A servo permanent magnet synchronous motor fault detection method based on artificial intelligence

Abstract

The application discloses a kind of servo permanent magnet synchronous motor fault detection methods based on artificial intelligence, belong to servo motor detection technical field, including data acquisition and windowing, feature extraction and abnormality degree calculation, static threshold determination, adaptive window and probability determination and continuous trigger strategy and alarm output, solve servo permanent magnet synchronous motor fault detection, realize the technical problem of strong detection real-time, predictable processing process and low false alarm rate, the sampling data processing mode of windowing is used in the application, continuous data is divided into fixed length window, feature is extracted in each window and abnormality degree is determined, both ensure the determinacy of processing time, and the sensitivity to abnormal fluctuation can be considered through adaptive window mechanism.Combined with quantile threshold determination and continuous trigger strategy, false alarm rate is significantly reduced, and only abnormal data is output, which is convenient for subsequent expansion to specific fault type diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of servo motor detection, and particularly relates to a servo permanent magnet synchronous motor fault detection method based on artificial intelligence. BACKGROUND

[0002] Servo permanent magnet synchronous motors are widely used in intelligent manufacturing and motion control fields, but problems such as current distortion, voltage fluctuation and harmonic distortion often occur in the running process due to power grid fluctuations, driver faults or winding abnormalities. In order to avoid the expansion of faults, the running state of the motor needs to be detected and alarmed in time.

[0003] Existing fault detection technologies mainly include two categories:

[0004] One category is a threshold-based monitoring method, which usually directly sets upper and lower limits for current, voltage and other measured data for judgment. This method is simple to implement, but is easily affected by transient disturbances, leading to misjudgment.

[0005] The other category is a traditional machine learning-based classification method, which usually needs to extract features from full-quantity raw data and train models. The data processing process lacks a unified time division mechanism, resulting in high computational complexity and unpredictable processing time, making it difficult to meet the real-time detection needs of the motor.

[0006] Therefore, the existing methods still have deficiencies in detection real-time, processing efficiency and result predictability. SUMMARY

[0007] The purpose of the present application is to provide a servo permanent magnet synchronous motor fault detection method based on artificial intelligence, which solves the technical problems of strong detection real-time, predictable processing process and low false alarm rate in servo permanent magnet synchronous motor fault detection.

[0008] To achieve the above purpose, the present application adopts the following technical solutions:

[0009] A servo permanent magnet synchronous motor fault detection method based on artificial intelligence includes the following steps:

[0010] Step 1: The data acquisition module synchronously acquires the raw data generated in the servo motor drive system, including three-phase current and DC bus voltage , cuts the continuous sampling frames of the raw data into several windows for analyzing data;

[0011] Step 2: The feature extraction module calculates the feature vector of each window, including time domain and sequence components, inputs the feature vector into a one-class SVM model to obtain the original decision output, and maps the original decision output to the normalized window abnormality score .

[0012] Step 3: The window analysis module sets static quantile threshold intervals and scores the anomaly level. Make a judgment: if the anomaly score is... If the score is less than the lower limit of the quantile threshold interval, it is considered normal; if the upper limit of the quantile threshold interval is less than the abnormality score... If this occurs, it is considered an anomaly, and the window's start time is taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. Combine the data in the window and end; if the anomaly score is... If the value falls within the quantile threshold range, then proceed to step 4 for probabilistic adaptive determination.

[0013] Step 4: Window Analysis Module Characteristic Volatility ,according to The length of the window is automatically selected to obtain an adaptive window. The feature vector is then extracted and the anomaly score is recalculated on the adaptive window using the method in step 2, resulting in a feature vector based on the adaptive window. and anomaly score Anomaly score for adaptive window Anomaly score compared to the corresponding original window Perform a weighted summation to obtain the final anomaly score. Regarding the final anomaly degree Perform probabilistic judgment to obtain the probability of failure of the adaptive window. ;

[0014] Step 5: The fault diagnosis module establishes a continuous triggering strategy based on the K / M principle, and assigns the probability of fault for each adaptive window. The data is summarized, and the faults are checked against the continuous triggering strategy. If the triggering conditions in the continuous triggering strategy are met, an alarm message is generated and output; the alarm message includes the fault occurrence time. The data in the adaptive window is now complete.

[0015] Preferably, step 1 includes the following steps:

[0016] Step 1-1: The data acquisition module reads the synchronously acquired three-phase current from the servo motor drive system. and DC bus voltage Generate a continuous sampling sequence :

[0017] ;

[0018] in, These are the phase a current value, phase b current value, and phase c current value at the nth sampling point, respectively. This represents the total number of sampling points; The DC bus voltage at the nth sampling point;

[0019] Step 1-2: Preset a fixed number of periods, and sample the sequence. The data is divided into several windows for analysis based on a fixed number of periods, and a dataset is constructed for each window. Each window contains One sampling point:

[0020] ;

[0021] ;

[0022] in, This represents the data set of the t-th window; and These are the indices of the starting and ending sampling points of the t-th window, respectively. These represent the phase a current values ​​at the nth sampling point in the t-th window, respectively. b-phase current value c-phase current value DC bus voltage value .

[0023] Preferably, step 2 includes the following steps:

[0024] Step 2-1: Calculate the feature vector for each window, specifically including temporal feature calculation and order component feature calculation:

[0025] The time-domain feature calculation is as follows:

[0026] Calculate the window value using the following formulas. In root mean square value :

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] in, , , and These represent the current in the t-th window, respectively. Root mean square value, current Root mean square value, current Root mean square value and DC bus voltage Root mean square value;

[0032] The calculation of the ordinal component features is as follows:

[0033] Calculate the magnitude of the negative-order component :

[0034] , ;

[0035] Calculate the zero-sequence current amplitude The specific formula is as follows:

[0036] ;

[0037] Step 2-2: Constructing Feature Vectors The details are as follows:

[0038] ;

[0039] in, This represents the feature vector of the t-th window;

[0040] Steps 2-3: ... The input is fed into a class of SVM models, and the output is anomaly score. :

[0041] ;

[0042] in, Represents a class of SVM model mapping functions. This represents the original anomaly score for the t-th window. The value of is normalized to the interval [0,1].

[0043] Preferably, step 3 includes the following steps:

[0044] Step 3-1: Select a set of normal data from the original data as training data. Based on the training data and the methods in Steps 1 to 2, obtain a set of anomaly scores based on the training data. Take set average As the baseline score for normal data, a quantile threshold interval is calculated based on the baseline score. The specific calculation formula is as follows:

[0045] ;

[0046] in, Indicates will The The percentile value is used as the static lower limit threshold. Indicates will The Percentile value, used as a static upper limit threshold; This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; The value is 95%. The value is 99%;

[0047] Step 3-2: In actual operation, for the new t-th window obtained through the methods in Step 1 and Step 2, score its anomaly. Perform static determination:

[0048] ;

[0049] in, This represents the judgment result for the t-th window;

[0050] Steps 3-5: If If it is not critical, then output. The result, if If the error is abnormal, the start time of the t-th window will be taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. End with the data in the window;

[0051] like The result is critical, so proceed to step 4.

[0052] Preferably, step 4 includes the following steps:

[0053] Step 4-1: Obtain For the t-th window that is at the critical point, obtain the feature vector of that window. Simultaneously, retrieve the feature vector of the (t-1)th window. Calculate the characteristic volatility of these two adjacent windows. :

[0054] ;

[0055] in, The second norm of a vector;

[0056] Step 4-2: Preset the upper limit of the rate of change and lower limit of rate of change Based on characteristic volatility Adaptive selection of window length :

[0057] like Then use a fixed number of points Shorten the window;

[0058] like Then use a fixed number of points Lengthen the window;

[0059] Step 4-3: For the adjusted adaptive window, use the methods from Steps 1 to 3 to re-extract the feature vector. And calculate the anomaly score of the adaptive window. ;

[0060] Step 4-4: Scoring the anomaly of the adaptive window Anomaly score compared to the corresponding original window The anomaly is then fused to obtain the final anomaly score. :

[0061] ;

[0062] in, These are adaptive weights used based on characteristic volatility. Adjust the weight of the original window of fixed length. ;

[0063] Steps 4-5: Determine the final anomaly level Perform probabilistic judgment to obtain the probability of failure of the adaptive window. :

[0064] , ;

[0065] in, This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; This indicates the probability that the adaptive window is faulty.

[0066] Preferably, step 5 includes the following steps:

[0067] Step 5-1: Retrieve the fault probability of any t0th adaptive window ;

[0068] Step 5-2: Establish a continuous triggering strategy Specifically, this refers to the K / M principle:

[0069] ;

[0070] in, To adapt the failure probability of the window, Here, K is the probability threshold, with a default value of 0.5. M is the total number of consecutive adaptive windows, and K is the threshold for the number of adaptive windows judged as abnormal. This indicates an indicator function that, when the condition is met... If the condition is met, the indicator function outputs 1; otherwise, it outputs 0. Indicates the starting index of the contiguous window, that is, the first... The first adaptive window is used as the starting window of the continuous adaptive window, and t is the index of the continuous adaptive window;

[0071] like If the window is not found to be faulty, then the adaptive window is confirmed to be faulty, i.e., the window is abnormal; otherwise, the next adaptive window is checked sequentially downwards.

[0072] Step 5-3: Place the first The start time of each adaptive window is used as the time of failure. Generate alarm information and output the time of fault occurrence. Generate alarm information ALM2 and output the fault occurrence time. And the data in the adaptive window.

[0073] This invention discloses an artificial intelligence-based fault detection method for servo permanent magnet synchronous motors, solving the technical problems of achieving high real-time detection, predictable processing, and low false alarm rate in servo permanent magnet synchronous motor fault detection. This invention employs a windowed sampling data processing method, dividing continuous data into fixed-length windows. Features are extracted and anomaly determination is performed within each window, ensuring deterministic processing time while maintaining sensitivity to abnormal fluctuations through an adaptive window mechanism. Combined with quantile threshold determination and a continuous triggering strategy, the false alarm rate is significantly reduced, and only abnormal data is output, facilitating subsequent expansion to diagnose specific fault types. Attached Figure Description

[0074] Figure 1 This is an overall flowchart of the present invention;

[0075] Figure 2 This is a flowchart of step 1 of the present invention;

[0076] Figure 3 This is a flowchart of step 2 of the present invention;

[0077] Figure 4 This is a flowchart of step 3 of the present invention;

[0078] Figure 5 This is a flowchart of step 4 of the present invention;

[0079] Figure 6This is a flowchart of step 5 of the present invention. Detailed Implementation

[0080] Depend on Figures 1-6 The method for fault detection of a servo permanent magnet synchronous motor based on artificial intelligence, as shown, includes the following steps:

[0081] Step 1: The data acquisition module synchronously acquires the raw data generated by the servo motor drive system, including the three-phase current. and DC bus voltage The continuous sampling frames of the original data are divided into several windows for data analysis;

[0082] Step 1 includes the following steps:

[0083] Step 1-1: The data acquisition module reads the synchronously acquired three-phase current from the servo motor drive system. and DC bus voltage Generate a continuous sampling sequence :

[0084] ;

[0085] in, These are the phase a current value, phase b current value, and phase c current value at the nth sampling point, respectively. This represents the total number of sampling points. The value is calculated based on the sampling time and sampling frequency; The DC bus voltage at the nth sampling point;

[0086] In this embodiment, within one sampling period, the data acquisition module reads a sequence of data that is continuous in time from the servo motor drive system. These data sequences generally include:

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] During this process, the three-phase current and the bus DC voltage are all acquired by the same AD chip in the servo motor drive system, and the time is based on the same timestamp of the servo motor drive system.

[0092] Within one period, n data points are collected, and the total number of data points collected is... The value is obtained by dividing the sampling time and sampling frequency by the predetermined sampling time of the data acquisition module. The sampling time and sampling frequency can be set according to different working conditions.

[0093] Step 1-2: Preset a fixed number of periods, and sample the sequence. The data is divided into several windows for analysis based on a fixed number of periods, and a dataset is constructed for each window. Each window contains One sampling point:

[0094] ;

[0095] ;

[0096] in, This represents the data set of the t-th window; and These are the indices of the starting and ending sampling points of the t-th window, respectively.

[0097] In this embodiment, to facilitate fixing the window length, the fixed number of cycles and The size of the window affects the speed of window processing. Generally, the sampling frequency Fs = 10kHz and the default window ≈ 10 electrical cycles.

[0098] These represent the phase a current values ​​at the nth sampling point in the t-th window, respectively. b-phase current value c-phase current value and DC bus voltage value . It also indicates the length of the window; it usually corresponds to a fixed number of electrical cycles × number of sampling points / cycle;

[0099] In this embodiment, all sampling points use a unified system timestamp to ensure that the three-phase current and the bus voltage correspond on the same time axis.

[0100] Step 2: The feature extraction module calculates the feature vector for each window. The feature vector includes temporal and order components. The feature vector is input into a class of SVM models to obtain the original decision output. The original decision output is then mapped to a normalized anomaly score for the window. ;

[0101] One type of SVM model is existing technology, so it will not be described in detail.

[0102] Step 2 includes the following steps:

[0103] Step 2-1: Calculate the feature vector for each window, specifically including temporal feature calculation and order component feature calculation:

[0104] The time-domain feature calculation is as follows:

[0105] Calculate the window value using the following formulas. In root mean square value :

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] in, , , and These represent the current in the t-th window, respectively. Root mean square value, current Root mean square value, current Root mean square value and DC bus voltage Root mean square value;

[0111] The calculation of the ordinal component features is as follows:

[0112] Calculate the magnitude of the negative-order component :

[0113] , ; It is the imaginary unit.

[0114] Calculate the zero-sequence current amplitude The specific formula is as follows:

[0115] ;

[0116] Step 2-2: Constructing Feature Vectors The details are as follows:

[0117] ;

[0118] in, This represents the feature vector of the t-th window;

[0119] In this embodiment, it is not limited to extracting only the root mean square value. Sum of components characteristics and Users can change the feature extraction model in the feature extraction module according to the working conditions to achieve the effect of adapting to various working conditions.

[0120] This embodiment only applies to... Several typical features were extracted.

[0121] Steps 2-3: ... The input is fed into a class of SVM models, and the output is anomaly score. :

[0122] ;

[0123] in, Represents a class of SVM model mapping functions. This represents the original anomaly score for the t-th window. The value of is normalized to the interval [0,1].

[0124] When using a class of SVM models, multiple sets of historical normal data should be selected from the original data as the training set.

[0125] When a new window is used as a sample and a class of SVM models is used to classify the sample, if the output of the class of SVM models is positive, the sample is within the normal boundary, and the closer it is to 0, the more critical it is.

[0126] If the value is negative, the sample is outside the boundary, indicating an anomaly.

[0127] In this embodiment, normalization can be used to map the output values ​​of a class of SVM models to the [0,1] interval, which facilitates subsequent processing.

[0128] In this embodiment, for The normalization of the values ​​is calculated as follows:

[0129] ;

[0130] in, The function is a normalization function, where k and c are scaling parameters. The normalization calculation is a current technique and will not be described in detail.

[0131] Step 3: The window analysis module sets static quantile threshold intervals and scores the anomaly level. Make a judgment: if the anomaly score is... If the score is less than the lower limit of the quantile threshold interval, it is considered normal; if the upper limit of the quantile threshold interval is less than the abnormality score... If this occurs, it is considered an anomaly, and the window's start time is taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. Combine the data in the window and end; if the anomaly score is... If the value falls within the quantile threshold range, then proceed to step 4 for probabilistic adaptive determination.

[0132] Step 3 includes the following steps:

[0133] Step 3-1: Select a set of normal data from the original data as training data. Based on the training data and the methods in Steps 1 to 2, obtain a set of anomaly scores based on the training data. Take set average As the baseline score for normal data, a quantile threshold interval is calculated based on the baseline score. The specific calculation formula is as follows:

[0134] ;

[0135] in, Indicates will The The percentile value is used as the static lower limit threshold. Indicates will The Percentile value, used as a static upper limit threshold; This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; The value is 95%. The value is 99%;

[0136] In this embodiment, features are extracted from data under normal operating conditions in steps 1 and 2, and an anomaly score Ds for each window is obtained through a type of SVM.

[0137] Since the training data itself is "normal", these Ds values ​​should mostly be concentrated in the low range (close to normal).

[0138] When analyzing the distribution of these Ds, we find that they exhibit a certain "long-tail" characteristic:

[0139] Most data points are concentrated in a small area (stable and normal operating conditions).

[0140] A small number of data points may be overestimated due to noise, disturbances, or other reasons.

[0141] To distinguish between "obviously normal" and "obviously abnormal", this embodiment sets a threshold range:

[0142] lower limit (i.e., 95%): Ensure that the majority of truly normal data is included;

[0143] upper limit (i.e., 99%): Used to remove data that deviates significantly from the normal range.

[0144] In this embodiment, 95% and 99% are selected using quantiles:

[0145] The 95th percentile means that after sorting all the training data by Ds, 95% of the data are less than this value;

[0146] The 99th percentile means that 99% of the normal data are less than this value.

[0147] Therefore: If the anomaly score of a new data window is lower than This can almost be considered normal. If a new window's rating exceeds... It can be judged as abnormal with a high degree of confidence.

[0148] Data falling within this range is uncertain whether it is a minor disturbance or a potential fault, and step 4 needs to be performed for further judgment.

[0149] This ensures that the data processing process has both stability (avoiding false alarms) and sensitivity (avoiding missed detections).

[0150] Step 3-2: In actual operation, for the new t-th window obtained through the methods in Step 1 and Step 2, score its anomaly. Perform static determination:

[0151] ;

[0152] in, This represents the judgment result for the t-th window;

[0153] Steps 3-5: If If it is not critical, then output. The result, if If the error is abnormal, the start time of the t-th window will be taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. End with the data in the window;

[0154] like The result is critical, so proceed to step 4.

[0155] Step 4: Window Analysis Module Characteristic Volatility ,according to The length of the window is automatically selected to obtain an adaptive window. The feature vector is then extracted and the anomaly score is recalculated on the adaptive window using the method in step 2, resulting in a feature vector based on the adaptive window. and anomaly score Anomaly score for adaptive window Anomaly score compared to the corresponding original window Perform a weighted summation to obtain the final anomaly score. Regarding the final anomaly degree Perform probabilistic judgment to obtain the probability of failure of the adaptive window. ;

[0156] Step 4 includes the following steps:

[0157] Step 4-1: Obtain For the t-th window that is at the critical point, obtain the feature vector of that window. Simultaneously, retrieve the feature vector of the (t-1)th window. Calculate the characteristic volatility of these two adjacent windows. :

[0158] ;

[0159] in, The second norm of a vector;

[0160] Step 4-2: Preset the upper limit of the rate of change and lower limit of rate of change Based on characteristic volatility Adaptive selection of window length :

[0161] like Then use a fixed number of points Shorten the window;

[0162] like Then use a fixed number of points Lengthen the window;

[0163] Step 4-3: For the adjusted adaptive window, use the methods from Steps 1 to 3 to re-extract the feature vector. And calculate the anomaly score of the adaptive window. ;

[0164] Step 4-4: Scoring the anomaly of the adaptive window Anomaly score compared to the corresponding original window The anomaly is then fused to obtain the final anomaly score. :

[0165] ;

[0166] in, These are adaptive weights used based on characteristic volatility. Adjust the weight of the original window of fixed length. ;

[0167] Steps 4-5: Determine the final anomaly level Perform probabilistic judgment to obtain the probability of failure of the adaptive window. :

[0168] ;

[0169] in, This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; This indicates the probability that the adaptive window is faulty. The closer it is to 1, the more likely it is to be faulty, i.e., the window is abnormal.

[0170] When performing windowed analysis on motor operation data, fixed-length windows may have problems with insufficient or redundant information: if the window is too short, feature extraction is unstable and easily affected by noise; if the window is too long, abnormal signals will be smoothed out, reducing detection sensitivity.

[0171] Therefore, this embodiment introduces characteristic volatility. As a basis for adaptive adjustment The larger the value, the more drastic the differences in characteristics between adjacent windows, which means that the motor's operating state fluctuates significantly. The smaller the value, the smaller the difference between adjacent windows, and the more stable the operation.

[0172] This embodiment also presets an upper limit value for the rate of change. and lower limit of rate of change ,

[0173] like This indicates that the characteristics are fluctuating drastically, and the window needs to be shortened to capture rapidly changing anomalies. In this case, a fixed number of points is used. Shorten the window, that is, reduce the window to half its original size;

[0174] like This indicates that the feature changes are gradual, so the window can be extended to reduce the impact of noise. Therefore, a fixed number of points is used. Lengthen the window;

[0175] Typically, with Fs = 10kHz and the default window ≈ 10 electrical cycles around the default sampling period. , .

[0176] This embodiment re-extracts features from the adjusted adaptive window and calculates the anomaly score, then fuses it with the corresponding original window anomaly score, based on the fused result. To adjust adaptive weights Generally speaking:

[0177] High volatility → Increase the weight of the adaptive window and decrease the weight of the original window;

[0178] When volatility is low, increase the weight of the original window to maintain stability.

[0179] The scope of adjustment is This fusion ensures that the detection can respond quickly to anomalies while avoiding false alarms.

[0180] After obtaining the fusion results Then, this embodiment calculates its probabilistic representation to obtain the probability of failure of the adaptive window. , This reflects the probability that the window is faulty. The closer the value is to 1, the higher the confidence level of the fault. This embodiment uses a probabilistic flexible judgment, which is convenient for subsequent cross-window aggregation.

[0181] Step 5: The fault diagnosis module establishes a continuous triggering strategy based on the K / M principle, and assigns the probability of fault for each adaptive window. The data is summarized, and the faults are checked against the continuous triggering strategy. If the triggering conditions in the continuous triggering strategy are met, an alarm message is generated and output; the alarm message includes the fault occurrence time. The data in the adaptive window is now complete.

[0182] Step 5 includes the following steps:

[0183] Step 5-1: Retrieve the fault probability of any t0th adaptive window ;

[0184] Step 5-2: Establish a continuous triggering strategy Specifically, this refers to the K / M principle:

[0185] ;

[0186] in, To adapt the failure probability of the window, Here, K is the probability threshold, with a default value of 0.5. M is the total number of consecutive adaptive windows, and K is the threshold for the number of adaptive windows judged as abnormal. This indicates an indicator function that, when the condition is met... If the condition is met, the indicator function outputs 1; otherwise, it outputs 0. Indicates the starting index of the contiguous window, that is, the first... The first adaptive window is used as the starting window of the continuous adaptive window, and t is the index of the continuous adaptive window;

[0187] like If the window is not found to be faulty, then the adaptive window is confirmed to be faulty, i.e., the window is abnormal; otherwise, the next adaptive window is checked sequentially downwards.

[0188] In this embodiment, to improve the reliability of the judgment, a continuous triggering principle is introduced: in M ​​consecutive adaptive windows, if at least K windows have a failure probability... Exceeding the threshold If the condition is "continuously abnormal", a fault alarm will be output. Here, M is the total number of continuous adaptive windows, which is usually 10 by default, and K is 5 by default.

[0189] Step 5-3: Place the first The start time of each adaptive window is used as the time of failure. Generate alarm information and output the time of fault occurrence. Generate alarm information ALM2 and output the fault occurrence time. And the data in the adaptive window.

[0190] In this embodiment, the final output includes an anomaly adaptive window (data set) confirmed by the continuous triggering strategy and the corresponding fault occurrence time. This output is a filtered and probability-determined set of anomaly data, containing raw data such as three-phase current, DC bus voltage, and timestamps. This anomaly data can serve as input for subsequent fault analysis, diagnosis, maintenance decisions, or further training of deep learning models, participating in processes such as health assessment of the servo motor drive system. The method of this invention enables rapid data tracing, and because the data windows are modular, their retrieval speeds are generally consistent, allowing for effective prediction of processing time before tracing, greatly improving the visualization of data processing.

[0191] This invention discloses an artificial intelligence-based fault detection method for servo permanent magnet synchronous motors, solving the technical problems of achieving high real-time detection, predictable processing, and low false alarm rate in servo permanent magnet synchronous motor fault detection. This invention employs a windowed sampling data processing method, dividing continuous data into fixed-length windows. Features are extracted and anomaly determination is performed within each window, ensuring deterministic processing time while maintaining sensitivity to abnormal fluctuations through an adaptive window mechanism. Combined with quantile threshold determination and a continuous triggering strategy, the false alarm rate is significantly reduced, and only abnormal data is output, facilitating subsequent expansion to diagnose specific fault types.

Claims

1. A fault detection method for servo permanent magnet synchronous motors based on artificial intelligence, characterized in that: Includes the following steps: Step 1: The data acquisition module synchronously acquires the raw data generated by the servo motor drive system, including the three-phase current. and DC bus voltage The continuous sampling frames of the original data are divided into several windows for data analysis; Step 2: The feature extraction module calculates the feature vector for each window. The feature vector includes temporal and order components. The feature vector is input into a class of SVM models to obtain the original decision output. The original decision output is then mapped to a normalized anomaly score for the window. ; Step 3: The window analysis module sets static quantile threshold intervals and scores the anomaly level. Make a judgment: if the anomaly score is... If the score is less than the lower limit of the quantile threshold interval, it is considered normal; if the upper limit of the quantile threshold interval is less than the abnormality score... If this occurs, it is considered an anomaly, and the window's start time is taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. Combine the data in the window and end; if the anomaly score is... If the value falls within the quantile threshold range, then proceed to step 4 for probabilistic adaptive determination. Step 4: Analyze characteristic volatility using the window analysis module. ,according to The length of the window is automatically selected to obtain an adaptive window. Following the methods in steps 1 to 3, feature vector extraction and anomaly scoring are recalculated on the adaptive window to obtain feature vectors based on the adaptive window. and anomaly score Anomaly score for adaptive window Anomaly score compared to the corresponding original window Perform a weighted summation to obtain the final anomaly score. Regarding the final anomaly degree Perform probabilistic judgment to obtain the probability of failure of the adaptive window. : ， ； in, This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; This indicates the probability that the adaptive window is faulty; Step 5: The fault detection module retrieves the fault probability of any t0-th adaptive window. , will the The start time of each adaptive window is used as the time of failure. Establish a continuous triggering strategy based on the K / M principle, and assign the probability of failure to each adaptive window. The data is summarized, and the faults are checked against the continuous triggering strategy. If the triggering conditions in the continuous triggering strategy are met, an alarm message is generated and output; the alarm message includes the fault occurrence time. The data in the adaptive window is now complete.

2. The fault detection method for a servo permanent magnet synchronous motor based on artificial intelligence as described in claim 1, characterized in that: Step 1 includes the following steps: Step 1-1: The data acquisition module reads the synchronously acquired three-phase current from the servo motor drive system. and DC bus voltage Generate a continuous sampling sequence : ； in, These are the phase a current value, phase b current value, and phase c current value at the nth sampling point, respectively. This represents the total number of sampling points; The DC bus voltage at the nth sampling point; Step 1-2: Preset a fixed number of periods, and sample the sequence. The data is divided into several windows for analysis based on a fixed number of periods, and a dataset is constructed for each window. Each window contains One sampling point: ； ； in, This represents the data set of the t-th window; and These are the indices of the starting and ending sampling points of the t-th window, respectively. These represent the phase a current values ​​at the nth sampling point in the t-th window, respectively. b-phase current value c-phase current value DC bus voltage value .

3. The fault detection method for a servo permanent magnet synchronous motor based on artificial intelligence as described in claim 2, characterized in that: Step 2 includes the following steps: Step 2-1: Calculate the feature vector for each window, specifically including temporal feature calculation and order component feature calculation: The time-domain feature calculation is as follows: Calculate the window value using the following formulas. In root mean square value : ； ； ； ； in, , , and These represent the current in the t-th window, respectively. Root mean square value, current Root mean square value, current Root mean square value and DC bus voltage Root mean square value; The calculation of the ordinal component features is as follows: Calculate negative order components Amplitude: ， ； Calculate zero-sequence current The amplitude is calculated using the following formula: ； Step 2-2: Constructing Feature Vectors The details are as follows: ； in, This represents the feature vector of the t-th window; Steps 2-3: ... The input is fed into a class of SVM models, and the output is anomaly score. : ； in, Represents a class of SVM model mapping functions. This represents the original anomaly score for the t-th window. The value of is normalized to the interval [0,1].

4. The fault detection method for a servo permanent magnet synchronous motor based on artificial intelligence as described in claim 3, characterized in that: Step 3 includes the following steps: Step 3-1: Select a set of normal data from the original data as training data. Based on the training data and the methods in Steps 1 to 2, obtain a set of anomaly scores based on the training data. Take set average As the baseline score for normal data, a quantile threshold interval is calculated based on the baseline score. The specific calculation formula is as follows: ; in, Indicates will The The percentile value is used as the static lower limit threshold. Indicates will The Percentile value, used as a static upper limit threshold; This represents the lower limit of the quantile threshold interval. This represents the upper limit of the quantile threshold range; The value is 95%. The value is 99%; Step 3-2: In actual operation, for the new t-th window obtained through the methods in Step 1 and Step 2, score its anomaly. Perform static determination: ； in, This represents the judgment result for the t-th window; Step 3-3: If If it is not critical, then output. The result, if If the error is abnormal, the start time of the t-th window will be taken as the time of the fault occurrence. Generate alarm information ALM1 and output the fault occurrence time. End with the data in the window; like The result is critical, so proceed to step 4.

5. The fault detection method for a servo permanent magnet synchronous motor based on artificial intelligence as described in claim 4, characterized in that: Step 4 includes the following steps: Step 4-1: Obtain For the t-th window that is at the critical point, obtain the feature vector of that window. Simultaneously, retrieve the feature vector of the (t-1)th window. Calculate the characteristic volatility of these two adjacent windows. : ； in, The second norm of a vector; Step 4-2: Preset the upper limit of the rate of change and lower limit of rate of change Based on characteristic volatility Adaptive selection of window length : like Then use a fixed number of points Shorten the window; like Then use a fixed number of points Lengthen the window; Step 4-3: For the adjusted adaptive window, use the methods from Steps 1 to 3 to re-extract the feature vector. And calculate the anomaly score of the adaptive window. ; Step 4-4: Scoring the anomaly of the adaptive window Anomaly score compared to the corresponding original window The anomaly is then fused to obtain the final anomaly score. : ； in, These are adaptive weights used based on characteristic volatility. Adjust the weight of the original window of fixed length. ; Steps 4-5: Determine the final anomaly level Perform probabilistic judgment to obtain the probability of failure of the adaptive window. .

6. The fault detection method for a servo permanent magnet synchronous motor based on artificial intelligence as described in claim 5, characterized in that: Step 5 includes the following steps: Step 5-1: Retrieve the fault probability of any t0th adaptive window Step 5-2: Establish a continuous triggering strategy Specifically, this refers to the K / M principle: ； in, To adapt the failure probability of the window, Here, K is the probability threshold, with a default value of 0.

5. M is the total number of consecutive adaptive windows, and K is the threshold for the number of adaptive windows judged as abnormal. This indicates an indicator function that, when the condition is met... If the condition is met, the indicator function outputs 1; otherwise, it outputs 0. Indicates the starting index of the contiguous window, that is, the first... A first adaptive window is used as the starting window of a continuous adaptive window, and t is the index of the continuous adaptive window; if If the window is not found to be faulty, then the adaptive window is confirmed to be faulty, i.e., the window is abnormal; otherwise, the next adaptive window is checked sequentially downwards. Step 5-3: Place the first The start time of each adaptive window is used as the time of failure. Generate alarm information and output the time of fault occurrence. Generate alarm information ALM2 and output the fault occurrence time. And the data in the adaptive window.

Images