Motor fault feature extraction and fault diagnosis method

Abstract

The invention discloses a motor fault feature extraction and fault diagnosis method. The method comprises the following steps: step 1, collecting an MCAVFCNN training data set, wherein motor fault types comprise two major classes, namely mechanical faults and electrical faults, namely 9 subclasses; 2, an MCAVFCNN model is built, a dynamic random erasure layer, a self-adaptive weighted sound and vibration fusion layer and a self-adaptive POOLING layer are provided, and the data set collected in the step 1 is used for model training till convergence; and step 3, the trained model is used online, and the trained MCAVFCNN model is applied online. According to the method, state monitoring and fault detection can be accurately carried out on the running motor in real time, and unnecessary loss is reduced while safe running of the motor is guaranteed.

Description

Technical Field

[0001] The invention relates to the field of motor state monitoring and fault identification, and in particular to a motor fault feature extraction and fault diagnosis method. Background Art

[0002] In the modern industrial field, motors, as key power equipment, are widely used in various production links, from production lines in large factories to daily household appliances. The reliability and stability of their operation are directly related to the normal operation of the entire system.

[0003] With the continuous improvement of industrial automation, the working environment of motors is becoming increasingly complex and diverse, facing the influence of different loads, temperatures, humidity, power grid fluctuations, etc. Traditional motor maintenance methods are mostly regular inspections, which has obvious disadvantages. It not only consumes a lot of manpower and material resources, but also makes it difficult to accurately predict motor failures, which easily leads to over-maintenance or under-maintenance.

[0004] In recent years, the rapid development of sensor technology, signal processing technology and artificial intelligence algorithms has brought new opportunities for motor status monitoring and fault diagnosis. More and more high-precision sensors are used to collect motor operating parameters, such as vibration sensors, temperature sensors, current sensors, etc., which can obtain multi-dimensional data of motors in real time. At the same time, fault diagnosis methods based on big data analysis and machine learning are gradually emerging. By learning and training a large amount of historical data and real-time monitoring data, intelligent diagnosis models are constructed, which can more accurately identify the type and degree of motor faults.

[0005] Effective motor status monitoring and fault diagnosis can detect potential faults in advance, arrange maintenance in time, avoid production interruptions caused by sudden failures, and reduce huge economic losses caused by downtime. For motors in some key production processes, such as chemical and electric power industries, the timeliness of fault diagnosis is even related to production safety and environmental safety. In addition, accurate status monitoring helps to optimize the operation and management of motors, improve energy efficiency, and extend the service life of motors. It is in line with the concept of high efficiency, energy saving, and sustainable development of modern industry, and is an indispensable part of promoting the process of industrial intelligent upgrading.

[0006] Through research on relevant literature and analysis of the current research status, it is found that there are two major problems in motor status monitoring and diagnosis: 1. The trained model is prone to overfitting and has poor generalization performance in practical applications. This is mainly because the characteristics of the training data are more obvious, and no disturbance is added during model training, resulting in poor generalization of the final model. 2. The trained model is difficult to comprehensively diagnose both mechanical and electrical faults of the motor. This is mainly limited by the data category of the training. Mechanical faults have obvious characteristics in vibration data, while electrical faults have obvious characteristics in sound signals. Most of the current training data are single vibration signals and sound signals. There are certain challenges in how to fuse vibration signals and sound signals. Summary of the invention

[0007] To solve the above problems, the present invention proposes a motor fault feature extraction and fault diagnosis method based on the Convolutional Neural Network (CNN). In response to problem 1, this patent proposes a dynamic random erasure algorithm, which adopts a random idea and pays attention to signals with obvious features. The dynamic erasure of acoustic vibration signals is achieved by designing calculation formulas and reasonably formulating algorithm thresholds, which brings randomness to the training of the model; at the same time, an adaptive POOLING layer is proposed, which enables the model to adaptively extract features for data with different features, and can also flexibly handle the diversity of data in online application scenarios, thereby improving the generalization of the model; in response to problem 2, this patent proposes an adaptive weighted acoustic vibration fusion algorithm, which introduces information entropy to calculate the weight coefficients of different channel data under different fault types, which can well highlight the characteristic signals and suppress interference signals, and finally complete the fusion of acoustic vibration data and improve the model's ability to diagnose various motor faults. To achieve this purpose, the present invention provides a motor fault feature extraction and fault diagnosis method, the specific steps are as follows, characterized in that:

[0008] Step 1: Collect the MCAVFCNN training data set, in which the motor fault types include two major categories of mechanical faults and electrical faults, with a total of 9 subcategories. Mechanical faults include local overheating, vibration faults, and air gap eccentricity faults; electrical faults include turn-to-turn short circuits, broken bars, end ring cracks, internal discharges, local insulation damage, and winding open circuits; then make corresponding category labels for the collected data.

[0009] Step 2: Build the MCAVFCNN model, which proposes a dynamic random erasing layer, an adaptive weighted acoustic vibration fusion layer, and an adaptive POOLING layer, and use the data set collected in step 1 to train the model until convergence;

[0010] Step 3: Use the trained model online. Apply the trained MCAVFCNN model online to perform real-time status monitoring and early fault identification on the running motor to reduce unnecessary losses.

[0011] As a further improvement of the present invention, in step 1, data set collection is performed, wherein the sensors of the collection system include an SSF-VIB-Z300 three-axis vibration sensor and an INV9206 high-precision ICP sound pressure sensor, wherein the three-axis vibration sensor is installed just above the motor drive end, and then a three-dimensional coordinate system is established with the circular plane where the three-axis sensor and the motor are located, the origin of the coordinate system coincides with the center of the motor rotor, the distance between the installation point of the sound pressure sensor and the origin of the coordinate system is the motor diameter, and the circular plane where the x-axis and y-axis are located coincides with the plane where the vibration sensor is located, and then the collection system is used to collect the acoustic vibration data of the x-axis, y-axis and z-axis under different fault states of the motor and make corresponding labels;

[0012] As a further improvement of the present invention, the specific steps of offline training of the MCAVFCNN model in step 2 are:

[0013] Step 2.1. A dynamic random erasure algorithm is proposed to erase the collected sound and vibration signals to improve the robustness of the model. The erasure rule is: if the data is selected for erasure, the value will be set to 0. A set of sampled data {x1, x2, ..., x N}, where N represents the dimension of the collected data, for the i-th data x i Perform erase calculation, where x i The corresponding erasure probability P(x i )The expression is designed as:

[0014]

[0015] Where r represents a random number that satisfies uniform distribution, that is, r~U[0,1], x max and xave Respectively represent the data {x1,x2,...,x N} maximum and average values.

[0016] Step 2.2. Use the FFT algorithm to calculate the signal processed in step 2.1, where the order of the acoustic vibration multi-channel data is: x-direction vibration signal, x-direction sound signal, y-direction vibration signal, y-direction sound signal, z-direction vibration signal and z-direction sound signal. After obtaining the 6-channel acoustic vibration signal, the proposed adaptive weighted acoustic vibration fusion algorithm is used to fuse the acoustic vibration signal. The fusion calculation process is as follows:

[0017] Step 2.2.1. Multi-channel acoustic vibration signal matrix (M) corresponding to the j-th type fault j , which can be expressed as:

[0018]

[0019] In the formula, q represents the total number of channels, and the value of q in this patent is 6.

[0020] Step 2.2.2. Calculate the information entropy of a single channel data, where the calculation formula is:

[0021]

[0022] In the formula, (E l ) k represents the information entropy corresponding to the collected data of the lth channel under the kth type of fault, (M il ) k Represents the i-th data point of the l-th channel of the k-th type fault.

[0023] Step 2.2.3. Considering the differences in the sensitivity of fault features in the three directions of X, Y and Z, a weighted weight algorithm is proposed to perform weighted processing on the data of each channel to highlight the signals with obvious features as much as possible and suppress the signals with interference. The specific calculation formula is as follows:

[0024]

[0025] In the formula, (w l ) k Represents the fusion weight corresponding to the lth channel data of the kth type of fault.

[0026] Step 2.2.4. Based on the weight coefficients corresponding to each channel, solve the fusion result of the k-th type of fault data (f) k , and its calculation formula is:

[0027]

[0028] Step 2.3. Build a convolutional neural network to extract and classify the data in step 2.2. The specific model architecture is: convolution layer 1 - adaptive POOLING layer 1 - convolution layer 2 - dropout layer - convolution layer 3 - adaptive POOLING layer 2 - fully connected layer - SOFTMAX layer. The algorithm calculation formula of the proposed adaptive POOLING layer is as follows:

[0029]

[0030] In the formula, (AC il ) krepresents the pooling value calculated based on the starting point of the ith data point of the lth channel of the kth fault, ξ represents the pooling adjustment factor, t represents the tth pooling kernel, v represents the convolution kernel length, and b represents the bias coefficient.

[0031] Among them, the calculation formula of the pooling adjustment factor ξ is:

[0032]

[0033] In the formula, η represents the characteristic coefficient, δ represents the correction coefficient, and its value range is (0,1), c and v max They respectively represent the average value calculated after removing the maximum value in the pooling domain where the convolution kernel is located and the maximum value in the pooling domain.

[0034] The expression of the characteristic coefficient η is as follows:

[0035]

[0036] where n epoch Indicates the number of rounds of network training.

[0037] Step 2.4. Use the cross entropy loss function and Adam parameter optimization algorithm to train the MCAVFCNN model proposed in this patent until the model converges, and the iteration termination condition is set to the loss function less than 0.001.

[0038] The present invention provides a motor fault feature extraction and fault diagnosis method, which collects acoustic vibration signals of different fault categories based on an acoustic vibration data acquisition system, and then builds a MCAVFCNN model, wherein a dynamic random erasure layer, an adaptive weighted acoustic vibration fusion layer and an adaptive POOLING layer are proposed, and through continuous training iterations, a MCAVFCNN model that can diagnose various motor faults and has excellent generalization performance is finally obtained. The overall beneficial effects are:

[0039] 1. The dynamic random erasing algorithm proposed in this patent adopts random ideas and focuses on signals with obvious features. It realizes dynamic erasing of acoustic vibration signals by designing calculation formulas and reasonably formulating algorithm thresholds, which brings randomness to the training of the model. At the same time, an adaptive POOLING layer is proposed, which enables the model to adaptively extract features for data with different features, and can also flexibly handle the diversity of data in online application scenarios, thereby improving the generalization of the model;

[0040] 2. The adaptive weighted acoustic-vibration fusion algorithm proposed in this patent introduces information entropy to calculate the weight coefficients of different channel data under different fault types. This can well highlight the characteristic signals and suppress the interference signals, and finally complete the fusion of acoustic-vibration data and improve the model's ability to diagnose various motor faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the sensor layout of the acoustic vibration data acquisition system in this patent;

[0042] Figure 2 is a flow chart of the present invention;

[0043] Figure 3 This is a schematic diagram of weighted fusion of acoustic vibration signals;

[0044] Figure 4 This is the network architecture diagram of the MCAVFCNN model in this patent. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0046] The present invention proposes a motor fault feature extraction and fault diagnosis method, which aims to perform real-time and accurate state monitoring and fault detection on a running motor, thereby ensuring the safe operation of the motor and reducing unnecessary losses.

[0047] Figure 1 This is a schematic diagram of the sensor layout of the acoustic vibration data acquisition system in this patent, where the sensors of the acquisition system include the SSF-VIB-Z300 three-axis vibration sensor and the INV9206 high-precision ICP sound pressure sensor. The three-axis vibration sensor is installed just above the motor drive end, and then a three-dimensional coordinate system is established with the circular plane where the three-axis sensor and the motor are located. The origin of the coordinate system coincides with the center of the motor rotor, the distance between the installation point of the sound pressure sensor and the coordinate origin is the motor diameter, and the circular plane where the x-axis and y-axis are located coincides with the plane where the vibration sensor is located. Then, the acquisition system is used to collect the acoustic vibration data of the x-axis, y-axis and z-axis under different fault states of the motor and make corresponding labels.

[0048] Figure 2 It is a flow chart of the present invention. It can be seen from the figure that the specific implementation steps of the method are:

[0049] Step 1: Collect the MCAVFCNN training data set, in which the motor fault types include two major categories of mechanical faults and electrical faults, with a total of 9 subcategories. Mechanical faults include local overheating, vibration faults, and air gap eccentricity faults; electrical faults include turn-to-turn short circuits, broken bars, end ring cracks, internal discharges, local insulation damage, and winding open circuits; then make corresponding category labels for the collected data.

[0050] Step 2: Build the MCAVFCNN model, which proposes a dynamic random erasing layer, an adaptive weighted acoustic vibration fusion layer, and an adaptive POOLING layer, and use the data set collected in step 1 to train the model until convergence;

[0051] Step 3: Use the trained model online. Apply the trained MCAVFCNN model online to perform real-time status monitoring and early fault identification on the running motor to reduce unnecessary losses.

[0052] The specific steps of offline training of the MCAVFCNN model in step 2 are:

[0053] Step 2.1. A dynamic random erasure algorithm is proposed to erase the collected sound and vibration signals to improve the robustness of the model. The erasure rule is: if the data is selected for erasure, the value will be set to 0. A set of sampled data {x1, x2, ..., x N}, where N represents the dimension of the collected data, for the i-th data x i Perform erase calculation, where x i The corresponding erasure probability P(x i )The expression is designed as:

[0054]

[0055] Where r represents a random number that satisfies uniform distribution, that is, r~U[0,1], x max and xave Respectively represent the data {x1,x2,...,x N} maximum and average values.

[0056] Step 2.2. Use the FFT algorithm to calculate the signal processed in step 2.1, where the order of the acoustic vibration multi-channel data is: x-direction vibration signal, x-direction sound signal, y-direction vibration signal, y-direction sound signal, z-direction vibration signal and z-direction sound signal. After obtaining the 6-channel acoustic vibration signal, the proposed adaptive weighted acoustic vibration fusion algorithm is used to fuse the acoustic vibration signal. The fusion calculation process is as follows:

[0057] Step 2.2.1. Multi-channel acoustic vibration signal matrix (M) corresponding to the j-th type fault j , which can be expressed as:

[0058]

[0059] In the formula, q represents the total number of channels, and the value of q in this patent is 6.

[0060] Step 2.2.2. Calculate the information entropy of a single channel data, where the calculation formula is:

[0061]

[0062] In the formula, (E l ) krepresents the information entropy corresponding to the collected data of the lth channel under the kth type of fault, (M il ) k Represents the i-th data point of the l-th channel of the k-th type fault.

[0063] Step 2.2.3. Considering the differences in the sensitivity of fault features in the three directions of X, Y and Z, a weighted weight algorithm is proposed to perform weighted processing on the data of each channel to highlight the signals with obvious features as much as possible and suppress the signals with interference. The specific calculation formula is as follows:

[0064]

[0065] In the formula, (w l ) k Represents the fusion weight corresponding to the lth channel data of the kth type of fault.

[0066] Step 2.2.4. Based on the weight coefficients corresponding to each channel, solve the fusion result of the k-th type of fault data (f) k , and its calculation formula is:

[0067]

[0068] Step 2.3. Build a convolutional neural network to extract and classify the data in step 2.2. The specific model architecture is: convolution layer 1 - adaptive POOLING layer 1 - convolution layer 2 - dropout layer - convolution layer 3 - adaptive POOLING layer 2 - fully connected layer - SOFTMAX layer. The algorithm calculation formula of the proposed adaptive POOLING layer is as follows:

[0069]

[0070] In the formula, (AC il ) k represents the pooling value calculated based on the starting point of the ith data point of the lth channel of the kth fault, ξ represents the pooling adjustment factor, t represents the tth pooling kernel, v represents the convolution kernel length, and b represents the bias coefficient.

[0071] Among them, the calculation formula of the pooling adjustment factor ξ is:

[0072]

[0073] In the formula, η represents the characteristic coefficient, δ represents the correction coefficient, and its value range is (0,1), c and v max They respectively represent the average value calculated after removing the maximum value in the pooling domain where the convolution kernel is located and the maximum value in the pooling domain.

[0074] The expression of the characteristic coefficient η is as follows:

[0075]

[0076] where n epoch Indicates the number of rounds of network training.

[0077] Step 2.4. Use the cross entropy loss function and Adam parameter optimization algorithm to train the MCAVFCNN model proposed in this patent until the model converges, and the iteration termination condition is set to the loss function less than 0.001.

[0078] Figure 3 This is a schematic diagram of the weighted fusion of acoustic and vibration signals. It can be clearly seen that for data from different channels, the fusion weighted coefficient is calculated using information entropy to finally obtain the fused data. Through this operation, the characteristics of sound and vibration signals are fused without increasing the data dimension, which greatly enhances the fault characteristics of the training data.

[0079] Figure 4 This is the architecture diagram of MCAVFCNN in this patent. It can be seen from the figure that the specific network architecture is: acoustic vibration data acquisition layer - dynamic random erasing layer - FFT data processing layer - adaptive weighted acoustic vibration fusion layer - convolution layer 1 - adaptive POOLING layer 1 - convolution layer 2 - dropout layer - convolution layer 3 - adaptive POOLING layer 2 - fully connected layer - SOFTMAX layer. It can be seen that the collected acoustic vibration data, after dynamic random erasing and FFT transformation, uses the adaptive weighted fusion algorithm to fuse features, and then extracts features through the convolution layer and pooling layer of the convolutional neural network, and incorporates adaptive ideas in the extraction process. Finally, the recognition and classification of different motor faults are realized through the SOFTMAX layer. Through the adjustment of the above model architecture, the classification ability and generalization performance of the training model are improved.

Claims

1. A method for extracting motor fault features and diagnosing faults, characterized in that: Step 1: Collect the MCAVFCNN training data set, where the motor fault types include two categories of mechanical faults and electrical faults, with a total of 9 subcategories. Mechanical faults include local overheating, vibration faults, and air gap eccentricity faults; electrical faults include turn-to-turn short circuits, broken bars, end ring cracks, internal discharges, local insulation damage, and winding open circuits; then make corresponding category labels for the collected data; Step 2: Build the MCAVFCNN model, which proposes a dynamic random erasing layer, an adaptive weighted acoustic vibration fusion layer, and an adaptive POOLING layer, and use the data set collected in step 1 to train the model until convergence; Step 3: Use the trained model online. Apply the trained MCAVFCNN model online to perform real-time status monitoring and early fault identification on the running motor to reduce unnecessary losses.

2. According to the method for extracting and diagnosing motor fault features and faults according to claim 1, the algorithm model MCAVFCNN is characterized in that: in step 1, data sets are collected, wherein the sensors of the collection system include SSF-VIB-Z300 three-axis vibration sensor and INV9206 high-precision ICP sound pressure sensor, wherein the three-axis vibration sensor is installed just above the motor drive end, and then a three-dimensional coordinate system is established with the circular plane where the three-axis sensor and the motor are located, the origin of the coordinate system coincides with the center of the motor rotor, the distance between the installation point of the sound pressure sensor and the coordinate origin is the motor diameter, and the circular plane where the x-axis and y-axis are located coincides with the plane where the vibration sensor is located, and then the collection system is used to collect the acoustic vibration data of the x-axis, y-axis and z-axis under different fault states of the motor and make corresponding labels.

3. According to the method for extracting motor fault features and diagnosing faults in claim 1, the algorithm model MCAVFCNN is characterized in that: the specific steps of offline training of the MCAVFCNN model in step 2 are: Step 2.

1. A dynamic random erasure algorithm is proposed to erase the collected sound and vibration signals to improve the robustness of the model. The erasure rule is: if the data is selected for erasure, the value will be set to 0. A set of sampled data {x1, x2, ..., x N }, where N represents the dimension of the collected data, for the i-th data x i Perform erase calculation, where x i The corresponding erasure probability P(x i )The expression is designed as: Where r represents a random number that satisfies uniform distribution, that is, r~U[0,1], x max and xave Respectively represent the data {x1,x2,...,x N } maximum and average values. Step 2.

2. Use the FFT algorithm to calculate the signal processed in step 2.1, where the order of the acoustic vibration multi-channel data is: x-direction vibration signal, x-direction sound signal, y-direction vibration signal, y-direction sound signal, z-direction vibration signal and z-direction sound signal. After obtaining the 6-channel acoustic vibration signal, the proposed adaptive weighted acoustic vibration fusion algorithm is used to fuse the acoustic vibration signal. The fusion calculation process is as follows: Step 2.2.

1. Multi-channel acoustic signal matrix (M) corresponding to the j-th type fault j , which can be expressed as: In the formula, q represents the total number of channels, and the value of q in this patent is 6. Step 2.2.

2. Calculate the information entropy of a single channel data, where the calculation formula is: In the formula, (E l ) k represents the information entropy corresponding to the collected data of the lth channel under the kth type of fault, (M il ) k Represents the i-th data point of the l-th channel of the k-th type fault. Step 2.2.

3. Considering the differences in the sensitivity of fault features in the three directions of X, Y and Z, a weighted weight algorithm is proposed to perform weighted processing on the data of each channel to highlight the signals with obvious features as much as possible and suppress the signals with interference. The specific calculation formula is as follows: In the formula, (w l ) k Represents the fusion weight corresponding to the lth channel data of the kth type of fault. Step 2.2.

4. Based on the weight coefficients corresponding to each channel, solve the fusion result of the k-th type of fault data (f) k , and its calculation formula is: Step 2.

3. Build a convolutional neural network to extract and classify the data in step 2.

2. The specific model architecture is: convolution layer 1 - adaptive POOLING layer 1 - convolution layer 2 - dropout layer - convolution layer 3 - adaptive POOLING layer 2 - fully connected layer - SOFTMAX layer. The algorithm calculation formula of the proposed adaptive POOLING layer is as follows: In the formula, (AC il ) k represents the pooling value calculated based on the starting point of the ith data point of the lth channel of the kth fault, ξ represents the pooling adjustment factor, t represents the tth pooling kernel, v represents the convolution kernel length, and b represents the bias coefficient. in, The calculation formula of the pooling adjustment factor ξ is: In the formula, η represents the characteristic coefficient, δ represents the correction coefficient, and its value range is (0,1), c and v max They respectively represent the average value calculated after removing the maximum value in the pooling domain where the convolution kernel is located and the maximum value in the pooling domain. The expression of the characteristic coefficient η is as follows: where n epoch Indicates the number of rounds of network training. Step 2.

4. Use the cross entropy loss function and Adam parameter optimization algorithm to train the MCAVFCNN model proposed in this patent until the model converges, and the iteration termination condition is set to the loss function less than 0.001.

Images