Adaptive feature weighted fusion fault diagnosis method for multi-source heterogeneous data of motor

Abstract

The invention relates to an adaptive feature weighted fusion fault diagnosis method for multi-source heterogeneous data of a motor, and the method comprises the steps: constructing a multi-head deep convolutional neural network, designing a plurality of parallel convolutional channels in the neural network, deeply mining the abnormal features of original signals of current and vibration, and focusing on the signal features in different dimensions, thereby achieving the fault diagnosis of the multi-source heterogeneous data of the motor. And comprehensive capture of key feature information is ensured. And adaptive weighted fusion is performed on the extracted multi-dimensional features by adopting a correlation analysis method to highlight abnormal key features, so that the feature fusion effect of the multi-source heterogeneous data and the accuracy of anomaly detection are improved. An SELU activation function is adopted and a hyper-parameter optimization strategy of a dynamic attenuation learning rate is combined so that an overfitting problem which is easy to occur in a model training process can be effectively solved and convergence stability of the model and high precision of fault identification can be enhanced. Through adaptive fusion of multi-source heterogeneous data features, key fault features are effectively extracted, and the diagnosis accuracy in an electromechanical coupling fault scene is improved.

Description

Technical Field

[0001] The present invention relates to the field of motor fault diagnosis, and in particular to a multi-source heterogeneous data adaptive feature weighted fusion method for improving the accuracy of compound fault diagnosis of a motor in a multi-fault coupling scenario. Background Art

[0002] With the rapid development of industrial automation, the operation status of motors, as core power equipment, directly affects the stability and safety of production. However, in the actual working environment, motors face inevitable stresses such as mechanical, electrical, thermal and environmental stresses due to long-term operation and external load fluctuations, power supply fluctuations, overheating, insufficient lubrication, poor sealing, product defects and natural aging. These stresses may eventually cause motor failures, which will not only interfere with the normal operation of the equipment, but also reduce the economic benefits of the enterprise. In addition, motor failures may also cause safety hazards and even lead to accidents such as personal injury. Therefore, it is particularly important to ensure the continuous and stable operation of the motor.

[0003] Traditional motor maintenance mainly relies on manual experience and technical processes. Technicians usually judge the operating status and fault conditions of the motor by visual inspection, listening or measurement. However, this method has lag. During the operation of the motor, wear or damage of mechanical parts can cause abnormal vibration or current fluctuations. These abnormal signals are usually early warning signs of potential faults. By analyzing the vibration and current signals of the motor, condition-based maintenance (CBM) fault diagnosis can effectively identify mechanical or electrical faults of the motor. For example, through Fourier transform, the vibration signal can be decomposed into different frequency components, from which frequency peaks related to specific fault types can be found. Using time domain or frequency domain analysis methods, fault features can be extracted from the current signal, and then the fault type can be accurately determined to achieve early diagnosis of the motor.

[0004] However, under actual working conditions, motors often have multiple fault coupling composite faults due to long-term operation under complex and harsh working conditions. Such faults usually involve multiple independent or interrelated fault points, and each fault point may show different characteristic information in different types of signals. The intelligent diagnosis method using a single signal may miss the key characteristic information contained in other signals, resulting in low diagnostic accuracy. In order to solve this problem, the present invention proposes an intelligent diagnosis method based on adaptive fusion and deep learning of multi-source heterogeneous data. This method makes full use of deep neural networks to deeply explore the fault features in the original signal. At the same time, based on the differences in the features of different signals, it adaptively fuses them, so that the fused feature information is richer and covers more comprehensive fault features, thereby optimizing the combination of multi-source data. Through this technical means, the accuracy and robustness of composite fault diagnosis of motors have been significantly improved, showing a broad application prospect. Summary of the invention

[0005] In view of the above problems, the present invention discloses a motor multi-source heterogeneous data adaptive feature weighted fusion fault diagnosis method. The purpose of the present invention is to provide a method for constructing a deep convolutional neural network model (Multi-Source Adaptive Feature Fusion Deep Convolutional Neural Network, MSAFF-DCNN) based on multi-source heterogeneous data adaptive feature weighted fusion, so as to solve the motor compound fault classification and diagnosis problems under complex and harsh working conditions.

[0006] An adaptive feature weighted fusion fault diagnosis method for motor multi-source heterogeneous data, the method comprising the following steps:

[0007] S1: Collect data;

[0008] S2: data preprocessing;

[0009] S3: Construct a network model for motor composite fault diagnosis;

[0010] S4: Optimize and design network hyperparameters;

[0011] S5: Evaluate model performance.

[0012] Furthermore, step S1 is specifically as follows:

[0013] Collect current signals and vibration acceleration signals of historical healthy motors and actual faulty motors.

[0014] S1-1: Use the vibration sensor to collect the vibration signal y0(t) when the healthy motor is running and the vibration signal y when the faulty motor is running i(t) Therefore, the current clamp is used to collect the current signal x0(t) when the healthy motor is running and the current signal x0(t) when the faulty motor is running. i (t);

[0015] Furthermore, step S2 is specifically as follows:

[0016] S2-1: pre-processing healthy vibration acceleration signal, current signal, fault vibration acceleration signal, current signal;

[0017] The preprocessing method is wavelet processing or FIR filter, and the purpose of data preprocessing is to reduce noise in the signal.

[0018] Furthermore, step S3 is specifically as follows:

[0019] The specific construction steps are as follows:

[0020] S3-1: Building a multi-head convolutional feature extraction module

[0021] S3-1-1: Construct current signal feature extraction channel:

[0022] The input current signal inputs1 is used, and the one-dimensional convolution layer Conv1D is used for feature extraction.

[0023] Multiple convolution blocks are stacked, including convolutional layers and maximum pooling layers MaxPooling1D. The convolution operation is expressed by the following formula:

[0024] x1=MaxPooling(Activation(Conv1D(x)))

[0025] In the formula, x represents the input current signal, and x1 represents the current signal after the convolution operation.

[0026] S3-1-2: Construct vibration signal feature extraction channel:

[0027] Input the vibration signal inputs2, and also use the one-dimensional convolution layer Conv1D for feature extraction.

[0028] Multiple convolution blocks are stacked, including convolutional layers and maximum pooling layers MaxPooling1D. The convolution operation is expressed by the following formula:

[0029] y1=MaxPooling(Activation(Conv1D(y)))

[0030] Where y represents the input vibration signal, and y1 represents the vibration signal after convolution operation.

[0031] The multi-head feature extraction module finally extracts the features of the two input signals into two feature tensors: x1 and y1 for subsequent feature weighted fusion.

[0032] S3-2: Constructing feature weighting module

[0033] S3-2-1: Calculate the Pearson correlation coefficient

[0034] Formula (1) is used to calculate the running current signal x0(t) of the healthy motor and the current signal x0(t) of the different fault motors. i Pearson correlation coefficient of (t);

[0035]

[0036] In formula (1), current_corr i Represents the healthy current signal x0(t) and different fault current signals x i The Pearson correlation coefficient between (t). Its value range is between -1 and 1. x0(n) represents the value of the healthy motor current signal x0 at the nth time point. i (n) represents the value of the current signal of the i-th fault motor at the n-th time point. Indicates signal N represents the number of samples of the current signal, that is, the length of the signal.

[0037] S3-2-2: Calculate the cross-correlation coefficient

[0038] Formula (2) is used to calculate the vibration signal y0(t) of the healthy motor and the vibration signal y0(t) of the different fault motors. i (t) the mutual correlation coefficient;

[0039]

[0040] In formula (2), Represents the vibration signal y0(t) of a healthy motor and the vibration signal y0(t) of a faulty motor i (t) is the cross-correlation function at time delay k, indicating their similarity at different time offsets. y0(n) represents the value of the healthy motor vibration signal y0(t) at the nth time point. i (n+k) represents the fault motor vibration signal y i (t) is the value at the n+kth time point. N represents the number of samples of the motor vibration signal, that is, the length of the signal. k represents the delay, which represents the time offset of one signal relative to another signal. For calculating the mutual correlation coefficient, k=0 is usually taken to represent the similarity without delay.

[0041] S3-2-3: Normalization of the cross-correlation coefficient

[0042] The mutual correlation coefficient is obtained by normalizing using formula (3);

[0043]

[0044] In formula (3), and It represents the cross-correlation value between the vibration signal of a healthy motor and the vibration signal of a faulty motor. i The value range of is -1 to 1, indicating the linear correlation between the two vibration signals.

[0045] S3-2-4: Calculate current signal weight

[0046] According to the correlation coefficient, the weight of the motor current signal in the feature fusion under the i-th fault can be obtained by normalizing it. The formula (1), (3), (4) is used to calculate w(x i );

[0047]

[0048] S3-2-5: Calculate vibration signal weight

[0049] According to the correlation coefficient, the weight of the motor current signal in the feature fusion under the i-th fault can be obtained by normalizing it. The formula (1)(3)(5) is used to calculate w(y i );

[0050]

[0051] S3-2-6: Fusion weighted features

[0052] The weighted features are fused using the K.concatenate function through the Lambda layer to complete the weighted fusion of features. K.concatenate can be expressed as:

[0053] Concat(X1,X2,...,X n ,axis)=Y

[0054] where X1,X2,...,X n Represents the weighted feature tensor, axis represents concatenation on the axis-th dimension, and Y represents the output tensor.

[0055] S3-3: Building a fault diagnosis module

[0056] The fault diagnosis of the motor is completed by using the weighted fusion features of S3-2-6. The fault diagnosis is performed on the trained model using the test set, the data in the test set is input into the model, and the output of the model is observed. The accuracy and performance of the model are evaluated by comparing the results predicted by the model with the actual labels. The classification results of each type of fault are obtained by SoftMax. Among them, SoftMax can be expressed as:

[0057]

[0058] Where C refers to the classification category, f(u) refers to the classification output, and exp refers to the exponential function. This allows the diagnosis and classification of motor compound faults to be realized.

[0059] Further, step S4 is specifically as follows:

[0060] S4-1: Adjust the dynamic learning decay learning rate

[0061] Adjusting the dynamic decay learning rate can improve the stability of model training and effectively avoid model crash. The dynamic decay learning rate decays exponentially. As the training progresses, the learning rate gradually decreases. The formula is:

[0062]

[0063] Where lr0 represents the initial learning efficiency, decay factor represents the decay factor, and step represents the cycle length.

[0064] S4-2 Apply Selu activation function

[0065] The formula and derivative formula of Selu activation function are as follows:

[0066]

[0067] Where λ represents the scaling factor and α represents the coefficient for adjusting the negative part.

[0068] Furthermore, step S5 is specifically as follows:

[0069] The trained network model was tested in the test set, and the comparative experiments verified that the multi-source heterogeneous data adaptive feature weighted fusion method (MSAFF-DCNN) model has a higher diagnostic accuracy in diagnosing complex faults, while the DCNN model based on the current single source data source has insufficient diagnostic accuracy in multiple faults.

[0070] Compared with the prior art, the present invention has the following effects:

[0071] The basic design idea of ​​the deep convolutional neural network model based on adaptive weighted fusion of multi-source heterogeneous data features: In this method, vibration signals are collected by acceleration sensors, and current data are obtained by current clamps, and these multimodal signals are used to diagnose motor faults. First, the deep convolutional neural network architecture is designed to extract the features of current signals and vibration signals respectively through multi-head convolution to capture the abnormal mutation feature information in different signals. Secondly, in order to strengthen and highlight the feature information in different signals, the correlation coefficient between different signals is calculated, and the feature weights are adaptively allocated based on the correlation coefficient. Thirdly, in the weighted fusion layer, the weighted current signal features are fused with the vibration signal features, which can not only effectively remove redundant information, but also highlight the key features that are most helpful in identifying abnormalities, so as to ensure the complementarity of multi-source information and the best fusion method. Finally, the fused feature vector is used for fault diagnosis, and the final fault classification result is output through the SoftMax classifier. This method clarifies the basis for weight allocation, and can dynamically adjust the allocation method of feature weights according to the feature changes of the motor under different operating conditions, so that the fused features are closer to the actual operating conditions. This method significantly enhances the robustness and precision of diagnosis, effectively improves the accuracy of complex fault diagnosis, and overcomes the limitations of single signal intelligent diagnosis methods in complex fault scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a flow chart of an adaptive feature weighted fusion fault diagnosis method for multi-source heterogeneous data of a motor according to the present invention.

[0073] Figure 2 This is a schematic diagram of the structure of a deep convolutional neural network model based on weighted fusion of adaptive features of multi-source heterogeneous data in the present invention.

[0074] Figure 3 It is the diagnosis result based on multi-source heterogeneous data adaptive feature weighted fusion method (MSAFF-DCNN).

[0075] Figure 4 It is the DCNN diagnosis result based on the current single source data source. DETAILED DESCRIPTION

[0076] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] An adaptive feature weighted fusion fault diagnosis method for motor multi-source heterogeneous data, the method comprising the following steps:

[0078] S1: Collecting Data

[0079] Collect current signals and vibration acceleration signals of historical healthy motors and actual faulty motors.

[0080] S1-1: If Figure 1 As shown in the figure, the vibration sensor is used to collect the vibration signal y0(t) when the healthy motor is running and the vibration signal y0(t) when the faulty motor is running. i (t) Therefore, the current clamp is used to collect the current signal x0(t) when the healthy motor is running and the current signal x0(t) when the faulty motor is running. i (t);

[0081] S2: Data preprocessing

[0082] S2-1: pre-processing healthy vibration acceleration signal, current signal, fault vibration acceleration signal, current signal;

[0083] The preprocessing method is wavelet processing or FIR filter, and the purpose of data preprocessing is to reduce noise in the signal.

[0084] S3: Constructing a network model for motor complex fault diagnosis

[0085] The present invention proposes a deep neural network model based on adaptive weighted fusion of multi-source heterogeneous data (Multi-Source Adaptive Feature Fusion Deep Convolutional Neural Network, MSAFF-DCNN). The model aims to adaptively weight the fusion of multi-source data features, so that the fused feature information is richer, covers more comprehensive fault features, and realizes the optimal combination of multi-source data. The construction of the model includes three core modules: multi-head convolution feature extraction module, feature weighted fusion module, and data set classification module.

[0086] The specific construction steps are as follows:

[0087] S3-1: Building a multi-head convolutional feature extraction module

[0088] The multi-head convolution feature layer captures the features of the input vibration and current signals at multiple scales and dimensions by applying convolution kernels of different sizes. Each convolution head extracts different types of pattern information, which can provide richer representation capabilities for subsequent feature processing after splicing. Figure 1 As shown in the figure, this structure is suitable for data types with multiple feature patterns. Through multi-scale analysis, it can improve the model's learning ability and robustness for complex data. The convolutional layer of each channel is constructed as follows:

[0089] S3-1-1: Construct current signal feature extraction channel:

[0090] The input current signal inputs1 is used, and the one-dimensional convolution layer Conv1D is used for feature extraction.

[0091] Multiple convolution blocks are stacked, including convolutional layers and maximum pooling layers MaxPooling1D. The convolution operation is expressed by the following formula:

[0092] x1=MaxPooling(Activation(Conv1D(x)))

[0093] In the formula, x represents the input current signal, and x1 represents the current signal after the convolution operation.

[0094] S3-1-2: Construct vibration signal feature extraction channel:

[0095] Input the vibration signal inputs2, and also use the one-dimensional convolution layer Conv1D for feature extraction.

[0096] Multiple convolution blocks are stacked, including convolutional layers and maximum pooling layers MaxPooling1D. The convolution operation is expressed by the following formula:

[0097] y1=MaxPooling(Activation(Conv1D(y)))

[0098] Where y represents the input vibration signal, and y1 represents the vibration signal after convolution operation.

[0099] The multi-head feature extraction module finally extracts the features of the two input signals into two feature tensors: x1 and y1 for subsequent feature weighted fusion.

[0100] S3-2: Constructing feature weighting module

[0101] When the motor is in different fault states, the abnormal features contained in the vibration signal and the current signal are inconsistent. In order to capture more comprehensive fault information, it is necessary to assign different weights according to the importance of the fault features contained in different signals, so as to emphasize important features and suppress unimportant features, thereby ensuring that the final fusion features contain more important fault features. For this purpose, an adaptive weight assignment method based on correlation analysis is introduced. Figure 2 As shown in the weighted fusion of features, the specific process is as follows:

[0102] S3-2-1: Calculate the Pearson correlation coefficient

[0103] Formula (1) is used to calculate the running current signal x0(t) of the healthy motor and the current signal x0(t) of the different fault motors.i Pearson correlation coefficient of (t);

[0104]

[0105] In formula (1), current_corr i Represents the healthy current signal x0(t) and different fault current signals x i The Pearson correlation coefficient between (t). Its value range is between -1 and 1. x0(n) represents the value of the healthy motor current signal x0 at the nth time point. i (n) represents the value of the current signal of the i-th fault motor at the n-th time point. Indicates signal N represents the number of samples of the current signal, that is, the length of the signal.

[0106] S3-2-2: Calculate the cross-correlation coefficient

[0107] Formula (2) is used to calculate the vibration signal y0(t) of the healthy motor and the vibration signal y0(t) of the different fault motors. i (t) the mutual correlation coefficient;

[0108]

[0109] In formula (2), Represents the vibration signal y0(t) of a healthy motor and the vibration signal y0(t) of a faulty motor i (t) is the cross-correlation function at time delay k, indicating their similarity at different time offsets. y0(n) represents the value of the healthy motor vibration signal y0(t) at the nth time point. i (n+k) represents the fault motor vibration signal y i (t) is the value at the n+kth time point. N represents the number of samples of the motor vibration signal, that is, the length of the signal. k represents the delay, which represents the time offset of one signal relative to another signal. For calculating the mutual correlation coefficient, k=0 is usually taken to represent the similarity without delay.

[0110] S3-2-3: Normalization of the cross-correlation coefficient

[0111] In order to prevent the correlation coefficient from being unstable due to being too large or too small, normalization is required. The correlation coefficient is obtained by normalization using formula (3);

[0112]

[0113] In formula (3), and It represents the cross-correlation value between the vibration signal of a healthy motor and the vibration signal of a faulty motor. i The value range of is -1 to 1, indicating the linear correlation between the two vibration signals.

[0114] S3-2-4: Calculate current signal weight

[0115] According to the correlation coefficient, the weight of the motor current signal in the feature fusion under the i-th fault can be obtained by normalizing it. The formula (1), (3), (4) is used to calculate w(x i );

[0116]

[0117] S3-2-5: Calculate vibration signal weight

[0118] According to the correlation coefficient, the weight of the motor current signal in the feature fusion under the i-th fault can be obtained by normalizing it. The formula (1)(3)(5) is used to calculate w(y i );

[0119]

[0120] S3-2-6: Fusion weighted features

[0121] The weighted features are fused using the K.concatenate function through the Lambda layer to complete the weighted fusion of features. Figure 2 As shown in the weighted fusion layer, K.concatenate can be expressed as:

[0122] Concat(X1,X2,...,X n ,axis)=Y

[0123] where X1,X2,...,X n Represents the weighted feature tensor, axis represents concatenation on the axis-th dimension, and Y represents the output tensor.

[0124] S3-3: Building a fault diagnosis module

[0125] like Figure 2 As shown in the figure, the S3-2-6 weighted fusion feature is used to complete the fault diagnosis of the motor. The test set is used to perform fault diagnosis on the trained model, the data in the test set is input into the model, and the output of the model is observed. The accuracy and performance of the model are evaluated by comparing the results predicted by the model with the actual labels. And the classification results of each type of fault are obtained through SoftMax. Among them, SoftMax can be expressed as:

[0126]

[0127] Where C refers to the classification category, f(u) refers to the classification output, and exp refers to the exponential function. This allows the diagnosis and classification of motor compound faults to be realized.

[0128] S4: Optimize network hyperparameters

[0129] S4-1: Adjust the dynamic learning decay learning rate

[0130] Adjusting the dynamic decay learning rate can improve the stability of model training and effectively avoid model crash. The dynamic decay learning rate decays exponentially. As the training progresses, the learning rate gradually decreases. The formula is:

[0131]

[0132] Where lr0 represents the initial learning efficiency, decay factor represents the decay factor, and step represents the cycle length.

[0133] S4-2 Apply Selu activation function

[0134] The activation function can complete the nonlinear mapping of data, which is a nonlinear transformation. Unlike the insufficient ability of linear operation models to express features, it can improve the ability of neural networks to learn complex data. In the process of back propagation of errors to update parameters, continuous multiplication of activation function derivatives will appear, which is prone to gradient vanishing and gradient explosion problems. The activation function directly affects the convergence problem in the gradient descent process, and also directly affects the parameter update and the learning efficiency of the network. The Selu activation function effectively controls the distribution of output values ​​in deep neural networks by introducing λ and α, avoiding the situation where the activation value deviates from zero, which is conducive to more stable training and prevents the model from falling into local optimality. The formula and derivative formula of the Selu activation function are as follows:

[0135]

[0136] Where λ represents the scaling factor and α represents the coefficient for adjusting the negative part.

[0137] S5: Evaluate model performance

[0138] like Figure 3 , Figure 4 As shown in the figure, the trained network model is tested in the test set. The comparative experiments verify that the multi-source heterogeneous data adaptive feature weighted fusion method (MSAFF-DCNN) model has higher diagnostic accuracy in diagnosing complex faults, while the DCNN model based on the current single source data source has insufficient diagnostic accuracy in multiple faults.

[0139] The present invention takes adaptive weight allocation under different operating states as design factors, takes fusion features as optimization goals, and uses correlation analysis and normalization processing methods to calculate the weights of vibration signals and current signals under different operating states. The features of current signals and vibration signals are extracted respectively through multi-head convolution, combined with the calculated signal weights, and through the weighted fusion layer, the features of current and vibration signals are adaptively fused to remove redundant information and highlight the key features that are most helpful in identifying abnormalities, so as to ensure the complementarity of multi-source information and the best fusion method. Finally, the fused features are used for fault diagnosis, and SoftMax is used to obtain accurate classification results.

[0140] The present invention can calculate different signal weights according to different types of motor faults, so that the fused features are more in line with the actual fault conditions, the diagnosis results are more reliable, and the diagnostic accuracy is significantly improved. This solves the problem of insufficient accuracy of the single signal fault intelligent diagnosis method in composite fault diagnosis.

[0141] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. An adaptive feature weighted fusion fault diagnosis method for multi-source heterogeneous data of electric motors, characterized in that: The method comprises the following steps: S1: Collect data; S2: data preprocessing; S3: Construct a network model for motor composite fault diagnosis; S4: Optimize and design network hyperparameters; S5: Evaluate model performance.

Images