Data-driven motor stator and rotor temperature online estimation method

Abstract

The invention discloses a data-driven motor stator and rotor temperature on-line estimation method, which comprises the following steps of: firstly, extracting a signal strongly related to motor temperature to perform temperature estimation so as to avoid input redundancy and reduce model calculation amount; secondly, constructing a TimesNet temperature estimation model, and efficiently extracting multi-scale and multi-cycle dynamic characteristics in the input signal by utilizing the unique multi-cycle time sequence decomposition capability of the TimesNet temperature estimation model; higher estimation precision is shown by using fewer data training samples, and a transfer learning framework is constructed in order to improve the generalization ability of the model. A permanent magnet synchronous motor temperature estimation model of a source domain is obtained by combining conventional working condition data with TimesNET model training, and knowledge is migrated to temperature estimation of the same motor under different operation conditions and different motors under different operation conditions through a fine tuning strategy. Therefore, the stator temperature and the rotor temperature of the motor under two different operation conditions of the same motor and different operation conditions of different motors are respectively and accurately estimated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of temperature monitoring, and particularly relates to a data-driven online estimation method for motor stator and rotor temperature. BACKGROUND

[0002] Permanent Magnet Synchronous Motor (PMSM) is widely used in new energy vehicle drive system, industrial robot and aerospace precise drive field due to its high power density, high efficiency, wide speed regulation range and excellent dynamic response characteristics. However, the temperature-sensitive components (such as stator winding and rotor permanent magnet) in the motor are prone to failure under high temperature stress: the stator winding may melt due to overheating, and the rotor permanent magnet may irreversibly demagnetize when overloading, which seriously affects the performance and service life of the motor. Therefore, accurate and real-time temperature estimation of the internal "hot spots" of the motor is of great significance to ensure the safe operation of the PMSM, optimize the control performance, realize accurate parameter identification, and maximize the power density while enabling self-protection.

[0003] For a long time, the industry has mainly relied on physical sensors for temperature monitoring. However, sensors are expensive and their functions will degrade over time. More importantly, due to the rotation of the rotor and the complex internal structure, it is extremely difficult and costly to deploy temperature sensors in the rotor. This lack of monitoring capability forces manufacturers to adopt expensive oversized motor designs as a safety margin, resulting in underutilization of the motor's overload potential.

[0004] Traditional temperature estimation methods that do not rely on physical sensors mainly include lumped parameter thermal network (LPTNs) and electrical parameter-based estimation methods. LPTNs approximate internal heat transfer through equivalent circuit diagrams, have the advantages of non-invasiveness and no need for additional hardware, but its main weakness is the trade-off between physical interpretability and real-time performance. After simplifying the degrees of freedom to meet real-time computing requirements, its interpretability will be weakened. In addition, designing a high-performance, low-order LPTN requires rich field expertise, and its parameterization is highly dependent on the geometry of the motor and the cooling system. This means that each different motor requires time-consuming and labor-intensive professional design and experimental data validation. Electrical parameter-based estimation methods include techniques based on fundamental flux observers and high-frequency signal injection. These methods also require complex modeling and accurate identification of the motor and inverter. Moreover, the fundamental flux observer has reduced estimation accuracy at low speeds, while high-frequency signal injection can cause additional losses and electromagnetic interference in the system, making it unsuitable for some applications.

[0005] The limitations of all the above traditional methods directly lead to increased production costs, material waste, and underutilization of motor performance. With the development of artificial intelligence, data-driven temperature estimation methods show great potential. Such methods do not require an accurate motor physical model, but instead directly collect easily measured voltage, current, speed, and other data from the running process to mine the nonlinear mapping relationship between them and temperature. However, models trained only on a single machine type and single operating condition often have a significant change in thermal feature distribution when faced with different loads, speeds, or different motors, resulting in a decrease in estimation accuracy. Therefore, the feature heterogeneity of data distribution needs to be considered in model design to achieve temperature estimation across operating conditions and across machine types. Data-driven methods automatically learn nonlinear mapping relationships from a large amount of running data and show great potential in dealing with complex operating condition changes. Combined with transfer learning, the problem of data differences between source and target domains can be further solved, thereby improving the adaptability of the model to new machines or new operating conditions. There are few existing solutions using data-driven methods for motor temperature estimation. The patent "Permanent Magnet Synchronous Motor Temperature Estimation Method Based on Model Data Dual Drive" (CN202411798888.8) estimates motor temperature by designing three modules: model-driven, data-driven, and state estimation. The design of the model-driven module relies heavily on the geometric and material parameters of the motor, which are difficult to obtain in actual applications, and the temperature estimation accuracy of this method is not mentioned. The literature "A model-based and data-driven integrated temperature estimation method for PMSM" uses a model-driven and data-driven integrated temperature estimation method. Although the model-driven part is simplified, reducing the dependence on parameters, the data-driven model part requires 10 signal inputs and 120000 data samples for training, increasing the computational burden of the data-driven model and posing challenges for actual deployment. The literature "Gradient Boosting Decision Tree for Rotor Temperature Estimation in Permanent Magnet Synchronous Motors" uses a pure data-driven method based on Gradient Boosting Decision Tree (GBDT) to estimate motor temperature. The model also requires 9 electrical and thermal related signals as input, and the generalization ability of the model on different motors has not been verified.

[0006] Therefore, there are the following defects in the current motor temperature estimation methods such as PMSM:

[0007] (1) The estimation accuracy of motor temperature does not match the model training cost. Generally, good estimation accuracy is based on multiple input signals and a large number of training samples. In existing motor estimation methods, the training of the model often needs multiple signal sources such as voltage, current, speed, torque, stator temperature (tooth, winding, yoke) or rotor temperature as input. Among them, the acquisition of the temperature source not only needs additional sensors, but also belongs to the thermal signal and has the same trend (as shown in Figure 1 ), which increases the additional cost and weakens the temperature estimation ability of the model. In addition, the training sample of the model is as high as hundreds of thousands, which increases the calculation burden of the model.

[0008] (2) The model generalization ability is insufficient, which seriously restricts its engineering practicability. First, in actual operation, the motor working condition presents a high dynamic characteristic, and the speed and torque change frequently and dramatically, showing complex multi-time scale characteristics (both slow change and transient impact). This complex time sequence dependence leads to a significant decrease in prediction accuracy of the model trained based on a specific working condition when facing new working conditions, making it difficult to meet the demand for full working condition coverage in engineering applications. Second, in engineering practice, even the same batch and same type of motor products, the key electrical parameters (stator resistance, inductance, permanent magnet flux linkage, etc.) will have a deviation of 5-15% due to manufacturing tolerance, assembly error and running aging. These seemingly small parameter differences will be significantly amplified through thermal accumulation effect during long-term operation, eventually leading to a temperature prediction deviation of 10-20℃. Therefore, the model trained for a specific motor sample cannot be directly transplanted to other motors, which seriously limits the batch deployment ability of the model.

[0009] Therefore, it is urgent to design a data-driven motor stator and rotor temperature online estimation method. SUMMARY

[0010] (1) The technical problem to be solved

[0011] Based on the above defects, the present application proposes a data-driven motor stator and rotor temperature online estimation method, which is a data-driven and transfer learning-based motor stator and rotor average temperature estimation method, solving the problems of existing motor temperature estimation methods such as mismatch between estimation accuracy and training cost, strong dependence on external parameters, and weak model batch deployment ability.

[0012] (2) Technical scheme

[0013] The present application discloses a data-driven motor stator and rotor temperature online estimation method, comprising the following steps:

[0014] Step S1: using the correlation analysis method, extracting the input signals strongly related to the temperature of the permanent magnet synchronous motor for temperature estimation, the input signals of the permanent magnet synchronous motor temperature estimation model including d-axis current i d , d-axis voltage u d , q-axis voltage u q and motor speed n; the output signals include rotor permanent magnet temperature θ pm and stator winding temperature θ sw ;

[0015] Step S2: for the input signals in step S1, a TimesNET data-driven model is constructed, and the extraction of multi-scale features, the conversion of two-dimensional tensors, the extraction of two-dimensional convolution features and the conversion of two-dimensional tensors to one-dimensional space are completed in turn, so as to fold the one-dimensional time series of the input signals according to the detected main period length, form multiple two-dimensional tensors, and finally convert the two-dimensional tensors to one-dimensional tensors;

[0016] Step S3: based on different working conditions and different motors, a transfer learning framework is constructed and combined with the TimesNET data-driven model of step S2, and the input signals with strong correlation and easy collection are used to train the source domain permanent magnet synchronous motor temperature estimation model;

[0017] Step S4: using fine-tuning strategy, training the target domain permanent magnet synchronous motor temperature estimation model, verifying the model accuracy, and finally forming the target domain temperature estimation model based on the input signals and output signals, and realizing the training of the target domain temperature estimation model across working conditions and the training of the target domain temperature estimation model across motors;

[0018] Step S5: using the target domain temperature estimation model obtained in step S4 to realize the online estimation of rotor permanent magnet temperature θ pm and stator winding temperature θ sw in the output signal, and estimating the temperatures of the motor stator and rotor under the remaining 70% original data of the target domain across working conditions and across motor types.

[0019] Preferably, step S2 specifically includes the following step S210:

[0020] Step S210: multi-scale feature extraction: in the case that the input signals of d-axis current i d , d-axis voltage u d , q-axis voltage u q and motor speed n have obvious multi-scale features under known motor operating conditions, TimesNet data-driven model analyzes the frequency spectrum of the input signals by fast Fourier transform FFT, identifies the main period length, and automatically extracts high-frequency-voltage current, low-frequency-speed dual time scale;

[0021] The input signal selected by step S1 constructs an input signal matrix

[0022] (7)

[0023] The time series is subjected to a fast Fourier transform (FFT) to obtain an amplitude A corresponding to each frequency motor ∈R T is:

[0024]

[0025] In the formula, FFT(·) represents a fast Fourier transform, Amp(·) represents an amplitude calculation, and Avg(·) represents an averaging calculation;

[0026] The first k frequencies with the largest amplitudes are selected, thereby obtaining a set of most significant frequencies {f1 motor ,…,f k motor}:

[0027]

[0028] In the formula, represents a set of key frequencies

[0029] Based on the selected set of most significant frequencies {f1 motor ,…,f k motor}, the corresponding k period lengths {p1,…,p k} are determined as:

[0030]

[0031] The four frequency parameters are, in turn, the reciprocals of the electrical period T e , the mechanical period T m , the switching period T w , and the thermal time period T th of the permanent magnet synchronous motor in various periodic mode operations.

[0032] Preferably, step S2 further comprises the following step S220:

[0033] Step S220: Conversion of a two-dimensional tensor: the one-dimensional time series X 1D motor ∈R T×4 is reconstructed into multiple two-dimensional tensors by formula (11):

[0034]

[0035] where Padding(·) function pads the time series with zeros along the time dimension to make it compatible with the matrix reconstruction function based on the fixed number of elements where p is the period length i as the number of rows of the transformed two-dimensional tensor, represents the variation within one period, capturing the fine dynamics within one complete operation period; as the number of columns of the transformed two-dimensional tensor, represents the evolution trend between adjacent periods, the two-dimensional tensor represents the frequency based reconstruction time series.

[0036] Preferably, step S2 further comprises the following step S230:

[0037] Step S230: Extraction of two-dimensional convolution features: for the two-dimensional tensor , an efficient Inception convolution module is used to construct multi-scale thermal perception convolution kernels combined with thermal response time constants , which dynamically adjusts the feature extraction process according to the current temperature state. This module can learn the typical features of each operating condition of the motor (such as smooth running or rapid acceleration loading operation), as well as the common patterns and subtle differences between these features, which is represented as:

[0038]

[0039] The multi-scale thermal response convolution is represented as:

[0040]

[0041] where the thermal state vector , represents the thermal response time function, represents the thermal perception convolution kernel of the s-th scale, and each convolution kernel size is associated with the thermal time constant , is f i motor corresponding to the i-th period length, is the interval between adjacent two sampling periods, and the large kernel size 5x5 and 7x7 correspond to the long time constant of the motor / thermal system, and the small kernel size 1x1 and 3x3 correspond to the short time constant, is a scalar modulation factor in the range [0, 1].

[0042] Preferably, step S2 further comprises the following step S240:

[0043] Step S240: Conversion of two-dimensional tensor to one-dimensional space: Map the extracted two-dimensional features back to one-dimensional space for subsequent information fusion; the conversion process is as follows:

[0044]

[0045] In the formula, TempTrunc(·) removes the zero added by the Padding(·) operation in formula (4), and restores the feature sequence to the original time step;

[0046] According to the amplitude spectrum intensity corresponding to each periodic component, the one-dimensional features output by formula (14) are weighted and summed, and the process is as follows:

[0047]

[0048]

[0049] wherein, is the frequency f i motor The corresponding amplitude A motor By stacking multiple TimesBlocks described above, the model can capture the two-dimensional variation patterns of the multi-time scale input signal in layers, and can extract more rich and effective feature information compared with directly modeling on one-dimensional sequence.

[0050] Preferably, in step S3, the rationality and difficulty of data acquisition need to be considered: assuming that the A motor a working condition data is easy to collect, and the A motor b working condition and B motor c working condition data are difficult to collect due to complex working conditions and experimental condition limits, therefore, the A motor a working condition data is taken as the source domain data, the A motor b working condition and B motor c working condition data are taken as the target domain, and the temperature transfer scene condition is set as the source domain transferring to the target domain;

[0051] According to the input signals of the d-axis current i d , d-axis voltage u d , q-axis voltage u q and motor speed n and stator winding temperature θ sw extracted in step S1 under the A motor a working condition which are strongly related to the stator winding temperature, a permanent magnet synchronous motor temperature estimation model based on TimesNET data driven model and transfer learning framework of the source domain is pre-trained as a basic stator temperature estimation model;

[0052] The training of the rotor temperature estimation source model is the same, according to the d-axis current i​d d-axis voltage u d q-axis voltage u q and motor speed n and rotor permanent magnet temperature θ pm Pre-train a source domain permanent magnet synchronous motor temperature estimation model based on TimesNET data-driven model and transfer learning framework as the basic rotor temperature estimation model as a training set.

[0053] Preferably, in step S4, a new TimesNET data-driven model is initialized as a target model in the target domain, the parameters of which are copied from the source model pre-trained in step S3, the first 30% of the original data in the target domain divided in step S3 is used as the training set of the target domain, and a fine-tuning strategy is adopted to obtain the target domain TimesNET temperature estimation model with a smaller learning rate and fewer training rounds. During the fine-tuning process, all parameters outside the other layers of the TimesNet data-driven model are frozen, only the prediction layer parameters are updated, including reducing the learning rate and the number of iterations, the learning rate is reduced from 0.001 to 0.0001, and the number of training rounds is reduced by 50%.

[0054] Preferably, in step S4, it further includes:

[0055] Training of the target domain temperature estimation model across operating conditions: A motor a operating condition→A motor b operating condition;

[0056] Based on the basic stator temperature estimation model trained in step S3, the target domain motor stator temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy;

[0057] Based on the basic rotor temperature estimation model trained in step S3, the target domain motor rotor temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy;

[0058] Training of the target domain temperature estimation model across motors: A motor a operating condition→B motor c operating condition

[0059] Based on the basic stator temperature estimation model trained in step S3, the target domain motor stator temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy;

[0060] Based on the basic rotor temperature estimation model trained in step S3, the target domain motor rotor temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy.

[0061] In another aspect, the application also discloses a data-driven motor stator and rotor temperature online estimation system, comprising:

[0062] at least one processor; and at least one memory connected in communication with the processor, wherein:

[0063] The memory stores program instructions executable by the processor, and the processor invoking the program instructions can execute the data-driven motor stator and rotor temperature online estimation method of any one of the above.

[0064] (III) Beneficial effects

[0065] 1. In order to improve the generalization ability of the model and the estimation accuracy of the motor temperature, the application improves and designs a data-driven motor stator and rotor temperature online estimation system, first analyzes the relationship between the existing measurable signals and the thermal signals, extracts the signals strongly related to the motor temperature for temperature estimation, avoids input redundancy and reduces the model calculation amount. Secondly, in order to adapt to the multiscale characteristics of the speed / torque signal of the motor operating condition, more effectively extract the key information for temperature estimation, and the application adopts TimesNET specially used for estimating motor temperature as the driving model, with fewer data training samples, and shows higher estimation accuracy. At the same time, in order to improve the generalization ability of the model, a transfer learning framework is constructed. The TimesNET model pre-trained by the conventional working condition data is used to transfer the knowledge to the temperature estimation of the same motor under different operating conditions and different motors under different operating conditions through the fine-tuning strategy, and the generalization performance of the model is enhanced. Finally, the trained model is used to estimate the motor stator temperature and rotor temperature under two different operating conditions of the same motor and under different operating conditions of different motors.

[0066] 2. The motor temperature estimation method proposed in the application is based on the TimesNET data-driven model, which is used to realize the extraction of multiscale features, the conversion of two-dimensional tensors, the extraction of two-dimensional convolution features and the conversion of two-dimensional tensors to one-dimensional space in sequence. It uses an efficient Inception convolution module combined with a thermal response time constant to form a multiscale thermal perception convolution kernel, dynamically adjusts the feature extraction process according to the current temperature state, and in order to fuse the features extracted from different periodic scales, the one-dimensional features are weighted and summed according to the amplitude spectrum intensity corresponding to each periodic component, so that the estimated temperature has obvious data efficiency and estimation accuracy advantage. This method only relies on four external input variables and a lightweight training cost of 9000 training samples, and can achieve higher estimation accuracy. Its outstanding "small sample learning" ability makes it very suitable for rapid deployment and iteration, providing strong support for engineering practice. More importantly, by combining the transfer learning technology and the fine-tuning strategy, the model can still maintain good estimation performance when facing different operating conditions and machine types, meeting the needs of full-condition coverage and cross-machine batch deployment in engineering applications. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the present application or the prior art, the drawings required for use in the embodiments will be briefly introduced as follows:

[0068] Figure 1 For the temperature variation trend of different parts of the motor in the prior art,

[0069] Figure 2 For the overall flowchart of the data-driven motor stator and rotor temperature online estimation method in the present application,

[0070] Figure 3 is a schematic diagram of the motor speed, torque, rotor permanent magnet / stator winding temperature data under various operating conditions in the present application, (a) corresponds to the A motor a operating condition, (b) corresponds to the A motor b operating condition, and (c) corresponds to the B motor c operating condition;

[0071] Figure 4 For the correlation analysis result of the input and output signals in the present application,

[0072] Figure 5 For the working process schematic diagram of the TimesNet data-driven model in the present application,

[0073] Figure 6 is a temperature estimation result diagram of operating condition B in the embodiment of the present application, (a) is a motor temperature estimation result diagram, and (b) is an estimation error result diagram;

[0074] Figure 7 is a temperature estimation result diagram of operating condition C in the embodiment of the present application, (a) is a motor temperature estimation result diagram, and (b) is an estimation error result diagram. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0076] The overall flow of the data-driven motor stator and rotor temperature online estimation method in the present application is shown in Figure 2 The present application extracts the motor temperature estimation signal, constructs the TimesNET data-driven model, constructs the transfer learning framework, and estimates the motor temperature, etc. In addition, the present application takes the motor (A motor) temperature data set provided by the LEA department of Paderborn University in Germany and the temperature data set of another different motor (B motor) in the laboratory as examples, and the two data sets constitute three operating conditions abc, and each operating condition is shown in Figure 3.

[0077] As Figure 2 shown, the data-driven motor stator and rotor temperature online estimation method of the application specifically comprises the following steps:

[0078] Step S1: using a correlation analysis method, extracting input signals strongly related to the temperature of the permanent magnet synchronous motor for temperature estimation, the input signals of the permanent magnet synchronous motor temperature estimation model including d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n; the output signals including rotor permanent magnet temperature θ pm and stator winding temperature θ sw .

[0079] Specifically, in step S1, in order to select appropriate input parameters to accurately estimate the output signals, thereby reducing the difficulty and cost of sensor deployment while ensuring the estimation accuracy of the estimation model, the application uses a correlation analysis method, including Pearson, Kendall, and Spearman, to analyze the relationship between the existing measurable input signals and the output thermal signals, extract signals strongly related to the stator winding / rotor permanent magnet temperature for temperature estimation, the input and output signals of the permanent magnet synchronous motor temperature estimation model are shown in Table 1, and the analysis results are shown in Figure 4 .

[0080] Table 1 Input and output signals

[0081]

[0082] As Figure 4 can be seen, because the stator tooth temperature θ st , the stator yoke temperature θ sy , the rotor permanent magnet temperature θ pm , and the stator winding temperature θ sw all belong to thermal signals and have the same change trend (which has been proved in Figure 1 ), in any correlation analysis, their correlation coefficients occupy the top two positions, and their use as model inputs has not been fully verified. In addition, the measurement of temperature requires additional sensors, which increases the additional cost and difficulty of system setup. In summary, the d-axis current i d , the dq-axis voltage u dq (u dq including u d and u q ), and the motor speed n are finally selected as the inputs of the model.

[0083] Step S2: For the input signal in step S1, a TimesNET data-driven model is constructed, and the extraction of multi-scale features, the conversion of two-dimensional tensors, the extraction of two-dimensional convolution features, and the conversion of two-dimensional tensors to one-dimensional space are sequentially completed, so as to fold the one-dimensional time series of the input signal according to the detected main period length, form multiple two-dimensional tensors, and finally convert the two-dimensional tensors into one-dimensional tensors.

[0084] Specifically, in step S2, considering that there are multiple periodic patterns in the operation of the motor:

[0085]

[0086] Therefore, the dynamic thermal model of the motor temperature can be expressed as:

[0087]

[0088] Wherein:

[0089]

[0090]

[0091]

[0092]

[0093] In the formula, R th is the equivalent thermal resistance, C th is the equivalent heat capacity, is the motor temperature, is the ambient temperature, P loss (t) is the total loss of the motor, P cu (t) is the copper loss of the motor, P fe (t) is the iron loss of the motor, P mech (t) is the mechanical loss, R s is the motor resistance, k h is the hysteresis loss coefficient, k e is the eddy current loss coefficient, k ex is the abnormal loss coefficient, B is the magnetic flux density, is the Steinmetz coefficient, k fr is the friction coefficient, k wind is the wind resistance coefficient.

[0094] It can be seen that the motor temperature is a typical physical quantity with large inertia and slow change characteristics, and the motor loss (copper loss, iron loss, mechanical loss, etc.) driving its change is determined by high-frequency changing electrical quantities (current, voltage, speed). In practical applications, such as WLTP working conditions of electric vehicles, repetitive actions of industrial robots, etc., permanent magnet synchronous motors usually operate in working conditions with obvious periodicity.

[0095] The core idea of the TimesNet model in the prior art is to fold the one-dimensional time series according to the detected main period length to form multiple two-dimensional tensors. This method breaks through the limitations of traditional one-dimensional time series modeling, so as to model the change patterns within and between periods at the same time, thereby significantly improving the modeling capability for permanent magnet synchronous motor temperature estimation. The inventors found that the above-mentioned multi-scale and multi-period features are highly consistent with the modeling idea of TimesNet extracting multi-periods and folding into 2D, so that the TimesNet model has high applicability and unique advantages in the field of temperature estimation of the permanent magnet synchronous motor, which helps to capture both instantaneous and trend changes.

[0096] The specific network structure and configuration of TimesNet include the input layer, the number of TimesBlock layers, the convolution layer configuration, the fully connected layer, and the output layer. The main working process of TimesNet is as shown in Figure 5 The TimesNET data-driven model constructed in the application specifically includes the following steps S210-S240:

[0097] Step S210: Extraction of multi-scale features: In the above formulas 1-6, it has been proved that the input signals of the motor operating conditions d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n have obvious multi-scale features. The TimesNet data-driven model analyzes the frequency spectrum of the input signal by fast Fourier transform (FFT), identifies the main period length, and automatically extracts high-frequency-voltage current and low-frequency-speed dual time scales.

[0098] In another embodiment, step S210 specifically includes the following calculation process:

[0099] The input signal selected by step S1 constructs the input signal matrix X

[0100]

[0101] The fast Fourier transform (FFT) is performed on the time series to obtain the amplitude A motor ∈R T corresponding to each frequency:

[0102]

[0103] where FFT(·) denotes the fast Fourier transform, Amp(·) denotes the amplitude calculation, and Avg(·) denotes the averaging calculation.

[0104] To avoid the noise caused by meaningless high frequencies, the first k largest amplitudes are selected, thereby obtaining the most significant frequency set {f1 motor ,…,f k motor}:

[0105]

[0106] where denotes the key frequency set , whose four frequency parameters are the reciprocals of the electrical period T e , the mechanical period T m , the switching period T w , and the thermal time period T th when the permanent magnet synchronous motor operates in multiple periodic modes.

[0107] Based on the selected most significant frequency set {f1 motor ,…,f k motor}, the corresponding k period lengths {p1,…,p k} are determined as:

[0108]

[0109] Step S220: Conversion of two-dimensional tensors: the one-dimensional time series X 1D motor ∈R T×4 is reconstructed into multiple two-dimensional tensors by formula (11):

[0110]

[0111] where the Padding(·) function performs zero extension of the time series along the time dimension to make it compatible with the matrix reconstruction function based on the number of elements unchanged, where the period length p i is the number of rows of the transformed two-dimensional tensor, representing the change within a period and capturing the fine dynamics within a complete operation period; is the number of columns of the transformed two-dimensional tensor, representing the evolution trend between adjacent periods, and the two-dimensional tensor denotes the frequency-based representation the jth reconstructed time series.

[0112] Step S230: Extraction of two-dimensional convolution features: for the two-dimensional tensor , a multi-scale thermal perception convolution kernel is constructed using the efficient Inception convolution module combined with the thermal response time constant , which dynamically adjusts the feature extraction process according to the current temperature state. This module can learn the typical features of each operating condition of the motor (such as smooth running or rapid acceleration loading operation), as well as the common patterns and subtle differences between these features, which is represented as:

[0113]

[0114] The multi-scale thermal response convolution is represented as:

[0115]

[0116] where the thermal state vector , represents the thermal response time function, denotes the thermal perception convolution kernel of the st scale, and each convolution kernel size is associated with the thermal time constant , is f i motor the corresponding ith cycle length, is the interval between adjacent two sampling periods, and the large kernel size (5x5, 7x7) corresponds to the long time constant (slow thermal characteristics) of the motor / thermal system, and the small kernel size (1x1, 3x3) corresponds to the short time constant (current pulsation), is a scalar modulation factor in the range [0, 1].

[0117] Step S240: Conversion of two-dimensional tensor to one-dimensional space: map the extracted two-dimensional features back to one-dimensional space for subsequent information fusion; the conversion process is as follows:

[0118]

[0119] where TempTrunc(·) removes the zeros added by the Padding(·) operation in equation (4) to restore the feature sequence to the original time step.

[0120] To fuse the features extracted from different cycle scales, the one-dimensional features output by (14) are weighted and summed according to the amplitude spectrum intensity of each cycle component, as follows:

[0121]

[0122]

[0123] wherein, is the frequency f i motor The corresponding amplitude A motor , by stacking multiple TimesBlock (s) described above (three in the Figure 5 , can be set to other quantities according to actual conditions), the model can capture the two-dimensional variation pattern of the multi-time scale input signal (d-axis current i d , dq-axis voltage u dq , motor speed n) in layers, which can extract more rich and effective feature information compared with directly modeling on one-dimensional sequence.

[0124] Step S3: Based on different working conditions and different motors, a transfer learning framework is constructed and combined with the TimesNET data-driven model of step S2, and a permanent magnet synchronous motor temperature estimation model of the source domain is trained using strongly correlated and easily collected input signals.

[0125] Specifically, in step S3, because different working conditions and different motors are considered, the rationality and difficulty of data acquisition need to be considered in actual application: assuming that A motor a working condition data is easy to collect, and A motor b working condition and B motor c working condition data are difficult to collect due to complex working conditions and experimental condition limits. Therefore, the number of A motor a working condition data is taken as the source domain data, the A motor b working condition and B motor c working condition data are taken as the target domain, and the temperature transfer scenario condition is set as the source domain transferring to the target domain.

[0126] It is worth mentioning that, for example, the source domain data such as A motor a working condition data only needs to come from one controllable and simple working condition, which can provide reliable feature reference for the model. This patent emphasizes the small sample learning ability and transfer learning generalization ability across working conditions and across models of the algorithm, and the algorithm can transfer this most basic knowledge to different motors and complex working conditions to realize the cross-domain generalization of temperature estimation. Therefore, this method has significant data efficiency and cross-scene adaptability, and even in the case of only the simplest source domain working condition data, it can still accurately predict the temperature change under complex working conditions.

[0127] According to the A motor a working condition d-axis current i d , d-axis voltage u d , q-axis voltage u qand the input signal of motor speed n and motor stator temperature together as a training set to pre-train a source domain permanent magnet synchronous motor temperature estimation model based on TimesNET data-driven model and transfer learning framework as a basic stator temperature estimation model (stator temperature estimation source model).

[0128] According to the training of the rotor temperature estimation source model, the d-axis current i d , d-axis voltage u d , q-axis voltage u q and motor speed n and motor rotor temperature together as a training set to pre-train a source domain permanent magnet synchronous motor temperature estimation model based on TimesNET data-driven model and transfer learning framework as a basic rotor temperature estimation model (rotor temperature estimation source model).

[0129] It should be pointed out that TimesNet data-driven model is also a temperature estimation model, but because it does not have the temperature estimation ability across conditions and across models, it needs to be combined with the transfer learning framework to form a permanent magnet synchronous motor temperature estimation model to train the source domain and the target domain.

[0130] Step S4: Using fine-tuning strategy, training the permanent magnet synchronous motor temperature estimation model of the target domain, verifying the model accuracy to finally form the target domain temperature estimation model based on input signal and output signal, and realizing the training of the target domain temperature estimation model across conditions and the training of the target domain temperature estimation model across motors.

[0131] Specifically, in step S4, a new TimesNET data-driven model is initialized as a target model in the target domain, and its parameters are copied from the source model pre-trained in step S3. The first 30% of the target domain original data divided in step S3 is used as the target domain training set, and a fine-tuning strategy is used to obtain the permanent magnet synchronous motor temperature estimation model of the target domain based on TimesNET data-driven model and transfer learning framework with smaller learning rate and fewer training rounds. During the fine-tuning process, all parameters outside the other layers of TimesNet data-driven model are frozen, only the prediction layer parameters are updated, including reducing the learning rate and the number of iterations, the learning rate is reduced from 0.001 to 0.0001, and the number of training rounds is reduced by 50%, further reducing the fine-tuning calculation overhead. Finally, the permanent magnet synchronous motor temperature estimation model of the target domain that meets the accuracy verification condition is used as the target domain temperature estimation model.

[0132] 1. Training of the target domain temperature estimation model across conditions: A motor a condition → A motor b condition

[0133] On the basis of the stator temperature estimation model trained in step S3, the target domain motor stator temperature estimation model is obtained by fine-tuning strategy using the target domain data (d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n and motor stator temperature data) in the first 30%.

[0134] The target domain motor rotor temperature estimation model is obtained as above. On the basis of the rotor temperature estimation model trained in step S3, the target domain motor rotor temperature estimation model is obtained by fine-tuning strategy using the target domain data (d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n and motor rotor temperature data) in the first 30%.

[0135] 2. Training of cross-motor target domain temperature estimation model: A motor a condition → B motor c condition

[0136] On the basis of the stator temperature estimation model trained in step S3, the target domain motor stator temperature estimation model is obtained by fine-tuning strategy using the target domain data (d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n and motor stator temperature data) in the first 30%.

[0137] The target domain motor rotor temperature estimation model is obtained as above. On the basis of the rotor temperature estimation model trained in step S3, the target domain motor rotor temperature estimation model is obtained by fine-tuning strategy using the target domain data (d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n and motor rotor temperature data) in the first 30%.

[0138] Step S5: Using the target domain temperature estimation model obtained in step S4 to realize online estimation of rotor permanent magnet temperature θ pm and stator winding temperature θ sw , the remaining 70% of the original data in the target domain under cross-condition cross-model are used to estimate the motor stator / rotor temperature.

[0139] In step S5 of the present application, taking the stator temperature estimation simulation result as an example, the simulation experiment result is shown in Figs. 6-7. As shown in the figures, there is an obvious difference in the change trend of the stator temperature of the A motor b working condition and the B motor c working condition, but the stator temperature estimation model can obtain accurate estimation result. The stator temperature estimation error range of the A motor b working condition is-5.27-2.13℃, and the stator temperature estimation error range of the B motor c working condition is-1.25-1.93℃. Therefore, the generalization of the temperature estimation model is proved, and the method can be effectively applied to the temperature estimation across working conditions and across motors to meet the demand of the same motor full working condition coverage and the direct transplantation of the temperature estimation model between different motors in engineering application.

[0140] The motor temperature estimation method based on data driving and transfer learning provided in the present application firstly extracts the d-axis current i d , the dq-axis voltage u dq and the speed signal n which are strongly related to the temperature of the permanent magnet synchronous motor as the model input, avoids input redundancy and reduces the model calculation amount. Then, the TimesNET data driving module and the transfer learning framework are constructed, the target domain TimesNET temperature estimation model is obtained by using the source domain data and the target domain data through the fine-tuning strategy. Finally, the generalization performance of the model is verified under the same motor different running conditions and different motor different running conditions, and the method can be deployed in the modern motor control system to monitor the temperature state of the motor.

[0141] Finally, it should be noted that each step S1-S5 of the above-mentioned temperature online estimation method can be converted into software program instructions, which can be implemented by using a system including a processor and a memory, or by using computer instructions stored in a non-transitory computer readable storage medium. The above-mentioned software function unit is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0142] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data-driven online estimation method of motor stator and rotor temperature, characterized in that, Comprising the following steps: Step S1: using a correlation analysis method, extracting input signals strongly related to the temperature of the permanent magnet synchronous motor for temperature estimation, the input signals of the permanent magnet synchronous motor temperature estimation model include d-axis current i d , d-axis voltage u d , q-axis voltage u q , and motor speed n; the output signals include rotor permanent magnet temperature θ pm and stator winding temperature θ sw ; Step S2: For the input signal in step S1, a TimesNET data-driven model is constructed, and the extraction of multi-scale features, the conversion of two-dimensional tensors, the extraction of two-dimensional convolution features, and the conversion of two-dimensional tensors to one-dimensional space are sequentially completed, so as to fold the one-dimensional time series of the input signal according to the detected main period length to form multiple two-dimensional tensors, and finally convert the two-dimensional tensors to one-dimensional tensors; Step S3: Based on different working conditions and different motors, a transfer learning framework is constructed and combined with the TimesNET data-driven model of step S2, and a strongly correlated and easily collected input signal is used to train a permanent magnet synchronous motor temperature estimation model of the source domain; Step S4: A fine-tuning strategy is used to train the permanent magnet synchronous motor temperature estimation model of the target domain, and after verifying the model accuracy, the target domain temperature estimation model based on the input signal and the output signal is finally formed, and the training of the target domain temperature estimation model across working conditions and the training of the target domain temperature estimation model across motors are realized; Step S5: Realize the online estimation of the rotor permanent magnet temperature θ pm and the stator winding temperature θ sw of the target domain based on the target domain temperature estimation model obtained in step S4, and estimate the temperatures of the motor stator and rotor based on the remaining 70% of the original data across working conditions and across models.

2. The data-driven motor stator and rotor temperature online estimation method of claim 1, wherein, Step S2 specifically comprises the following step S210: Step S210: Extraction of multi-scale features: In the known motor operating conditions, the input signals of d-axis current i d , d-axis voltage u d , q-axis voltage u q , motor speed n have obvious multi-scale characteristics. In this case, the TimesNet data-driven model analyzes the frequency spectrum of the input signal by fast Fourier transform (FFT), identifies its main period length, and automatically extracts high-frequency-voltage current, low-frequency-speed dual time scale. The input signal matrix is constructed from the input signal selected in step S1 (7) Fast Fourier transform (FFT) is performed on the time series to obtain the amplitude A corresponding to each frequency motor ∈R T is: where FFT( ) denotes a fast Fourier transform, Amp( ) denotes a calculation of amplitude, and Avg( ) denotes an averaging calculation. The first k largest amplitude frequencies are selected, thereby obtaining the most significant frequency set {f1 motor ,…,f k motor} wherein, denotes the set of key frequencies based on the selected set of most significant frequencies {f1 motor ,…,f k motor}, determine their corresponding k periodicity lengths {p1, …, p k} as: its four frequency parameters the electrical period T of the various periodic modes of operation of the permanent magnet synchronous machine e , the mechanical period T m , the switching period T w , the thermal time period T th in inverse.

3. The data-driven motor stator and rotor temperature online estimation method of claim 2, wherein, Step S2 specifically further comprises the following step S220: Step S220: Conversion of the two-dimensional tensor: The one-dimensional time series X 1D motor ∈R T×4 reconstructed into multiple two-dimensional tensors by formula (11): where the Padding(·) function pads the time series with zeros along the time dimension to make it compatible with the matrix reconstruction function based on the invariable number of elements where p is the period length i as the number of rows of the transformed two-dimensional tensor, represents the variation within a period, capturing the fine dynamics within a complete operation period; as the number of columns of the transformed two-dimensional tensor, represents the evolution trend between adjacent periods, the two-dimensional tensor represents the jth reconstructed time series based on the frequency .

4. The data-driven motor stator and rotor temperature online estimation method of claim 3, wherein, Step S2 specifically further comprises the following step S230: Step S230: Extraction of two-dimensional convolution features: for two-dimensional tensor , using an efficient Inception convolution module combined with a thermal response time constant to form a multi-scale thermal perception convolution kernel , dynamically adjusting the feature extraction process according to the current temperature state. This module can learn the typical features of each operating condition of the motor (such as smooth running or sudden acceleration loading operation), as well as the common patterns and subtle differences between these features. The process is represented as: The multi-scale heat response convolution is represented as: where the thermal state vector , represents the thermal response time function, denotes the thermal perception convolution kernel of the s-th scale, each kernel size associated with the thermal time constant , f i motor the corresponding i-th cycle length, is the interval between two adjacent sampling periods, the large kernel size 5x5 and 7x7 correspond to the long time constant of the motor / thermal system, and the small kernel size lxl and 3x3 correspond to the short time constant, is a scalar modulation factor, ranging in [0, 1].

5. The data-driven motor stator and rotor temperature online estimation method of claim 4, wherein, Step S2 specifically further comprises the following step S240: Step S240: Conversion of two-dimensional tensor to one-dimensional space: The extracted two-dimensional features are mapped back to one-dimensional space to facilitate subsequent information fusion; the conversion process is as follows: where TempTrunc( ) removes the zeros added by the Padding( ) operation in Equation (4) to restore the feature sequence to the original time steps; According to the amplitude spectrum intensity corresponding to each periodic component, the one-dimensional features output by formula (14) are weighted and summed, and the process is as follows: in, For frequency f i motor The corresponding amplitude A motor By stacking multiple TimesBlocks, the model can capture the two-dimensional variation patterns of input signals at multiple time scales in a hierarchical manner, which can extract richer and more effective feature information compared to modeling directly on a one-dimensional sequence.

6. The data-driven motor stator and rotor temperature online estimation method of claim 1, wherein, In step S3, the rationality and difficulty of data acquisition need to be considered: assuming that A motor a working condition data is easy to collect, while A motor b working condition and B motor c working condition data are difficult to collect due to complex working conditions and experimental condition limits, therefore, the A motor a working condition data is taken as the source domain data, the A motor b working condition and B motor c working condition data are taken as the target domain, and the temperature transfer scenario condition is set as the source domain transferring to the target domain; The d-axis current i extracted according to step S1 is strongly related to the stator winding temperature under the a operating condition of motor A d , d-axis voltage u d , q-axis voltage u q and the input signal of motor speed n and stator winding temperature θ sw Together as a training set, a source domain permanent magnet synchronous motor temperature estimation model based on TimesNET data driven model and transfer learning framework is pre-trained as a basic stator temperature estimation model; The training of the rotor temperature estimation source model is the same as the extraction of the d-axis current i under the a working condition of the motor A which is strongly related to the rotor permanent magnet temperature according to step S1 d , the d-axis voltage u d , the q-axis voltage u q , and the motor speed n and the rotor permanent magnet temperature θ pm Together, a TimesNET data-driven model-based source domain permanent magnet synchronous motor temperature estimation model based on a transfer learning framework is pre-trained as a basic rotor temperature estimation model.

7. The data-driven motor stator and rotor temperature online estimation method of claim 6, wherein, In step S4, a new TimesNET data-driven model is initialized as a target model in the target domain, the parameters of which are copied from the pre-trained source model in step S3, the first 30% of the target domain original data divided in step S3 is used as the target domain training set, and a fine-tuning strategy is used to obtain the target domain TimesNET temperature estimation model with a smaller learning rate and fewer training rounds. During the fine-tuning process, all parameters outside the other layers of the TimesNet data-driven model are frozen, only the prediction layer parameters are updated, including reducing the learning rate and the number of iterations, the learning rate is reduced from 0.001 to 0.0001, and the number of training rounds is reduced by 50%.

8. The data-driven motor stator and rotor temperature online estimation method of claim 7, wherein, In step S4, it also includes: Training of the target domain temperature estimation model across working conditions: A motor a working condition → A motor b working condition; Based on the stator temperature estimation model trained in step S3, the target domain motor stator temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy; Based on the rotor temperature estimation model trained in step S3, the target domain motor rotor temperature estimation model is obtained by using the first 30% of the target domain data through the fine-tuning strategy; Training of the target domain temperature estimation model across motors: A motor a working condition → B motor c working condition On the basis of the stator temperature estimation model trained in step S3, the first 30% of the target domain data is used to obtain the target domain motor stator temperature estimation model through fine-tuning strategy; On the basis of the rotor temperature estimation model trained in step S3, the first 30% of the target domain data is used to obtain the target domain motor rotor temperature estimation model through fine-tuning strategy.

9. A data-driven online estimation system of motor stator and rotor temperature, characterized in that, Comprise: At least one processor; And at least one memory connected in communication with the processor, wherein: The memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the data-driven motor stator and rotor temperature online estimation method according to any one of claims 1 to 8.

Images