A method for detecting early inter-turn short circuit fault of permanent magnet synchronous motor based on multi-scale sliding window

Abstract

The application discloses a kind of permanent magnet synchronous motor early interturn short circuit fault detection methods based on multiscale sliding window, comprising: after motor is started and runs to steady state in healthy state, using multiscale sliding window, different scale sliding window is along the positive direction of time axis sliding, and different scale data set and reference data set are generated in turn and are temporarily stored in respective database.Making difference between each data set and reference data set, residual signal is obtained, which is used as fault feature, and is temporarily stored in respective database;Set the significance level threshold of residual signal, the difference between the data sets of residual signal is quantified by combining statistical significance analysis and significance probability value, compare significance probability value and significance level threshold, and determine whether motor occurs interturn short circuit fault by combining multiscale joint decision mechanism, realize fault phase positioning simultaneously, and update reference data set adaptively.

Description

Technical Field

[0001] This invention pertains to fault diagnosis methods for permanent magnet synchronous motors, and particularly relates to a lightweight method for high-sensitivity detection of early-stage inter-turn short-circuit faults in permanent magnet synchronous motors, designed for industrial online monitoring scenarios. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in aerospace, new energy transportation, and wind power generation due to their high power density and efficiency. However, under long-term high-load and complex operating conditions, the stator windings of the motor are highly susceptible to insulation degradation caused by high temperature, vibration, and electromagnetic stress, leading to inter-turn short circuit faults (ISCF). The initial characteristics of this fault are extremely subtle; if not detected and addressed in time, the fault area will rapidly escalate due to localized overheating, eventually evolving into phase-to-phase or short-to-ground circuits, or even burning out the motor, causing irreparable economic losses and safety accidents. Therefore, detecting early inter-turn short circuit faults in PMSMs is of paramount practical importance.

[0003] For fault detection of inter-turn short circuits in PMSMs, existing methods mainly include analytical modeling, signal processing, and data-driven methods. Analytical modeling relies on constructing precise mathematical equations containing fault parameters. However, in complex real-world environments, motor physical parameters (such as inductance and resistance) drift with temperature rise and magnetic saturation, causing deviations between the theoretical model and the actual system, easily leading to false alarms or missed alarms. Signal processing methods currently primarily rely on time-frequency analysis. These methods require prior analysis of the motor model to determine the characteristic frequencies of inter-turn short circuit faults before fault diagnosis can be completed. This not only heavily depends on preliminary model analysis but also consumes significant computational resources. Furthermore, in practical industrial scenarios, these methods require specialized hardware systems. While data-driven methods have strong generalization capabilities, they heavily rely on massive amounts of labeled fault samples for model training. However, in practical industrial scenarios, actively conducting destructive inter-turn short circuit experiments is both dangerous and expensive, resulting in a severe shortage of high-quality fault data. Additionally, the complex network structure leads to long inference times, making it difficult to meet the timeliness requirements of online real-time monitoring.

[0004] In summary, the current field of PMSM fault diagnosis urgently needs a lightweight early fault detection method that is suitable for industrial online monitoring scenarios, without relying on massive fault samples or complex modeling and hyperparameter selection, and that can achieve this by monitoring the statistical distribution drift of the motor's operating status online.

[0005] A patent document published on December 12, 2025, with publication number CN121114757A, discloses a method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors based on high-frequency negative-sequence components. The steps are as follows: A high-frequency voltage injection signal is obtained based on the three-phase current of the motor; a high-frequency current signal is obtained through space vector pulse width modulation and injected into the motor to obtain the current three-phase voltage and current; the high-frequency components are extracted; high-frequency negative-sequence voltage and current are obtained based on the extracted high-frequency components, and fault features are constructed; the fault features are compared with a significance level threshold. If the fault features are greater than the significance level threshold, the permanent magnet synchronous motor is judged to be healthy; if the fault features are less than or equal to the significance level threshold, the permanent magnet synchronous motor is judged to be faulty. The flow of the above method for diagnosing inter-turn short-circuit faults in permanent magnet synchronous motors is as follows: Figure 3 As shown, the main drawbacks of this technical solution are: First, when using a second-order generalized integrator to extract the frequency response (including the high-frequency components in the two-phase voltage and the two-phase current of the motor), its parameters need to be precisely tuned. Fixing the parameters can easily lead to distortion in the extraction of high-frequency components, introducing amplitude / phase errors, and requires a large amount of computing power. Second, when obtaining the phase of the high-frequency negative-sequence current and the phase of the high-frequency phase voltage, the phase calculation method is complex, and the extraction of the phase of the high-frequency phase voltage and negative-sequence current is easily affected by inverter nonlinearity, sampling noise, and sensor accuracy, leading to incorrect phase difference calculations and consequently misjudging the faulty phase. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method for detecting early inter-turn short-circuit faults in permanent magnet synchronous motors (PMSMs) based on a multi-scale sliding window. This method accurately identifies early inter-turn short-circuit faults in PMSMs while simultaneously locating the faulty phase. It solves the problems of complex modeling required for fault feature extraction and the reliance on large-scale sample training for hyperparameter selection and fault detection. This method enables online real-time fault monitoring under low computing power conditions and is particularly suitable for motor health diagnosis scenarios in complex industrial environments.

[0007] To address the aforementioned technical problems, this invention proposes an early-stage inter-turn short-circuit fault detection method for permanent magnet synchronous motors based on a multi-scale sliding window. The core of this method lies in constructing a multi-scale sliding window joint decision mechanism oriented towards motor state evolution: firstly, a multi-scale sliding window dataset is constructed under healthy motor operating conditions; then, real-time operating data is acquired through the sliding window mechanism, and residual datasets at different time scales are constructed; further, statistical significance analysis methods (t-test and p-value calculation) are used to quantify the degree of distribution drift of the residual signal; finally, fault detection and location are achieved through the multi-scale sliding window joint decision mechanism. Simultaneously, this invention introduces an adaptive update mechanism for the benchmark dataset, enabling the system to dynamically correct the reference model based on long-term healthy operating data, thereby effectively suppressing the impact of operating condition drift, load changes, and temperature rise effects on fixed benchmark data, and improving the stability and robustness of long-term online monitoring. The method includes the following steps:

[0008] Step 1: Use Hall current sensors to collect the instantaneous values ​​of the three-phase current of the stator of the motor under test in real time throughout the entire operation process;

[0009] Step 2: After the motor under test starts up in a healthy state and runs to a steady state, construct two multi-scale sliding windows with different time resolutions. Under the premise of ensuring phase synchronization of sliding windows of the same scale, sample the current signal in segments to form a multi-scale sliding window benchmark dataset. Among them, sliding windows with different time resolutions are used to characterize the signal statistical characteristics at different time scales.

[0010] Step 3: During motor operation, the real-time current signal is synchronously segmented using the multi-scale sliding window corresponding to Step 2, and the multi-scale sliding window datasets are obtained respectively. The difference calculation is performed between the datasets and their respective benchmark datasets to construct the residual datasets under the multi-scale sliding window.

[0011] Step 4: Perform statistical significance analysis on the residual dataset and the residual benchmark dataset at different time scales, and calculate the corresponding statistical test quantity and significance probability value to characterize the distribution shift of the real-time signal relative to the residual benchmark dataset.

[0012] Step 5: Construct a multi-scale sliding window joint decision mechanism: Compare the most recent effective statistical results of each scale with their corresponding significance level thresholds. If the most recent effective statistical results of each scale are not higher than their corresponding significance level thresholds, it is determined that an early inter-turn short circuit fault has occurred in the current motor phase; otherwise, it is determined to be in a healthy state.

[0013] The method and steps of this invention are followed to simultaneously detect inter-turn short-circuit faults of phases A, B, and C with different fault degrees, thereby improving the accuracy of motor fault detection.

[0014] Furthermore, in the method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to the present invention, wherein:

[0015] In step 1, the Hall current sensor has a current rating of 500A, a rated operating voltage of ±15V, and a maximum supply voltage of ±16.5V.

[0016] In step 2, the multi-scale sliding window includes a short-timescale sliding window and a long-timescale sliding window. The short-timescale sliding window is used to quickly detect transient statistical anomalies within a short period of time, ensuring the capture of the spike pulse at the moment of the fault. The long-timescale sliding window is designed to cover at least one complete short-timescale sliding window, suppressing misjudgments caused by random noise and operating condition fluctuations, while simultaneously assessing the steady-state impact of the fault on the current. Different timescale sliding windows employ different window lengths and update cycles. The window scales of the short-timescale sliding window and the long-timescale sliding window are w1 and w2, respectively, and their benchmark datasets are named Set. s_0 Set m_0 And store them in databases C1 and C2 respectively.

[0017] In step 3, the sliding window adopts an overlapping sliding method. During the window sliding process, the sliding step sizes of the short-timescale sliding window and the long-timescale sliding window are set to step1 and step2, respectively, to achieve dynamic update sampling of the continuous-time signal. The windows slide along the positive direction of the time axis, and the signal intercepted by the short-timescale sliding window generates a dataset Set. s_i Datasets generated from signals captured by a long-term sliding window m_j Where i = 1, 2, ..., e, j = 1, 2, ..., f, and are temporarily stored in databases C1 and C2 respectively. When the motor stops running, if the data volume of the last window of the short-timescale sliding window is less than w1, and the data volume of the last window of the long-timescale sliding window is less than w2, then the data included in that window is discarded. The dataset Set... s_0 With Set s_i Set m_0 With Set m_j By subtracting the values ​​from each, we obtain the residual benchmark dataset Set, which consists of the residual signals. sf_0 Set mf_0 and residual dataset Set sf_x Set mf_y Where x=1,2,..,e-1, y=1,2,..,f-1, and all residual datasets corresponding to the short-time-scale sliding window and the long-time-scale sliding window are temporarily stored in databases CF1 and CF2 respectively.

[0018] In step 4, the statistical significance analysis method preferably employs the t-test method. This is used for Set... sf_0 With Set sf_x Set mf_0 With Set mf_y The t-test is used to obtain the t-statistic, and the significance probability value p is output. s p m , where p s p m They are used to characterize Set sf_0 With Set sf_x Set mf_0 With Set mf_y The degree of difference in signal distribution between turns. Because inter-turn short circuits cause amplitude bias and waveform distortion in the phase current at a specific phase, this distortion manifests as a sudden change in the mean within a short-timescale sliding window and an increase in the standard deviation within a long-timescale sliding window. Therefore, the p-value calculated by the t-test essentially quantifies the cumulative effect of this physical distortion.

[0019] In step 5, the significance level thresholds for the short-timescale sliding window and the long-timescale sliding window are α, respectively. sf α mf At the same time, set p s Updated every T1 time interval, p m Update the value every time interval T2, and output the most recent significance probability value p every time interval T2. s_last p m_last ; will p s_last p m_last Significance level thresholds α for sliding windows at different time scales sf α mf Comparison, if p is satisfied simultaneously s_last ≤α sf p m_last ≤α mf If the fault is found, it is determined that a short circuit between turns has occurred in that phase of the motor; otherwise, it is determined that no short circuit between turns has occurred in that phase of the motor.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] (1) The present invention uses the p value to characterize the difference between fault characteristics and normal operating conditions, ensuring high sensitivity of inter-turn short circuit fault detection. It realizes fault location while performing fault detection, without the need for additional calculation of high frequency phase difference, which greatly saves computing resources.

[0022] (2) It abandons the cumbersome model training and hyperparameter tuning process of traditional deep learning, and does not require complex time-frequency domain transformation. It directly uses a lightweight statistical significance analysis method on the time-domain residual signal, gets rid of the dependence on large-scale samples, and saves computing resources.

[0023] (3) The present invention further constructs a dynamic health reference evolution mechanism. During the long-term operation of the motor, the benchmark dataset is adaptively updated based on the statistical significance results within the continuous time window, thereby suppressing false alarms caused by load fluctuations, temperature rise drift and changes in operating conditions, and improving the reliability of long-term online monitoring in complex industrial environments.

[0024] (4) The present invention uses sliding windows of different time scales to construct fault features and performs joint statistical significance analysis, thereby taking into account both the sensitivity of weak fault detection and the reliability of long-term operation data detection and judgment. Attached Figure Description

[0025] Figure 1 This is the equivalent model of the inter-turn short-circuit fault method for permanent magnet synchronous motors of the present invention;

[0026] Figure 2 This is a flowchart of the inter-turn short-circuit fault detection process for permanent magnet synchronous motors according to the present invention;

[0027] Figure 3 This is a flowchart of a method for diagnosing inter-turn short-circuit faults in a permanent magnet synchronous motor in the prior art;

[0028] Figure 4 This is a schematic diagram of the sliding window algorithm in the method of this invention;

[0029] Figure 5 These are residual signals under different fault levels of the motor, where: (a) is the residual signal waveform diagram with an inter-turn short-circuit ratio of 0.05, (b) is the residual signal waveform diagram with an inter-turn short-circuit ratio of 0.08, (c) is the residual signal waveform diagram with an inter-turn short-circuit ratio of 0.1, and (d) is the residual signal waveform diagram with an inter-turn short-circuit ratio of 0.15.

[0030] Figure 6 This is a partial current waveform diagram of phase A of the motor when it is operating in a steady state under healthy conditions. Detailed Implementation

[0031] The design concept of this invention, a method for detecting early inter-turn short-circuit faults in permanent magnet synchronous motors based on a multi-scale sliding window, is as follows: Hall current sensors are used to sample the instantaneous current values ​​of each phase of the motor, obtaining the instantaneous phase current values ​​at any moment during motor operation. A multi-scale sliding window and a t-test are applied together to detect and locate the faulty phases in the early stages of an inter-turn short circuit. The multi-scale sliding window is used to truncate the dataset, discarding data when the amount of data at the end is insufficient, thus avoiding detection errors caused by insufficient data and ensuring detection accuracy. The residual signal is used as a fault feature, solving the problems of complex modeling and hyperparameter selection required for fault feature extraction, highlighting the differences in current characteristics caused by the fault, and reducing the interference of external noise on the accuracy of fault detection. A significance level threshold is set, and the dataset differences of the residual signal are quantified by combining the t-test and p-value method. This achieves fault phase localization while performing fault detection, and also solves the problem of fault detection relying on large-scale sample training.

[0032] To fully understand this invention, the principles of the t-test and p-value methods mentioned herein will be described in detail.

[0033] The specific expression for the t-test is:

[0034]

[0035] In this invention, X gj X 0j D represents the data at corresponding positions in the residual dataset and the residual benchmark dataset, both at the same scale. j This represents the difference between corresponding data in these two datasets. S represents the mean of n differences. D The standard deviation is represented by T, and the t-statistic is represented by T.

[0036] After obtaining the t-statistic, the corresponding p-value can be calculated. The p-value method can then be used to quantify the difference between residual datasets and residual benchmark datasets of the same scale. The equation for calculating the p-value is:

[0037]

[0038] Where Γ(z) is the gamma function, t is the random variable of the t-statistic, and f t (t;v) is the probability density function of the t-distribution, F t Let df = n-1 represent the cumulative distribution function, where p ∈ [0, 1].

[0039] This invention employs statistical significance analysis to overcome the inherent problems of both traditional signal processing methods and data-driven fault diagnosis methods. On the one hand, it avoids the drawback of manually selecting hyperparameters during fault feature extraction in signal processing methods; on the other hand, it overcomes the limitation of data-driven diagnostic models relying on large-scale labeled samples for training. In engineering scenarios where real fault samples are scarce, the proposed method can effectively detect early-stage inter-turn short-circuit faults in motors without the need for large-scale labeled sample training, while ensuring the accuracy and reliability of the fault detection results.

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.

[0041] Based on the above principles, this invention proposes a method for detecting early-stage inter-turn short-circuit faults in permanent magnet synchronous motors based on multi-scale sliding window statistical significance analysis. The detection process is as follows: Figure 2 As shown in Table 1, the basic parameters of the motor are as follows.

[0042] Table 1 Basic parameters of permanent magnet synchronous motor

[0043]

[0044] In this embodiment, the Hall current sensor has a current rating of 500A, a rated operating voltage of ±15V, and a maximum supply voltage of ±16.5V.

[0045] In this embodiment, the sampling frequency is configured to 10kHz. The three-phase permanent magnet synchronous motor is set to operate stably at a speed of 2500 r / min and a torque of 100 Nm. At 2.5 seconds, a sudden inter-turn short-circuit fault is simulated in phase A of the motor. To demonstrate the effectiveness of this method, simulations of inter-turn short-circuit faults with inter-turn short-circuit ratios of 0.05, 0.08, 0.1, and 0.15 are performed. The specific steps are as follows:

[0046] Step 1: Use Hall current sensors to collect the instantaneous values ​​of the A, B, and C phase currents of the stator in real time throughout the entire operation of the motor.

[0047] Step 2: After the motor under test starts up in a healthy state and runs to a steady state, two sliding windows with different time scales are used. Taking phase A as an example, Figure 1 An equivalent model of a three-phase permanent magnet synchronous motor with phase A being the phase experiencing an inter-turn short-circuit fault is shown. Figure 1 As can be seen from the diagram, this circuit uses the common node o as the reference neutral point. Phase A is the phase with an inter-turn short-circuit fault, while phases B and C are the healthy phases of the motor. This includes the resistance R of the healthy windings in phases B and C. b R c With inductor L b Lc The short-circuit resistance R of the faulty winding in phase A f and short-circuit current i f R a1 L a1 The resistance and inductance of the A-phase winding that is not faulty, R a2 L a2 The model represents the resistance and inductance of the winding where the fault occurred in phase A. Additionally, the model includes the mutual inductance between the windings, denoted as M. a1a2 M a2b M a1b M a2c M a1c . Figure 6 The figure shows the partial current waveform of phase A of a three-phase permanent magnet synchronous motor under healthy conditions during steady-state operation. Figure 6 As can be seen, the intercepted period is 1.6-1.8s. During this period, the motor has entered a steady-state operation and no inter-turn short circuit fault has occurred. The current of phase A shows periodic sinusoidal fluctuations with an amplitude stable within the range of -50-50A.

[0048] Before phase A is operating stably and no inter-turn short circuit occurs, and while ensuring phase synchronization within the same time-scale sliding window, a set of benchmark datasets is extracted from both the short-time-scale and long-time-scale sliding windows, named Set. s_0 Set m_0 And store them in databases C1 and C2 respectively.

[0049] Step 3: During motor operation, a short-timescale sliding window and a long-timescale sliding window synchronously segment the real-time current signal. Both windows of different time scales continuously slide along the positive direction of the time axis using an overlapping sliding method. The signal intercepted by the short-timescale sliding window generates a dataset Set. s_i Datasets generated from signals captured by a long-term sliding window m_j Where i = 1, 2, ..., e, j = 1, 2, ..., f, and these values ​​are temporarily stored in databases C1 and C2 respectively. Figure 4The schematic diagram of the sliding window algorithm is shown. To ensure rapid capture of instantaneous anomalies by the short-timescale sliding window, the time resolution is set to 100ms. To optimize the balance between computational power consumption and memory usage, and to update the effective statistical results of the long-timescale sliding window with high frequency and accuracy, the time resolution is set to 200ms, i.e., window sizes w1 and w2 are set to 1000 and 2000 respectively. During the window sliding process, it is necessary to ensure smooth data transition and continuity while reducing computational redundancy and overhead, and to enhance the robustness of statistical features. Therefore, the sliding step size (step1) and step2 for both the short-timescale and long-timescale sliding windows are set to 500. Furthermore, when the motor stops running, if the data volume of the last window of the short-timescale sliding window is less than 1000, and the data volume of the last window of the long-timescale sliding window is less than 2000, the data included in that window is discarded to avoid interference with the accuracy of the detection results due to insufficient data, thus ensuring the accuracy of the detection.

[0050] Set the dataset s_i With Set s_0 Set m_j With Set m_0 By subtracting the values ​​from each, we obtain the residual benchmark dataset Set, which consists of the residual signals. sf_0 Set mf_0 and residual dataset Set sf_x Set mf_y Where x=1,2,..,e-1, y=1,2,..,f-1, and all residual datasets corresponding to the short-time-scale sliding window and the long-time-scale sliding window are temporarily stored in databases CF1 and CF2 respectively.

[0051] Step 4, respectively for Set sf_0 With Set sf_x Set mf_0 With Set mf_y The t-test is used to obtain the t-statistic, and the significance probability value p is output. s p m , where p s p m They are used to characterize Set sf_0 With Set sf_x Set mf_0 With Set mf_yThe degree of difference in signal distribution between turns. Because inter-turn short circuits cause amplitude bias and waveform distortion in the phase current at a specific phase, this distortion manifests as a sudden change in the mean within a short-timescale sliding window and an increase in the standard deviation within a long-timescale sliding window. Therefore, the p-value calculated by the t-test essentially quantifies the cumulative effect of this physical distortion.

[0052] Step 5: The significance level thresholds for the short-timescale sliding window and the long-timescale sliding window are α, respectively. sf α mf If the significance level threshold is higher than 0.5, any minute sensor noise, electromagnetic interference, or normal load fluctuations will trigger a short-window alarm, leading to frequent false alarms. This not only wastes computational resources of the long-term sliding window but also causes the final joint decision logic to fail. To ensure high robustness while retaining sensitivity to weak fault characteristics such as early inter-turn short-circuit faults, the lower limit of the significance level threshold is set at 0.05. Therefore, α sf α mf The value range is 0.05 to 0.5. Therefore, we set α... sf It is 0.5, α mf It is 0.05.

[0053] Constructing a multi-scale sliding window joint decision mechanism: The p-value method is used to quantify the degree of difference between various residual signal datasets. The p-values ​​of the residual signals at different scales of a certain phase of the motor are respectively compared with α. sf α mf A comparison was made one by one. To balance the real-time performance and computational efficiency of short-timescale sliding windows, and to avoid increased detection latency and masking short-term signal abrupt changes while evaluating steady-state statistical characteristics in long-timescale sliding windows, and to ensure alignment between the last data point of each short-timescale sliding window and the last data point of each long-timescale sliding window while maintaining high update efficiency, a p-value was set for the first three... s Updated every 150ms, the first time p m Update after 450ms, that is, output the most recent valid statistical result p after 450ms. s_last p m_last ; followed by p s With p m Updated every 150ms, meaning the most recent valid statistical result p is output every 150ms. s_last p m_last If p is satisfied at the same time s_last ≤0.5, p m_last If the value is ≤0.05, then the phase is determined to have an inter-turn short circuit fault; otherwise, the phase of the motor is determined not to have an inter-turn short circuit fault.

[0054] This invention constructs a dynamic health reference evolution mechanism and a high-confidence health interval. During continuous motor operation, if the effective statistical results of the sliding window for multiple consecutive cycles of the short-scale and long-scale windows remain within the high-confidence health interval, the system is determined to be in a fault-free, healthy, and stable interval. The residual signal within the corresponding time interval is then updated to the residual benchmark dataset to achieve adaptive evolution and update of the residual benchmark dataset. If the p-value shows a slow decreasing trend but does not reach the significance level threshold, benchmark updates are paused, and potential degradation trends are recorded to prevent early weak fault characteristics from being masked by the benchmark update mechanism. To avoid masking early signs of motor degradation and to prevent the high-confidence health interval from being too narrow and thus affecting the dynamic health reference evolution mechanism, in this embodiment, the high-confidence health interval is set to [0.8-1]. During continuous motor operation, when p... s p m When all values ​​are within the range of [0.8-1], the system is considered to be in a healthy and stable range without faults. Therefore, Set... sf_0 Set mf_0 The data in the dataset is updated to the corresponding new steady-state window data to achieve adaptive evolution update of the residual benchmark dataset; if p m If a slow downward trend is observed but the significance level threshold is not reached, baseline updates are paused and potential degradation trends are recorded to prevent early, subtle fault characteristics from being masked by the baseline update mechanism. When a statistically significant anomaly occurs at any scale, healthy baseline updates are stopped and fault detection mode is entered.

[0055] By processing the phase current of all phases of the motor according to the above detection procedure, the inter-turn short circuit fault can be located.

[0056] In this embodiment, data from time periods of 1.7s-1.8s and 1.6s-1.8s are used as the baseline datasets. A multi-scale sliding window is then used to extract data, and residual signals are obtained by subtracting the data from the time periods of 2.1s-2.2s and 2.0s-2.2s. This case uses data from time periods of 2.1s-2.2s, 2.3s-2.4s, 2.6s-2.7s, 2.0s-2.2s, 2.2s-2.4s, and 2.5s-2.7s as examples, denoted as Dataset Set. sf_0 Set sf_1 Set sf_2 Set mf_0 Set mf_1 Set mf_2 The t-test and p-value method were used to analyze the Set. sf_0 With Set sf_1 Set sf_0 With Set sf_2 Setmf_0 With Set mf_1 Set mf_0 With Set mf_2 Process them separately to obtain the corresponding p-values. s0_1 p s0_2 p m0_1 p m0_2 The results are shown in Table 2, which shows the original signals and the signals after subtraction for different fault degrees. Figure 5 As shown.

[0057] Table 2. t-test results for different fault levels of the motor.

[0058]

[0059] Figure 5 The residual signals of the motor under different fault levels are shown. Table 2 shows the Set signals under different window sizes and different fault levels. sf_0 With Set sf_1 Set sf_0 With Set sf_2 Set mf_0 With Set mf_1 Set mf_0 With Set mf_2 The corresponding p value p s0_1 p s0_2 p m0_1 p m0_2 .Depend on Figure 5 It can be seen that, Figure 5 In the diagram, (a), (b), (c), and (d) all select the time period from 1.8 to 3.0 seconds, corresponding to the residual signals under fault severity levels of 0.05, 0.08, 0.1, and 0.15, respectively. Since the motor experiences an inter-turn short-circuit fault at 2.5 seconds, the 1.8-2.5 seconds period represents the residual signal under healthy motor operation, while the period after 2.5 seconds represents the residual signal under the inter-turn short-circuit fault condition. Before the fault occurs at 2.5 seconds, the motor remains in a healthy operating condition, and Set... sf_0 With Set sf_1 Set mf_0 With Set mf_1 The signal differences between them are relatively small, so the p values ​​corresponding to each fault level in Table 2 are... s0_1 p m0_1 Both are greater than 0.5 and 0.05 respectively, and their values ​​are similar within the same scale. After a 2.5s fault trigger, the motor enters an inter-turn short-circuit operating state. When the fault severity is 0.05, 0.08, 0.1, and 0.15, the forward bias of the fault current reaches 3.5A, 4.4A, 7.1A, and 8.5A respectively. It can be seen that as the fault severity increases, the forward bias of the current at the moment of fault occurrence gradually increases.sf_0 With Set sf_2 Set mf_0 With Set mf_2 The signal differences between them also increase accordingly. Correspondingly, the p values ​​under each fault level in Table 2... s0_2 p m0_2 The values ​​gradually decrease as the fault severity increases, and are all less than 0.5 and 0.05, respectively. This verifies the feasibility of the early-stage inter-turn short-circuit fault detection method for permanent magnet synchronous motors proposed in this invention.

[0060] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are preferred application examples that demonstrate the core technical ideas of the present invention, and are merely illustrative and not restrictive. Those skilled in the art can make many improvements and changes under the guidance of the present invention without departing from the spirit of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for detecting early inter-turn short-circuit faults in permanent magnet synchronous motors based on a multi-scale sliding window, characterized in that, Includes the following steps: Step 1: Use Hall current sensors to collect the instantaneous values ​​of the three-phase current of the stator of the motor under test in real time throughout the entire operation process; Step 2: Construct two sliding windows with different time resolutions. The two sliding windows with different time resolutions include a short time scale sliding window and a long time scale sliding window. The long time scale sliding window covers 1 to 50 short time scale sliding windows. After the motor under test starts up in a healthy state and runs to a steady state, the current signal is sampled in segments to form a reference dataset for a short-time-scale sliding window and a long-time-scale sliding window, provided that the phase of the sliding window of the same scale is synchronized. The sliding window with different time resolution is used to characterize the signal statistical characteristics at different time scales. Step 3: During motor operation, the short-time-scale sliding window and the long-time-scale sliding window synchronously segment the real-time current signal to obtain the datasets of the short-time-scale sliding window and the long-time-scale sliding window respectively, and perform difference calculation with their respective corresponding benchmark datasets to construct the residual dataset and the residual benchmark dataset under the multi-scale sliding window. Step 4: Perform statistical significance analysis on the residual dataset and the residual benchmark dataset at different time scales, and calculate the corresponding statistical test quantity and significance probability value to characterize the distribution shift of the real-time signal relative to the residual benchmark dataset. Step 5: Construct a multi-scale sliding window joint decision mechanism: Compare the most recent effective statistical results of each scale with their corresponding set significance level thresholds. If the most recent effective statistical results of each scale are not higher than their set significance level thresholds, it is determined that the current motor phase has an early inter-turn short circuit fault; otherwise, it is determined that the current motor phase is in a healthy state. Step 6: Following steps 1 to 5, simultaneously complete the inter-turn short circuit fault detection for the three phases A, B, and C of the motor.

2. The method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 1, characterized in that, In step 1, the Hall current sensor has a current rating of 500A, a rated operating voltage of ±15V, and a maximum supply voltage of ±16.5V.

3. The method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 1, characterized in that, In step 2, the short-timescale sliding window is used to quickly detect transient statistical anomalies within a short period of time, ensuring the capture of the spike pulse at the moment of the fault; the long-timescale sliding window is used to suppress misjudgments caused by random noise and operating condition fluctuations, while simultaneously evaluating the steady-state impact of the fault on the current; wherein, the window scales of the short-timescale sliding window and the long-timescale sliding window are denoted as w1 and w2, respectively, and the benchmark datasets for the short-timescale sliding window and the long-timescale sliding window are denoted as Set, respectively. s_0 Set m_0 The benchmark dataset Set s_0 and Set m_0 Stored in databases C1 and C2 respectively.

4. The method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 3, characterized in that, In step 3, two sliding windows of different scales are used in an overlapping sliding manner. During the window sliding process, both windows slide along the positive direction of the time axis. The sliding step sizes of the short-time-scale sliding window and the long-time-scale sliding window are step1 and step2, respectively, to achieve dynamic update sampling of the continuous-time signal. The signal intercepted by the short-time-scale sliding window generates a dataset Set. s_i The dataset Set generated by the signal captured by the long-term sliding window. m_j Where i = 1, 2, ..., e, j = 1, 2, ..., f, the dataset Set s_i and Set m_j The data is temporarily stored in databases C1 and C2 respectively. When the motor stops running, if the short-timescale sliding window or the long-timescale sliding window slides to the last window and the amount of data in it is insufficient for the window size, then the data included in that window is discarded; the benchmark dataset Set is then used. s_0 With dataset Set s_i By subtracting the values, we obtain the residual benchmark dataset Set, which consists of the residual signals. sf_0 and residual dataset Set sf_x Set the benchmark dataset m_0 With dataset Set m_j By subtracting the values, we obtain the residual benchmark dataset Set, which consists of the residual signals. mf_0 and residual dataset Set mf_y Where x=1,2,..,e-1, y=1,2,..,f-1; the residual benchmark dataset and residual dataset corresponding to the short-timescale sliding window and the long-timescale sliding window are temporarily stored in databases CF1 and CF2, respectively.

5. The method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 4, characterized in that, In step 4, the statistical significance analysis uses the t-test method; For residual datasets and residual benchmark datasets at short time scales, the following residual benchmark dataset is obtained: Set sf_0 With residual dataset Set sf_x t-statistic and significance probability value p s The significance probability value p s Set used to characterize residual benchmark datasets sf_0 With residual dataset Set sf_x The degree of difference in signal distribution between them, where p s The range of values ​​for is the closed interval [0,1]. For the residual dataset and residual benchmark dataset over a long time scale, we obtain: Residual Benchmark Dataset Set mf_0 With residual dataset Set mf_y t-statistic and significance probability value p m The significance probability value p m Set used to characterize residual benchmark datasets mf_0 With residual dataset Set mf_y The degree of difference in signal distribution between them, where p m The range of values ​​is a closed interval [0,1].

6. The method for detecting early inter-turn short-circuit faults in a permanent magnet synchronous motor according to claim 1, characterized in that, In step 5, the significance level thresholds for the short-timescale sliding window and the long-timescale sliding window are α, respectively. sf and α mf Meanwhile, the significance probability value p s The significance probability value p is updated every time interval T1. m Update the value every time interval T2, and output the significance probability p of the most recent short-timescale sliding window every time interval T2. s_last The significance probability value p of the long-term sliding window m_last The significance probability value p s_last and p m_last Compared with the significance level threshold α respectively sf and α mf Compare; if p is satisfied simultaneously s_last ≤α sf p m_last ≤α mf If the fault is found, the motor phase is determined to have an inter-turn short circuit fault; otherwise, the motor phase is determined to be in a healthy state.

Images