Permanent magnet synchronous motor fault diagnosis method and system

Abstract

The invention relates to the field of motor fault diagnosis, and particularly discloses a permanent magnet synchronous motor fault diagnosis method and system, which dynamically update motor physical model parameters through an online identification algorithm so as to reflect the characteristic change of a motor in real time. Thirdly, predicting the theoretical voltage of the healthy motor under the working condition by using the self-adaptive model and the real-time operation data, and calculating a residual sequence subjected to working condition normalization between the theoretical predicted voltage and the actual voltage; the residual signal is essentially free from the influence of working condition change, and only the abnormal characteristics caused by the fault are highlighted. And finally, inputting the high-robustness residual error sequence into a convolutional recurrent neural network, deeply fusing space-time fault features, and realizing accurate judgment on a motor fault type and confidence thereof, thereby effectively improving variable working condition adaptability and weak fault detection capability of diagnosis.

Description

Technical Field

[0001] This application relates to the field of motor fault diagnosis, and more specifically, to a method and system for diagnosing faults in permanent magnet synchronous motors. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs), with their high power density, high efficiency, and excellent dynamic response performance, have become key components in national strategic core fields such as new energy vehicle drive systems, industrial robots, aerospace, and precision manufacturing. However, these motors often operate under complex and demanding conditions, including frequent start-stop cycles, wide speed range regulation, and severe load fluctuations. They are subjected to multiple stress couplings from electrical, thermal, magnetic, and mechanical factors over long periods, making them prone to various faults such as stator winding inter-turn short circuits, permanent magnet demagnetization, and bearing wear. If these faults are not detected and addressed in a timely manner, they can lead to production stoppages and economic losses, or even catastrophic safety accidents. Therefore, researching and developing a fault diagnosis technology for PMSMs that enables early, accurate, and reliable fault diagnosis under complex operating conditions is of paramount importance for ensuring the safe and stable operation of critical equipment.

[0003] Currently, existing fault diagnosis technologies for permanent magnet synchronous motors can be mainly divided into mechanistic model-based methods and data-driven methods. Mechanism-based methods establish accurate mathematical models of the motor and compare the difference (i.e., residual) between actual measured signals and model-predicted signals to determine faults. This method has clear physical meaning, but its performance is highly dependent on the accuracy of the model parameters. In actual operation, motor parameters drift due to factors such as temperature, magnetic saturation, and aging. Fixed models struggle to adapt to these changes, leading to decreased diagnostic accuracy. On the other hand, data-driven methods, represented by deep learning, can automatically learn fault characteristics from massive amounts of data, avoiding reliance on accurate models. However, these methods have a critical bottleneck: their diagnostic performance is severely constrained by changes in motor operating conditions. Because the characteristics of motor signals (such as amplitude and frequency) vary greatly under different speeds and loads, a model trained under one operating condition will drastically degrade in performance under another, generating numerous false alarms and false negatives, making it difficult to meet the core requirement of robustness to varying operating conditions in industrial settings.

[0004] Therefore, an optimized fault diagnosis scheme for permanent magnet synchronous motors is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for diagnosing faults in a permanent magnet synchronous motor.

[0006] According to one aspect of this application, a fault diagnosis method for a permanent magnet synchronous motor is provided, comprising: Obtain the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle; The three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are preprocessed to obtain the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity. The theoretical signal prediction of the dq-axis voltage is obtained by performing theoretical signal prediction based on an adaptive physical model on the dq-axis current and the filtered rotor mechanical angular velocity. Calculate the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage; Extracting deeply fused fault feature vectors from the dq-axis voltage residual sequence; The deeply fused fault feature vectors are input into a classifier to obtain the fault type and its confidence level.

[0007] According to another aspect of this application, a fault diagnosis system for a permanent magnet synchronous motor is provided, comprising: The motor operation data acquisition module is used to acquire three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle. The data preprocessing module is used to preprocess the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity and rotor electrical angle to obtain dq-axis current, dq-axis voltage and filtered rotor mechanical angular velocity. The theoretical signal prediction module is used to perform theoretical signal prediction based on an adaptive physical model on the dq axis current and the filtered rotor mechanical angular velocity to obtain the theoretically predicted dq axis voltage. The voltage residual calculation module is used to calculate the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltages; The fault feature extraction module is used to extract deeply fused fault feature vectors from the dq axis voltage residual sequence; The fault classification module is used to input the deeply fused fault feature vector into the classifier to obtain the fault type and its confidence level.

[0008] Compared with existing technologies, this application provides a fault diagnosis method and system for permanent magnet synchronous motors. It dynamically updates the physical model parameters of the motor through an online identification algorithm to reflect changes in the motor's own characteristics in real time. Next, it uses this adaptive model and real-time operating data to predict the theoretical voltage of a healthy motor under the specified operating condition, and calculates the residual sequence between the theoretically predicted voltage and the actual voltage after operating condition normalization. This residual signal essentially removes the influence of operating condition changes, highlighting only the abnormal features caused by the fault. Finally, this highly robust residual sequence is input into a convolutional recurrent neural network to deeply fuse spatiotemporal fault features, achieving accurate judgment of the motor fault type and its confidence level, thereby effectively improving the diagnostic adaptability to changing operating conditions and the ability to detect weak faults. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a fault diagnosis method for a permanent magnet synchronous motor according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the permanent magnet synchronous motor fault diagnosis method according to an embodiment of this application; Figure 3 This is a flowchart illustrating the theoretical signal prediction of the dq-axis voltage based on an adaptive physical model in the fault diagnosis method for permanent magnet synchronous motors according to embodiments of this application, whereby the dq-axis current and the filtered rotor mechanical angular velocity are used to obtain the theoretically predicted dq-axis voltage. Figure 4 This is a flowchart of the dq-axis voltage residual sequence between the dq-axis voltage and the dq-axis voltage predicted by the computational theory of the permanent magnet synchronous motor fault diagnosis method according to the embodiments of this application; Figure 5 This is a block diagram of a permanent magnet synchronous motor fault diagnosis system according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] To address the technical challenge of severely reduced diagnostic accuracy and robustness of existing permanent magnet synchronous motor (PMSM) fault diagnosis technologies under varying operating conditions due to signal characteristic drift, this application proposes a PMSM fault diagnosis method. The implementation process first uses Clarke and Park transforms to convert the acquired three-phase current and voltage to a more easily analyzed dq rotating coordinate system. Furthermore, instead of directly analyzing these operating condition-coupled signals, a self-correcting adaptive physical model is constructed. This model utilizes online identification algorithms such as recursive least squares to continuously update estimates of key parameters such as stator resistance, dq-axis inductance, and permanent magnet flux linkage based on real-time dq-axis voltage, current, and speed data, thereby generating a dynamic health benchmark that accurately reflects the motor's current health status and operating characteristics. Subsequently, based on this updated model and real-time operating condition inputs (dq-axis current and speed), the theoretical healthy voltage vector is predicted, and the normalized modulus deviation rate and phase residual between the measured voltage vector and the theoretical benchmark vector are innovatively calculated. The sequence formed by these two complementary residual features effectively isolates the influence of speed and load fluctuations, greatly enhancing the signal-to-noise ratio of the fault signal. Finally, this condition-insensitive residual sequence is fed into a hybrid neural network composed of a one-dimensional convolutional network and a long short-term memory network to deeply mine the hidden local spatial structure and global temporal dependencies in the residuals, ultimately outputting accurate fault types and confidence levels, thereby achieving high-reliability fault diagnosis under complex and variable operating conditions.

[0017] The present application proposes a fault diagnosis method for permanent magnet synchronous motors. Figure 1 This is a flowchart of a fault diagnosis method for a permanent magnet synchronous motor according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in a permanent magnet synchronous motor fault diagnosis method according to an embodiment of this application. Figure 1 and Figure 2As shown, the fault diagnosis method for a permanent magnet synchronous motor according to an embodiment of this application includes the following steps: S100, acquiring three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle; S200, preprocessing the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle to obtain dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity; S300, performing theoretical signal prediction based on an adaptive physical model on the dq-axis current and filtered rotor mechanical angular velocity to obtain the theoretically predicted dq-axis voltage; S400, calculating the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage; S500, extracting a deeply fused fault feature vector from the dq-axis voltage residual sequence; S600, inputting the deeply fused fault feature vector into a classifier to obtain the fault type and its confidence level.

[0018] Specifically, in step S100, the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are acquired. It should be understood that the electromagnetic dynamic behavior and internal state of a permanent magnet synchronous motor are entirely determined by four fundamental physical quantities: three-phase voltage, three-phase current, rotor speed, and rotor position. These raw signals are the fundamental basis for constructing an accurate mathematical model of the motor and reflecting its complete operating state. Therefore, in the technical solution of this application, the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are further acquired to provide complete and necessary raw data input for subsequent coordinate transformation, adaptive physical model prediction, and residual calculation. This ensures that the basis for diagnostic analysis is an accurate mapping of the motor's true and comprehensive electromagnetic and mechanical operating state at any given time, laying a data foundation for subsequently accurately removing the influence of operating conditions and highlighting fault characteristics.

[0019] More specifically, in a specific example of this application, the method includes: First, three high-precision Hall effect current sensors are respectively installed on the three-phase power line connecting the inverter and the motor to measure the three-phase stator current in real time in a non-invasive manner; simultaneously, the proportional signal of the three-phase terminal voltage is obtained through a resistor voltage divider network and measured after processing by an isolation amplifier; second, a high-resolution absolute photoelectric encoder is coaxially mounted with the motor rotor, which directly outputs a high-precision rotor mechanical angle. This mechanical angle is multiplied by the number of motor pole pairs to calculate the rotor electrical angle, and the rotor mechanical angular velocity is obtained by performing real-time differential calculation on the mechanical angle; finally, the analog or digital signals of all the aforementioned sensors are synchronously sampled and converted from analog to digital using the microcontroller unit inside the motor controller or an external data acquisition card at a fixed sampling frequency of 20kHz, and the acquired discrete time-series data stream is stored in a buffer for subsequent steps.

[0020] Specifically, in step S200, the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are preprocessed to obtain the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity. It should be understood that because the current and voltage signals of a permanent magnet synchronous motor in a three-phase stationary coordinate system are mutually coupled and periodically time-varying AC quantities, their dynamic characteristics are complex and difficult to directly use for constructing a simple linear physical model or for accurate analysis of fault characteristics. At the same time, the originally acquired rotor mechanical angular velocity signal is easily affected by sensor noise and measurement disturbances, which, if not processed, will introduce errors into the model prediction. Therefore, in the technical solution of this application, the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are further preprocessed to obtain the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity. This converts the complex alternating three-phase physical quantities into mutually decoupled dq-axis components that appear as quasi-DC quantities in steady state, and eliminates high-frequency noise in the mechanical angular velocity. This provides a high-quality, stable, and easy-to-model data foundation for subsequent theoretical signal prediction based on an adaptive physical model. This significantly simplifies the mathematical expression of the motor model, improves the model's computational efficiency and prediction accuracy, and ensures that the generation of residual signals more purely reflects the motor's true abnormal state rather than the complexity of measurement or coordinate system transformation.

[0021] More specifically, in the embodiments of this application, the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are preprocessed to obtain dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity. This includes: performing Clarke transformation on the three-phase stator current and three-phase terminal voltage to obtain two-phase stationary coordinate system current and two-phase stationary coordinate system voltage; performing Park transformation on the two-phase stationary coordinate system current and two-phase stationary coordinate system voltage based on the rotor electrical angle to obtain dq-axis current and dq-axis voltage; and performing signal filtering processing on the rotor mechanical angular velocity to obtain filtered rotor mechanical angular velocity.

[0022] More specifically, in a concrete example of this application, the implementation of this step includes: first, performing Clarke transformation on the real-time acquired three-phase stator current and three-phase terminal voltage data respectively. This transformation will convert the three-dimensional ( , , ) vector and ( , , Vector projection onto two-dimensional stationary -β Coordinate system, respectively obtain the current in the two-phase stationary coordinate system. and the voltage in the two-phase stationary coordinate system This condenses the original three-phase information while preserving equal power or equal amplitude characteristics. Then, based on the synchronously acquired rotor electrical angle, the... and Perform the Park transformation to bring the stationary state to rest. -β The vector in the coordinate system is rotated to the dq rotating coordinate system synchronized with the rotor, and the dq-axis current is finally obtained. and dq axis voltage This transformation ensures that, ideally, the stator current and voltage become DC or slowly varying quantities in the dq coordinate system, greatly simplifying the analysis of the motor model. Simultaneously, the original rotor mechanical angular velocity signal is processed in a low-pass filter. High-frequency noise components are effectively filtered out using methods such as moving average filtering or IIR digital filtering, resulting in a smooth and stable filtered rotor mechanical angular velocity. These three sets of processed data—dq-axis current, dq-axis voltage, and the filtered rotor mechanical angular velocity—will serve as direct inputs to the subsequent adaptive physical model prediction stage.

[0023] Specifically, in step S300, the dq-axis current and the filtered rotor mechanical angular velocity are theoretically predicted using an adaptive physical model to obtain the theoretically predicted dq-axis voltage. It should be understood that the core physical parameters of a permanent magnet synchronous motor, such as stator resistance, dq-axis inductance, and permanent magnet flux linkage, are not constant. They slowly drift with increasing motor operating temperature, magnetic saturation caused by changes in load current, and aging due to long-term service. Using fixed nominal parameters for model prediction would inevitably introduce significant systematic errors. Therefore, in the technical solution of this application, the dq-axis current and the filtered rotor mechanical angular velocity are further theoretically predicted using an adaptive physical model to obtain the theoretically predicted dq-axis voltage. This constructs a high-fidelity health state benchmark model that can dynamically track changes in the motor's own characteristics. This ensures that the generated theoretically predicted voltage accurately represents the electrical response of a healthy motor under any operating condition, providing a possibility for subsequent calculations of a pure residual signal that truly reflects fault information, rather than one caused by model mismatch.

[0024] Figure 3 This is a flowchart illustrating the process of predicting the theoretically predicted dq-axis voltage by performing theoretical signal prediction based on an adaptive physical model on the dq-axis current and the filtered rotor mechanical angular velocity according to the fault diagnosis method for permanent magnet synchronous motors according to embodiments of this application. Figure 3As shown, step S300 includes: S310, based on the dq-axis current, dq-axis voltage and filtered rotor mechanical angular velocity, performing online identification of motor parameters based on recursive least squares method on the parameter estimation vector of the previous moment to obtain updated motor model parameters; S320, based on the dq-axis current, updated motor model parameters and filtered rotor mechanical angular velocity, performing theoretical voltage prediction to obtain theoretically predicted dq-axis voltage.

[0025] Accordingly, in step S310, based on the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity, the parameter estimation vector from the previous moment is subjected to online identification of motor parameters using recursive least squares to obtain updated motor model parameters. It should be understood that the core physical parameters of a permanent magnet synchronous motor, such as stator resistance, dq-axis inductance, and permanent magnet flux linkage, are the cornerstone of the diagnostic model's accuracy. However, these parameters change over time due to factors such as temperature, magnetic saturation, and aging during actual operation, causing fixed model parameters to fail to continuously and accurately characterize the motor's health status. Therefore, in the technical solution of this application, the parameter estimation vector from the previous moment is further subjected to online identification of motor parameters using recursive least squares to obtain updated motor model parameters, thereby enabling real-time and adaptive tracking of the dynamic changes in the motor's key physical parameters. This ensures that subsequent theoretical voltage predictions are based on a model that closely matches the actual state of the motor, thus providing a fundamental guarantee for generating a high-fidelity health reference signal and greatly improving the robustness of the diagnostic system to parameter drift.

[0026] In a specific example of this application, the online parameter identification is implemented by algebraically reconstructing the discrete-time dq-axis voltage equations of a permanent magnet synchronous motor to form a standard linear regression model. , where the measurement vector The dq-axis voltage at the current moment, and the parameter vector to be identified. The observation matrix is ​​a set containing stator resistance, dq-axis inductance, and permanent magnet flux linkage. This consists of known quantities such as the real-time dq-axis current, its discrete difference value, and the filtered electric angular velocity. Based on this, within each preset identification period, the recursive least squares algorithm is executed: first, using the parameter vector from the previous time step... and the observation matrix at the current time Calculate the predicted voltage and compare it with the actual measured voltage. The prediction error is obtained by subtracting the values ​​from the previous value. Next, a gain matrix is ​​calculated based on a recursively updated covariance matrix and the current observation matrix. This gain matrix dynamically adjusts the correction weights for the prediction error. Finally, the parameter vector from the previous time step is added to the product of this gain matrix and the prediction error to obtain the updated motor model parameters for the current time step. Simultaneously, the covariance matrix is ​​also updated, preparing for the identification in the next cycle. The final output of this iterative process is the updated motor model parameters containing the latest estimates.

[0027] Accordingly, in step S320, a theoretical voltage prediction is performed based on the dq-axis current, the updated motor model parameters, and the filtered rotor mechanical angular velocity to obtain the theoretically predicted dq-axis voltage. It should be understood that fault diagnosis of permanent magnet synchronous motors requires an accurate health reference signal for comparison. This reference signal not only needs to reflect real-time operating conditions but also needs to dynamically adapt to changes in the motor's own characteristics. If the original, uncalibrated model parameters are used directly for prediction, the prediction results will deviate from the actual health state, leading to mixed model mismatch errors in the fault residuals. Therefore, in the technical solution of this application, a theoretical voltage prediction is further performed based on the dq-axis current, the updated motor model parameters, and the filtered rotor mechanical angular velocity to obtain the theoretically predicted dq-axis voltage. This allows for the accurate calculation of a theoretical terminal voltage value that a healthy motor should have under the current operating conditions, with parameters adaptively corrected. This generates a high-fidelity health reference point, greatly eliminating spurious residuals caused by parameter drift and model inaccuracies, enabling subsequent residual analysis to more purely and sensitively capture the real abnormal signals caused by actual faults.

[0028] In a specific example of this application, this step includes: First, acquiring the real-time dq-axis current data stream from the preceding preprocessing stage, along with the updated motor model parameters (including stator resistance, dq-axis inductance, and permanent magnet flux linkage) provided by the online identification process, while simultaneously inputting the filtered rotor mechanical angular velocity. Next, based on the acquired real-time mechanical angular velocity and the number of motor pole pairs, the electrical angular velocity is first calculated. Subsequently, the dq-axis current sequence is numerically differentiated to obtain the current rate of change, which is typically estimated using backward differential or other digital filters in discrete systems. Finally, all these input variables—real-time dq-axis current, current rate of change, electrical angular velocity, and the updated stator resistance, dq-axis inductance, and permanent magnet flux linkage—are completely substituted into the dq-axis voltage equation of the permanent magnet synchronous motor. That is, the theoretical voltage is predicted using the following formula: in, , , and For the updated motor model parameters, the stator resistance, q-axis inductance, d-axis inductance, and permanent magnet flux linkage are... and For dq axis current, The filtered rotor mechanical angular velocity, The theoretically predicted d-axis voltage. This is the theoretically predicted q-axis voltage.

[0029] Specifically, in step S400, the residual sequence of the dq-axis voltage between the theoretically predicted dq-axis voltage and the actual dq-axis voltage is calculated. It should be understood that because the amplitude base of the electrical signals of a motor is inherently high when operating under heavy load or high speed, this leads to a corresponding increase in the absolute residual caused by small model errors or noise, even under healthy conditions. Conversely, under light load and low speed conditions, this absolute residual will be very small. This heteroscedasticity of the residual signal with varying operating conditions makes it difficult for subsequent diagnostic models to learn a unified fault judgment benchmark. In other words, generating the residual signal by directly subtracting the measured voltage from the theoretically predicted voltage has its technical defects. This method fails to fully consider the dynamic coupling and scale dependence between the residual signal and the real-time operating conditions of the motor. Specifically, when a permanent magnet synchronous motor operates in the heavy load or high speed range, the amplitude base of its various electrical signals is inherently high. At this time, even under healthy conditions, the amplitude of the absolute residual caused by small model errors or sensor noise will increase accordingly; conversely, under light load and low speed conditions, the amplitude of the absolute residual will be very small. The heteroscedasticity of the residual signal's variance, which varies with operating conditions, makes it extremely difficult for subsequent fault diagnosis models to learn a unified and stable fault judgment benchmark. This can easily amplify normal signal fluctuations into false alarms under high loads. Furthermore, this method treats the dq-axis voltage as two independent scalar channels. This element-wise subtraction mathematically severs their intrinsic connection, completely ignoring the spatial angular shift information that the dq-axis voltage, as a two-dimensional spatial vector, may exhibit under specific fault conditions. Some early faults may not significantly change the overall amplitude of the voltage vector, but they can cause a slight phase shift. This crucial geometric feature is completely lost in the original mechanism, thus reducing sensitivity to specific fault types. To accurately extract fault-related features from a strongly time-varying operating condition background, a dual-channel residual feature construction mechanism based on vector projection and adaptive operating condition normalization is constructed in the preferred embodiment of this application. In the technical solution of this application, the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage is further calculated to generate a fault feature indicator that is insensitive to operating condition changes and is scale-normalized. This effectively eliminates the direct impact of operating conditions on residual amplitude, enabling the generated residual sequence to stably and consistently characterize the relative severity of faults, thereby greatly reducing the false alarm rate and missed alarm rate under drastic operating conditions.

[0030] Figure 4 This is a flowchart illustrating the dq-axis voltage residual sequence predicted by the computational theory of the permanent magnet synchronous motor fault diagnosis method according to embodiments of this application. Figure 4As shown, step S400 includes: S410, calculating the original residual vector and reference vector magnitude of the theoretically predicted dq-axis voltage and the dq-axis voltage to obtain the original residual vector and the predicted voltage magnitude; S420, calculating the normalized magnitude deviation rate of the original residual vector based on the predicted voltage magnitude to obtain the normalized magnitude deviation rate; S430, calculating the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage based on the normalized magnitude deviation rate.

[0031] Accordingly, in step S410, the original residual vector and reference vector magnitude of the theoretically predicted dq-axis voltage and the dq-axis voltage are calculated to obtain the original residual vector and the predicted voltage magnitude. It should be understood that treating the dq-axis voltage as two independent scalar channels and performing element-wise subtraction mathematically severs their inherent connection as a unified spatial entity, completely ignoring the minute spatial angular offset information that the dq voltage may exhibit under specific early fault conditions, which is a two-dimensional spatial vector. Therefore, in the technical solution of this application, the original residual vector and reference vector magnitude of the theoretically predicted dq-axis voltage and the dq-axis voltage are further calculated to obtain the original residual vector and the predicted voltage magnitude. This elevates the analytical perspective from scalar to vector, treating the residual as a unified physical quantity with direction and magnitude, and simultaneously calculating a dynamic reference representing the current operating condition energy level. This fully preserves all geometric information regarding the deviation of the actual measured value from the theoretical health reference, laying a solid foundation for all subsequent normalization and phase analyses based on vector relationships.

[0032] Specifically, in a specific example of this application, the implementation of this step includes: first, at each sampling moment, the calculation module receives the measured dq-axis voltage vector from the data bus. and the theoretically predicted dq-axis voltage vector from the output of the adaptive physical model [ Next, a vector subtraction operation is performed, that is, the d and q components of the measured voltage vector are subtracted from the d and q components of the theoretically predicted voltage vector, respectively, and the difference is obtained. This constitutes the original residual vector at the current moment. This lays the foundation for all subsequent analyses based on geometric relationships. Its significance lies in the fact that it abandons the independent, separate analysis of the d and q channels in the original mechanism, instead treating the residual as a unified physical quantity with direction and magnitude. In other words, it completely preserves all information about the deviation of the actual measured value from the theoretical health benchmark.

[0033] Simultaneously, the calculation of the reference vector magnitude is performed in parallel. This involves taking the square root of the sum of the squares of the two components of the theoretically predicted dq-axis voltage vector, i.e., calculating its Euclidean magnitude, to obtain a scalar value, which is the predicted voltage magnitude. In a physical scenario, it represents the ideal voltage signal energy scale that a healthy motor should have under the current speed and load conditions, providing a dynamic and adaptive benchmark for subsequent normalization operations to eliminate the influence of operating conditions. It is expressed by the following formula: in, To predict the voltage magnitude, and These are the theoretically predicted d-axis and q-axis voltage components. Finally, the calculated original residual vector and predicted voltage magnitude are output as intermediate results for use in subsequent steps.

[0034] Accordingly, in step S420, based on the predicted voltage magnitude, the normalized magnitude deviation rate is calculated for the original residual vector to obtain the normalized magnitude deviation rate. It should be understood that since the absolute amplitude of the original residual signal fluctuates drastically with changes in motor operating conditions (such as speed and load), normal signal fluctuations under high load are easily amplified into false alarm fault characteristics, while weak fault characteristics under low load are easily masked. This heteroscedasticity makes it extremely difficult for subsequent fault diagnosis models to learn a unified and stable fault judgment benchmark. Therefore, in order to completely solve the technical defect of the residual signal amplitude fluctuating drastically with operating conditions in the original mechanism, the technical solution of this application further calculates the normalized magnitude deviation rate for the original residual vector based on the predicted voltage magnitude to obtain the normalized magnitude deviation rate. This allows for dynamic normalization of the residual vector magnitude, representing fault signal energy, and the predicted voltage magnitude, representing background operating condition signal energy, through a signal-to-noise ratio formula with clear physical meaning. In this way, a relatively stable, scale-invariant deviation rate can be generated. Regardless of the operating conditions of the motor, the deviation rate can be constrained to a stable numerical range, thus providing a robust fault indication feature for downstream diagnostic models that can be directly compared across operating conditions.

[0035] Specifically, in a specific example of this application, the implementation of this step includes: first, obtaining the d-axis and q-axis components of the original residual vector output from the previous calculation step. and predicted voltage magnitude At the same time, two pre-calibrated parameters are read from the system configuration: one is the expected noise value of the system in a healthy state. One is used to subtract the basis noise in the calculation to improve the sensitivity to detect weak faults; the other is a very small normal number. This is used to avoid zero denominators in the calculation. Next, the square root of the sum of squares is performed on the two components of the original residual vector to obtain its Euclidean magnitude, and the expected noise value is subtracted from it. This forms the numerator of the normalization formula. Then, the predicted voltage magnitude is... With the smallest normal number The sums form the denominator of the normalized formula. Finally, a division operation is performed between the numerator and denominator to obtain the normalized modulus deviation rate at the current moment. That is, the modulus deviation rate of the original residual vector is calculated using the following formula for normalizing the operating conditions: in, Normalized modulus deviation rate; and These are the initial voltage residual components along the d-axis and q-axis of the original residual vector; The pre-calibrated expected noise value of the system under healthy conditions. The value is a very small positive constant to avoid a denominator of zero. This mechanism compares the energy of the fault signal (numerator) with the energy of the background operating condition signal (denominator) to obtain a relatively stable deviation rate. Thus, regardless of the motor's operating conditions, the generated normalized modulus deviation rate remains consistent. All of these can be constrained within a stable numerical range, thus providing downstream diagnostic models with scale-invariant fault indication features that can be directly compared.

[0036] Accordingly, in step S430, based on the normalized modulus deviation rate, the residual sequence of the dq-axis voltage between the theoretically predicted dq-axis voltage and the dq-axis voltage is calculated. It should be understood that, since only the normalized modulus deviation rate, which is insensitive to changes in operating conditions, is used as a single feature, although the problem of residual amplitude drift with operating conditions is solved, the spatial angle information of the voltage vector is still lost. Furthermore, some early faults, such as encoder eccentricity and changes in inductance parameters, are characterized more by spatial angle deviations of the actual electromagnetic field relative to the ideal state, rather than amplitude changes. Therefore, in order to capture those fault features that mainly affect the spatial relationship of the motor's electromagnetic field and to compensate for the deficiency of the original mechanism in losing vector spatial angle information, the technical solution of this application further calculates the residual sequence of the dq-axis voltage between the theoretically predicted dq-axis voltage and the dq-axis voltage based on the normalized modulus deviation rate. This combines the normalized modulus deviation rate under operating conditions with a phase residual feature that can directly quantify phase shift, forming a physically orthogonal dual-channel residual feature sequence. In this way, fault information can be captured completely from both amplitude and phase dimensions, making up for the lack of dimension in the single mode length deviation feature, and greatly enhancing the detection sensitivity for specific fault types (especially faults that affect the orientation accuracy of electromagnetic fields).

[0037] Specifically, in a concrete example of this application, this step includes: First, at each sampling moment, the system calculates the phase residual feature in parallel. This calculation utilizes the geometric definition of the vector dot product to directly calculate the angle between the measured voltage vector and the theoretically predicted voltage vector using the inverse cosine function. This angle defines the residual from a geometric dimension, intuitively reflecting the degree of spatial angular deviation of the actual electromagnetic field. Thus, defining the residual from a geometric rather than an algebraic dimension generates a new feature that is highly sensitive to faults related to the decoupling accuracy of the motor controller and the accuracy of the magnetic field orientation (such as encoder eccentricity and changes in inductance parameters). .

[0038] The calculation process of the phase residual characteristics can be expressed by the formula: in, The phase residual is represented by ·, which indicates the vector dot product operation. The Euclidean modulus representing the vector. It is an inverse cosine function. To measure the dq-axis voltage vector, To theoretically predict the dq-axis voltage vector.

[0039] Finally, the normalized modulus deviation rate under the operating conditions is calculated. With phase residual These two physically orthogonal features are combined into a dual-channel residual feature sequence, which is then sliced ​​using a sliding window to form the final input sample for downstream deep learning model analysis. Specifically, the normalized modulus deviation rate calculated in the previous step is used as the feature of the first channel, and the phase residual calculated in this step is used as the feature of the second channel, combined to form a two-dimensional feature vector. Further, the continuous two-dimensional feature vector stream is fed into a sliding window module, which slices the data stream with a fixed window length (e.g., 2048 sampling points) and step size. Each slice constitutes a dq-axis voltage residual sequence sample with dimensions [2, 2048], and is then output to the subsequent feature extraction process.

[0040] Through the above-described preferred embodiments, the robustness, sensitivity, and accuracy of the permanent magnet synchronous motor fault diagnosis system can be significantly improved in complex and variable industrial operating environments. By constructing the normalized modulus deviation rate as a core feature, the problem of residual signal scale drift caused by motor speed and load fluctuations is effectively solved. This allows the diagnostic model to judge faults based on a stable standard with cross-operating conditions, thereby significantly reducing the false alarm rate and false negative rate in scenarios with drastic changes in operating conditions. Simultaneously, by introducing the phase residual as a novel feature dimension, this mechanism can capture early fault information hidden in the angular changes of the voltage vector space, which is imperceptible by traditional methods. This greatly enhances the detection sensitivity for specific fault types (especially those affecting the accuracy of electromagnetic field orientation). Finally, the dual-channel residual feature sequence provided to the diagnostic model, compared to the original dq-axis residual, contains higher-dimensional, more stable, and physically more clearly defined fault information, providing a solid data foundation for achieving highly reliable intelligent health management of motors.

[0041] Specifically, in step S500, a deeply fused fault feature vector is extracted from the dq-axis voltage residual sequence. It should be understood that although the dq-axis voltage residual sequence generated in the aforementioned steps has been stripped of the influence of operating condition changes, it is still a raw time-series signal. The fault information contained within it often exists in the form of complex local patterns (such as specific harmonic components and transient impacts) and temporal evolution laws. Directly inputting it into a simple classifier makes it difficult to fully exploit its diagnostic value and is easily affected by noise. Therefore, in the technical solution of this application, a deeply fused fault feature vector is further extracted from the dq-axis voltage residual sequence to automatically learn and integrate the local spatial structure information and long-range temporal dependencies in the residual signal, thereby forming a highly abstract, information-density, and highly discriminative fault feature representation. This significantly improves the accuracy of fault mode identification and the robustness of the classifier, providing high-quality input for accurately determining the fault type and its confidence level.

[0042] More specifically, in this embodiment, extracting a deeply fused fault feature vector from the dq-axis voltage residual sequence includes: performing 1D-CNN-based local spatial feature extraction and dimensionality reduction on the dq-axis voltage residual sequence to obtain a dq-axis voltage feature sequence; and performing LSTM-based sequence temporal dependency modeling and feature fusion on the dq-axis voltage feature sequence to obtain a deeply fused fault feature vector. That is, more specifically, this step includes: first, performing 1D-CNN-based local spatial feature extraction and dimensionality reduction on the dq-axis voltage residual sequence. This residual sequence, as a dual-channel input (a sequence of normalized modulus deviation rate and phase residual changing over time), is fed into a multi-layer one-dimensional convolutional neural network. Each convolutional kernel slides along the time dimension, automatically capturing transient impacts, specific frequency components, or pattern fluctuations in the sequence through a local receptive field, such as specific harmonic modes caused by inter-turn short circuits. Convolutional layers are typically followed by activation functions to enhance nonlinear expressiveness, and pooling layers are used for feature dimensionality reduction and abstraction, making the extracted features robust to small time shifts. The final output is a compressed and refined dq-axis voltage feature sequence. This dq-axis voltage feature sequence is then fed into a Long Short-Term Memory (LSTM) network. As a recurrent neural network, LSTM incorporates gating mechanisms, enabling it to effectively learn the temporal dependencies and evolutionary trends of feature sequences, possessing the ability to learn from long-term memory. The network processes the feature sequence point by point, memorizing and updating its internal state to capture the dynamic patterns of fault features persisting, decaying, or appearing periodically over time. Finally, the hidden state at the last time step of the LSTM network is taken. This state integrates the local structure and temporal evolution information of the entire input sequence, forming a highly condensed single vector, the deeply fused fault feature vector, which will be used for final fault classification.

[0043] Specifically, in step S600, the deeply fused fault feature vector is input into a classifier to obtain the fault type and its confidence level. It should be understood that, since the aforementioned steps have refined the complex original sensor signals and high-dimensional residual sequences into a highly abstract and discriminative deeply fused fault feature vector, this vector encapsulates the essential information of the fault, but it does not yet possess the function of directly indicating the specific fault type and its probability of occurrence. Therefore, in the technical solution of this application, the deeply fused fault feature vector is further input into a classifier to obtain the fault type and its confidence level, thereby mapping the abstract fault features to a predefined fault category space and quantifying the reliability of the diagnostic results. In this way, user-readable fault information can be directly output, realizing intelligent and automated diagnosis of permanent magnet synchronous motor faults, providing a clear basis for early warning and maintenance decisions, and significantly improving the level of equipment health management.

[0044] More specifically, in a concrete example of this application, this step includes: First, obtaining a deep fusion fault feature vector representing the current motor state from the output of the preceding feature extraction module. This vector is a fixed-length numerical sequence, where each value carries the spatiotemporal information of the fault. Next, this feature vector is input into a classifier pre-trained with a large amount of labeled fault data. This classifier typically consists of one or more fully connected neural networks, which are responsible for performing complex nonlinear transformations and aggregations on the input feature vector. The final layer is typically connected to a Softmax activation function, which receives the logical value output from the previous layer and converts it into a probability distribution vector. Each element of this probability distribution vector corresponds to a predefined fault type (e.g., healthy, inter-turn short circuit, bearing fault, permanent magnet demagnetization, etc.), and its value represents the probability that the current feature vector belongs to that fault type. The system then selects the fault type with the highest probability as the final diagnostic result and outputs its corresponding probability value as the confidence level of the diagnostic result.

[0045] In summary, the fault diagnosis method for permanent magnet synchronous motors according to the embodiments of this application is explained. It dynamically updates the physical model parameters of the motor through an online identification algorithm to reflect changes in the motor's own characteristics in real time. Next, the adaptive model and real-time operating data are used to predict the theoretical voltage of a healthy motor under the specified operating condition, and the residual sequence between the theoretically predicted voltage and the actual voltage is calculated after operating condition normalization. This residual signal essentially removes the influence of operating condition changes, highlighting only the abnormal features caused by the fault. Finally, this highly robust residual sequence is input into a convolutional recurrent neural network to deeply fuse spatiotemporal fault features, achieving accurate judgment of the motor fault type and its confidence level, thereby effectively improving the diagnostic adaptability to changing operating conditions and the ability to detect weak faults.

[0046] Furthermore, a fault diagnosis system for permanent magnet synchronous motors is also provided.

[0047] Figure 5 This is a block diagram of a permanent magnet synchronous motor fault diagnosis system according to an embodiment of this application. Figure 5As shown, the permanent magnet synchronous motor fault diagnosis system 100 according to an embodiment of this application includes: a motor operation data acquisition module 110, used to acquire three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle; a data preprocessing module 120, used to preprocess the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle to obtain dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity; a theoretical signal prediction module 130, used to perform theoretical signal prediction based on an adaptive physical model on the dq-axis current and filtered rotor mechanical angular velocity to obtain theoretically predicted dq-axis voltage; a voltage residual calculation module 140, used to calculate the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage; a fault feature extraction module 150, used to extract a deeply fused fault feature vector from the dq-axis voltage residual sequence; and a fault classification module 160, used to input the deeply fused fault feature vector into a classifier to obtain the fault type and its confidence level.

[0048] As described above, the permanent magnet synchronous motor fault diagnosis system 100 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with permanent magnet synchronous motor fault diagnosis algorithms. In one possible implementation, the permanent magnet synchronous motor fault diagnosis system 100 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the permanent magnet synchronous motor fault diagnosis system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the permanent magnet synchronous motor fault diagnosis system 100 can also be one of many hardware modules of the wireless terminal.

[0049] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for diagnosing faults in a permanent magnet synchronous motor, characterized in that, include: Obtain the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle; The three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle are preprocessed to obtain the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity. The theoretical signal prediction of the dq-axis voltage is obtained by performing theoretical signal prediction based on an adaptive physical model on the dq-axis current and the filtered rotor mechanical angular velocity. Calculate the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltage and the dq-axis voltage; Extracting deeply fused fault feature vectors from the dq-axis voltage residual sequence; The deeply fused fault feature vectors are input into a classifier to obtain the fault type and its confidence level.

2. The fault diagnosis method for permanent magnet synchronous motors according to claim 1, characterized in that, Preprocessing of three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle yields dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity, including: Clarke transformation is performed on the three-phase stator current and three-phase terminal voltage to obtain the two-phase stationary coordinate system current and two-phase stationary coordinate system voltage; Based on the rotor electrical angle, Park transformation is performed on the two-phase stationary coordinate system current and the two-phase stationary coordinate system voltage to obtain the dq-axis current and dq-axis voltage. The rotor mechanical angular velocity is processed by signal filtering to obtain the filtered rotor mechanical angular velocity.

3. The fault diagnosis method for permanent magnet synchronous motors according to claim 1, characterized in that, The theoretically predicted dq-axis voltage is obtained by performing theoretical signal prediction based on an adaptive physical model on the dq-axis current and the filtered rotor mechanical angular velocity, including: Based on the dq-axis current, dq-axis voltage, and filtered rotor mechanical angular velocity, the parameter estimation vector of the previous moment is used to perform online identification of motor parameters based on recursive least squares method to obtain updated motor model parameters; Theoretical voltage prediction is performed based on the dq-axis current, updated motor model parameters, and filtered rotor mechanical angular velocity to obtain the theoretically predicted dq-axis voltage.

4. The fault diagnosis method for permanent magnet synchronous motors according to claim 3, characterized in that, The theoretical voltage prediction for the dq-axis is performed based on the dq-axis current, updated motor model parameters, and filtered rotor mechanical angular velocity. This includes the following formula for theoretical voltage prediction: in, , , and For the updated motor model parameters, the stator resistance, q-axis inductance, d-axis inductance, and permanent magnet flux linkage are... and For dq axis current, This represents the filtered rotor mechanical angular velocity.

5. The fault diagnosis method for permanent magnet synchronous motors according to claim 1, characterized in that, The calculation of the theoretically predicted dq-axis voltage and the dq-axis voltage residual sequence includes: The original residual vector and the magnitude of the reference vector are calculated from the theoretically predicted dq-axis voltage and the dq-axis voltage to obtain the original residual vector and the predicted voltage magnitude; Based on the predicted voltage magnitude, the normalized magnitude deviation rate is calculated by normalizing the original residual vector under operating conditions to obtain the normalized magnitude deviation rate. Based on the normalized modulus deviation rate, the residual sequence of dq-axis voltages between the theoretically predicted dq-axis voltages is calculated.

6. The fault diagnosis method for a permanent magnet synchronous motor according to claim 5, characterized in that, Based on the predicted voltage magnitude, the normalized magnitude deviation rate is calculated by normalizing the original residual vector under operating conditions. This includes: calculating the normalized magnitude deviation rate of the original residual vector under operating conditions using the following formula: in, To predict the voltage modulus, and This is the theoretically predicted dq-axis voltage. Normalized modulus deviation rate; and The initial dq-axis voltage residual of the original residual vector; The pre-calibrated expected noise value; It is a very small positive number.

7. The fault diagnosis method for permanent magnet synchronous motors according to claim 1, characterized in that, Deeply fused fault feature vectors are extracted from the dq-axis voltage residual sequence, including: Local spatial feature extraction and dimensionality reduction based on 1D-CNN are performed on the dq-axis voltage residual sequence to obtain the dq-axis voltage feature sequence; LSTM-based sequence time dependency modeling and feature fusion are performed on the dq axis voltage feature sequence to obtain a deeply fused fault feature vector.

8. A fault diagnosis system for a permanent magnet synchronous motor, characterized in that, include: The motor operation data acquisition module is used to acquire three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity, and rotor electrical angle. The data preprocessing module is used to preprocess the three-phase stator current, three-phase terminal voltage, rotor mechanical angular velocity and rotor electrical angle to obtain dq-axis current, dq-axis voltage and filtered rotor mechanical angular velocity. The theoretical signal prediction module is used to perform theoretical signal prediction based on an adaptive physical model on the dq axis current and the filtered rotor mechanical angular velocity to obtain the theoretically predicted dq axis voltage. The voltage residual calculation module is used to calculate the dq-axis voltage residual sequence between the theoretically predicted dq-axis voltages; The fault feature extraction module is used to extract deeply fused fault feature vectors from the dq axis voltage residual sequence; The fault classification module is used to input the deeply fused fault feature vector into the classifier to obtain the fault type and its confidence level.

Images