Motor predictive maintenance system and method based on digital twinning

Abstract

The application relates to the technical field of motor state monitoring, and particularly discloses a motor predictive maintenance system and method based on digital twinning, which collects three-phase currents, bus voltages, rotor positions and winding end temperatures of a motor; according to the actually measured temperature and historical loss calculated from the currents, fractional calculus is used to reverse the order and heat capacity parameters, and a power law weighted accumulation process of the historical loss is formed; the historical loss is input into the process, the transient temperature distribution of each virtual position point in the winding is calculated, and a dynamic thermal reference containing a memory effect is obtained; the dynamic thermal reference is compared with the actually measured temperature at each time point to obtain a temperature residual sequence, and a sensitivity attenuation index is solved; a current harmonic residual sequence is extracted, is fused with the temperature residual sequence to obtain a fusion characteristic value, and when the fusion characteristic value exceeds a threshold and the sensitivity attenuation index is lower than a threshold value, an early diagnosis result of inter-turn short circuit is output; the application realizes accurate early diagnosis of the inter-turn short circuit under high dynamic conditions.

Description

Technical Field

[0001] This invention relates to the field of motor condition monitoring technology, and specifically to a predictive maintenance system and method for motors based on digital twins. Background Technology

[0002] Electric motors are core actuators in industrial drives, transportation, and energy equipment, and their operational reliability directly affects the safety and economy of the entire system. Inter-turn short circuits, one of the most common early faults in motor stator windings, are typically caused by insulation aging, overvoltage, or accumulated thermal stress. Early diagnosis before such faults evolve into phase-to-phase short circuits or ground faults can significantly reduce maintenance costs and downtime losses.

[0003] Under high dynamic conditions, traditional integer-order thermal network models cannot characterize the fractional-order relaxation characteristics and historical loss memory effect of motor winding heat between copper conductors, insulation layers and iron cores. As a result, the residual between the dynamic thermal reference and the measured transient temperature cannot distinguish between the persistent thermal anomaly caused by inter-turn short circuit and the transient temperature fluctuation caused by load transients, making it difficult to achieve accurate early diagnosis in the stage of slight carbonization of inter-turn insulation. Summary of the Invention

[0004] The purpose of this invention is to provide a predictive maintenance system and method for motors based on digital twins to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A predictive maintenance method for motors based on digital twins includes the following steps: S1: Collect the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends; S2: Based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform, a thermal diffusion characterization is constructed using fractional calculus, and the fractional order and heat capacity parameters are deduced to form a power-law weighted accumulation process of historical loss. S3: Input the historical loss sequence into the power-law weighted accumulation process to calculate the transient temperature distribution of each virtual position point inside the winding in real time, and obtain a dynamic thermal reference with memory effect; S4: Compare the dynamic thermal reference with the measured transient temperature sequence point by point to obtain the temperature residual sequence, and calculate the sensitivity decay index of the temperature residual sequence to the historical loss sequence. S5: Extract the harmonic residual sequence from the three-phase current time-domain waveform, fuse the temperature residual sequence with the harmonic residual sequence, and distinguish between transient load fluctuations and persistent thermal anomalies based on the sensitivity attenuation index, and output the early diagnosis results of inter-turn short circuits.

[0006] As a further aspect of the present invention: S2 specifically includes: Extract the temperature rise response segment after step excitation from the measured transient temperature sequence; The predicted temperature rise is obtained by performing a discrete convolution of the historical loss sequence in reverse chronological order with a power-law weighting function. With the goal of predicting the temperature rise to approximate the temperature rise response segment, the fractional order and heat capacity parameters are iteratively adjusted. Substitute the converged fractional order and the heat capacity parameter into the power-law weighted accumulation process.

[0007] As a further aspect of the present invention: the construction and use of the power-law weighting function are specifically as follows: Extract a predetermined number of loss values ​​prior to the current moment from the historical loss sequence to form a loss vector; The predicted temperature rise is obtained by multiplying the loss vector point by point with the initial power-law weight vector and then summing the results. Calculate the temperature rise difference between the predicted temperature rise and the temperature rise difference at the corresponding moment in the measured transient temperature sequence, and correct the decay exponent of each element in the power law weight vector in reverse based on the corresponding difference. The modified power-law weight vector is used as the power-law weight function at the current moment.

[0008] As a further aspect of the present invention: S3 specifically includes: The interior of the winding is divided radially into multiple virtual location points, and each virtual location point corresponds to a thermal diffusion depth. Based on the thermal diffusion depth and heat source distance of each virtual location, different power-law weighting coefficients are assigned to the same historical loss sequence. The historical loss sequence is input into the power-law weighted accumulation process of each virtual location point, and the transient temperature value of each point is calculated in parallel. The transient temperature values ​​of all virtual location points are combined into a dynamic thermal reference with a memory effect.

[0009] As a further aspect of the present invention: the step of combining the transient temperature values ​​of all virtual location points into a dynamic thermal reference containing a memory effect specifically includes: Calculate the heat propagation delay time between each virtual location point and the measured temperature point at the end of the winding. The delay time is determined by both the heat diffusion depth and the fractional order. Based on the heat propagation delay time, the transient temperature value sequence of each virtual location point is shifted backward in time so that each sequence is aligned with the measured transient temperature sequence in the time dimension; The aligned sequences are weighted and summed according to the reciprocal of the thermal diffusion depth of their respective virtual locations to obtain the fusion temperature sequence. The fused temperature sequence is used as a dynamic thermal reference with memory effect.

[0010] As a further aspect of the present invention: S4 specifically includes: The transient sensitivity sequence is obtained by dividing the residual value at each time point in the temperature residual sequence by the historical loss value at the corresponding time point. The logarithmic sensitivity sequence is obtained by taking the natural logarithm of the transient sensitivity sequence in chronological order. Linear fitting is performed on the logarithmic sensitivity sequence, and the decreasing slope of the fitted line is extracted; The opposite value of the descent slope is used as the sensitivity decay index, which is used to quantify the rate at which the temperature residual decays as the historical loss time shifts.

[0011] As a further aspect of the present invention: the extraction of the descending slope of the fitted straight line specifically includes: The effective attenuation range is obtained by extracting the time interval in which the temperature residual is continuously positive from the logarithmic sensitivity sequence, removing the initial transient fluctuation segment and the final noise segment; The logarithmic sensitivity values ​​at each time point within the effective decay interval are arranged in chronological order, and each time point is assigned a confidence weight proportional to the absolute value of the residual. We use the weighted least squares method to fit a straight line, minimizing the weighted sum of squared residuals; The slope of the fitted line is output as the descending slope.

[0012] As a further aspect of the present invention: S5 specifically includes: The fundamental component is subtracted from the three-phase current time-domain waveform periodically to obtain the harmonic component sequence of each phase. The three-phase harmonic residual sequence is obtained by subtracting the harmonic component sequence of each phase from the corresponding phase health status reference harmonic sequence at each time point. The temperature residual sequence and the three-phase harmonic residual sequence are multiplied point by point along the same time coordinate and then summed to obtain the fused characteristic value; When the fused feature value exceeds the preset threshold and the sensitivity attenuation index is lower than the attenuation threshold, the early diagnosis result of inter-turn short circuit is output.

[0013] A predictive maintenance system for motors based on digital twins includes: The data acquisition module collects the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends. The fractional-order parameter inversion module uses fractional-order calculus to construct a thermal diffusion characterization based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform. It then inverts the fractional-order order and heat capacity parameters, forming a power-law weighted accumulation process of historical losses. The dynamic thermal reference generation module inputs the historical loss sequence into the power-law weighted accumulation process, calculates the transient temperature distribution of each virtual position point inside the winding in real time, and obtains a dynamic thermal reference with memory effect. The temperature residual and sensitivity calculation module compares the dynamic thermal reference with the measured transient temperature sequence point by point to obtain the temperature residual sequence, and calculates the sensitivity decay index of the temperature residual sequence to the historical loss sequence. The integrated diagnostic module extracts the harmonic residual sequence from the three-phase current time-domain waveform, merges the temperature residual sequence with the harmonic residual sequence, and distinguishes between transient load fluctuations and persistent thermal anomalies based on the sensitivity decay index, outputting early diagnosis results for inter-turn short circuits.

[0014] The beneficial effects of this invention are: (1) This invention constructs a thermal diffusion characterization based on fractional calculus to form a power-law weighted accumulation process of historical losses, which can capture the memory effect of heat transfer in motors under dynamic conditions such as frequent start-stop, rapid acceleration and regenerative braking, and obtain a dynamic thermal reference containing the memory effect. On this basis, the temperature residual sequence and the three-phase current harmonic residual sequence are fused to obtain fused characteristic values, and the continuous thermal anomaly and transient load fluctuation are distinguished according to the sensitivity decay index, so that early diagnosis results can be output at the stage of slight carbonization of inter-turn insulation, effectively reducing the false alarm rate.

[0015] (2) This invention divides multiple virtual location points along the radial direction, calculates the transient temperature value of each point, and combines the heat propagation delay time for time-reverse translation and weighted fusion to form a dynamic thermal reference. This reference does not require additional sensors to be embedded inside the motor, and can reflect the heat diffusion process from the surface of the copper conductor to different depths of the iron core yoke, providing multi-dimensional temperature information support for fault location, while avoiding the impact of built-in sensors on the original structure and reliability of the motor. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a predictive maintenance method for motors based on digital twins, comprising the following steps: S1: Collect the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends; S2: Based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform, a thermal diffusion characterization is constructed using fractional calculus, and the fractional order and heat capacity parameters are deduced to form a power-law weighted accumulation process of historical loss. S3: Input the historical loss sequence into the power-law weighted accumulation process to calculate the transient temperature distribution of each virtual position point inside the winding in real time, and obtain a dynamic thermal reference with memory effect; S4: Compare the dynamic thermal reference with the measured transient temperature sequence point by point to obtain the temperature residual sequence, and calculate the sensitivity decay index of the temperature residual sequence to the historical loss sequence. S5: Extract the harmonic residual sequence from the three-phase current time-domain waveform, fuse the temperature residual sequence with the harmonic residual sequence, and distinguish between transient load fluctuations and persistent thermal anomalies based on the sensitivity attenuation index, and output the early diagnosis results of inter-turn short circuits.

[0020] In S1, the operating data of the motor is collected. This operating data includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends. Specifically, it includes: During motor operation, data acquisition is performed first. The motor in question is a permanent magnet synchronous motor. Three K-type thermocouples are pre-embedded at the ends of the stator windings, positioned at the midpoints of the three-phase winding ends to acquire the measured transient temperature sequence at the winding ends. Three Hall effect current sensors are used to acquire the three-phase current time-domain waveforms, each connected to the motor's three-phase power supply lines, with a range twice the rated current, synchronously recording the instantaneous value of each phase current. The DC bus voltage is obtained from the inverter's DC side using a resistive voltage divider sensor. The rotor position angle is acquired using an incremental photoelectric encoder mounted on the motor shaft, with the output angle value expressed as an electrical angle. All analog signals output from the sensors are converted to digital quantities by a 16-bit analog-to-digital converter and synchronized using a GPS-timed timestamp, forming a time-aligned multi-channel data stream, stored in a local buffer for subsequent steps.

[0021] In S2, based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform, fractional calculus is used to construct a thermal diffusion characterization, inverting the fractional order and heat capacity parameters to form a power-law weighted accumulation process of historical losses, specifically including: First, the temperature rise response segment after a step excitation is extracted from the measured transient temperature sequence. Specifically, at the moment of motor startup or after a step current command with an amplitude of 1.2 times the rated current is applied, the measured transient temperature values ​​at the winding ends are recorded for two consecutive seconds from that moment, forming the temperature rise response segment. The starting point of this response segment corresponds to the moment the step excitation is applied, and the ending point is selected when the temperature rise rate drops to 10% of the initial rate.

[0022] Secondly, a historical loss sequence is constructed. Based on the collected three-phase current time-domain waveforms, the sum of the squares of the effective values ​​of the three-phase currents at the current moment is calculated every 0.1 milliseconds. This sum is then multiplied by the reference value of the stator winding resistance to obtain the copper loss value at that moment. Simultaneously, an approximate value of the iron loss is calculated based on the current change rate and the rotor position angle. The copper loss and iron loss are then added together to obtain the total loss value. The total loss values ​​at each moment are arranged in chronological order to form a historical loss sequence, with the sequence length being 500 sampling points prior to the current moment.

[0023] Then, the historical loss sequence is discretely convolved with the power-law weighting function in reverse chronological order to obtain the predicted temperature rise. Specifically, the 500 loss values ​​traced back from the current moment in the historical loss sequence are arranged in ascending order of time, forming a reverse-order vector. The power-law weighting function is defined as a power-law sequence with the decay exponent as the independent variable, where the... Each weight value equals The negative alpha power is given by the formula, where alpha is the fractional order to be inverted, initially set to 0.8. Each loss value in the inverted vector is multiplied by the corresponding power-law weight value, and then all products are summed to obtain the predicted temperature rise.

[0024] Next, with the goal of approximating the predicted temperature rise to the temperature rise response segment, the fractional order alpha and heat capacity parameters are iteratively adjusted. The specific iterative process is as follows: Calculate the sum of squares of the absolute differences between the predicted temperature rise and the measured temperature rise at the corresponding time in the temperature rise response segment, using this as the objective function. Employ the simplex search algorithm to sequentially adjust the fractional order alpha and heat capacity parameters, where the heat capacity parameter is defined as the heat capacity per unit mass of the winding, with an initial value of 380 joules per Kelvin per kilogram. After each adjustment, the predicted temperature rise and objective function value are recalculated. Iteration stops when the objective function value is less than one percent of the average measured temperature rise in the temperature rise response segment, and the converged alpha and heat capacity parameters are recorded.

[0025] The converged fractional order and heat capacity parameter are substituted into the power-law weighted accumulation process. This process is used in subsequent real-time calculations: for each new historical loss value, a power-law weight sequence is constructed using the converged alpha according to the aforementioned discrete convolution method, and then weighted and accumulated in reverse order with the most recent 500 historical loss values. The result is the predicted temperature rise at the current moment.

[0026] The specific implementation of the power-law weighting function is as follows: The first step is to extract a preset number of loss values ​​from the historical loss sequence up to the current moment to form a loss vector. The preset number is 500, which means taking the total loss values ​​of the 500 consecutive sampling moments before the current moment and arranging them in chronological order into a one-dimensional vector.

[0027] The second step involves multiplying the loss vector point-by-point with the initial power-law weight vector and then summing the results to obtain the predicted temperature rise. The initial power-law weight vector is constructed as follows: The initial decay exponent is set to 0.8, and for the first element in the vector… Locations ( (from 1 to 500), its weight value is equal to The negative 0.8th power. The th element in the loss vector. The loss value multiplied by the corresponding first... Each weighted value is used to calculate the predicted temperature rise. The 500 product results are then summed to obtain the predicted temperature rise.

[0028] The third step involves calculating the temperature rise difference between the predicted temperature rise and the corresponding moment in the measured transient temperature sequence. This difference is then used to reverse-correct the decay exponent of each element in the power-law weight vector. Specifically, the correction method is as follows: subtract the increase in measured temperature relative to the initial temperature from the predicted temperature rise to obtain the temperature rise difference. If the difference is positive, it indicates that the prediction is too high, so the decay exponent is increased; if the difference is negative, it indicates that the prediction is too low, so the decay exponent is decreased. Each correction step is two percent of the current decay exponent. The decay exponent corresponding to each position in the weight vector is corrected independently, and the corrected decay exponent is used to recalculate the weight value at that position.

[0029] The fourth step involves using the corrected power-law weight vector as the power-law weight function for the current moment, which is then used in the subsequent power-law weighted accumulation process of historical losses. That is, in the next calculation of the predicted temperature rise, the weight values ​​obtained from this correction are used to replace the initial weight values, and steps two through four are repeated to achieve online adaptive updating of the weight function.

[0030] In S3, the historical loss sequence is input into the power-law weighted accumulation process to calculate the transient temperature distribution at each virtual location point inside the winding in real time, thereby obtaining a dynamic thermal reference with memory effect, specifically including: First, the interior of the motor stator winding is evenly divided radially from the bottom of the slot to the opening into five virtual position points. Each virtual position point corresponds to a heat diffusion depth, defined as the radial distance from the virtual position point to the surface of the copper conductor at the bottom of the winding slot. The first virtual position point has a heat diffusion depth of 0 mm and is located on the surface of the copper conductor; the second virtual position point has a heat diffusion depth of 0.3 mm and is located inside the main insulation layer; the third virtual position point has a heat diffusion depth of 0.6 mm and is located at the junction of the insulation layer and the slot wall; the fourth virtual position point has a heat diffusion depth of 0.9 mm and is located inside the core slot wall; and the fifth virtual position point has a heat diffusion depth of 1.2 mm and is located at the core yoke. The thermal capacity and thermal diffusivity of each virtual location are set according to the material properties of that location. The thermal capacity of the copper conductor region is 385 joules per kilogram and the thermal diffusivity is 117 square millimeters per second; the thermal capacity of the insulation layer region is 1200 joules per kilogram and the thermal diffusivity is 0.25 square millimeters per second; and the thermal capacity of the iron core region is 450 joules per kilogram and the thermal diffusivity is 12 square millimeters per second.

[0031] Secondly, based on the thermal diffusion depth and heat source distance of each virtual location point, different power-law weighting coefficients are assigned to the same historical loss sequence. The heat source is defined as the heat generated by internal losses within the copper conductor, and the heat source distance is the thermal diffusion depth of the virtual location point. For the... For each virtual location point, the decay exponent alpha in the power-law weighting coefficients is determined as follows: alpha equals the baseline alpha value obtained in step S2 multiplied by a depth correction factor, which is equal to the sum of the thermal diffusion depth and 0.5 mm. That is, the greater the thermal diffusion depth, the greater the alpha, indicating that the memory effect of heat decays faster at that depth. Simultaneously, the power-law weighted accumulation process for each virtual location point uses the same historical loss sequence, but the weight vector... The element value equals The negative alpha power. Specifically, in the calculation, for the ... For each virtual location point, 500 loss values ​​prior to the current moment are extracted from the historical loss sequence, arranged in reverse chronological order, and multiplied point by point with the power-law weight vector and accumulated to obtain the predicted temperature rise at that location point. Then, the predicted temperature rise is multiplied by the reciprocal of the heat capacity parameter at that location point to obtain the transient temperature value at that location point.

[0032] Then, the historical loss sequence is input into the power-law weighted accumulation process at each virtual location point to calculate the transient temperature values ​​of the five points in parallel. The parallel calculation is implemented as follows: at each sampling time (sampling frequency of 10 kHz), five independent accumulation calculation threads are started simultaneously. Each thread processes its own power-law weighting coefficients and heat capacity parameters according to the aforementioned steps, and calculates the transient temperature value of the corresponding virtual location point at the current time. After all threads have completed their calculations, the five transient temperature values ​​are arranged in order of virtual location points to form a one-dimensional temperature vector. Each element in this vector represents the transient temperature at the corresponding thermal diffusion depth.

[0033] Finally, the transient temperature values ​​of all virtual location points are combined into a dynamic thermal reference with a memory effect. This combination process specifically includes the following sub-steps: The first step is to calculate the heat propagation delay time between each virtual location point and the measured temperature point at the winding end. The heat propagation delay time is determined by both the heat diffusion depth and the fractional order, and the calculation formula is as follows: ; in, For the first The thermal propagation delay time of each virtual location point, in seconds; The thermal diffusion depth of this virtual location point, in millimeters; The thermal diffusivity is taken as 117 square millimeters per second for the copper conductor region. Let S2 be the fractional order obtained by convergence, with a value of 0.82. For gamma function, when When it is 0.82, The value of (1.82) is approximately 0.93. In the above formula, the delay time is obtained by dividing the square of the thermal diffusion depth by the product of twice the thermal diffusion coefficient and the gamma function value. For the 5 virtual location points, the delay time is calculated sequentially: the delay time of point 1 is 0 seconds, the delay time of point 2 is 0.0005 seconds, the delay time of point 3 is 0.002 seconds, the delay time of point 4 is 0.0045 seconds, and the delay time of point 5 is 0.008 seconds.

[0034] The second step involves performing a time-reverse shift on the transient temperature value sequence at each virtual location point based on the heat propagation delay time, aligning each sequence with the measured transient temperature sequence in the time dimension. Specifically, for the first... A virtual location point is used, and the transient temperature value calculated at that point at the current moment is assigned to the current moment plus a delay time. That is, if the current moment is t, then in the shifted temperature value sequence, t is added... The temperature value at any given moment is equal to the transient temperature value calculated at that moment. After translation, the temperature value sequence of all virtual location points is based on the time marker of the measured transient temperature sequence on the time axis, so that the temperature value of each virtual location point at the same moment corresponds to the actual temperature reached by heat propagation at that moment.

[0035] The third step involves weighting and summing the aligned sequences according to the reciprocal of the thermal diffusion depth at each virtual location point to obtain the fusion temperature sequence. The reciprocal of the thermal diffusion depth serves as a weighting coefficient, its physical meaning being that the closer a location is to the heat source, the greater its contribution to the measured end temperature. The specific weighted summation is calculated as follows: For each time point, the transient temperature values ​​of the five virtual locations after translation at that time point are multiplied by their respective reciprocals of the thermal diffusion depth. Since the reciprocal of the thermal diffusion depth (0 mm) is mathematically undefined, the weight of the first virtual location point is set to 1.0, the weight of the second virtual location point is 1 divided by 0.3 (3.33), the weight of the third virtual location point is 1 divided by 0.6 (1.67), the weight of the fourth virtual location point is 1 divided by 0.9 (1.11), and the weight of the fifth virtual location point is 1 divided by 1.2 (0.83). All products are summed, and then divided by the sum of the weights to obtain the fusion temperature value at that time point. The sum of the weights is 1.0 + 3.33 + 1.67 + 1.11 + 0.83, which equals 7.94. This calculation is performed sequentially for each time point in the above manner to form a fusion temperature sequence.

[0036] The fourth step involves using the fused temperature sequence as a dynamic thermal benchmark with a memory effect. This dynamic thermal benchmark reflects the comprehensive temperature response of the historical loss sequence after fractional power-law weighted accumulation, radial multi-point thermal diffusion delay, and end-fusion correction. It is used for time-by-time comparison with the measured transient temperature sequence in subsequent steps.

[0037] In S4, the dynamic thermal reference is compared with the measured transient temperature series time-by-time to obtain the temperature residual series, and the sensitivity decay index of the temperature residual series to the historical loss series is calculated, specifically including: First, the dynamic thermal reference obtained in step S3 is subtracted from the measured transient temperature sequence of the winding end collected in step S1 at each time point on the same time coordinate to obtain the temperature residual sequence. Specifically, for each sampling time (sampling frequency of 10 kHz), the dynamic thermal reference value at that time is subtracted from the measured transient temperature value at that time; the difference is the temperature residual value at that time. If the difference is positive, it indicates that the temperature predicted by the dynamic thermal reference is higher than the measured value, suggesting that the actual loss is lower than expected or the heat dissipation capacity is enhanced; if the difference is negative, it indicates that the temperature predicted by the dynamic thermal reference is lower than the measured value, suggesting the existence of an abnormal heat source. The temperature residual values ​​at all times are arranged in chronological order to form the temperature residual sequence.

[0038] Secondly, the residual values ​​at each time point in the temperature residual sequence are divided by the corresponding historical loss value to obtain the transient sensitivity sequence. The historical loss value is taken from the total loss value at the same moment in the historical loss sequence calculated in S2. During the division operation, if the historical loss value is zero, the transient sensitivity value at that moment is set to zero. The physical meaning of the transient sensitivity sequence is that the magnitude of the temperature residual generated per unit loss reflects the sensitivity of the current temperature deviation to changes in loss.

[0039] Then, the natural logarithm of the transient sensitivity sequence is taken in chronological order to obtain the logarithmic sensitivity sequence. Specifically, for each positive sensitivity value in the transient sensitivity sequence, its natural logarithm is calculated; if the sensitivity value is zero or negative, its natural logarithm is set to a sufficiently small negative number, such as -10. The purpose of taking the natural logarithm is to transform the originally power-law decaying sensitivity curve into a linear form, facilitating subsequent fitting.

[0040] Next, the time interval where the temperature residual is consistently positive is extracted from the logarithmic sensitivity sequence. The specific method is as follows: starting from the beginning of the temperature residual sequence, the search continues until the first ten consecutive sampling points with positive temperature residuals are identified as the starting point. This search continues until five consecutive sampling points with non-positive temperature residuals are identified as the ending point. The interval between the starting and ending points is the time interval where the temperature residual is consistently positive. Within this interval, the first 10% of sampling points (considered as the transient fluctuation segment) and the last 10% of sampling points (considered as the noise segment) are removed, leaving the remaining middle portion as the effective attenuation interval.

[0041] Subsequently, the logarithmic sensitivity values ​​at each time point within the effective attenuation interval are arranged in chronological order, and a confidence weight proportional to the absolute value of the residual is assigned to each time point. The confidence weight is calculated as follows: first, the sum of the absolute values ​​of all temperature residuals within the effective attenuation interval is calculated; then, the absolute value of the temperature residual at each time point is divided by this sum to obtain the initial weight for that time point. Finally, the initial weights at each time point are normalized so that the sum of all weights within the effective attenuation interval equals one.

[0042] Next, a weighted least squares method is used to fit a straight line to minimize the weighted sum of squared residuals. The process of fitting the straight line is as follows: the equation of the straight line is set so that the logarithmic sensitivity value equals the intercept plus the slope multiplied by the time index. The time index is numbered starting from the first sampling point in the effective decay interval as 1, the second as 2, and so on. The weighted sum of squared residuals is calculated as follows: for each sampling point, the difference between the fitted value and the actual logarithmic sensitivity value at that point is calculated first, the difference is squared and multiplied by the confidence weight of that point, and then the weighted results of all sampling points are added together. By traversing and searching the possible values ​​of the slope and intercept (the slope search range is from -1 to 0, with a step size of -0.01; the intercept search range is from -5 to 5, with a step size of 0.1), the slope value that minimizes the weighted sum of squared residuals is found, and this slope value is output as the descending slope of the fitted straight line.

[0043] Finally, the inverse value of the descent slope is used as the sensitivity decay index. That is, if the fitted descent slope is negative 0.3, the sensitivity decay index is positive 0.3. The sensitivity decay index is used to quantify the rate at which the temperature residual decays with increasing historical loss over time: the larger the index value, the faster the dependence of the temperature residual on historical loss decays, and the current temperature residual is mainly contributed by recent loss; the smaller the index value, the stronger the memory effect of the temperature residual, and the longer the contribution of historical loss to the current temperature residual lasts.

[0044] In S5, harmonic residual sequences are extracted from the three-phase current time-domain waveforms. The temperature residual sequences are fused with the harmonic residual sequences, and transient load fluctuations and persistent thermal anomalies are distinguished based on the sensitivity decay index. Early diagnosis results for inter-turn short circuits are output, specifically including: First, the fundamental component is subtracted periodically from the three-phase current time-domain waveform to obtain the harmonic component sequence for each phase. Specifically, for each phase current, the fundamental amplitude and phase of the current cycle are calculated using a discrete Fourier transform. The fundamental waveform is then reconstructed, and this fundamental waveform is subtracted from the original current waveform point by point to obtain the harmonic component sequence. The cycle length is determined based on the rotor position angle acquired in step S1, with each 360-degree electrical angle constituting one cycle. The sampling frequency is 10 kHz, and each cycle contains approximately 200 sampling points. This operation is performed on each of the three phase currents to obtain the harmonic component sequences for phases A, B, and C.

[0045] Secondly, the harmonic component sequences of each phase are subtracted from the corresponding phase's health-state reference harmonic sequence at each time point to obtain the three-phase harmonic residual sequences. The health-state reference harmonic sequence is obtained as follows: under the healthy condition of the motor after factory manufacturing or maintenance, it is continuously run at rated speed and rated load for 30 seconds. The harmonic component sequences of each phase are extracted using the aforementioned method. Then, the harmonic amplitudes of the same phase angle in all cycles within 30 seconds are arithmetically averaged to obtain an average harmonic waveform, which serves as the reference harmonic sequence for that phase. The harmonic component sequences of each phase at the current moment are subtracted from the corresponding phase's reference harmonic sequence; the difference is the harmonic residual sequence. Each of the three phases is calculated independently, resulting in three harmonic residual sequences, each with a length consistent with the number of sampling points.

[0046] Then, the temperature residual sequence obtained in S4 is multiplied point-by-point with the above three-phase harmonic residual sequence along the same time coordinate, and then accumulated to obtain the fusion characteristic value. The specific calculation method is as follows: for each sampling time, the temperature residual value at that time is multiplied by the harmonic residual values ​​of phase A, phase B, and phase C at that time, respectively, to obtain three products; these three products are then added together to obtain a sum; then, the sums of all sampling times are accumulated in chronological order, and the final accumulated result is the fusion characteristic value. The time interval for accumulation is the time interval during which the temperature residual is continuously positive, as determined in S4, i.e., the effective attenuation interval. The physical meaning of the fusion characteristic value is: the combined response of temperature deviation and current harmonic deviation over time; the larger the value, the stronger the correlation between thermal anomalies and electromagnetic anomalies.

[0047] Finally, when the fused characteristic value exceeds the preset threshold and the sensitivity attenuation index obtained in step S4 is lower than the attenuation threshold, the early diagnosis result of inter-turn short circuit is output. The preset threshold is determined as follows: under the healthy condition of the motor, 10 sets of data within the effective attenuation range are continuously collected, 10 fused characteristic values ​​are calculated in the above manner, and the arithmetic mean of these 10 values ​​is multiplied by 3 to obtain the preset threshold. The attenuation threshold is set to 0.2. This value comes from the analysis of measured data of multiple motors in the early stage of inter-turn short circuit. When the sensitivity attenuation index is lower than 0.2, it indicates that the dependence of the temperature residual on historical losses attenuates slowly, that is, there is a persistent thermal anomaly. When both conditions are met, the fused characteristic value exceeds the preset threshold and the sensitivity attenuation index is lower than 0.2, it is judged as an early inter-turn short circuit fault, and a diagnosis result including the fault type (inter-turn short circuit) and confidence level is output; if only one condition is met, it is diagnosed as no abnormality or transient load fluctuation, and no inter-turn short circuit alarm is output.

[0048] Please see Figure 2 As shown, the predictive maintenance system for motors based on digital twins includes: The data acquisition module collects the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends. The fractional-order parameter inversion module uses fractional-order calculus to construct a thermal diffusion characterization based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform. It then inverts the fractional-order order and heat capacity parameters, forming a power-law weighted accumulation process of historical losses. The dynamic thermal reference generation module inputs the historical loss sequence into the power-law weighted accumulation process, calculates the transient temperature distribution of each virtual position point inside the winding in real time, and obtains a dynamic thermal reference with memory effect. The temperature residual and sensitivity calculation module compares the dynamic thermal reference with the measured transient temperature sequence point by point to obtain the temperature residual sequence, and calculates the sensitivity decay index of the temperature residual sequence to the historical loss sequence. The integrated diagnostic module extracts the harmonic residual sequence from the three-phase current time-domain waveform, merges the temperature residual sequence with the harmonic residual sequence, and distinguishes between transient load fluctuations and persistent thermal anomalies based on the sensitivity decay index, outputting early diagnosis results for inter-turn short circuits.

[0049] The working principle of this invention is as follows: First, the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends of the motor are acquired. Then, based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform, fractional-order calculus is used to inversely derive the fractional-order order and heat capacity parameters, forming a power-law weighted accumulation process of historical losses. Next, the historical loss sequence is input into this accumulation process to calculate the transient temperature distribution of multiple virtual position points inside the winding in real time, obtaining a dynamic thermal reference with memory effect. Then, the dynamic thermal reference and the measured transient temperature sequence are compared hourly to obtain the temperature residual sequence, and the sensitivity attenuation index of the temperature residual sequence to the historical loss sequence is calculated. Finally, the harmonic residual sequence is extracted from the three-phase current time-domain waveform, and the temperature residual sequence and the harmonic residual sequence are fused to obtain a fused feature value. When the fused feature value exceeds a preset threshold and the sensitivity attenuation index is lower than the attenuation threshold, the early diagnosis result of inter-turn short circuit is output, thereby realizing accurate prediction of early motor faults.

[0050] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A predictive maintenance method for motors based on digital twins, characterized in that, Includes the following steps: S1: Collect the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends; S2: Based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform, a thermal diffusion characterization is constructed using fractional calculus, and the fractional order and heat capacity parameters are deduced to form a power-law weighted accumulation process of historical loss. S3: Input the historical loss sequence into the power-law weighted accumulation process to calculate the transient temperature distribution of each virtual position point inside the winding in real time, and obtain a dynamic thermal reference with memory effect; S4: Compare the dynamic thermal reference with the measured transient temperature series time-by-time to obtain the temperature residual series, and calculate the sensitivity decay index of the temperature residual series to the historical loss series, specifically including: Divide the residual value at each time point in the temperature residual sequence by the historical loss value at the corresponding time point to obtain the transient sensitivity sequence. The logarithmic sensitivity sequence is obtained by taking the natural logarithm of the transient sensitivity sequence in chronological order. Linear fitting is performed on the logarithmic sensitivity sequence, and the decreasing slope of the fitted line is extracted; The opposite value of the downward slope is used as the sensitivity decay index; S5: Extract the harmonic residual sequence from the three-phase current time-domain waveform, fuse the temperature residual sequence with the harmonic residual sequence, and distinguish between transient load fluctuations and persistent thermal anomalies based on the sensitivity decay index, outputting early diagnosis results for inter-turn short circuits, specifically including: The harmonic component sequence of each phase is obtained by subtracting the fundamental component from the three-phase current time-domain waveform periodically. The three-phase harmonic residual sequence is obtained by subtracting the harmonic component sequence of each phase from the corresponding phase health status reference harmonic sequence at each time point. The temperature residual sequence and the three-phase harmonic residual sequence are multiplied point by point along the same time coordinate and then summed to obtain the fused characteristic value; When the fused feature value exceeds the preset threshold and the sensitivity attenuation index is lower than the attenuation threshold, the early diagnosis result of inter-turn short circuit is output.

2. The predictive maintenance method for motors based on digital twins according to claim 1, characterized in that, S2 specifically includes: Extract the temperature rise response segment after step excitation from the measured transient temperature sequence; The predicted temperature rise is obtained by performing a discrete convolution of the historical loss sequence in reverse chronological order with a power-law weighting function. With the goal of predicting the temperature rise to approximate the temperature rise response segment, the fractional order and heat capacity parameters are iteratively adjusted. Substitute the converged fractional order and the heat capacity parameter into the power-law weighted accumulation process.

3. The predictive maintenance method for motors based on digital twins according to claim 2, characterized in that, The power-law weighting function is constructed and used in the following ways: Extract a predetermined number of loss values ​​prior to the current moment from the historical loss sequence to form a loss vector; The predicted temperature rise is obtained by multiplying the loss vector point by point with the initial power-law weight vector and then summing the results. Calculate the temperature rise difference between the predicted temperature rise and the temperature rise difference at the corresponding moment in the measured transient temperature sequence, and correct the decay exponent of each element in the power law weight vector in reverse based on the corresponding difference. The modified power-law weight vector is used as the power-law weight function at the current moment.

4. The predictive maintenance method for motors based on digital twins according to claim 1, characterized in that, S3 specifically includes: The interior of the winding is divided radially into multiple virtual location points, and each virtual location point corresponds to a thermal diffusion depth. Based on the thermal diffusion depth and heat source distance of each virtual location, different power-law weighting coefficients are assigned to the same historical loss sequence. The historical loss sequence is input into the power-law weighted accumulation process of each virtual location point, and the transient temperature value of each point is calculated in parallel. The transient temperature values ​​of all virtual location points are combined into a dynamic thermal reference with a memory effect.

5. The predictive maintenance method for motors based on digital twins according to claim 4, characterized in that, The process of combining the transient temperature values ​​of all virtual location points into a dynamic thermal reference with a memory effect specifically includes: Calculate the heat propagation delay time between each virtual location point and the measured temperature point at the end of the winding. The delay time is determined by both the heat diffusion depth and the fractional order. Based on the heat propagation delay time, the transient temperature value sequence of each virtual location point is shifted backward in time so that each sequence is aligned with the measured transient temperature sequence in the time dimension; The aligned sequences are weighted and summed according to the reciprocal of the thermal diffusion depth of their respective virtual locations to obtain the fusion temperature sequence. The fused temperature sequence is used as a dynamic thermal reference with memory effect.

6. The predictive maintenance method for motors based on digital twins according to claim 1, characterized in that, The extraction of the descending slope of the fitted straight line specifically includes: The effective attenuation range is obtained by extracting the time interval in which the temperature residual is continuously positive from the logarithmic sensitivity sequence, removing the initial transient fluctuation segment and the final noise segment; The logarithmic sensitivity values ​​at each time point within the effective decay interval are arranged in chronological order, and each time point is assigned a confidence weight proportional to the absolute value of the residual. We use the weighted least squares method to fit a straight line, minimizing the weighted sum of squared residuals; The slope of the fitted line is output as the descending slope.

7. A predictive maintenance system for motors based on digital twins, characterized in that, The method for performing the predictive maintenance of motors based on digital twins as described in any one of claims 1-6 includes: The data acquisition module collects the motor's operating data, which includes the three-phase current time-domain waveform, DC bus voltage, rotor position angle, and measured transient temperature sequence at the winding ends. The fractional-order parameter inversion module uses fractional-order calculus to construct a thermal diffusion characterization based on the measured transient temperature sequence and the historical loss sequence calculated from the three-phase current time-domain waveform. It then inverts the fractional-order order and heat capacity parameters, forming a power-law weighted accumulation process of historical losses. The dynamic thermal reference generation module inputs the historical loss sequence into the power-law weighted accumulation process, calculates the transient temperature distribution of each virtual position point inside the winding in real time, and obtains a dynamic thermal reference with memory effect. The temperature residual and sensitivity calculation module compares the dynamic thermal reference with the measured transient temperature sequence point by point to obtain the temperature residual sequence, and calculates the sensitivity decay index of the temperature residual sequence to the historical loss sequence. The integrated diagnostic module extracts the harmonic residual sequence from the three-phase current time-domain waveform, merges the temperature residual sequence with the harmonic residual sequence, and distinguishes between transient load fluctuations and persistent thermal anomalies based on the sensitivity decay index, outputting early diagnosis results for inter-turn short circuits.

Images