A temperature early warning method and system for motor stator winding

Abstract

The application relates to the technical field of data processing, in particular to a temperature early warning method and system for a motor stator winding; the method comprises the following steps: acquiring multi-dimensional data related to winding temperature when the motor is running in real time, and acquiring real-time prediction error by using a prediction model; obtaining a drift index at each moment according to the difference between the real-time prediction error and a health error benchmark; based on the autocorrelation characteristics of the historical sequence of the drift index, a plurality of monitoring windows of different time scales are constructed; the abnormal score of each monitoring window is calculated according to the drift index in the monitoring window; the abnormal scores of the plurality of monitoring windows are weighted and fused to obtain a comprehensive abnormal index for temperature early warning of the motor stator winding. The application has the effect of improving the accuracy of temperature early warning of the motor stator winding.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for early warning of temperature of motor stator windings. Background Technology

[0002] In modern industrial production, the temperature of the motor stator windings is a core indicator of their operational health. Abnormally high temperatures accelerate the aging of insulation materials, shorten motor lifespan, and in severe cases, even lead to winding burnout, causing unplanned downtime and significant economic losses. Therefore, accurate and reliable predictive maintenance of motor winding temperatures has become a critical issue in the industry.

[0003] Therefore, existing temperature early warning methods mainly rely on experience-based threshold judgments or classification models trained on historical fault samples. For example, collecting multi-dimensional operating data related to motor temperature (such as current, load, vibration, etc.) and using machine learning models for intelligent prediction. Alternatively, algorithms based on recurrent neural networks such as gated recurrent units (GRUs) can be used to construct time-series prediction models, learning the complex mapping relationship between multi-dimensional inputs and temperature to achieve motor state identification and temperature early warning.

[0004] However, temperature warning methods in related technologies have some shortcomings: First, methods based on fixed thresholds cannot adapt to the dynamic differences in prediction errors under different operating conditions, and are prone to generating a large number of false alarms during load changes or start-up and shutdown processes; Second, methods that rely on historical fault data require the collection of sufficient typical fault samples, but in actual engineering, it is often difficult to obtain rich fault data for motors, resulting in insufficient model generalization ability and a lack of detection capability for unknown anomalies; Finally, during long-term operation, the thermal characteristics of motors will gradually drift due to material aging, changes in cooling conditions, etc., and traditional static models lack adaptive update mechanisms, making it difficult to maintain long-term stable warning performance. Summary of the Invention

[0005] To improve the accuracy of motor temperature warning, this application provides a method and system for warning the temperature of motor stator windings.

[0006] Firstly, this application provides a method for early warning of temperature in motor stator windings, employing the following technical solution: A method for early warning of motor stator winding temperature includes: acquiring historical operating data related to winding temperature during motor operation, and constructing a prediction model for predicting motor stator winding temperature based on the operating data; inputting operating data related to winding temperature during healthy motor operation into the prediction model to obtain prediction error; dividing the normal operating state of the motor into multiple operating conditions, grouping the prediction error generated by the prediction model according to the operating conditions to obtain multiple error state groups, and calculating the health error benchmark for each error state group. Multidimensional data related to winding temperature during real-time motor operation are acquired, and a prediction model is used to obtain the real-time prediction error. The drift index at each time point is obtained based on the difference between the real-time prediction error and the health error benchmark. Based on the autocorrelation characteristics of the historical sequence of the drift index, multiple monitoring windows at different time scales are constructed. Based on the drift index within each monitoring window, the anomaly score of each monitoring window is calculated. The anomaly scores of multiple monitoring windows are weighted and fused to obtain a comprehensive anomaly index for temperature warning of the motor stator winding.

[0007] On the one hand, this application divides the normal operation of the motor into multiple operating states and establishes an independent health error benchmark for each operating state. This step can adapt to the fluctuations in the normal temperature prediction error under different operating conditions. If the prediction error corresponding to the data deviates from the normal fluctuation during real-time monitoring, it indicates that the data may be abnormal, effectively solving the problem of a large number of false alarms caused by the inability to adapt to dynamic errors during load changes or start-stop processes in existing technologies. On the other hand, this method can establish a prediction model for motor winding temperature by collecting motor health operation data. By monitoring the drift index of the real-time prediction error relative to the health baseline, anomalies are identified, thereby eliminating the dependence on difficult-to-obtain fault samples and enabling it to generalize the detection of unknown anomaly types. In addition, by constructing multiple monitoring windows with different time scales and weighted fusion of anomaly scores, this method can simultaneously capture instantaneous temperature changes and slow deterioration trends caused by equipment aging, achieving long-term adaptive tracking of motor thermal characteristics. This overcomes the shortcomings of traditional static models in dealing with concept drift, thus providing more accurate, robust, and proactive temperature warnings.

[0008] Optionally, the step of dividing the normal operating state of the motor into multiple operating conditions includes: acquiring the operating condition data of the motor, which includes operating condition load rate and load change rate data, and clustering the operating condition data to obtain multiple operating conditions.

[0009] By performing cluster analysis on the two key indicators of operating condition load rate and load change rate, various typical operating states of the motor can be automatically identified more objectively and accurately, such as steady state, transient state, high load, and low load. This refined classification of operating conditions provides a solid foundation for establishing more targeted health error benchmarks in the future.

[0010] Optionally, the steps for obtaining the prediction error include: using a sensor to collect the temperature of the motor windings as the actual temperature value, obtaining the predicted value from the prediction model, and using the difference between the actual temperature value and the predicted value as the prediction error.

[0011] The deviation of the model prediction is directly quantified by subtracting the actual temperature value collected by the sensor from the predicted value output by the prediction model.

[0012] Optionally, the step of calculating the health error benchmark for each error state group includes: for any operating condition, calculating the mean and standard deviation of the prediction error corresponding to that operating condition, and using the mean and standard deviation of the prediction error as the health error benchmark.

[0013] The mean of the prediction error reflects the level of normal prediction error of the prediction model under a certain operating condition, while the standard deviation of the prediction error reflects the fluctuation of normal prediction error of the prediction model under a certain operating condition. By using the two, a healthy error benchmark is determined to reflect the distribution of the error of the prediction model under normal conditions.

[0014] Optionally, the step of obtaining the drift index at each time point based on the difference between the real-time prediction error and the health error benchmark includes: for any given time point, obtaining the corresponding operating condition data, determining the operating condition state at that time point, using the absolute difference between the prediction error at that time point and the mean of the health error benchmark for the corresponding operating condition state as the deviation, and using the ratio of the deviation to the standard deviation in the health error benchmark as the drift index.

[0015] The deviation is obtained by subtracting the mean of the real-time prediction error from the mean of the corresponding health error benchmark, and then standardized by the standard deviation to obtain the drift index that reflects the degree of deviation of the current data.

[0016] Optionally, the steps of constructing multiple monitoring windows at different time scales based on the autocorrelation characteristics of the historical series of the drift index include: for any given time, constructing a historical observation window based on that time, obtaining the base time scale based on the integral value of the autocorrelation function of the drift index in the historical observation window; scaling the base time scale to different degrees to obtain a set of candidate windows, and selecting multiple candidate windows from the set of candidate windows as monitoring windows.

[0017] The base time scale is determined and scaled based on the autocorrelation function of the historical drift index series. This adaptive windowing method can determine the analysis scale according to the inherent temporal correlation characteristics of the data itself, ensuring that the length of the monitoring window matches the period of information change in the data, thereby more effectively capturing abnormal patterns at different time scales.

[0018] Optionally, the step of selecting multiple candidate windows as monitoring windows from the candidate window set includes: calculating the significance index of the degree of change of drift index in each candidate window; constructing multiple subsets in the candidate window set, calculating the evaluation index of the subset based on the significance index of each window in the subset, selecting the subset with the largest evaluation index as the optimal subset, and selecting the windows in the optimal subset as monitoring windows.

[0019] To reduce the significant overhead and potential noise interference from computation across all possible candidate windows, this invention selects the window combination that best reflects abnormal changes by calculating the significance index of each window and constructing a subset for optimization. This method improves computational efficiency while maintaining early warning sensitivity.

[0020] Optionally, the step of calculating the evaluation index of the subset based on the saliency index of each window of the subset includes: for any subset, setting a scoring penalty based on the number of windows in the subset; for any window, taking the difference between the saliency index corresponding to the window and the scoring penalty as the local score, and taking the sum of the local scores corresponding to each window as the evaluation index of the subset.

[0021] Optionally, in the step of weighted fusion of anomaly scores from multiple monitoring windows, for any given moment, the window difference between the monitoring window used at the current moment and the previous moment is calculated. If the window difference is greater than a preset threshold, the monitoring window is updated. The anomaly score is calculated using the monitoring window corresponding to the current moment, and then weighted fusion is performed.

[0022] Secondly, this application provides a temperature warning system for motor stator windings, employing the following technical solution: A temperature warning system for motor stator windings includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the temperature warning method for motor stator windings as described above.

[0023] The above-mentioned method for temperature warning of motor stator winding is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. Thus, a system can be built based on the memory and processor for convenient use.

[0024] This application has the following technical effects: By establishing dynamic health error benchmarks for different operating conditions, the problem of high false alarm rate and inability to adapt to changes in operating conditions in traditional fixed threshold methods is overcome. At the same time, the autocorrelation of the drift index is used to dynamically construct multi-scale monitoring windows, and information from each scale is integrated for comprehensive judgment, enabling it not only to detect sudden anomalies, but also to effectively capture slow performance degradation caused by equipment aging. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for temperature early warning of motor stator windings according to an embodiment of this application. Detailed Implementation

[0026] This application discloses a method for early warning of stator winding temperature in a motor. First, a high-precision time-series prediction model is trained using data from the motor's normal operating conditions to predict stator winding temperature in real time. Second, an adaptive conditional error reference distribution is created to dynamically set reasonable prediction error benchmarks for different operating states (such as steady state, transient state, and different loads). By capturing the winding temperature change characteristics under healthy conditions, the method determines whether the current temperature is abnormal based on the difference between the temperature change characteristics of the real-time collected data and those under healthy conditions, thus eliminating reliance on historical abnormal data.

[0027] By constructing a multi-scale, adaptively adjusting monitoring system, the drift index between the current prediction error and the health error benchmark is calculated. Finally, by fusing drift evidence from multiple scales, a stable and reliable comprehensive anomaly index is output, thereby achieving generalized capture and early warning of unknown abnormal heat sources. This effectively overcomes the concept drift problem caused by equipment aging and achieves long-term, stable, and highly intelligent early warning.

[0028] Reference Figure 1 This includes steps S1-S5.

[0029] S1: Obtain historical operating data related to the winding temperature of the motor, and build a prediction model based on the operating data to predict the stator winding temperature of the motor.

[0030] The system acquires multi-dimensional raw signals that comprehensively characterize the historical operation of the motor in real time and with high synchronization. In this embodiment, the raw signals mainly include: stator winding temperature, three-phase current, three-phase voltage, and vibration signals.

[0031] As an example, the acquisition methods for various signals are as follows: For winding temperature T: directly utilize the platinum resistance thermometer embedded in the motor stator slot, combined with a high-precision temperature transmitter, to acquire data at a frequency of 1Hz. For three-phase current / voltage: employ a high-bandwidth Hall effect sensor or Rogowski coil, combined with a high-speed data acquisition card (DAQ), to synchronously acquire instantaneous waveform data of three-phase current and voltage at a sampling rate of not less than 10kHz. For vibration signals: install IEPE (Integrated Electronics Piezo-Electric) triaxial accelerometers in the vertical, horizontal, and axial directions on the bearing housings at both the motor drive end and non-drive end, and acquire vibration signals at a sampling rate of not less than 20kHz.

[0032] All collected data are accompanied by precise timestamps to ensure strict temporal alignment of multidimensional data in subsequent analysis.

[0033] The acquired raw signals are preprocessed. For high-frequency signals, such as three-phase current / voltage and vibration signals, wavelet threshold denoising or bandpass filters can be used to filter out power frequency interference and high-frequency white noise that are unrelated to the motor's operating state.

[0034] Subsequently, within a 1-second time window, the mean, variance, root mean square (RMS), kurtosis, and other time-domain statistical characteristics of each preprocessed signal (current, vibration) are calculated. These time-domain statistical characteristics are naturally aligned in time with the winding temperature collected at a frequency of 1 Hz, together forming the running data used for model training.

[0035] To eliminate the dimensional differences between different features, this embodiment uses the Min-Max method to linearly scale all feature values ​​to the [0,1] interval.

[0036] A prediction model is trained based on operational data to establish a nonlinear mapping relationship between operational state characteristics (i.e., three-phase voltage, three-phase current, vibration signals, etc.) and winding temperature. In this embodiment, a sequence-to-vector regression model based on gated cyclic units (GRUs) is used.

[0037] The GRU model is implemented as follows: It aims to learn to predict the winding temperature at the next moment by utilizing multi-dimensional operating characteristics of the motor over a past period. The model's input is a time series segment, for example, containing multi-dimensional feature data from the past 60 time points (i.e., 60 seconds); the model's output is the predicted winding temperature at the next time point.

[0038] The GRU network can be configured to contain two hidden layers, with each layer preferably containing 128 hidden units. During training, the Adam optimizer is used, with an initial learning rate of 0.001 and a loss function of mean squared error (MSE). By training on a large amount of normal operating condition data, a predictive model capable of accurately predicting normal temperature changes is finally obtained. This model is a conventional technique in this field and will not be elaborated further here.

[0039] S2: Input the operating data related to the winding temperature during the healthy operation of the motor into the prediction model to obtain the prediction error; divide the normal operating state of the motor into multiple operating states, group the prediction error generated by the prediction model according to the operating state to obtain multiple error state groups, and calculate the health error benchmark for each error state group.

[0040] The predicted value of the winding temperature is obtained by using the prediction model trained in step S1, and the difference between the predicted value and the actual temperature value collected by the sensor is used as the prediction error.

[0041] The normal operating state of the motor is divided into multiple operating conditions: the operating condition data of the motor is obtained, including the operating condition load rate and load change rate data, and the operating condition data is clustered to obtain multiple operating conditions.

[0042] In the operating condition data, the load rate can be directly obtained from the SCADA / PLC system; the load change rate is obtained by smoothing the time series of the load rate using a Savitzky-Golay filter and then calculating its first difference. This indicator can robustly reflect the trend and rate of load change. Together, they reflect different operating conditions under normal motor operation. Next, the K-Means clustering algorithm is used to cluster the operating condition data during historical normal operation. For example, the value of K can be set to 5, thereby dividing the motor operation process into several typical operating condition clusters, such as: low load steady state, medium load steady state, heavy load steady state, rapid load increase transient (this state reflects the motor start-up process), and rapid load decrease transient (this state reflects the motor shutdown process or the transition from high load to low load conditions), etc.

[0043] Based on the above analysis, each moment corresponds to a working condition and a prediction error. Therefore, the prediction error can also be classified according to the working condition corresponding to the prediction error, thus obtaining multiple error state groups. Each error state group corresponds to the prediction error at multiple moments.

[0044] For any operating condition, calculate the mean and standard deviation of the prediction error corresponding to that condition, and use the mean and standard deviation of the prediction error as the health error benchmark. The mean and standard deviation reflect the overall level and fluctuation of the prediction error in that error state group, and are used to reflect the health error benchmark of the prediction model under that operating condition when the motor is working normally.

[0045] S3: Obtain multi-dimensional data related to winding temperature during real-time motor operation, use a prediction model to obtain real-time prediction error, and obtain the drift index at each time point based on the difference between the real-time prediction error and the health error benchmark.

[0046] In this step, the multidimensional data related to the winding temperature are the same as the data obtained in step S1, and the acquisition of the prediction error is the same as in step S2, so it will not be repeated here.

[0047] For any given moment, acquire the corresponding operating condition data, determine the operating condition state at that moment, and use the absolute difference between the prediction error at that moment and the mean of the health error benchmark for the corresponding operating condition state as the deviation. Use the ratio of the deviation to the standard deviation in the health error benchmark as the drift index. For any given time, the corresponding operating condition data can be obtained. The Euclidean distance between the operating condition data at that time and the cluster centers corresponding to each operating condition state in step S2 is calculated. The operating condition state with the smallest Euclidean distance is taken as the operating condition state to which that time belongs. Then, the drift index is obtained based on the health baseline corresponding to the operating condition state.

[0048] The formula for calculating the drift index can be expressed as: In the formula, Indicates time The drift index; Indicates time Corresponding prediction error; Indicates operating status The mean of the prediction error in the corresponding healthy baseline; Operating conditions The standard deviation of the prediction error in the corresponding healthy baseline. The probability of the standard deviation being 0 is extremely small, so we will not consider this case here. Of course, to avoid the denominator being 0, a positive number can be set in the denominator (e.g., ...). (This value is an empirical value set by those skilled in the art based on experience) added to the standard deviation.

[0049] S4: Based on the autocorrelation characteristics of the historical series of the drift index, multiple monitoring windows with different time scales are constructed.

[0050] The reason for constructing multiple monitoring windows with different time scales in this step is to enable long-term, medium-term, and short-term analysis of the drift index, thereby improving the accuracy of subsequent early warning of motor winding temperature.

[0051] For any given moment, a historical observation window is constructed based on that moment. The base time scale is obtained by integrating the autocorrelation function of the drift exponent within the historical observation window. Different scaling degrees are applied to the base time scale to obtain a set of candidate windows. Multiple candidate windows are then selected from this set as monitoring windows. The length of the historical observation window in this step is determined based on the experience of those skilled in the art.

[0052] In the formula, time The effective relevant time; Sample time interval (e.g., 1 second or 1 minute, the value of which is preset by those skilled in the art based on experience); : Autocorrelation function in lag The estimated value at the location (using unbiased or fast FFT estimation); The cutoff lag for autocorrelation calculations is usually set to such that... The hysteresis when the value first drops to 0 or below a threshold (e.g., 0.05).

[0053] The autocorrelation function of a time series describes the correlation between the series at different time lags. Integrating (or summing) the autocorrelation function yields a comprehensive index that measures the memory length of the series. The larger this index, the more the current value of the series is correlated with historical values ​​from a more distant time period, requiring a longer observation window to smooth out noise.

[0054] For any given time, a set of candidate windows is constructed based on the corresponding effective correlation time, and multiple candidate windows constitute a candidate window set. In this embodiment, the effective correlation time is scaled using a preset scaling factor to generate windows of different lengths. Specifically, the scaling factor set in this embodiment is a positive integer from 1 to 10. Multiple scaling factors are multiplied by the effective correlation time to obtain a set of window lengths, and thus a set of candidate windows of different lengths.

[0055] To prevent candidate windows from being too long or too short, an upper limit and a lower limit can be set to remove candidate windows that are longer than the upper limit or shorter than the lower limit.

[0056] When analyzing the degree of data anomalies, not all windows can reflect the anomalies well. Therefore, in order to reduce the amount of computation and ensure the accuracy of subsequent analysis of data anomalies, candidate windows in the candidate window set are filtered here.

[0057] The steps for selecting multiple candidate windows as monitoring windows from the candidate window set include: calculating the significance index of the drift index change in each candidate window; constructing multiple subsets in the candidate window set; calculating the evaluation index of the subset based on the significance index of each window in the subset; selecting the subset with the largest evaluation index as the optimal subset; and selecting the windows in the optimal subset as monitoring windows.

[0058] In this embodiment, for any candidate window, if the offset index within that window is relatively large, it indicates that anomalies in the data can be detected more effectively within that window. If the offset index in a candidate window is small, there are two possibilities: either the data in the current window is normal, or short-term anomalies are being overwhelmed by data in a longer window. Therefore, in this embodiment, the normalized sum of the offset indices at each time point within the candidate window is used as the significance index of that candidate window.

[0059] Then, a monitoring window is selected based on the significance index of the candidate windows.

[0060] Specifically, multiple candidate windows in the candidate window set are further grouped to form multiple subsets. The grouping rules are random, meaning that a subset can consist of any number of different candidate windows. The evaluation index of each subset is calculated based on the significance index of each candidate window. Based on the above analysis, the larger the significance index of a candidate window, the better it can distinguish outlier data, and thus the larger the evaluation index of the subset formed by such candidate windows. To balance computational efficiency, a scoring penalty is set based on the number of candidate windows in the subset. For any window, the difference between the significance index corresponding to that window and the scoring penalty is taken as the local score, and the sum of the local scores corresponding to all windows is taken as the evaluation index of that subset.

[0061] By following the steps above, we can obtain the evaluation index of each subset, select the subset with the largest evaluation index as the optimal subset, and select the candidate windows in the optimal subset as the monitoring windows to complete the selection of the monitoring windows.

[0062] S5: Calculate the anomaly score for each monitoring window based on the drift index within each monitoring window; weight and fuse the anomaly scores from multiple monitoring windows to obtain a comprehensive anomaly index for temperature warning of the motor stator winding.

[0063] Based on the above steps, each time point corresponds to a different set of monitoring windows. To prevent the window length from frequently changing due to minor fluctuations, window updates are restricted. For any given time point, the window difference between the current time point and the previous time point is calculated. If the window difference exceeds a preset threshold, the monitoring window is updated. Anomaly scores are calculated using the monitoring window corresponding to the current time point and then weighted and fused.

[0064] The step of calculating the window difference between the monitoring windows used at the current time and the previous time includes a length difference and a quantity difference. The length difference is obtained by: for any two time points, calculating the total length of the corresponding monitoring windows, and taking the absolute value of the difference between the total lengths at the two time points as the length difference. The quantity difference is obtained by: for any two time points, obtaining the total number of monitoring windows, and taking the absolute value of the difference between the total number of monitoring windows at the two time points as the quantity difference. The sum of the normalized results of the length difference and the quantity difference is taken as the window difference.

[0065] If the window difference at the current moment is greater than a preset threshold, the monitoring window is updated, meaning the monitoring window at the current moment is used to issue a warning for the motor winding temperature. If the window difference is less than the preset threshold, the monitoring window is not updated, and the monitoring window used at the previous moment continues to be used to issue a warning for the motor winding temperature.

[0066] Anomaly scores can then be calculated based on the drift index in the monitoring window. In this embodiment, the normalized result of the sum of the drift indices in the detection window is used as the anomaly score corresponding to the monitoring window.

[0067] The weight of any monitoring window is determined based on its significance index. According to the analysis in step S4 above, a larger significance index indicates that the window is more effective at distinguishing abnormal data, and therefore the anomaly score obtained by the monitoring window is more reliable. Therefore, for any given monitoring window, the ratio of its significance index to the sum of the significance indices of all monitoring windows is used as its weight.

[0068] The anomaly scores for each monitoring window are weighted and summed based on their respective weights to obtain a comprehensive score. This comprehensive score is then used to issue early warnings for the motor stator winding temperature. For example, an anomaly threshold can be set; when the comprehensive score exceeds the threshold, an alarm is triggered, reminding staff to monitor the motor's operating data and the stator winding's condition.

[0069] This application also discloses a temperature warning system for motor stator windings, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a temperature warning method for motor stator windings according to this application is implemented.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0071] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for early warning of temperature in motor stator windings, characterized in that, include: Obtain historical operating data related to the winding temperature of the motor, and build a predictive model based on the operating data to predict the stator winding temperature of the motor. The operating data related to winding temperature during the healthy operation of the motor are input into the prediction model to obtain the prediction error; the normal operating state of the motor is divided into multiple operating states, and the prediction error generated by the prediction model is grouped according to the operating state to obtain multiple error state groups, and the health error benchmark of each error state group is calculated. Multidimensional data related to winding temperature during real-time motor operation are acquired, and a prediction model is used to obtain the real-time prediction error. The drift index at each time point is obtained based on the difference between the real-time prediction error and the health error benchmark. Based on the autocorrelation characteristics of the historical sequence of the drift index, multiple monitoring windows at different time scales are constructed. Based on the drift index within each monitoring window, the anomaly score of each monitoring window is calculated. The anomaly scores from multiple monitoring windows are weighted and fused to obtain a comprehensive anomaly index, which is used for temperature warning of the motor stator winding.

2. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, The steps to divide the normal operating state of the motor into multiple operating conditions include: acquiring the operating condition data of the motor, which includes operating condition load rate and load change rate data, and clustering the operating condition data to obtain multiple operating conditions.

3. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, The steps to obtain the prediction error include: using a sensor to collect the temperature of the motor windings as the actual temperature value, obtaining the predicted value from the prediction model, and taking the difference between the actual temperature value and the predicted value as the prediction error.

4. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, The steps for calculating the health error benchmark for each error state group include: for any operating condition, calculating the mean and standard deviation of the prediction error corresponding to that operating condition, and using the mean and standard deviation of the prediction error as the health error benchmark.

5. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, The steps for obtaining the drift index at each time point based on the difference between the real-time prediction error and the health error benchmark include: for any given time point, obtaining the corresponding operating condition data, determining the operating condition state at that time point, using the absolute difference between the prediction error at that time point and the mean of the health error benchmark for the corresponding operating condition state as the deviation, and using the ratio of the deviation to the standard deviation in the health error benchmark as the drift index.

6. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, Based on the autocorrelation characteristics of the historical series of the drift index, the steps for constructing multiple monitoring windows at different time scales include: for any given time, constructing a historical observation window based on that time; obtaining the base time scale based on the integral value of the autocorrelation function of the drift index in the historical observation window; scaling the base time scale to different degrees to obtain a set of candidate windows; and selecting multiple candidate windows from the set of candidate windows as monitoring windows.

7. The method for early warning of temperature of motor stator winding according to claim 6, characterized in that, The steps for selecting multiple candidate windows as monitoring windows from the candidate window set include: calculating the significance index of the drift index change in each candidate window; constructing multiple subsets in the candidate window set; calculating the evaluation index of the subset based on the significance index of each window in the subset; selecting the subset with the largest evaluation index as the optimal subset; and selecting the windows in the optimal subset as monitoring windows.

8. The method for early warning of temperature of motor stator winding according to claim 7, characterized in that, The steps for calculating the evaluation index of a subset based on the saliency index of each window of the subset include: for any subset, setting a scoring penalty based on the number of windows in the subset; for any window, taking the difference between the saliency index corresponding to the window and the scoring penalty as the local score, and taking the sum of the local scores corresponding to each window as the evaluation index of the subset.

9. The method for early warning of temperature of motor stator winding according to claim 1, characterized in that, In the step of weighted fusion of anomaly scores from multiple monitoring windows, for any given moment, the window difference between the monitoring window used at the current moment and the previous moment is calculated. If the window difference is greater than a preset threshold, the monitoring window is updated. The anomaly score is calculated using the monitoring window corresponding to the current moment, and then weighted fusion is performed.

10. A temperature warning system for motor stator windings, characterized in that, include: The processor and memory, the memory storing computer program instructions, when executed by the processor, implement a temperature warning method for motor stator windings according to any one of claims 1-9.

Images