A mechanism-data hybrid driving-based motor stator temperature field reduction method

Abstract

The application discloses a motor stator temperature field reduction method based on mechanism-data hybrid driving, and belongs to the field of motor equipment digital twinning and intelligent operation and maintenance. The method establishes a close coupling between physical mechanism and data driving through a four-stage progressive technical path of "regional gridding dimension reduction-loss tracing parameter modeling-Gaussian process sparse correction-IDW spatial reconstruction", and can realize the reconstruction of low-dimensional grid temperature to high-dimensional full-field temperature distribution. The application deeply integrates physical mechanism and data driving, and the time consumption of single inference is in the order of milliseconds. After verification of 8 actual running main motors, the average absolute percentage error after correction is controlled within 1.5%, realizing the rapid prediction of motor stator temperature field in a wide speed regulation range, the full-field adaptive correction driven by sparse measured data, and the high-fidelity spatial temperature distribution reconstruction, thereby providing technical support for real-time temperature monitoring and insulation life assessment of the motor digital twinning system.

Description

Technical Field

[0001] This invention relates to the field of digital twins and intelligent operation and maintenance of motor equipment, and in particular to a method for reducing the stator temperature field of a wide-range variable frequency synchronous motor based on mechanism-data hybrid drive. Background Technology

[0002] Large variable frequency synchronous motors, as the core drive equipment of water pump units, are widely used in water conservancy projects. Given the high complexity and dynamic variability of water pump operation, the main motor must possess a wide speed range to adapt to load fluctuations. This type of motor has a highly complex internal structure, integrating key components such as the stator core, rotor body, excitation windings, and multi-layered ventilation channels. Traditional single-point temperature measurement techniques (such as thermal resistors embedded in bearings or stator windings) can only capture localized, discrete temperature data, making it difficult to reconstruct the continuous thermal distribution characteristics within the motor. This results in the inability to accurately capture the location and evolution of "hot spots" in the core area. Therefore, constructing a high-precision coupled model of the main motor's temperature and flow fields, and comprehensively analyzing its internal thermal-fluid physical mechanisms, is the core path to breaking the "black box" state of the digital twin main motor. This approach not only provides theoretical support for real-time visualization of the thermal / flow field state but is also key to realizing full life-cycle thermal safety monitoring and intelligent operation and maintenance of the motor.

[0003] For the calculation of temperature field twins in main motors, common methods are mainly divided into two categories: theoretical methods and data-driven methods. The most common theoretical calculation method is finite element method (FEM). FEM offers small calculation errors and rich detail, but it is highly dependent on modeling parameters. For fluid simulations, it suffers from high computational costs, lack of universality and poor generalization ability, and difficulty in lightweighting. In contrast, data-driven methods exclude ventilation and heat dissipation theory and derive regression relationships based solely on existing data. They require a large amount of data, are typically trained once, and are easy to call up after subsequent regressions, resulting in low usage costs. However, they suffer from key problems such as a lack of mechanistic support and poor interpretability. In cutting-edge research areas, researchers have mainly used data-driven methods for temperature field order reduction in main motors in recent years, with relatively little research on model (mechanism)-data hybrid-driven methods for temperature field order reduction.

[0004] For the construction of digital twins of the temperature field of the main motor, existing calculation methods are mainly divided into two paradigms: theoretical mechanism-driven and data-driven.

[0005] Theoretical mechanism-driven methods, represented by finite element method (FEM) simulation, offer clear physical mechanisms, high computational accuracy, and the ability to reveal rich field distribution details. However, they are highly dependent on precise modeling parameters and boundary conditions. Especially when dealing with complex fluid coupling simulations, they face significant bottlenecks such as enormous computational time, weak model generalization ability, and difficulties in lightweight deployment, making it difficult to meet real-time requirements.

[0006] Data-driven approaches: These methods abandon explicit modeling of the physical mechanisms of ventilation and heat dissipation, instead constructing a regression mapping relationship between inputs and outputs based on historical operational data. Their advantages lie in fast inference speed and low call cost after training; however, their disadvantages are equally obvious: they are highly dependent on the scale and quality of data, and due to the lack of physical mechanism support, the model has poor interpretability and insufficient robustness when extrapolating operating conditions.

[0007] Looking at the forefront of scientific research, existing studies mostly focus on single data-driven reduction models, aiming to improve computational efficiency. However, research on hybrid-driven temperature field reduction methods that integrate physical mechanisms and data-driven approaches is still lacking. How to break through the limitations of a single paradigm and construct a hybrid model that combines physical interpretability and data adaptability is a key scientific problem that urgently needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies. For variable frequency synchronous motors with a wide speed range, in order to take into account the advantages of classical ventilation and heat dissipation theory, traditional theoretical simulation methods and data-driven methods, as well as the requirement of fast calculation, this invention provides a mechanism-data hybrid driven method for reducing the order of the stator temperature field of variable frequency synchronous motors with a wide speed range. This method achieves lightweight reduction and reconstruction of the stator temperature field of variable frequency synchronous motors with a wide speed range and real-time unreachable mapping.

[0009] The technical solution to achieve the above objective is: a method for reducing the order of the motor stator temperature field based on mechanism-data hybrid drive, comprising the following steps: S1, Discretize the space into a grid according to the actual motor operation: Perform grid clustering on all temperature nodes of the motor stator finite element model according to spatial location, material properties or functional partitions, divide into N grid regions, and use the average temperature of nodes in each region as the characteristic temperature of the region to compress the high-dimensional temperature field into an N-dimensional feature vector. S2, the least squares-driven temperature reduction modeling steps based on ventilation mechanism: Based on the simulation data of multi-condition parameter scanning, with various losses and cooling air volume as input, the system identification or regression method is used to establish an explicit reduction formula between the characteristic temperature and the input parameters for each region, and obtain a fast prediction model (also known as the price reduction model) of the characteristic temperature of N grid regions. S3, Global Correction Mapping Step Driven by Joint Features of Working Condition and Space: In real-time operation, acquire a limited number of measured temperature measurement points, calculate the residual between the measured value and the predicted value of the fast prediction model in step S2, model the spatial correlation of the residual using a Gaussian process, estimate the residual correction value of all N grid areas based on the squared exponential kernel function and through the posterior inference formula, and superimpose the correction value with the predicted value to complete the global adaptive correction. S4, Spatial reconstruction steps based on sparse block support by inverse distance weighted interpolation: The corrected N grid region feature temperatures are used as sparse control points, and the inverse distance weighted interpolation method is used to perform spatial interpolation on all finite element nodes to realize the reconstruction of the low-dimensional grid temperature to the high-dimensional full-field temperature distribution.

[0010] In the above-mentioned mechanism-data hybrid drive-based motor stator temperature field order reduction method, in step S1, the value of N is 38, and the spatial location includes stator teeth, stator yoke, winding ends and air domain.

[0011] In the above-mentioned mechanism-data hybrid drive-based motor stator temperature field order reduction method, step S2 includes various losses such as stator and rotor copper loss, total iron loss, mechanical loss and stray loss.

[0012] In the above-mentioned mechanism-data hybrid driving method for reducing the order of motor stator temperature field, in step S2, the regression method adopts the least squares method.

[0013] In the above-mentioned mechanism-data hybrid drive-based motor stator temperature field order reduction method, in step S2, the range of the multi-condition parameter scanning is: frequency 25~52.5Hz, voltage 5000~10500V, sampling intervals of 5Hz and 500V respectively, forming more than 50 sets of operating conditions.

[0014] In the aforementioned mechanism-data hybrid driving method for reducing the order of motor stator temperature field, step S3, the quadratic exponential kernel function is: ; In the formula, , Let G be the coordinates of two points in space. For signal variance, This is the length scale parameter.

[0015] In the aforementioned mechanism-data hybrid driving method for reducing the order of motor stator temperature field, step S3, the posterior inference formula is:

[0016] In the formula, Points to be predicted The residual estimate at the location, K The covariance matrix of the training points, Let be the covariance vector between the test points and the training points. for transpose, For noise variance, It is the identity matrix. r This is the measured residual vector; It is the inverse of the covariance matrix.

[0017] In the aforementioned mechanism-data hybrid driving method for reducing the order of the motor stator temperature field, step S3, the formula for global adaptive correction is as follows: ; In the formula, For the first i The final temperature prediction value after correction for each region For the first i Initial temperature forecast values ​​for each region For the first i The residual correction value for the entire region.

[0018] In the aforementioned mechanism-data hybrid driving method for reducing the order of motor stator temperature field, step S4, the formula for the inverse distance weighted interpolation is: ; In the formula, Points to be interpolated The final temperature value at that location, For the spatial points to be interpolated, i The area is numbered, with a total of 38 grid areas; For the first i The weight coefficient of each region is determined by the distance from the interpolation point to the representative point of that region; the closer the distance, the greater the weight. For the first i The final temperature prediction values ​​after correction for each region; Weight The calculation formula is: ; In the formula, Let be the Euclidean distance from the node to be interpolated to the representative point of the i-th region; q The distance exponent is 2; distance The calculation formula is: ; In the formula, Points to be interpolated To the i The Euclidean distance of each region's representative point. x and y Points to be interpolated coordinates x i and y i For the first i Each grid area represents the coordinates of a point.

[0019] In the above-mentioned mechanism-data hybrid driving method for reducing the temperature field of a motor stator, in step S4, the number of nodes in the reconstructed temperature field after interpolation is more than 6,000, covering the full field temperature distribution of the two ends and the middle section of the motor stator.

[0020] The mechanism-data hybrid drive-based method for reducing the order of motor stator temperature field of the present invention has the following advantages compared with the prior art: (1) This invention provides a lightweight, order-reduced reconstruction calculation method for the stator temperature field of a variable frequency synchronous motor that combines theoretical support with the advantages of artificial intelligence. It can quickly and relatively accurately calculate the stator temperature field of the variable frequency synchronous motor under a wide speed range, and can adaptively correct the global temperature field based on the measured data of individual nodes, realizing the global unreachable mapping of the stator temperature field and digital twin feedback. Compared with directly using machine learning or POD models for end-to-end prediction of the temperature field, the architecture proposed in this invention establishes a tighter coupling between physical mechanisms and data-driven approaches. This architecture not only profoundly reveals the inherent physical mapping mechanism between the motor stator temperature and various motor losses, speed / airflow, and significantly enhances the robustness and interpretability of the model under complex operating conditions, but also compresses the time consumed by a single inference to the millisecond level, perfectly meeting the stringent requirements of digital twin systems for real-time response.

[0021] (2) This invention provides a key technical path for constructing a twin model of the factory condition mechanism of a wide speed range variable frequency synchronous motor, and establishes a new paradigm of stator temperature field twinning based on mechanism-data hybrid drive. By observing the overall twin temperature of the stator temperature field, the highest temperature and dangerous location are determined, thus laying a solid technical foundation for real-time temperature monitoring and insulation life assessment of this type of motor, and further laying a solid theoretical and engineering foundation for intelligent operation and maintenance of this type of motor throughout its entire life cycle. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall invocation process of the method of this invention. Figure 2 This is a flowchart illustrating the step-by-step implementation of the method of the present invention; Figure 3 This is a schematic diagram of the finite element model of a single slot stator at section B in an embodiment of the present invention; Figure 4 This is a schematic diagram of the grid region division at section A in an embodiment of the present invention; Figure 5 This is a comparison curve of the temperature twin value of the main motor measuring point before and after correction in an embodiment of the present invention; Figure 6 This is a comparison curve of the temperature twin values ​​of the main motor at measuring point 1 before and after correction in an embodiment of the present invention; Figure 7This is a diagram showing the error distribution before and after correction of the temperature twin values ​​of one or two measuring points of the main motor in an embodiment of the present invention; Figure 8 This is a reconstructed contour map of the temperature field at section B in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0026] Please see Figure 1 and Figure 2 According to an embodiment of the present invention, a method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive, the key technologies mainly include: least squares identification technology of grid temperature reduction formula based on regional gridding and loss tracing, global correction mapping based on joint feature drive of operating condition and space, and calculation method of inverse distance weighted interpolation space reconstruction based on sparse block support.

[0027] 1. Least squares identification technique for grid temperature reduction formula based on regional rasterization and loss tracing: To address the challenges of high complexity and massive data volume in real-time temperature field simulation of large-scale finite element models, a region-partitioning-based method for dimensionality reduction and reconstruction is proposed. This method employs a "regional rasterization dimensionality reduction - parameter correlation modeling" approach, which significantly reduces computational and storage overhead while preserving the key characteristics of the physical field, and achieves an effective mapping from macroscopic features to microscopic distributions.

[0028] First, based on the spatial location, material properties, or functional partitioning of the nodes, the original 6000+ finite element scan temperature nodes are spatially clustered into 38 grid regions (logical regions). Each grid region is considered a basic unit, and its thermal state is characterized by the average temperature of all nodes within the region, defined as the "characteristic temperature" of that region. This step compresses the high-dimensional node temperature data (commonly ranging from several thousand to tens of thousands of dimensions) into a low-dimensional feature vector (38 dimensions), completing the first effective dimensionality reduction of the temperature field.

[0029] To establish a rapid mapping relationship between system operating parameters and regional temperatures, parametric modeling of each region is necessary. Based on historical operating data or computational analysis, system identification or regression analysis methods are used to establish a mathematical relationship model for each region. This model takes key operating parameters (such as voltage, current, and rotational speed) as input, obtains the loss values ​​of each region according to empirical formulas, and outputs the characteristic temperature of the region—a reduced-order formula. Ultimately, 38 specialized formulas characterizing the thermal response of each region are obtained. Using this set of formulas, the characteristic temperatures of the 38 regions can be directly and quickly derived from real-time operating parameters, thus avoiding complex and time-consuming global finite element calculations and achieving efficient temperature field prediction.

[0030] This method has the following characteristics: (1) High efficiency: Through rasterization dimensionality reduction and formulaic modeling, the computational efficiency is greatly improved and the data storage requirements are reduced.

[0031] (2) Physical interpretability: The regional division is related to physical parameters or actual operating conditions, which makes the reduced-order model have clear physical meaning and facilitates engineering analysis and application.

[0032] 2. Global correction mapping driven by joint features of working conditions and space: Based on actual operating data of variable frequency motors collected under different working conditions, an artificial intelligence approach is used to establish a complex mapping relationship between the factory-corrected simulated temperature rise value and the actual operating value. A Gaussian process algorithm based on Bayesian theory is employed to continuously and adaptively adjust the established mapping relationship based on subsequently updated operating data. This enables an adaptive correction method for the digital twin main motor temperature real-time monitoring and insulation life real-time assessment model based on measured data.

[0033] Building upon the rapid temperature field prediction provided by the previous regional order reduction model, an innovative adaptive correction mechanism based on Gaussian process and spatial correlation propagation is introduced. This mechanism aims to significantly improve the accuracy and reliability of the overall temperature prediction using only a small amount of field measurement data.

[0034] The core technology lies in treating the prediction residual (the difference between the measured temperature and the model's predicted temperature) between two regions with measured values ​​as a spatial signal containing model system bias and unknown interference. This residual signal is learned and modeled by establishing a Gaussian process model. As a non-parametric Bayesian model, the core of the Gaussian process lies in a kernel function that defines spatial correlation. For two points in space... The physical intuition that “the closer the locations are, the more similar the residuals” can be naturally represented by the following quadratic exponential kernel function.

[0035] The quadratic exponential kernel function is: ; In the formula, , Let G be the coordinates of two points in space. For signal variance, This is the length scale parameter.

[0036] Based on this, the model can optimally estimate any unmeasured point in space using only two measured residual points and the following posterior inference formula. The residuals. This process not only provides the best estimate of the residuals, but the derived prediction variance also quantifies the uncertainty of the estimate.

[0037] The posterior inference formula is:

[0038] In the formula, Points to be predicted The residual estimate at the location, K The covariance matrix of the training points, Let be the covariance vector between the test points and the training points. for transpose, For noise variance, It is the identity matrix. r This is the measured residual vector; It is the inverse of the covariance matrix.

[0039] Subsequently, the calculated global residual correction field is used. Compared with the initial predicted temperature field Superposition, i.e., completes adaptive correction.

[0040] The formula for global adaptive correction is:

[0041] In the formula, For the first i The final temperature prediction value after correction for each region For the firsti Initial temperature forecast values ​​for each region For the first i The residual correction value for the entire region.

[0042] This step cleverly combines data-driven bias correction with mechanism-driven initial prediction, enabling adaptive model updates.

[0043] 3. A computational method for inverse distance weighted interpolation space reconstruction based on sparse block support: Finally, to obtain a high-resolution full-field temperature distribution, the corrected 38 regional characteristic temperatures were treated as control points, and spatial reconstruction was performed using inverse distance weighted interpolation. For any node in the entire field... The formula for inverse distance weighted interpolation of its temperature is: ; In the formula, Points to be interpolated The final temperature value at that location, For the spatial points to be interpolated, i The area is numbered, with a total of 38 grid areas; For the first i The weight coefficient of each region is determined by the distance from the interpolation point to the representative point of that region; the closer the distance, the greater the weight. For the first i The final temperature prediction value after correction for each region.

[0044] Weight The calculation formula is: ; In the formula, Let be the Euclidean distance from the node to be interpolated to the representative point of the i-th region; q The distance power exponent. The value can be set to indicate how quickly the weight decreases as distance increases; a value of 2 is recommended.

[0045] distance The calculation formula is: ; In the formula, Points to be interpolated To the i The Euclidean distance of each region's representative point. x and y Points to be interpolated coordinates x i and y i For the first i Each grid area represents the coordinates of a point.

[0046] This method ensures that the temperature field is spatially continuous and rationally distributed.

[0047] The entire method constitutes a complete technical closed loop of "rapid prediction by physical model - sparse data-driven correction - full-field reconstruction of spatial correlation". It ensures computational efficiency with front-end order reduction model and improves prediction accuracy and robustness under real complex working conditions with back-end Gaussian process adaptive correction, providing an effective hybrid modeling approach for solving the problem of "limited model accuracy" and "scarcity of measured data" in engineering.

[0048] Explanation of key steps: (1) Discrete measurement point aggregation: Based on the material properties or functional zones where the simulation nodes are located, discrete simulation points are aggregated into an initial temperature field through statistical rasterization (i.e., from tens of thousands of nodes in the simulation output to several raster regions). (2) Reduction modeling of mechanism constraints: Based on simulation data, the least squares identification is used to realize grid reduction and temperature modeling of operating condition response under ventilation mechanism constraints (the above grid region formula identification). (3) Multi-feature global correction: integrate working condition features and spatial coordinates, and perform global correction based on sparse measured grid (a limited number of temperature measurement points are corrected to the entire grid area). (4) Sparse dimensionality upscaling: Based on sparse block-supported IDW (Inverse Distance Weighting) interpolation, high-fidelity dimensionality upscaling reconstruction of low-resolution physical fields is achieved (extending the above grid area to tens of thousands of temperature points, global interpolation).

[0049] This invention selects a core piece of equipment in a pumping station—the TLKS9000-24 series vertical variable frequency synchronous motor manufactured by Shanghai Electric Machinery Plant—as the empirical object. Relying on high-precision operating data collected by an online monitoring platform, this invention applies the proposed innovative method for empirical analysis, and systematically compares the calculation results with temperature data corresponding to two locations—the motor stator core and windings—collected by the monitoring platform. Through multi-dimensional verification analysis, the aim is to quantitatively evaluate the advantages of this method in terms of fault feature extraction accuracy, response timeliness, and engineering applicability, thereby verifying the effectiveness and reliability of the technical solution.

[0050] The core basic parameters of the main motor are detailed in the table below.

[0051] Table 1. Main foundation parameters of the pump station's main motor:

[0052] This invention selects the original operating data of the motor from June 2024 to August 2025 as the analysis sample. To ensure the validity and typicality of the data, the original dataset was rigorously cleaned.

[0053] according to Figure 2 The flowchart illustrates the step-by-step implementation of the method. The implementation process of order reduction of the motor stator temperature field in the example is as follows: 1. CFD-based multi-condition temperature field parameter scanning within the operating range: Taking this type of motor as an example, its frequency conversion operating range is 25~52.5Hz, and its voltage operating range is 5000~10500V. With 5Hz and 500V as sampling intervals, and extreme conditions such as oversaturation being filtered out, more than 50 sets of operating conditions are obtained.

[0054] Considering the stator's cylindrical structure and its axisymmetric periodic array characteristics, we can extract sections from both ends and the middle of the stator, labeled A and C respectively, to establish a single-slot finite element model. Taking section B as an example... Figure 3 As shown, more than 50 sets of operating conditions were formed by taking points at intervals within the above operating range using parametric scanning, resulting in temperature field distribution cloud maps and all 6000 output node data for each set of operating conditions.

[0055] 2. Least squares identification technique for grid temperature reduction formula based on regional gridding and loss tracing: Based on the physical structure (such as stator teeth, yoke, and winding ends), material properties, and heat flow paths, the discrete model of the motor temperature field (such as a finite element model containing more than 6000 nodes) is divided into several (e.g., 38) logical regions. The thermal state of each region is characterized by a region characteristic temperature, usually taken as the average of the temperatures of all nodes within that region. Thus, the high-dimensional temperature field is compressed into a low-dimensional eigenvector.

[0056] Please see Figure 4 Taking section A as an example, the axisymmetric model of the stator core is divided into regional grids according to the function of each component. The output of more than 6,000 nodes is divided into 38 grid regions. The coordinates of the midpoint of each grid region approximately represent the average temperature of the entire small grid region. The entire grid division is as follows: Figure 4 As shown: Regions 36-38 are the air domain in the simulation analysis, and regions 33-35 are the stator core yoke with relatively low temperature and not very important in temperature monitoring, so the grid is larger.

[0057] For each region, based on the parametric scan simulation data, a linear explicit mathematical relationship is established between its characteristic temperature and various losses and cooling air volume Q (which is directly related to the rotational speed and can be expressed by the rotational speed through a classical fitting formula) using system identification or regression methods (here, the least squares method is used). This yields a reduced-order formula for the temperature field at the center point of each gridded region.

[0058] 3. Global correction mapping driven by joint features of working conditions and space: Step 3.1: Residual Acquisition and GP Modeling: During real-time operation, temperature sensors are deployed in a few key areas to obtain their measured temperatures. The reduced-order model (fast prediction model) from step S2 is called synchronously to obtain the initial predicted temperatures for these areas, and the residual vector is calculated.

[0059] Step 3.2: Gaussian Process Residual Field Inference: Treating the residuals as a random field varying with spatial location, a Gaussian process model is established. A squared exponential kernel function is used to encode spatial correlation. Based on sparse measured residuals, the residual correction values ​​for all 38 regions can be predicted through posterior inference of the Gaussian process.

[0060] Error analysis of the correction effect in the example: To verify the effectiveness and engineering applicability of the aforementioned model twin and adaptive correction method, this section will systematically analyze the performance of the constructed temperature field digital twin model based on the measured data of eight actual operating main motors. The aim is to evaluate the twin accuracy, robustness and adaptability of the method to different operating conditions from multiple dimensions.

[0061] (1) Temperature model twin effect: Using historical operating data of all current main motors as input, twin values ​​of temperature at two historical measuring points are obtained, and their average absolute percentage error is calculated and summarized in Table 2.

[0062] Table 2, MAPE statistics of temperature twin model at measurement points:

[0063] An analysis of the prediction performance (measured by Mean Absolute Percentage Error, MAPE) of the twin models for eight motors reveals significant differences in prediction accuracy among the main motors. Motors 1, 2, and 3 show relatively high MAPE values ​​at both measurement points (all significantly higher than 5%), indicating relatively low accuracy in predicting the state of these three motors. In contrast, motors 4 through 8 generally have lower MAPE values ​​at both measurement points (mostly below 6%, primarily concentrated in the 2%-5% range), indicating higher accuracy in predicting the operating states of these five motors. Motor 4, in particular, exhibits the best accuracy among all measurement points, with a MAPE value as low as 2.21% at measurement point 2. To improve the generalizability of the model's prediction performance and overcome the current model's limitations in applicability to specific main motors (such as motors 1, 2, and 3), it is necessary to conduct a specific analysis of the main motors with higher errors.

[0064] (2) Verification of the adaptive field correction effect of Gaussian process driven by sparse measured data: To systematically evaluate the actual performance of the Gaussian process adaptive field correction algorithm driven by sparse measured data, this study randomly selected 2000 data points from each of the eight main motors as the training set and 500 data points as the test set, constructing a performance comparison test set before and after correction. The quantitative analysis of the prediction errors at key measurement points is summarized in Table 3.

[0065] Table 3. Comparison of mean absolute percentage error before and after sample correction:

[0066] The results show that after Gaussian process adaptive field correction, the prediction errors of all eight motors were significantly reduced, with the corrected MAPE values ​​all controlled within 1.5%, a reduction of more than 50% compared to before correction. Before correction, the MAPE values ​​of each motor were distributed in a wide range of 2.43%-4.43%, with motor 8 having the highest prediction error (4.43%). After correction, the MAPE values ​​of each motor were concentrated in a narrow range of 1.22%-1.41%, indicating an overall improvement in prediction accuracy and a trend towards consistency. Particularly noteworthy is that motors with relatively low accuracy before correction (such as motors 5, 6, and 8) achieved the most significant improvement after correction, with motor 8 experiencing an error reduction of nearly 70%.

[0067] Taking the main motor as an example, the temperature comparison curves before and after correction at the two measuring points are as follows: Figure 5 , Figure 6 As shown, the residual distribution before and after correction is as follows: Figure 7 The revised green curve more closely matches the upward trend of the true value, and the 95% confidence interval (light green shading) is significantly narrowed, indicating that the dispersion of the prediction results has been significantly reduced. Figure 6 In the (residual comparison), the residuals before correction (red) have a wider fluctuation range (some exceeding ±4℃), while the residuals after correction (purple) are more concentrated near 0, with a significant reduction in fluctuation amplitude. The stability and symmetry of the residuals are significantly improved. Figure 7 The residual distribution is presented visually. Before correction, the residuals (orange histogram) are more dispersed, with low peaks and obvious tails on both sides. After correction, the residuals (green histogram) have sharper peaks and are concentrated near 0, indicating that the normality and stability of the residuals are significantly improved after correction.

[0068] Therefore, it can be concluded that the Gaussian process adaptive field correction method significantly improves the accuracy and stability of the temperature prediction of the main motor: the corrected prediction curve is closer to the true value, the residual fluctuation range is greatly reduced and the distribution is more concentrated, effectively reducing the dispersion and randomness of the prediction error, and verifying the effectiveness of the correction algorithm at the level of a single main motor.

[0069] 4. A computational method for inverse distance weighted interpolation space reconstruction based on sparse block support: Using the inverse distance weighted interpolation method, 38 grid regions were interpolated into 6000 points in the plane to plot the temperature field contour maps of sections A / B / C of the stator axisymmetric model. Taking section B as an example, its corrected and reconstructed temperature field contour map is shown below. Figure 8 As shown.

[0070] In summary, the mechanism-data hybrid-driven method for reducing the stator temperature field of a motor, as presented in this invention, enables rapid prediction of the stator temperature field over a wide speed range, full-field adaptive correction driven by sparse measured data, and high-fidelity reconstruction of the spatial temperature distribution. This provides technical support for real-time temperature monitoring and insulation life assessment of motor digital twin systems. It has the following characteristics: (1) A lightweight temperature field calculation architecture that deeply integrates physical mechanisms and data-driven approaches is constructed, combining computational efficiency and physical interpretability: This invention establishes a tight coupling between physical mechanisms and data-driven approaches through a four-level progressive technical path of "regional rasterization dimensionality reduction - loss source parameter modeling - Gaussian process sparse correction - IDW space reconstruction". The front end is based on the least squares identification of regional rasterization and loss source, using various motor losses and cooling airflow as inputs and regional characteristic temperatures as outputs, to establish a physically interpretable explicit reduction formula, which profoundly reveals the intrinsic causal mapping mechanism between motor stator temperature and copper loss, iron loss, and speed / airflow, and enhances the robustness and extrapolation generalization ability of the model in a wide speed range (25~52.5Hz) and a wide voltage range (5000~10500V). Compared with the black box method of directly using pure machine learning or POD model for end-to-end temperature field prediction, the reduction formula of this invention has clear physical meaning, which is convenient for engineers to perform fault source tracing and thermal state diagnosis. Meanwhile, regional gridding compresses the temperature field dimension from over 6,000 nodes to a 38-dimensional feature vector, reducing the time for a single inference to the millisecond level. This perfectly meets the stringent requirements of digital twin systems for real-time response, thus laying an efficient foundation for real-time monitoring of the stator temperature of variable frequency synchronous motors.

[0071] (2) An adaptive correction method based on Gaussian process and spatial correlation propagation was proposed, achieving a breakthrough in "sparse measured data-driven global correction": This invention addresses the realistic constraint of extremely sparse engineering measured data (usually only 2 measurement points), innovatively modeling the prediction residual as a spatially correlated random field, and using the squared exponential kernel function to encode the physical intuition that "the closer the location, the more similar the residuals," enabling the Gaussian process model to achieve optimal estimation of the residuals of all 38 regions through posterior inference based on a very small number of measured residuals. System verification was performed on 8 actual operating main motors, each with 2000 training samples and 500 test samples. After correction, the MAPE values ​​of all motors were controlled within 1.5%, generally decreasing by more than 50% compared to before correction, with motor No. 8 showing an error reduction of nearly 70%. The peak shape of the residual distribution after correction was sharper and concentrated near zero, with a significantly narrowed 95% confidence interval. This effect means that precise correction of the global temperature field can be achieved by deploying only two temperature sensors, significantly reducing the cost of sensor deployment and maintenance complexity in engineering, and providing an efficient hybrid modeling solution for the engineering dilemma of "limited model accuracy" and "scarcity of measured data".

[0072] (3) A high-fidelity spatial reconstruction mechanism based on "macroscopic features - microscopic distribution" was established to ensure accurate identification of hazardous locations: After correcting the temperature characteristics of 38 regions, this invention uses an inverse distance weighted interpolation method to reconstruct the low-dimensional grid temperature into a full-field temperature distribution with 6000 nodes. By setting the distance power exponent (recommended value 2), the interpolation results are made spatially continuous, smooth, and physically reasonable, avoiding the step effect of traditional nearest neighbor interpolation or the local distortion of linear interpolation. This mechanism ensures the reliable reconstruction of the stator cross-section temperature field cloud map, enabling maintenance personnel to accurately determine the location of the highest stator temperature and potential thermal fault areas, providing an intuitive and reliable decision-making basis for insulation life assessment and intelligent operation and maintenance throughout the entire life cycle.

[0073] (4) A new paradigm for digital twinning of stator temperature field of wide-range speed-regulating variable frequency synchronous motor has been established, laying a technical foundation for intelligent operation and maintenance throughout the entire life cycle: This invention systematically integrates the entire chain of technologies of "multi-condition CFD parameter scanning - regional rasterization order reduction - Gaussian process sparse correction - IDW spatial reconstruction", and constructs a complete technical closed loop of "rapid prediction of physical model - sparse data-driven correction - spatial correlation full-field reconstruction". Taking the TLKS9000-24 series 9000kW vertical variable frequency synchronous motor as the empirical object, covering the frequency operation range of 25~52.5Hz and the voltage operation range of 5000~10500V, the effectiveness and engineering applicability of the method under wide-range speed-regulating complex operating conditions have been verified. This invention provides a key technical path for constructing a mechanism twin model of variable frequency synchronous motor at the factory state, and establishes a new paradigm for stator temperature field twinning based on mechanism-data hybrid drive. By observing the overall twin temperature of the stator temperature field in real time, the highest temperature and dangerous location can be determined, thus laying a solid technical foundation for real-time monitoring of motor temperature and insulation life assessment, and providing theoretical and engineering support for condition warning, maintenance decision-making and intelligent operation and maintenance of this type of motor throughout its entire life cycle.

[0074] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for reducing the order of a motor stator temperature field based on mechanism-data hybrid drive, characterized in that, Includes the following steps: S1, Discretize the space into a grid according to the actual motor operation: Perform grid clustering on all temperature nodes of the motor stator finite element model according to spatial location, material properties or functional partitions, divide into N grid regions, and use the average temperature of nodes in each region as the characteristic temperature of the region to compress the high-dimensional temperature field into an N-dimensional feature vector. S2, the least squares-driven temperature reduction modeling steps based on ventilation mechanism: Based on the simulation data of multi-condition parameter scanning, with various losses and cooling air volume as input, the system identification or regression method is used to establish an explicit reduction formula between the characteristic temperature and the input parameters for each region, and obtain a fast prediction model of the characteristic temperature of N grid regions. S3, Global Correction Mapping Step Driven by Joint Features of Working Condition and Space: In real-time operation, acquire a limited number of measured temperature measurement points, calculate the residual between the measured value and the predicted value of the fast prediction model in step S2, model the spatial correlation of the residual using a Gaussian process, estimate the residual correction value of all N grid areas based on the squared exponential kernel function and through the posterior inference formula, and superimpose the correction value with the predicted value to complete the global adaptive correction. S4, Spatial reconstruction steps based on sparse block support by inverse distance weighted interpolation: The corrected N grid region feature temperatures are used as sparse control points, and the inverse distance weighted interpolation method is used to perform spatial interpolation on all finite element nodes to realize the reconstruction of the low-dimensional grid temperature to the high-dimensional full-field temperature distribution.

2. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S1, N is 38, and the spatial location includes stator teeth, stator yoke, winding ends, and air domain.

3. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S2, the various losses include stator and rotor copper losses, total iron losses, mechanical losses, and stray losses.

4. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S2, the regression method uses the least squares method.

5. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S2, the range of the multi-condition parameter scanning is: frequency 25~52.5Hz, voltage 5000~10500V, and sampling intervals of 5Hz and 500V respectively, forming more than 50 sets of operating conditions.

6. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S3, the quadratic exponential kernel function is: ; In the formula, , Let G be the coordinates of two points in space. For signal variance, This is the length scale parameter.

7. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 6, characterized in that, In step S3, the posterior inference formula is: In the formula, Points to be predicted The residual estimate at the location, K The covariance matrix of the training points, Let be the covariance vector between the test points and the training points. for transpose, For noise variance, It is the identity matrix. r This is the measured residual vector; It is the inverse of the covariance matrix.

8. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 7, characterized in that, In step S3, the formula for global adaptive correction is: In the formula, For the first i The final temperature prediction value after correction for each region For the first i Initial temperature forecast values ​​for each region For the first i The residual correction value for the entire region.

9. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S4, the formula for the inverse distance weight interpolation is: ； In the formula, Points to be interpolated The final temperature value at that location, For the spatial points to be interpolated, i The area is numbered, with a total of 38 grid areas; For the first i The weight coefficient of each region is determined by the distance from the interpolation point to the representative point of that region; the closer the distance, the greater the weight. For the first i The final temperature prediction values ​​after correction for each region; Weight The calculation formula is: ； In the formula, Let be the Euclidean distance from the node to be interpolated to the representative point of the i-th region; q The distance exponent is 2; distance The calculation formula is: ； In the formula, Points to be interpolated To the i The Euclidean distance of each region's representative point. x and y Points to be interpolated coordinates x i and y i For the first i Each grid area represents the coordinates of a point.

10. The method for reducing the order of motor stator temperature field based on mechanism-data hybrid drive according to claim 1, characterized in that, In step S4, the number of nodes in the reconstructed temperature field after interpolation is more than 6,000, covering the entire field temperature distribution of the two ends and the middle section of the motor stator.

Images