Induction motor fault diagnosis method based on multi-source signal fusion deep learning

Through the multi-source signal fusion deep learning method, combining convolutional network and multi-source data attention mechanism, and fusing current and vibration signals, the problem of insufficient accuracy and sensitivity of traditional induction motor fault diagnosis methods is solved, achieving more accurate and efficient Troubleshooting.

CN118940216APending Publication Date: 2024-11-12JIANGSU UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411029913.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Traditional induction motor fault diagnosis methods rely on regular manual inspections and single parameter monitoring, and cannot reflect the motor operating status in real time, making it difficult to predict potential faults, resulting in frequent unplanned downtime, and insufficient accuracy and sensitivity of single signal diagnosis, which cannot meet intelligent diagnosis. need.

Method used

Using a deep learning method based on multi-source signal fusion, the current signal and vibration signal are fused through convolutional networks and multi-source data attention mechanisms, more comprehensive fault characteristics are extracted, and the memory neural network is used to capture timing dependencies to achieve more accurate Troubleshooting.

Benefits of technology

It improves the accuracy of identifying various fault modes, avoids information redundancy, enhances the generalization ability of the model, and achieves more accurate and efficient induction motor fault diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940216A_ABST
    Figure CN118940216A_ABST
Patent Text Reader

Abstract

The invention discloses an induction motor fault diagnosis method based on multi-source signal fusion deep learning, and the method comprises the following steps: A), collecting a current signal and a vibration signal of an induction motor of a known fault type, marking the current signal and the vibration signal of the motor according to the type, and obtaining a training sample; step B), dividing the collected data set into a training set, a test set and a verification set, wherein the proportions are 70%, 20% and 10% respectively; and step C), building a network model based on a convolutional network and multi-source signal fusion, and carrying out fusion and feature extraction on the current and vibration signals. And D) training the network by using the training set until the model converges, and storing model parameters. And loading the trained model parameters, and performing motor fault diagnosis by using the test set. According to the method, priori knowledge is not needed, one-dimensional original current signals and vibration signals can be fully utilized, hidden information can be mined, and more accurate signal features can be obtained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and in particular to an induction motor fault diagnosis technology, and in particular to an induction motor fault diagnosis method based on multi-source signal fusion deep learning. Background Art

[0002] With the advancement of industrial automation and the rapid development of intelligent manufacturing, efficient and reliable production equipment has become one of the key factors in corporate competitiveness. As a core power equipment widely used in the industrial field, the stability and efficiency of induction motors directly affect the operating cost and efficiency of the entire production system. However, traditional monitoring methods mostly rely on regular manual inspections and passive repairs after failures, which not only cannot reflect the motor's operating status in real time, but also cannot effectively predict potential failures, resulting in frequent unplanned downtime, increasing maintenance costs and reducing overall productivity.

[0003] The traditional fault diagnosis method that relies on manual experience is limited by the uncertainty of subjective judgment, the limitation of detection cycle and the lack of ability to identify subtle fault signs, and it is difficult to meet the current needs of intelligent and refined management. Most of the existing induction motor monitoring systems are limited to the monitoring of a single parameter, such as spectrum analysis and pattern recognition based on signal processing. Although they can identify specific fault types to a certain extent, they are often limited by the subjectivity of manual feature extraction, the lack of ability to identify complex fault patterns and the sensitivity to environmental interference, and it is difficult to meet the growing needs of real-time monitoring and intelligent diagnosis. In addition, a single sensor signal can often only provide partial information about the motor status, and there are limitations for the comprehensive diagnosis of faults. Especially when facing early minor faults or compound faults, the diagnostic accuracy and sensitivity of a single signal are often unsatisfactory.

[0004] With the rapid progress of information technology, Internet of Things technology and artificial intelligence technology, especially the rise of deep learning technology, the field of induction motor fault diagnosis has brought innovative breakthroughs. By constructing a multi-level neural network model, deep learning can automatically learn and extract fault features from massive data, realize the effective identification of complex and nonlinear fault modes, and show significant advantages in processing nonlinear and high-dimensional induction motor fault signals, greatly improving the accuracy and efficiency of diagnosis. However, the construction and application of deep learning models are not easy, especially in the field of multi-source signal fusion, facing challenges such as large data volume, noise interference, high feature dimension, and complex and changeable fault modes. Effectively integrating these highly heterogeneous signals, avoiding information redundancy, extracting key fault features, and ensuring the generalization ability of the model are the key to achieving efficient fault diagnosis. Summary of the invention

[0005] In view of the above problems, the present invention proposes an induction motor fault diagnosis method based on multi-source signal fusion deep learning. The current signal and vibration signal of the induction motor are multi-source fused and trained with deep learning to obtain better signal features. The extracted signal features are fused with the attention mechanism, and then the timing dependency is captured through the memory neural network. The training set is further used to train the network until convergence, the model parameters are obtained, and finally the test set is used to perform induction motor fault diagnosis to achieve more accurate diagnosis results.

[0006] The technical solution of the present invention is:

[0007] An induction motor fault diagnosis method based on multi-source signal fusion deep learning includes the following steps:

[0008] Step A): Collect current signals and vibration signals of induction motors with known fault categories, and mark the current signals and vibration signals of the motors according to categories to obtain training samples.

[0009] Step B): Divide the collected data set into training set, test set and validation set, with the proportions of 70%, 20% and 10% respectively.

[0010] Step C): Build a network model based on convolutional network and multi-source signal fusion to fuse current and vibration signals and extract features.

[0011] Step D): Use the training set to train the network until the model converges and save the model parameters. Load the trained model parameters and use the test set to perform motor fault diagnosis.

[0012] Preferably, the step B is further performed as follows:

[0013] Organize the collected signals into data sets m represents the number of samples, X i represents the i-th sample, l i Represents X i Corresponding data labels. Since the recurrent network performs time series analysis, the training set and test set must be arranged in chronological order and cannot be randomly selected.

[0014] Preferably, the step C is further performed as follows:

[0015] Step C1 takes the current signal and vibration signal during the operation of the induction motor as input, and uses a one-dimensional neural network to extract features from the training set. Each one-dimensional neural network sub-block contains a convolutional layer Each submodule is designed to extract features from the original current and vibration signals, converting the original signals into multiple feature maps. Downsampling is performed through the maximum pooling layer to retain important features and remove redundant information.

[0016] Furthermore, the first layer of the neural network is used to process the time series data, and the dimensions of the feature map are rearranged using the Permute layer to adjust it to a format suitable for attention mechanism processing. Secondly, the multi-source data attention module is used to adaptively allocate the weights of the current signal a1 and the vibration signal a2 features, and then the re-weighted features are fused to obtain a new feature extraction layer output g.

[0017] Furthermore, step C2: the fused signal g is subjected to two different convolution-pooling operations to extract more abstract and more distinguishing feature representations, and the fault features extracted from each signal are flattened through the Flatten layer to reduce the dimension of the features. The feature size is reduced to the SoftMax output layer through the fully connected layer, thereby realizing the fault diagnosis of the broken rotor bar of the induction motor.

[0018] Preferably, the step D is further as follows: determining whether the model has converged by calculating the loss, training the network model using the training set until convergence, saving the converged model parameters, loading the trained model parameters, and using the test set to perform motor fault diagnosis.

[0019] The beneficial effects of the present invention are:

[0020] 1. Different sensor signals (such as current, vibration, temperature, etc.) may have different importance in characterizing equipment status or fault mode. The present invention adopts multi-source fusion technology. Current signals usually reflect the internal electrical conditions of the motor, such as winding short circuit, unbalanced load, etc., while vibration signals are closely related to mechanical conditions, such as bearing damage, rotor imbalance, etc. By fusing these two signals with different physical properties, more comprehensive motor operation status information can be obtained, thereby improving the accuracy of identifying various fault modes.

[0021] 2. The multi-source data attention mechanism adopted in the present invention can be used interspersed in the neural network, so as to avoid the problem of low neural network recognition accuracy caused by equal-weight fusion of multi-source data features, and perform weighted processing on the signals extracted by the convolutional neural network to improve the characterization ability of fault feature signals.

[0022] 3. The present invention does not require prior knowledge, can make full use of one-dimensional original current signals and vibration signals and mine hidden information to obtain more accurate signal characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described in detail below in conjunction with the accompanying drawings;

[0024] Figure 1 Flowchart for induction motor fault diagnosis.

[0025] Figure 2This is the neural network fault feature extraction module diagram.

[0026] Figure 3 This is the framework diagram of the feature fusion module of the multi-source signal attention mechanism.

[0027] Figure 4 Diagram of feature signal splicing and fusion mapping process. DETAILED DESCRIPTION

[0028] The present invention will be further described in detail below in conjunction with the accompanying drawings;

[0029] Step A): First, current signals and vibration signals of induction motors with known fault categories are collected, and the current signals and vibration signals of the motors are marked according to categories to obtain training samples.

[0030] like Figure 1 As shown, the induction motor fault diagnosis method proposed in the present invention. First, the current signal and vibration signal of the induction motor with known fault category are collected by sensors to obtain raw data. These signal data are then marked and organized into a structured data set, where each record contains signal features and their corresponding fault categories. These signals cover the performance of the motor under different fault categories, such as bearing wear, rotor imbalance, electrical short circuit, etc. The signal data is captured in real time or periodically, and then pre-processed by digital signal processing technology, including but not limited to noise filtering, signal enhancement and feature extraction, to eliminate irrelevant interference and highlight key information.

[0031] The preprocessed signal data is converted into a structured format, usually presented in a table, where each row represents an observation sample, including but not limited to signal features such as time series data, spectrum analysis results, and the corresponding specific fault category label. This process ensures the standardization of the data set, which facilitates subsequent data analysis and machine learning model training.

[0032] Step B): Further, the collected data set is divided into training set, test set and validation set, with the proportions of 70%, 20% and 10% respectively.

[0033] In order to ensure the objectivity and effectiveness of model evaluation, this paper adopts a multi-stage data set partitioning strategy, which aims to divide the entire data set into three independent and complementary subsets in proportion through a random sampling algorithm: training set, validation set and test set. This partitioning process strictly adheres to the principles of randomness and independence to avoid the introduction of bias.

[0034] The first step is preliminary segmentation, using a random sampling algorithm to split the original data set into two subsets. The larger subset (about 80%) is used as the initial training and validation set for model training and hyperparameter tuning; the smaller subset (about 20%) is used as the test set for the final evaluation of the generalization ability of the model.

[0035] Then, a second split was performed, and a second random sampling was performed on the preliminary training and validation sets, which were further subdivided into training sets and validation sets. The training set accounted for 87.5% of the preliminary subset, equivalent to 70% of the total data set, and was used for the actual model training process; the validation set accounted for 12.5% ​​of the preliminary subset, equivalent to 10% of the total data set, and was used to monitor performance changes during the model training process to avoid overfitting.

[0036] In the process of data set division, the present invention particularly emphasizes the consistency of data distribution to ensure that each subset can reflect the overall characteristics of the original data set. To this end, a stratified sampling strategy is adopted, that is, the distribution of fault categories is considered in the random sampling process to ensure that the training set, validation set, and test set all contain all fault types, and the proportion of each fault category in different subsets is as close as possible to the proportion in the original data set.

[0037] The training set is divided into Figure 2 The learning of the signal feature extraction model is carried out by adjusting the model parameters through the optimization algorithm to minimize the loss function. The test set is independent of the training and verification process and is used for the final fault diagnosis of the motor. The verification set is used regularly during the model training process to optimize the model architecture and prevent overfitting by adjusting hyperparameters and monitoring model performance indicators (such as accuracy, recall, etc.). The test set is used once after the model training is completely completed to provide a data set that has not been "seen" by the model, which is used to evaluate the model's predictive ability and generalization level on unknown data, thereby verifying the effectiveness and practicality of the model.

[0038] Step C): Further, a network model based on convolutional network and multi-source signal fusion is constructed to fuse the current and vibration signals and extract features.

[0039] The present invention uses a one-dimensional convolutional neural network to extract features from the input motor current signal and vibration signal. Figure 2 As shown in FIG. 1 , the one-dimensional convolutional neural network used in the present invention only includes a convolution layer and a pooling layer. In the convolution layer, the convolution kernel performs a convolution operation on the input data and constructs the output features using a nonlinear activation function. The output result of each convolution layer is obtained by formula 1

[0040]

[0041] in, is the input signal of the jth convolution kernel in layer l+1, is the weight of the i-th convolution kernel in layer l; x l (j) is the receptive field of the convolution kernel; is the bias of the convolution kernel.

[0042] The activation function is used to perform a nonlinear transformation on the output after each convolution operation. The nonlinear transformation is given by Formula 2

[0043]

[0044] Among them, f(g) is the operator of the activation function; is the output value of the convolution operation; for The target value obtained by the activation function.

[0045] In the feature extraction stage, the maximum pooling operation is used to take the maximum value in the perception domain as the output. The maximum pooling layer is given by Formula 3

[0046]

[0047] Among them, P i l+1 (j) is the input value of the neuron in the l+1th layer, is the value of the tth neuron of the i-th feature in the l-th layer, t∈[(j-1)H,jH], H is the width of the pooling area.

[0048] The current signal and vibration signal of the motor are respectively passed through convolution layer 1 and convolution layer 2 to capture the local features of the input signal, and the convolution results are output and sent to the maximum pooling layer 1 and maximum pooling layer 2 respectively. In the maximum pooling, the maximum value of the feature value in the perception domain is calculated by formula 2 as the output. After being processed by the convolution layer and pooling layer, the local features of the time series and spectrum domain of the current signal and vibration signal are extracted. Before fusing them, the Permute layer is used to adjust the dimensional order of the data, and finally reach the output layer 1 and output layer 2.

[0049] In order to more efficiently utilize the effective information of different channel features and suppress the influence of irrelevant components, the present invention adopts Figure 3 The multi-source data attention mechanism module shown adaptively assigns weights to the data features of different input signal sources extracted by the neural network, and then fuses the two weighted features to obtain the output of a new feature extraction layer g.

[0050] First, two multi-layer perceptrons are used to compress the input current signal feature value a1 and the vibration signal feature value a2. Each perceptron compresses the input feature into a 1′1 output, and then the compressed features are spliced. The normalized exponential function Softmax function is used, as given by Formula 4, to calculate the spliced ​​features.

[0051]

[0052] Among them, ω i is the weight of the ith signal source, z i is the eigenvalue of the compressed ith signal source.

[0053] The obtained weights are then multiplied element-by-element with the original features of each signal source, or the features after preliminary processing, as shown in Formula 5, and then all weighted features are added to form the final fusion feature. This process is equivalent to weighted summation of the original features according to the relative importance of the signal source, thereby achieving effective signal fusion.

[0054]

[0055] in, is the feature output by the tth neuron in the ith feature of the lth layer of the motor current signal source, is the feature output by the tth neuron in the ith feature of the lth layer of the motor vibration signal source; is the corresponding weighted current signal feature, is the corresponding weighted vibration signal feature; γ1 is the weight of the current signal feature corresponding to the vibration signal source, and γ2 is the weight of the vibration signal feature corresponding to the current signal source. It can be calculated by formula 6.

[0056]

[0057] Among them, P ML1 (·) and P ML1 (·) are the multi-layer perceptrons of the compression current and vibration characteristic signals respectively; a1 and a2 are the whole sound feature set and vibration feature set respectively; s(·) is the operator of the Softmax function.

[0058] After obtaining the weighted sound signal features and vibration signal features, in order to provide more sufficient information, it is necessary to perform feature fusion operation on the two weighted feature signals. The process is as follows: Figure 4 As shown. Let F1 and F2 be the one-dimensional expansion sequence of all weighted sound signal feature sets and the one-dimensional expansion sequence of weighted vibration signal feature sets respectively; g is the feature signal used for classification after fusion, which is given by Formula 7.

[0059] g = map[stitch(F1,F2)] (7)

[0060] Among them, stitch is the feature signal dimension expansion and splicing operation, F1 j and F1 j are the j-th local eigenvalue in the weighted current signal feature set and the j-th local eigenvalue in the weighted vibration signal feature set, respectively. map is the mapping operation in the fully connected layer, which is given by Formula 8.

[0061]

[0062] Furthermore, if Figure 2 As shown in the figure, the signal g obtained by the multi-source data attention mechanism fusion module is used as a new feature extraction layer output and as the input of the convolution layer 3. After two convolution-pooling operations, a more abstract and discriminative feature representation is extracted from the feature g of the fused current and vibration signals.

[0063] The features of the convolutional network output are used as the input of the long short-term memory network. After the network transformation, the last hidden layer output state is used as the feature of the final signal for the diagnosis of the motor state. A standard long short-term memory network unit consists of an input vector, a storage unit, an output vector, and several gate units. The input gate and output gate control the unit information flowing in and out, and the forget gate can selectively delete information from the repeated connections of the previous unit.

[0064] The input of the LSTM network is the output feature of the convolutional network. For the motor fusion feature data output by the fully connected layer at time t, it first passes through the forget gate of the LSTM network and is given by Formula 9

[0065] f t =σ(W f ×[h t-1 ,x t ]+b f (9)

[0066] Among them, f t is the output of the forget gate, W f is the corresponding weight matrix, b f is the bias term, h t-1 is the hidden state of the previous time step, x t Is the input of the current time step. The forget gate outputs a value between 0 and 1 through a sigmoid activation function σ, where 0 means complete forgetting and 1 means complete retention.

[0067] i t =σ(W i ×[h t-1,x t ]+b i (10)

[0068]

[0069] Among them, i t is the output of the input gate, is the candidate cell state, W i and W c is the corresponding weight matrix, b i and b c is the bias term.

[0070] Then, according to the output of the forget gate and the output gate, the cell state is updated through formula 12, and the new candidate cell state is merged.

[0071]

[0072] Among them, c t is the updated cell state, c t-1 is the cell state at the previous time step, and e represents the element-wise product.

[0073] Equation 13 shows which cell state contents should be used to generate the hidden state of the current time step.

[0074] o t =σ(W o ·[h t-1 ,x t ]+b0) (13)

[0075] Among them, t is the output of the output gate, W o is the weight matrix, b o is the bias term.

[0076] Then take the weighted sum of the output gate and the cell state obtained in the previous step, and get the hidden state by formula 14

[0077] h t =o t etanh(c t ) (14)

[0078] Among them, h t is the hidden state of the current time step, tanh(c t ) is the cell state c t The tanh activation function is used to obtain the last state of the hidden layer, which is used as the final signal feature for motor status diagnosis.

[0079] Before the fused feature signal is input to the fully connected layer, the Flatten layer is used to flatten the features and compress these multi-dimensional features into one dimension. The fully connected layer expands the output of the last pooling layer into a one-dimensional vector, which is also used as the input of the fully connected layer, and then a full connection is performed between the input and output. The fully connected layer can merge the differential local information of the convolutional layer or the pooling layer, which is given by Formula 15

[0080]

[0081] Among them, z l+1 (j) is the output value of the jth neuron in the l+1 layer, is the weight between the tth neuron in the i-th layer feature and the jth neuron in the l+1th layer, is the output value of the tth neuron in the ith feature signal of the previous layer, is the total bias of the network, and f(g) is the operator of the activation function.

[0082] Step D): Further, the network is trained using the training set until the model converges, and the model parameters are saved. The trained model parameters are loaded, and the motor fault diagnosis is performed using the test set.

[0083] In the present invention, the input signal learns the fault feature representation through the constructed network structure. After the effective signal with fault features is extracted, the features are finally normalized using the Softmax function, and the fault features are classified by the Softmax full connection.

[0084] The Softmax function receives the output vector of the fully connected layer, where each element corresponds to a potential fault category. The function applies an exponential operation to each element in this vector and then divides it by the sum of the exponential operation results of all elements to obtain a new vector where each element represents the probability value of the corresponding position in the original vector and the sum of all probabilities is 1. The Softmax function for classification is given by Formula 16:

[0085]

[0086] Among them, z k is the raw score of the kth fault category, that is, the unnormalized output, and P(y=k|X) is the probability of belonging to the kth fault category given the input feature X.

[0087] The index corresponding to the maximum value in the Softmax output vector is the fault category with the highest probability that the network finally predicts, which represents the most likely fault type considered by the model. During training, the output of the Softmax layer is usually used together with the actual fault label to calculate the cross entropy loss function.

[0088] During training, the proposed loss function is the cross entropy between the output probability distribution and the target probability distribution, which is given by Formula 17

[0089]

[0090] Where C is the total number of categories output by the model, and q c The probability distribution of the prediction generated by the Softmax function, p c is the true probability distribution.

[0091] The accompanying drawings and detailed description of the present invention mentioned above as examples of the present invention are used to explain the present invention, but do not limit the meaning or scope of the present invention described in the claims. For those skilled in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is limited by the attached claims and their equivalents.

[0092] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. An induction motor fault diagnosis method based on multi-source signal fusion deep learning, characterized in that: The following steps are involved: Step A: collecting current signals and vibration signals of induction motors with known fault categories, and marking the current signals and vibration signals of the motors according to categories to obtain training samples; Step B: Divide the collected data set into training set, test set and validation set, with the proportions of 70%, 20% and 10% respectively; Step C: Build a network model based on convolutional network and multi-source signal fusion to fuse current and vibration signals and extract features; Step D: Use the training set to train the network until the model converges, save the model parameters, load the trained model parameters, and use the test set to perform motor fault diagnosis.

2. The induction motor fault diagnosis method based on multi-source signal fusion deep learning according to claim 1 is characterized in that: In the step B: Organize the collected signals into data sets m represents the number of samples, X i represents the i-th sample, l i Represents X i Corresponding data labels, since the recurrent network performs time series analysis, the training set and test set need to be arranged in chronological order and cannot be randomly extracted.

3. The induction motor fault diagnosis method based on multi-source signal fusion deep learning according to claim 1 is characterized in that: The specific process of step C is: Step C1, taking the current signal and vibration signal during the operation of the induction motor as input, a one-dimensional neural network is used to extract features from the training set. Each one-dimensional neural network sub-block contains a convolutional layer in, is the input signal of the jth convolution kernel in layer l+1, is the weight of the i-th convolution kernel in layer l; x l (j) is the receptive field of the convolution kernel; is the deviation of the convolution kernel; each submodule is designed to extract features from the original current and vibration signals, convert the original signals into multiple feature maps, downsample through the maximum pooling layer, retain important features and remove redundant information; The first layer of the neural network is used to process the time series data. The Permute layer is used to rearrange the dimensions of the feature map and adjust it to a format suitable for attention mechanism processing. Secondly, the multi-source data attention module is used to adaptively allocate the weights of the current signal a1 and the vibration signal a2 features. Then, the re-weighted features are fused to obtain a new feature extraction layer output g. In step C2, the fused signal g is subjected to two different convolution-pooling operations to extract more abstract and discriminative feature representations. The fault features extracted from each signal are flattened through the Flatten layer to reduce the dimension of the features; the feature size is reduced to the SoftMax output layer through the fully connected layer, thereby realizing the fault diagnosis of the broken rotor bar of the induction motor.

4. The induction motor fault diagnosis method based on multi-source signal fusion deep learning according to claim 1 is characterized in that: In the step D: determine whether the model has converged by calculating the loss, use the training set to train the network model until convergence, save the model parameters after convergence, load the trained model parameters, and use the test set to perform motor fault diagnosis.

Citation Information

Cited By

  • Antenna servo motor fault diagnosis method and device for satellite communication, and medium

    CN119646613A