Insulating material performance evaluation and formula optimization method, equipment and medium

Abstract

The invention discloses an insulating material performance evaluation and formula optimization method and device and a medium, and the method comprises the steps: extracting multi-dimensional data features according to the multi-physical field data of an insulating material, carrying out the feature fusion, and obtaining multi-modal fusion features; constructing a material performance prediction model based on the multi-modal fusion features, and responding to real-time multi-physical field data to obtain a residual life prediction result of the insulating material; determining a key factor influencing the life based on the residual life prediction result, and mapping data characteristics of the key factor to an insulation material failure mechanism to determine a failure factor weight; based on the failure factor weight, establishing a formula optimization objective function by adopting a multi-objective reinforcement learning algorithm; solving the formula optimization objective function based on constraint conditions to obtain a material formula optimization result; the accuracy of the residual life prediction result of the insulating material is improved, and meanwhile, the reliability and pertinence of the formula optimization result of the insulating material are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of insulation material performance evaluation and process optimization, in particular to an insulation material performance evaluation and formula optimization method, device and medium. BACKGROUND

[0002] Currently widely used materials such as epoxy resin, silicone rubber and polyimide are prone to aging failure under complex working conditions such as high temperature, corona and mechanical stress. Traditional offline detection methods rely on destructive sampling and periodic sampling, which have defects such as long detection period, large sample loss and poor data continuity. Especially, it is difficult to meet the real-time monitoring needs of the crosslinking degree change of thermosetting materials such as epoxy resin during high temperature aging, resulting in low material performance evaluation efficiency.

[0003] Under complex working conditions, the coupling effect of multiple physical fields aggravates the material aging process, and the traditional single stress test method cannot simulate the actual operating environment, the test data is insufficient, and it is difficult to track the micro-chemical structure evolution of the material under the coupling effect of electric, thermal, mechanical and other stresses in real time. It is also impossible to establish a quantitative mapping relationship between the micro-characteristics such as molecular chain rupture and crosslinking degree change and the macro-electrical performance parameters such as breakdown voltage and dielectric loss factor. The traditional single stress test method cannot simulate the actual operating environment, and it is difficult to reveal the internal mechanism of material aging, resulting in insufficient accuracy of the life prediction model, and unable to meet the demand of intelligent power grid for accurate evaluation of equipment state.

[0004] In summary, the current technical system has the problem of closed loop fracture of "performance evaluation-material optimization". Although the existing technology can achieve rough material performance evaluation, the evaluation results cannot effectively guide the material formula optimization, and there is a lack of intelligent algorithm support based on multi-dimensional performance data, which makes it difficult to realize precise regulation of material composition and structure through reverse engineering, resulting in long development cycle of new insulation materials, high trial and error cost, and inability to dynamically respond to the differentiated needs of actual working conditions on material performance. SUMMARY

[0005] The purpose of the present application is to solve the problems of low efficiency and poor effect of conventional insulation material performance evaluation and inability to effectively link material formula defect optimization. A kind of insulation material performance evaluation and formula optimization method, equipment and medium are provided, by real-time acquisition of multi-physical field data, extraction of time sequence signal characteristics and microstructure degradation characteristics, multi-modal feature fusion is carried out, and the remaining life of insulation material is predicted based on material performance prediction model;Based on the remaining life prediction result, the key factor affecting the life is determined, the related data characteristics of the key factor are mapped to the material failure mechanism to determine the failure factor weight, and the formula optimization objective function is established, and the material formula optimization data is obtained by solving the optimization objective function, which overcomes the problems of low efficiency and poor effect of conventional insulation material performance evaluation and inability to effectively link material formula defect optimization, realizes the predictive maintenance of material performance, and improves the prediction efficiency and material formula optimization efficiency.

[0006] To achieve the above purpose, the technical scheme adopted by the embodiments of the present application is as follows: In a first aspect, the embodiments of the present application provide an insulation material performance evaluation and formula optimization method, which comprises: extracting multi-dimensional data features from multi-physical field data of insulation materials and performing feature fusion to obtain multi-modal fusion features; Based on the multi-modal fusion features, a material performance prediction model is constructed, and the remaining life prediction result of the insulation material is obtained in response to real-time multi-physical field data; Based on the remaining life prediction result, the key factor affecting the life is determined, and the data characteristics of the key factor are mapped to the insulation material failure mechanism to determine the failure factor weight; Based on the failure factor weight, a formula optimization objective function is established by using multi-objective reinforcement learning algorithm; Based on the constraint condition, the formula optimization objective function is solved to obtain the material formula optimization result.

[0007] In the scheme, multi-dimensional data of the insulating material under multiple physical fields is acquired to fully obtain the relationship between the chemical structure and the macroscopic electrical performance of the material, thereby providing sufficient and accurate basic data for subsequent material residual life prediction and formula optimization; a model is constructed based on multi-modal fusion features, and the model is used to predict the residual life of the insulating material according to real-time multi-physical field data, thereby realizing real-time prediction of the residual life of the insulating material, reflecting the performance change and residual life of the insulating material in real time, and improving the performance evaluation efficiency of the insulating material; the residual life prediction result is associated with the material formula, the data features of the determined key factors are mapped to the failure mechanism of the insulating material, the weight of the failure factor is determined, and it can be clearly understood which factors play a key role in the performance and life of the insulating material, thereby improving the effectiveness of subsequent formula optimization; further, a multi-objective reinforcement learning algorithm is used to establish a formula optimization objective function, thereby providing a clear direction for formula optimization, enabling targeted adjustment and optimization of key factors affecting material performance, taking into account multiple optimization objectives at the same time, avoiding one-sidedness caused by considering only a single objective, and improving the efficiency and accuracy of formula optimization. The formula optimization objective function is solved under given constraints to ensure the feasibility and rationality of the formula optimization result.

[0008] Preferably, the multi-dimensional data features are extracted from the multi-physical field data of the insulating material and are fused, multi-modal fusion features are acquired, including: The data of the high-voltage insulating material under multiple physical fields and environmental parameters are collected, wherein the multi-physical field data at least includes: dielectric spectrum, partial discharge pulse, SEM image and spectrum signal, and the environmental parameters at least include: temperature, humidity and mechanical load; multi-modal physical information neural network is used to extract and fuse features of the multi-physical field data, to acquire time sequence features and microstructure features; Synchronously, the temperature, humidity and mechanical load are normalized, and environmental stress factors and material time sequence cumulative damage amounts are calculated to acquire environmental features; The time sequence features, microstructure features and environmental features are fused as the multi-modal fusion features.

[0009] Preferably, the multi-modal physical information neural network is used to extract and fuse features of the multi-physical field data, to acquire time sequence features and microstructure features, including: Based on a long short-term memory-convolutional neural hybrid network, time-frequency domain local features of the partial discharge pulse and time sequence dependent features of the dielectric spectrum are extracted; The time-frequency domain local features and the time sequence dependent features are weighted and combined to obtain the time sequence feature vector; extracting crack features of the SEM image and chemical structure features of the spectrum signal based on a residual-convolutional neural hybrid network, dimensionally reducing and splicing the crack features and the chemical structure features to obtain the microstructure feature vector.

[0010] Preferably, the material performance prediction optimization model is constructed based on the multi-modal fusion features, and a remaining life prediction result of the insulating material is obtained in response to real-time multi-physical field data. The multi-modal fusion features are used as input parameters of a physical information neural network to construct a material performance prediction model. The material performance prediction model is optimized based on physical constraints to construct the material performance prediction optimization model. The real-time multi-physical field data and environmental parameters are predicted based on the material performance prediction optimization model, and the mean and variance of the prediction are calculated to obtain a remaining life prediction probability of the insulating material by Gaussian mixture calculation, thereby obtaining the remaining life prediction result.

[0011] Preferably, the material performance prediction optimization model is constructed by optimizing the material performance prediction model based on physical constraints. The physical constraints include a space charge equation residual constraint and an Arrhenius aging dynamic constraint, wherein the space charge equation residual constraint is used as a regularization term to establish a loss function for one-time fine-tuning optimization of the material performance prediction model. The material performance prediction model is fine-tuned twice based on the Arrhenius aging dynamic constraint to complete the optimization of the material performance prediction model and obtain the material performance prediction optimization model.

[0012] Preferably, the key factor affecting the life is determined based on the remaining life prediction result, and the data features of the key factor are mapped to the failure factor weight of the insulating material failure mechanism. The SHAP value representing the contribution of each modal feature in the multi-modal fusion features is obtained by performing SHAP analysis on the remaining life prediction result combined with the corresponding modal features. The SHAP values are arranged in descending order to determine the key factor affecting the life of the insulating material. The SHAP values of the key factor are mapped to the insulating material failure mechanism, and the SHAP values are corrected based on the actual contribution value of the insulating material failure mechanism. The corrected SHAP values are converted into the initial weight corresponding to the key factor to obtain the failure factor weight.

[0013] Preferably, the formula optimization objective function is established by using a multi-objective reinforcement learning algorithm based on the failure factor weight, including: The multi-modal fusion features are fused with the formula parameters of the insulating material as a state space of the multi-objective reinforcement learning algorithm. Synchronously, the formula parameters of the insulating material are encoded, and the encoded data are used as an action space of the multi-objective reinforcement learning algorithm; a weighted multi-objective optimization function is constructed based on the failure factor weight, combined with the state space and the action space, and is used as the formula optimization objective function.

[0014] Preferably, the formula optimization objective function is solved based on a constraint condition to obtain a material formula optimization result, including: The formula optimization objective function is solved based on the physical property constraint and the process constraint of the insulating material under the multi-physical field to obtain the formula optimization parameters of the insulating material.

[0015] In a second aspect, an embodiment of the present application provides a computer device, including a processor, a memory and a network interface, the memory stores machine readable instructions executable by the processor, when the computer device is running, the processor and the memory communicate through the network interface, and the processor executes the machine readable instructions to perform the steps of the insulating material performance evaluation and formula optimization method in the first aspect.

[0016] In a third aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is run by a processor to perform the steps of the insulating material performance evaluation and formula optimization method in the first aspect.

[0017] The present application has the following beneficial effects: 1. By calculating the environmental stress factor and the time sequence cumulative damage of the material, the environmental characteristics are obtained, which can convert complex environmental factors into quantifiable ones, help to comprehensively evaluate the influence of the environment on the insulation material, provide a quantitative basis for evaluating the performance change of the insulation material in the actual operating environment, and further provide a more accurate data basis for the residual life prediction of the insulation material; 2. The residual error of the space charge equation is added to the loss function as a regularization term, which can limit the value range of the model parameters, make the model pay more attention to the characteristics and relationships that conform to the physical laws, and improve the generalization ability of the model. By embedding the Arrhenius aging law characteristic aging time into the LSTM network hidden state update, according to the material aging condition at the current temperature, the memory cell state is dynamically adjusted to better remember and transfer the information related to thermal aging, and the influence of temperature on the thermal aging of the insulation material is reflected in a quantitative way, the thermal aging sensitivity is accurately reflected, the model captures the thermal aging sensitivity, and ensures that the model can effectively process the insulation material aging data under different temperature conditions, and adapt to the temperature fluctuation working conditions in actual operation, and provide reliable support for the state evaluation and life prediction of the insulation material under different temperature environments; 3. The modified SHAP value is converted into the initial weight of the key factor to obtain the failure factor weight, which can quantitatively measure the influence degree of each key factor on the life of the insulation material. The formula optimization objective function is established, the insulation material formula optimization problem is converted into a mathematical optimization problem, which is beneficial to calculate and solve the target optimization function according to the physical constraints and process constraints of the insulation material, improve the calculation efficiency, and then efficiently ensure that the insulation material has appropriate mechanical properties and electrical properties under the premise of determining reasonable raw material ratio and processing parameters, and realizing the balance between insulation material performance and production feasibility. BRIEF DESCRIPTION OF DRAWINGS

[0018] Other features, objects, and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, read in conjunction with the accompanying drawings. The drawings are only for the purpose of illustrating preferred embodiments and are not to be considered as limiting the application. Moreover, like reference numerals designate like parts throughout the drawings.

[0019] Figure 1 A flow chart of an insulation material performance evaluation and formula optimization method is provided for the embodiments of the application.

[0020] Figure 2 A structural schematic diagram of a computer device is provided for the embodiments of the application. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and examples. It should be understood that the specific implementation described herein is only one of the best embodiments of the present application, which is used to explain the present application and does not limit the protection scope of the present application. All other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0022] Embodiment 1: As shown in the figure, a method for evaluating the performance of an insulating material and optimizing the formula, comprising the following steps: step S101, extracting multi-dimensional data features and performing feature fusion according to multi-physical field data of the insulating material, and obtaining multi-modal fusion features. Figure 1

[0023] Specifically, the above step S101 comprises: Collecting data and environmental parameters of high-voltage insulating materials under multi-physical fields, wherein the multi-physical field data at least includes: dielectric spectrum, partial discharge pulse, SEM image and spectrum signal, and the environmental parameters at least includes: temperature, humidity and mechanical load; using a multi-modal physical information neural network to extract and fuse features of the multi-physical field data, to obtain time series features and microstructure features; Synchronously, normalizing the temperature, humidity and mechanical load, and calculating the environmental stress factor and the material time series cumulative damage amount to obtain the environmental features; Fusing the time series features, the microstructure features and the environmental features as the multi-modal fusion features.

[0024] In some embodiments, the time series feature vector extracts time series features according to the collected dielectric spectrum and partial discharge pulse, and converts the extracted features into a fixed-dimensional time series vector, thereby obtaining the time series features; and the microstructure feature vector is obtained after feature extraction according to the collected SEM image and spectrum signal. Before extracting the features, the collected time series data can be standardized, for example, Z-score standardization, to eliminate the dimensional differences of physical quantities such as voltage and temperature.

[0025] Further, by extracting features from the SEM image and the spectrum signal, the changes of the physical properties and the chemical properties of the insulating material under the multi-physical field are obtained, thereby obtaining the microstructure features.

[0026] In some embodiments, the scalar parameters such as temperature, humidity and mechanical load are normalized to a unified numerical interval, so as to map the discretized parameters to a continuous vector space, and then calculate the environmental stress factor and the material time series cumulative damage amount with the numerical parameters, to ensure the accuracy of the calculation results and further ensure the reliability of the environmental features. ​

[0027] Specifically, the environmental stress factor S is calculated according to the temperature parameter, humidity parameter and stress amplitude of the environment, and the calculation formula is as follows: S = aT + b ln s + g RH max (1) In the formula, T is the highest temperature in the monitoring period, s is the mechanical vibration stress amplitude, RH is the root mean square value of the environmental humidity, g represents the coupling effect of humidity, a, b and g are respectively the degradation sensitivity coefficients of the insulating material, and the coefficient values can be calibrated through an accelerated aging test. rms max rms

[0028] Specifically, the time sequence cumulative damage D is calculated by combining the environmental parameters with the time sequence characteristic data, and the calculation formula is as follows: In the formula, m is the material fatigue index, N is the partial discharge frequency, AT is the temperature fluctuation amplitude of the tth time step, T is the limit tolerance temperature of the insulating material, and N(PD) is the partial discharge frequency of the tth time step. The time sequence cumulative damage D quantifies the multi-stress coupling effect and drives the subsequent reinforcement learning optimization formula. t lim t

[0029] It can be understood that the performance of high-voltage insulating materials is affected by multiple factors, and multi-source data contains information in different aspects. The multi-modal physical information neural network can extract effective features from different types of data such as time sequence signals, microscopic images and spectra, and fuse these features to obtain a fusion feature vector, which is a multi-modal fusion feature here, to comprehensively reflect the characteristics of the material. Compared with a single feature extraction method, the state of the insulating material can be more comprehensively and accurately described. For example, dielectric spectrum, partial discharge pulse sequence and other time sequence signals can reflect the dynamic response of the material under electrical excitation, and the microscopic features of SEM images and FTIR spectra can reveal the internal structure and molecular level changes of the material. By fusing these features, material information can be obtained from different scales and angles, and the accuracy of material performance evaluation can be improved.

[0030] It can also be understood that different environmental parameters have different dimensions and value ranges, and normalization can convert these parameters into a unified dimensionless value range, eliminating the influence of different parameters on the calculation results due to differences in dimensions and value ranges, avoiding the situation where a parameter value is too large or too small to dominate the calculation results, and ensuring that the model reasonably considers different environmental factors.

[0031] ​​​​​​In this embodiment, the environmental stress factor reflects the comprehensive influence of environmental factors on the performance of the material, and the time-sequential cumulative damage quantity quantifies the damage accumulation of the material over time under the action of the environment. By calculating these two quantities to obtain the environmental characteristics, the complex environmental factors can be converted into quantifiable and comparable characteristic values for evaluating the influence of the environment on the insulating material. This provides a quantitative basis for evaluating the performance changes of the insulating material in the actual operating environment. For example, humidity and salt spray concentration can accelerate material aging, and mechanical load can cause cracks in the material. By considering the influence of these factors on the service life of the material through the environmental stress factor and the time-sequential cumulative damage quantity, it is helpful to accurately predict the remaining life of the material.

[0032] Specifically, the above-mentioned multi-modal physical information neural network is used to extract features and fuse features of the multi-physical field data, to obtain time-sequential features and microstructure features, which include: Based on a long short-term memory-convolutional neural hybrid network, time-frequency domain local features of the partial discharge pulse and time-sequential dependent features of the dielectric spectrum are extracted; The time-frequency domain local features and the time-sequential dependent features are weighted and combined to obtain the time-sequential feature vector; Based on a residual-convolutional neural hybrid network, crack features of the SEM image and chemical structure features of the spectrum signal are extracted, and the crack features and the chemical structure features are dimensionally reduced and spliced to obtain the microstructure feature vector.

[0033] In some embodiments, in the process based on the long short-term memory-convolutional neural hybrid network, the long short-term memory network extracts local features of the partial discharge pulse time series data through a sliding window framing method, including at least discharge frequency, discharge interval time distribution entropy, and pulse amplitude, to capture the time-sequential dependent relationship of the discharge pulse. The convolutional neural network analyzes the dielectric spectrum curve to extract time-sequential dependent features of the dielectric spectrum, including at least feature frequency peak and slope of dielectric loss with frequency.

[0034] In some embodiments, when the SEM image is processed, the image can be threshold segmented and edge detected to extract morphological features of the material surface cracks, such as length, fractal dimension, and crack density, and then a residual network is used to extract deep semantic features, such as crack propagation direction correlation, to obtain the final crack features. Meanwhile, the spectrum signal is first baseline corrected and noise filtered, and then principal component analysis is used to extract the first percentage of principal components in the signal, such as the first 30%, based on which the feature absorption peak change of the molecular bond is represented to obtain the chemical structure features. Further, the crack features and the chemical structure features are dimensionally reduced by PCA to obtain the microstructure feature vector.

[0035] In the embodiment, by fusing the extracted time sequence features, microstructure features and environmental features, a multi-modal fusion feature capable of comprehensively reflecting the state of the insulation material is obtained, avoiding the limitations of single feature or single data type, considering the material properties, microstructure changes and environmental factors, providing more accurate information for subsequent insulation material remaining life prediction, and improving the accuracy and reliability of the insulation material remaining life prediction.

[0036] Step S102, constructing a material performance prediction optimization model based on the multi-modal fusion feature, obtaining the remaining life prediction result of the insulation material in response to real-time multi-physical field data.

[0037] Specifically, the above step S102 includes: The multi-modal fusion feature is used as an input parameter of a physical information neural network to construct a material performance prediction model. The material performance prediction model is optimized based on physical constraints to construct the material performance prediction optimization model. Based on the material performance prediction optimization model, real-time multi-physical field data and environmental parameters are predicted, and the mean and variance of the prediction are calculated. The remaining life prediction probability of the insulation material is calculated by Gaussian mixture to obtain the remaining life prediction result.

[0038] Specifically, the above optimization of the material performance prediction model based on physical constraints to construct the material performance prediction optimization model includes: The physical constraints include space charge equation residual constraint and Arrhenius aging dynamic constraint. The space charge equation residual constraint is used as a regularization term to establish a loss function, and the material performance prediction model is optimized once. The material performance prediction model is optimized twice according to the Arrhenius aging dynamic constraint to complete the optimization of the material performance prediction model and obtain the material performance prediction optimization model.

[0039] In some embodiments, when constructing a prediction model based on a physical information neural network (PINN), the multi-modal fusion feature is used as an input parameter, and the multi-modal fusion feature is represented as f1 represents the time sequence feature, f2 represents the microstructure feature, and S represents the environmental stress factor. represents feature splicing.

[0040] In the PINN architecture, the space charge equation residual constraint and the Arrhenius aging dynamic constraint are embedded. During training, the space charge equation residual constraint is introduced as a regularization term into the loss function to balance the feature data and the material mechanism. The Arrhenius aging dynamic constraint is introduced into the update of the hidden state of the LSTM network to reflect the thermal aging sensitivity of the insulating material through the characteristic aging time at temperature T. The space charge equation residual constraint is expressed as follows: wherein, ρ is the space charge density, the higher the space charge density, the faster the aging speed of the insulating material and the lower the insulating performance, μ is the carrier mobility, d is the diffusion coefficient of the space charge, d reflects the difficulty and rate of charge diffusion in the material, and E is the electric field intensity.

[0041] Further, the loss function is obtained and expressed as follows: wherein, λ represents the physical constraint weight, and the value range is (0.1, 0.5), y pred is the remaining life prediction value, and y true is the true value of the remaining life. In this embodiment, the residual term of the space charge equation is added to the loss function as a regularization term to constrain the neural network prediction to comply with the charge transport law.

[0042] The Arrhenius aging dynamic constraint based on the time series cumulative damage D is expressed as follows: wherein, τ(T) represents the characteristic aging time of the material at temperature T, D ref represents the stress normalized reference value, E a represents the activation energy of the material, k B represents the Boltzmann constant, and T is the absolute temperature. Among them, the Arrhenius equation is modified by replacing the traditional single temperature parameter with the environmental stress factor S to more comprehensively characterize the aging rate under multi-stress coupling.

[0043] Further, the update equation of the LSTM cell state c t is as follows: wherein, f t is the output value of the forgetting gate, and the value range is between 0 and 1. The forgetting gate determines which multi-modal fusion feature information in the cell state c t-1 at the previous time will be retained or forgotten. When f t is close to 1, most of the information in the cell state at the previous time will be retained; and when f tWhen approaching 0, most of the information in the cell state at the previous time will be forgotten. t The output value of the input gate determines how much of the current input x t (i.e., multi-modal feature) and the previous hidden state h t-1 In the combined information, how much new information will be added to the cell state. tanh(W c · [h t-1 , x t ]+b c ) represents a nonlinear transformation, specifically, the multi-modal fusion feature vector is obtained by splicing the previous hidden state h t-1 and the current input x t , W c is a weight matrix for linear transformation of the spliced vector, and b c is a bias term. After nonlinear transformation by the tanh function, a multi-modal fusion feature vector with a value range of -1 to 1 is obtained, representing the newly generated candidate information at the current time, which will be partially added to the cell state under the control of the input gate; T t is the temperature input at time step t. Further, by embedding the effect of temperature on material aging into the LSTM cell state update process, the higher the temperature, the faster the material feature aging time at temperature T, the cell state update is accelerated, reflecting the high temperature accelerating material aging, which has a more significant effect on the LSTM memory unit state update, enabling the model to better capture temperature-related aging information and further ensure the accuracy of the material remaining life prediction results.

[0044] It can be understood that the space charge equation reflects the physical processes such as charge transport inside the insulating material, and adding its residual as a regularization term to the loss function can make the model consider both the physical mechanism of material internal charge behavior and the relationship between feature data and remaining life when learning, realizing the fusion of physical knowledge and data-driven model. By adding the residual constraint of the space charge equation, the model parameter value range is limited, making the model pay more attention to features and relationships that conform to physical laws, improving the model's generalization ability. By embedding the Arrhenius aging law feature aging time into the LSTM network hidden state update, the memory unit state is dynamically adjusted according to the material aging at the current temperature, better remembering and transmitting information related to thermal aging, realizing the quantitative representation of the effect of temperature on the thermal aging of insulating materials, accurately reflecting thermal aging sensitivity, enabling the model to capture thermal aging sensitivity and ensuring that the model can effectively process insulating material aging data under different temperature conditions, adapting to temperature fluctuations in actual operation, providing reliable support for insulating material state assessment and life prediction under different temperature environments.

[0045] In some embodiments, the Monte Carlo Dropout uncertainty quantification formula can be used to quantify the probability distribution of the remaining life of the material, wherein Dropout is enabled for multiple predictions, and each time a part of the neurons is randomly turned off to simulate the uncertainty of the model; and the Gaussian distribution of multiple predictions is averaged to form a mixed Gaussian distribution p(L), which is represented as follows: wherein K represents the number of forward propagations, for example, K = 100, indicating that 100 forward propagations with random Dropout are performed to generate 100 different prediction results; μ k is the mean of the remaining life output by the kth forward propagation, is the variance of the remaining life output by the kth forward propagation; represents the Gaussian distribution of a single prediction, and represents the possible value range of the remaining life.

[0046] Specifically, by calculating the mean of the remaining life as the expected value to obtain the most likely predicted remaining life, and calculating the standard deviation of the prediction output to reflect the fluctuation range of the prediction and measure the prediction error caused by the uncertainty of the model. wherein the mean μ L and the variance are represented as follows, respectively: wherein L0 is the life of the material under ideal conditions (no stress, no damage, no charge), r s , r D , r ρ are the attenuation coefficients of stress, damage, and charge on the mean life, respectively, and η is a nonlinear correction index.

[0047] In this embodiment, the confidence interval of the prediction result of the remaining life of the insulating material is calculated according to the mean and the standard deviation, and a reasonable range of the remaining life is obtained, for example, the confidence interval = [μ L -1.96σ L , μ L +1.96σ L ].

[0048] In this embodiment, by calculating the prediction mean and the standard deviation, accidental uncertainty caused by multi-modal feature data noise and the like and cognitive uncertainty of the prediction model parameters can be avoided, that is, the accidental uncertainty and the cognitive uncertainty of the neural network prediction are combined, and the probability distribution of the remaining life of the material is output instead of a single value, ensuring the accuracy and reliability of the prediction result of the remaining life of the material.

[0049] In step S103, a key factor affecting the service life is determined based on the remaining life prediction result, and data characteristics of the key factor are mapped to a failure factor weight determined by an insulation material failure mechanism.

[0050] Specifically, step S103 includes: SHAP analysis is performed on the remaining life prediction result in combination with the corresponding modal characteristics to obtain SHAP values representing the contribution degrees of the modal characteristics in the multi-modal fusion characteristics; The SHAP values are arranged in descending order to determine a key factor affecting the service life of the insulation material; The SHAP values of the key factor are mapped to an insulation material failure mechanism, and the SHAP values are corrected based on actual contribution values of the insulation material failure mechanism; The corrected SHAP values are converted into initial weights corresponding to the key factor to obtain the failure factor weight.

[0051] In this embodiment, the SHAP values of the key factor are mapped to the insulation material failure mechanism, and a feature-mechanism mapping table can be established, as shown in Table 1. Table 1 Feature-mechanism mapping table Multi-modal features Failure mechanism Physical / chemical correlation basis Partial discharge frequency Electrochemical aging Discharge-induced molecular chain scission, increased dielectric loss Fractal dimension of crack Mechanical fatigue crack propagation Crack tip field concentration, reduced breakdown field FTIR oxidation index Oxidative degradation Increased C=0 bonds indicate oxidative chain scission Root mean square value of humidity RH rms ]]> Surface contamination / hydrolysis reaction Moisture penetration induces interfacial debonding, increased leakage current The failure factor is the corresponding failure mechanism type affecting the modal mapping of the remaining life of the insulation material, such as electrochemical aging and mechanical fatigue, which is determined by the SHAP values calculated from the prediction result and the feature data sample.

[0052] In step S104, a multi-objective reinforcement learning algorithm is used to establish a formula optimization objective function based on the failure factor weight.

[0053] Specifically, step S104 includes: The multi-modal fusion characteristics are fused with the formula parameters of the insulation material to serve as a state space of the multi-objective reinforcement learning algorithm; Simultaneously, the formula parameters of the insulation material are encoded, and the encoded data are used as an action space of the multi-objective reinforcement learning algorithm; a weighted multi-objective optimization function is constructed based on the failure factor weight, in combination with the state space and the action space, and is used as the formula optimization objective function.

[0054] In some embodiments, the formula parameters of the insulation material include a base material type, an additive ratio, and a process condition. These data are fused with multi-modal characteristic data such as environmental temperature, humidity, mechanical vibration amplitude, partial discharge frequency, and crack fractal dimension to obtain a representation of the state space, for example: state space s t = [s1, s2] ∈ R dWherein, s1 represents a formula parameter, and s2 represents a multi-modal fusion feature.

[0055] Further, the proportion of formula ingredients and process parameters are dynamically adjusted, and an additive type is selected, and the action space preset is completed by encoding the action, for example, the action space a t = [Δx1, Δx2, …, Δx n , additive type] ∈ R n+1 , so as to construct a multi-objective reward function, which is represented as follows: Wherein, ω j is a failure factor weight, j = 1, 2, 3, …, M, M represents the category of failure factors; R j is a sub-reward item of the target reward function, and Cost(a t ) represents a cost penalty item; wherein: Wherein, E b represents a residual breakdown field strength, Wherein, E0 represents an initial breakdown field strength, e is an exponential function, ι is a degradation sensitivity coefficient of the material, D f is a crack fractal dimension, and the greater the value, the more complex the crack is, E min and E max are minimum and maximum values of the breakdown field strength respectively; υ represents a tensile strength, υ min and υ max represent minimum and maximum values of the tensile strength respectively; OI is an oxidation index of the insulating material, OI min and OI max are minimum and maximum values of the oxidation index respectively; θ is a contact angle, θ min and θ max are minimum and maximum values of the contact angle respectively. Wherein, c i is a raw material unit price, and c′ is a process energy consumption coefficient.

[0056] In step S105, the formula optimization objective function is solved based on the constraint condition, and a material formula optimization result is obtained.

[0057] Specifically, the above step S105 includes: The formula optimization objective function is solved based on the physical property constraint and the process constraint of the insulating material under the multi-physical field, and a formula optimization parameter of the insulating material is obtained.

[0058] In some embodiments, the key performance indicators to be optimized are determined according to the failure factor weight, and a failure mechanism-optimization target mapping table is established, as shown in Table 2, Table 2 Failure mechanism-optimization target mapping table Failure mechanism Corresponding performance indicator Optimization direction Electrochemical aging breakdown field strength E b ]]> Maximize (inhibit charge injection / accumulation) Mechanical fatigue Tensile strength υ Maximize (improve toughness) Oxidative degradation Oxidation index OI Minimize (add antioxidant) Surface contamination Contact angle θ Maximize (enhance hydrophobicity) In some embodiments, the physical constraints include the elongation at break and the dielectric loss, for example, the elongation at break ≥ 500%, and the dielectric loss tan δ ≤ 0.005; the process constraints include the curing temperature range and the pressure value range, for example, 100℃ ≤ T cure ≤ 180℃, and 0.1MPa ≤ P pressure ≤ 5MPa. According to the physical constraints, the multi-objective function is solved, the specific reward value of each failure factor is calculated, the optimization direction is determined according to the reward value, and the formula parameters of the insulating material are correspondingly dynamically adjusted, for example, the proportion of nano-TiO2 is increased to inhibit charge accumulation, reduce local field strength, optimize the proportion of crosslinking agent to improve mechanical strength, and add hydrophobic additives to improve the contact angle.

[0059] In this embodiment, by performing SHAP analysis on the prediction results, the contribution of each feature to the model prediction results can be quantified in a unified standard, so that the importance of time sequence, microstructure, environment and other multi-modal features in predicting the remaining life of the insulating material can be clearly understood; the SHAP values are arranged in descending order, and the key factors that significantly affect the life of the insulating material can be intuitively screened out; further mapping the SHAP values to the failure mechanism of the insulating material can establish the connection between the feature contribution and the actual physical process, so as to correct the SHAP values according to the actual contribution of the failure mechanism, and make the feature contribution more consistent with the actual situation. The corrected SHAP values are converted into initial weights of the key factors, and the failure factor weights are obtained, which can quantitatively measure the influence degree of each key factor on the life of the insulating material. The formula optimization objective function is established, the insulating material formula optimization problem is converted into a mathematical optimization problem, which is conducive to calculating and solving the target optimization function according to the physical constraints and process constraints of the insulating material, improves the calculation efficiency, and further efficiently ensures that the insulating material has appropriate mechanical properties and electrical properties, determines reasonable raw material ratio and processing parameters, and balances the performance of the insulating material and the feasibility of production.

[0060] The embodiments of the present application also provide a computer device, as shown in Figure 2 The computer device provided by the embodiments of the present application includes a processor 21, a memory 22 and a network interface 23. The memory 22 stores machine readable instructions executable by the processor 21. When the computer device is running, the processor 21 communicates with the memory 22 through the network interface 23. When the machine readable instructions are executed by the processor 21, the steps of the insulating material performance evaluation and formula optimization method in the above embodiments are performed.

[0061] The embodiment of the present application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is run by a processor, steps of the insulation material performance evaluation and formula optimization method in the above embodiment are executed.

[0062] The above detailed description is a preferred embodiment of the present application, and is not intended to limit the specific implementation range of the present application. The range of the present application includes but is not limited to the above detailed description. Any equivalent changes made according to the shape, structure and method of the present application are within the protection scope of the present application.

Claims

1. A method for evaluating and optimizing insulation material performance, characterized by: The steps include: Extract multidimensional data features based on the multi-physical field data of insulating materials and perform feature fusion to obtain multimodal fusion features; Building a material performance prediction optimization model based on the multimodal fusion features, and obtaining a remaining life prediction result of the insulation material in response to real-time multi-physics field data; Determining key factors affecting the lifespan based on the remaining lifespan prediction result, and mapping data features of the key factors to the failure mechanism of the insulation material to determine failure factor weights; Based on the failure factor weights, a multi-objective reinforcement learning algorithm is used to establish a recipe optimization objective function; The formulation optimization objective function is solved based on the constraint conditions to obtain the material formulation optimization result.

2. The insulation material performance evaluation and formulation optimization method according to claim 1, characterized in that: The method of extracting multidimensional data features based on the multi-physical field data of the insulating material and performing feature fusion to obtain multimodal fusion features includes: Collecting data and environmental parameters of high-voltage insulation materials under multi-physical fields, where the multi-physical field data includes at least dielectric spectrum, partial discharge pulses, SEM images, and spectral signals, and the environmental parameters include at least temperature, humidity, and mechanical load; using a multimodal physical information neural network to perform feature extraction and feature fusion on the multi-physical field data to obtain temporal features and microstructural features; Synchronously, normalizing the temperature, humidity, and mechanical load, and calculating an environmental stress factor and a time-series cumulative damage amount of a material to obtain environmental characteristics; The temporal features, the microstructure features and the environmental features are fused as the multimodal fusion features.

3. The insulation material performance evaluation and formulation optimization method according to claim 2, characterized in that: The multimodal physical information neural network is used to perform feature extraction and feature fusion on the multi-physical field data to obtain time series features and microstructure features, including: Extracting the local time-frequency characteristics of the partial discharge pulse and the time-dependent characteristics of the dielectric spectrum based on a long short-term memory-convolutional neural network hybrid; Performing weighted merging of the time-frequency domain local features and the time series dependent features to obtain the time series feature vector; The crack features of the SEM image and the chemical structure features of the spectral signal are extracted based on a residual-convolutional neural hybrid network, and the crack features and the chemical structure features are reduced in dimension and spliced ​​to obtain the microstructure feature vector.

4. The method for evaluating and optimizing insulation material performance according to claim 1 or 2, wherein: The method of constructing a material performance prediction optimization model based on the multimodal fusion feature and obtaining a remaining life prediction result of the insulation material in response to real-time multi-physical field data includes: Using the multimodal fusion features as input parameters of a physical information neural network to construct a material performance prediction model; The material performance prediction model is optimized based on physical constraints to construct the material performance prediction optimization model; real-time multi-physical field data and environmental parameters are predicted based on the material performance prediction optimization model, and the predicted mean and variance are calculated, and the remaining life prediction probability of the insulating material is obtained using Gaussian mixture calculation to obtain the remaining life prediction result.

5. The insulation material performance evaluation and formulation optimization method according to claim 4, characterized in that: The step of optimizing the material property prediction model based on physical constraints to construct the material property prediction optimization model includes: The physical constraints include a space charge equation residual constraint and an Arrhenius aging dynamic constraint, wherein the space charge equation residual constraint is used as a regularization term to establish a loss function, and a fine-tuning optimization is performed on the material property prediction model; The material property prediction model is subjected to secondary fine-tuning optimization according to the Arrhenius aging dynamic constraint to complete the optimization of the material property prediction model and obtain the material property prediction optimization model.

6. The insulation material performance evaluation and formulation optimization method according to claim 4, characterized in that: The determining of key factors affecting the lifespan based on the remaining lifespan prediction result, and mapping data features of the key factors to the failure mechanism of the insulation material to determine failure factor weights, includes: Performing SHAP analysis on the remaining life prediction result in combination with the corresponding modal features to obtain a SHAP value representing the contribution of each modal feature in the multimodal fusion feature; Arrange the SHAP values ​​in descending order to determine the key factors affecting the life of the insulation material; Mapping the SHAP value of the key factor to the insulation material failure mechanism, and correcting the SHAP value based on the actual contribution value of the insulation material failure mechanism; The corrected SHAP value is converted into the initial weight corresponding to the key factor to obtain the failure factor weight.

7. The insulation material performance evaluation and formulation optimization method according to claim 1, characterized in that: The formulation optimization objective function is established based on the failure factor weights using a multi-objective reinforcement learning algorithm, including: The multimodal fusion features are integrated with the formulation parameters of the insulation material to serve as the state space of the multi-objective reinforcement learning algorithm; Synchronously, the recipe parameters of the insulating material are encoded, and the encoded data is used as the action space of the multi-objective reinforcement learning algorithm; based on the failure factor weight, a weighted multi-objective optimization function is constructed by combining the state space and the action space, and is used as the recipe optimization objective function.

8. The method for evaluating and optimizing insulation material performance according to claim 7, wherein: Solving the formulation optimization objective function based on the constraint conditions to obtain the material formulation optimization result includes: Based on the physical property constraints and process constraints of the insulating material under multi-physical fields, the formulation optimization objective function is solved to obtain the formulation optimization parameters of the insulating material.

9. A computer device, characterized in that: include: A processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate via the bus, and the processor executes the machine-readable instructions to perform the insulation material performance evaluation and formulation optimization method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is run by a processor, the method for evaluating the performance of insulating materials and optimizing the formulation according to any one of claims 1 to 8 is executed.

Images