A reasoning priori knowledge construction method and device for motor parameter identification

Abstract

The present application belongs to the field of motor engineering, and particularly relates to a reasoning type prior knowledge construction method and device for motor parameter identification, comprising: obtaining basic physical constraint prior based on multi-source parallel data of a target motor; obtaining hot start prior by cross-task learning of multi-source parallel data of the target motor between different motor types and different working conditions; obtaining non-independent and identically distributed adaptive prior based on one or both of the basic physical constraint prior and the hot start prior; obtaining reinforced non-independent and identically distributed adaptive prior by collaborative iterative optimization of the non-independent and identically distributed adaptive prior; integrating the basic physical constraint prior, the hot start prior and the reinforced non-independent and identically distributed adaptive prior to obtain reasoning type prior knowledge for motor parameter identification; the present application can obtain reasoning type prior knowledge with physical rationality, cross-domain adaptability and reasoning reliability, and improve the generalization ability and efficiency of an AI model in a motor parameter identification scene.

Description

Technical Field

[0001] This invention belongs to the field of electrical engineering technology, and relates to the application of artificial intelligence technology in industrial motors, and particularly to a reasoning-based prior knowledge construction method and device for motor parameter identification. Background Technology

[0002] Industrial motors, as core power equipment in industrial production, are widely used in manufacturing, energy, transportation, and other fields, serving as fundamental equipment to ensure continuous and stable industrial production. Due to the wide variety of industrial motors, their design principles and structural parameters differ significantly. Furthermore, the diverse operating conditions, such as load, operating speed, and ambient temperature, result in varying motor operating data modes. Despite these differences, all industrial motors operate according to unified underlying physical laws and are subject to similar engineering constraints in their applications. These common characteristics provide a crucial foundation for condition monitoring, fault diagnosis, and performance optimization of industrial motors.

[0003] With the development of industrial intelligent technology, artificial intelligence (AI) technology is increasingly widely used in scenarios such as condition monitoring, fault diagnosis, and performance optimization of industrial motors. However, how to make full use of the limited and heterogeneous industrial data in the industrial field and extract prior knowledge that can support AI models to perform reliable reasoning has become the key to improving the generalization ability of AI models and the efficiency of engineering implementation. For example, the reasoning performance of AI models for motor parameter identification scenarios such as condition monitoring, fault diagnosis, and performance optimization of industrial motors is highly dependent on the quality of prior knowledge.

[0004] Currently, AI models for industrial motor scenarios mainly rely on manual summarization or extraction from a single data source to obtain experiential prior knowledge. However, due to the characteristics of experiential prior knowledge, such as fragmentation, poor adaptability, low construction efficiency, and inability to be reused across domains, it cannot support the reliable reasoning of AI models.

[0005] Specifically, firstly, traditional methods for acquiring empirical prior knowledge mostly rely on the assumption of independent and identically distributed (i.i.d.) data, extracting statistical priors from homogeneous data from a single data source. However, industrial motor operating data often exhibits time-series dependencies, significant differences in operating conditions, and heterogeneous data noise, resulting in a clear non-independent and identically distributed characteristic. Consequently, empirical prior knowledge cannot accurately reflect the actual operating patterns of industrial motors, leading to a significant decrease in the inference accuracy and reliability of AI models. Secondly, the construction of empirical prior knowledge often adopts a "cold start" model, meaning that the construction process lacks any cross-domain common constraints and initial benchmarks, relying entirely on a single data source. Exploring and refining prior knowledge from scratch not only results in an excessively long construction cycle and high trial-and-error costs, but also makes it prone to deviations from actual physical constraints due to a lack of support from common physical laws and engineering constraints, thus affecting the reliability of AI model inference. Furthermore, the empirical prior knowledge extracted in existing technologies is often confined to specific motor types or specific operating scenarios, making it impossible to achieve cross-motor type and cross-operating scenario flow and reuse. This not only causes huge knowledge waste but also creates serious data barriers, hindering the large-scale application and development of industrial AI technology in the field of motors. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a reasoning-based prior knowledge construction method and apparatus for motor parameter identification, in order to solve the technical problem that experience-based prior knowledge generally suffers from fragmentation, poor adaptability, low construction efficiency, and inability to be reused across domains, making it unable to support reliable reasoning of AI models.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] This invention provides a reasoning-based prior knowledge construction method for motor parameter identification, including:

[0009] Acquire multi-source parallel data of the target motor; wherein, the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor;

[0010] Based on the multi-source parallel data of the target motor, the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor are determined, and the basic physical constraint prior is obtained.

[0011] The multi-source parallel data of the target motor is used for cross-task learning across different motor types and operating conditions to obtain hot start priors.

[0012] Based on one or both of the fundamental physical constraint priors and hot start priors, non-independent and identically distributed adaptive priors are obtained through transfer learning and processed by statistical analysis and distribution fitting analysis.

[0013] Using the hot-start prior as the initial benchmark, the non-independent and identically distributed fitness prior is optimized through collaborative iteration to obtain the enhanced non-independent and identically distributed fitness prior.

[0014] By integrating the basic physical constraint priors, the hot start priors, and the enhanced non-independent identically distributed adaptation priors, inference-based prior knowledge for motor parameter identification is obtained.

[0015] Furthermore, the low-fidelity simulation data of the target motor includes electromagnetic simulation data, thermodynamic simulation data, and dynamic simulation data; the actual test data of the target motor includes bench test data and monitoring data from preset sensors in the target motor; and the high-fidelity simulation data of the target motor includes simulation analysis data from the finite element model or physical model of the target motor.

[0016] Furthermore, based on the multi-source parallel data of the target motor, the process of determining the value range, qualitative correlation, and physical equation constraints of the preset key parameters in the target motor, and obtaining the prior knowledge of the basic physical constraints, is as follows:

[0017] Data cleaning and feature extraction are performed on the multi-source parallel data of the target motor to obtain the preset key parameters of the target motor;

[0018] Statistical analysis was performed on the preset key parameters in the target motor to obtain the value range of the preset key parameters in the target motor;

[0019] Using a predetermined correlation analysis method, a correlation analysis is performed on the preset key parameters in the target motor to obtain the qualitative correlation relationship of the preset key parameters in the target motor.

[0020] Based on the fundamental physical laws of motors, and combined with the preset key parameters in the target motor, physical equation constraints for the preset key parameters in the target motor are established.

[0021] By integrating the value range of the preset key parameters in the target motor, the qualitative correlation of the preset key parameters in the target motor, and the physical equation constraints of the preset key parameters in the target motor, the basic physical constraint prior is obtained.

[0022] Furthermore, the process of obtaining hot-start priors by performing cross-task learning on the multi-source parallel data of the target motor across different motor types and operating conditions is as follows:

[0023] By using a pre-determined meta-learning agent, the multi-source parallel data of the target motor is used to perform cross-task learning across different motor types and operating conditions to obtain the pre-defined common laws of motors.

[0024] By performing a structured mapping of the common laws of preset motors, a hot start prior can be obtained.

[0025] Furthermore, the common laws of motors are preset, including common parameter constraints across motor types, adaptation rules across operating conditions, and universal expressions of physical laws; the hot start prior includes common parameter constraint equations, cross-domain adaptation rules, and initial migration coefficients.

[0026] Furthermore, based on one or both of the fundamental physical constraint priors and hot-start priors, the process of obtaining non-independent and identically distributed adaptive priors through transfer learning and processing by statistical analysis and distribution fitting analysis is as follows:

[0027] Calculate the fidelity of the fundamental physical constraint priors and the hot start priors respectively;

[0028] Based on the fidelity of computational fundamental physical constraint priors and hot start priors, one or both of the fundamental physical constraint priors and hot start priors are selected as source domain priors.

[0029] Transfer learning is performed on source domain priors to obtain cross-domain transfer priors;

[0030] Statistical analysis was performed on cross-domain migration priors to obtain the data time dependencies of cross-domain migration priors;

[0031] Distribution fitting analysis was performed on cross-domain migration priors to identify the differences in data distribution of cross-domain migration priors;

[0032] Based on the data time dependency and data distribution differences of cross-domain migration priors, an initial non-independent and identically distributed adaptation prior is obtained.

[0033] Based on the prior dynamic adaptation mechanism of the extended Kalman filter method, the initial non-independent and identically distributed adaptation prior is corrected by data distribution offset to obtain the non-independent and identically distributed adaptation prior.

[0034] Furthermore, using the hot-start prior as the initial benchmark, the non-independent and identically distributed (i.e., identically distributed) fitness prior is iteratively optimized collaboratively to obtain the enhanced non-independent and identically distributed fitness prior, as follows:

[0035] Based on the reinforcement learning optimization framework, a hot-start prior is used as the initial benchmark to perform cooperative iterative optimization on the non-independent and identically distributed (i.e., identically distributed) adaptation prior, so as to obtain the reinforced non-independent and identically distributed adaptation prior. The termination condition of the cooperative iterative optimization is that the adaptation error of the non-independent and identically distributed adaptation prior is less than the preset adaptation error threshold.

[0036] Furthermore, based on the reinforcement learning optimization framework, using the warm-start prior as the initial benchmark, the non-independent and identically distributed (IOD) fitting prior is iteratively optimized collaboratively to obtain the reinforced non-independent and identically distributed (IOD) fitting prior, as follows:

[0037]

[0038] in, The action value function; The state space before iterative optimization Compared with the action space before iterative optimization The value of the action function under the following conditions; For the reward function; The state space before iterative optimization Compared with the action space before iterative optimization The reward function value below; The state space before iterative optimization; The action space before iterative optimization; Discount factor; It is a function for maximizing the value; The state space after iterative optimization; The action space after iterative optimization; For the iteratively optimized state space With the iteratively optimized action space The value of the action value function.

[0039] Furthermore, the process of integrating the basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors to obtain inferential prior knowledge for motor parameter identification is as follows:

[0040] The basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors are stored in a structured manner, and an index of inference-based prior knowledge is established according to preset indexing rules to obtain inference-based prior knowledge for motor parameter identification.

[0041] This invention also provides a reasoning-based prior knowledge construction system for motor parameter identification, comprising:

[0042] The multi-source data acquisition module is used to acquire multi-source parallel data of the target motor; the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor.

[0043] The basic physical constraint prior construction module is used to determine the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor based on the multi-source parallel data of the target motor, and obtain the basic physical constraint prior.

[0044] The hot start constraint prior construction module is used to learn the hot start prior by performing cross-task learning on the multi-source parallel data of the target motor across different motor types and different operating conditions.

[0045] The non-independent and identically distributed adaptive prior construction module is used to obtain non-independent and identically distributed adaptive priors based on one or both of the basic physical constraint priors and hot-start priors through transfer learning and processing by statistical analysis and distribution fitting analysis.

[0046] The non-independent and identically distributed adaptation prior enhancement module is used to perform collaborative iterative optimization of the non-independent and identically distributed adaptation prior based on the hot-start prior, so as to obtain the enhanced non-independent and identically distributed adaptation prior.

[0047] The prior integration module is used to integrate the basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors to obtain inferential prior knowledge for motor parameter identification.

[0048] The present invention also provides an electronic device, comprising:

[0049] A processor is used to execute computer programs;

[0050] A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the inference-based prior knowledge construction method for motor parameter identification.

[0051] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the inference-based prior knowledge construction method for motor parameter identification.

[0052] The present invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the inference-based prior knowledge construction method for motor parameter identification.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] The inference-based prior knowledge construction method for motor parameter identification provided by this invention obtains inference-based prior knowledge that combines physical rationality, cross-domain adaptability, and inference reliability by starting from the physical essence and data characteristics of motor operation. This effectively improves the generalization ability and engineering implementation efficiency of AI models in motor parameter identification scenarios. Specifically, firstly, by acquiring multi-source parallel data including low-fidelity simulation data, real test data, and high-fidelity simulation data, comprehensive, realistic, and accurate data support is provided for the subsequent extraction of basic physical constraint priors and the learning of hot-start priors, avoiding the limitations of a single data source. Secondly, based on the multi-source parallel data, the determined... The preset value ranges, qualitative correlations, and physical equation constraints of key parameters in the target motor form the basic physical constraint priors. This ensures that the basic physical constraint priors align with the deep physical laws and engineering constraints of motor operation, fundamentally solving the problem of empirical prior knowledge deviating from actual physical laws. Furthermore, multi-source parallel data of the target motor is used for cross-task learning across different motor types and operating conditions to obtain hot-start priors. This provides a reliable starting point for the subsequent construction and optimization of non-independent, identically distributed adaptive priors, replacing the cold-start knowledge construction mode in existing technologies. By using common cross-task knowledge as the initial benchmark, the construction of reasoning-based prior knowledge is significantly shortened. The system reduces construction time and trial-and-error costs. Furthermore, based on one or both of the fundamental physical constraint priors and hot-start priors, non-independent and identically distributed (IOD) adaptive priors are obtained through transfer learning and processed by statistical analysis and distribution fitting analysis. This achieves accurate adaptation to the non-independent and identically distributed characteristics caused by the time-series dependence of motor data, significant differences in operating conditions, and heterogeneous noise. This addresses the deficiency of empirical prior knowledge based on the IOD assumption, which cannot accurately reflect the actual operating laws of the motor, and significantly improves the adaptability of inference-based prior knowledge. Subsequently, using the hot-start prior as the initial benchmark, the non-independent and identically distributed adaptive priors are iteratively optimized collaboratively, achieving complementary reinforcement of different types of prior knowledge, effectively... This approach enhances the reliability and accuracy of reasoning-based prior knowledge. Finally, by integrating basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors, a comprehensive and systematic reasoning-based prior knowledge is formed. This breaks through the limitations of single-type prior knowledge, avoids the fragmentation problem of empirical prior knowledge, and enables the flow and reuse of prior knowledge across motor types and operating scenarios. This reduces knowledge waste, breaks down data barriers, and provides high-quality, comprehensive prior support for AI models. It significantly improves the reasoning accuracy and reliability of AI models in motor parameter identification, and promotes the large-scale application and development of industrial AI technology in the motor field.

[0055] Furthermore, by employing a pre-determined meta-learning agent for cross-task learning, the universality and adaptability of the hot-start prior are improved, enabling it to adapt to motor parameter identification scenarios of different motor types and operating conditions. In addition, by transforming the pre-defined common laws of motors into hot-start priors through structured mapping, the hot-start priors are made more practical and can be directly used as the initial benchmark for subsequent prior optimization, significantly shortening the prior knowledge construction cycle and improving construction efficiency. At the same time, it provides a high-quality source domain foundation for the establishment of cross-domain transfer priors.

[0056] Furthermore, by calculating the fidelity of the basic physical constraint prior and the hot-start prior and selecting the source domain prior, the high quality and adaptability of the source domain prior are ensured, avoiding the problem of poor cross-domain transfer effect caused by improper selection of the source domain prior. Secondly, through statistical analysis and distribution fitting analysis, the data temporal dependency, operating condition difference and noise heterogeneity of the cross-domain transfer prior are accurately mined, which can be specifically adapted to the non-independent and identically distributed characteristics of industrial motor data, solving the core defect that empirical prior knowledge cannot adapt. Finally, the prior dynamic adaptation mechanism of the extended Kalman filter method is introduced to correct the data distribution offset of the initial non-independent and identically distributed adapted prior, which can respond to the dynamic changes in data distribution during motor operation in real time, so that the non-independent and identically distributed adapted prior always maintains high adaptability and accuracy, further improving the reliability of inference-type prior knowledge, providing the AI ​​model with prior support that is more in line with the actual operation scenario, and ensuring that the AI ​​model can still achieve reliable inference in non-independent and identically distributed data scenarios.

[0057] Furthermore, by adopting an optimization framework based on reinforcement learning, using a hot-start prior as the initial benchmark, dynamic optimization of the non-independent and identically distributed (IOD) fitting prior can be achieved. By utilizing the feedback regulation effect of the reward mechanism in reinforcement learning, the optimization process can be advanced in the direction of improving the fit and reliability of the non-independent and identically distributed fitting prior. At the same time, the complementary reinforcement of the hot-start prior and the non-independent and identically distributed fitting prior can be achieved, effectively improving the overall quality of reasoning-based prior knowledge.

[0058] Furthermore, the basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors are stored in a structured manner, avoiding the fragmentation and chaos of reasoning prior knowledge and ensuring its systematicity and integrity. At the same time, by establishing an index for reasoning prior knowledge, rapid retrieval and accurate calling of reasoning prior knowledge can be achieved, improving the reuse efficiency of reasoning prior knowledge.

[0059] The reasoning-based prior knowledge construction system, electronic device, computer-readable storage medium, and computer program product for motor parameter identification provided by this invention possess all the advantages of the aforementioned reasoning-based prior knowledge construction method for motor parameter identification. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 A flowchart of the reasoning-based prior knowledge construction method for motor parameter identification provided in Example 1;

[0062] Figure 2 This is a graph showing the reward trajectory when performing cooperative iterative optimization of non-independent identically distributed fitting priors using different initial benchmarks in Example 1.

[0063] Figure 3 This is a graph showing the relationship between the final reward value and the final violation rate when performing collaborative iterative optimization of non-independent identically distributed fitting priors using different initial benchmarks in Example 1.

[0064] Figure 4 Here is a structural block diagram of the reasoning-based prior knowledge construction system for motor parameter identification provided in Example 2;

[0065] Figure 5 This is a structural block diagram of the electronic device provided in Example 3. Detailed Implementation

[0066] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0067] Before describing the specific embodiments of this application, some of the technical terms involved in the embodiments of this application are explained as follows:

[0068] AI models refer to computer programs or data frameworks that are built using algorithms and data to simulate human intelligent behavior.

[0069] Prior knowledge refers to the known background information or regularities related to the target task that are known before AI model training or inference in artificial intelligence and industrial intelligence systems.

[0070] This invention provides a reasoning-based prior knowledge construction method for motor parameter identification, including:

[0071] Step 100: Obtain multi-source parallel data of the target motor; wherein, the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor.

[0072] Step 200: Based on the multi-source parallel data of the target motor, determine the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor, and obtain the basic physical constraint prior.

[0073] Step 300: Perform cross-task learning on the multi-source parallel data of the target motor across different motor types and operating conditions to obtain hot start prior.

[0074] Step 400: Based on one or both of the fundamental physical constraint priors and hot start priors, obtain non-independent and identically distributed adaptive priors through transfer learning and processing by statistical analysis and distribution fitting analysis.

[0075] Step 500: Using the hot-start prior as the initial benchmark, perform collaborative iterative optimization on the non-independent and identically distributed adaptation prior to obtain the enhanced non-independent and identically distributed adaptation prior.

[0076] Step 600: Integrate the basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors to obtain reasoning-based prior knowledge for motor parameter identification.

[0077] In the above implementation, by acquiring multi-source parallel data including low-fidelity simulation data, real test data and high-fidelity simulation data, comprehensive, real and accurate data support is provided for the subsequent extraction of basic physical constraint priors and the learning of hot-start priors. This effectively avoids the one-sidedness and limitations of a single data source and lays the foundation for the high-quality construction of reasoning-based prior knowledge.

[0078] Secondly, based on the multi-source parallel data of the target motor, the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor are determined, forming a basic physical constraint prior. This ensures that the basic physical constraint prior strictly conforms to the deep physical laws and engineering constraints of motor operation, and fundamentally solves the core problem of empirical prior knowledge deviating from the actual operating law of the motor.

[0079] Furthermore, by performing cross-task learning on the multi-source parallel data of the target motor across different motor types and operating conditions to obtain hot-start priors, and replacing the cold-start knowledge construction mode in the existing technology, the common knowledge across tasks is used as the initial benchmark for constructing inference-based prior knowledge. This provides a reliable starting point for the subsequent construction and optimization of non-independent and identically distributed adaptive priors, which can significantly shorten the construction cycle of inference-based prior knowledge and effectively reduce trial and error costs.

[0080] Furthermore, by utilizing fundamental physical constraint priors or hot-start priors and combining them with transfer learning methods to establish cross-domain transfer priors, and then processing them through statistical analysis and distribution fitting analysis, non-independent and identically distributed adaptive priors are obtained. These priors can accurately adapt to the non-independent and identically distributed characteristics of motor operating data caused by time-series dependence, significant differences in operating conditions, and heterogeneous noise. This effectively solves the defect that empirical prior knowledge based on the assumption of independent and identical distribution cannot accurately reflect the actual operating law of the motor, and significantly improves the scenario adaptability of inference-based prior knowledge. Using the hot-start prior as the initial benchmark, the non-independent and identically distributed adaptive priors are collaboratively iteratively optimized, realizing the complementary reinforcement of different types of prior knowledge, and effectively improving the reliability and accuracy of inference-based prior knowledge.

[0081] Finally, by integrating basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors, a comprehensive and systematic reasoning-based prior knowledge is formed. This breaks the limitations of single-type prior knowledge, avoids the fragmentation problem of empirical prior knowledge, and enables the flow and reuse of prior knowledge across motor types and operating scenarios. This reduces knowledge waste, breaks down industry data barriers, and provides high-quality, comprehensive prior support for AI models, significantly improving the reasoning accuracy and reliability of AI models in motor parameter identification.

[0082] The following specific embodiments further explain the reasoning-based prior knowledge construction method for motor parameter identification provided by the present invention:

[0083] Example 1

[0084] As attached Figure 1 As shown, this embodiment 1 provides a reasoning-based prior knowledge construction method for motor parameter identification, including the following steps:

[0085] Step 1: Obtain multi-source parallel data of the target motor; wherein, the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor.

[0086] The target motor refers to the motor application object determined according to the actual application scenario. The target motor can be a specific single motor, a series of motor products of a certain type, or different types of motors. The process of acquiring multi-source parallel data of the target motor adopts either an offline acquisition mode or an online acquisition mode. In the offline acquisition mode, the multi-source parallel data of the target motor is obtained by batch collecting historical operating data of the target motor. In the online acquisition mode, the multi-source parallel data of the target motor is obtained by real-time collecting real-time operating data of the target motor. The multi-source parallel data of the target motor obtained in the online acquisition mode can be used to dynamically supplement existing multi-source parallel data of the target motor, and then used in the update process of existing reasoning-based prior knowledge.

[0087] Specifically, the low-fidelity simulation data of the target motor includes electromagnetic simulation data, thermodynamic simulation data, and dynamic simulation data; the real test data of the target motor includes bench test data and monitoring data from preset sensors in the target motor; among which, the real test data of the target motor includes, for example, phase current, phase voltage, output torque, winding temperature, and stator speed; the high-fidelity simulation data of the target motor includes simulation analysis data based on the finite element model or physical model of the target motor; among which, the simulation analysis data based on the finite element model or physical model of the target motor includes, for example, magnetic saturation simulation results and eddy current loss simulation results.

[0088] It should be noted that by acquiring multi-source parallel data including low-fidelity simulation data, real test data, and high-fidelity simulation data, the multi-source parallel data is made comprehensive, realistic, and accurate. This provides accurate and comprehensive data support for the subsequent extraction of basic physical constraint priors and the learning of hot-start priors, further improving the reliability and adaptability of subsequent prior knowledge and reducing the cost of data acquisition and screening. Specifically, the low-fidelity simulation data of the target motor is limited to include electromagnetic simulation data, thermodynamic simulation data, and dynamic simulation data of the target motor to comprehensively cover the core physical processes of motor operation and ensure that the multi-source parallel data can reflect the different dimensions of the motor's operating characteristics. The real test data of the target motor is clearly defined to include bench test data and monitoring data from preset sensors in the target motor, achieving data authenticity that takes into account both controlled test scenarios and actual operating scenarios, avoiding the one-sidedness of single test data. The high-fidelity simulation data of the target motor is limited to simulation analysis data based on finite element models or physical models, ensuring the accuracy of high-fidelity simulation data and compensating for the lack of extreme working conditions and microscopic characteristic data that are difficult to obtain in real test data.

[0089] Step 2: Based on the multi-source parallel data of the target motor, determine the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor, and obtain the basic physical constraint prior.

[0090] Among them, the value range of the preset key parameters in the target motor refers to the numerical range or value space of the preset key parameters in the target motor; the qualitative correlation of the preset key parameters in the target motor refers to the mutual influence law between the preset key parameters in the target motor; the physical equation constraint of the preset key parameters in the target motor refers to the quantitative relationship between the preset key parameters in the target motor based on the basic physical laws of motors and expressed through mathematical models, which can be directly used for actual engineering calculations.

[0091] Specifically, the steps are as follows:

[0092] Step 21: Perform data cleaning and feature extraction on the multi-source parallel data of the target motor to obtain the preset key parameters of the target motor. The preset key parameters of the target motor include stator resistance, direct-axis inductance, quadrature-axis inductance, flux linkage characteristics, number of pole pairs, rated power, stator speed, and output torque; the flux linkage characteristics are the permanent magnet flux linkage of permanent magnet motors or the equivalent flux linkage or excitation flux linkage of non-permanent magnet motors.

[0093] The data cleaning process includes missing value imputation using linear interpolation, outlier removal, and smoothing using Gaussian filtering. The feature extraction process includes feature extraction methods based on time-domain statistics, frequency-domain statistics, or motor operating characteristics.

[0094] Step 22: Perform statistical analysis on the preset key parameters in the target motor to obtain the value range of the preset key parameters in the target motor. Specifically, firstly, use the Kolmogorov-Smirnov (KS) test method or the Shapiro-Wilk (SW) test method to perform a normality test on the preset key parameters in the target motor; then, use the maximum likelihood estimation method to fit the parameter distribution to obtain the value range of the preset key parameters in the target motor.

[0095] Step 23: Using a predetermined correlation analysis method, perform correlation analysis on the preset key parameters in the target motor to obtain the qualitative correlation relationship of the preset key parameters in the target motor. Preferably, the predetermined correlation analysis method is the Pearson correlation analysis method.

[0096] Step 24: Based on the fundamental physical laws of motors and combined with the preset key parameters in the target motor, establish the physical equation constraints for the preset key parameters in the target motor. Specifically, firstly, normalize the preset key parameters in the target motor using a normalization method to obtain the normalized results; then, determine the constraint coefficients of the preset key parameters in the target motor through regression analysis, obtaining the constraint coefficients for each preset key parameter; based on the fundamental physical laws of motors and combined with the normalized results of the preset key parameters and the constraint coefficients of each preset key parameter, establish the constraint equations for the preset key parameters in the target motor; finally, verify the physical rationality of the constraint equations for the preset key parameters in the target motor to ensure that the constraint equations for the preset key parameters in the target motor satisfy the conditions of energy conservation and power balance, thus obtaining the physical equation constraints for the preset key parameters in the target motor; among them, the fundamental physical laws of motors include, for example, the law of electromagnetic induction, the law of conservation of energy, and the laws of thermodynamics.

[0097] Step 25: Integrate the value range of the preset key parameters in the target motor, the qualitative correlation of the preset key parameters in the target motor, and the physical equation constraints of the preset key parameters in the target motor to obtain the basic physical constraint prior.

[0098] Step 3: Perform cross-task learning on the multi-source parallel data of the target motor across different motor types and operating conditions to obtain hot start priors.

[0099] It is important to note that "warm-start prior" refers to prior knowledge obtained by performing cross-task learning from the multi-source parallel data of the target motor to generate preset common motor rules, and then performing structured mapping. The warm-start prior serves as a constraint and initial benchmark for the construction and optimization of subsequent non-independent and identically distributed (IMDG) adaptation priors, enabling rapid construction and collaborative optimization of IMDG adaptation priors. Specifically, using a pre-determined meta-learning agent, the multi-source parallel data of the target motor is used for cross-task learning across different motor types and operating conditions to obtain preset common motor rules. These preset common motor rules are then structured and mapped to obtain the warm-start prior. The pre-determined meta-learning agent is deployed as an initial model benchmark without preset domain knowledge; that is, the pre-determined meta-learning agent does not pre-load motor-specific pre-trained knowledge. Preferably, the pre-determined meta-learning agent adopts a meta-learning agent based on a Model-Agnostic Meta-Learning (MAML) algorithm architecture.

[0100] Specifically, the steps are as follows:

[0101] Step 31: Using a pre-determined meta-learning agent, cross-task learning is performed on the multi-source parallel data of the target motor across different motor types and operating conditions to obtain the preset common laws of motors.

[0102] Specifically, the process of using a pre-determined meta-learning agent to perform cross-task learning on multi-source parallel data of the target motor across different motor types and operating conditions to obtain pre-defined common laws of the motor includes:

[0103] Step 311: Divide the multi-source parallel data of the target motor into task groups according to the division strategy of motor type and operating condition to obtain several independent sub-tasks.

[0104] Step 312: Input each independent subtask into the predetermined meta-learning agent to perform local gradient descent training on the predetermined meta-learning agent using each independent subtask, and obtain the training log corresponding to each independent subtask; wherein, the training log corresponding to each independent subtask includes the local loss curve corresponding to each independent subtask, the parameter optimization trajectory of the meta-learning agent, and the optimal meta-model output within each independent subtask.

[0105] Step 313: Based on the local loss curve corresponding to each independent subtask, perform a global meta-update on the pre-determined meta-learning agent to obtain the optimized meta-model parameters. It is worth noting that local gradient descent training only enables the meta-learning agent to adapt to each independent subtask, failing to form common knowledge across tasks. Therefore, performing a global meta-update on the meta-learning agent based on the local loss curve corresponding to each independent subtask allows for the extraction of common knowledge across different independent subtasks, strengthening the representational ability of the meta-learning agent and enabling it to quickly adapt across tasks.

[0106] Step 314: Synchronously output the training logs and optimized meta-model parameters corresponding to each independent subtask to obtain the meta-training results.

[0107] Step 315: Extract the common features of each independent subtask from the meta-training results to obtain the preset common laws of motors; among which, the preset common laws of motors include common parameter constraints across motor types, adaptation laws across operating conditions, and universal expressions of physical laws; common parameter constraints across motor types, such as the range of stator resistance values ​​for all motors; adaptation laws across operating conditions, such as the correlation pattern between load torque and speed; and universal expressions of physical laws, such as the simplified constraint logic of the law of electromagnetic induction.

[0108] Step 32: Perform structured mapping on the extracted common laws of the motor to obtain the hot start prior.

[0109] Specifically, firstly, the extracted common laws of motors are grouped according to a classification strategy based on parameter constraints, engineering adaptation, and physical laws to obtain the law grouping results. Next, based on the law grouping results, the common laws of motors are transformed into mathematical expressions or logical rules to obtain formalized common laws. Furthermore, the formalized common laws are assigned preset engineering parameters to obtain a hot-start prior that includes common parameter constraint equations, cross-domain adaptation rules, and initial migration coefficients. Among them, the initial migration coefficients refer to the migration weights determined based on the common parameter constraints across motor types, which are used in the subsequent step 4 to determine the migration coefficients.

[0110] Step 4: Based on one or both of the fundamental physical constraint priors and hot start priors, obtain non-independent and identically distributed adaptive priors through transfer learning and processing by statistical analysis and distribution fitting analysis.

[0111] Specifically, the steps are as follows:

[0112] Step 41: Calculate the fidelity of the fundamental physical constraint prior and the hot-start prior, respectively. The process for calculating the fidelity of the fundamental physical constraint prior and the hot-start prior is as follows:

[0113]

[0114]

[0115] in, This is to ensure the fidelity of fundamental physical constraint priors or hot-start priors. ; This is the physical consistency weighting coefficient; This is a score for the physical consistency of the underlying physical constraint priors or hot-start priors, calculated based on the compliance of physical laws. ; This refers to the numerical precision weighting coefficient; For numerical accuracy scoring of basic physical constraint priors or hot-start priors based on measurement error assessment, ; For inference adaptability weighting coefficients; The reasoning fitness score is based on the underlying physical constraint priors or hot-start priors used for logical reasoning consistency assessment. ; This is a weighting coefficient for cross-domain reusability; For cross-domain reusability scoring of basic physical constraint priors or hot-start priors based on motor type coverage assessment, .

[0116] Step 42: Based on the fidelity of the fundamental physical constraint prior and the hot-start prior, select one or both of the fundamental physical constraint prior and the hot-start prior as the source domain prior. Specifically, compare the fidelity of the fundamental physical constraint prior or the hot-start prior with a preset fidelity threshold; if the fidelity of the fundamental physical constraint prior or the hot-start prior is greater than the preset fidelity threshold, then the corresponding fundamental physical constraint prior or hot-start prior is used as the source domain prior; preferably, the preset fidelity threshold is 8.

[0117] Step 43: Perform transfer learning on the source domain prior to obtain cross-domain transfer prior. Specifically, the process is as follows:

[0118]

[0119] in, This is a prerequisite for cross-domain migration; For the source domain prior; The migration coefficient; This represents the difference in motor parameters before and after transfer learning; it should be noted that the transfer coefficient... It is calculated based on the fidelity of the basic physical constraint prior or the fidelity of the hot start prior calculated in step 41, and combined with the initial migration coefficient in the hot start prior in step 3.

[0120] Step 44: Perform statistical analysis on the cross-domain migration priors to obtain the data time dependencies of the cross-domain migration priors; perform distribution fitting analysis on the cross-domain migration priors to identify the data distribution differences of the cross-domain migration priors; among which, the data distribution differences of the cross-domain migration priors include the operating condition differences and noise heterogeneity of the cross-domain migration priors; based on the data time dependencies and data distribution differences of the cross-domain migration priors, obtain the initial non-independent identically distributed fit priors.

[0121] The process of performing statistical analysis on cross-domain migration priors to obtain the data time dependencies of cross-domain migration priors is as follows:

[0122]

[0123] in, Priorities for cross-domain migration in time lag The autocorrelation coefficient of the data at that time; It is the covariance function; For cross-domain migration prerequisites Real-time data samples at any given moment and prior knowledge of cross-domain migration The covariance of real-time data samples at any given time; For cross-domain migration prerequisites Real-time data samples at any given moment; For cross-domain migration prerequisites Real-time data samples at any given moment; It is the variance function; For cross-domain migration prerequisites The variance of real-time data samples at any given time.

[0124] It should be noted that in the prior knowledge of cross-domain migration Real-time data samples at any given moment For example, the target motor in Measured values ​​of phase current, phase voltage, or winding temperature at any given time; in the prior knowledge of cross-domain migration. Real-time data samples at any given moment In order to align with the prior knowledge of cross-domain migration Real-time data samples at any given moment Measured values ​​of the same data type at lag time; priors for cross-domain migration Real-time data samples at any given moment and prior knowledge of cross-domain migration Covariance of real-time data samples at time 1 It is used to quantify the degree of linear correlation of data at different times; in cross-domain transfer priors Variance of real-time data samples at time 1 It is used to quantify the dispersion of data at a specific moment.

[0125] Step 45: Calculate the Kullback-Leibler (KL) divergence of the initial non-independent and identically distributed (IOD) fitted prior to obtain the divergence calculation result of the initial non-independent and identically distributed fitted prior; compare the divergence calculation result of the initial non-independent and identically distributed fitted prior with the preset divergence threshold to determine whether it is necessary to perform data distribution offset correction on the initial non-independent and identically distributed fitted prior.

[0126] The process of calculating the Kolbec-Leibler divergence of the initial non-independent and identically distributed fitted prior is as follows:

[0127]

[0128] in, To adapt priors for initial non-independent and identically distributed systems Compared with the preset baseline data distribution The Körbek-Leibler divergence between them; For the initial non-independent and identically distributed adaptation prior; For summation functions; The baseline data distribution is based on fundamental physical constraints, hot-start priors, and initial non-independent identically distributed adaptation priors. Determined comprehensively; To adapt priors for initial non-independent and identically distributed systems Data sample points in real-time data samples The probability density at that location; Preset baseline data distribution Data sample points in real-time data samples The probability density at that location; These are data sample points for real-time data samples, such as measured values ​​of phase current, phase voltage, or winding temperature.

[0129] It is worth noting that when the divergence calculation result of the initial non-independent and identically distributed (IOD) fitting prior is greater than the preset divergence threshold, the data distribution offset correction of the initial non-independent and identically distributed (IOD) fitting prior needs to be performed, and the process jumps to step 46; when the divergence calculation result of the initial non-independent and identically distributed (IOD) fitting prior is less than or equal to the preset divergence threshold, the initial non-independent and identically distributed (IOD) fitting prior is output as the non-independent and identically distributed (IOD) fitting prior.

[0130] Step 46: Based on the prior dynamic adaptation mechanism of the Extended Kalman Filter (EKF) method, the initial non-independent and identically distributed adaptation prior is corrected for data distribution offset to obtain the non-independent and identically distributed adaptation prior.

[0131] The prior dynamic adaptation mechanism based on the extended Kalman filter method involves correcting the data distribution shift of the initial non-independent and identically distributed adaptation prior, including a prediction step, an update step, and a prior adaptation step. The specific principle is as follows:

[0132] (1) In the prediction step, based on The prior parameters that have been adapted at each time point, combined with the state transition matrix, are used to... Predicting is done using prior parameters at time points to obtain the predicted values. The prior parameters at each time point are used to dynamically deduce the prior parameters; simultaneously, the prediction error covariance matrix is ​​updated; the specific process is as follows:

[0133]

[0134]

[0135] in, For prediction Prior parameters at time; for The state transition matrix at time step 1 is used to describe the change in prior parameters from... Time's up The dynamic evolutionary relationship at any given moment; for Prior parameters that have been adapted at the time; The input matrix is ​​used to characterize the influence of the external control input of the target motor on the prior parameters; The control input vector includes external control inputs to the target motor, such as the direct-axis voltage reference command and quadrature-axis voltage reference command of the target motor. This is the updated prediction error covariance matrix used in the prediction step to quantify the prediction. Prior parameters at time Uncertainty; for The prediction error covariance matrix at time step is used for quantization. Uncertainty of prior parameters that have been adapted at the given time; for The transpose of the state transition matrix at time step 1; for The process noise covariance matrix at time step is used to quantify the prior parameters from... Time's up Random disturbances during time transitions.

[0136] (2) In the update step, first calculate Kalman gain at time step; then, based on The Kalman gain at time k, combined with the measured data and observation matrix at time k, is used to predict... Prior parameters at time Make corrections to obtain The highest priority arithmetic parameters at time t are used as the corrected highest priority arithmetic parameters; simultaneously, the prediction error covariance matrix is ​​updated; the specific process is as follows:

[0137]

[0138]

[0139]

[0140] in, for Kalman gain matrix at time step; for The observation matrix at each time point; for The transpose of the observation matrix at time t; for The observation noise covariance matrix at time step; for The highest priority parameter at time; for Real-time measured data; It is the identity matrix; for The prediction error covariance matrix at time 1.

[0141] It should be noted that, Kalman gain matrix at time step Used to measure predictions Prior parameters at time and Real-time measured data The weighting relationship; Observation matrix at time Used to characterize the prediction Prior parameters at time and Real-time measured data The mapping relationship between them; Observation noise covariance matrix at time 1 , used to measure Real-time measured data Measurement error; Real-time measured data ,include The phase current, phase voltage, and output torque of the target motor are collected at all times. Prediction error covariance matrix at time step Used for quantification The highest priority parameter at time Uncertainty.

[0142] (3) In the prior fitting step, the parameter boundaries and constraint rules of the initial non-independent and identically distributed fitting prior are dynamically adjusted based on the modified highest priority prior parameters so that they match the data distribution of the measured data, thereby obtaining the non-independent and identically distributed fitting prior.

[0143] Step 5: Using the hot start prior as the initial benchmark, perform collaborative iterative optimization on the non-independent and identically distributed fitting prior to obtain the enhanced non-independent and identically distributed fitting prior.

[0144] Specifically, using a reinforcement learning-based optimization framework, with a hot-start prior as the initial benchmark, the non-independent and identically distributed (i.i.d.) adaptation prior is collaboratively iteratively optimized to obtain the reinforced non-independent and identically distributed (i.i.d.) adaptation prior. The termination condition for the collaborative iterative optimization is that the adaptation error of the non-independent and identically distributed (i.i.d.) adaptation prior is less than a preset adaptation error threshold.

[0145] Specifically, the collaborative iterative optimization process is as follows:

[0146]

[0147] in, The action value function; The state space before iterative optimization Compared with the action space before iterative optimization The value of the action function under the following conditions; For the reward function; The state space before iterative optimization Compared with the action space before iterative optimization The reward function value below; The state space before iterative optimization; The action space before iterative optimization; Discount factor; It is a function for maximizing the value; The state space after iterative optimization; The action space after iterative optimization; For the iteratively optimized state space With the iteratively optimized action space The value of the action value function.

[0148] It is worth noting that the state space includes the parameter states of the hot-start prior, the adaptation states of the non-independent and identically distributed (ISD) adaptation prior, the real-time operating state of the target motor, and the data distribution offset states of the ISD adaptation prior; the action space is based on the hot-start prior as the initial reference and includes the weight correction of the ISD adaptation prior, the data distribution offset rules, and the transition coefficient; and the discount factor. It is used to balance immediate reduction with long-term cross-task optimization goals.

[0149] It should also be noted that designing the termination condition of collaborative iterative optimization as having the adaptation error of the non-independent identically distributed adapting prior less than a preset adaptation error threshold can effectively control the optimization cycle and trial-and-error costs, avoid resource waste caused by over-optimization, and ensure that the optimized non-independent identically distributed adapting prior can reach the preset adaptation standard, providing a high-quality core component for the integration of reasoning-based prior knowledge.

[0150] Step 6: Integrate the basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors to obtain inferential prior knowledge for motor parameter identification. Specifically, the basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptation priors are stored in a structured manner, and an index of inferential prior knowledge is established according to preset indexing rules to obtain inferential prior knowledge for motor parameter identification.

[0151] The detailed explanation is the process of structurally storing the basic physical constraint priors, the hot start priors, and the enhanced non-independent and identically distributed adaptation priors. Specifically, based on the basic physical constraint priors, the hot start priors, and the enhanced non-independent and identically distributed adaptation priors, a motor physical constraint prior library, a hot start prior library, and an enhanced non-independent and identically distributed adaptation prior library are constructed respectively.

[0152] In the motor physical constraint prior library, basic physical constraint priors are stored in a data format of parameter type-constraint type-constraint content-fidelity level; in the hot start prior library, hot start priors are stored in a data format of motor type-operating condition label-prior content-fidelity level; in the enhanced non-independent and identically distributed adaptation prior library, enhanced non-independent and identically distributed adaptation priors are stored in a data format of non-independent and identically distributed characteristic type-correction rule-trigger threshold-hot start prior benchmark association identifier.

[0153] The hot-start prior benchmark association identifier refers to the identifier of the hot-start prior used as the initial benchmark in the collaborative iterative optimization process, ensuring that the association between the enhanced non-independent identically distributed adaptation prior and the hot-start prior used as the initial benchmark is traceable; among them, the hot-start prior benchmark association identifier is encoded in the format of motor type-operating condition-hot-start prior identifier to ensure the uniqueness of the association traceability.

[0154] The data format for the index of reasoning-based prior knowledge is specifically motor type-operating condition-parameter type-fidelity level-hot start association identifier, so as to support the AI ​​model to quickly call the corresponding prior knowledge according to the real-time scenario, and ensure the synergistic empowerment of the AI ​​model by basic physical constraint prior, hot start prior and enhanced non-independent identically distributed adaptation prior.

[0155] Performance verification:

[0156] To verify the performance of the reasoning-based prior knowledge for motor parameter identification obtained in Example 1, the rewards and violations of enhanced non-independent and identically distributed (ISD) adaptive priors obtained through different methods are used for verification and explanation. Specifically, four different prior data are used as initial benchmarks, and collaborative iterative optimization is performed on the ISD adaptive priors to obtain four different enhanced ISD adaptive priors. The reward values ​​and violation rates of the four different enhanced ISD adaptive priors are calculated and statistically analyzed to obtain performance verification results, as shown in the appendix. Figures 2-3 As shown; among them, the four different prior data are hot start prior, low-fidelity simulation data of the target motor, real test data and high-fidelity simulation data.

[0157] As attached Figure 2 As shown, attached Figure 2 The figure shows the reward trajectory curves for collaborative iterative optimization of non-independent identically distributed fitting priors using different initial benchmarks; Appendix Figure 2 The thin solid line in the figure represents the reward trajectory curve when using low-fidelity simulation data of the target motor as an initial benchmark for collaborative iterative optimization of non-independent identically distributed adaptive priors. (See attached figure.) Figure 2The dashed line in the figure represents the reward trajectory curve when using a hot-start prior as an initial benchmark for collaborative iterative optimization of a non-independent, identically distributed fit prior. Figure 2 The dotted line in the figure represents the reward trajectory curve when using real test data as the initial baseline for collaborative iterative optimization of non-independent and identically distributed fitting priors. Figure 2 The dotted lines in the figure represent the reward trajectory curves when using high-fidelity simulation data as the initial benchmark to perform collaborative iterative optimization of non-independent identically distributed adaptive priors.

[0158] From the appendix Figure 2 As can be seen, when the number of rounds of collaborative iterative optimization is less than or equal to 20, the reward value for collaborative iterative optimization using low-fidelity simulation data of the target motor as the initial benchmark and non-independent identically distributed (ISD) fitting priors is in the low exploration range of around -80,000. However, when using a hot-start prior as the initial benchmark and performing collaborative iterative optimization on ISD fitting priors, the reward value rapidly recovers and stabilizes above -10,000 when the number of rounds of collaborative iterative optimization is 20, significantly moving away from the low-reward trial-and-error stage of the cold-start knowledge building mode. Comparing the reward trajectory curves for collaborative iterative optimization using real test data as the initial benchmark and ISD fitting priors with those using high-fidelity simulation data as the initial benchmark, it can be intuitively verified that using a hot-start prior as the initial benchmark for collaborative iterative optimization on ISD fitting priors can significantly shorten the training cycle and avoid the waste of resources from exploring from scratch.

[0159] As attached Figure 3 As shown, attached Figure 3 The figure shows the relationship between the final reward value and the final violation rate when performing cooperative iterative optimization of non-independent and identically distributed fitting priors using different initial benchmarks; among them, the appendix... Figure 3 The rectangular symbols in the figure represent the final reward value and the final violation rate when using a hot-start prior as the initial baseline for collaborative iterative optimization of a non-independent and identically distributed fitting prior. Figure 3 The circular symbols in the figure represent the final reward value and the final violation rate when using low-fidelity simulation data of the target motor as the initial benchmark for collaborative iterative optimization of non-independent identically distributed adaptive priors. Figure 3 The triangle symbol in the diagram represents the final reward value and the final violation rate when using real test data as the initial benchmark for collaborative iterative optimization of non-independent and identically distributed fitting priors. Figure 3 The diamond symbol in the figure represents the final reward value and the final violation rate when using high-fidelity simulation data as the initial benchmark to perform collaborative iterative optimization of non-independent identically distributed adaptive priors.

[0160] From the appendix Figure 3As can be seen, when using the hot-start prior as the initial benchmark to perform collaborative iterative optimization of the non-independent identically distributed fitting prior, the final reward value is better and the constraint violation rate is 0. Compared with the other three initial benchmark cases, the hot-start prior can guide the AI ​​model to spontaneously internalize the physical laws of motors and engineering constraints while pursuing high performance or high rewards, achieving a unity of performance improvement and constraint compliance. This marks the upgrade of the AI ​​model from passively following rules to actively internalizing common sense.

[0161] The reasoning-based prior knowledge construction method for motor parameter identification described in Embodiment 1 integrates the constructed basic physical constraint prior, hot-start prior, and enhanced non-independent identically distributed (IOD) adaptation prior as reasoning-based prior knowledge for motor parameter identification. This generates high-quality prior knowledge that is deeply adapted to logical reasoning and can be reused across motor types, greatly alleviating the problem of industrial data scarcity. It provides plug-and-play, continuously evolving reasoning empowerment for various motor parameter identification models, improving the credibility and generalization ability of industrial AI models from the source. Specifically, based on one or both of the basic physical constraint prior and hot-start prior, the non-independent identically distributed (IOD) adaptation prior is obtained through transfer learning and processed by statistical analysis and distribution fitting analysis. Using the hot-start prior as the initial benchmark, the non-independent identically distributed (IOD) adaptation prior is iteratively optimized to obtain the enhanced non-independent identically distributed (IOD) adaptation prior. By introducing explicit modeling and dynamic adaptation mechanisms for the non-independent identically distributed characteristics, it ensures that the prior knowledge resonates with the distribution of real industrial data, avoiding prior distortion caused by invalid assumptions.

[0162] Example 2

[0163] As attached Figure 4 As shown, this embodiment 2 provides a reasoning-based prior knowledge construction system for motor parameter identification, including a multi-source data acquisition module, a basic physical constraint prior construction module, a hot start constraint prior construction module, a non-independent and identically distributed adaptation prior construction module, a non-independent and identically distributed adaptation prior reinforcement module, and a prior integration module.

[0164] The multi-source data acquisition module is used to acquire multi-source parallel data of the target motor; the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor.

[0165] The basic physical constraint prior construction module is used to determine the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor based on the multi-source parallel data of the target motor, and obtain the basic physical constraint prior.

[0166] The hot start constraint prior construction module is used to learn the target motor's multi-source parallel data across different motor types and operating conditions to obtain the hot start prior.

[0167] The non-independent and identically distributed adaptive prior construction module is used to obtain non-independent and identically distributed adaptive priors based on one or both of the basic physical constraint priors and hot-start priors, through transfer learning and processing by statistical analysis and distribution fitting analysis.

[0168] The non-independent and identically distributed adaptation prior enhancement module is used to perform collaborative iterative optimization of the non-independent and identically distributed adaptation prior, using the hot-start prior as the initial benchmark, to obtain the enhanced non-independent and identically distributed adaptation prior.

[0169] The prior integration module is used to integrate the basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors to obtain inferential prior knowledge for motor parameter identification.

[0170] Example 3

[0171] As attached Figure 5 As shown, this embodiment 3 provides an electronic device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the reasoning-based prior knowledge construction method for motor parameter identification; or, the processor executing the computer program to implement the functions of each module in the above-mentioned reasoning-based prior knowledge construction system for motor parameter identification.

[0172] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a preset function, the instruction segments describing the execution process of the computer program in the electronic device.

[0173] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of electronic devices and do not constitute a limitation on the electronic device. It may include more components than described above, or combine certain components, or different components. For example, the electronic device may also include a communication interface, input / output devices, network access devices, and a bus.

[0174] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the electronic device via various communication interfaces and lines.

[0175] The memory can be used to store the computer program and / or module. The processor implements various functions of the electronic device by running or executing the computer program and / or module stored in the memory and by calling the data stored in the memory.

[0176] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards, secure digital cards, flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0177] Example 4

[0178] This embodiment 4 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the inference-based prior knowledge construction method for motor parameter identification.

[0179] If the module / unit of the reasoning-based prior knowledge construction system for motor parameter identification is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0180] Based on this understanding, the present invention can implement all or part of the processes in the above-described reasoning-based prior knowledge construction method for motor parameter identification. This can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described reasoning-based prior knowledge construction method for motor parameter identification. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0181] The computer-readable storage medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0182] Example 5

[0183] This embodiment 5 provides a computer product, which includes a computer program stored in a computer-readable storage medium. The processor of the electronic device reads the computer program from the computer-readable storage medium and executes the computer program, so that the electronic device can execute the reasoning-based prior knowledge construction method for motor parameter identification described in embodiment 1, which will not be repeated here.

[0184] It should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods.

[0185] The inference-based prior knowledge construction method for motor parameter identification described in this invention acquires multi-source parallel data including low-fidelity simulation data, real test data, and high-fidelity simulation data, providing comprehensive, realistic, and accurate data support for the subsequent extraction of basic physical constraint priors and the learning of hot-start priors. Based on the multi-source parallel data, the value ranges, qualitative correlations, and physical equation constraints of the predetermined key parameters in the target motor are determined to form basic physical constraint priors, ensuring that the basic physical constraint priors can conform to the deep physical laws and engineering constraints of motor operation. The multi-source parallel data of the target motor is used for cross-task learning across different motor types and operating conditions to obtain hot-start priors, providing a reliable starting point for the construction and optimization of subsequent non-independent and identically distributed adaptive priors. Using common knowledge across tasks as the initial benchmark, the construction cycle of inference-based prior knowledge is significantly shortened, and the trial-and-error cost is reduced. One or both of the basic physical constraint priors and hot-start priors are used to obtain non-independent and identically distributed (IOD) adapted priors through transfer learning and statistical analysis and distribution fitting analysis. This achieves accurate adaptation to the non-independent and identically distributed characteristics caused by the time-series dependence of motor data, significant differences in operating conditions, and heterogeneous noise, significantly improving the adaptability of inference-based prior knowledge. Using the hot-start prior as the initial benchmark, the non-independent and identically distributed adapted priors are iteratively optimized collaboratively, realizing the complementary reinforcement of different types of prior knowledge and effectively improving the reliability and accuracy of inference-based prior knowledge. By integrating the basic physical constraint priors, hot-start priors, and reinforced non-independent and identically distributed adapted priors, a comprehensive and systematic inference-based prior knowledge is formed, enabling the flow and reuse of prior knowledge across motor types and operating conditions. This provides high-quality and comprehensive prior support for AI models, significantly improving the inference accuracy and reliability of AI models in motor parameter identification.

[0186] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

Claims

1. A reasoning-based prior knowledge construction method for motor parameter identification, characterized in that, include: Acquire multi-source parallel data of the target motor; wherein, the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor; Based on the multi-source parallel data of the target motor, the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor are determined, and the basic physical constraint prior is obtained. The multi-source parallel data of the target motor is used for cross-task learning across different motor types and operating conditions to obtain hot start priors. Based on one or both of the fundamental physical constraint priors and hot start priors, non-independent and identically distributed adaptive priors are obtained through transfer learning and processed by statistical analysis and distribution fitting analysis. Using the hot-start prior as the initial benchmark, the non-independent and identically distributed fitness prior is optimized through collaborative iteration to obtain the enhanced non-independent and identically distributed fitness prior. By integrating the basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors, inference-based prior knowledge for motor parameter identification is obtained. Based on one or both of fundamental physical constraint priors and hot-start priors, the process of obtaining non-independent and identically distributed adaptive priors through transfer learning and processing with statistical analysis and distribution fitting analysis is as follows: Calculate the fidelity of the fundamental physical constraint priors and the hot start priors respectively; Based on the fidelity of computational fundamental physical constraint priors and hot start priors, one or both of the fundamental physical constraint priors and hot start priors are selected as source domain priors. Transfer learning is performed on source domain priors to obtain cross-domain transfer priors; Statistical analysis was performed on cross-domain migration priors to obtain the data time dependencies of cross-domain migration priors; Distribution fitting analysis was performed on cross-domain migration priors to identify the differences in data distribution of cross-domain migration priors; Based on the data time dependency and data distribution differences of cross-domain migration priors, an initial non-independent and identically distributed adaptation prior is obtained. Based on the prior dynamic adaptation mechanism of the extended Kalman filter method, the initial non-independent and identically distributed adaptation prior is corrected by data distribution offset to obtain the non-independent and identically distributed adaptation prior.

2. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 1, characterized in that, Low-fidelity simulation data of the target motor, including electromagnetic simulation data, thermodynamic simulation data, and dynamic simulation data; real test data of the target motor, including bench test data and monitoring data from preset sensors in the target motor; high-fidelity simulation data of the target motor, including simulation analysis data of the finite element model or physical model of the target motor.

3. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 1, characterized in that, Based on multi-source parallel data of the target motor, the process of determining the value range, qualitative correlation, and physical equation constraints of the preset key parameters in the target motor, and obtaining the prior knowledge of the basic physical constraints, is as follows: Data cleaning and feature extraction are performed on the multi-source parallel data of the target motor to obtain the preset key parameters of the target motor; Statistical analysis was performed on the preset key parameters in the target motor to obtain the value range of the preset key parameters in the target motor; Using a predetermined correlation analysis method, a correlation analysis is performed on the preset key parameters in the target motor to obtain the qualitative correlation relationship of the preset key parameters in the target motor. Based on the fundamental physical laws of motors, and combined with the preset key parameters in the target motor, physical equation constraints for the preset key parameters in the target motor are established. By integrating the value range of the preset key parameters in the target motor, the qualitative correlation of the preset key parameters in the target motor, and the physical equation constraints of the preset key parameters in the target motor, the basic physical constraint prior is obtained.

4. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 1, characterized in that, The process of obtaining hot-start prior by performing cross-task learning on multi-source parallel data of the target motor across different motor types and operating conditions is as follows: By using a pre-determined meta-learning agent, the multi-source parallel data of the target motor is used to perform cross-task learning across different motor types and operating conditions to obtain the pre-defined common laws of motors. By performing a structured mapping of the common laws of preset motors, a hot start prior can be obtained.

5. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 4, characterized in that, The pre-defined common laws of motors include common parameter constraints across motor types, adaptation rules across operating conditions, and universal expressions of physical laws; the hot start prior includes common parameter constraint equations, cross-domain adaptation rules, and initial migration coefficients.

6. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 1, characterized in that, The process of obtaining the enhanced non-independent and identically distributed (i.e., identically distributed) fitness prior, using the hot-start prior as the initial benchmark, and performing cooperative iterative optimization on the non-independent and identically distributed fitness prior is as follows: Based on the reinforcement learning optimization framework, a hot-start prior is used as the initial benchmark to perform cooperative iterative optimization on the non-independent and identically distributed (i.e., identically distributed) adaptation prior, so as to obtain the reinforced non-independent and identically distributed adaptation prior. The termination condition of the cooperative iterative optimization is that the adaptation error of the non-independent and identically distributed adaptation prior is less than the preset adaptation error threshold.

7. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 6, characterized in that, Based on the reinforcement learning optimization framework, using a hot-start prior as the initial benchmark, the non-independent and identically distributed (IOD) fitness prior is iteratively optimized collaboratively to obtain the reinforced non-independent and identically distributed (IOD) fitness prior. The process is as follows: in, The action value function; The state space before iterative optimization Compared with the action space before iterative optimization The value of the action function under the following conditions; For the reward function; The state space before iterative optimization Compared with the action space before iterative optimization The reward function value below; This refers to the state space before iterative optimization. The action space before iterative optimization; Discount factor; It is a function for maximizing the value; The state space after iterative optimization; The action space after iterative optimization; For the iteratively optimized state space With the iteratively optimized action space The value of the action function.

8. The reasoning-based prior knowledge construction method for motor parameter identification according to claim 1, characterized in that, The process of integrating basic physical constraint priors, hot-start priors, and enhanced non-independent identically distributed adaptive priors to obtain inferential prior knowledge for motor parameter identification is as follows: The basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors are stored in a structured manner, and an index of inference-based prior knowledge is established according to preset indexing rules to obtain inference-based prior knowledge for motor parameter identification.

9. A reasoning-based prior knowledge construction system for motor parameter identification, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source parallel data of the target motor; the multi-source parallel data of the target motor includes low-fidelity simulation data, real test data and high-fidelity simulation data of the target motor. The basic physical constraint prior construction module is used to determine the value range, qualitative correlation and physical equation constraints of the preset key parameters in the target motor based on the multi-source parallel data of the target motor, and obtain the basic physical constraint prior. The hot start constraint prior construction module is used to learn the hot start prior by performing cross-task learning on the multi-source parallel data of the target motor across different motor types and different operating conditions. The non-independent and identically distributed adaptive prior construction module is used to obtain non-independent and identically distributed adaptive priors based on one or both of the basic physical constraint priors and hot-start priors through transfer learning and processing by statistical analysis and distribution fitting analysis. The non-independent and identically distributed adaptation prior enhancement module is used to perform collaborative iterative optimization of the non-independent and identically distributed adaptation prior based on the hot-start prior, so as to obtain the enhanced non-independent and identically distributed adaptation prior. The prior integration module is used to integrate the basic physical constraint priors, hot start priors, and enhanced non-independent identically distributed adaptation priors to obtain reasoning-based prior knowledge for motor parameter identification. Based on one or both of fundamental physical constraint priors and hot-start priors, the process of obtaining non-independent and identically distributed adaptive priors through transfer learning and processing with statistical analysis and distribution fitting analysis is as follows: Calculate the fidelity of the fundamental physical constraint priors and the hot start priors respectively; Based on the fidelity of computational fundamental physical constraint priors and hot start priors, one or both of the fundamental physical constraint priors and hot start priors are selected as source domain priors. Transfer learning is performed on source domain priors to obtain cross-domain transfer priors; Statistical analysis was performed on cross-domain migration priors to obtain the data time dependencies of cross-domain migration priors; Distribution fitting analysis was performed on cross-domain migration priors to identify the differences in data distribution of cross-domain migration priors; Based on the data time dependency and data distribution differences of cross-domain migration priors, an initial non-independent and identically distributed adaptation prior is obtained. Based on the prior dynamic adaptation mechanism of the extended Kalman filter method, the initial non-independent and identically distributed adaptation prior is corrected by data distribution offset to obtain the non-independent and identically distributed adaptation prior.

10. An electronic device, characterized in that, include: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the reasoning-based prior knowledge construction method for motor parameter identification as described in any one of claims 1-8.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the reasoning-based prior knowledge construction method for motor parameter identification as described in any one of claims 1-8.

12. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the reasoning-based prior knowledge construction method for motor parameter identification as described in any one of claims 1-8.

Images