data set in a random sampling mode; 4, an MAF-Transform model is established, and the model is trained through the training data set; and 5, inputting a test data set to verify the quality of the extracted features and the performance of the MAF-Transform model. According to the diagnosis method, local and global features in a time-frequency domain can be extracted at the same time, and multi-dimensional feature representation is carried out on a fault state.

Description

Technical Field

[0001] The invention belongs to the technical field of rotating machinery fault diagnosis, and relates to a rotating machinery fault diagnosis method based on artificial intelligence. Background Art

[0002] Rotating machinery faults are usually manifested through sensor signals, but sensor signal measurements are easily affected by environmental background noise, especially under complex working conditions, and may contain a large amount of interference information, which poses a great challenge to the fault diagnosis of rotating machinery. Although traditional rotating machinery fault diagnosis methods have made certain progress, they cannot automatically extract data features and require manual feature design for scenarios, which is costly. In addition, CNN has achieved remarkable success in rotating machinery fault diagnosis, but due to the inherent noise of the signal, it is difficult to extract effective global features. At the same time, the local properties of the convolution kernel make it impossible to capture remote contexts, resulting in a lack of global feature information. Transformer focuses on learning global features and applies them to time series classification and prediction, achieving good results, but the effect in the field of fault diagnosis still cannot meet the requirements of implementation. The literature Y. Ding, M. Jia, Q. Miao, and Y. Cao, "A novel time-frequency transformer based on self-attention mechanism and its application in fault diagnosis of rolling bearings," Mechanical Systems and Signal Processing, vol. 168, p. 108616, 2022. proposed a new end-to-end fault diagnosis framework based on time-frequency feature Transformer to extract effective features from the time-frequency representation of vibration signals, but the anti-interference ability still needs to be improved. In addition, the literature X. Tang, Z. Xu, and Z. Wang, "A novel fault diagnosis method of rolling bearing based on integrated vision transformer model," Sensors, vol. 22, no. 10, p. 3878, 2022. The Vision Transformer model is used to classify faults according to one-dimensional original fault signals or fault time-frequency two-dimensional images. Compared with CNN, Transformer lacks the ability to extract spatial local features and is not suitable for situations with limited data. Summary of the invention

[0003] To solve the above technical problems, the purpose of the present invention is to provide a rotating machinery fault diagnosis method based on artificial intelligence, which can simultaneously extract local and global features in the time and frequency domains and perform multi-dimensional feature characterization of the fault state.

[0004] The present invention provides a rotating machinery fault diagnosis method based on artificial intelligence, comprising:

[0005] Step 1: Collect vibration signals from the rotating machinery to be detected;

[0006] Step 2: Preprocess the collected vibration signal;

[0007] Step 3: Divide the preprocessed signal into 80% training data set and 20% test data set by random sampling;

[0008] Step 4: Build the MAF-Transformer model and train the model using the training dataset;

[0009] Step 5: Input the test dataset to verify the quality of the extracted features and the performance of the MAF-Transformer model.

[0010] The present invention provides a method for diagnosing faults of rotating machinery based on artificial intelligence, which has the following features:

[0011] Beneficial effects:

[0012] The present invention uses fast Fourier transform to reconstruct the signal dimension, and converts the one-dimensional signal into a two-dimensional signal according to the period and frequency, which retains the time domain characteristics within a period and can reflect the frequency domain characteristics. Then, a two-dimensional backbone network block is used to extract the local structural features in the two-dimensional signal, and then the global time series features after deconstruction are extracted through the Transformer coding layer. The above method can effectively overcome the noise interference under different working conditions and realize the high robustness and high accuracy fault diagnosis of rotating machinery under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of a rotating machinery fault diagnosis method based on artificial intelligence of the present invention;

[0014] Figure 2 This is the network structure diagram of the MAF-Transformer model. DETAILED DESCRIPTION

[0015] The signals collected by sensors of rotating machinery under complex working conditions such as variable load and variable speed have non-stationary and nonlinear characteristics. The local and global features of a single time domain or frequency domain cannot fully describe the fault state of the rotating machinery. The existing manually designed feature diagnosis methods are costly. The feature extraction method based on the CNN network cannot capture global context features. The feature extraction method based on the Transformer network is not suitable for scenarios with large data requirements and limited data volume. In order to solve the above problems, the present invention proposes a rotating machinery fault diagnosis method based on artificial intelligence, such as Figure 1 The steps shown include:

[0016] Step 1: Collect vibration signals from the rotating machinery to be detected;

[0017] Step 2: Preprocess the collected vibration signal, specifically:

[0018] Step 2.1: Perform outlier removal processing on the collected vibration signal;

[0019] Step 2.2: Normalize the vibration signal after removing outliers;

[0020] Step 2.3: And add additional Gaussian white noise to the normalized signal.

[0021] Step 3: Divide the preprocessed signal into 80% training data set and 20% test data set by random sampling;

[0022] Step 4: Build the MAF-Transformer model and train the model using the training dataset;

[0023] like Figure 2 As shown, the MAF-Transformer model includes an input layer, an embedding layer, a slicing layer, a Transformer encoding layer, an adaptive fusion layer and a fully connected layer connected in sequence, and the outputs of the slicing layer and the position encoding embedding layer are channel-concatenated and then input into the Transformer encoding layer;

[0024] The input layer is used to receive the one-dimensional fault signal in the training data set; the embedding layer is used to reconstruct the one-dimensional fault signal into a two-dimensional signal and then extract its local features; the slicing layer is used to divide the local features into multiple signal sequence segments; the signal sequence segments and the position coding information output by the position coding embedding layer are superimposed and then input into the Transformer coding layer for global feature extraction, multi-channel feature fusion is performed through the adaptive fusion layer, and finally the diagnosis result is obtained by classification through the fully connected layer.

[0025] In specific implementation, the embedding layer reconstructs the one-dimensional fault signal into a two-dimensional signal as follows:

[0026] Step 4.1: Let X N is a one-dimensional fault signal with a length of N in the training data set, and X is obtained by fast Fourier transform FFT. N The frequency and its corresponding amplitude are:

[0027]

[0028] Among them, F and A are the frequency set and amplitude set obtained by fast Fourier transform;

[0029] Step 4.2: Find the largest first k amplitudes A' and their corresponding frequencies F':

[0030] A'={A1',A'2,…A' k}=Top k (A), F'={F1',F2',...F k '}

[0031] Among them, Top k () function means returning the largest first k values ​​in set A, k∈2~5;

[0032] Step 4.3: Calculate the corresponding period based on the first k frequencies found:

[0033]

[0034] The above is to obtain the corresponding period by dividing the length of the one-dimensional fault signal by the corresponding frequency. During the calculation process, round up to retain all the information in the one-dimensional signal to obtain k periods P i ;

[0035] Step 4.4: Use the Padding() function to fill the one-dimensional signal of length N with 0 to a length of P. i ×F i 'One-dimensional signal, the length of P is reshaped by Reshape() function i ×F i 'The one-dimensional signal is reconstructed into k P i ×F i 'The two-dimensional signal

[0036] In the specific implementation, a two-dimensional backbone block is selected to extract local features from the two-dimensional signal. The two-dimensional backbone block can be ResNet Block, Inception Block, ConvNeXt Block and Swin Transformer Block. Taking ConvNeXt Block as an example, the local feature extraction is as follows:

[0037]

[0038] in, To process the i-th two-dimensional signal through the two-dimensional backbone network block The two-dimensional feature vector obtained by local feature extraction is then expanded into the i-th one-dimensional feature vector Z by Reshape. i , because Padding() was used before, its length becomes P i ×F i ', and finally the first N data are extracted through the Truc() function to obtain a one-dimensional feature vector with a length of N.

[0039] The local features are divided into multiple signal sequence segments as follows: the expanded one-dimensional feature vector is divided into multiple overlapping or non-overlapping sequence segments of length L according to the step size S, so as to reduce the number of input tokens and reduce the complexity of model training and reasoning:

[0040]

[0041] Among them, Z L is a set of sequence segments, Indicates the Nth L sequence segment.

[0042] The signal sequence segment and the position coding information output by the position coding embedding layer are channel-concatenated and then input into the Transformer coding layer for global feature extraction and multi-channel feature fusion. Specifically:

[0043] The position encoding information is directly superimposed on the sequence segment, and then the i-th one-dimensional feature vector Z i N p The sequence segments are input into the encoder to encode and obtain the global features The k amplitudes A' are used as channel weights to perform multi-channel adaptive fusion of the extracted k global features, and then the classification results are obtained through a multi-layer perceptron:

[0044]

[0045] R Class Classification results obtained for the MAF-Transformer model.

[0046] Step 5: Input the test dataset to verify the quality of the extracted features and the performance of the MAF-Transformer model.

[0047] Example:

[0048] In order to verify the effectiveness of the method of the present invention, an experimental platform consisting of a loading system and a data acquisition system was built. Five common types of harmonic reducer faults in actual production lines were selected, namely F0: normal, F1: flexible bearing fracture in the wave generator, F2: input shaft alignment, input shaft fatigue failure due to long-term operation; F3-F4 flexible wheel tooth root and tooth tip pitting, simulating the failure of the flexible wheel with low load capacity steel wheel under high load conditions. Four data sets under working conditions were selected: constant load and constant speed Normal, constant speed and variable speed Vel, constant load and variable speed Load, and variable load and variable speed Mixture. The sampling frequency of the data acquisition system was set to 25600Hz, with 1s sampling each time, and 10 groups of samples were collected for each case. Finally, 5000 groups of data sets were collected for each working condition, of which 70% of the data were used as training sets and 30% of the data were used as test sets. Multiple Transformer-based algorithms FEDformer, MICN, TimesNet, and 1D-Vit in the field were selected as comparative experiments. Table 1 shows the accuracy table of each algorithm.

[0049] Table 1

[0050]

[0051] It can be seen that under the premise of using the same data preprocessing and the same data set, the algorithm of the present invention achieves the best accuracy compared with other algorithms, and the effectiveness of adaptive fusion is proved by the ablation experimental results with and without adaptive fusion.

[0052] In order to verify the generalization ability of the proposed fault diagnosis method in various rotating machinery fault diagnosis, we conducted experiments on the well-known CWRU bearing dataset. The specific composition of the selected dataset is shown in Table 2. The CWRU bearing dataset has three fault locations: rotor Ball, inner race, and outer race. Each fault location corresponds to three types of faults, plus no fault, a total of ten types. The training data is distributed as 640 no-fault data, 1920 rotor fault data, 1920 inner race fault data, and 1920 outer race fault data. The test data has the same distribution as the training data, and the number is 1 / 4 of that.

[0053] Table 2

[0054]

[0055] Similar to the homemade harmonic reducer dataset, we selected a high-performance algorithm based on Transformer as the comparison algorithm. The performance of the MAF-Transformer method proposed in this invention on the CWRU dataset is shown in Table 3. The MAF-Transformer method proposed in this invention achieved the best accuracy on the CWRU dataset.

[0056] Table 3

[0057]

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A rotating machinery fault diagnosis method based on artificial intelligence, It is characterized in that include: Step 1: Collect vibration signals from the rotating machinery to be detected; Step 2: Preprocess the collected vibration signal; Step 3: Divide the preprocessed signal into 80% training data set and 20% test data set by random sampling; Step 4: Build the MAF-Transformer model and train the model using the training dataset; Step 5: Input the test dataset to verify the quality of the extracted features and the performance of the MAF-Transformer model.

2. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 1, It is characterized in that The step 2 is specifically as follows: Step 2.1: Perform outlier removal processing on the collected vibration signal; Step 2.2: Normalize the vibration signal after removing outliers; Step 2.3: And add additional Gaussian white noise to the normalized signal.

3. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 1, It is characterized in that The MAF-Transformer model includes an input layer, an embedding layer, a slice layer, a Transformer encoding layer, an adaptive fusion layer and a fully connected layer connected in sequence, and the outputs of the slice layer and the position encoding embedding layer are channel-concatenated and then input into the encoding layer; The input layer is used to receive the one-dimensional fault signal in the training data set; the embedding layer is used to reconstruct the one-dimensional fault signal into a two-dimensional signal and then extract its local features; The slicing layer is used to divide local features into multiple signal sequence segments; the signal sequence segments and the position coding information output by the position coding embedding layer are superimposed and then input into the Transformer coding layer for global feature extraction, multi-channel feature fusion is performed through the adaptive fusion layer, and finally classification is performed through the fully connected layer to obtain the diagnosis result.

4. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 3, It is characterized in that The embedding layer reconstructs the one-dimensional fault signal into a two-dimensional signal as follows: Step 4.1: Let X N is a one-dimensional fault signal with a length of N in the training data set, and X is obtained by fast Fourier transform FFT. N The frequency and its corresponding amplitude: Among them, F and A are the frequency set and amplitude set obtained by fast Fourier transform; Step 4.2: Find the largest first k amplitudes A' and their corresponding frequencies F': A'={A 1 ',A' 2 ,…A' k }=Top k (A),F'={F 1 ',F 2 ',…F k '} Among them, Top k () function means returning the largest first k values ​​in set A, k∈2~5; Step 4.3: Calculate the corresponding period based on the first k frequencies found: The above is to obtain the corresponding period by dividing the length of the one-dimensional fault signal by the corresponding frequency. During the calculation process, round up to retain all the information in the one-dimensional signal to obtain k periods P i ; Step 4.4: Use the Padding() function to fill the one-dimensional signal of length N with 0 to a length of P. i ×F i 'One-dimensional signal, the length of P is reshaped by Reshape() function i ×F i 'The one-dimensional signal is reconstructed into k P i ×F i 'The two-dimensional signal 5. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 4, It is characterized in that Select a 2D backbone block to extract local features from the 2D signal, specifically: in, To process the i-th two-dimensional signal through the two-dimensional backbone network block The two-dimensional feature vector obtained by local feature extraction is then expanded into the i-th one-dimensional feature vector Z by Reshape. i , because Padding() was used before, its length becomes P i ×F i ', and finally the first N data are extracted through the Truc() function to obtain a one-dimensional feature vector with a length of N.

6. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 5, It is characterized in that The local features are divided into multiple signal sequence segments as follows: the expanded one-dimensional feature vector is divided into multiple overlapping or non-overlapping sequence segments of length L according to the step size S, so as to reduce the number of input tokens and reduce the complexity of model training and reasoning: Among them, Z L is a set of sequence segments, Indicates the Nth L sequence segment.

7. The rotating machinery fault diagnosis method based on artificial intelligence as claimed in claim 6, It is characterized in that The signal sequence segment and the position coding information output by the position coding embedding layer are channel-concatenated and then input into the Transformer coding layer for global feature extraction and multi-channel feature fusion. Specifically: The position encoding information is directly superimposed on the sequence segment, and then the i-th one-dimensional feature vector Z i N p The sequence segments are input into the encoder to encode and obtain the global features The k amplitudes A' are used as channel weights to perform multi-channel adaptive fusion of the extracted k global features, and then the classification results are obtained through a multi-layer perceptron: R Class Classification results obtained for the MAF-Transformer model.

Images