A permanent magnet synchronous motor reliability state monitoring system for a collaborative robot

Abstract

This invention provides a reliability status monitoring system for permanent magnet synchronous motors used in collaborative robots, belonging to the field of physical testing. It includes a data acquisition module, a physical information hybrid degradation modeling module, a finite element simulation and dataset construction module, a CEPIN network module, a dynamic risk assessment module, and a multi-level status assessment module. The system first collects degradation data through accelerated degradation testing, establishing a temperature-degradation stochastic model that combines deterministic physical trends with stochastic fluctuations. A multi-source feature dataset is constructed using finite element simulation combined with Latin hypercube sampling and K-means clustering strategies. Then, a dual-branch CEPIN network, fusing residual feature extraction and structural causal inference branches, is employed to achieve high-precision, highly interpretable prediction of component health status parameters. Subsequently, a time-varying parameter probability distribution model dynamically quantifies the comprehensive operational risk of the motor. Finally, a progressive health status assessment result is output, enabling continuous online monitoring and early warning of the motor's reliability status.

Description

Technical Field

[0001] This invention relates to the field of physical testing, and more specifically to a reliability status monitoring system for permanent magnet synchronous motors used in collaborative robots. Background Technology

[0002] Collaborative robots are widely used in industrial automation, medical assistance, and precision manufacturing, and their joint drive systems typically employ permanent magnet synchronous motors (PMSMs). During actual operation, collaborative robots frequently encounter complex dynamic conditions such as start-stop shocks, instantaneous overloads, and continuous load variations. This makes the motor's critical components—the permanent magnets and stator winding insulation—extremely susceptible to both thermal and electromagnetic stresses, leading to progressive degradation and ultimately, motor performance decline or even functional failure, jeopardizing the overall reliability and safety of the robot system. Therefore, effective reliability monitoring of PMSMs used in collaborative robots is of significant engineering importance.

[0003] In existing technologies, methods for monitoring the reliability status of permanent magnet synchronous motors (PMSMs) can be mainly divided into three categories. The first category is data-driven methods, including random forests, convolutional neural networks, and long short-term memory networks. These methods learn degradation features from large amounts of historical data to achieve status monitoring, resulting in high accuracy. However, the models lack physical interpretability, and their generalization ability significantly decreases when there are differences between actual operating conditions and the distribution of training data, making them unsuitable for adapting to the dynamic operating conditions of collaborative robots with frequent load variations. The second category is physical model-driven methods, represented by finite element analysis and analytical electromagnetic models. These methods establish degradation models based on the electromagnetic mechanism of the motor, providing clear physical meaning. However, the model construction process relies on numerous simplifying assumptions, making it difficult to efficiently handle high-dimensional heterogeneous measured data, and the modeling cost is high, making it difficult to meet the needs of real-time monitoring under dynamic operating conditions. The third category is causal reasoning methods. These methods infer the motor state by constructing causal relationships between variables, offering strong interpretability. However, the construction of causal graphs highly depends on the prior knowledge of domain experts, making it difficult to automatically learn degradation features from data, thus significantly limiting their applicability in practical engineering applications.

[0004] In summary, all three methods described above have significant limitations when used individually: data-driven methods offer high accuracy but suffer from poor interpretability and weak generalization ability; physical model-driven methods have clear mechanisms but struggle to handle complex measured data; and causal reasoning methods offer strong interpretability but lack sufficient feature learning capabilities. Furthermore, existing methods generally lack multi-level comprehensive assessment capabilities, ranging from component degradation states to system-level health levels, making it difficult to provide sufficient quantitative basis for collaborative robot operation and maintenance decisions. Therefore, it is necessary to propose a reliability status monitoring system for permanent magnet synchronous motors used in collaborative robots that integrates the advantages of physical knowledge and data-driven approaches, balances accuracy and interpretability, and possesses multi-level health assessment capabilities. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings by proposing a reliability status monitoring system for permanent magnet synchronous motors used in collaborative robots.

[0006] The present invention adopts the following technical solution: A reliability status monitoring system for a permanent magnet synchronous motor for collaborative robots includes a data acquisition module, a physical information hybrid degradation modeling module, a finite element simulation and dataset construction module, a CEPIN network module, a dynamic risk assessment module, and a multi-level status assessment module. The data acquisition module is used to collect accelerated degradation test data of permanent magnets and winding insulation, as well as online operation data of motors, providing a data foundation for subsequent modeling and online inference. The physical information hybrid degradation modeling module is used to combine the Arrhenius equation with Gaussian process regression based on accelerated degradation test data to establish temperature-degradation stochastic models for permanent magnets and winding insulation, respectively, and to quantify the degradation law of key components. The finite element simulation and dataset construction module is used to construct a multi-source feature dataset covering permanent magnet state, winding state and torque characteristics based on the degradation state quantization results and through an efficient sampling strategy combining finite element simulation with Latin hypercube sampling and K-means clustering. The CEPIN network module is used to predict the health status parameters of key components such as permanent magnet demagnetization rate and winding insulation state by using a multi-source feature dataset as training input and a dual-branch fusion architecture of residual feature extraction branch and structural causal reasoning branch. The dynamic risk assessment module is used to construct a time-varying parameter probability distribution risk model based on the component health status parameters output by the CEPIN network module, dynamically quantify the comprehensive operating risk of the motor, and output a normalized comprehensive risk value. The multi-level status assessment module is used to integrate component status parameters, system performance parameters, and normalized comprehensive risk values, and output progressive health status assessment results from a binary classification of "reliable / unreliable" to a four-level classification of "excellent / good / average / needs improvement", thereby realizing online continuous monitoring and early warning of motor reliability status.

[0007] Furthermore, the accelerated degradation test data collected by the data acquisition module includes: the degradation data of the surface magnetic field strength over time obtained by baking and aging the permanent magnet sample at multiple constant temperature levels, and the degradation data of the insulation resistance over time obtained by isothermal aging the winding insulation sample at multiple constant temperature levels. The online operating data of the motor includes: permanent magnet temperature signal collected by a temperature sensor installed near the permanent magnet, winding temperature signal collected by a temperature sensor installed near the winding, output torque signal collected by a torque sensor, speed signal collected by an incremental encoder, and phase current signals collected by a current sensor. The above signals are output after being synchronously collected by a multi-channel data acquisition unit.

[0008] Furthermore, the specific method by which the physical information hybrid degradation modeling module establishes the degradation model is as follows: A quantitative relationship between degradation rate and temperature is established based on the Arrhenius relation, and a physical trend term for the magnetic field strength of permanent magnets and winding insulation resistance is established using an exponential decay function. The test time at each temperature level was converted into a uniform equivalent time at room temperature using an acceleration factor, and the residual between the measured degradation data and the trend term was calculated. Gaussian process regression is used to learn the mapping relationship between equivalent time and residual. A frequency-aware kernel function is used for permanent magnets, and an additional fluctuation frequency term based on temperature calculation is introduced to capture the frequency difference of magnetic field strength fluctuation with temperature and time. A kernel function that considers the joint correlation of time difference and temperature difference is used for winding insulation. The final predicted output is the sum of the physical trend term and the Gaussian process residual term.

[0009] Furthermore, the specific implementation method of the finite element simulation and dataset construction module is as follows: A three-dimensional finite element model of a permanent magnet synchronous motor was established. Uniform demagnetization conditions were set in the model. The maximum, minimum and average values ​​of the output torque under different demagnetization degrees were obtained by gradually reducing the residual magnetic induction intensity of the permanent magnet. Inject inter-turn short-circuit faults with different numbers of turns into the model and obtain the corresponding phase current waveform data; Latin hypercube sampling is used to uniformly generate candidate sample points in a multidimensional feature space. Then, K-means clustering is used to screen representative typical working point points for finite element simulation, so as to significantly reduce the number of simulations while ensuring data representativeness. Finally, a multi-source feature dataset was constructed, which includes the surface magnetic field strength of permanent magnets, demagnetization rate, winding insulation resistance, output torque characteristics, and phase current characteristics.

[0010] Furthermore, the specific implementation method of the residual feature extraction branch of the CEPIN network module is as follows: First, the original feature vector is standardized by mean-variance to eliminate the numerical scale difference between features of different dimensions. The standardized features are fed into a residual learning module constructed from multiple fully connected layers. Each residual unit adopts a skip connection structure, which superimposes the transformation result of the current layer with the unit input and outputs the result. The activation function of each layer is ReLU. The output of the residual module is transformed by the linear output layer to predict key health parameters such as permanent magnet demagnetization rate and winding insulation status.

[0011] Furthermore, the specific implementation of the structural causal reasoning branch of the CEPIN network module is as follows: Based on a predefined causal graph of a permanent magnet synchronous motor structure, the causal graph includes nodes such as temperature stress, permanent magnet demagnetization rate, winding insulation resistance, output torque, phase current distortion, and system reliability, as well as directed causal edges between nodes that conform to physical laws. Using demagnetization rate, surface magnetic field strength, and ambient temperature as inputs, the causal characteristics of the output torque of the permanent magnet state-guided theory are established through multiple linear regression. At the same time, winding state characterization quantities are directly extracted from the original winding data as winding causal features; After the two types of causal features are fused according to the set weights, they are concatenated with the original feature matrix of the residual feature extraction branch to form the final input feature matrix, which is then sent to the prediction output layer.

[0012] Furthermore, the specific implementation method of the dynamic risk assessment module is as follows: The current operating risk state of the motor is modeled as a time-varying normal distribution, with the mean of the distribution decaying exponentially over time and the variance increasing linearly over time. A time-varying distribution is also established for the risk assessment threshold, so that the assessment boundary can dynamically adapt to the actual degradation process of the motor. The risk state space is divided into multiple continuous intervals. The probability of the motor state falling into each risk interval at each time moment is calculated using the cumulative distribution function of the standard normal distribution. The probability of the system being in a high-risk state is represented by the sum of the probabilities of all intervals exceeding the preset risk threshold. A normalized comprehensive risk value is generated and output by combining the probability of high-risk intervals and the probability of threshold crossing.

[0013] Furthermore, the specific implementation method of the multi-level state evaluation module is as follows: The reliability criteria are a demagnetization rate of no more than 30% and an insulation resistance of no less than 5 megohms. If either index exceeds the above threshold, it is judged as unreliable. A system state function is introduced to give a unified time-varying mathematical expression to the above criteria. Under the premise of being deemed reliable, the comprehensive performance score is further calculated by weighted summation of five sub-scores: demagnetization score, winding insulation score, torque stability score, comprehensive risk score, and time degradation item. The weight of each sub-score is determined by regression fitting of historical degradation data. The overall score is divided into four health levels according to preset threshold ranges: above 0.8 is "excellent", between 0.6 and 0.8 is "good", between 0.4 and 0.6 is "average", and below 0.4 is "needs improvement". Ultimately, the reliability status, health level, and early warning information will be output in real time.

[0014] The beneficial effects achieved by this invention are: This invention combines the Arrhenius equation with Gaussian process regression to establish a temperature-degradation stochastic hybrid model that combines deterministic physical trends with the ability to describe random fluctuations. This overcomes the shortcomings of existing physical model-driven methods in handling the random uncertainties of degradation processes and achieves accurate quantitative characterization of the degradation laws of permanent magnets and winding insulation.

[0015] The proposed CEPIN dual-branch network architecture organically integrates the residual feature extraction branch with the structural causal reasoning branch. While fully leveraging the powerful feature extraction capabilities of data-driven methods, it applies physical and logical constraints to the network reasoning process through a predefined physical causal graph. This overcomes the shortcomings of existing data-driven methods, such as poor interpretability and weak generalization ability, enabling the model output to have clear physical meaning and supporting multi-level comprehensive evaluation from component-level degradation state to system-level health level.

[0016] This invention constructs a dynamic risk assessment model based on the probability distribution of time-varying parameters. The mean and variance of the risk distribution change dynamically with time, and the risk judgment threshold also adopts a time-varying form. This overcomes the shortcomings of existing methods that use fixed thresholds for risk judgment, which are prone to misjudgment in long-term operation scenarios, and realizes dynamic and refined quantification of motor degradation risk.

[0017] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structural framework of the present invention; Figure 2 This is a schematic diagram of the data flow in the data acquisition module of the present invention; Figure 3 This is a schematic diagram of the dual-branch architecture of the CEPIN network of the present invention; Figure 4 This is a schematic diagram of the multi-level state evaluation logic of the present invention; Figure 5 This is a schematic diagram showing the prediction results of the permanent magnet magnetic field strength degradation model of the present invention; Figure 6This is a schematic diagram showing the prediction results of the winding insulation resistance degradation model of the present invention. Detailed Implementation

[0019] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0020] Example 1: This embodiment uses a permanent magnet synchronous motor for joint drive in a collaborative robot as an example to fully demonstrate the specific implementation of the reliability status monitoring system, combined with... Figure 1 The system consists of a data acquisition module, a physical information hybrid degradation modeling module, a finite element simulation and dataset construction module, a CEPIN network module, a dynamic risk assessment module, and a multi-level state assessment module working together in sequence.

[0021] When the system is working, the data acquisition module collects accelerated degradation test data and motor online operation data respectively, providing a data foundation for subsequent modeling and online inference; The physical information hybrid degradation modeling module establishes a temperature-degradation stochastic model of permanent magnets and winding insulation based on degradation test data, quantifying the degradation law of key components; The finite element simulation and dataset construction module constructs a multi-source feature dataset based on the degradation state quantification results by combining finite element simulation with an efficient sampling strategy. The CEPIN network module uses this dataset as training input and achieves high-precision interpretable prediction of component health status parameters through a dual-branch fusion architecture of residual feature extraction branch and structural causal reasoning branch. The dynamic risk assessment module constructs a time-varying probability distribution risk model based on the output of the CEPIN network, and dynamically quantifies the comprehensive operational risk of the motor. The multi-level condition assessment module integrates component condition parameters and comprehensive risk values, and outputs progressive health condition assessment results from a binary classification of "reliable / unreliable" to a four-level classification of "excellent / good / average / needs improvement", realizing online continuous monitoring and early warning of motor reliability status.

[0022] Combination Figure 2 The data acquisition module is responsible for providing the system with two types of raw input data.

[0023] The first category is accelerated degradation test data: For permanent magnets, NdFeB:35 material samples with dimensions of 20mm×15mm×2.5mm were selected. Twenty magnets from the same batch were baked and aged at three constant temperature levels of 100℃, 130℃, and 160℃. The surface magnetic field strength was measured every 24 hours using a gaussmeter. The total test duration was 500 hours, and data on the decay and degradation of magnetic field strength with time and temperature were obtained. For winding insulation, F-class insulated enameled wire (copper core diameter 1.2mm) was selected. Stranded wire samples were prepared according to relevant IEC standards. Constant temperature aging was carried out at three temperature levels of 230℃, 260℃, and 290℃. The insulation resistance was measured every 12 hours. The total test duration was 300 hours, and data on the degradation of insulation resistance with time and temperature were obtained.

[0024] The second category is motor online operation data: the temperature signal of the permanent magnet is collected by a temperature sensor installed near the permanent magnet, the winding temperature signal is collected by a temperature sensor installed near the winding, the output torque signal is collected by a torque sensor, the speed signal is collected by an incremental encoder, and the current signal of each phase is collected by a current sensor. The above signals are synchronously collected by a multi-channel data acquisition unit and then sent to the subsequent processing modules.

[0025] Based on the accelerated degradation test data provided by the data acquisition module, the physical information hybrid degradation modeling module combines the Arrhenius equation with Gaussian process regression to establish physical information hybrid degradation models for permanent magnets and winding insulation, respectively.

[0026] For the permanent magnet degradation model, a quantitative relationship between degradation rate and temperature is first established based on the Arrhenius relation. The degradation rate increases exponentially with increasing temperature and is determined by the activation energy, gas constant, and absolute temperature. Based on this, a physical trend term for the change of magnetic field strength with time at different temperatures is established using an exponential decay function. The decay rate of the trend term is a function of temperature.

[0027] Specifically, a model is constructed based on the degradation test data of permanent magnets. The degradation function of the magnetic field strength on the surface of the permanent magnet is: ; Where B0(T) is the initial magnetic field strength at different temperatures, and k(T) is the degradation rate at different temperatures. This embodiment introduces Gaussian process regression (GPR) to address the uncertainty in results caused by small fluctuations in data from different groups at the same temperature. The specific steps are as follows.

[0028] Given magnetic field degradation data at different temperatures T {t i ,B(t i Let the prior physical model be exponentially degenerate, then: ; In the formula, GP(t, θ) is a Gaussian process that characterizes random fluctuations, and θ is a kernel hyperparameter.

[0029] Convert the high-temperature test time to the equivalent time at room temperature: ; The magnetic field strength degradation at 100℃, 130℃, and 160℃ was simulated, its exponential trend was learned, and random noise was added to simulate experimental fluctuations. ; The equivalent time and magnetic field strength at different temperatures are concatenated into a training set, and GPR is used to learn t. eq The relationship between the residual and the magnetic field strength. The residual is defined as: ; In this embodiment, a frequency sensing term is added to incorporate the temperature-dependent fluctuation frequency of the magnetic induction intensity degradation on the surface of the permanent magnet. The square exponential kernel is selected as the GPR kernel function, which allows the model to more accurately capture the frequency differences in the magnetic induction intensity fluctuation on the surface of the permanent magnet caused by temperature and time.

[0030] ; The first term of this equation is the squared exponential kernel used to characterize the smooth correlation between temperature and time, ω(T i ,T j (Based on temperature T) i ,T j The second term in the calculated fluctuation frequency term is a white noise kernel σ, which represents the random error of uncertainty in the data. f , l, σ n , l f This is a kernel hyperparameter used to characterize the correlation and noise characteristics between data.

[0031] Finally, combining the trend term and the GP residual, the final prediction model is: ; Prediction results are as follows Figure 5 As shown.

[0032] Considering the random fluctuations caused by differences in material homogeneity and measurement errors among different sample groups at the same temperature level, Gaussian process regression is introduced to model the random residuals based on the deterministic trend term. An acceleration factor is used to uniformly convert the experimental time at each temperature level to the equivalent time at room temperature, eliminating time scale differences between different temperatures. The residuals between the measured magnetic field strength and the trend term at each moment are calculated. A frequency-aware kernel function is used for Gaussian process regression, consisting of a squared exponential kernel superimposed with a white noise kernel. An additional fluctuation frequency term based on temperature calculation is introduced to accurately capture the frequency differences in magnetic field strength fluctuations with temperature and time. The final predicted output is the sum of the physical trend term and the Gaussian process residual term.

[0033] For the winding insulation degradation model, the same modeling framework is used. The physical trend term of insulation resistance is established with an exponential decay function. The decay rate of the trend term also satisfies the Arrhenius relation. Gaussian process regression is introduced to model the random residuals of insulation resistance. The training features include both equivalent time and temperature dimensions. The kernel function considers the joint correlation between time difference and temperature difference and includes a white noise term to avoid overfitting. The final prediction output is the sum of the trend term and the Gaussian process prediction residual term.

[0034] Specifically, a model is constructed based on the insulation winding degradation test data, and the degradation function of the winding insulation resistance is: ; Where R0 is the initial insulation resistance and A is a parameter to be determined, the value of which is independent of temperature.

[0035] Similar to the derivation process for permanent magnet modeling, the expression for the insulating winding in the Gaussian process regression process with fluctuations is: ; For the insulation winding data at each temperature, first calculate the trend term, i.e., its theoretical value: ; Then calculate the corresponding residual, which is the difference between the actual value and the trend term: ; The original expression can then be expressed as: ; The training features of GPR are equivalent time and temperature, and the training objective is the residual ΔR. Considering the correlation between time and temperature, a custom kernel function is used: ; Where x=[t eq ,T], Δt and ΔT are the differences between time and temperature, respectively, σ f , l t , σn , l T These are hyperparameters, determined through optimization. Similarly, GPR learns the distribution of the residuals to avoid overfitting.

[0036] Finally, the insulation resistance values ​​at different temperatures are predicted. First, the trend term is calculated: ; Then, GPR is used to predict the residual ΔR(t,T), with the input feature [t,T]. The final predicted value is: ; Prediction results are as follows Figure 6 As shown.

[0037] Based on degradation modeling, the finite element simulation and dataset construction module uses ANSYS Maxwell software to establish a three-dimensional finite element model of the permanent magnet synchronous motor, obtaining electromagnetic performance data under different degradation states. A uniform demagnetization condition is set in the finite element model, and the electromagnetic characteristics under different demagnetization degrees are simulated by gradually reducing the residual magnetic induction intensity of the permanent magnet. The simulation obtains the maximum, minimum, and average values ​​of the output torque under each demagnetization degree. The increasing degree of demagnetization leads to a monotonically decreasing trend in each torque characteristic value. Inter-turn short-circuit faults with different numbers of turns are injected into the finite element model, and the corresponding phase current waveforms are obtained. Inter-turn short circuits in phase A cause significant distortion of the phase current, and the more short-circuited turns, the more severe the distortion. A mapping relationship between the winding insulation state and the phase current characteristics is established.

[0038] To address the issues of high computational load and significant time consumption in 3D finite element simulation, a strategy combining Latin hypercube sampling and K-means clustering is employed to generate training samples. First, candidate sample points are uniformly generated in the multi-dimensional feature space, including permanent magnet demagnetization rate, winding insulation resistance, and temperature, ensuring uniform coverage of the entire feature space. Then, K-means clustering is used to select representative typical operating point from the candidate points and submit them for finite element simulation. This approach significantly reduces the number of simulations while ensuring data representativeness, effectively reducing data distribution bias. Finally, a dataset containing multi-source features such as permanent magnet surface magnetic field strength, demagnetization rate, winding insulation resistance, output torque characteristics (maximum, minimum, and average values), and phase current characteristics is constructed.

[0039] Based on the above dataset, a reliability status monitoring standard and a multi-level evaluation system were further established. The reliability binary classification judgment standard was determined based on the finite element simulation results: when the permanent magnet demagnetization rate reaches 30%, the output torque drops to outside the allowable deviation range of the rated value, and the motor output performance deteriorates significantly; when the winding insulation resistance drops to 5 megohms, the phase current distortion rate increases significantly, and the risk of insulation failure enters an unacceptable range. Therefore, a demagnetization rate of no more than 30% and an insulation resistance of no less than 5 megohms are used as reliability criteria. If any indicator exceeds the above threshold, it is judged as unreliable. A system state function is introduced to give a unified time-varying mathematical expression to the above criteria.

[0040] Based on the binary classification judgment, a four-level health scoring system is further established. The comprehensive performance score is obtained by weighted summation of five sub-scores: demagnetization score, winding insulation score, torque stability score, comprehensive risk score, and time degradation item. The weight of each sub-score is determined by regression fitting of historical degradation data. The demagnetization score decreases monotonically with the increase of demagnetization rate; the winding insulation score decreases monotonically with the decrease of insulation resistance; the torque stability score is calculated based on the maximum, minimum, and average values ​​of output torque to determine the current torque fluctuation degree, with a higher score for smaller fluctuations; the comprehensive risk score is converted from the normalized comprehensive risk value output by the dynamic risk assessment module; and the time degradation item decreases with the cumulative operating time, reflecting the cumulative effect of degradation over time.

[0041] The overall score is divided into four health levels according to preset threshold ranges: above 0.8 is "excellent", between 0.6 and 0.8 is "good", between 0.4 and 0.6 is "average", and below 0.4 is "needs improvement".

[0042] Combination Figure 3The CEPIN network module uses the aforementioned multi-source feature dataset as training input and adopts a dual-branch network architecture, integrating data-driven feature extraction capabilities with interpretable reasoning capabilities constrained by physical causality. The residual feature extraction branch uses multi-source typical factor data as input, first performing mean-variance standardization on the original feature vector to eliminate numerical scale differences between features of different dimensions. The standardized features are then fed into the residual learning module constructed from multiple fully connected layers. Each residual unit uses a skip connection structure, superimposing the transformation result of the current layer with the unit input before outputting, effectively alleviating the gradient vanishing problem in deep network training. The activation function for each layer is ReLU. The output of the residual module is transformed by a linear output layer to predict key health state parameters such as permanent magnet demagnetization rate and winding insulation state. The structural causal reasoning branch is based on the pre-defined... A causal graph for the defined permanent magnet synchronous motor structure is constructed. The causal graph includes nodes such as temperature stress, permanent magnet demagnetization rate, winding insulation resistance, output torque, phase current distortion, and system reliability, as well as directed causal edges between nodes that conform to physical laws, fully reflecting the causal transmission relationship between various factors during motor degradation. The branch takes demagnetization rate, surface magnetic field strength, and ambient temperature as inputs, and establishes the causal characteristics of the output torque of the permanent magnet state-guided theory through multiple linear regression. At the same time, the winding state characterization quantity is directly extracted from the original winding data as the winding causal characteristics. The two types of causal features are weighted and fused according to the set weights, and then concatenated with the original feature matrix of the residual feature extraction branch to form the final input feature matrix that integrates deep data-driven features and physical causal constraint features. This matrix is ​​then sent to the prediction output layer for state parameter inference.

[0043] Based on the component health status parameters output by the CEPIN network module, the dynamic risk assessment module constructs a dynamic risk quantification model based on the time-varying parameter probability distribution. The module models the current motor's operational risk state as a time-varying normal distribution, with the distribution mean decaying exponentially over time, reflecting the dynamic drift of the risk center as the degradation process progresses; the distribution variance increases linearly over time, reflecting the continuous increase in degradation uncertainty with accumulated operating time. A time-varying distribution is also established for the risk judgment threshold, enabling the judgment boundary to dynamically adapt to the actual degradation process of the motor, avoiding misjudgments caused by fixed thresholds in long-term operating scenarios. The risk state space is divided into multiple continuous intervals, and the probability of the motor state falling into each risk interval at each moment is calculated using the cumulative distribution function of the standard normal distribution. The probability of the system being in a high-risk state is represented by the sum of the probabilities of all intervals exceeding the preset risk threshold. Combining the high-risk interval probability and the threshold crossing probability, a normalized comprehensive risk value is generated and output to the multi-level state assessment module for comprehensive health score calculation.

[0044] Combination Figure 4The multi-level state assessment module integrates the component state parameters and system performance parameters output by the CEPIN network module, as well as the normalized comprehensive risk value output by the dynamic risk assessment module. First, it performs a reliability binary classification judgment and outputs a judgment result of "reliable" or "unreliable". Under the premise of being judged as reliable, it further calculates the comprehensive performance score and outputs one of four health levels: "excellent", "good", "average" or "needs improvement". Finally, the reliability status, health level and corresponding early warning information are output in real time for the collaborative robot control system to call, so as to realize online continuous monitoring and early warning of motor reliability status.

[0045] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. A reliability status monitoring system for a permanent magnet synchronous motor used in collaborative robots, characterized in that, It includes a data acquisition module, a physical information hybrid degradation modeling module, a finite element simulation and dataset construction module, a CEPIN network module, a dynamic risk assessment module, and a multi-level state assessment module; The data acquisition module is used to collect accelerated degradation test data of permanent magnets and winding insulation, as well as online operation data of motors, providing a data foundation for subsequent modeling and online inference. The physical information hybrid degradation modeling module is used to combine the Arrhenius equation with Gaussian process regression based on accelerated degradation test data to establish temperature-degradation stochastic models for permanent magnets and winding insulation, respectively, and to quantify the degradation law of key components. The finite element simulation and dataset construction module is used to construct a multi-source feature dataset covering permanent magnet state, winding state and torque characteristics based on the degradation state quantization results and through an efficient sampling strategy combining finite element simulation with Latin hypercube sampling and K-means clustering. The CEPIN network module is used to predict the health status parameters of key components such as permanent magnet demagnetization rate and winding insulation state by using a multi-source feature dataset as training input and a dual-branch fusion architecture of residual feature extraction branch and structural causal reasoning branch. The dynamic risk assessment module is used to construct a time-varying parameter probability distribution risk model based on the component health status parameters output by the CEPIN network module, dynamically quantify the comprehensive operating risk of the motor, and output a normalized comprehensive risk value. The multi-level status assessment module is used to integrate component status parameters, system performance parameters, and normalized comprehensive risk values, and output progressive health status assessment results from a binary classification of "reliable / unreliable" to a four-level classification of "excellent / good / average / needs improvement", thereby realizing online continuous monitoring and early warning of motor reliability status.

2. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 1, characterized in that, The accelerated degradation test data collected by the data acquisition module includes: the degradation data of the surface magnetic field strength over time obtained by baking and aging permanent magnet samples at multiple constant temperature levels, and the degradation data of insulation resistance over time obtained by isothermal aging of winding insulation samples at multiple constant temperature levels. The online operating data of the motor includes: permanent magnet temperature signal collected by a temperature sensor installed near the permanent magnet, winding temperature signal collected by a temperature sensor installed near the winding, output torque signal collected by a torque sensor, speed signal collected by an incremental encoder, and phase current signals collected by a current sensor. The above signals are output after being synchronously collected by a multi-channel data acquisition unit.

3. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 2, characterized in that, The specific method by which the physical information hybrid degradation modeling module establishes the degradation model is as follows: A quantitative relationship between degradation rate and temperature is established based on the Arrhenius relation, and a physical trend term for the magnetic field strength of permanent magnets and winding insulation resistance is established using an exponential decay function. The test time at each temperature level was converted into a uniform equivalent time at room temperature using an acceleration factor, and the residual between the measured degradation data and the trend term was calculated. Gaussian process regression is used to learn the mapping relationship between equivalent time and residual. A frequency-aware kernel function is used for permanent magnets, and an additional fluctuation frequency term based on temperature calculation is introduced to capture the frequency difference of magnetic field strength fluctuation with temperature and time. A kernel function that considers the joint correlation of time difference and temperature difference is used for winding insulation. The final predicted output is the sum of the physical trend term and the Gaussian process residual term.

4. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 3, characterized in that, The specific implementation method of the finite element simulation and dataset construction module is as follows: A three-dimensional finite element model of a permanent magnet synchronous motor was established. Uniform demagnetization conditions were set in the model. The maximum, minimum and average values ​​of the output torque under different demagnetization degrees were obtained by gradually reducing the residual magnetic induction intensity of the permanent magnet. Inject inter-turn short-circuit faults with different numbers of turns into the model and obtain the corresponding phase current waveform data; Latin hypercube sampling is used to uniformly generate candidate sample points in a multidimensional feature space. Then, K-means clustering is used to screen representative typical working point points for finite element simulation, so as to significantly reduce the number of simulations while ensuring data representativeness. Finally, a multi-source feature dataset was constructed, which includes the surface magnetic field strength of permanent magnets, demagnetization rate, winding insulation resistance, output torque characteristics, and phase current characteristics.

5. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 4, characterized in that, The specific implementation method of the residual feature extraction branch of the CEPIN network module is as follows: First, the original feature vector is standardized by mean-variance to eliminate the numerical scale difference between features of different dimensions. The standardized features are fed into a residual learning module constructed from multiple fully connected layers. Each residual unit adopts a skip connection structure, which superimposes the transformation result of the current layer with the unit input and outputs the result. The activation function of each layer is ReLU. The output of the residual module is transformed by the linear output layer to predict key health parameters such as permanent magnet demagnetization rate and winding insulation status.

6. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 5, characterized in that, The specific implementation method of the structural causal reasoning branch of the CEPIN network module is as follows: Based on a predefined causal graph of a permanent magnet synchronous motor structure, the causal graph includes nodes such as temperature stress, permanent magnet demagnetization rate, winding insulation resistance, output torque, phase current distortion, and system reliability, as well as directed causal edges between nodes that conform to physical laws. Using demagnetization rate, surface magnetic field strength, and ambient temperature as inputs, the causal characteristics of the output torque of the permanent magnet state-guided theory are established through multiple linear regression. At the same time, winding state characterization quantities are directly extracted from the original winding data as winding causal features; After the two types of causal features are fused according to the set weights, they are concatenated with the original feature matrix of the residual feature extraction branch to form the final input feature matrix, which is then sent to the prediction output layer.

7. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 6, characterized in that, The specific implementation method of the dynamic risk assessment module is as follows: The current operating risk state of the motor is modeled as a time-varying normal distribution, with the mean of the distribution decaying exponentially over time and the variance increasing linearly over time. A time-varying distribution is also established for the risk assessment threshold, so that the assessment boundary can dynamically adapt to the actual degradation process of the motor. The risk state space is divided into multiple continuous intervals. The probability of the motor state falling into each risk interval at each time moment is calculated using the cumulative distribution function of the standard normal distribution. The probability of the system being in a high-risk state is represented by the sum of the probabilities of all intervals exceeding the preset risk threshold. A normalized comprehensive risk value is generated and output by combining the probability of high-risk intervals and the probability of threshold crossing.

8. The reliability status monitoring system for a permanent magnet synchronous motor for a collaborative robot as described in claim 7, characterized in that, The specific implementation method of the multi-level state assessment module is as follows: The reliability criteria are a demagnetization rate of no more than 30% and an insulation resistance of no less than 5 megohms. If either index exceeds the above threshold, it is judged as unreliable. A system state function is introduced to give a unified time-varying mathematical expression to the above criteria. Under the premise of being deemed reliable, the comprehensive performance score is further calculated by weighted summation of five sub-scores: demagnetization score, winding insulation score, torque stability score, comprehensive risk score, and time degradation item. The weight of each sub-score is determined by regression fitting of historical degradation data. The overall score is divided into four health levels according to preset threshold ranges: above 0.8 is "excellent", between 0.6 and 0.8 is "good", between 0.4 and 0.6 is "average", and below 0.4 is "needs improvement". Ultimately, the reliability status, health level, and early warning information will be output in real time.

Images