Permanent magnet motor demagnetization fault diagnosis method based on deep learning

Abstract

The invention discloses a permanent magnet motor demagnetization fault diagnosis method based on deep learning, and relates to the technical field of permanent magnet motor fault diagnosis, and the method comprises the steps: collecting the operation state data of a permanent magnet synchronous motor and a fan in real time through a sensor network, and generating a basic variable; multi-modal features in the basic variables are extracted, a health index sequence is constructed, and a demagnetization risk score is generated; setting a self-adaptive diagnosis opportunity control strategy, and determining the opportunity of next diagnosis detection according to the current demagnetization risk score; establishing a demagnetization risk prediction model, predicting a demagnetization risk score at the next diagnosis time, generating a dynamic risk threshold value, and judging whether a demagnetization early warning decision is triggered or not; allocating a maintenance task priority according to the current demagnetization risk score and the historical mode; comprehensively calculating the overall health index of the fan system, and predicting the residual life of the fan permanent magnet motor. According to the method, the problem of resource waste caused by unreasonable diagnosis time in the demagnetization fault diagnosis of the permanent magnet motor is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of permanent magnet motor fault diagnosis, and specifically to a permanent magnet motor demagnetization fault diagnosis method based on deep learning. Background Art

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in electric vehicles, aerospace, and high-end manufacturing due to their high power density, high efficiency, and excellent controllability. However, over long-term operation, PMSMs are susceptible to irreversible permanent magnet demagnetization due to factors such as armature reaction, overload current, high temperatures, manufacturing defects, and material aging, leading to performance degradation or even motor failure.

[0003] Traditional demagnetization fault diagnosis methods, such as those based on mathematical models, signal processing, or high-frequency signal injection, suffer from strong dependence on operating conditions, limited diagnostic accuracy, and poor real-time performance, making them unable to meet the high reliability and intelligent needs of modern industry. Permanent magnets are particularly prone to irreversible demagnetization under conditions such as high temperature and high current. The initial symptoms are not obvious and difficult to detect in time, which may lead to motor performance degradation or even shutdown. With the development of sensor technology and artificial intelligence, demagnetization fault prediction and diagnosis methods based on multimodal data fusion and deep learning have gradually become a research hotspot. However, existing deep learning diagnostic models are mostly trained under single or controlled operating conditions. Their performance degrades significantly when faced with changes in speed, load, and ambient temperature, making it difficult to meet the online diagnosis needs of the changing operating conditions in industrial sites. Typical fault monitoring uses fixed-cycle sampling and analysis, evenly distributing computing resources regardless of the equipment status. This leads to excessive monitoring waste during normal operation and insufficient response to sudden anomalies.

[0004] To this end, the present invention provides a permanent magnet motor demagnetization fault diagnosis method based on deep learning. Summary of the Invention

[0005] The purpose of the present invention is to provide a permanent magnet motor demagnetization fault diagnosis method based on deep learning to solve the existing problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing demagnetization faults of permanent magnet motors based on deep learning, comprising the following steps: S1. Real-time data on the operation status of the permanent magnet synchronous motor and fan is collected through the sensor network to generate basic variables. S2. extracting multimodal features from the basic variables, constructing a health index sequence, and generating a demagnetization risk score; S3. Setting an adaptive diagnostic timing control strategy to determine the timing of the next diagnostic test based on the current demagnetization risk score; S4. Establish a demagnetization risk prediction model to predict the demagnetization risk score at the next diagnostic opportunity and generate a dynamic risk threshold to determine whether a demagnetization warning decision is triggered. If so, execute S5; if not, return to S2. S5. Assign maintenance task priorities based on current demagnetization risk scores and historical patterns; S6. Calculate the overall health index of the fan system by combining the next cycle demagnetization risk score and historical data to predict the remaining life of the fan permanent magnet motor.

[0007] The present invention is further improved in that the sensor network includes a three-phase current sensor, a motor terminal voltage sensor, a speed sensor, a temperature sensor and a vibration sensor; the multimodal characteristics are obtained through the sensor network, including filtering the collected current and voltage signals in the controller, performing Park transformation on the three-phase current, and obtaining the dq axis current value 、 , perform FFT calculation on voltage and current to calculate harmonic components and obtain current harmonic characteristics and voltage harmonic characteristics ; Calculate the root mean square error of the vibration signal to obtain the vibration characteristics , the temperature rise characteristics are obtained by calculating the temperature rise rate , and extract the speed fluctuation rate to obtain the speed fluctuation characteristics ; The above features are normalized and weighted averaged to calculate the demagnetization risk score.

[0008] The present invention is further improved in that the adaptive diagnosis timing control strategy includes the system entering the normal monitoring mode by default after power-on or reset. At this time, the control strategy puts the internal state machine into the normal state and sets the normal monitoring interval. The system synchronously acquires the stator dq axis current, voltage, electromechanical speed and temperature signals at the first sampling frequency. At this time, the monitoring window for calculating the speed fluctuation rate and temperature rise rate is the regular monitoring interval; the safety thresholds corresponding to all indexes in the health index sequence are set. When any index in the health index sequence is greater than or equal to its corresponding safety threshold, it switches to the high-risk monitoring mode, otherwise the next regular monitoring interval is calculated. , where IHS stands for demagnetization risk score, Represents the sensitivity coefficient; at this time, the sampling frequency is increased to the second sampling frequency.

[0009] A further improvement of the present invention is that the adaptive diagnostic timing control strategy also includes synchronously sampling all indices in the health index sequence when the sampling frequency is increased to the second sampling frequency; calling a pre-trained multimodal deep network to perform fine-grained analysis of all indices, performing deep feature extraction, and recalculating the demagnetization risk score; if in the high-risk monitoring mode, any index in the health index sequence is less than its corresponding safety threshold for K consecutive times, the state machine is switched back to the normal monitoring mode, and the normal monitoring interval is reset.

[0010] A further improvement of the present invention is that the process of establishing the demagnetization risk prediction model includes: based on the U-NET neural network model, historical multimodal features are spliced ​​in chronological order and input into the network, and the input data includes the demagnetization health index sequence of the historical nth diagnostic time, the three-phase stator current amplitude spectrum vector collected at the same time, the electromechanical speed fluctuation rate signal and its spectral characteristics, the vibration root mean square signal, and the temperature rise curve at the same time; after normalizing the above-mentioned modal signals, cross-layer jump connections are combined in the U-NET encoder stage; the demagnetization fault prediction value output by the last layer of the network is used as the demagnetization risk score of the next cycle, and the training goal is to minimize the sum of the prediction accuracies of all predicted demagnetization risk scores, and stop training until the sum of the prediction accuracies reaches convergence. The prediction accuracy is obtained by the square of the difference between the predicted value and the actual value of the demagnetization risk prediction model.

[0011] A further improvement of the present invention is that the demagnetization risk prediction model further includes defining the average of the demagnetization risk scores within m historical diagnostic cycles as a dynamic risk threshold; comparing the demagnetization risk score predicted by the cycle of the current diagnostic opportunity with the dynamic risk threshold; if the demagnetization risk score predicted by the cycle of the current diagnostic opportunity is greater than the dynamic risk threshold, the system issues a demagnetization warning safety signal and returns to S2; if the demagnetization risk score predicted by the cycle of the current diagnostic opportunity is less than or equal to the dynamic risk threshold, a demagnetization warning danger signal is issued, triggering a demagnetization warning decision, at which time a demagnetization recovery expected value is calculated, and the demagnetization recovery expected value is returned to S3 before executing S5. In the high-risk monitoring mode of the adaptive diagnostic opportunity control strategy in S3, if any index in the K consecutive health index sequences is less than its corresponding safety threshold and the demagnetization recovery expected value is less than the set recovery expected threshold, the state machine is switched back to the normal monitoring mode and the normal monitoring interval is reset; the demagnetization recovery expected value is obtained by calculating the number of times in the historical diagnostic cycles that the demagnetization risk score predicted by the cycle of the current diagnostic opportunity is greater than the dynamic risk threshold and the ratio of the number of historical diagnostic times to all historical diagnostic times.

[0012] The present invention is further improved in that step S5 specifically includes the following steps: listing the demagnetization risk scores obtained at m historical diagnosis times into the historical risk score sequence , as the target task; the demagnetization risk score prediction value obtained at the current diagnosis time As a bidder, in the rth round of bidding, the bidder has Make a quote: ; Among them, U represents the profit magnification coefficient, extract the highest bid in this round Corresponding target , and its execution right is awarded to the current diagnosis cycle risk assessment task; if there are still historical target tasks that have not won the bid, the remaining targets will enter the next round of bidding until all historical target tasks have obtained the execution right; finally, the demagnetization risk of each historical cycle is given priority according to the bidding rounds and bid levels of each target, and the demagnetization risk score prediction value obtained at the current diagnosis time is output and maintain sorted results.

[0013] The present invention is further improved in that step S6 updates the overall health index by introducing an exponentially weighted moving average: ; in, represents the smoothing coefficient, initial health index ;Will Append to the overall health index sequence; perform least squares linear fitting on the overall health index sequence to obtain the slope a, and set the overall health critical value The remaining life of the fan permanent magnet motor is .

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention first introduces a state machine driven by demagnetization risk scoring to dynamically switch the sampling frequency between normal and high-risk modes, ensuring high-resolution monitoring at critical moments while reducing the sampling load during stable periods, thus achieving resource optimization from "less sampling → normal sampling → more sampling"; 2. Dynamic thresholds are adaptively generated based on the average risk scores of multiple historical diagnostic cycles to avoid false positives or missed negatives caused by static thresholds. Only persistently high risks that deviate from normal levels trigger alerts, thus balancing the contradiction between "timely detection" and "avoiding excessive alerts." 3. Introduce recovery expectation value for secondary confirmation. Switch back to regular monitoring only when the health index continues to exceed expectations, avoiding frequent mode switching and overdiagnosis caused by short-term fluctuations. 4. Through the task bidding and auction mechanism, current risks are matched with historical patterns, and execution priorities are assigned to various maintenance tasks, so that maintenance resources can be focused on the most urgent risk points and maintenance efficiency can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to the present invention; Figure 2 This is a framework diagram of a demagnetization risk prediction model for a permanent magnet motor demagnetization fault diagnosis method based on deep learning in the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0017] The term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " generally indicates an "or" relationship between the related objects.

[0018] Example 1 Figure 1 A flowchart of a method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning disclosed in this embodiment is shown. The steps are as follows: S1. Real-time data collection of permanent magnet synchronous motor and fan operating status is performed through a sensor network to generate basic variables, while also collecting external operating condition data. The sensor network includes a three-phase current sensor, a motor terminal voltage sensor, a speed sensor, a temperature sensor, and a vibration sensor. The external operating condition data includes fan load. Multimodal features are obtained through the sensor network, including filtering the collected current and voltage signals in the controller, performing Park transformation on the three-phase current, and obtaining the dq axis current values. 、 , perform FFT calculation on voltage and current to calculate harmonic components and obtain current harmonic characteristics and voltage harmonic characteristics ; Calculate the root mean square error of the vibration signal to obtain the vibration characteristics , the temperature rise characteristics are obtained by calculating the temperature rise rate , and extract the speed fluctuation rate to obtain the speed fluctuation characteristics ; The above features are normalized and weighted averaged to calculate the demagnetization risk score.

[0019] In a permanent magnet synchronous motor, the magnetic field generated by the three-phase stator current interacts with the rotor magnetic field to produce torque. However, the mathematical model of the motor in the three-phase stationary coordinate system (abc axes) is complex and includes time-varying parameters (such as inductance that varies with rotor position), making it difficult to directly use for high-performance control. Using a coordinate transformation (Park transform), the three-phase stationary coordinate system can be converted into a two-phase rotating coordinate system (dq axes), simplifying the mathematical model and achieving decoupled control similar to that of a DC motor.

[0020] S2. Extract the multimodal features in the basic variables, construct a health index sequence, and generate a demagnetization risk score. In this step, the diagnostic path selection based on working condition perception can be carried out. First, a preliminary fault category judgment is made based on the characteristic indicators in the preprocessed signal. If the three-phase current of the motor is found to be unbalanced or the harmonic components are significantly abnormal, it is marked that there may be an electrical fault (such as inter-turn short circuit, inter-turn open circuit or demagnetization). If the vibration spectrum shows bearing frequency or imbalance frequency characteristics, it points to a mechanical fault (such as bearing damage or rotor imbalance). If the speed does not match the wind pressure, it may be a magnetic circuit demagnetization or a pneumatic abnormality fault.

[0021] For different fault types, the system activates the corresponding diagnostic sub-processes in sequence; when the fault is a demagnetization fault, it immediately executes permanent magnet flux estimation and demagnetization index update, and inputs the results into the demagnetization trend prediction module; when the fault is a bearing fault, it switches to the vibration analysis branch to refine the high-frequency vibration feature extraction and spectrum comparison.

[0022] In each sub-process, a preset fuzzy rule base or Bayesian decision matrix is ​​used for judgment. This rule base integrates the fan operating conditions (speed, load) and detection results (current harmonic rate, vibration amplitude, temperature change rate) to generate a confidence score. Specifically, the confidence of different fault hypotheses is calculated through Dempster-Shafer evidence fusion, and subsequent actions are triggered based on the confidence level. The fault type with the highest confidence is identified as the diagnostic target, and the corresponding fault confirmation test (such as AC excitation test and single-phase failure simulation test) is initiated.

[0023] This multi-branch decision-making mechanism tightly couples the diagnostic process with the current operating conditions, activating only necessary detection steps and avoiding blind full-scale diagnosis, thereby shortening diagnostic time and reducing computational burden; each sub-process can be executed in parallel or sequentially, with resources and priorities coordinated by the diagnostic control logic.

[0024] S3. Setting an adaptive diagnostic timing control strategy to determine the timing of the next diagnostic test based on the current demagnetization risk score; The adaptive diagnostic timing control strategy includes the system entering the normal monitoring mode by default after power-on or reset. At this time, the control strategy puts the internal state machine into the normal state and sets the normal monitoring interval. , the system synchronously acquires the stator dq axis current, voltage, electromechanical speed and temperature signals at the first sampling frequency (low sampling frequency), such as 1kHz. At this time, the monitoring window for calculating the speed fluctuation rate and temperature rise rate is the regular monitoring interval; set the safety thresholds corresponding to all indexes in the health index sequence. When any index in the health index sequence is greater than or equal to its corresponding safety threshold, switch to the high-risk monitoring mode, otherwise calculate the next regular monitoring interval , where IHS stands for demagnetization risk score, Indicates the sensitivity coefficient; at this time, the sampling frequency is increased to the second sampling frequency (high sampling frequency), such as 10kHz.

[0025] In either mode, the system continuously maintains a real-time overall health index. When the multimodal characteristics decrease significantly, the regular time interval will be shortened accordingly to increase the frequency of monitoring and diagnosis.

[0026] The adaptive diagnosis timing control strategy also includes synchronously sampling all indices in the health index sequence when the sampling frequency is increased to the second sampling frequency; calling a pre-trained multimodal deep network to perform fine-grained analysis on all indices, performing deep feature extraction, and recalculating the demagnetization risk score; specifically including: Stratified random sampling is performed according to the working conditions to ensure that the training and validation sets contain multiple working conditions; each signal channel is independently normalized to zero mean and unit variance; the normalized multi-channel signal is divided into overlapping windows of length L (e.g., 256 ms) with a window sliding step of S (e.g., 64 ms), and each sample is a tensor , where C represents the number of channels; load the trained network weights into the inference engine of the high-risk diagnosis module; acquire all channel signals in real time at a high sampling rate (e.g., 10 kHz); normalize the signals using the same normalization parameters used during training; divide the signals into windows of length L and input them into the model for forward inference. The inferred data is used as the demagnetization risk score for this window, and smoothed using a sliding window or exponentially weighted average to obtain a recalculated demagnetization risk score; If any index in the K consecutive health index sequences is less than its corresponding safety threshold in the high-risk monitoring mode, the state machine will be switched back to the normal monitoring mode and the normal monitoring interval will be reset.

[0027] S4. Establish a demagnetization risk prediction model to predict the demagnetization risk score at the next diagnostic opportunity and generate a dynamic risk threshold to determine whether a demagnetization warning decision is triggered. If so, execute S5; if not, return to S2. S5. Assign maintenance task priorities based on current demagnetization risk scores and historical patterns; S6. Calculate the overall health index of the fan system by combining the next cycle demagnetization risk score and historical data to predict the remaining life of the fan permanent magnet motor.

[0028] Example 2 Based on the inventive concept of Example 1, this embodiment proposes a specific construction process of the demagnetization risk prediction model in step S4. Figure 2 The present invention shows a framework diagram of a demagnetization risk prediction model for a permanent magnet motor demagnetization fault diagnosis method based on deep learning, which specifically includes: Based on the U-NET neural network model, historical multimodal features are spliced ​​in chronological order and input into the network. The input data includes the demagnetization health index sequence of the nth historical diagnostic time, the three-phase stator current amplitude spectrum vector collected at the same time, the electromechanical speed fluctuation rate signal and its spectral characteristics, the vibration root mean square signal, and the temperature rise curve. After normalizing the above modal signals, cross-layer jump connections are combined in the U-NET encoder stage to achieve high-resolution demagnetization fault feature mapping. The demagnetization fault prediction value output by the last layer of the network is used as the demagnetization risk score of the next cycle. The training goal is to minimize the sum of the prediction accuracies of all predicted demagnetization risk scores. Training is stopped until the sum of the prediction accuracies reaches convergence. The prediction accuracy is obtained by the square of the difference between the predicted value and the actual value of the demagnetization risk prediction model.

[0029] The demagnetization risk prediction model further includes defining the average demagnetization risk score within the historical m diagnostic cycles as a dynamic risk threshold; comparing the demagnetization risk score predicted by the cycle at the current diagnostic opportunity with the dynamic risk threshold; if the demagnetization risk score predicted by the cycle at the current diagnostic opportunity is greater than the dynamic risk threshold, the system issues a demagnetization warning safety signal and returns to S2; if the demagnetization risk score predicted by the cycle at the current diagnostic opportunity is less than or equal to the dynamic risk threshold, a demagnetization warning danger signal is issued to trigger a demagnetization warning decision, at which time the demagnetization recovery expected value is calculated, and the demagnetization recovery expected value is returned to S3 and then S5 is executed, and then the adaptive demagnetization recovery decision in S3 is executed. In the high-risk monitoring mode of the diagnostic timing control strategy, when any index in the health index sequence is less than its corresponding safety threshold for K consecutive times and the demagnetization recovery expected value is less than the set recovery expected threshold, the state machine will be switched back to the normal monitoring mode and the normal monitoring interval will be reset; the demagnetization recovery expected value is obtained by calculating the number of times the demagnetization risk score of the current diagnostic timing cycle is greater than the dynamic risk threshold and the ratio of the number of demagnetization risks to all historical diagnostic times in the historical diagnostic cycle. It is used to control the diagnostic frequency of the system to return to normal, avoiding the problem of the system maintaining high-frequency diagnosis and excessive resource consumption due to short-term fluctuations or occasional minor anomalies.

[0030] This embodiment also provides a specific method for allocating maintenance task priorities in step S5 of embodiment 1, which specifically includes the following steps: The demagnetization risk scores obtained at the m historical diagnosis times are included in the historical risk score sequence , as the target task; the demagnetization risk score prediction value obtained at the current diagnosis time As a bidder, in the rth round of bidding, the bidder has Make a quote: ; Among them, U represents the profit magnification coefficient. The bids of the winning rounds are calculated by cumulative sum, and the highest bid of this round is extracted. Corresponding target , and its execution right is awarded to the current diagnosis cycle risk assessment task; if there are still historical target tasks that have not won the bid, the remaining targets will enter the next round of bidding until all historical target tasks have obtained the execution right; finally, the demagnetization risk of each historical cycle is given priority according to the bidding rounds and bid levels of each target, and the demagnetization risk score prediction value obtained at the current diagnosis time is output and maintain sorted results.

[0031] Based on the inventive concept in Example 1, this embodiment further provides a specific implementation of step S6, including: updating the overall health index by introducing an exponentially weighted moving average: ; in, Indicates the smoothing coefficient, the larger the value, the more emphasis is placed on the latest state; initial health index ;Will Append to the overall health index series for subsequent analysis and model retraining; perform least squares linear fitting on the overall health index series to obtain the slope a, and set the overall health critical value The remaining life of the fan permanent magnet motor (number of cycles) ; Regularly add new multimodal features, prediction results, and overall health index to the training set, fine-tune the U-NET network, and ensure consistency between the model and health index calculation; This method combines real-time features, prediction results, and historical data to calculate the overall health index of the fan system. It then updates the health index sequence, providing a reference for U-Net model training and threshold adjustment. By continuously monitoring the health index's changing trends, it can help determine the remaining lifespan of the system or when maintenance is needed.

[0032] The thresholds, weights and other setting values ​​may be set by default according to the present invention, or may be set by those skilled in the art.

[0033] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0034] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0035] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0036] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0037] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

Claims

1. A method for diagnosing demagnetization faults of permanent magnet motors based on deep learning, characterized by: The following steps are involved: S1. Real-time data on the operation status of the permanent magnet synchronous motor and fan is collected through the sensor network to generate basic variables. S2. extracting multimodal features from the basic variables, constructing a health index sequence, and generating a demagnetization risk score; S3. Setting an adaptive diagnostic timing control strategy to determine the timing of the next diagnostic test based on the current demagnetization risk score; S4. Establish a demagnetization risk prediction model to predict the demagnetization risk score at the next diagnostic opportunity and generate a dynamic risk threshold to determine whether a demagnetization warning decision is triggered. If so, execute S5; if not, return to S2. S5. Assign maintenance task priorities based on current demagnetization risk scores and historical patterns; S6. Calculate the overall health index of the fan system by combining the next cycle demagnetization risk score and historical data to predict the remaining life of the fan permanent magnet motor.

2. A method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 1, characterized in that: The sensor network includes a three-phase current sensor, a motor terminal voltage sensor, a speed sensor, a temperature sensor, and a vibration sensor; the multimodal features are obtained through the sensor network, including filtering the collected current and voltage signals in the controller, performing Park transformation on the three-phase current, and obtaining the dq axis current value. 、 , perform FFT calculation on voltage and current to calculate harmonic components and obtain current harmonic characteristics and voltage harmonic characteristics ; Calculate the root mean square error of the vibration signal to obtain the vibration characteristics , the temperature rise characteristics are obtained by calculating the temperature rise rate , and extract the speed fluctuation rate to obtain the speed fluctuation characteristics ; The above features are normalized and weighted averaged to calculate the demagnetization risk score.

3. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 1, characterized in that: The adaptive diagnostic timing control strategy includes the system entering the normal monitoring mode by default after power-on or reset. At this time, the control strategy puts the internal state machine into the normal state and sets the normal monitoring interval. The system synchronously acquires the stator dq axis current, voltage, electromechanical speed and temperature signals at the first sampling frequency. At this time, the monitoring window for calculating the speed fluctuation rate and temperature rise rate is the regular monitoring interval; the safety thresholds corresponding to all indexes in the health index sequence are set. When any index in the health index sequence is greater than or equal to its corresponding safety threshold, it switches to the high-risk monitoring mode, otherwise the next regular monitoring interval is calculated. , where IHS represents the demagnetization risk score, Represents the sensitivity coefficient; at this time, the sampling frequency is increased to the second sampling frequency.

4. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 3, characterized in that: The adaptive diagnosis timing control strategy further includes synchronously sampling all indices in the health index sequence when the sampling frequency is increased to the second sampling frequency; calling a pre-trained multimodal deep network to perform fine-grained analysis on all indices, performing deep feature extraction, and recalculating the demagnetization risk score; If any index in the K consecutive health index sequences is less than its corresponding safety threshold in the high-risk monitoring mode, the state machine will be switched back to the normal monitoring mode and the normal monitoring interval will be reset.

5. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 1, characterized in that: The establishment process of the demagnetization risk prediction model includes: based on the U-NET neural network model, historical multimodal features are spliced ​​in chronological order and input into the network, and the input data includes the demagnetization health index sequence of the historical nth diagnostic time, the three-phase stator current amplitude spectrum vector collected at the same time, the electromechanical speed fluctuation rate signal and its spectral characteristics, the vibration root mean square signal, and the temperature rise curve at the same time; after normalizing the above-mentioned modal signals, cross-layer jump connections are combined in the U-NET encoder stage; the demagnetization fault prediction value output by the last layer of the network is used as the demagnetization risk score of the next cycle, and minimizing the sum of the prediction accuracy of all predicted demagnetization risk scores is used as the training goal. Training is stopped until the sum of the prediction accuracy reaches convergence. The prediction accuracy is obtained by the square of the difference between the predicted value and the actual value of the demagnetization risk prediction model.

6. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 5, characterized in that: The demagnetization risk prediction model further includes defining an average of the demagnetization risk scores in m historical diagnostic cycles as a dynamic risk threshold; comparing the demagnetization risk score predicted for the current diagnostic cycle with the dynamic risk threshold; If the demagnetization risk score predicted by the cycle of the current diagnostic opportunity is greater than the dynamic risk threshold, the system issues a demagnetization warning safety signal and returns to S2; if the demagnetization risk score predicted by the cycle of the current diagnostic opportunity is less than or equal to the dynamic risk threshold, a demagnetization warning danger signal is issued to trigger a demagnetization warning decision. At this time, the demagnetization recovery expected value is calculated, and the demagnetization recovery expected value is returned to S3 and then S5 is run. Then, in the high-risk monitoring mode of the adaptive diagnostic opportunity control strategy of S3, if any index in the K consecutive health index sequences is less than its corresponding safety threshold and the demagnetization recovery expected value is less than the set recovery expected threshold, the state machine is switched back to the normal monitoring mode and the normal monitoring interval is reset; the demagnetization recovery expected value is obtained by calculating the number of times the demagnetization risk score predicted by the cycle of the current diagnostic opportunity in the historical diagnostic cycle is greater than the dynamic risk threshold and the ratio to all historical diagnostic times.

7. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 1, characterized in that: Step S5 specifically includes the following steps: listing the demagnetization risk scores obtained at m historical diagnosis times into the historical risk score sequence , as the target task; the demagnetization risk score prediction value obtained at the current diagnosis time As a bidder, in the rth round of bidding, the bidder has Make a quote: ； Among them, U represents the profit magnification coefficient, extract the highest bid in this round Corresponding target , and its execution right is awarded to the current diagnosis cycle risk assessment task; if there are still historical target tasks that have not won the bid, the remaining targets will enter the next round of bidding until all historical target tasks have obtained the execution right; finally, the demagnetization risk of each historical cycle is given priority according to the bidding rounds and bid levels of each target, and the demagnetization risk score prediction value obtained at the current diagnosis time is output and maintain sorted results.

8. The method for diagnosing demagnetization faults of a permanent magnet motor based on deep learning according to claim 1, characterized in that: Step S6 updates the overall health index by introducing an exponentially weighted moving average: ； in, represents the smoothing coefficient, initial health index ;Will Append to the overall health index sequence; perform least squares linear fitting on the overall health index sequence to obtain the slope a, and set the overall health critical value The remaining life of the fan permanent magnet motor is .

Images