A method for diagnosing demagnetization fault of permanent magnet motor based on two-dimensional dynamic torque distribution prediction

Abstract

The application discloses a kind of two-dimensional dynamic torque distribution prediction permanent magnet motor demagnetization fault diagnosis methods, belong to computer data processing and application technical field, the specific steps of this method are as follows: first, collect motor multi-source electric signal to construct dynamic time series data window, form data set, according to this, train the two sparse observation models of dual-channel flux observer, input real-time data is handled and outputs integrated flux estimation value and integrated predicted torque, temperature compensation flux and reference compensation torque are obtained by calculation, and characteristic deviation sequence is generated from the difference between the two, to determine demagnetization fault;The application builds dual-channel flux observer, peels off temperature interference, optimizes parameters relying on selective memory update strategy;Collaborate inverter compensation link to improve the accuracy of electrical parameters, generate characteristic deviation sequence through two-dimensional dynamic torque prediction, adopt interlock determination mechanism, combined with multi-source signal acquisition, improve the accuracy of permanent magnet motor demagnetization fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of computer data processing and application technology, and in particular to a method and system for diagnosing demagnetization faults in permanent magnet motors based on two-dimensional dynamic torque distribution prediction. Background Technology

[0002] As a high-performance drive component, the operating status of permanent magnet synchronous motors is directly related to the safety and reliability of the entire system. Monitoring the health status of its core component, permanent magnets, especially the online diagnosis of irreversible demagnetization faults, has always been a key technology in this field. Such diagnostic methods usually involve complex analysis and processing of multi-source, high-dimensional, nonlinear electrical parameter time series data generated during motor operation in order to extract weak features related to the fault.

[0003] Application number CN202010986508.9 discloses a "method and device for diagnosing demagnetization faults of permanent magnet synchronous motors based on information fusion". This method uses a torque sensor to collect the torque signal generated by the tangential electromagnetic force and multiple acceleration sensors to collect the vibration signal generated by the radial electromagnetic force. The average value of the two is used as the input of a BP neural network to output the degree of demagnetization. Although this method improves the diagnostic accuracy to some extent through information fusion, it still has obvious shortcomings: First, this method requires the additional installation of torque sensors and multiple acceleration sensors, which increases the hardware cost and structural complexity of the system. First, the method has several drawbacks. First, the sensor's installation location, accuracy, and long-term stability directly affect the diagnostic results. Second, the method relies on offline training of a BP neural network, requiring a large amount of labeled sample data with different degrees of demagnetization, which is very difficult to obtain in actual engineering. Third, once the model is trained, it cannot be updated online adaptively; when the motor's operating conditions change or a new demagnetization mode appears, the diagnostic accuracy will significantly decrease. Fourth, the method does not consider the interference of temperature on the permanent magnet flux linkage, and in actual operation, flux linkage drift caused by temperature changes is difficult to distinguish from flux linkage attenuation caused by demagnetization, easily leading to misjudgment.

[0004] Existing demagnetization fault diagnosis methods are based on fixed-parameter models. These typically involve establishing a mathematical model describing the electromagnetic relationship of the motor offline, such as a flux linkage model or torque model. Fault diagnosis is then made by comparing the deviation between the model's predicted values ​​and the actual observed values. However, once established, these models remain unchanged, making it difficult to adapt to the dynamic characteristic drift caused by load changes, magnetic circuit saturation, and other factors during actual motor operation. This leads to decreased prediction accuracy under complex operating conditions. Signal processing-based diagnostic methods analyze the current or voltage signals of the motor to identify specific harmonic components generated by demagnetization faults. However, the fault characteristic signals relied upon by these methods are often weak and easily drowned out by strong background noise and dynamic disturbances during operating condition switching, resulting in insufficient stability and reliability of the diagnostic results. Some solutions introduce physical temperature sensors to compensate for the temperature effect on flux linkage. This not only increases hardware costs and potential fault points but also means that the temperature measured by the sensor may not accurately reflect the true internal temperature of the permanent magnet, resulting in measurement delays due to thermal inertia and making it difficult to clearly distinguish the temperature effect from the actual demagnetization effect. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a two-dimensional dynamic torque distribution prediction method for diagnosing demagnetization faults in permanent magnet motors. It constructs a dynamic time-series dataset by collecting multi-source electrical signals from the motor, establishes a dual-channel flux linkage observer for precise flux linkage monitoring, employs a selective memory update strategy to dynamically optimize model parameters, and combines temperature-induced flux linkage calculation with a collaborative compensation mechanism to effectively counteract temperature interference with flux linkage and torque. Simultaneously, it corrects signal errors caused by inverter nonlinear distortion, generates a characteristic deviation sequence through torque differences, and completes dual-logic fault determination using an interlocking judgment mechanism. Furthermore, it enhances diagnostic confidence through key basis function weight analysis. This solution significantly improves the accuracy and anti-interference capability of fault diagnosis under complex operating conditions, enabling online precise monitoring of permanent magnet motor demagnetization faults and adapting to the safety operation and maintenance needs of permanent magnet motors in various fields.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for diagnosing demagnetization faults in permanent magnet motors based on two-dimensional dynamic torque distribution prediction, the method comprising: The process of reading multi-source electrical signals from a motor involves acquiring the phase current, terminal voltage, and rotor speed signals of the motor, performing coordinate transformation on the phase current and terminal voltage to obtain direct-axis current, quadrature-axis current, direct-axis voltage, and quadrature-axis voltage, which together with the rotor speed constitute multi-source electrical signals. A dynamic time-series data window is constructed based on the temporal relationship to obtain a multi-source time-series dataset, providing a standardized and time-unified original data foundation for subsequent model training and electromagnetic parameter calculation; The phase current, terminal voltage, and rotor speed signals of the motor are acquired. Coordinate transformation is performed on the phase current and terminal voltage to obtain direct-axis current, quadrature-axis current, direct-axis voltage, and quadrature-axis voltage, which, together with the rotor speed, constitute a multi-source electrical signal. A dynamic time-series data window is constructed based on the timing relationship to obtain a multi-source time-series dataset. Based on a multi-source time-series dataset, the flux linkage observation values ​​obtained through offline calibration calculation of the voltage equation are used as the training target. A first sparse observation model is trained through a contribution competition mechanism, and a second sparse observation model is trained based on the temperature perturbation flux linkage sample signals pre-collected and physically calculated during the training phase. The two models are combined to form a dual-channel flux linkage observer, which realizes independent calculation and cross-validation of the comprehensive flux linkage and the temperature perturbation flux linkage, thereby improving the accuracy of flux linkage calculation. The temperature disturbance flux signal calculated by the physical calculation is specifically to synchronously collect the direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current and rotor speed electrical parameters of the motor at the current running time and the previous sampling time. The two sets of electrical parameters are substituted into the voltage equation in the rotating coordinate system of the permanent magnet motor to calculate the intermediate flux value at the current time and the intermediate flux value at the previous time. The difference operation is performed on the two intermediate flux values ​​to obtain the temperature disturbance flux signal caused only by temperature change. The real-time collected direct-axis current, quadrature-axis current, and rotor speed are input into the first sparse observation model. After dynamically adjusting the model parameters using a selective memory update strategy, the comprehensive flux linkage estimate is output. Combined with the real-time current solution, the two-dimensional dynamic comprehensive prediction torque is calculated to complete the dynamic prediction of electromagnetic torque across the entire operating condition range. Based on the electrical parameters at adjacent sampling times, the first temperature disturbance flux linkage value is calculated using the differential transformation of the permanent magnet motor voltage equation. Based on the temperature drift relationship of stator resistance, the stator resistance term in the voltage equation is corrected, and the difference between the inverse calculation results of flux linkage before and after the correction is taken as the first temperature disturbance flux linkage value.

[0008] After inputting the first temperature perturbation flux linkage value into the second sparse observation model to complete the verification and fusion, the second temperature perturbation flux linkage value is output, and the flux linkage drift component caused solely by temperature is accurately extracted and purified. Temperature-compensated flux linkage is obtained by performing temperature-coordinated compensation on the comprehensive flux linkage estimate using the second temperature disturbance flux linkage value. Combined with the real-time current calculation benchmark compensation torque, all interference of temperature factors on torque calculation is eliminated, and an interference-free fault judgment torque benchmark is established. The difference between the two-dimensional dynamic comprehensive predicted torque and the reference compensation torque is used to generate a demagnetization fault characteristic deviation sequence. Combined with the long-term degradation trend of the reference compensation torque, the demagnetization fault is determined through an interlocking judgment mechanism. The calculation of the adaptive warning threshold introduces the load factor of the current operating condition. The load factor is defined as the ratio of the actual torque under the current operating condition to the rated torque of the motor, so as to realize dynamic threshold adjustment under different loads and reduce the probability of fault misjudgment under complex operating conditions.

[0009] Optionally, the first sparse observation model is trained through a contribution competition mechanism, the basis function library contains four types of nonlinear transformation functions: polynomial function, exponential function, Gaussian function, and radial basis function, and the initial design matrix is ​​a high-dimensional mathematical matrix that perfectly matches the dimension of the multi-source time series dataset. The training process is as follows: Based on the data dimensions of the multi-source time series dataset, a corresponding number of candidate basis functions are selected from the basis function library, and the initial design matrix Φ is constructed based on the candidate basis functions; Using the flux linkage observations obtained from offline calibration based on the voltage equation in a multi-source time-series dataset as the model training target values, the expectation-maximization algorithm is used to iteratively update the weight hyperparameters corresponding to each candidate basis function in the initial design matrix. The formula is as follows:

[0010] in Let be the weight hyperparameter updated for the i-th candidate basis function. The value of this parameter is negatively correlated with the contribution of the basis function. Let be the average weight value of the contribution of the i-th candidate basis function to the model output. The variance estimated for the weights of the i-th candidate basis function; Based on the updated weight hyperparameter values, all candidate basis functions are ranked by contribution level. Redundant basis functions with weight hyperparameters exceeding the preset engineering threshold are removed. The candidate basis functions with the highest contribution level are retained to form the first sparse observation model. The model structure is simplified by basis function screening, which reduces computational complexity while ensuring the accuracy and generalization ability of magnetic flux observation.

[0011] Optionally, the dynamic parameter tuning adopts a selective memory update strategy, the prediction variance is the core indicator for quantifying the uncertainty of new data points, and the incremental Bayesian learning framework is the mathematical algorithm basis for online incremental update of the model. The update process is as follows: Based on the current parameter covariance matrix of the first sparse observation model, the system measurement noise accuracy parameters, and the kernel function vector corresponding to the new data points containing multi-source electrical signals including real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed, the prediction variance of the new data points is calculated. The formula for calculating the prediction variance is as follows:

[0012] in The prediction variance for the new data points. The noise accuracy is the reciprocal of the system's noise variance. The kernel function vector is calculated by preserving the basis functions in the first sparse observation model for the new data points. ε is the transpose of the kernel function vector, and ε is the covariance matrix of the current parameters of the model; The predicted variance is compared with a preset dynamic change threshold, and new data points with uncertainty below the threshold are defined as informative data points. Informational data points are converted into corresponding design vectors, which are then spliced ​​and merged with the design vectors corresponding to historical valid data points to construct an augmented design matrix with expanded dimensions. Based on the incremental Bayesian learning framework, the covariance matrix and weight mean vector of the first sparse observation model are updated synchronously using the augmented design matrix. The parameter update formula is as follows:

[0013]

[0014] in, The inverse of the covariance matrix of the model before the update. The inverse of the covariance matrix of the updated model. For noise accuracy, Design vectors generated for informative data points. To design the transpose of a vector, This is the weight mean vector before the update. This is the updated weight mean vector. These are the actual magnetic flux observations corresponding to the informative data points. This is the updated model covariance matrix; Output the comprehensive flux linkage estimate based on the updated first sparse observation model; The updated model parameters output a comprehensive flux linkage estimate, enabling the model to continuously adapt to the dynamic electromagnetic characteristics of the motor under varying operating conditions and magnetic circuit saturation, maintaining observation accuracy without the need for offline retraining.

[0015] Optionally, the calculation process for the first temperature disturbance flux value is as follows: The direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed electrical parameters of the motor are synchronously acquired at the current and previous sampling times. Inverter nonlinear distortion compensation is performed on the direct-axis voltage and quadrature-axis voltage. The compensated voltage data, current data, and speed data are substituted into the reconstructed voltage equation transformation, which is the voltage equation of the permanent magnet synchronous motor in the rotating coordinate system.

[0016] in, For direct-axis and quadrature-axis voltages, For direct-axis and quadrature-axis currents, For stator resistance, For direct-axis and quadrature-axis inductors, The rotor's electric angular velocity, For permanent magnet flux linkage, the voltage equation is transformed into a reconstructed algebraic expression that can completely cancel out the interference from winding resistance and inductance parameters, allowing for the calculation of the current intermediate flux linkage value. Intermediate flux linkage value compared to the previous time step The intermediate flux linkage value is the intermediate flux linkage value obtained by substituting the direct-axis and quadrature-axis voltage, current, and speed electrical parameters of the motor at the current moment and the previous moment into the preset voltage equation transformation formula. The difference between the current intermediate flux linkage value and the previous intermediate flux linkage value is calculated to obtain the first temperature disturbance flux linkage value. The calculation formula is as follows:

[0017] in This is the first temperature-perturbed flux linkage value, used to quantify the flux linkage drift caused by the temperature change of the permanent magnet within adjacent sampling intervals. This represents the current intermediate flux linkage value. The intermediate flux linkage value at the previous moment is used to eliminate winding parameter interference through differential operation, and the flux linkage drift component caused by pure temperature is extracted.

[0018] Optionally, the verification and fusion process of the first temperature perturbation flux value is as follows: The first temperature perturbation flux value is used as input to the trained second sparse observation model, and the corresponding predicted temperature perturbation flux value is obtained through the calculation of the second sparse observation model. Calculate the absolute difference between the first temperature perturbation flux linkage value and the predicted temperature perturbation flux linkage value, and define the absolute difference as the degree of difference between the two. The dynamic fusion weighting coefficient W within the 0-1 range is matched based on the difference value. The matching method is as follows: the smaller the difference, the greater the weight of the predicted temperature perturbation flux linkage value; the weight is 1 when the difference is zero, and 0 when the difference is greater than a preset upper limit. Linear or exponential interpolation is used in between. This weighting coefficient is negatively correlated with the difference. The two flux linkage values ​​are weighted using the dynamic fusion weighting coefficient to obtain the second temperature perturbation flux linkage value. The weighted fusion formula is as follows:

[0019] in Here, W represents the second temperature-induced magnetic flux linkage value, and W is the dynamic fusion weighting coefficient. The first temperature disturbance flux value, To predict temperature perturbation flux linkage values ​​for the model, weighted fusion is used to suppress instantaneous noise in physical calculations, combining the directness of the physical model with the robustness of the data model to improve the accuracy of temperature perturbation flux linkage values.

[0020] Optionally, the implementation process of the interlock judgment mechanism is as follows: Extract the historical statistical mean of the feature deviation sequence under healthy motor operating conditions. Historical statistical standard deviation Combined with the load factor of the motor under current operating conditions The adaptive early warning threshold is calculated using the following formula:

[0021] in The adaptive warning threshold is given by k, where k is the confidence coefficient. This represents the historical statistical mean of the characteristic deviation sequence under healthy operating conditions. The historical statistical standard deviation of the characteristic deviation sequence under healthy operating conditions. is the load factor of the motor under the current operating condition, and is the ratio of the actual torque under the current operating condition to the rated torque of the motor. The actual torque can be calculated from the real-time current and flux linkage through the motor electromagnetic torque equation. The characteristic deviation sequence values ​​of multiple control cycles are continuously collected, and it is determined whether the characteristic deviation sequence values ​​continuously exceed the adaptive early warning threshold to form the first judgment condition. A time-series trend analysis is performed on the benchmark compensation torque sequence within a preset time window. The fitting slope S of the sequence is calculated by linear regression. If the fitting slope S is negative and its absolute value is greater than the preset noise threshold, a second judgment condition is formed. The first and second judgment conditions are logically ANDed, and the result is used to determine whether a demagnetization fault has occurred. By verifying both instantaneous anomalies and long-term degradation, the system avoids misjudgments of faults caused by transient operating disturbances and noise.

[0022] Optionally, the method further includes a fault determination confidence enhancement step: From all the basis functions of the first sparse observation model, the basis functions that are directly related to the magnetic field strength of the permanent magnet are selected and defined as key basis functions. The weights of these basis functions can directly characterize the magnetic field strength of the permanent magnet. Specifically, the key basis functions are the Pearson correlation coefficient between the output value of each basis function in the first sparse observation model and the measured value of the permanent magnet flux linkage, and the weight contribution of each basis function in the flux linkage estimation. The basis functions whose Pearson correlation coefficient is higher than a preset correlation threshold, whose weight contribution is higher than a preset weight threshold, and whose output value decreases monotonically with the decay of the permanent magnet magnetic field strength are defined as key basis functions that are directly related to the magnetic field strength of the permanent magnet. Extract the weight values ​​of key basis functions in continuous running cycles and construct a time series sequence of key basis function weights; Trend fitting analysis is performed on the time series of key basis function weights. When the weight time series shows a continuous decay trend and the absolute value of the decay slope is greater than the preset noise threshold, the confidence of demagnetization fault determination is improved. When the trend is stable or there is no significant decay, the original confidence level is maintained. An evidence loop is formed by the changes in internal model parameters and external torque characteristics, which strengthens the accuracy of fault determination.

[0023] Optionally, after constructing the dual-channel flux linkage observer and before outputting the comprehensive flux linkage estimate, an inverter nonlinear distortion compensation step is also included: The direct-axis voltage compensation of the inverter under different current conditions was obtained through offline calibration. Cross-axis voltage compensation Establish a one-to-one mapping relationship between current operating conditions and voltage compensation quantities; Match the corresponding direct-axis voltage compensation amount according to the real-time current condition of the motor. Cross-axis voltage compensation ; The matched voltage compensation amount is used to compare the real-time acquired raw direct-axis voltage. Original quadrature axis voltage Numerical correction is performed to obtain the compensated direct-axis voltage. With cross-axis voltage The voltage compensation formula is:

[0024]

[0025] in To compensate for the direct-axis voltage, This is the raw direct-axis voltage acquired in real time.

[0026] This is the direct-axis voltage compensation amount. To compensate for the quadrature-axis voltage, The raw quadrature-axis voltage is acquired in real time. This is the quadrature axis voltage compensation amount; The compensated voltage data is used to calculate the flux linkage value for subsequent temperature disturbances, eliminating nonlinear distortions caused by inverter switching dead zones and power device voltage drops, and ensuring that the voltage data is consistent with the actual electromagnetic state of the motor.

[0027] Optionally, the multi-source electrical signal acquisition method is as follows: Hall effect current sensors, resistive voltage divider voltage sensors, and rotary transformer speed sensors are used to collect analog signals of the motor's phase current, terminal voltage, and rotor speed, respectively. The analog-to-digital conversion of all acquired signals is synchronously triggered by the unified master clock signal of the motor controller, ensuring that the timing of multi-channel signal acquisition is fully aligned. The digital signal after analog-to-digital conversion is synchronously sampled at a preset sampling frequency higher than the fundamental frequency of the motor to form a multi-source electrical signal sequence, avoiding signal aliasing and channel delay errors, and providing high-fidelity raw data for subsequent calculations.

[0028] Based on the same inventive concept, this invention also provides a two-dimensional dynamic torque distribution prediction system for permanent magnet motor demagnetization fault diagnosis, the system comprising: Signal acquisition module: Collects multi-source electrical signals such as motor direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed, constructs a dynamic timing data window, and outputs a timing-aligned multi-source timing dataset; Observer building module: The first sparse observation model is trained based on a multi-source time series dataset through a contribution competition mechanism, and the second sparse observation model is trained based on the temperature perturbation flux signal calculated by physics, thus building a dual-channel flux observer. Model update and output module: Input the real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed multi-source electrical signals into the first sparse observation model, adjust the model parameters using a selective memory update strategy, and output the comprehensive flux linkage estimate. Torque prediction module: Combines the comprehensive flux linkage estimate with the real-time current to calculate the two-dimensional dynamic comprehensive prediction torque; Temperature perturbation calculation module: Calculates the first temperature perturbation flux value based on the electrical parameters at adjacent time points through voltage equation differential transformation, and outputs the second temperature perturbation flux value after verification and fusion by the second sparse observation model; Temperature compensation module: The temperature compensation flux is obtained by using the second temperature disturbance flux value to compensate the comprehensive flux estimation value, and the reference compensation torque is calculated in combination with the real-time current solution. Fault determination module: Generates a demagnetization fault characteristic deviation sequence, completes the demagnetization fault determination through an interlocking judgment mechanism, and improves the confidence of fault determination by combining the weight trend of key basis functions.

[0029] Compared with the prior art, the present invention has the following advantages: I. This invention constructs a dual-channel flux linkage observer, integrating a dual-sparse observation model trained with multi-source time-series data and temperature-disturbance flux linkage signals. This enables accurate observation and fusion processing of flux linkage signals, effectively eliminating the interference of temperature factors on flux linkage observation results, avoiding diagnostic biases caused by temperature disturbances at the source. Utilizing a selective memory update strategy, the model dynamically adjusts parameters, quantifies data uncertainty, filters effective information points, and continuously optimizes model parameters and output accuracy. This allows the flux linkage estimate to better reflect the actual operating state of the motor. Combined with inverter nonlinear distortion compensation, it corrects errors in voltage acquisition data, further improving the accuracy of electrical parameter calculations. This makes the flux linkage and torque calculation results more reliable, significantly enhancing the anti-interference capability and data reliability of the fault diagnosis process, ensuring that diagnostic results are unaffected by system nonlinear factors and external disturbances.

[0030] II. This invention combines two-dimensional dynamic torque distribution prediction with temperature-coordinated compensation to achieve temperature compensation of the flux linkage and calculation of the reference torque. It generates a characteristic deviation sequence based on the difference between the comprehensively predicted torque and the reference compensated torque, accurately capturing abnormal parameter changes caused by demagnetization faults. An interlocking judgment mechanism integrates deviation over-limit judgment with torque time-series trend analysis. Fault determination is completed through dual logic verification, avoiding missed or false judgments caused by single-feature judgments. The combination of key basis function weighted time-series analysis enhances the confidence of fault determination. Furthermore, multi-source signal synchronous acquisition and time-series data window construction ensure the time-series consistency and integrity of operating data, comprehensively improving the accuracy, response speed, and adaptability of permanent magnet motor demagnetization fault diagnosis, providing reliable fault monitoring assurance for stable motor operation.

[0031] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0032] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating a two-dimensional dynamic torque distribution prediction method for diagnosing demagnetization faults in permanent magnet motors, according to an embodiment of the present invention.

[0034] Figure 2 This is a time-series comparison diagram of the feature deviation sequence and the adaptive threshold in an embodiment of the present invention.

[0035] Figure 3 This is a biaxial graph showing the long-term degradation trend of the benchmark compensation torque and the verification of model parameters in an embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the structure of the permanent magnet motor demagnetization fault diagnosis system for two-dimensional dynamic torque distribution prediction according to an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0038] Reference Figure 1 One embodiment of the present invention proposes a two-dimensional dynamic torque distribution prediction method for diagnosing demagnetization faults in permanent magnet motors. It employs a data processing mechanism that integrates a dual-channel self-verifying observer, differential transformation of physical equations, and multi-dimensional interlocking decision-making. This method can accurately separate temperature interference and demagnetization signals without relying on additional temperature sensors, and provides highly robust and high-confidence online diagnosis of motor demagnetization faults. It is adaptable to the online safety monitoring and fault early warning needs of permanent magnet motors in various fields such as industrial transmission, new energy vehicles, and rail transit.

[0039] The method described in this embodiment, according to the specific implementation process of each step, mathematical operation logic, hardware and software cooperation method, and exemplary implementation details, is as follows: Synchronous acquisition of multi-source electrical signals and construction of dynamic time-series datasets: This step is used to complete the high-precision acquisition of multi-dimensional electrical signals of the permanent magnet motor's operating status, synchronous timing calibration, and standardized dataset construction. By using a unified clock trigger, high-frequency sampling, and sliding window timing alignment, signal aliasing, channel delay, and abnormal noise interference are eliminated, providing a time-uniform and data-fidelity-preserving original input foundation for subsequent sparse observation model training, flux linkage calculation, and torque calculation.

[0040] Detailed implementation process: Hardware deployment and signal type determination: Dedicated sensing units are deployed in the permanent magnet motor body and drive system: Hall effect current sensors are used to collect the phase current of the motor's three-phase windings; resistive voltage divider voltage sensors are used to directly collect the stator terminal voltage of the motor; and rotary transformer speed sensors are installed coaxially with the motor rotor to collect the rotor speed analog signal in real time.

[0041] Signal coordinate transformation processing: The coordinate transformation unit of the motor controller performs Park transformation on the three-phase currents, converting them into direct-axis currents in a rotating coordinate system. quadrature axis current Perform coordinate transformation on the stator terminal voltage to convert it into a direct-axis voltage. quadrature axis voltage ; and rotor speed The signals together constitute a multi-source electrical signal.

[0042] Synchronous analog-to-digital conversion trigger control: Using the unified master clock signal of the motor main controller as the synchronization reference, the analog-to-digital conversion operation of all sensor signals is triggered simultaneously to ensure that the five signals of direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed are sampled at the same time, so as to achieve complete timing alignment of multi-channel signals and eliminate phase errors caused by inter-channel delay.

[0043] High-frequency sampling and anti-aliasing processing: Sampling is performed at a preset sampling frequency of 10kHz, which is much higher than the fundamental frequency of the motor at its rated speed. This allows for the complete capture of dynamic characteristics such as current harmonics and voltage transients, thus avoiding signal aliasing. During the sampling process, high-frequency electromagnetic interference is filtered out by hardware low-pass filtering, while retaining the effective electromagnetic signal.

[0044] Dynamic time-series data window construction and dataset generation: A dynamic time-series data window is constructed according to the rule of a 100-point sliding window, and the sampled data is time-series aligned and filtered; the 3σ principle is used to remove abnormal data points that exceed the normal range, and the effective data is arranged in the order of sampling time. Finally, a multi-source time-series dataset containing direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed is generated. The dataset dimension is completely matched with the sampling duration and window size.

[0045] Exemplary implementation: In this embodiment, the test object is a permanent magnet synchronous motor with a rated power of 15kW and a rated speed of 3000r / min. A Hall effect current sensor with a bandwidth of 100kHz, a resistive voltage divider with an input impedance of 10MΩ, and a rotary transformer with a resolution of 4096 lines are deployed. The analog-to-digital conversion is synchronously triggered by the motor controller's 20MHz master clock, and the signal is continuously acquired at a sampling frequency of 10kHz. At a certain sampling moment, the direct-axis voltage of 310.5V, the quadrature-axis voltage of 180.2V, the direct-axis current of -50A, the quadrature-axis current of 200A, and the rotor speed of 1500r / min are synchronously captured. After time alignment and outlier removal, this set of data is included in the sliding window. After continuous acquisition for 1 second, a multi-source time-series dataset containing 10,000 valid data points is constructed. The timestamp error of all data points is less than 1μs, which meets the accuracy requirements for model training and parameter calculation.

[0046] Dual-channel flux linkage observer setup and sparse observation model training: This step trains the first sparse observation model and the second sparse observation model separately through a contribution competition mechanism. The two models are then combined to build a dual-channel flux linkage observer, enabling independent calculation and cross-validation of the integrated flux linkage and the temperature perturbation flux linkage. This simplifies the model structure, reduces computational complexity, and ensures the accuracy and generalization ability of flux linkage observation.

[0047] Detailed implementation process: Basis function library construction and initial design matrix generation: A basis function library containing four types of nonlinear transformation functions, namely polynomial functions, exponential functions, Gaussian functions, and radial basis functions, is built. This basis function library can cover the nonlinear electromagnetic characteristics of the motor under all operating conditions. According to the dimension of the multi-source time series dataset, the corresponding candidate basis function is matched for each data point, and a high-dimensional initial design matrix Φ with the same dimension as the dataset is constructed, providing a complete candidate pool for basis function selection.

[0048] Training the first sparse observation model: Using the flux linkage observations obtained from offline calibration based on the voltage equation in the multi-source time-series dataset as the model training target values, the expectation-maximization algorithm is used to iteratively update the weight hyperparameters of each candidate basis function in the initial design matrix. The formula is as follows:

[0049] in Let be the weight hyperparameter updated for the i-th candidate basis function. The value of this parameter is negatively correlated with the contribution of the basis function. Let be the average weight value of the contribution of the i-th candidate basis function to the model output. The variance of the weight estimate for the i-th candidate basis function is used to characterize the uncertainty in weight calculation. After iterative training until the weight hyperparameters converge or the maximum number of iterations is reached, the candidate basis functions are sorted by contribution level according to the weight hyperparameter values. Redundant basis functions with weight hyperparameters greater than a preset engineering threshold of 106 are removed, and the combination of basis functions with the highest contribution is retained to form the first sparse observation model.

[0050] Second sparse observation model training: In the model training stage, multi-source electrical signals of the motor are collected in advance under different ambient temperatures and different load conditions. The physical temperature perturbation flux sample signals under different temperature gradients are calculated by voltage equation differential transformation. The sample signal is used as the exclusive training target, and the contribution competition mechanism and expectation maximization algorithm are completely consistent with the first sparse observation model to complete the training. This model is only for learning temperature perturbation flux features and is specifically used for the verification, purification and noise suppression of temperature perturbation flux.

[0051] Dual-channel flux linkage observer integration: The first sparse observation model and the second sparse observation model are integrated in parallel. The first model is responsible for solving the comprehensive flux linkage that includes all electromagnetic effects, while the second model is responsible for solving the flux linkage drift caused by pure temperature. The output results of the two models are cross-validated to form a dual-channel flux linkage observer.

[0052] Exemplary implementation: For a multi-source time-series dataset containing 10,000 data points, an initial design matrix is ​​constructed, and the expectation-maximization algorithm is iteratively trained with the magnet link observation value as the objective. In the fifteenth iteration, the average weight value of candidate basis function number 0527 is... The variance of the weighted estimate is 0.0002. The value is 0.00001. Substituting this into the formula, we get... This value subsequently exceeded The engineering threshold was determined to be redundant and eliminated; finally, 600 high-contribution basis functions were retained from 10,000 candidate basis functions to form the first sparse observation model, which reduced the computational complexity of the model by 85%; the training of the second sparse observation model was completed simultaneously, and a dual-channel magnetron observer was built by combining them.

[0053] Inverter nonlinear distortion compensation, model dynamic updating, and predicted torque calculation: This step first uses a selective memory update strategy to dynamically optimize the parameters of the first sparse observation model, making the model adapt to the dynamic characteristics of the motor under varying operating conditions and magnetic circuit saturation, outputting a comprehensive flux linkage estimate and solving for the two-dimensional dynamic comprehensive predicted torque. Before the subsequent temperature disturbance flux linkage calculation, offline calibration is used to correct the voltage nonlinearity distortion caused by the inverter switching dead zone and the voltage drop of power devices, providing distortion-free voltage data input for the differential transformation of the voltage equation.

[0054] Detailed implementation process: Offline calibration and voltage compensation for inverter nonlinear distortion: Through an offline calibration program, the deviation between the inverter command voltage and the actual motor output voltage is measured under different current directions and amplitudes to obtain the direct-axis voltage compensation amount corresponding to each operating condition. Quadrature axis voltage compensation A one-to-one mapping table of current operating conditions and voltage compensation amounts is established and stored in the controller's non-volatile memory; during real-time operation, the corresponding compensation amount is matched according to the real-time current operating conditions of the motor, and the original direct-axis voltage is adjusted accordingly. Original quadrature axis voltage The correction and compensation formula is as follows:

[0055]

[0056] in To compensate for the direct-axis voltage, This is the raw direct-axis voltage acquired in real time. This is the direct-axis voltage compensation amount. To compensate for the quadrature-axis voltage, The raw quadrature-axis voltage is acquired in real time. This is the quadrature axis voltage compensation amount; The compensated voltage eliminates the inverter's nonlinear distortion and remains consistent with the actual electromagnetic state of the motor.

[0057] New data point prediction variance calculation and informative data filtering: based on the current parameter covariance matrix ε of the first sparse observation model and the accuracy of system measurement noise. New data point kernel function vector The formula for calculating the prediction variance of new data points (quantifying data uncertainty) is as follows:

[0058] in The prediction variance for the new data points. The noise accuracy is the reciprocal of the system's noise variance. The kernel function vector is calculated by preserving the basis functions in the first sparse observation model for the new data points. ε is the transpose of the kernel function vector, and ε is the covariance matrix of the current parameters of the model; The predicted variance is compared with a preset dynamic threshold of 0.002, and new data points with uncertainty below the threshold are selected and defined as informative data points. This type of data can be used for model refinement updates.

[0059] Online incremental model update: converting informative data points into corresponding design vectors The design vectors from historical valid data points are concatenated and fused to construct an augmented design matrix. Based on the incremental Bayesian learning framework, the model covariance matrix and weight mean vector are updated synchronously using the following formula:

[0060]

[0061] in, The inverse of the covariance matrix of the model before the update. The inverse of the covariance matrix of the updated model. For noise accuracy, Design vectors generated for informative data points. To design the transpose of a vector, This is the weight mean vector before the update. This is the updated weight mean vector. These are the actual magnetic flux observations corresponding to the informative data points. This is the updated model covariance matrix.

[0062] Two-dimensional dynamic comprehensive predicted torque solution: Input the real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed into the updated first sparse observation model, and output the comprehensive flux linkage estimate; combine the real-time current parameters, substitute them into the motor electromagnetic torque equation, and solve to obtain the two-dimensional dynamic comprehensive predicted torque over the entire operating condition range.

[0063] Exemplary implementation: In this embodiment, for a multi-source time-series dataset containing 10,000 data points, an initial design matrix of corresponding dimensions is constructed. The expectation-maximization algorithm is initiated with the magnet link observations as the training objective, and the maximum number of iterations is set to 20. During the fifteenth iteration, the average weight value of candidate basis function number 0527 is... Weighted estimation variance Substituting into the formula, we get This value exceeded [a certain value] in subsequent iterations. The engineering threshold was used to identify redundant basis functions for elimination. Finally, 600 high-contribution basis functions were retained from 10,000 candidate basis functions to form the first sparse observation model, which reduced the computational complexity of the model by 85% and controlled the flux linkage solution error to within 0.2%. The training of the second sparse observation model was completed simultaneously. After integrating the two models, the dual-channel flux linkage observer can simultaneously output the comprehensive flux linkage and the temperature perturbation flux linkage, with a cross-validation error of less than 0.1%.

[0064] Integration of physical calculations and model validation for temperature-induced magnetic flux disturbances: This step uses differential transformation of the motor voltage equation to eliminate interference from winding resistance and inductance parameters, and extracts the flux drift component caused by pure temperature (the first temperature disturbance flux value). Then, the second sparse observation model is used to verify and weightedly fuse this value to obtain a high-purity, low-noise second temperature disturbance flux value, which provides accurate input for subsequent temperature compensation.

[0065] Detailed implementation process: Physical calculation of the first temperature disturbance flux linkage value: Synchronously acquire the direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed electrical parameters of the motor at the current and previous sampling times. Perform inverter nonlinear distortion compensation on the direct-axis voltage and quadrature-axis voltage. Substitute the compensated voltage data, current data, and speed data into the reconstructed voltage equation transformation formula. The voltage equation transformation formula is the voltage equation of the permanent magnet synchronous motor in the rotating coordinate system.

[0066] in, For direct-axis and quadrature-axis voltages, For direct-axis and quadrature-axis currents, For stator resistance, For direct-axis and quadrature-axis inductors, The rotor's electric angular velocity, For permanent magnet flux linkage, the current value of the intermediate flux linkage is calculated. Intermediate flux linkage value at the previous moment The first temperature disturbance flux linkage value is obtained by performing a difference operation on the two sets of intermediate flux linkage values, using the following formula:

[0067] in This is the first temperature-perturbed flux linkage value, used to quantify the flux linkage drift caused by the temperature change of the permanent magnet within adjacent sampling intervals. The intermediate flux linkage value at the current moment is obtained by substituting the direct-axis and quadrature-axis voltage, current, and speed parameters of the motor at the current moment and the previous moment into a preset voltage equation and solving the transformation formula. The intermediate flux linkage value at the previous moment is used to eliminate winding parameter interference through differential operation and extract the flux linkage drift component caused by pure temperature. This value directly quantifies the flux drift caused by the temperature change of the permanent magnet within adjacent sampling intervals, without interference from winding parameters.

[0068] Temperature perturbation flux linkage model verification: The first temperature perturbation flux linkage value is input into the pre-trained second sparse observation model. Based on the learned temperature perturbation features, the model outputs the corresponding predicted temperature perturbation flux linkage value.

[0069] Dynamic weighted fusion and generation of the second temperature perturbation flux linkage value: The absolute difference between the first temperature perturbation flux linkage value and the predicted temperature perturbation flux linkage value is calculated, and this difference is defined as the degree of difference between the two. The degree of difference is negatively correlated with the dynamic fusion weight coefficient W, and the weight coefficient W in the range of 0-1 is matched according to the magnitude of the degree of difference. The two flux linkage values ​​are calculated using a weighted fusion formula to obtain the second temperature perturbation flux linkage value. The formula is as follows:

[0070] in Here, W represents the second temperature-induced magnetic flux linkage value, and W is the dynamic fusion weighting coefficient. The first temperature disturbance flux value, To predict temperature perturbation flux linkage values ​​for the model, weighted fusion is used to suppress instantaneous noise in physical calculations, combining the directness of the physical model with the robustness of the data model to improve the accuracy of temperature perturbation flux linkage values. The fused signal suppresses instantaneous noise in physical calculations, balancing the directness of the physical model with the robustness of the data model.

[0071] Exemplary implementation: In this embodiment, electrical parameters at the current moment and the previous sampling moment are collected simultaneously, and the intermediate flux linkage value at the current moment is calculated by substituting them into the voltage equation transformation formula. The intermediate flux value at the previous moment Substituting into the difference formula, we obtain the first temperature disturbance flux value. Input this value into the second sparse observation model to obtain the predicted temperature perturbation flux value. The degree of difference between the two is calculated as follows: The difference is small, and the dynamic matching weight coefficient is used. After weighted fusion calculation This value represents the purified temperature-perturbed flux linkage value, with a 60% reduction in noise amplitude, accurately characterizing the effect of temperature on flux linkage.

[0072] Calculation of flux linkage temperature co-compensation and reference compensation torque: This step uses the second temperature disturbance flux value to perform temperature-coordinated compensation on the comprehensive flux estimate, completely eliminating the interference of temperature factors on flux and torque calculation; then, based on the compensated flux solution benchmark, the torque is compensated, and a demagnetization fault judgment benchmark without temperature interference is established, realizing the accurate separation of temperature signal and demagnetization signal.

[0073] Detailed implementation process: Integrated flux linkage temperature co-compensation: The second temperature disturbance flux linkage value is used as the compensation correction amount and substituted into the flux linkage compensation calculation to correct the integrated flux linkage estimate output by the first sparse observation model point by point, eliminating the flux linkage drift caused by temperature and obtaining the temperature-compensated flux linkage; this flux linkage only contains the inherent magnetic properties of the permanent magnet and the electromagnetic coupling characteristics of the motor, without any temperature interference component.

[0074] Reference compensation torque calculation: Substitute the temperature compensation flux and the real-time direct-axis current and quadrature-axis current into the electromagnetic torque equation of the permanent magnet motor, and complete the flux-torque conversion calculation step by step to obtain the reference compensation torque; this torque eliminates all torque fluctuations caused by temperature changes, and can truly reflect the magnetic performance state of the permanent magnet itself, serving as the sole standard reference for judging demagnetization faults.

[0075] Exemplary implementation: In this embodiment, a second temperature disturbance flux linkage of 0.0007775Wb is used as a correction value to compensate the comprehensive flux linkage estimate output by the first sparse observation model point by point, resulting in a temperature-compensated flux linkage without temperature interference. This flux linkage, along with the real-time direct-axis current of -48A and quadrature-axis current of 205A, is substituted into the electromagnetic torque equation. After coordinate transformation and torque coefficient calculation, the reference compensation torque is obtained. This torque fluctuates by less than 0.1Nm during the process of the motor temperature rising from 25℃ to 120℃, completely eliminating the torque interference caused by temperature and accurately reflecting the true state of the permanent magnet's magnetic properties.

[0076] Demagnetization fault feature generation, interlock determination, and confidence enhancement: This step generates a demagnetization fault characteristic deviation sequence based on torque difference, combines adaptive threshold and dual logic interlocking judgment mechanism to complete fault determination, and then forms an evidence loop through key basis function weight trend analysis, significantly improving the confidence of fault diagnosis and eliminating misjudgment and missed judgment under complex operating conditions. Figure 2 , Figure 3 As shown.

[0077] Detailed implementation process: Demagnetization fault characteristic deviation sequence generation: The difference between the two-dimensional dynamic comprehensive predicted torque and the reference compensated torque is calculated in real time. The difference is arranged in the order of sampling time to generate the demagnetization fault characteristic deviation sequence. The amplitude change of this sequence directly corresponds to the torque abnormality caused by permanent magnet demagnetization and is completely unrelated to temperature interference.

[0078] Adaptive early warning threshold calculation: Extract the historical statistical mean of the feature deviation sequence under healthy motor operating conditions. Historical statistical standard deviation

[0079] in The adaptive warning threshold is given by k, where k is the confidence coefficient. This represents the historical statistical mean of the characteristic deviation sequence under healthy operating conditions. The historical statistical standard deviation of the characteristic deviation sequence under healthy operating conditions. This represents the load factor of the motor under its current operating conditions.

[0080] Interlock judgment mechanism implementation: First judgment condition: Collect the characteristic deviation sequence values ​​for 12 consecutive control cycles and determine whether the sequence values ​​continuously exceed the adaptive early warning threshold. The second judgment condition is to perform a time-series linear regression analysis on the benchmark compensation torque sequence for 20 driving hours, calculate the sequence fitting slope S, and determine that the slope S is negative and the absolute value is greater than 0.01 preset noise threshold. Fault determination: Perform a logical AND operation between the first determination condition and the second determination condition. Only when both conditions are met simultaneously, the motor is determined to have a demagnetization fault.

[0081] Enhanced Confidence in Fault Judgment: From the first sparse observation model, key basis functions directly related to the magnetic field strength of the permanent magnet (accounting for 1.5% of the total number of basis functions) are selected; the weight values ​​of the key basis functions in continuous operating cycles are extracted to construct a weight time series; linear regression trend analysis is performed on the weight series. When the weight series shows a continuous decay trend and the absolute value of the decay slope is greater than the 0.02% engineering threshold, the confidence of demagnetization fault judgment is increased from 90% to 99.5%, and the fault alarm level is upgraded from Level 1 warning to Level 2 fault; when the trend is stable or there is no significant decay, the original confidence level is maintained without changing the fault judgment result. The continuous decay of the key basis function weights and the external torque degradation form a complete evidence loop, enhancing the accuracy of fault judgment.

[0082] Exemplary implementation: In this embodiment, the historical mean of the feature deviation sequence under the motor's healthy state is extracted. Standard deviation Combined with the current load factor Substituting into the formula, the adaptive early warning threshold is calculated. The real-time monitoring characteristic deviation sequence value remained at 2.8 Nm and exceeded the threshold for 12 consecutive control cycles, meeting the first judgment condition; linear regression analysis was performed on the benchmark compensation torque sequence over 20 driving hours to obtain the fitting slope. If the absolute value of the noise level exceeds 0.01 for 20 hours, the second judgment condition is met. A logical AND operation is then performed to determine that the motor has a demagnetization fault. Subsequently, 150 key basis functions are extracted from the first sparse observation model, and their weight sequences over 20 hours of operation are analyzed. The average decay slope is found to be -0.08% / hour, exceeding the 0.02% engineering threshold. The continuous decay of the key basis function weights and the external torque degradation form a complete evidence loop, increasing the fault diagnosis confidence from 90% to 99.5%. Simultaneously, the fault alarm level is upgraded from Level 1 warning to Level 2 fault, providing maintenance personnel with accurate decision-making support.

[0083] Reference Figure 4 This embodiment also provides a two-dimensional dynamic torque distribution prediction system for permanent magnet motor demagnetization fault diagnosis, used to perform the above-mentioned diagnostic method. The system includes: Signal acquisition module: used to collect multi-source electrical signals from the motor, construct a dynamic time-series data window based on the timing relationship, and obtain a multi-source time-series dataset; Observer building module: used to train the first sparse observation model through a contribution competition mechanism, train the second sparse observation model based on the temperature perturbation flux signal calculated by physics, and build a dual-channel flux observer; Model update and output module: It is used to take in real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed multi-source electrical signals, and dynamically adjusts parameters using a selective memory update strategy, and outputs a comprehensive flux linkage estimate. Torque prediction module: used to calculate the comprehensive predicted torque by combining the comprehensive flux linkage estimate and the real-time current; Temperature disturbance calculation module: used to calculate the first temperature disturbance flux value through differential transformation of motor voltage equation, and output the second temperature disturbance flux value after verification and fusion by the second sparse observation model; Temperature compensation module: used to compensate the comprehensive flux estimate value using the second temperature disturbance flux value, and calculate the reference compensation torque; Fault determination module: Used to generate characteristic deviation sequence, complete demagnetization fault determination through interlocking judgment mechanism, and perform confidence enhancement operation.

[0084] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for diagnosing demagnetization faults in permanent magnet motors based on two-dimensional dynamic torque distribution prediction, characterized in that, The method includes: Multi-source electrical signals are collected from the motor and a dynamic time-series data window is constructed based on the timing relationship to form a multi-source time-series dataset. Based on a multi-source time-series dataset, the flux observation values ​​obtained through offline calibration calculation of the voltage equation are used as the training target to train the first sparse observation model, and the second sparse observation model is trained based on the temperature perturbation flux sample signals obtained in advance during the training phase, thus forming a dual-channel flux observer. The real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed multi-source electrical signals are input into the first sparse observation model. After dynamic parameter adjustment, the comprehensive flux linkage estimate is output, and then the comprehensive predicted torque is calculated by combining it with the real-time current. Based on the electrical parameters at adjacent time points after nonlinear distortion compensation by the inverter, the first temperature disturbance flux value is calculated by differential transformation of the motor voltage equation, and the second temperature disturbance flux value is output after verification and fusion by the second sparse observation model. The temperature-compensated flux is obtained by using the second temperature disturbance flux value to compensate the comprehensive flux estimate, and the base compensation torque is calculated by combining the real-time current. A characteristic deviation sequence is generated by the difference between the comprehensive predicted torque and the reference compensated torque. The demagnetization fault is determined by an interlocking judgment mechanism in combination with the trend of the reference compensated torque.

2. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The first sparse observation model is obtained through a contribution competition mechanism, the implementation process of which is as follows: A basis function library is constructed by selecting several nonlinear transformation functions, including polynomial functions, exponential functions, and Gaussian functions. Based on the data dimensions of the multi-source time series dataset, a corresponding number of candidate basis functions are selected from the basis function library. An initial design matrix matching the data dimensions is constructed based on the candidate basis functions. Using the flux linkage observations obtained from the offline calibration calculation based on the voltage equation in the multi-source time series dataset as the model training target value, the expectation-maximization algorithm is used to iteratively update the weight hyperparameters corresponding to each candidate basis function in the initial design matrix. Based on the updated weight hyperparameter values, all candidate basis functions are ranked by contribution level. Candidate basis functions whose weight hyperparameters exceed a preset threshold are removed, and the remaining candidate basis functions with the highest contribution levels are combined to form the first sparse observation model.

3. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The dynamic parameter tuning employs a selective memory update strategy, and the implementation process of the selective memory update strategy is as follows: Based on the current parameter covariance matrix of the first sparse observation model, the system measurement noise accuracy parameters, and the kernel function vector corresponding to the new data points containing real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed multi-source electrical signals, the prediction variance of the new data points is calculated to quantify the uncertainty. The predicted variance is compared with a preset dynamic change threshold, and new data points with uncertainty below the threshold are defined as informative data points. Informational data points are converted into corresponding design vectors, which are then spliced ​​and merged with the design vectors corresponding to historical valid data points to construct an augmented design matrix with expanded dimensions. Based on the incremental Bayesian learning framework, the covariance matrix and weight mean vector of the first sparse observation model are updated synchronously using the augmented design matrix, and the comprehensive magnet linkage estimate is output based on the updated model parameters.

4. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The calculation process for the first temperature disturbance flux value is as follows: Synchronously collect the direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, and rotor speed electrical parameters of the motor at the current running time and the previous sampling time; Substitute the electrical parameters of the current time and the previous time into the preset voltage equation transformation formula to calculate the intermediate magnetic flux value of the current time and the intermediate magnetic flux value of the previous time respectively. The difference between the current intermediate flux linkage value and the previous intermediate flux linkage value is calculated to obtain the first temperature disturbance flux linkage value.

5. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The verification and fusion process of the first temperature disturbance flux value is as follows: The first temperature perturbation flux value is used as input to the trained second sparse observation model, and the corresponding predicted temperature perturbation flux value is obtained through the calculation of the second sparse observation model. Calculate the absolute difference between the first temperature perturbation flux linkage value and the predicted temperature perturbation flux linkage value, and define the absolute difference as the degree of difference between the two. Based on the difference value, a dynamic fusion weighting coefficient within the 0-1 range is matched. The dynamic fusion weighting coefficient is used to perform a weighted operation on the first temperature perturbation flux linkage value and the predicted temperature perturbation flux linkage value to obtain the second temperature perturbation flux linkage value.

6. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The implementation process of the interlock judgment mechanism is as follows: Extract the historical statistical mean and historical statistical standard deviation of the feature deviation sequence under the motor's healthy operating condition, and combine them with the load factor of the motor's current operating condition to calculate and generate an adaptive early warning threshold; The characteristic deviation sequence values ​​are continuously collected for several control cycles. It is determined whether the characteristic deviation sequence values ​​continuously exceed the adaptive early warning threshold, thus forming the first judgment condition. A time-series trend analysis is performed on the benchmark compensation torque sequence within a preset time window. The fitting slope of the sequence is calculated by linear regression. The numerical attribute and absolute value of the fitting slope are determined to form the second judgment condition. Perform a logical AND operation between the first and second judgment conditions, and determine whether a demagnetization fault has occurred based on the result of the operation.

7. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 6, characterized in that, The method further includes a fault determination confidence enhancement step: From all the basis functions of the first sparse observation model, the basis functions that are directly related to the magnetic field strength of the permanent magnet are selected and defined as key basis functions; Extract the weight values ​​of key basis functions in continuous running cycles and construct a time series sequence of key basis function weights; Trend fitting analysis is performed on the time series sequence of key basis function weights. When the weight time series sequence shows a continuous decay trend and the absolute value of the decay slope is greater than the preset noise threshold, the confidence of demagnetization fault determination is improved. When the trend is stable or shows no significant decay, maintain the original confidence level.

8. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, Before calculating the first temperature disturbance flux value based on the electrical parameters at adjacent time points, an inverter nonlinear distortion compensation step is also included: The direct-axis voltage compensation and quadrature-axis voltage compensation of the inverter under different current conditions are obtained by offline calibration, and the corresponding mapping relationship between current conditions and voltage compensation is established. Match the corresponding direct-axis voltage compensation and quadrature-axis voltage compensation based on the real-time current conditions of the motor. The original direct-axis voltage and original quadrature-axis voltage acquired in real time are numerically corrected using the matched voltage compensation amount to obtain the compensated direct-axis voltage and quadrature-axis voltage. The compensated voltage data is then used for the subsequent calculation of the temperature disturbance flux value.

9. The method for diagnosing demagnetization faults of permanent magnet motors based on two-dimensional dynamic torque distribution prediction according to claim 1, characterized in that, The multi-source electrical signal acquisition method is as follows: Hall effect current sensors, resistive voltage divider voltage sensors, and rotary transformer speed sensors are used to collect the phase current, terminal voltage, and rotor speed signals of the motor, respectively. The analog-to-digital conversion of all acquired signals is synchronously triggered by the unified master clock signal of the motor controller. The digital signal after analog-to-digital conversion is synchronously sampled at a preset sampling frequency higher than the fundamental frequency of the motor. The sampled phase current and terminal voltage are transformed into direct-axis current, quadrature-axis current, direct-axis voltage, and quadrature-axis voltage, which together with the rotor speed form a multi-source electrical signal sequence.

10. A two-dimensional dynamic torque distribution prediction system for diagnosing demagnetization faults in permanent magnet motors, characterized in that, The system includes: Signal acquisition module: used to collect multi-source electrical signals from the motor and construct a multi-source time-series dataset; Observer building block: Used to build a dual-channel magnetic flux observer; Model update and output module: used to input real-time direct-axis voltage, quadrature-axis voltage, direct-axis current, quadrature-axis current, rotor speed multi-source electrical signals, dynamically adjust parameters, and output comprehensive flux linkage estimate; Torque prediction module: used to calculate the comprehensive predicted torque by combining the comprehensive flux linkage estimate with the real-time current; Temperature disturbance calculation module: used to calculate the first temperature disturbance flux linkage value and output the second temperature disturbance flux linkage value after verification and fusion; Temperature compensation module: used to perform magnetic flux temperature compensation and calculate the reference compensation torque; Fault determination module: used to generate characteristic deviation sequences and complete demagnetization fault determination through an interlocking judgment mechanism; Each module is configured to perform the method described in any one of claims 1 to 9.

Images