Permanent magnet synchronous motor fault identification method based on customized fusion and related equipment

Abstract

The application relates to the field of motor identification, and discloses a permanent magnet synchronous motor fault identification method based on customized fusion and related equipment, which comprises the following steps: determining the fusion weight corresponding to each pixel point in a first time-frequency graph and the fusion weight corresponding to each pixel point in a second time-frequency graph according to the first time-frequency graph, the second time-frequency graph, and the signal entropy corresponding to each pixel point in the first time-frequency graph; through the customized fusion mode, the respective advantages of short-time Fourier transform and synchronous compression wavelet transform can be fully utilized, different characteristic information can be effectively extracted and fused, the capturing capacity for early weak fault characteristics can be enhanced, the distinguishing degree of fault characteristics can be significantly improved, meanwhile, the real-time characteristics of the non-stationary signal of the permanent magnet synchronous motor can be better adapted, the fault prediction model can more accurately distinguish different early faults, and therefore, the accuracy, reliability and real-time performance of the permanent magnet synchronous motor fault identification are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of motor identification, and in particular to a permanent magnet synchronous motor fault identification method based on customized fusion and related equipment. BACKGROUND

[0002] Permanent magnet synchronous motors are widely used in various key scenarios due to their high efficiency, energy saving, high power density and other significant advantages. However, due to the long-term operation of the motor in complex and variable working conditions, combined with the influence of mechanical stress, electromagnetic interference, environmental factors and other factors, various faults are inevitable. If these faults cannot be identified and processed in a timely and accurate manner, not only will the performance of the motor decrease and the operating efficiency decrease, but also serious safety accidents may occur, causing huge economic losses and casualties. Therefore, timely and accurate fault identification of permanent magnet synchronous motors is of great practical significance.

[0003] Currently, the fault identification of permanent magnet synchronous motors faces many severe challenges. On the one hand, early fault features are extremely weak and fault degrees are difficult to distinguish. In the early stage of fault, the physical changes and electrical characteristics changes inside the motor are often in the embryonic stage, and the fault feature signals exhibited are extremely low in intensity and have little difference from normal signals. Moreover, the change in early fault degree is not obvious, especially the fault features that can represent the early fault evolution trend, which have even smaller distinguishability. This makes it difficult for traditional fault detection methods to effectively extract early fault features from complex signals, and thus unable to accurately determine the occurrence and development of faults. On the other hand, traditional permanent magnet synchronous motor fault prediction models have obvious shortcomings in dealing with the real-time characteristics of non-stationary signals. The signals generated by the permanent magnet synchronous motor during operation are non-stationary, with rapidly changing frequency and amplitude over time. Traditional models are often designed and trained based on the assumption of stationary signals, and are difficult to adapt to the real-time characteristics of non-stationary signals. They cannot effectively distinguish different types of early faults, resulting in a significant reduction in the accuracy and reliability of fault identification. SUMMARY

[0004] Therefore, it is necessary to propose a permanent magnet synchronous motor fault identification method based on customized fusion and related equipment to address the above problems. This method can fully utilize the advantages of short-time Fourier transform and synchronous compression wavelet transform, effectively extract and fuse different feature information, enhance the ability to capture early weak fault features, significantly improve the distinguishability of fault features, and better adapt to the real-time characteristics of non-stationary signals of permanent magnet synchronous motors. This makes the fault prediction model more accurate in distinguishing different early faults, thereby greatly improving the accuracy, reliability and real-time performance of permanent magnet synchronous motor fault identification, and providing strong support for the safe and stable operation of the motor.

[0005] To achieve the above objectives, the present invention provides, in a first aspect, a method for fault identification of permanent magnet synchronous motors based on customized fusion, the method comprising:

[0006] Obtain the vibration signal of the permanent magnet synchronous motor;

[0007] The vibration signal is subjected to a short-time Fourier transform to obtain a first time-frequency diagram, and the vibration signal is subjected to a synchronous compressed wavelet transform to obtain a second time-frequency diagram.

[0008] Based on the vibration signal, determine the signal entropy corresponding to each pixel in the first time-frequency diagram;

[0009] Based on the first time-frequency diagram, the second time-frequency diagram, and the signal entropy corresponding to each pixel in the first time-frequency diagram, determine the fusion weight corresponding to each pixel in the first time-frequency diagram and the fusion weight corresponding to each pixel in the second time-frequency diagram;

[0010] The first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image are input into a preset permanent magnet synchronous motor fault prediction model to obtain fault classification and identification results.

[0011] Optionally, determining the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the first time-frequency map, the second time-frequency map, and the signal entropy corresponding to each pixel in the first time-frequency map includes:

[0012] Based on the first time-frequency graph, determine the energy corresponding to each pixel in the first time-frequency graph;

[0013] Based on the second time-frequency graph, determine the local frequency change rate corresponding to each pixel in the second time-frequency graph;

[0014] The energy corresponding to each pixel in the first time-frequency graph is normalized to obtain the normalized energy corresponding to each pixel in the first time-frequency graph, and the normalized local frequency change rate corresponding to each pixel in the second time-frequency graph.

[0015] Based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency image, and the normalized local frequency change rate corresponding to each pixel in the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image are determined.

[0016] Optionally, determining the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency image and the normalized local frequency change rate corresponding to each pixel in the second time-frequency image includes:

[0017] Using formula Determine the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image;

[0018] in, Let be the fusion weight corresponding to each pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. Let be the normalized energy corresponding to each pixel in the first time-frequency graph. For adjustment coefficients, Let be the signal entropy corresponding to each pixel in the first time-frequency graph. Let be the normalized local frequency change rate corresponding to each pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. The fusion weight is the weight corresponding to each pixel in the second time-frequency graph.

[0019] Optionally, the preset permanent magnet synchronous motor fault prediction model includes a two-stream convolutional neural network module and an SVM classifier connected in sequence;

[0020] The dual-stream convolutional neural network module is used to extract key features from the first time-frequency map and the second time-frequency map to obtain a first key feature vector and a second key feature vector. Based on the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map, the first key feature vector and the second key feature vector are weighted and fused to obtain a dual-stream fused feature vector.

[0021] The SVM classifier is used to classify the dual-stream fused feature vectors to obtain the fault classification and identification results.

[0022] Optionally, the step of performing weighted feature fusion of the first key feature vector and the second key feature vector based on the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image to obtain a dual-stream fused feature vector includes:

[0023] Using formula The dual-stream fusion feature vector is obtained;

[0024] in, For the r-th feature in the dual-stream fusion feature vector, Let be the fusion weight corresponding to each pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. The nth feature in the first key feature vector Let be the fusion weight corresponding to each pixel in the second time-frequency map. Let m be the coordinates of the m-th pixel in the second time-frequency graph. It is the m-th feature in the second key feature vector.

[0025] Optionally, the dual-stream convolutional neural network module includes two convolutional neural network units, a fusion layer, and a fully connected layer. Both convolutional neural network units are connected to the fusion layer, and the fusion layer is connected to the fully connected layer.

[0026] The first convolutional neural network unit is used to extract key features from the first time-frequency graph to obtain the first key feature vector.

[0027] The second convolutional neural network unit is used to extract key features from the second time-frequency graph to obtain the second key feature vector;

[0028] The fusion layer is used to perform weighted feature fusion of the first key feature vector and the second key feature vector according to the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image, so as to obtain an initial dual-stream fusion feature vector.

[0029] The fully connected layer is used to perform feature mapping on the initial dual-stream fusion feature vector to obtain the dual-stream fusion feature vector.

[0030] Optionally, each convolutional neural network unit includes an input layer, a first feature extraction block, a second feature extraction block, a first connection layer, and a second connection layer connected in sequence. The first feature extraction block and the second feature extraction block each include a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, and a Dropout layer connected in sequence.

[0031] To achieve the above objectives, the present invention provides a second aspect of a fault identification device for permanent magnet synchronous motors based on customized fusion, the device comprising:

[0032] The acquisition module is used to acquire the vibration signal of the permanent magnet synchronous motor;

[0033] The transformation module is used to perform a short-time Fourier transform on the vibration signal to obtain a first time-frequency diagram, and to perform a synchronous compressed wavelet transform on the vibration signal to obtain a second time-frequency diagram;

[0034] The signal entropy determination module is used to determine the signal entropy corresponding to each pixel in the first time-frequency diagram based on the vibration signal.

[0035] The fusion weight determination module is used to determine the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the first time-frequency map, the second time-frequency map, and the signal entropy corresponding to each pixel in the first time-frequency map.

[0036] The model prediction module is used to input the first time-frequency map, the second time-frequency map, the fusion weight corresponding to each pixel in the first time-frequency map, and the fusion weight corresponding to each pixel in the second time-frequency map into a preset permanent magnet synchronous motor fault prediction model to obtain fault classification and identification results.

[0037] To achieve the above objectives, the present invention provides, in a third aspect, a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the fault identification method for permanent magnet synchronous motors based on customized fusion as described in any one aspect.

[0038] To achieve the above objectives, the present invention provides a computer device in a fourth aspect, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the fault identification method for permanent magnet synchronous motors based on customized fusion as described in any one of the first aspects.

[0039] The present invention has the following beneficial effects: The above method acquires the vibration signal of a permanent magnet synchronous motor, performs a short-time Fourier transform on the vibration signal to obtain a first time-frequency image, and performs a synchronous compressed wavelet transform on the vibration signal to obtain a second time-frequency image. Then, based on the vibration signal, it determines the signal entropy corresponding to each pixel in the first time-frequency image. Based on the first time-frequency image, the second time-frequency image, and the signal entropy corresponding to each pixel in the first time-frequency image, it determines the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image. Finally, it combines the first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image. The fusion weights corresponding to each pixel in the frequency spectrum are input into a preset permanent magnet synchronous motor fault prediction model to obtain fault classification and identification results. That is, through this customized fusion method, the advantages of short-time Fourier transform and synchronous compressed wavelet transform can be fully utilized to effectively extract and fuse different feature information, enhance the ability to capture early weak fault features, and significantly improve the distinguishability of fault features. At the same time, this method can better adapt to the real-time characteristics of non-stationary signals of permanent magnet synchronous motors, enabling the fault prediction model to more accurately distinguish different early faults, thereby greatly improving the accuracy, reliability, and real-time performance of permanent magnet synchronous motor fault identification, and providing strong support for ensuring the safe and stable operation of the motor. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] in:

[0042] Figure 1 This is a schematic diagram of the fault identification method for permanent magnet synchronous motors based on customized fusion in the embodiments of this application;

[0043] Figure 2 This is a schematic diagram of a permanent magnet synchronous motor fault identification device based on customized fusion in an embodiment of this application.

[0044] Figure 3 This is a diagram showing the internal structure of a computer device in some embodiments. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Permanent magnet synchronous motors (PMSMs) are widely used in various critical applications due to their significant advantages such as high efficiency, energy saving, and high power density. However, because these motors operate under complex and variable conditions for extended periods, and are affected by multiple factors including mechanical stress, electromagnetic interference, and environmental influences, various faults are inevitable. If these faults are not identified and addressed promptly and accurately, they can not only lead to decreased motor performance and reduced operating efficiency, but may also cause serious safety accidents, resulting in substantial economic losses and personal injury. Therefore, timely and accurate fault identification of PMSMs is of paramount practical importance.

[0047] Currently, fault identification in permanent magnet synchronous motors (PMSMs) faces numerous severe challenges. On the one hand, early fault characteristics are extremely weak, making it difficult to distinguish fault severity. In the early stages of a fault, the physical and electrical changes within the motor are often in their nascent stages, exhibiting extremely low fault characteristic signal strength, with negligible differences from normal signals. Moreover, the changes in early fault severity are not significant, especially those fault characteristics that characterize the early fault evolution trend; their distinguishability is extremely low. This makes it difficult for traditional fault detection methods to effectively extract early fault characteristics from complex signals, thus failing to accurately determine the occurrence and development of the fault. On the other hand, traditional PMSM fault prediction models are significantly inadequate in dealing with the real-time characteristics of non-stationary signals from the motor. The signals generated by PMSMs during operation are non-stationary, with their frequency and amplitude changing rapidly over time. Traditional models are often designed and trained based on the assumption of stationary signals, making it difficult to adapt to these real-time changing non-stationary signal characteristics. This makes it impossible to effectively distinguish different types of early faults, significantly reducing the accuracy and reliability of fault identification.

[0048] To address the aforementioned issues, this application proposes a fault identification method and related equipment for permanent magnet synchronous motors based on customized fusion. This method fully leverages the advantages of both short-time Fourier transform and synchronous compressed wavelet transform to effectively extract and fuse different feature information, enhancing the ability to capture early, weak fault features and significantly improving the distinguishability of fault features. Simultaneously, it better adapts to the real-time characteristics of non-stationary signals in permanent magnet synchronous motors, enabling the fault prediction model to more accurately distinguish different early faults. This greatly improves the accuracy, reliability, and real-time performance of permanent magnet synchronous motor fault identification, providing strong support for ensuring the safe and stable operation of the motor. The specific implementation principle will be described in detail in the following embodiments.

[0049] In its first aspect, this application provides a method for fault identification of permanent magnet synchronous motors based on customized fusion.

[0050] Please see Figure 1 This is a schematic diagram of a fault identification method for permanent magnet synchronous motors based on customized fusion in an embodiment of this application. The method includes:

[0051] Step 110: Obtain the vibration signal of the permanent magnet synchronous motor.

[0052] Regarding the method of acquiring vibration signals, in some embodiments, vibration sensors can be used to acquire the vibration signals of permanent magnet synchronous motors.

[0053] Furthermore, in some embodiments, the number of vibration sensors can be multiple, and the multiple vibration sensors can correspondingly acquire multiple vibration signals.

[0054] Step 120: Perform a short-time Fourier transform on the vibration signal to obtain the first time-frequency diagram, and perform a synchronous compressed wavelet transform on the vibration signal to obtain the second time-frequency diagram.

[0055] Regarding the determination of the first and second time-frequency maps, in some embodiments, the short-time Fourier transform and synchronous compressed wavelet transform can be replaced with other similar types of time-frequency analysis techniques to transform the vibration signal and obtain the first and second time-frequency maps; for example, the short-time Fourier transform can be replaced with Fourier transform, Hilbert-Huang transform, etc., and the synchronous compressed wavelet transform can be replaced with wavelet transform, continuous wavelet transform, etc.

[0056] Step 130: Determine the signal entropy corresponding to each pixel in the first time-frequency diagram based on the vibration signal.

[0057] Regarding the type of signal entropy, in some embodiments, the type of signal entropy includes, but is not limited to, Shannon entropy, sample entropy, etc. This application preferably uses Shannon entropy as the signal entropy.

[0058] It should be noted that, since the horizontal and vertical coordinates of the pixels in the first time-frequency graph are time and frequency, respectively, and the vibration signal is a signal that changes with time, in some embodiments, the signal entropy corresponding to each time can be determined based on the vibration signal, and matched with the time corresponding to the pixels in the first time-frequency graph to determine the signal entropy corresponding to each pixel in the first time-frequency graph; wherein, the signal entropy corresponding to each time can be the signal entropy for each moment, or the signal entropy for each time window.

[0059] Step 140: Based on the first time-frequency diagram, the second time-frequency diagram, and the signal entropy corresponding to each pixel in the first time-frequency diagram, determine the fusion weight corresponding to each pixel in the first time-frequency diagram and the fusion weight corresponding to each pixel in the second time-frequency diagram.

[0060] It should be noted that, since the horizontal and vertical coordinates of the pixels in the first and second time-frequency maps are time and frequency, respectively, in some embodiments, the frequency corresponding to each pixel in the first time-frequency map can be determined based on the first time-frequency map, and the frequency corresponding to each pixel in the second time-frequency map can be determined based on the second time-frequency map. Then, based on the frequency and signal entropy corresponding to each pixel in the first time-frequency map, and the frequency corresponding to each pixel in the second time-frequency map, the fusion weight corresponding to each pixel in the first time-frequency map is determined. Finally, the difference between 1 and the fusion weight corresponding to each pixel in the first time-frequency map is used as the fusion weight corresponding to each pixel in the second time-frequency map.

[0061] Step 150: Input the first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification results.

[0062] Here, the preset permanent magnet synchronous motor fault prediction model refers to a pre-trained model used to predict the output fault classification and identification results based on the input first time-frequency map, second time-frequency map, the fusion weight corresponding to each pixel in the first time-frequency map, and the fusion weight corresponding to each pixel in the second time-frequency map.

[0063] Regarding the acquisition method of the preset permanent magnet synchronous motor fault prediction model, in some embodiments, it can be obtained by the training method of the model in the prior art, which will not be elaborated here.

[0064] In some embodiments, the preset permanent magnet synchronous motor fault prediction model includes a dual-stream convolutional neural network module and an SVM classifier connected in sequence. The dual-stream convolutional neural network module is used to extract key features from the first time-frequency map and the second time-frequency map to obtain a first key feature vector and a second key feature vector. Based on the fusion weights corresponding to each pixel in the first time-frequency map and the fusion weights corresponding to each pixel in the second time-frequency map, the first key feature vector and the second key feature vector are weighted and fused to obtain a dual-stream fused feature vector. The SVM classifier is used to classify the dual-stream fused feature vector to obtain the fault classification and identification result.

[0065] It should be noted that, since the preset permanent magnet synchronous motor fault prediction model does not include a quantum heuristic neural network module, the data input to the preset permanent magnet synchronous motor fault prediction model is the first time-frequency plot and the second time-frequency plot. If the preset permanent magnet synchronous motor fault prediction model includes a quantum heuristic neural network module, then the data input to the preset permanent magnet synchronous motor fault prediction model can be the first global quantum state tensor and the second global quantum state tensor (or the first global customized quantum state tensor and the second global customized quantum state tensor), that is:

[0066] In other embodiments, when the preset permanent magnet synchronous motor fault prediction model includes a quantum heuristic neural network module, a two-stream convolutional neural network module, and an SVM classifier connected in sequence, the first global quantum state tensor and the second global quantum state tensor (or the first global customized quantum state tensor and the second global customized quantum state tensor), the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image are input into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification result. The quantum heuristic neural network module is used to perform rotation and entanglement operations on both the first global quantum state tensor and the second global quantum state tensor (or the first global customized quantum state tensor and the second global customized quantum state tensor) to obtain a first enhanced feature vector and a second enhanced feature vector. The two-stream convolutional neural network module is used to extract and fuse features from the first enhanced feature vector and the second enhanced feature vector to obtain a two-stream fused feature vector. The SVM classifier is used to classify the two-stream fused feature vector to obtain the fault classification and identification result.

[0067] It should be noted that the data input to the preset permanent magnet synchronous motor fault prediction model can be either the first global quantum state tensor and the second global quantum state tensor, or the first global customized quantum state tensor and the second global customized quantum state tensor. To avoid redundancy, this application will uniformly adopt the first global quantum state tensor and the second global quantum state tensor as the input data. However, when the input is the first global customized quantum state tensor and the second global customized quantum state tensor, the first global customized quantum state tensor can be used as the first global quantum state tensor, and the second global customized quantum state tensor can be used as the first global quantum state tensor, thereby avoiding redundancy.

[0068] In some embodiments, the determination of the first global quantum state tensor and the second global quantum state tensor can be achieved by mapping and encoding the first time-frequency graph to obtain the first global quantum state tensor, and by mapping and encoding the second time-frequency graph to obtain the second global quantum state tensor.

[0069] Furthermore, regarding the determination of the first global quantum state tensor and the second global quantum state tensor, in some embodiments, each pixel in the first time-frequency diagram can be mapped and encoded to obtain the quantum state corresponding to each pixel in the first time-frequency diagram, and each pixel in the second time-frequency diagram can be mapped and encoded to obtain the quantum state corresponding to each pixel in the second time-frequency diagram. Then, based on the quantum states corresponding to all pixels in the first time-frequency diagram, the first global quantum state tensor is determined, and based on the quantum states corresponding to all pixels in the second time-frequency diagram, the second global quantum state tensor is determined; wherein the horizontal and vertical coordinates of the pixels are time and frequency, respectively.

[0070] Furthermore, in some embodiments, the quantum states corresponding to all pixels in the first time-frequency image can be tensor-producted sequentially according to the coordinate order of the pixels to obtain the first global quantum state tensor, and the quantum states corresponding to all pixels in the second time-frequency image can be tensor-producted sequentially to obtain the second global quantum state tensor.

[0071] In this application, by mapping and encoding quantum states, pixels in the time-frequency graph are mapped to quantum states and a global quantum state tensor is constructed. This fully utilizes the unique properties of quantum states, enabling more effective extraction and characterization of weak features of early faults in permanent magnet synchronous motors. This significantly improves the detection capability for complex nonlinear fault modes, thereby enhancing the detection sensitivity of early weak faults and providing a stronger guarantee for the reliable operation of permanent magnet synchronous motors.

[0072] In some embodiments, the amplitude corresponding to each pixel in the first time-frequency image can be normalized to obtain the normalized amplitude corresponding to each pixel in the first time-frequency image, and the amplitude corresponding to each pixel in the second time-frequency image can be normalized to obtain the normalized amplitude corresponding to each pixel in the second time-frequency image. Then, based on the normalized amplitude and phase corresponding to each pixel in the first time-frequency image, the quantum state corresponding to each pixel in the first time-frequency image is determined, and based on the normalized amplitude and phase corresponding to each pixel in the second time-frequency image, the quantum state corresponding to each pixel in the second time-frequency image is determined.

[0073] In this application, the ability to characterize fault features of quantum states and the robustness of the encoding are significantly improved by using amplitude normalization and phase fusion encoding.

[0074] Furthermore, in some embodiments, the quantum state corresponding to each pixel in the first time-frequency graph can be determined according to the following formula:

[0075] ;

[0076] The quantum state corresponding to each pixel in the second time-frequency graph is determined using the following formula:

[0077] ;

[0078] in, This represents the quantum state corresponding to the nth pixel in the first time-frequency diagram. Let n be the coordinates of the nth pixel in the first time-frequency graph. This represents the normalized amplitude corresponding to the nth pixel in the first time-frequency graph. It is a natural constant. The imaginary unit, This represents the phase corresponding to the nth pixel in the first time-frequency diagram. and Let be the orthogonal basis vectors of the two-dimensional Hilbert space in quantum state encoding. Let m be the quantum state corresponding to the m-th pixel in the second time-frequency diagram. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This represents the normalized amplitude corresponding to the m-th pixel in the second time-frequency graph. This represents the phase corresponding to the m-th pixel in the second time-frequency diagram.

[0079] In this application, by normalizing the amplitude and phase parameters and using a combined quantum state encoding formula, the quantum characterization capability and pattern recognition accuracy of fault features are significantly improved.

[0080] In some embodiments, the first global quantum state tensor can be determined according to the following formula:

[0081] ;

[0082] The second global quantum state tensor is determined according to the following formula:

[0083] ;

[0084] in, This is the first global quantum state tensor. The symbol for tensor product is... Let N be the coordinates of the Nth pixel in the first time-frequency graph. This represents the total number of pixels in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. This represents the quantum state corresponding to the nth pixel in the first time-frequency diagram. This is the second global quantum state tensor. Here are the coordinates of the Mth pixel in the second time-frequency graph. This represents the total number of pixels in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This represents the quantum state corresponding to the m-th pixel in the second time-frequency diagram.

[0085] In this embodiment, a global quantum state tensor is constructed through tensor product, thereby enabling high-order correlation analysis and mode decoupling of fault characteristics.

[0086] It should be noted that the mapping encoding in the above embodiments uses quantum state mapping encoding. In other embodiments, customized quantum state mapping encoding can also be used, that is:

[0087] In other embodiments, the method may involve acquiring the rotational speed signal of the permanent magnet synchronous motor, determining the rotational speed corresponding to each pixel in the first time-frequency image and the rotational speed corresponding to each pixel in the second time-frequency image based on the rotational speed corresponding to each pixel in the first time-frequency image, mapping and encoding each pixel in the first time-frequency image to obtain a custom quantum state corresponding to each pixel in the first time-frequency image, and mapping and encoding each pixel in the second time-frequency image based on the rotational speed corresponding to each pixel in the second time-frequency image to obtain a custom quantum state corresponding to each pixel in the second time-frequency image, determining a first global custom quantum state tensor based on the custom quantum states corresponding to all pixels in the first time-frequency image, and determining a second global custom quantum state tensor based on the custom quantum states corresponding to all pixels in the second time-frequency image, and finally using the first global custom quantum state tensor as the first global quantum state tensor and the second custom quantum state tensor as the second global quantum state tensor; wherein the horizontal and vertical coordinates of the pixel are time and frequency, respectively.

[0088] Furthermore, similar to the method for determining the global quantum state tensor by mapping and encoding quantum states, in some embodiments, the custom quantum states corresponding to all pixels in the first time-frequency image can be sequentially multiplied by tensor according to the coordinate order of the pixels to obtain the first global custom quantum state tensor, and the custom quantum states corresponding to all pixels in the second time-frequency image can be sequentially multiplied by tensor to obtain the second global custom quantum state tensor.

[0089] It should be noted that, since the horizontal and vertical coordinates of the pixels in the first and second time-frequency maps are time and frequency, respectively, and the rotational speed signal is a signal that changes with time, in some embodiments, the time in the rotational speed signal can be matched with the time corresponding to the pixels in the first and second time-frequency maps to determine the rotational speed corresponding to each pixel in the first time-frequency map and the rotational speed corresponding to each pixel in the second time-frequency map.

[0090] It should also be noted that in practical applications, the rotational speed of a permanent magnet synchronous motor, as a dynamically changing parameter, has a significant impact on fault characteristics and can limit the accuracy and reliability of fault identification. Therefore, this application introduces the rotational speed corresponding to the pixel in the time-frequency diagram when mapping and encoding the pixel. That is, by combining the rotational speed corresponding to the pixel in the time-frequency diagram, the pixel in the time-frequency diagram is mapped and encoded, and a quantum state related to the rotational speed is customized for the pixel in the time-frequency diagram. This fully considers the influence of rotational speed on fault characteristics and can further improve the accuracy and reliability of fault identification.

[0091] In this application, by mapping and encoding such customized quantum states, quantum states related to rotational speed are customized for pixels in the time-frequency graph, and a globally customized quantum state tensor is constructed. This fully considers the influence of rotational speed on fault characteristics, and can more accurately extract and characterize early weak fault features related to rotational speed. This effectively improves the fault prediction model's ability to detect complex nonlinear fault modes, significantly improves the detection sensitivity of early weak faults and the accuracy of fault identification, and provides a more reliable guarantee for the safe and stable operation of permanent magnet synchronous motors.

[0092] Regarding the method for determining the custom quantum state, in some embodiments, a first maximum rotational speed can be determined based on the rotational speeds corresponding to all pixels in the first time-frequency image, and a second maximum rotational speed can be determined based on the rotational speeds corresponding to all pixels in the second time-frequency image; a rotational speed modulation weight corresponding to each pixel in the first time-frequency image can be determined based on the rotational speeds corresponding to each pixel in the first time-frequency image and the first maximum rotational speed, and a rotational speed modulation weight corresponding to each pixel in the second time-frequency image can be determined based on the rotational speeds corresponding to each pixel in the second time-frequency image and the second maximum rotational speed; each pixel in the first time-frequency image is mapped and encoded based on the rotational speed modulation weights corresponding to each pixel in the first time-frequency image to obtain the custom quantum state corresponding to each pixel in the first time-frequency image, and the same applies to the second time-frequency image.

[0093] Furthermore, in some embodiments, the quotient between the rotational speed corresponding to each pixel in the time-frequency graph and the maximum rotational speed can be used as the rotational speed modulation weight corresponding to each pixel in the time-frequency graph.

[0094] In this application, by optimizing the quantum state encoding through rotational speed modulation weights, the extraction accuracy of early weak fault features and the ability to distinguish fault modes are significantly improved.

[0095] Regarding the method for determining the rotational speed modulation weight, in some embodiments, the frequency corresponding to each pixel in the first time-frequency image and the frequency corresponding to each pixel in the second time-frequency image can be obtained; the rotational speed modulation weight corresponding to each pixel in the first time-frequency image can be determined based on the frequency and rotational speed corresponding to each pixel in the first time-frequency image and the first maximum rotational speed; and the rotational speed modulation weight corresponding to each pixel in the second time-frequency image can be determined based on the frequency and rotational speed corresponding to each pixel in the second time-frequency image and the second maximum rotational speed.

[0096] In this application, by integrating frequency and rotation speed information to optimize weight calculation, the ability of customized quantum state coding to characterize complex fault features is significantly improved, and the robustness of fault mode recognition is enhanced.

[0097] Furthermore, in some embodiments, the rotational speed modulation weight corresponding to each pixel in the first time-frequency graph can be determined according to the following formula:

[0098] ;

[0099] The rotational speed modulation weight corresponding to each pixel in the second time-frequency graph is determined according to the following formula:

[0100] ;

[0101] in, This represents the rotational speed modulation weight corresponding to the nth pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. The rotational speed corresponds to the nth pixel in the first time-frequency graph. The first maximum speed, It is a sine function. Pi This represents the frequency corresponding to the nth pixel in the first time-frequency graph. This represents the rotational speed modulation weight corresponding to the m-th pixel in the second time-frequency diagram. Let m be the coordinates of the m-th pixel in the second time-frequency graph. Let m be the rotational speed corresponding to the m-th pixel in the second time-frequency graph. The second maximum speed, This is the frequency corresponding to the m-th pixel in the second time-frequency diagram.

[0102] In some embodiments, for It can also be represented as the ordinate of the nth pixel in the first time-frequency graph. It can also be represented as the ordinate of the m-th pixel in the second time-frequency graph.

[0103] In this application, by introducing a sine function to construct a frequency-speed joint modulation weighting formula, the analytical capability and diagnostic robustness of the customized quantum state coding for complex fault characteristics are significantly improved.

[0104] Furthermore, in some embodiments, the amplitude corresponding to each pixel in the first time-frequency image can be normalized to obtain the normalized amplitude corresponding to each pixel in the first time-frequency image, and the amplitude corresponding to each pixel in the second time-frequency image can be normalized to obtain the normalized amplitude corresponding to each pixel in the second time-frequency image; the customized quantum state corresponding to each pixel in the first time-frequency image is determined based on the rotational speed modulation weight, normalized amplitude, and phase corresponding to each pixel in the first time-frequency image, and the customized quantum state corresponding to each pixel in the second time-frequency image is determined based on the rotational speed modulation weight, normalized amplitude, and phase corresponding to each pixel in the second time-frequency image.

[0105] In this application, by using rotation speed modulation weighting, amplitude normalization and phase fusion coding, the fault characteristic characterization capability and coding robustness of customized quantum states are significantly improved.

[0106] Furthermore, the custom quantum state corresponding to each pixel in the first time-frequency graph can be determined using the following formula:

[0107] ;

[0108] The custom quantum state corresponding to each pixel in the second time-frequency graph is determined according to the following formula:

[0109] ;

[0110] in, This represents the custom quantum state corresponding to the nth pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. This represents the rotational speed modulation weight corresponding to the nth pixel in the first time-frequency graph. This represents the normalized amplitude corresponding to the nth pixel in the first time-frequency graph. It is a natural constant. The imaginary unit, This represents the phase corresponding to the nth pixel in the first time-frequency diagram. and Let be the orthogonal basis vectors of the two-dimensional Hilbert space in quantum state encoding. This represents the custom quantum state corresponding to the m-th pixel in the second time-frequency diagram. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This represents the rotational speed modulation weight corresponding to the m-th pixel in the second time-frequency diagram. This represents the normalized amplitude corresponding to the m-th pixel in the second time-frequency graph. This represents the phase corresponding to the m-th pixel in the second time-frequency diagram.

[0111] In this application, a quantum state encoding formula is customized by jointly using three parameters: rotation speed modulation weight, normalized amplitude, and phase, which significantly improves the quantum characterization capability and pattern recognition accuracy of fault features.

[0112] In some embodiments, the first globally customized quantum state tensor can be determined according to the following formula:

[0113] ;

[0114] The second globally customized quantum state tensor is determined according to the following formula:

[0115] ;

[0116] in, Customize the first global quantum state tensor. The symbol for tensor product is... Let N be the coordinates of the Nth pixel in the first time-frequency graph. This represents the total number of pixels in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. This represents the custom quantum state corresponding to the nth pixel in the first time-frequency graph. Customize the second global quantum state tensor. Here are the coordinates of the Mth pixel in the second time-frequency graph. This represents the total number of pixels in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This is the custom quantum state corresponding to the m-th pixel in the second time-frequency diagram.

[0117] In this embodiment, a globally customized quantum state tensor is constructed through tensor product, thereby realizing high-order correlation analysis and mode decoupling of fault characteristics.

[0118] It should be noted that, since this application uses the first global customized quantum state tensor as the first global quantum state tensor and the second customized quantum state tensor as the second global quantum state tensor when employing the mapping encoding of customized quantum states, the first global quantum state tensor and the second global quantum state tensor can be obtained by quantum state mapping encoding or by customized quantum states. In the subsequent description of embodiments, if the processing methods of the first global quantum state tensor and the second global quantum state tensor obtained by quantum state mapping encoding differ from those obtained by customized quantum state mapping encoding in the preset permanent magnet synchronous motor fault prediction model, different embodiments can be used under the conditions of quantum state mapping encoding or customized quantum state mapping encoding for special explanation. Of course, if the processing methods are the same, no further special explanation will be given to avoid redundancy.

[0119] Regarding the determination of the first and second enhanced feature vectors, in some embodiments, the rotation matrix corresponding to each pixel in the first time-frequency image and the rotation matrix corresponding to each pixel in the second time-frequency image can be obtained. Using the rotation matrix corresponding to the nth pixel in the first time-frequency image, the nth quantum state in the first global quantum state tensor is rotated to obtain the nth rotated quantum state in the third global quantum state tensor. Similarly, using the rotation matrix corresponding to the mth pixel in the second time-frequency image, the mth quantum state in the second global quantum state tensor is rotated to obtain the mth rotated quantum state in the fourth global quantum state tensor. A convolution operation is used to simulate entanglement, and a cross-time and frequency entanglement operation is performed on the nth rotated quantum state in the third global quantum state tensor and the remaining rotated quantum states to obtain the first correlation between the nth rotated quantum state in the third global quantum state tensor and the remaining rotated quantum states, and the first correlation between the nth rotated quantum state in the third global quantum state tensor and the remaining rotated quantum states. The m-th rotated quantum state in the quantum state is entangled with the remaining rotated quantum states across time and frequency to obtain the second correlation between the m-th rotated quantum state and the remaining rotated quantum states in the fourth global quantum state tensor. Here, n takes successively greater than 0 integer values ​​until it equals the total number of pixels in the first time-frequency map, resulting in the third global quantum state tensor and the first correlation vector composed of all first correlations. Similarly, m takes successively greater than 0 integer values ​​until it equals the total number of pixels in the second time-frequency map, resulting in the fourth global quantum state tensor and the second correlation vector composed of all second correlations. The third global quantum state tensor and the first correlation vector are fused and concatenated to obtain the first concatenated feature vector. The fourth global quantum state tensor and the second correlation feature vector are fused and concatenated to obtain the second concatenated feature vector. The first concatenated feature vector is mapped and decoded to obtain the first enhanced feature vector. The second concatenated feature vector is mapped and decoded to obtain the second enhanced feature vector.

[0120] It should be noted that, regarding mapping decoding, the opposite of mapping encoding, that is, in some embodiments, during mapping decoding, the... and The coefficients are reassembled into a new time-frequency diagram, and the new time-frequency diagram is the spliced ​​feature vector.

[0121] Furthermore, regarding the determination method of the concatenated feature vector, in some embodiments, a fusion concatenation method can be used to fuse and concatenate the quantum states in the global quantum state tensor with the correlations in the correlation vector in a one-to-one correspondence to obtain the concatenated feature vector; of course, in other embodiments, the learnable weights obtained from pre-training can also be obtained, and then weighted fusion concatenation can be used to fuse and concatenate the quantum states in the global quantum state tensor with the correlations in the correlation vector in a one-to-one correspondence according to the learnable weights to obtain the concatenated feature vector; wherein, the value range of the learnable weights is [0, 1], the weighting of the quantum states in the global quantum state tensor is the learnable weight, and the correlation in the correlation vector is the difference between 1 and the learnable weight.

[0122] In this application, the ability to express fault features is significantly enhanced by the rotation and entanglement operations of quantum states. Combined with correlation analysis and fusion splicing strategies, high-sensitivity extraction and accurate classification of early weak fault features are achieved.

[0123] In some embodiments, the rotational speed signal of a permanent magnet synchronous motor can be obtained. Based on the rotational speed signal, the rotational speed corresponding to each pixel in the first time-frequency image and the rotational speed corresponding to each pixel in the second time-frequency image are determined. Based on the vibration signal, the signal entropy corresponding to each pixel in the first time-frequency image and the signal entropy corresponding to each pixel in the second time-frequency image are determined. Based on the energy corresponding to all pixels in the first time-frequency image, a first maximum energy is determined, and based on the energy corresponding to all pixels in the second time-frequency image, a second maximum energy is determined. Based on the rotational speed, signal entropy, and energy corresponding to each pixel in the first time-frequency image, and the first maximum energy, the rotational angle corresponding to each pixel in the first time-frequency image is determined, and based on the rotational speed, signal entropy, and energy corresponding to each pixel in the second time-frequency image, and the second maximum energy, the rotational angle corresponding to each pixel in the second time-frequency image is determined. Based on the rotational angle corresponding to each pixel in the first time-frequency image, a rotational matrix corresponding to each pixel in the first time-frequency image is constructed, and based on the rotational angle corresponding to each pixel in the second time-frequency image, a rotational matrix corresponding to each pixel in the second time-frequency image is constructed.

[0124] In this application, a dynamic rotation matrix is ​​constructed by fusing rotation speed, signal entropy, and energy, which significantly improves the fault feature enhancement capability of quantum state rotation operation and achieves high-sensitivity capture of weak fault features under complex working conditions.

[0125] It should be noted that the embodiment here uses a method for determining the rotation matrix under the condition of quantum state mapping encoding. If a method for determining the rotation matrix under the condition of customized quantum state mapping encoding is used, the following embodiment will be adopted:

[0126] In other embodiments, the rotational speed signal of the permanent magnet synchronous motor can be acquired. Based on the rotational speed signal, the rotational speed corresponding to each pixel in the first time-frequency image and the rotational speed corresponding to each pixel in the second time-frequency image can be determined. Then, based on the rotational speed corresponding to each pixel in the first time-frequency image, the rotational angle corresponding to each pixel in the first time-frequency image can be determined. Finally, based on the rotational angle corresponding to each pixel in the first time-frequency image, a rotation matrix corresponding to each pixel in the first time-frequency image can be constructed. And based on the rotational angle corresponding to each pixel in the second time-frequency image, a rotation matrix corresponding to each pixel in the second time-frequency image can be constructed.

[0127] In this application, a dynamic rotation matrix is ​​constructed by rotating the rotation speed, which significantly improves the fault feature enhancement capability of customized quantum state rotation operation and achieves high-sensitivity capture of weak fault features under complex working conditions.

[0128] Furthermore, in some embodiments, the rotation angle corresponding to each pixel in the first time-frequency image can be determined according to the following formula:

[0129] ;

[0130] The rotation angle corresponding to each pixel in the second time-frequency graph is determined according to the following formula;

[0131] ;

[0132] in, Let be the rotation angle corresponding to the nth pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. For speed coefficient, The rotational speed corresponds to the nth pixel in the first time-frequency graph. The signal entropy coefficient, Let be the signal entropy corresponding to the nth pixel in the first time-frequency graph. The energy coefficient, This represents the energy corresponding to the nth pixel in the first time-frequency graph. The first and greatest energy, Let be the rotation angle corresponding to the m-th pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. Let m be the rotational speed corresponding to the m-th pixel in the second time-frequency graph. Let be the signal entropy corresponding to the m-th pixel in the second time-frequency diagram. Let m be the energy corresponding to the m-th pixel in the second time-frequency graph. It is the second largest energy.

[0133] The values ​​of the speed coefficient, signal entropy coefficient, and energy coefficient can be obtained and preset by the operator based on extensive experience, experiments, or statistics. Preferably, the speed coefficient is set to 0.4, the signal entropy coefficient is set to 0.3, and the energy coefficient is set to 0.3.

[0134] In this application, the rotation angle is dynamically calculated by fusing rotation speed, signal entropy and energy to construct a rotation matrix, which significantly improves the fault feature enhancement capability of quantum state rotation operation and achieves high sensitivity capture and accurate classification of weak fault features under complex working conditions.

[0135] It should be noted that the embodiment here uses the rotation angle calculation formula under the condition of quantum state mapping encoding. If the rotation angle calculation formula under the condition of customized quantum state mapping encoding is used, the following embodiment will be used, namely:

[0136] In other embodiments, the rotation angle corresponding to each pixel in the first time-frequency image can also be determined according to the following formula:

[0137] ;

[0138] The rotation angle corresponding to each pixel in the second time-frequency graph is determined using the following formula:

[0139] ;

[0140] in, To customize the speed coefficient.

[0141] The value of the customized speed coefficient can be obtained and preset by the operator based on a large amount of experience, experiments or statistics; preferably, the customized speed coefficient is set to 0.7 in this application.

[0142] In this application, the rotation angle is dynamically calculated by rotation speed to construct a rotation matrix, which significantly improves the fault feature enhancement capability of customized quantum state rotation operation and achieves high-sensitivity capture and accurate classification of weak fault features under complex working conditions.

[0143] Before constructing the rotation matrix, in some embodiments, the method further includes: in the first time-frequency image and the second time-frequency image, if there is a pixel whose rotation angle does not meet the preset rotation angle range, then according to the rotation angle corresponding to the corresponding pixel and the preset rotation angle range, clamp the rotation angle corresponding to the corresponding pixel to the preset rotation angle range to obtain the clamped rotation angle corresponding to the corresponding pixel; and use the clamped rotation angle corresponding to the corresponding pixel as the rotation angle corresponding to the corresponding pixel again.

[0144] The preset rotation angle range can be obtained and set in advance by the operator based on a large amount of experience, experiments or statistics. Of course, it can also be set in advance by the operator according to actual needs. There is no limitation here.

[0145] Regarding the value of the preset rotation angle range, in some embodiments, this application preferably sets the preset rotation angle range to [0, π / 2].

[0146] In this application, a rotation angle clamping mechanism is used to ensure the stability of the rotation matrix construction, avoid the interference of extreme angle values ​​on quantum state rotation operations, and improve the convergence of model training and the reliability of fault feature enhancement.

[0147] Furthermore, regarding the method for determining the clamped rotation angle, in some embodiments, the rotation angle corresponding to the corresponding pixel can be clamped to the preset rotation angle range based on the rotation angle corresponding to the corresponding pixel and the preset rotation angle range to obtain the clamped rotation angle corresponding to the corresponding pixel. This includes: if the rotation angle corresponding to the corresponding pixel is less than the lower limit of the preset rotation angle range, the lower limit of the preset rotation angle range is used as the clamped rotation angle corresponding to the corresponding pixel; if the rotation angle corresponding to the corresponding pixel is less than the upper limit of the preset rotation angle range, the upper limit of the preset rotation angle range is used as the clamped rotation angle corresponding to the corresponding pixel.

[0148] In this application, by taking upper and lower limits to transform extreme rotation angles to the standard range, the interference of abnormal angle values ​​on quantum state rotation operations is effectively eliminated, significantly improving the stability of model training and the reliability of feature enhancement.

[0149] Before constructing the rotation matrix, in some other embodiments, the method further includes: in the first time-frequency image and the second time-frequency image, if there is a pixel whose rotation angle does not meet the preset rotation angle range, then according to the rotation angle corresponding to the corresponding pixel, the rotation angle corresponding to the corresponding pixel is transformed to the preset rotation angle range to obtain the transformed rotation angle corresponding to the corresponding pixel; and the transformed rotation angle corresponding to the corresponding pixel is used again as the rotation angle corresponding to the corresponding pixel.

[0150] In this application, the stability of the rotation matrix construction is ensured by the rotation angle transformation mechanism, avoiding the interference of extreme angle values ​​on the quantum state rotation operation, while improving the convergence of model training and the reliability of fault feature enhancement.

[0151] In one feasible implementation, the transformed rotation angle corresponding to the respective pixel is obtained according to the following formula:

[0152] ;

[0153] in, This represents the transformed rotation angle corresponding to the respective pixel. This represents the rotation angle corresponding to the respective pixel. The remainder symbol is used. Pi is the mathematical constant of a circle.

[0154] In this application, the formula transforms extreme rotation angles to the standard range by taking the remainder, effectively eliminating the interference of abnormal angle values ​​on quantum state rotation operations, and significantly improving the stability of model training and the reliability of feature enhancement.

[0155] Furthermore, regarding the method for determining the transformed rotation angle, in some embodiments, the nth rotated quantum state in the third global quantum state tensor can be obtained according to the following formula:

[0156] ;

[0157] The m-th rotated quantum state in the fourth global quantum state tensor is obtained using the following formula:

[0158] ;

[0159] in, This represents the nth rotated quantum state in the third global quantum state tensor. Let n be the coordinates of the nth pixel in the first time-frequency graph. Let be the rotation matrix corresponding to the nth pixel in the first time-frequency graph. This represents the quantum state corresponding to the nth pixel in the first time-frequency diagram. It is a cosine function. Let be the rotation angle corresponding to the nth pixel in the first time-frequency graph. It is a sine function. This represents the m-th rotated quantum state in the fourth global quantum state tensor. Let m be the coordinates of the m-th pixel in the second time-frequency graph. Let be the rotation matrix corresponding to the m-th pixel in the second time-frequency graph. Let m be the quantum state corresponding to the m-th pixel in the second time-frequency diagram. The rotation angle corresponds to the m-th pixel in the second time-frequency graph.

[0160] In this application, the dynamic enhancement and phase calibration of fault characteristics are achieved through the quantum state rotation formula, which significantly improves the detection sensitivity and feature discrimination of early weak faults.

[0161] Regarding the determination of the rotated quantum state, in some embodiments, the first correlation between the nth rotated quantum state and the remaining rotated quantum states in the third global quantum state tensor can be obtained according to the following formula:

[0162] ;

[0163] The second correlation between the m-th rotated quantum state and the remaining rotated quantum states in the fourth global quantum state tensor is obtained using the following formula:

[0164] ;

[0165] in, This represents the first correlation between the nth rotated quantum state and the remaining rotated quantum states in the third global quantum state tensor. Let N be the coordinates of the Nth pixel in the first time-frequency graph. This represents the total number of pixels in the first time-frequency graph. Let i be the coordinates of the i-th pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. For the kernel function of the entanglement operation, Let be the quantum state inner product between the nth rotated quantum state and the ith rotated quantum state in the third global quantum state tensor. Let n be the dual state of the nth rotated quantum state in the third global quantum state tensor. Let i be the i-th rotated quantum state in the third global quantum state tensor. This represents the second correlation between the m-th rotated quantum state and the remaining rotated quantum states in the fourth global quantum state tensor. Here are the coordinates of the Mth pixel in the second time-frequency graph. This represents the total number of pixels in the second time-frequency graph. Let j be the coordinates of the j-th pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. Let be the quantum state inner product between the m-th rotated quantum state and the j-th rotated quantum state in the fourth global quantum state tensor. Let m be the dual state of the m-th rotated quantum state in the fourth global quantum state tensor. Let j be the j-th rotated quantum state in the fourth global quantum state tensor.

[0166] In this application, by introducing the entanglement operation of the kernel function and the inner product of the quantum state, high-order correlation modeling across the time and frequency dimensions is realized, which significantly improves the feature extraction capability and classification accuracy of complex nonlinear fault modes.

[0167] In this embodiment, this customized fusion method can fully utilize the advantages of both short-time Fourier transform and synchronous compressed wavelet transform, effectively extract and fuse different feature information, enhance the ability to capture early weak fault features, and significantly improve the distinguishability of fault features. At the same time, this method can better adapt to the real-time characteristics of non-stationary signals of permanent magnet synchronous motors, enabling the fault prediction model to more accurately distinguish different early faults. This greatly improves the accuracy, reliability, and real-time performance of fault identification for permanent magnet synchronous motors, providing strong support for ensuring the safe and stable operation of the motor.

[0168] In addition to the effects mentioned above, the custom-fusion-based permanent magnet synchronous motor fault identification method and related equipment also have the following advantages: Compatibility with multiple time-frequency analysis techniques: When acquiring the first and second time-frequency maps, the method mentions that short-time Fourier transform and synchronous compressed wavelet transform can be replaced with other similar types of time-frequency analysis techniques, such as replacing short-time Fourier transform with Fourier transform, Hilbert-Huang transform, etc., and replacing synchronous compressed wavelet transform with wavelet transform, continuous wavelet transform, etc. This compatibility allows the system to flexibly select the most suitable time-frequency analysis method according to different application scenarios and needs, improving the system's adaptability and versatility; Expandable model structure: The preset permanent magnet synchronous motor fault prediction model includes different... The modular design, such as including only a dual-stream convolutional neural network module and an SVM classifier, or adding a quantum heuristic neural network module, allows the model structure to be expanded and adjusted according to the complexity and accuracy requirements of actual fault identification, facilitating subsequent model upgrades and optimizations. Multi-dimensional feature fusion: This method not only utilizes the time-frequency image information obtained from short-time Fourier transform and synchronous compressed wavelet transform, but also combines feature information from multiple dimensions such as signal entropy, frequency, and rotational speed. By calculating the signal entropy corresponding to each pixel in the first time-frequency image and determining the rotational speed corresponding to each pixel in the time-frequency image based on the rotational speed signal, these multi-dimensional features are fused, enabling a more comprehensive description of the motor's operating state. The system incorporates fault characteristics, improving the accuracy and reliability of fault identification; quantum state encoding enhances feature representation: by mapping or customizing quantum state mapping, pixels in the time-frequency graph are mapped to quantum states and a global quantum state tensor is constructed. This unique encoding method fully utilizes the superposition and entanglement properties of quantum states, enabling more effective extraction and characterization of weak features of early faults in permanent magnet synchronous motors. Compared with traditional methods, quantum state encoding can uncover deeper correlation information in the data, significantly improving the detection capability of complex nonlinear fault modes; correlation analysis uncovers potential information: in the quantum heuristic neural network module, convolution operations are used to simulate entanglement operations, performing cross-time and frequency entanglement operations on the rotated quantum states. The correlation between quantum states is obtained. This correlation analysis can uncover the potential connections between pixels at different positions in the time-frequency graph, discover some fault features that are difficult to detect by traditional methods, and further enrich the information dimension of fault features, which helps to improve the accuracy of fault classification. Dynamic weight adjustment improves model adaptability: When determining the fusion weight, the fusion weight of each pixel is dynamically calculated based on the frequency and signal entropy corresponding to each pixel in the first time-frequency graph and the frequency corresponding to each pixel in the second time-frequency graph. This dynamic weight adjustment mechanism enables the model to automatically allocate weights according to the importance of different features, which improves the model's adaptability to different working conditions and fault types, and helps the model converge to the optimal solution faster during training.Rotation angle clamping and transformation ensure stability: Before constructing the rotation matrix, the rotation angles of pixels that do not meet the preset rotation angle range are clamped or transformed. This operation avoids the interference of extreme angle values ​​on quantum state rotation operations, ensuring the stability of the rotation matrix construction. This improves the convergence of model training and the reliability of fault feature enhancement. A stable training process helps the model learn more accurate and generalizable fault feature representations, improving the model's performance in practical applications. Accurate early fault identification reduces economic losses: Because this method can significantly enhance the ability to capture early weak fault features and improve the discriminative power of fault features, This allows for accurate identification of fault occurrence and development in the early stages. Timely and accurate fault identification can prevent further deterioration of the fault, reduce economic losses such as production interruptions and equipment damage caused by motor failures, and lower maintenance costs and downtime. It also improves motor operational safety and reduces accident risks: accurate and reliable fault identification can promptly detect potential safety hazards in the motor, allowing for proactive maintenance measures to ensure the safe and stable operation of the motor. This is of great significance in critical application scenarios such as aerospace and energy power, effectively reducing the risk of serious safety accidents caused by motor failures and protecting personnel and equipment property.

[0169] In one feasible implementation, step 140 in the above embodiments, determining the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the first time-frequency map, the second time-frequency map, and the signal entropy corresponding to each pixel in the first time-frequency map, includes: determining the energy corresponding to each pixel in the first time-frequency map based on the first time-frequency map; determining the local frequency change rate corresponding to each pixel in the second time-frequency map based on the second time-frequency map; normalizing the energy corresponding to each pixel in the first time-frequency map to obtain the normalized energy corresponding to each pixel in the first time-frequency map and the normalized local frequency change rate corresponding to each pixel in the second time-frequency map; determining the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency map and the normalized local frequency change rate corresponding to each pixel in the second time-frequency map.

[0170] Regarding the method of determining energy, in some embodiments, the square of the modulus of each pixel in the first time-frequency image can be used as the energy corresponding to each pixel in the first time-frequency image.

[0171] Regarding the method for determining the local frequency change rate, in some embodiments, the instantaneous frequency corresponding to each pixel in the second time-frequency image can be obtained, and then time difference can be performed on the instantaneous frequency corresponding to each pixel in the second time-frequency image to obtain the local frequency change rate corresponding to each pixel in the second time-frequency image; of course, in other embodiments, frequency difference can also be performed on the instantaneous frequency corresponding to each pixel in the second time-frequency image to obtain the local frequency change rate corresponding to each pixel in the second time-frequency image.

[0172] In this embodiment, by extracting the energy of the first time-frequency map and the local frequency change rate of the second time-frequency map, and combining it with the signal entropy, the rationality and accuracy of the calculation of the fusion weight are improved, thereby enhancing the fault feature extraction effect and improving the reliability of fault identification.

[0173] Understandably, multi-dimensional feature utilization involves extracting the energy from the first time-frequency graph and the local frequency change rate from the second time-frequency graph, combined with signal entropy. This allows for a comprehensive evaluation of feature importance from three dimensions: time-frequency energy distribution, frequency dynamics, and signal complexity. Energy reflects the strength of the fault signal, the local frequency change rate captures transient frequency fluctuations, and signal entropy quantifies signal uncertainty. These three elements complement each other, enhancing the comprehensiveness of weight allocation. Normalization enhances comparability: normalizing the energy and local frequency change rate eliminates the influence of differences in feature dimensions, ensuring that the fusion weight calculation is based on a unified scale. For example, without normalization, high energy might mask the role of the frequency change rate; after normalization, both can synergistically contribute to the weight, improving feature fusion. Reasonableness of the combination; Dynamic weight allocation to optimize feature fusion: The weights are dynamically calculated based on normalized energy, signal entropy, and normalized frequency change rate, enabling the model to focus on weak but critical time-frequency features in the early stages of a fault, while suppressing noise interference. For example, high-frequency fault signals may have low energy but high frequency change rate. By adjusting the weights, such features can be highlighted, improving the sensitivity of early fault detection; Adaptation to non-stationary signal characteristics: By combining the local frequency change rate, the weight calculation can adapt to the real-time non-stationary changes of the motor signal, enhancing the model's adaptability to dynamic operating conditions. For example, when the load changes abruptly, the frequency change rate can quickly reflect signal fluctuations, and the weights are adjusted accordingly to capture transient fault features, avoiding the missed detections caused by the assumption of signal stability in traditional models.

[0174] In one feasible implementation, the determination of the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency image and the normalized local frequency change rate corresponding to each pixel in the second time-frequency image in the above embodiments includes:

[0175] Using formula Determine the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image;

[0176] in, Here are the fusion weights corresponding to each pixel in the first time-frequency image. Let n be the coordinates of the nth pixel in the first time-frequency graph. This represents the normalized energy corresponding to each pixel in the first time-frequency graph. For adjustment coefficients, Let be the signal entropy corresponding to each pixel in the first time-frequency graph. This represents the normalized local frequency change rate corresponding to each pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This represents the fusion weight corresponding to each pixel in the second time-frequency graph.

[0177] The value of the adjustment coefficient can be obtained and preset by the operator based on a large amount of experience, experimentation or statistics; preferably, the adjustment coefficient is set to 0.5 in this application.

[0178] In this embodiment, by introducing quantitative indicators such as normalized energy, signal entropy, and normalized local frequency change rate, and combining them with adjustment coefficients, a mathematically formulated weight calculation model is constructed. This realizes the scientific quantification and dynamic optimization of the fusion weight, significantly improving the accuracy and adaptability of fault feature fusion, thereby enhancing the reliability and real-time performance of fault identification.

[0179] Understandably, the following features are employed: Quantitative fusion weight calculation: By introducing quantitative indicators such as normalized energy, signal entropy, and normalized local frequency change rate, combined with adjustment coefficients, a mathematically formulated weight calculation model is constructed, avoiding the bias of subjective experience judgment and making weight allocation more objective and interpretable; Dynamic adaptation to changes in fault characteristics: In the formula, normalized energy reflects the fault signal strength, signal entropy quantifies signal complexity, and local frequency change rate captures transient frequency fluctuations. The dynamic combination of these three allows the weights to be adjusted in real time according to changes in fault characteristics; Multi-dimensional feature synergistic enhancement: By incorporating the three dimensions of time-frequency energy, frequency dynamics, and signal complexity into the weight calculation, the formula achieves multi-feature complementarity, avoiding missed detections or misjudgments; Real-time processing capability for non-stationary signals: The introduction of local frequency change rate enables the weight calculation to adapt to the real-time non-stationary changes of motor signals, enhancing the model's adaptability to dynamic operating conditions; Adjustment coefficient optimizes model flexibility: The settable nature of the adjustment coefficients allows the model to adjust the weight allocation strategy according to actual operating conditions or fault types.

[0180] In one feasible implementation, the preset permanent magnet synchronous motor fault prediction model of the above embodiment includes a dual-stream convolutional neural network module and an SVM classifier connected in sequence.

[0181] In some embodiments, the dual-stream convolutional neural network module is used to extract key features from the first time-frequency map and the second time-frequency map to obtain a first key feature vector and a second key feature vector. Based on the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map, the first key feature vector and the second key feature vector are weighted and fused to obtain a dual-stream fused feature vector. The SVM classifier is used to classify the dual-stream fused feature vector to obtain the fault classification and identification result.

[0182] It should be noted that the dual-stream convolutional neural network module of this application does not contain a softmax classifier. This application improves the accuracy of fault identification by introducing an SVM classifier to replace the softmax classifier in the dual-stream convolutional neural network module.

[0183] In this embodiment, the accuracy and reliability of fault classification are significantly improved by the collaborative design of the dual-stream convolutional neural network module and the SVM classifier.

[0184] Understandably, dual-stream feature extraction enhances complementarity: the dual-stream convolutional neural network module extracts key features from the first and second time-frequency maps respectively, generating first and second key feature vectors. Since the two time-frequency analysis methods (short-time Fourier transform and synchronous compressed wavelet transform) have different time-frequency localization characteristics, they can capture fault features from different dimensions, forming complementary feature representations and avoiding the limitations of single feature extraction methods. Weighted fusion optimizes feature representation: based on the fusion weights corresponding to each pixel in the first and second time-frequency maps, the first and second key feature vectors are weighted and fused to generate a dual-stream fused feature vector. This weighted fusion mechanism can dynamically adjust the contribution of different features in the final representation, highlighting features more important for fault classification and suppressing the influence of noise or irrelevant information, thereby improving the discriminative power of the feature representation. SVM The classifier improves classification robustness: An SVM classifier replaces the traditional softmax classifier, classifying the dual-stream fused feature vectors. SVM optimizes the decision boundary by maximizing the classification margin, exhibiting stronger generalization ability for small sample data and high-dimensional features. Especially when fault samples are unevenly distributed or feature dimensions are high, it significantly improves classification accuracy and robustness, reducing the risk of overfitting. End-to-end optimization enhances overall performance: The sequential connection design of the dual-stream convolutional neural network module and the SVM classifier achieves end-to-end optimization from feature extraction to classification decision. The dual-stream convolutional neural network module automatically learns fault features through deep learning, while the SVM classifier uses the principle of minimizing structured risk for the final decision. The combination of these two approaches fully leverages the feature learning capabilities of deep learning and the classification stability of traditional machine learning, resulting in an overall improvement in the accuracy and reliability of fault identification.

[0185] In one feasible implementation, the first key feature vector and the second key feature vector are weighted and fused according to the fusion weights corresponding to each pixel in the first time-frequency image and the fusion weights corresponding to each pixel in the second time-frequency image to obtain a dual-stream fused feature vector, including:

[0186] Using formula Obtain the dual-stream fusion feature vector;

[0187] in, For the r-th feature in the dual-stream fusion feature vector, Here are the fusion weights corresponding to each pixel in the first time-frequency image. Let n be the coordinates of the nth pixel in the first time-frequency graph. The nth feature in the first key feature vector. Here are the fusion weights corresponding to each pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. This is the m-th feature in the second key feature vector.

[0188] In this embodiment, a formulaic weighted fusion strategy is used to achieve dynamic optimization and integration of multi-dimensional features, which significantly improves the discriminativeness and classification reliability of fault feature representation.

[0189] Understandably, the dynamic weight allocation mechanism involves the formula dynamically weighting feature vectors by fusing weight parameters. This allows the model to adaptively adjust feature contributions based on the characteristics of different time-frequency analysis methods (the time-frequency resolution of short-time Fourier transform and the transient capture capability of synchronous compressed wavelet transform), avoiding information loss caused by a single dominant feature. Multimodal feature complementarity enhancement combines the complementary information of the first and second key feature vectors. Through pixel-level weight mapping, the formula achieves a coordinated expression of time-frequency energy distribution and transient frequency changes. Compared to simple splicing or averaging, this more accurately characterizes the spatiotemporal evolution of fault features. Improved noise resistance and robustness: During the weighted fusion process, low weights automatically suppress noise interference features (such as random fluctuations under normal operating conditions), while high weights focus on fault-sensitive features (such as energy mutations of specific frequency components). This enhancement mechanism enables the dual-stream fused feature vector to have stronger fault representation capabilities and anti-interference performance. End-to-end optimization compatibility: The weighted fusion formula and subsequent SVM classifier form a closed-loop optimization. The fusion weights can be automatically adjusted through backpropagation to ensure that the extracted dual-stream fused features are highly adapted to the classification decision boundary. Compared with the traditional hard-coded fusion method, it significantly improves the accuracy and generalization ability of fault classification.

[0190] In one feasible implementation, the dual-stream convolutional neural network module of the above embodiment includes two convolutional neural network units, a fusion layer, and a fully connected layer. Both convolutional neural network units are connected to the fusion layer, and the fusion layer is connected to the fully connected layer.

[0191] In some embodiments, a first convolutional neural network unit is used to extract key features from a first time-frequency image to obtain a first key feature vector; a second convolutional neural network unit is used to extract key features from a second time-frequency image to obtain a second key feature vector; a fusion layer is used to perform weighted feature fusion of the first key feature vector and the second key feature vector according to the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image to obtain an initial dual-stream fused feature vector; a fully connected layer is used to perform feature mapping on the initial dual-stream fused feature vector to obtain a dual-stream fused feature vector.

[0192] In this embodiment, the discriminative power and classification reliability of fault feature representation are significantly improved by using a two-stream convolutional neural network module for phased feature extraction and weighted fusion, combined with feature mapping optimization of the fully connected layer.

[0193] Understandably, dual-stream feature extraction enhances complementarity: two convolutional neural network units independently extract features from the first time-frequency map (short-time Fourier transform) and the second time-frequency map (synchronous compressed wavelet transform), generating first and second key feature vectors. Since the two time-frequency analysis methods have different time-frequency localization characteristics, the dual-stream structure can capture fault features from different dimensions, forming complementary feature representations and avoiding the limitations of single feature extraction methods. Weighted fusion optimizes feature representation: the fusion layer dynamically weights and fuses the first and second key feature vectors according to the fusion weights corresponding to each pixel in the first and second time-frequency maps, generating an initial dual-stream fused feature vector. This mechanism highlights features more important for fault classification through weight allocation, suppressing the influence of noise or irrelevant information, thereby improving the discriminative power of the feature representation. Fully connected layers optimize feature mapping: the fully connected layer optimizes the initial dual-stream feature vector. The dual-stream fusion feature vector is subjected to nonlinear feature mapping to further extract high-level abstract features and compress dimensions, generating the final dual-stream fusion feature vector. This process optimizes the fit between features and classification targets by learning the complex nonlinear relationships between features, enabling the subsequent SVM classifier to make decisions based on more compact and discriminative feature vectors, thereby improving classification accuracy and robustness. End-to-end optimization improves overall performance: The phased design of the dual-stream convolutional neural network module forms a closed-loop optimization with the SVM classifier. The convolutional neural network unit automatically learns fault features through deep learning, and the fusion layer and fully connected layer dynamically adjust weights and mapping parameters through backpropagation to ensure that the extracted dual-stream fusion features are highly adapted to the classification decision boundary. Compared with the traditional hard-coded fusion method, this structure significantly improves the accuracy and generalization ability of fault identification, and is especially suitable for dynamic working conditions under non-stationary signals.

[0194] In one feasible implementation, each convolutional neural network unit in the above embodiments includes an input layer, a first feature extraction block, a second feature extraction block, a first connection layer, and a second connection layer connected in sequence. The first feature extraction block and the second feature extraction block each include a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, and a Dropout layer connected in sequence.

[0195] In this embodiment, a convolutional neural network unit with hierarchical feature extraction and regularization design is used to achieve in-depth mining and efficient expression of time-frequency features, which significantly improves the robustness and discriminativeness of fault feature extraction.

[0196] Understandably, hierarchical feature extraction enhances representational capabilities: each convolutional neural network unit extracts low- to high-level features from the time-frequency map layer by layer through a cascaded structure of input layer → first feature extraction block → second feature extraction block. The first feature extraction block captures local time-frequency patterns, and the second feature extraction block further integrates global contextual information to form a multi-scale feature representation, avoiding information loss from single-layer features. Batch normalization layers accelerate training convergence: introducing batch normalization layers after convolutional layers standardizes the feature distribution of each batch of data, eliminating internal covariate bias. This design makes network training more stable, allows for larger learning rates, and reduces sensitivity to initial weights, making it particularly suitable for small-sample fault data scenarios. Activation function layers introduce nonlinear modeling capabilities: introducing nonlinear transformations through activation functions enables the network to fit the complex nonlinear relationship between fault features and classification labels, significantly improving the expressive power of feature extraction compared to linear models, especially... More sensitive to the nonlinear characteristics of early, weak faults; Max pooling layer achieves dimensionality reduction and translation invariance: Max pooling layer downsamples by the maximum value of local regions, compressing the data dimension while retaining key features and reducing subsequent computation. Its translation invariance makes the network insensitive to small time or frequency shifts in fault signals, enhancing the robustness of feature extraction; Dropout layer prevents overfitting: A Dropout layer is added at the end of the feature extraction block, randomly discarding some neuron connections and forcing the network to learn redundant feature representations. This regularization strategy effectively suppresses the risk of overfitting, especially when there are few fault samples, and can improve the model's generalization ability on the test set; Connection layer optimizes feature fusion: The first and second connection layers fuse multi-level features through concatenation or weighting to form a composite feature vector that combines local details and global semantics. This cross-level feature interaction mechanism makes the final feature representation more comprehensive, especially suitable for distinguishing multiple types of faults under complex working conditions.

[0197] In a second aspect, this application provides a fault identification device for permanent magnet synchronous motors based on customized fusion.

[0198] Please see Figure 2 This is a schematic diagram of a fault identification device for a permanent magnet synchronous motor based on customized fusion in an embodiment of this application. The device 210 includes:

[0199] The acquisition module 211 is used to acquire the vibration signal of the permanent magnet synchronous motor;

[0200] The transformation module 212 is used to perform short-time Fourier transform on the vibration signal to obtain a first time-frequency diagram, and to perform synchronous compressed wavelet transform on the vibration signal to obtain a second time-frequency diagram;

[0201] The signal entropy determination module 213 is used to determine the signal entropy corresponding to each pixel in the first time-frequency diagram based on the vibration signal.

[0202] The fusion weight determination module 214 is used to determine the fusion weight corresponding to each pixel in the first time-frequency diagram and the fusion weight corresponding to each pixel in the second time-frequency diagram based on the first time-frequency diagram, the second time-frequency diagram, and the signal entropy corresponding to each pixel in the first time-frequency diagram.

[0203] The model prediction module 215 is used to input the first time-frequency map, the second time-frequency map, the fusion weight corresponding to each pixel in the first time-frequency map, and the fusion weight corresponding to each pixel in the second time-frequency map into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification results.

[0204] In this embodiment, the relevant contents of the acquisition module 211, transformation module 212, signal entropy determination module 213, fusion weight determination module 214, and model prediction module 215 can be found in the following references. Figure 1 The contents of the illustrated embodiments will not be repeated here.

[0205] It should be noted that the device 210 of this application also includes other modules. It can be understood that the method of this application and the device 210 have a one-to-one correspondence. Therefore, the other modules of the device 210 of this application are the contents corresponding to the method of this application in the above embodiments.

[0206] In this embodiment, this customized fusion method fully leverages the advantages of both short-time Fourier transform and synchronous compressed wavelet transform, effectively extracting and fusing different feature information, enhancing the ability to capture early weak fault features, and significantly improving the distinguishability of fault features. Simultaneously, the device can better adapt to the real-time characteristics of non-stationary signals in permanent magnet synchronous motors, enabling the fault prediction model to more accurately distinguish different early faults. This greatly improves the accuracy, reliability, and real-time performance of fault identification in permanent magnet synchronous motors, providing strong support for ensuring the safe and stable operation of the motor.

[0207] In a third aspect, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform a custom-fusion-based fault identification method for permanent magnet synchronous motors as described in any of the first aspects.

[0208] This application provides a computer device in a fourth aspect, including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform a custom-fusion-based permanent magnet synchronous motor fault identification method as described in any of the first aspects.

[0209] Figure 3The diagram illustrates the internal structure of a computer device in some embodiments. This computer device may specifically be a terminal, a server, or a gateway. Figure 3 As shown, the computer device includes a processor, memory, and network interface connected via a system bus.

[0210] The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When executed by a processor, this computer program causes the processor to perform the steps in the above method embodiments. The internal memory may also store a computer program, which, when executed by a processor, causes the processor to perform the steps in the above method embodiments. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0211] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods.

[0212] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0213] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0214] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A fault identification method for permanent magnet synchronous motors based on customized fusion, characterized in that, The method includes: Acquire vibration and speed signals of the permanent magnet synchronous motor; The vibration signal is subjected to a short-time Fourier transform to obtain a first time-frequency diagram, and the vibration signal is subjected to a synchronous compressed wavelet transform to obtain a second time-frequency diagram. Based on the vibration signal, determine the signal entropy corresponding to each pixel in the first time-frequency diagram; Based on the first time-frequency diagram, the second time-frequency diagram, and the signal entropy corresponding to each pixel in the first time-frequency diagram, determine the fusion weight corresponding to each pixel in the first time-frequency diagram and the fusion weight corresponding to each pixel in the second time-frequency diagram; The first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image are input into a preset permanent magnet synchronous motor fault prediction model to obtain fault classification and identification results. Wherein, in the case that the preset permanent magnet synchronous motor fault prediction model includes a quantum heuristic neural network module, a two-stream convolutional neural network module, and an SVM classifier connected in sequence, the step of inputting the first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification result includes: Based on the rotation speed signal, determine the rotation speed corresponding to each pixel in the first time-frequency diagram and the rotation speed corresponding to each pixel in the second time-frequency diagram; Based on the rotational speed corresponding to each pixel in the first time-frequency diagram, each pixel in the first time-frequency diagram is mapped and encoded to obtain the customized quantum state corresponding to each pixel in the first time-frequency diagram. Similarly, based on the rotational speed corresponding to each pixel in the second time-frequency diagram, each pixel in the second time-frequency diagram is mapped and encoded to obtain the customized quantum state corresponding to each pixel in the second time-frequency diagram. The horizontal and vertical coordinates of the pixel are time and frequency, respectively. Based on the custom quantum states corresponding to all pixels in the first time-frequency graph, a first global custom quantum state tensor is determined, and based on the custom quantum states corresponding to all pixels in the second time-frequency graph, a second global custom quantum state tensor is determined. The first globally customized quantum state tensor, the second globally customized quantum state tensor, the fusion weight corresponding to each pixel in the first time-frequency graph, and the fusion weight corresponding to each pixel in the second time-frequency graph are input into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification results.

2. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 1, characterized in that, The step of determining the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the first time-frequency map, the second time-frequency map, and the signal entropy corresponding to each pixel in the first time-frequency map includes: Based on the first time-frequency graph, determine the energy corresponding to each pixel in the first time-frequency graph; Based on the second time-frequency graph, determine the local frequency change rate corresponding to each pixel in the second time-frequency graph; The energy corresponding to each pixel in the first time-frequency graph is normalized to obtain the normalized energy corresponding to each pixel in the first time-frequency graph, and the normalized local frequency change rate corresponding to each pixel in the second time-frequency graph. Based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency image, and the normalized local frequency change rate corresponding to each pixel in the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image are determined.

3. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 2, characterized in that, The step of determining the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image based on the normalized energy and signal entropy corresponding to each pixel in the first time-frequency image and the normalized local frequency change rate corresponding to each pixel in the second time-frequency image includes: Using formula Determine the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image; in, Let be the fusion weight corresponding to each pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. Let be the normalized energy corresponding to each pixel in the first time-frequency graph. For adjustment coefficients, Let be the signal entropy corresponding to each pixel in the first time-frequency graph. Let be the normalized local frequency change rate corresponding to each pixel in the second time-frequency graph. Let m be the coordinates of the m-th pixel in the second time-frequency graph. The fusion weight is the weight corresponding to each pixel in the second time-frequency graph.

4. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 1, characterized in that, In the case where the preset permanent magnet synchronous motor fault prediction model includes the dual-stream convolutional neural network module and the SVM classifier connected in sequence: The dual-stream convolutional neural network module is used to extract key features from the first time-frequency map and the second time-frequency map to obtain a first key feature vector and a second key feature vector. Based on the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map, the first key feature vector and the second key feature vector are weighted and fused to obtain a dual-stream fused feature vector. The SVM classifier is used to classify the dual-stream fused feature vectors to obtain the fault classification and identification results.

5. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 4, characterized in that, The step of performing weighted feature fusion of the first key feature vector and the second key feature vector based on the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image to obtain a dual-stream fused feature vector includes: Using formula The dual-stream fusion feature vector is obtained; in, For the r-th feature in the dual-stream fusion feature vector, Let be the fusion weight corresponding to each pixel in the first time-frequency graph. Let n be the coordinates of the nth pixel in the first time-frequency graph. The nth feature in the first key feature vector. Let be the fusion weight corresponding to each pixel in the second time-frequency image. Let m be the coordinates of the m-th pixel in the second time-frequency graph. It is the m-th feature in the second key feature vector.

6. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 4, characterized in that, The dual-stream convolutional neural network module includes two convolutional neural network units, a fusion layer, and a fully connected layer. Both convolutional neural network units are connected to the fusion layer, and the fusion layer is connected to the fully connected layer. The first convolutional neural network unit is used to extract key features from the first time-frequency graph to obtain the first key feature vector. The second convolutional neural network unit is used to extract key features from the second time-frequency graph to obtain the second key feature vector; The fusion layer is used to perform weighted feature fusion of the first key feature vector and the second key feature vector according to the fusion weight corresponding to each pixel in the first time-frequency image and the fusion weight corresponding to each pixel in the second time-frequency image, so as to obtain an initial dual-stream fusion feature vector. The fully connected layer is used to perform feature mapping on the initial dual-stream fusion feature vector to obtain the dual-stream fusion feature vector.

7. The fault identification method for permanent magnet synchronous motors based on customized fusion according to claim 6, characterized in that, Each convolutional neural network unit includes an input layer, a first feature extraction block, a second feature extraction block, a first connection layer, and a second connection layer connected in sequence. The first feature extraction block and the second feature extraction block each include a convolutional layer, a batch normalization layer, an activation function layer, a max pooling layer, and a dropout layer connected in sequence.

8. A fault identification device for permanent magnet synchronous motors based on customized fusion, characterized in that, The device includes: The acquisition module is used to acquire the vibration signal and speed signal of the permanent magnet synchronous motor; The transformation module is used to perform a short-time Fourier transform on the vibration signal to obtain a first time-frequency diagram, and to perform a synchronous compressed wavelet transform on the vibration signal to obtain a second time-frequency diagram; The signal entropy determination module is used to determine the signal entropy corresponding to each pixel in the first time-frequency diagram based on the vibration signal. The fusion weight determination module is used to determine the fusion weight corresponding to each pixel in the first time-frequency map and the fusion weight corresponding to each pixel in the second time-frequency map based on the first time-frequency map, the second time-frequency map, and the signal entropy corresponding to each pixel in the first time-frequency map. The model prediction module is used to input the first time-frequency map, the second time-frequency map, the fusion weight corresponding to each pixel in the first time-frequency map, and the fusion weight corresponding to each pixel in the second time-frequency map into a preset permanent magnet synchronous motor fault prediction model to obtain fault classification and identification results. Wherein, in the case that the preset permanent magnet synchronous motor fault prediction model includes a quantum heuristic neural network module, a two-stream convolutional neural network module, and an SVM classifier connected in sequence, the step of inputting the first time-frequency image, the second time-frequency image, the fusion weight corresponding to each pixel in the first time-frequency image, and the fusion weight corresponding to each pixel in the second time-frequency image into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification result includes: Based on the rotation speed signal, determine the rotation speed corresponding to each pixel in the first time-frequency diagram and the rotation speed corresponding to each pixel in the second time-frequency diagram; Based on the rotational speed corresponding to each pixel in the first time-frequency diagram, each pixel in the first time-frequency diagram is mapped and encoded to obtain the customized quantum state corresponding to each pixel in the first time-frequency diagram. Similarly, based on the rotational speed corresponding to each pixel in the second time-frequency diagram, each pixel in the second time-frequency diagram is mapped and encoded to obtain the customized quantum state corresponding to each pixel in the second time-frequency diagram. The horizontal and vertical coordinates of the pixel are time and frequency, respectively. Based on the custom quantum states corresponding to all pixels in the first time-frequency graph, a first global custom quantum state tensor is determined, and based on the custom quantum states corresponding to all pixels in the second time-frequency graph, a second global custom quantum state tensor is determined. The first globally customized quantum state tensor, the second globally customized quantum state tensor, the fusion weight corresponding to each pixel in the first time-frequency graph, and the fusion weight corresponding to each pixel in the second time-frequency graph are input into the preset permanent magnet synchronous motor fault prediction model to obtain the fault classification and identification results.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the fault identification method for permanent magnet synchronous motors based on custom fusion as described in any one of claims 1 to 7.

10. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the fault identification method for permanent magnet synchronous motors based on custom fusion as described in any one of claims 1 to 7.

Images