Multiphase permanent magnet synchronous motor fault diagnosis method and system
By generating the residual sequence of a multiphase permanent magnet synchronous motor and a Gaussian white noise vector sequence combined with a neural network training classifier, the low accuracy and false alarm problems of multiphase motor fault diagnosis are solved, and higher diagnostic accuracy and type judgment capabilities are achieved.
Patent Information
- Application Number
- CN202510288640.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-07-04
AI Technical Summary
The existing multiphase motor fault diagnosis methods have low accuracy and are prone to false alarms when the number of fault samples is small, making it difficult to accurately judge the fault type.
By obtaining the working condition data sequences of each phase of the motor, comparing them with the simulation data sequence to generate residual sequences, using Gaussian white noise to generate vector sequences, and troubleshooting is performed through the discriminating network training classifier, including a combination of recursive neural network, multi-head self-attention neural network, time sampling layer and fully connected neural network, combining high-dimensional vector generation network and discriminating network for troubleshooting.
Improve the accuracy of fault diagnosis, reduce the false alarm rate, and improve the accurate judgment ability of fault types.
Smart Images
Figure CN120254597A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of motor control, and in particular, to a fault diagnosis method and system for a multiphase permanent magnet synchronous motor. Background Art
[0002] Currently, the industrial community has put forward higher requirements and goals for motor control. Due to its higher control freedom, excellent fault tolerance performance, and higher power density in control, multiphase motors have been widely used in multiple industries. The fault tolerance performance of multiphase motors is the realization of the design of fault tolerance schemes, and the fault diagnosis method can better improve the fault tolerance scheme.
[0003] However, the existing diagnostic methods usually train classifiers based on fault samples. In the case of a small number of fault samples, the classification accuracy is low; while only based on the difference between the predicted data sequence and the actual data sequence of normal operating conditions for fault diagnosis, it is difficult to accurately judge the fault type and is prone to false alarms. Summary of the Invention
[0004] In view of the above problems existing in the prior art, the present invention proposes a fault diagnosis method and system for a multiphase permanent magnet synchronous motor, mainly solving the problems of low accuracy and easy false alarms in existing motor fault diagnosis.
[0005] In order to achieve the above object and other objects, the technical solutions adopted by the present invention are as follows.
[0006] The present application provides a fault diagnosis method for a multiphase permanent magnet synchronous motor, including: obtaining the working condition data sequences corresponding to each phase of the motor, and comparing the working condition data sequences with the simulation data sequences corresponding to the respective phases to obtain a residual sequence, where the simulation data sequences are output by a dynamic parameter model, and the dynamic parameter model is modeled based on the historical working condition data of the motor; generating a vector sequence corresponding to each phase at the target time based on the residual sequence and a preset Gaussian white noise; inputting the vector sequence and the residual sequence into a discrimination network to train a classifier, and diagnosing the working condition data of each phase of the motor through the classifier to obtain a fault diagnosis result.
[0007] In an embodiment of the present application, the dynamic parameter model includes: a recurrent neural network, a multi-head self-attention neural network, a time sampling layer, and a fully connected neural network; the working condition data sequences are respectively input into the recurrent neural network and the multi-head self-attention neural network, and the outputs of the recurrent neural network and the multi-head self-attention neural network are randomly sampled in time through the time sampling layer, and then the simulation data sequences corresponding to the sequences after time sampling are output through the fully connected neural network.
[0008] In an embodiment of the present application, the number of samples taken by the time sampling layer from the output of the recurrent neural network is greater than the number of samples taken from the output of the multi-head self-attention neural network.
[0009] In an embodiment of the present application, before generating the vector sequence corresponding to each phase at the target moment based on the residual sequence and the preset Gaussian white noise, the method further includes: using a preset time length as a sliding window and decoupling the residual sequence according to the phases of the motor to obtain a first sequence; performing trigonometric decoupling on the first sequence with the included angle between the phases of the motor as the vector included angle to obtain the vector components corresponding to the number of phases, so as to generate the vector sequence based on the vector components and the Gaussian white noise.
[0010] In an embodiment of the present application, the step of generating the vector sequence corresponding to each phase at the target moment based on the residual sequence and the preset Gaussian white noise includes: inputting the vector components and the Gaussian white noise into a high-dimensional vector generation network to obtain the vector sequence, where the high-dimensional vector generation network includes: a dynamic feature capture module, a spatio-temporal convolutional layer, a sequence decoding layer, and a vector coupling connection association layer; the dynamic feature capture module includes a plurality of neural network units equal in number to the motors and in parallel, and each neural network unit receives the vector components and the Gaussian white noise; the spatio-temporal convolutional layer is used to capture the associated features in the outputs of the neural network units to generate a spatio-temporal topology matrix sequence; the sequence decoding layer decodes the spatio-temporal topology matrix sequence to generate the generated feature seeds corresponding to the motor phases; the vector coupling connection association layer constrains the phase differences between the generated feature seeds to generate the corresponding vector sequence, so that the phase differences between the generated feature seeds corresponding to different phases satisfy the corresponding motor phase differences.
[0011] In an embodiment of the present application, the discrimination network includes a long short-term memory neural network, a fully connected layer, and a normalization layer connected in sequence.
[0012] In an embodiment of the present application, the training step of the classifier includes: inputting the output of the high-dimensional vector generation network and the working condition data sequence into the discrimination network, so as to form a cross-entropy output through the discrimination network as the training target of the high-dimensional vector generation network, and adjusting the parameter weights of the high-dimensional vector generation network through gradient descent to obtain the classifier.
[0013] In an embodiment of the present application, the cross-entropy output is determined based on the probability of the working condition data sequence having a fault, the probability of the vector sequence having a fault, the probability of the working condition data sequence having no fault, and the probability of the vector sequence having no fault.
[0014] The present application also provides a fault diagnosis system for a multiphase permanent magnet synchronous motor, including: a data comparison module, configured to obtain the operating condition data sequences corresponding to each phase of the motor, and compare the operating condition data sequences with the simulation data sequences of the corresponding phases to obtain residual sequences, wherein the simulation data sequences are output by a dynamic parameter model, and the dynamic parameter model is established based on the historical operating condition data of the motor; a vector generation module, configured to generate vector sequences corresponding to each phase at a target moment based on the residual sequences and a preset Gaussian white noise; and a fault diagnosis module, configured to input the vector sequences and the residual sequences into a discrimination network to train a classifier, and diagnose the operating condition data of each phase of the motor through the classifier to obtain a fault diagnosis result.
[0015] As described above, a fault diagnosis method and system for a multiphase permanent magnet synchronous motor proposed by the present application have the following beneficial effects.
[0016] By adding Gaussian white noise to expand the residual sequences, the residual sequences in the case of faults can be enriched, and the accuracy of the fault diagnosis classifier can be improved. Description of the Drawings
[0017] Figure 1 It is a schematic flowchart of a fault diagnosis method for a multiphase permanent magnet synchronous motor in an embodiment of the present application.
[0018] Figure 2 It is a schematic architecture diagram of a dynamic parameter model in an embodiment of the present application.
[0019] Figure 3 It is a schematic overall architecture diagram of high-dimensional vector spatio-temporal generative adversarial learning in an embodiment of the present application.
[0020] Figure 4 It is a schematic architecture diagram of a discrimination network in an embodiment of the present application.
[0021] Figure 5 It is a module diagram of a fault diagnosis system for a multiphase permanent magnet synchronous motor in an embodiment of the present application. Detailed Embodiments
[0022] The following specifically illustrates the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.
[0023] It should be noted that the illustrations provided in the following embodiments only schematically illustrate the basic concept of the present invention. Therefore, only the components related to the present invention are shown in the drawings, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.
[0024] It has been found by the inventor through research that:
[0025] Currently, the industrial community has put forward higher requirements and goals for motor control. Due to its higher control freedom, excellent fault tolerance performance, and higher power density in control, multi-phase motors have been widely used in multiple industries. The fault tolerance performance of multi-phase motors is to realize the design of fault tolerance schemes, and the fault diagnosis method can better improve the fault tolerance scheme; taking the five-phase motor as an example, the output torque of the five-phase synchronous motor is smooth, the number of phases changes from three phases of the traditional motor to five phases, the structure becomes more complex, and variables such as current, speed, position, temperature, output torque, induced electromotive force, and magnetic flux are coupled, increasing the difficulty of motor control and fault diagnosis. The five-phase motor has a complex structure and is non-linear, making it difficult to model through mechanism. Therefore, a data-driven method is proposed, such as training a fault classifier based on fault samples. However, it is difficult to construct an accurate data-driven fault diagnosis classifier with small fault samples using the traditional fault diagnosis method based on fault samples. Another technical route is to construct a dynamic parameter model of the motor under normal operating conditions through the normal operating condition data of the motor system, and then compare the output of this model (such as speed, torque, current, voltage, temperature field) with the actual motor operating conditions to determine whether the motor has a fault by comparing the differences. However, the problem with this type of method is that it is difficult to accurately determine the fault type, and it is difficult to determine the comparison difference threshold for diagnosing faults, which is prone to false alarms of faults.
[0026] Based on the above problems existing in the prior art, this application proposes a fault diagnosis method and system for a multi-phase permanent magnet synchronous motor. The technical solution of this application will be elaborated in detail below with specific embodiments.
[0027] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the fault diagnosis method for a multi-phase permanent magnet synchronous motor in an embodiment of this application. The method includes the following steps:
[0028] Step S100: Obtain the operating condition data sequences corresponding to each phase of the motor, and compare the operating condition data sequences with the simulation data sequences corresponding to the respective phases to obtain residual sequences, where the simulation data sequences are output by a dynamic parameter model, and the dynamic parameter model is established based on the historical operating condition data of the motor. Specifically, the operating conditions of the motor within a preset time period before the target time point can be collected to obtain the operating condition data sequences corresponding to each phase of the motor. Based on this operating condition data sequence, the operating condition of the motor at the next moment is predicted to obtain the corresponding simulation data sequence, and the difference between the simulation data sequence and the actual operating condition data sequence at the corresponding moment is calculated to obtain the residual sequence. Among them, the simulation data sequence can be obtained through the dynamic parameter model. First, a dynamic parameter model of the motor is established based on the historical operating condition data of the motor, and then the simulation data sequence, such as the sequences of rotational speed, torque, current, voltage, and temperature field, is output based on the established model, and the difference between this sequence and the actual motor operating condition data at the corresponding moment is calculated to obtain the residual sequence.
[0029] In one embodiment, the dynamic parameter model includes: a recurrent neural network, a multi-head self-attention neural network, a time sampling layer, and a fully connected neural network; the operating condition data sequences are respectively input into the recurrent neural network and the multi-head self-attention neural network, and the time sampling layer performs time random sampling on the outputs of the recurrent neural network and the multi-head self-attention neural network, and then the fully connected neural network outputs the simulation data sequence corresponding to the sequence after time sampling. Please refer to Figure 2 , Figure 2 which is a schematic diagram of the architecture of the dynamic parameter model in an embodiment of the present application. The dynamic parameter model is composed of a recurrent neural network (RNN), a Transformer neural network with multi-head self-attention, a time sampling layer, and a fully connected neural network output layer. The RNN is responsible for modeling the short-term dynamic process of the motor (such as the electrical rapid change fault process of current, voltage, etc.), and the Transformer is responsible for modeling the long-term dynamic process of the motor (such as the mechanical and electrical slow change fault processes reflected by torque and temperature field). The operating condition data sequences from K - T to K moments are respectively input into the RNN and the Transformer neural network, and after time random sampling by the time sampling layer, they are input into the fully connected neural network to obtain the operating condition data at the K + 1 moment. Among them, the ratio of the number of time samples output by the RNN and the Transformer is approximately 10:1, that is, 10 samples are taken from the output of the RNN and 1 sample is taken from the output of the Transformer, because the LSTM will capture the information of the rapid change process.
[0030] Step S110: Generate the vector sequences corresponding to each phase at the target moment based on the residual sequence and the preset Gaussian white noise.
[0031] In one embodiment, before generating the vector sequences corresponding to each phase at the target moment based on the residual sequence and the preset Gaussian white noise, the following steps are further included: using a preset time length as a sliding window and decoupling the residual sequence according to the phases of the motor to obtain a first sequence; performing trigonometric decoupling on the first sequence with the included angle between the phases of the motor as the vector included angle to obtain the vector components corresponding to the number of phases, so as to generate the vector sequences based on the vector components and the Gaussian white noise. Specifically, the generated residual sequence data is decoupled by phase in the form of a sliding window with a time length of T, and the first sequence is calculated by trigonometric decoupling with a vector included angle of 72 degrees (taking a five-phase motor as an example) to obtain 5 vector components of trigonometric functions. Assuming that the working condition data is S(k) = (s(k - T), …, s(k)), the specific calculation is shown in the following formula:
[0032]
[0033] Wherein:
[0034]
[0035] In one embodiment, the step of generating the vector sequences corresponding to each phase at the target moment based on the residual sequence and the preset Gaussian white noise includes: inputting the vector components and the Gaussian white noise into a high-dimensional vector generation network to obtain the vector sequences, where the high-dimensional vector generation network includes: a dynamic feature capture module, a spatio-temporal convolutional layer, a sequence decoding layer, and a vector coupling connection association layer; the dynamic feature capture module includes a plurality of neural network units equal to the number of motors and in parallel, and each neural network unit receives the vector components and the Gaussian white noise; the spatio-temporal convolutional layer is used to capture the correlation features in the outputs of the neural network units to generate a spatio-temporal topology matrix sequence; the sequence decoding layer decodes the spatio-temporal topology matrix sequence to generate the generated feature seeds corresponding to the motor phases; the vector coupling connection association layer constrains the phase differences between the generated feature seeds to generate the corresponding vector sequences, so that the phase differences between the generated feature seeds corresponding to different phases satisfy the corresponding motor phase differences.
[0036] Please refer to Figure 3 , Figure 3This is a schematic diagram of the overall architecture of high-dimensional vector spatio-temporal generative adversarial learning in an embodiment of this application. Taking a five-phase motor as an example, in the high-dimensional vector spatio-temporal generative adversarial learning architecture, 5 Transformers form a dynamic feature module, corresponding to the five phases of the motor. The high-dimensional vector spatio-temporal generative adversarial learning model G consists of a high-dimensional vector dynamic feature capture module, vector coupling connection, spatio-temporal convolution, sequence decoding generation, and a discrimination network. Among them, the dynamic feature capture module, vector coupling connection, spatio-temporal convolution, and sequence decoding generation constitute the high-dimensional vector generation network. Input the working condition data sequence S from k-T to k moment and high-dimensional Gaussian white noise into the high-dimensional vector generation network. The 5 Transformers receive the input, elaborate the sequence features and output F=(f1, f2, f3, f4, f5), and give them to the spatio-temporal convolution layer, where f1, f2, f3, f4, f5 correspond to the features of the five phases;
[0037] The spatio-temporal convolution layer captures the correlation features in the spatio-temporal topology. Assume that a spatio-temporal topology matrix sequence of phases is formed by F. The specific calculation formula is as shown:
[0038]
[0039] Input the output o of the spatio-temporal convolution layer into the sequence decoding layer. The decoding layer is a multi-head self-attention module, which generates spatio-temporal generation feature seeds s1, s2, s3, s4, s5 of the five phases of the five-phase motor, and inputs the spatio-temporal generation feature seeds into the vector coupling connection correlation layer;
[0040] The main function of the coupling correlation layer is to constrain the correlation mode of the five phases, that is, the phase difference is 72 degrees, to obtain the generated vector sequence The calculation formula of the coupling correlation layer is as follows:
[0041]
[0042] Mix the generated vector sequence into the working condition data sequence and input it into the discrimination network D,
[0043] Step S120, input the vector sequence and the residual sequence into the discrimination network to train a classifier, and use the classifier to diagnose the working condition data of each phase of the motor to obtain a fault diagnosis result. Among them, the discrimination network is a network composed of a fully connected neural network, an LSTM network, and a softmax layer. Its architecture schematic diagram is as Figure 4 shown.
[0044] In one embodiment, the training step of the classifier includes: inputting the output of the high-dimensional vector generation network and the working condition data sequence into the discrimination network, so as to form a cross-entropy output through the discrimination network as the training target of the high-dimensional vector generation network, and adjusting the parameter weights of the high-dimensional vector generation network through gradient descent to obtain the classifier. The discrimination network forms a cross-entropy output as the training target of the generation network, and trains and adjusts the parameter weights of the generation network by means of stochastic gradient descent. There are multiple discriminations in the discrimination network. Assume that a fault is represented by the subscript f and no fault is represented by the subscript n. There are four types of vector data sets: one is the real vector data V with a fault f , one is the generated vector data under a fault One is the real vector data V without a fault n , one is the generated vector data without a fault
[0045] That is, for the input vector data V, the output of the discrimination network has the mutually exclusive probabilities D(V)=(p1, p2, p3, p4) in 4 cases and p1 + p2 + p3 + p4 = 1. The calculation formula of the cross-entropy function is
[0046]
[0047] The training process is the adversarial training of D and G, and the respective goals are to maximize the objective function and minimize the objective function.
[0048] The obtained discriminator can be used as a classifier for fault diagnosis. The classifier outputs the probability distribution in 4 cases, where p1 + p2 is the detection probability of a fault, which is the detection result.
[0049] Based on the technical solution of the embodiment of the present application above, the multi-vector spatio-temporal generation adversarial method is used to generate and expand the working condition residual data of the five phases of the five-phase motor, and the vector data corresponding to the generated and expanded multi-phases are synthesized, thereby enriching the residual data sequence in the case of a fault, which helps to improve the accuracy of the fault diagnosis classifier.
[0050] Please refer to Figure 5 , Figure 5It is a module diagram of a fault diagnosis system for a multiphase permanent magnet synchronous motor in an embodiment of the present application. The system includes: a data comparison module 50, configured to obtain the operating condition data sequences corresponding to each phase of the motor, and compare the operating condition data sequences with the simulation data sequences of the corresponding phases to obtain a residual sequence, wherein the simulation data sequences are output by a dynamic parameter model, and the dynamic parameter model is established based on the historical operating condition data of the motor; a vector generation module 51, configured to generate a vector sequence corresponding to each phase at the target moment based on the residual sequence and a preset Gaussian white noise; a fault diagnosis module 52, configured to input the vector sequence and the residual sequence into a discrimination network to train a classifier, and diagnose the operating condition data of each phase of the motor through the classifier to obtain a fault diagnosis result.
[0051] The execution process of the specific system has been elaborated in detail in the foregoing method embodiment and will not be repeated here. The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.
Claims
1. A fault diagnosis method for a multiphase permanent magnet synchronous motor, characterized in that Including: Obtain the working condition data sequences corresponding to each phase of the motor, compare the working condition data sequences with the simulation data sequences of the corresponding phases, and obtain the residual sequences, where the simulation data sequences are output by a dynamic parameter model, and the dynamic parameter model is established based on the historical working condition data of the motor; Based on the residual sequences and a preset Gaussian white noise, generate vector sequences corresponding to each phase at the target moment; Input the vector sequences and the residual sequences into a discrimination network to train a classifier, and use the classifier to diagnose the working condition data of each phase of the motor to obtain a fault diagnosis result.
2. The fault diagnosis method of the multiphase permanent magnet synchronous motor according to claim 1, wherein The dynamic parameter model includes: a recurrent neural network, a multi-head self-attention neural network, a time sampling layer, and a fully connected neural network; the working condition data sequences are respectively input into the recurrent neural network and the multi-head self-attention neural network, and the time sampling layer performs time random sampling on the outputs of the recurrent neural network and the multi-head self-attention neural network, and then outputs the simulation data sequences corresponding to the sequences after time sampling through the fully connected neural network.
3. The fault diagnosis method for a multiphase permanent magnet synchronous motor according to claim 2, wherein The number of samples taken by the time sampling layer for the output of the recurrent neural network is greater than the number of samples taken for the output of the multi-head self-attention neural network.
4. The fault diagnosis method of the multiphase permanent magnet synchronous motor according to claim 1, wherein Before generating the vector sequences corresponding to each phase at the target moment based on the residual sequences and the preset Gaussian white noise, it further includes: Use a preset time length as a sliding window and decouple the residual sequences according to the phases of the motor to obtain a first sequence; Perform trigonometric decoupling on the first sequence with the included angle between the phases of the motor as the vector included angle to obtain the vector components corresponding to the number of phases, so as to generate the vector sequences based on the vector components and the Gaussian white noise.
5. The fault diagnosis method for a multiphase permanent magnet synchronous motor according to claim 4, wherein The steps of generating the vector sequences corresponding to each phase at the target moment based on the residual sequences and the preset Gaussian white noise include: Input the vector components and the Gaussian white noise into a high-dimensional vector generation network to obtain the vector sequences, where the high-dimensional vector generation network includes: a dynamic feature capture module, a spatio-temporal convolutional layer, a sequence decoding layer, and a vector coupling connection association layer; the dynamic feature capture module includes a plurality of neural network units equal to the number of motors and in parallel, and each neural network unit receives the vector components and the Gaussian white noise; the spatio-temporal convolutional layer is used to capture the associated features in the outputs of each neural network unit to generate a spatio-temporal topology matrix sequence; the sequence decoding layer decodes the spatio-temporal topology matrix sequence to generate generation feature seeds corresponding to the phases of the motor; the vector coupling connection association layer constrains the phase differences between the generation feature seeds to generate the corresponding vector sequences, so that the phase differences between the generation feature seeds corresponding to different phases satisfy the corresponding motor phase differences.
6. The fault diagnosis method for the polyphase permanent magnet synchronous motor according to claim 5, characterized in that, The discrimination network includes a long short-term memory neural network, a fully connected layer, and a normalization layer.
7. The fault diagnosis method of the polyphase permanent magnet synchronous motor according to claim 5, characterized in that, The training steps of the classifier include: Input the output of the high-dimensional vector generation network and the working condition data sequence into the discrimination network, so as to form a cross-entropy output through the discrimination network as the training target of the high-dimensional vector generation network, and adjust the parameter weights of the high-dimensional vector generation network through gradient descent to obtain the classifier.
8. The fault diagnosis method for the multiphase permanent magnet synchronous motor according to claim 7, characterized in that, The cross-entropy output is determined based on the probability of a fault existing in the working condition data sequence, the probability of a fault existing in the vector sequence, the probability of no fault in the working condition data sequence, and the probability of no fault in the vector sequence.
9. A multi-phase permanent magnet synchronous motor fault diagnosis system, characterized in that, It includes: A data comparison module, configured to obtain the working condition data sequence corresponding to each phase of the motor, and compare the working condition data sequence with the simulation data sequence of the corresponding phase to obtain a residual sequence, wherein the simulation data sequence is output by a dynamic parameter model, and the dynamic parameter model is built based on the historical working condition data of the motor; A vector generation module, configured to generate a vector sequence corresponding to each phase at the target moment based on the residual sequence and a preset Gaussian white noise; A fault diagnosis module, configured to input the vector sequence and the residual sequence into a discrimination network to train and obtain a classifier, and diagnose the working condition data of each phase of the motor through the classifier to obtain a fault diagnosis result.