A multi-scale feature-focused fault diagnosis method for electromechanical equipment

Abstract

The present invention discloses a multi-scale feature-focused electromechanical equipment fault diagnosis method, comprising: 1) inputting a current signal data set collected from a reducer; 2) performing simple feature engineering to construct a new data set; 3) dividing the newly constructed data set into a training set and a test set and labeling them; 4) using the training set divided in step 3) to train a multi-scale dual-attention fault diagnosis model based on a convolutional neural network; 5) performing fault diagnosis on the test set to obtain a final diagnosis prediction result. The present invention utilizes the powerful feature extraction capability of the convolutional neural network while adding a dual-attention mechanism combined with a multi-scale convolution method to improve the feature extraction capability of the convolutional neural network, adaptively enhance information related to the fault, and suppress features unrelated to the fault. It has a significant improvement effect on the current data set of the electromechanical equipment reducer where the fault features are not obvious, thereby improving the diagnostic efficiency of the electromechanical equipment.

Description

Technical Field

[0001] The present invention belongs to the field of automation technology, and in particular relates to a multi-scale feature-focused electromechanical equipment fault diagnosis method. Background Art

[0002] With the improvement of the automation level of the manufacturing industry, electromechanical equipment has developed rapidly in recent decades. At present, various types of electromechanical equipment such as industrial robots, elevators, CNC machine tools, etc. are widely used in various tasks in the manufacturing industry such as machining, assembly, welding, and injection molding. In the operation of electromechanical equipment, the reducer is the core power component. Once the reducer fails, it will cause great losses to production. Accurate fault diagnosis of the reducer can effectively shorten the maintenance time and maintenance cost of the equipment. Therefore, the realization of fault diagnosis of the reducer of electromechanical equipment plays a huge practical significance to ensure the normal operation of production.

[0003] Traditional maintenance methods for electromechanical equipment are mainly implemented by two passive strategies: post-maintenance and regular maintenance, which cannot meet the needs of modern large-scale and highly complex electromechanical equipment maintenance. In the context of the rapid development of artificial intelligence and big data technology, data-driven fault diagnosis has received great attention. However, the current fault diagnosis of reducers is mainly based on collecting vibration data for analysis. However, during the operation of electromechanical equipment, vibration data cannot be obtained because it is difficult to install sensors. Signals such as feedback current and feedback torque can be collected from the motor tightly coupled with the reducer, which can also be used to reveal the health status of the reducer. However, there is still no research on the fault diagnosis of reducers triggered by motor signals. Therefore, it is of great significance to study and develop a fault diagnosis method for reducers of electromechanical equipment, analyze the operating data of the motor in the electromechanical equipment, and thus judge the operating status of the reducer in the electromechanical equipment, so as to provide a basis for whether the electromechanical equipment needs maintenance.

[0004] Zhi Zhuo et al. proposed a joint wavelet regional correlation threshold denoising (WRCTD) algorithm and convolutional neural network long short-term memory (CNN-LSTM) fault detection method to analyze and diagnose the vibration signal of the harmonic reducer. (Zhi Z, Liu L, Liu D, et al. Fault Detection of the Harmonic Reducer Based on CNN-LSTM With a Novel Denoising Algorithm[J]. IEEE Sensors Journal, Institute of Electrical and Electronics Engineers Inc., 2022, 22(3): 2572–2581.) During the operation of the harmonic reducer, in addition to collecting vibration signals, a large amount of mechanical vibration noise and environmental noise are also collected. The WRCTD algorithm first uses the regional correlation of wavelet decomposition coefficients and the 3σ criterion to suppress the noise of the original sensor data; then, the CNN-LSTM model can mine the hidden features of the processed sensor data and correctly identify the fault. This method can analyze and diagnose the vibration signal of the harmonic reducer very well, but it has the following disadvantages: the vibration signal generated by the operation of the harmonic reducer has the characteristics of non-stationarity, nonlinearity, impact and modulation, which usually brings difficulties to fault diagnosis. The required feature engineering is too complicated and requires a lot of professional knowledge, which raises the threshold of deep learning in the field of fault diagnosis and is not conducive to application. In addition, LSTM lacks a special feature mining step and is insensitive to key features when processing data with deep hidden features.

[0005] Wang Xin et al. proposed a deep learning multimodal sensor fusion method for bearing fault diagnosis (Wang X, Mao D, Li X. Bearing fault diagnosis based on vibro-acoustic data fusion and 1D-CNN network [J]. Measurement, Elsevier, 2021, 173: 108518.) called the vibration acoustic sensor data fusion (VAF) algorithm based on one-dimensional convolutional neural network. The algorithm processes the signals collected simultaneously by two different modal sensors, the accelerometer and the microphone, and extracts features from the signals in the feature extraction stage based on 1d-CNN. The extracted features are then fused in the fusion stage, and finally classified by the SoftMax classifier to determine the type of bearing fault. However, this method does not perform data preprocessing before inputting into 1d-CNN, and directly inputs the raw data into the model. Although good results have been achieved in specific environments, noise may have a greater impact on the model in other complex environments. In addition, the signal acquisition process is relatively complicated and requires vibration signals and acoustic signals. Summary of the invention

[0006] The present invention proposes a multi-scale dual attention electromechanical equipment fault diagnosis method based on convolutional neural network. By extracting features from the current data collected by the sensor on the reducer of the electromechanical equipment and constructing a multi-scale convolutional neural network model combined with the attention mechanism, the fault diagnosis model has universality and high accuracy, and can effectively realize the fault diagnosis of electromechanical equipment:

[0007] (1) Solve the problem of feature extraction of reducers for electromechanical equipment. Overly complex feature engineering will lead to a waste of resources. Less feature processing or no feature processing will cause the noise in the signal to have a greater impact on the model. Current signals are usually easier to obtain than vibration signals, and fault diagnosis can be performed by performing simple feature engineering.

[0008] (2) Solve the problem of identifying the feedback current characteristics of the reducer of electromechanical equipment. When processing data with deep hidden features, networks such as LSTM and CNN are insensitive to key features and have insufficient ability to distinguish fault features.

[0009] The technical solution adopted by the present invention is as follows:

[0010] A multi-scale feature-focused electromechanical equipment fault diagnosis method comprises the following steps:

[0011] (1) Collect the current data of the normal operation of electromechanical equipment to form a data set;

[0012] (2) Data preprocessing and feature engineering of the data collected in step (1);

[0013] (3) Shuffle the dataset, divide the newly constructed dataset into training set and test set and label them;

[0014] (4) Construct a multi-scale feature attention convolution module.

[0015] (5) Construct a deep convolutional neural network fault diagnosis model.

[0016] (6) adding the multi-scale feature attention convolution module constructed in step (4) to the fault diagnosis model constructed in step 5, and training the fault diagnosis model using the data set input constructed in step (3);

[0017] (7) Input the test set constructed in step (3) into the fault diagnosis model with trained network parameters to evaluate the fault diagnosis effect.

[0018] The data preprocessing in step (2) includes the following steps:

[0019] Step 21. Construct features, take the mean (Formula 1), standard deviation (Formula 2) and kurtosis (Formula 3) for every b data, with a step size of s;

[0020]

[0021]

[0022]

[0023] Where data i represents the i-th data, and n represents the total number of samples.

[0024] Step 22. Normalize and map the new sequence to [-1, 1]; the Min-Max normalization formula used is:

[0025]

[0026] Among them, i represents the row where the data is located, and j represents the column where the data is located.

[0027] Step 23. Build a sliding window for the data set, with each time window size being N tw ×N ft , N tw Represents the time series dimension, N ft Represents the number of features.

[0028] Step (4) comprises the following steps:

[0029] A. Deep convolutional neural networks are mainly composed of multiple convolutional neural networks and fully connected layers, and each input data sample is a time window;

[0030] B. Three of the convolutional layers are used for feature extraction and have the same configuration: the number of convolution kernels is F N , size F L × 1. The zero padding method is used to keep the feature map size unchanged. So far, the output of the convolutional layer is F N ×N tw ×N ft The last convolutional layer aggregates the previous features into a feature map with a kernel size of 3×1 to obtain a high-level representation of the original features.

[0031] C. Flatten the 2D feature map and connect a fully connected layer. The dropout technology is used for the flattening layer to prevent overfitting. The output layer consists of a fully connected layer and a SoftMax function, which is used to classify the output features of the feature extraction layer.

[0032] Furthermore, although a 2D convolutional network is used for feature extraction, the convolution operation is actually 1D. Step (5) includes the following steps:

[0033] A. Spatial attention module. For local features A∈R C×H×W , first send it to a convolutional layer to generate two new feature maps B, C∈R C×H×W , and then reshaped into B, C ∈ R C×N , N = H × W, then perform matrix multiplication on the transpose of B and C, and use the SoftMax layer to calculate the spatial attention map S∈R N×N , as shown in formula (5):

[0034]

[0035] At the same time, A is sent to a convolutional layer to obtain a new feature map D∈R C×H×W , and then reshaped into D∈R C ×N , then perform matrix multiplication of the transpose of S and D, and reshape the result into R C×H×W Finally, the result is combined with a learnable parameter

[0036] Multiply α and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (6):

[0037]

[0038] B. Channel attention module. Different from the position attention module, it directly uses the original feature map A∈RC×H×W To calculate the channel attention map X∈R C×C , specifically reshape A∈R C×N , then perform matrix multiplication on A by taking the transpose of A and reshape the result into R C×H×W , and use the SoftMax layer to calculate the channel attention map X∈R N×N , as shown in formula (7).

[0039]

[0040] Then perform matrix multiplication of the transpose of X and A and reshape the result into R C×H×W Finally, multiply the result by a learnable parameter β and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (8):

[0041]

[0042] C. Then, the output feature maps of the two attentions are concatenated and input into the multi-scale convolution module. The multi-scale convolution module contains 4 branches, each of which contains different convolution layers. Conv2 is 1×1 convolution, and Conv3 is 3×3 convolution. The outputs of the last 4 branches are concatenated and then passed through the BN layer to prevent overfitting.

[0043] Step (6) comprises the following steps:

[0044] A. All convolutional layers use tanh as the activation function, use the Xaviernormal initializer to initialize the network weights, use the BP algorithm to adjust the network through the cross entropy loss, and the Adam algorithm is used as the optimizer.

[0045] B. Use a learning rate of 0.001 for fast optimization in the first 100 cycles, and then use a learning rate of 0.0001 to make the iteration converge smoothly.

[0046] The beneficial effects brought about by the technical solution of the present invention are:

[0047] 1. The method of the present invention realizes the reducer fault diagnosis by automatically extracting fault features in a hierarchical manner without the need for a complicated feature extraction process.

[0048] 2. The dual self-attention multi-scale convolution module designed by the method of the present invention can selectively learn fault-related features, adaptively enhance fault-related information, and suppress features unrelated to the fault, so that the model pays more attention to features that are valuable for classification, and adds a multi-scale method to further improve the model's fault recognition and generalization capabilities.

[0049] 3. The method of the present invention can effectively improve the model fault diagnosis accuracy while increasing fewer network parameters, solving the problem of poor generalization performance of traditional convolutional neural network models. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 This is the dual attention multi-scale convolution module architecture of the present invention. DETAILED DESCRIPTION

[0052] The specific technical solution of the present invention is explained in conjunction with embodiments.

[0053] Figure 1 It is an overall flow chart of the electromechanical equipment fault diagnosis method focusing on multi-scale features, including the following steps:

[0054] (1) Collect the current data of the normal operation of electromechanical equipment to form a data set;

[0055] (2) Data preprocessing and feature engineering of the data collected in step (1);

[0056] (3) Shuffle the dataset, divide the newly constructed dataset into training set and test set and label them;

[0057] (4) Construct a multi-scale feature attention convolution module.

[0058] (5) Construct a deep convolutional neural network fault diagnosis model.

[0059] (6) adding the multi-scale feature attention convolution module constructed in step (4) to the fault diagnosis model constructed in step 5, and training the fault diagnosis model using the data set input constructed in step (3);

[0060] (7) Input the test set constructed in step (3) into the fault diagnosis model with trained network parameters to evaluate the fault diagnosis effect.

[0061] The data preprocessing in step (2) includes the following steps:

[0062] Step 21. Construct features, take the mean (Formula 1), standard deviation (Formula 2) and kurtosis (Formula 3) for every b data, with a step size of s;

[0063]

[0064]

[0065]

[0066] Where datai represents the i-th data, and n represents the total number of samples.

[0067] Step 22. Normalize and map the new sequence to [-1, 1]; the Min-Max normalization formula used is:

[0068]

[0069] Among them, i represents the row where the data is located, and j represents the column where the data is located.

[0070] Step 23. Build a sliding window for the data set, with each time window size being N tw ×N ft , N tw Represents the time series dimension, N ft Represents the number of features.

[0071] Step (4) comprises the following steps:

[0072] A. Deep convolutional neural networks are mainly composed of multiple convolutional neural networks and fully connected layers, and each input data sample is a time window;

[0073] B. Three of the convolutional layers are used for feature extraction and have the same configuration: the number of convolution kernels is F N , size F L × 1. The zero padding method is used to keep the feature map size unchanged. So far, the output of the convolutional layer is F N ×N tw ×N ft The last convolutional layer aggregates the previous features into a feature map with a kernel size of 3×1 to obtain a high-level representation of the original features.

[0074] C. Flatten the 2D feature map and connect a fully connected layer. The dropout technology is used for the flattening layer to prevent overfitting. The output layer consists of a fully connected layer and a SoftMax function, which is used to classify the output features of the feature extraction layer.

[0075] Furthermore, although a 2D convolutional network is used for feature extraction, the convolution operation is actually 1D.

[0076] Figure 2 This is the structure diagram of the dual self-attention multi-scale convolution module.

[0077] Step (5) comprises the following steps:

[0078] A. Spatial attention module. For local features A∈R C×H×W , first send it to a convolutional layer to generate two new feature maps B, C∈R C×H×W , and then reshaped into B, C ∈ RC×N , N = H × W, then perform matrix multiplication on the transpose of B and C, and use the SoftMax layer to calculate the spatial attention map S∈R N×N , as shown in formula (5):

[0079]

[0080] At the same time, A is sent to a convolutional layer to obtain a new feature map D∈R C×H×W , and then reshaped into D∈R C ×N , then perform matrix multiplication of the transpose of S and D, and reshape the result into R C×H×W Finally, multiply the result by a learnable parameter α and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (6):

[0081]

[0082] B. Channel attention module. Different from the position attention module, it directly uses the original feature map A∈R C×H×W To calculate the channel attention map X∈R C×C , specifically reshape A∈R C×N , then perform matrix multiplication on A by taking the transpose of A and reshape the result into R C×H×W , and use the SoftMax layer to calculate the channel attention map X∈R N×N , as shown in formula (7).

[0083]

[0084] Then perform matrix multiplication of the transpose of X and A and reshape the result into R C×H×W Finally, multiply the result by a learnable parameter β and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (8):

[0085]

[0086] C. Then, the output feature maps of the two attentions are concatenated and input into the multi-scale convolution module. Figure 2 As shown in the figure, the multi-scale convolution module contains 4 branches, each of which contains different convolution layers. Conv2 is 1×1 convolution, and Conv3 is 3×3 convolution. The outputs of the last 4 branches are concatenated and then passed through the BN layer to prevent overfitting.

[0087] Step (6) comprises the following steps:

[0088] A. All convolutional layers use tanh as the activation function, use the Xaviernormal initializer to initialize the network weights, use the BP algorithm to adjust the network through the cross entropy loss, and the Adam algorithm is used as the optimizer.

[0089] B. Use a learning rate of 0.001 for fast optimization in the first 100 cycles, and then use a learning rate of 0.0001 to make the iteration converge smoothly.

Claims

1. A multi-scale feature-focused electromechanical equipment fault diagnosis method, characterized in that: The steps include: (1) Collect the current data of the normal operation of electromechanical equipment to form a data set; (2) Data preprocessing and feature engineering of the data collected in step (1); (3) Shuffle the dataset, divide the newly constructed dataset into training set and test set and label them; (4) Construct a multi-scale feature attention convolution module; Step (4) comprises the following steps: A. A deep convolutional neural network consists of multiple convolutional neural networks and fully connected layers, and each input data sample is a time window; B. Three of the convolutional layers are used for feature extraction and have the same configuration: the number of convolution kernels is F N , size F L ×1; use zero padding to keep the feature map size unchanged. So far, the output of the convolutional layer is F N ×N tw ×N ft ; The last convolutional layer aggregates the previous features into a feature map with a kernel size of 3×1 to obtain a high-level representation of the original features; C. Flatten the 2D feature map and connect a fully connected layer. The dropout technique is used for the flattening layer to prevent overfitting. The output layer consists of a fully connected layer and a SoftMax function to classify the output features of the feature extraction layer. (5) Construct a deep convolutional neural network fault diagnosis model; Step (5) comprises the following steps: A. Spatial attention module; for local features A∈R C×H×W , first send it to a convolutional layer to generate two new feature maps B, C∈R C×H×W , and then reshaped into B, C ∈ R C×N , N = H × W, then perform matrix multiplication on the transpose of B and C, and use the SoftMax layer to calculate the spatial attention map S∈R N×N , as shown in formula (5): At the same time, A is sent to a convolutional layer to obtain a new feature map D∈R C×H×W , and then reshaped into D∈R C×N , then perform matrix multiplication of the transpose of S and D, and reshape the result into R C×H×W Finally, multiply the result by a learnable parameter α and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (6): B. Channel attention module; Different from the position attention module, it directly uses the original feature map A∈R C×H×W To calculate the channel attention map X∈R C×C , specifically reshape A∈R C×N , then perform matrix multiplication on A by taking the transpose of A and reshape the result into R C×H×W , and use the SoftMax layer to calculate the channel attention map X∈R N×N , as shown in formula (7); Then perform matrix multiplication of the transpose of X and A and reshape the result into R C×H×W Finally, multiply the result by a learnable parameter β and perform element-wise addition operation with the original feature A to obtain the final result E∈R C×H×W , as shown in formula (8): C. Then, the output feature maps of the two attentions are concatenated and input into the multi-scale convolution module; the multi-scale convolution module contains 4 branches, each of which contains different convolution layers Conv2 is 1×1 convolution, Conv3 is 3×3 convolution; the outputs of the last 4 branches are concatenated and then passed through the BN layer to prevent overfitting; (6) adding the multi-scale feature attention convolution module constructed in step (4) to the fault diagnosis model constructed in step 5, and training the fault diagnosis model using the data set input constructed in step (3); (7) Input the test set constructed in step (3) into the fault diagnosis model with trained network parameters to evaluate the fault diagnosis effect.

2. The multi-scale feature-focused electromechanical equipment fault diagnosis method according to claim 1, characterized in that: The data preprocessing in step (2) includes the following steps: Step 21. Construct features, take the mean (Formula 1), standard deviation (Formula 2) and kurtosis (Formula 3) for every b data, with a step size of s; Where data i represents the i-th data, and n represents the total number of samples; Step 22. Normalize and map the new sequence to [-1, 1]; the Min-Max normalization formula used is: Among them, i represents the row where the data is located, and j represents the column where the data is located; Step 23. Build a sliding window for the data set, with each time window size being N tw ×N ft , N tw Represents the time series dimension, N ft Represents the number of features.

3. The multi-scale feature-focused electromechanical equipment fault diagnosis method according to claim 1, characterized in that: Step (6) comprises the following steps: A. All convolutional layers use tanh as the activation function, use the Xaviernormal initializer to initialize the network weights, use the BP algorithm to adjust the network through cross entropy loss, and the Adam algorithm is used as the optimizer; B. Use a learning rate of 0.001 for fast optimization in the first 100 cycles, and then use a learning rate of 0.0001 to make the iteration converge smoothly.

Images