Multi-signal fusion induction motor fault diagnosis method based on knowledge and data double-wheel driving

Abstract

The invention discloses a multi-signal fusion induction motor fault diagnosis method based on knowledge and data double-wheel driving, and relates to the technical field of motor fault diagnosis. Comprising the steps of obtaining an induction motor simulation model, actual motor operation data and a motor knowledge base; generating a simulation data set by using the induction motor simulation model, inputting the simulation data set into the neural network model to be migrated for training, and obtaining a fault diagnosis baseline model of the induction motor in an ideal environment; inputting actual motor operation data into the fault diagnosis baseline model, and finely adjusting parameters of the fault diagnosis baseline model to obtain a fault diagnosis model; inputting the to-be-classified data into the fault diagnosis model to obtain a motor diagnosis prediction result; and inputting the motor diagnosis prediction result into a motor knowledge base to obtain a fault reason and a maintenance suggestion corresponding to the motor diagnosis prediction result. According to the method, a more accurate and stable diagnosis result can be obtained, and the fault decision time of the induction motor is effectively shortened.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly to a multi-signal fusion induction motor fault diagnosis method based on dual-wheel drive of knowledge and data. Background Technique

[0002] Common faults of induction motors mainly include bearing faults, stator faults, and rotor faults. Among them, the proportion of bearing faults is 40%, the proportion of stator faults is 36%, the proportion of rotor faults is 10%, and the proportion of other various faults is about 10%. Classified from the diagnostic technology, existing fault diagnoses are usually divided into three categories: methods based on analytical models, methods based on signal processing, and methods based on knowledge. The method based on an analytical model needs to establish a mathematical model for a certain fault of the object to be detected, and it is difficult to implement on-site; the method based on signal processing processes the signals collected on-site and compares them with normal signals to determine whether the motor has a fault; the knowledge-based diagnostic method uses artificial intelligence algorithms such as neural networks and fuzzy logic, trains the model first, and then inputs the processed fault signals into the model for fault classification. With the development of motor fault diagnosis technology, these diagnostic technologies are no longer used alone, but multiple technologies are used in combination. Classified according to the type of sampled signal, it can be divided into: electromagnetic field detection, temperature detection, infrared identification, radio frequency detection, noise detection, vibration analysis, chemical analysis, current analysis, etc. Traditional mechanical fault diagnosis first processes the original signal, extracts fault features, and inputs the extracted features into a classifier for fault diagnosis. The fault diagnosis method based on deep learning simplifies the preprocessing process, inputs the processed original signal into the trained model, and implicitly performs feature selection.

[0003] Existing mainstream motor fault diagnosis technical solutions are all data-driven. However, pure data-driven technical solutions require a large amount of reliable labeled data. In an industrial environment, the signals of induction motors are affected by multiple factors, and it is difficult to effectively guarantee the signal quality, which further increases the implementation difficulty of existing fault diagnosis technical solutions. Secondly, existing motor fault diagnosis technical solutions with low requirements for labeled data are all based on single-signal modeling. In the real world, when an induction motor fails, its fault characteristics often reflect in different signals. For example, when the bearing of an induction motor breaks, both its vibration signal and current signal will be affected; while the single-signal modeling solution ignores the internal physical connection between different signals and cannot model fault signals from multiple aspects and dimensions, resulting in its susceptibility to external factors such as sensors and the environment, as well as its own quality defects. Finally, there is a large amount of induction motor knowledge in physical enterprises (including the common knowledge of induction motor equipment and the maintenance knowledge of induction motors recorded by old masters). Existing technical solutions all ignore the value of this knowledge, and induction motor knowledge cannot fully play its due value in these solutions.

[0004] Therefore, how to provide a multi-signal fusion induction motor fault diagnosis method driven by both knowledge and data to solve the difficulties existing in the prior art is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a multi-signal fusion induction motor fault diagnosis method driven by both knowledge and data, which can obtain more accurate and stable diagnosis results and effectively reduce the induction motor fault decision-making time.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A multi-signal fusion induction motor fault diagnosis method driven by both knowledge and data includes the following steps:

[0008] S1. Obtain the induction motor simulation model, actual motor operation data, and motor knowledge base;

[0009] S2. Generate a simulation data set using the induction motor simulation model, input the simulation data set into the neural network model to be migrated for training, and obtain a fault diagnosis baseline model of the induction motor in an ideal environment;

[0010] S3. Input the actual motor operation data into the fault diagnosis baseline model to fine-tune the parameters of the fault diagnosis baseline model and obtain a fault diagnosis model;

[0011] S4. Input the data to be classified into the fault diagnosis model to obtain the motor diagnosis prediction result;

[0012] S5. Input the motor diagnosis prediction result into the motor knowledge base to obtain the fault cause and maintenance suggestions corresponding to the motor diagnosis prediction result.

[0013] In the above method, optionally, the simulation data set in S2 includes: the vibration signal and current signal of the induction motor in the normal simulation state, and the vibration signal and current signal obtained from various fault states of the induction motor simulation.

[0014] In the above method, optionally, the neural network model to be migrated in S2 includes: a data preprocessing unit, an iTransformer neural network, and a Softmax classifier.

[0015] In the above method, optionally, the specific content of the training of the neural network model to be migrated in S2 is:

[0016] S21. Obtain the vibration signal and current signal in the motor simulation data set, and perform timestamp alignment on the vibration signal and current signal respectively to obtain the vibration signal and current signal after timestamp alignment;

[0017] S22. Respectively perform reversible instance normalization on the vibration signal and the current signal after aligning the timestamps to obtain the normalized vibration signal and the current signal;

[0018] S23. After feature splicing the normalized vibration signal and the current signal, input them into the iTransformer neural network for training to obtain the output data of the iTransformer neural network;

[0019] S24. Input the output data of the iTransformer neural network into the Softmax classifier to obtain the probability distribution of the induction motor fault, and determine the motor fault type based on the probability distribution value.

[0020] For the above method, optionally, in S22, the reversible instance normalization includes normalization and inverse normalization:

[0021] Normalization: Perform normalization processing on the mean and variance of each input data in each original data to obtain the normalized data;

[0022] Inverse normalization: Perform inverse normalization processing on the normalized data to restore it to the original distribution.

[0023] For the above method, optionally, the iTransformer neural network is an encoder structure, including: an embedding layer, a projection layer, and multiple stackable Transformer modules;

[0024] The embedding layer and the projection layer are implemented using a multi-layer perceptron;

[0025] The Transformer module includes: a layer normalization block, a feed-forward network block, and a self-attention block.

[0026] For the above method, optionally, the number of attention heads of the iTransformer neural network is 8, the number of layers is 6, and the word embedding dimension is equal to the dimension of the spliced signal.

[0027] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a multi-signal fusion induction motor fault diagnosis method based on dual-wheel drive of knowledge and data, which has the following beneficial effects: 1) Most of the samples required to construct the induction motor fault diagnosis model of the present invention are generated by the induction motor simulation model, effectively reducing the demand for real-label data samples of the induction motor; 2) The present invention comprehensively utilizes the characteristics of the induction motor vibration signal and current signal, which can reduce the influence of environmental factors on a single signal and obtain a more accurate and stable diagnosis result than a single-signal model; 3) The present invention combines the induction motor fault diagnosis algorithm model with the knowledge of the induction motor. Through the recommendation of the induction motor fault causes and maintenance suggestions, the induction motor fault decision time can be effectively reduced, and the production line capacity loss cost caused by the induction motor fault can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative work.

[0029] Figure 1 It is a flowchart of a multi-signal fusion induction motor fault diagnosis method based on dual-wheel drive of knowledge and data disclosed by the present invention;

[0030] Figure 2 It is a schematic diagram of a multi-signal fusion induction motor fault diagnosis method based on dual-wheel drive of knowledge and data disclosed by the present invention;

[0031] Figure 3 It is a diagram of the network model to be migrated disclosed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0033] Referring to Figure 1 and Figure 2 as shown, the present invention discloses a multi-signal fusion induction motor fault diagnosis method based on dual-wheel drive of knowledge and data, including the following steps:

[0034] S1. Obtain the induction motor simulation model, actual motor operation data, and motor knowledge base;

[0035] S2. Generate a simulation dataset using the induction motor simulation model, input the simulation dataset into the neural network model to be migrated for training, and obtain a fault diagnosis baseline model for the induction motor in an ideal environment;

[0036] S3. Input the actual motor operation data into the fault diagnosis baseline model to fine-tune the parameters of the fault diagnosis baseline model and obtain a fault diagnosis model;

[0037] S4. Input the data to be classified into the fault diagnosis model to obtain the motor diagnosis prediction result;

[0038] S5. Input the motor diagnosis prediction result into the motor knowledge base to obtain the fault cause and maintenance suggestions corresponding to the motor diagnosis prediction result.

[0039] Furthermore, the induction motor simulation model in S1 includes: establishing simulation models of the motor in normal state and various fault states through simulation software such as Matlab Simulink and Ansys Maxwell.

[0040] Furthermore, the simulation dataset in S2 includes: vibration signals and current signals in the normal state of the induction motor simulation, and vibration signals and current signals obtained from various fault states of the induction motor simulation.

[0041] Furthermore, referring to Figure 3 as shown, the neural network model to be migrated in S2 includes: a data preprocessing unit, an iTransformer neural network, and a Softmax classifier.

[0042] Furthermore, the specific content of training the neural network model to be migrated in S2 is:

[0043] S21. Obtain the vibration signals and current signals in the motor simulation dataset, and perform timestamp alignment on the vibration signals and current signals respectively to obtain the vibration signals and current signals after timestamp alignment;

[0044] S22. Perform reversible instance normalization on the vibration signals and current signals after timestamp alignment to obtain the normalized vibration signals and current signals;

[0045] S23. Perform feature splicing on the normalized vibration signals and current signals and then input them into the iTransformer neural network for training to obtain the output data of the iTransformer neural network;

[0046] S24. Input the output data of the iTransformer neural network into the Softmax classifier to obtain the probability distribution of induction motor faults, and determine the motor fault type based on the probability distribution value.

[0047] Further, in S22, the invertible instance normalization includes normalization and inverse normalization:

[0048] Normalization: Normalize the mean and variance of each input data in each original data to obtain the normalized data;

[0049] Inverse normalization: Perform inverse normalization on the normalized data to restore it to the original distribution.

[0050] Specifically, in the normalization process, for the given input data x ∈ R B×T×N , where B is the batch size, T is the number of time steps, and N is the number of features. First, calculate the mean μ of the input data x i,t :

[0051]

[0052] In normalization, only the information in the time (T) dimension is concerned, and the obtained standard deviation σ is:

[0053]

[0054] where ε is a very small value, and then perform normalization on the data to obtain the normalized input data

[0055]

[0056] To enhance the expressive power of the model, RevIN introduces the affine transformation parameters γ and β, and the output data y of the normalization layer i,t is:

[0057]

[0058] In inverse normalization, restoring the normalized data to the original distribution includes:

[0059]

[0060] Further, the iTransformer neural network is an encoder structure, including: an embedding layer, a projection layer, and multiple stackable Transformer modules;

[0061] The embedding layer and the projection layer are implemented using a multi-layer perceptron;

[0062] The Transformer module includes: a layer normalization block, a feed-forward network block, and a self-attention block.

[0063] Specifically, the layer normalization block was initially designed to improve the training stability and convergence of deep networks. In previous Transformers, this module normalized multiple variables at the same time, making each variable indistinguishable. Once the collected data is not aligned in time, this operation will also introduce interactive noise between non-causal or delayed processes. In the inverted version, layer normalization is applied to the feature representations of each variable, bringing the feature channels of all variables into a relatively uniform distribution. This normalization idea has been widely proven to be effective in dealing with non-stationary time series problems and can be naturally implemented through layer normalization in iTransformer. In addition, since the feature representations of all variables are normalized to a normal distribution, the differences caused by different variable value ranges can be weakened. In contrast, in the previous structure, the feature representations (Temporal Tokens) of all timestamps were uniformly standardized, resulting in an over-smoothed time series actually seen by the model.

[0064] The feed-forward network block is used to encode word vectors. In previous models, word vectors were formed from multiple variables collected at the same time, and their generation times may not be consistent, and it is difficult for a word reflecting a time step to provide sufficient semantics. In the inverted version, the entire sequence of the same variable forms the word vector. Based on the universal approximation theorem of multi-layer perceptrons, it has a large enough model capacity to extract time features shared in historical observations and future predictions and uses feature extrapolation for prediction results.

[0065] The self-attention block is used to model the correlations between different variables. The entire attention map can reveal the correlations between variables to a certain extent, and in subsequent weighted operations based on the attention map, highly correlated variables will obtain greater weights in their interactions with their Value vectors.

[0066] Furthermore, the iTransformer neural network has 8 attention heads, 6 layers, and the word embedding dimension is equal to the dimension of the concatenated signals.

[0067] In a specific embodiment, motor technical materials and motor maintenance materials are obtained to generate a fact database, which includes a motor equipment knowledge base constructed based on motor technical specifications, a motor equipment fault base constructed based on bearing fracture faults, and a motor repair case base constructed based on motor short-circuit repairs; motor equipment common knowledge bases and induction motor equipment proprietary knowledge bases are constructed by motor equipment experts as expert knowledge bases. The current signal of an actual induction motor is input into the fault diagnosis model, and the fault type is determined to be stator winding short circuit. The fault type is input into the fact database and the expert database. The fact database obtains the fault cause as motor stator insulator aging, and the expert database obtains the corresponding solution as replacing the stator winding. The maintenance personnel perform corresponding processing on the motor according to the obtained solution.

[0068] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive, characterized in that: The following steps are involved: S1, obtaining an induction motor simulation model, actual motor operation data and a motor knowledge base; S2. Generate a simulation data set using the induction motor simulation model, input the simulation data set into the neural network model to be transferred for training, and obtain a fault diagnosis baseline model of the induction motor under an ideal environment; S3, inputting actual motor operation data into the fault diagnosis baseline model, fine-tuning the parameters of the fault diagnosis baseline model, and obtaining a fault diagnosis model; S4, inputting the data to be classified into a fault diagnosis model to obtain a motor diagnosis prediction result; S5. Input the motor diagnosis prediction result into the motor knowledge base to obtain the fault cause and maintenance suggestion corresponding to the motor diagnosis prediction result.

2. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 1 is characterized in that: The simulation data set in S2 includes: the vibration signal and current signal of the induction motor in the normal state simulation, and the vibration signal and current signal obtained by simulating various fault states of the induction motor.

3. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 2 is characterized in that: The neural network models to be migrated in S2 include: data preprocessing unit, iTransformer neural network and Softmax classifier.

4. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 3 is characterized by: The specific content of the training of the neural network model to be transferred in S2 is: S21, obtaining a vibration signal and a current signal in a motor simulation data set, and performing timestamp alignment on the vibration signal and the current signal respectively, to obtain a vibration signal and a current signal after the timestamps are aligned; S22, performing reversible instance normalization on the vibration signal and the current signal after the alignment time stamp, respectively, to obtain a normalized vibration signal and a normalized current signal; S23, performing feature splicing on the normalized vibration signal and current signal and inputting the resultant signal into the iTransformer neural network for training to obtain the iTransformer neural network output data; S24. Input the output data of the iTransformer neural network into the Softmax classifier to obtain the probability distribution of the induction motor fault, and determine the motor fault type based on the probability distribution value.

5. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 4 is characterized in that: In S22, reversible instance normalization includes normalization and inverse normalization: Normalization is: normalizing the mean and variance of each input data in each original data to obtain normalized data; Denormalization is to perform denormalization on the normalized data to restore it to its original distribution.

6. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 4 is characterized in that: The iTransformer neural network is an encoder structure, including: an embedding layer, a projection layer, and multiple stackable Transformer modules; The embedding layer and projection layer are implemented using multi-layer perceptron; The Transformer module includes: layer normalization block, feedforward network block and self-attention block.

7. The multi-signal fusion induction motor fault diagnosis method based on knowledge and data dual-wheel drive according to claim 6 is characterized in that: The iTransformer neural network has 8 attention heads and 6 layers, and the word embedding dimension is equal to the concatenated signal dimension.

Images