Crossover capacitance balance type PCS converter

Abstract

The invention relates to the technical field of power electronics, and discloses a bridge connection capacitance balance type PCS converter comprising a data acquisition module used for acquiring PCS system multi-mode sensing data and calculating multi-scale entropy features; the feature extraction module is used for constructing an initial time sequence causal graph model; the structure optimization module is used for optimizing a causal graph structure by adopting a mixed structure learning algorithm; the anti-fact reasoning module is used for establishing a structural causal model and a situation simulator and realizing anti-fact query and analysis; the causal evaluation module is used for calculating an average causal effect and a comprehensive causal influence index and quantifying the influence of each factor on the service life of the capacitor; the health assessment module is used for constructing a system health index in combination with nonlinear conversion and a time decay effect; the maintenance decision module is used for generating an optimal maintenance strategy and evaluating cost effectiveness under a multi-constraint condition according to the health index; according to the method, a deep causal mechanism of capacitance degradation is disclosed, correlation and causality are effectively distinguished, and accurate maintenance decision support is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronics, more particularly, it relates to a cross-capacitor balanced PCS converter. BACKGROUND

[0002] With the rapid development of power electronics technology, power conversion systems (PCS) play an increasingly important role in energy conversion and management. As a key component in the PCS system, the cross-capacitor bears the important functions of smoothing voltage fluctuations, reducing ripple and providing energy buffer. However, during long-term operation, the cross-capacitor faces performance degradation and failure risks, which directly affects the reliability and safety of the entire PCS system.

[0003] Traditional predictive maintenance methods are mainly based on statistical analysis of historical data, lacking understanding of the deep causal mechanism in the degradation process of the capacitor, only focusing on surface phenomena and correlation. This method is difficult to effectively distinguish between correlation and causality, leading to inaccurate identification of key factors affecting capacitor life, lack of targeted maintenance decisions, and inability to develop precise maintenance strategies based on specific causal relationships.

[0004] In addition, traditional methods are difficult to evaluate the actual impact of different operating conditions and interventions on capacitor life, which leads to the fact that the maintenance of cross-capacitors in PCS systems is often not precise, either prematurely maintained causing resource waste, or not timely maintained leading to sudden system failure, increasing maintenance costs and reducing system availability. Therefore, a new method is needed that can deeply understand the degradation mechanism of the cross-capacitor, accurately predict the health status and optimize maintenance decisions. SUMMARY

[0005] The present application provides a cross-capacitor balanced PCS converter, which solves the technical problems in the related art that lack understanding of the deep causal mechanism in the degradation process of the capacitor, cannot effectively distinguish between correlation and causality, and leads to lack of targeted maintenance decisions and inability to accurately evaluate the actual impact of different operating conditions and interventions on capacitor life.

[0006] The present application provides a cross-capacitor balanced PCS converter, comprising: a data acquisition module for acquiring PCS system multi-modal sensor data, including electrical, temperature, vibration and capacitor parameter data, and calculating multi-scale entropy features; a feature extraction module for time series correlation analysis based on multi-scale entropy features to construct an initial time series causal graph model; a structure optimization module for optimizing the causal graph structure using a hybrid structure learning algorithm combined with conditional independence testing and Bayesian scoring; counterfactual reasoning module, establishing structural causal model and scenario simulator, realizing counterfactual query and analysis; causal assessment module, calculating average causal effect and comprehensive causal impact index, quantifying the influence of each factor on the capacitance life; health assessment module, based on causal contribution, combining nonlinear conversion and time decay effect, constructing system health index; maintenance decision module, according to the health index, generating the optimal maintenance strategy under multiple constraint conditions and evaluating the cost-benefit.

[0007] In a preferred embodiment, the multi-modal sensing data collected by the data acquisition module includes: electrical parameter data, temperature data, vibration data and capacitance parameter data, and the multi-scale entropy feature is obtained by calculating the sample entropy at different time scales.

[0008] In a preferred embodiment, the hybrid structure learning algorithm in the structure optimization module includes: constructing the skeleton structure based on conditional independence test; applying direction determination rules to determine the direction of the edge; using Bayesian scoring function to evaluate the overall structure and fine-tuning.

[0009] In a preferred embodiment, the counterfactual reasoning module includes: establishing structural causal model; designing counterfactual query mechanism; constructing counterfactual scenario simulator, wherein the counterfactual query mechanism is realized based on the three-step method of causal reasoning.

[0010] In a preferred embodiment, the causal assessment module quantifies the causal contribution by the following indicators: average causal effect, conditional average causal effect and causal impact index, wherein the causal impact index considers direct effect, indirect effect and mediating effect.

[0011] In a preferred embodiment, the causal impact index is calculated by the weighted combination of normalized direct effect, indirect effect and mediating effect, and the weight coefficient is optimized by maximizing the objective function related to the actual observed system degradation rate.

[0012] In a preferred embodiment, the health assessment module calculates the health index by the combination function, which uses the weighted combination of the causal contribution weight of each factor, the nonlinear conversion function and the dynamic adjustment function of the time decay effect.

[0013] In a preferred embodiment, the maintenance decision module generates the maintenance strategy based on the multi-objective optimization framework, considering the maintenance cost, system reliability and equipment remaining life, and is subject to budget constraints, reliability constraints and time window constraints.

[0014] In a preferred embodiment, the cost-benefit evaluation of the maintenance decision module includes: calculating the maintenance cost function, the system reliability function and the equipment remaining life function, and finding the optimal maintenance strategy that balances various objectives through a Pareto optimal solution screening mechanism.

[0015] In a preferred embodiment, a computer-readable storage medium for storing computer-readable instructions capable of running a cross-capacitor balanced PCS converter when read by a computer.

[0016] The beneficial effects of the present application are: Deep causal mechanism reveals that the cross-capacitor balanced PCS converter method can discover potential causal relationship structures from massive multi-modal sensor data through the combination of temporal causal graph model and hybrid structure learning algorithm, and reveal the deep causal mechanism in the capacitor degradation process. This structured causal understanding far exceeds the surface analysis of traditional methods relying only on statistical correlation, enabling the system to accurately identify the real key factors affecting capacitor life, rather than being misled by surface phenomena.

[0017] Effective distinction between correlation and causality, the invented counterfactual reasoning framework can strictly distinguish between correlation and causality through the "hypothetical intervention" method. By evaluating the effect of intervention operations rather than simply observing the conditional probability of variables, the identification of true causal relationships is achieved, avoiding the risk of being misled by pseudo-correlation phenomena. This enables the system to exclude factors that show high correlation with capacitor life but have no causal effect, thereby focusing maintenance resources on truly effective aspects.

[0018] Precise maintenance decision support, based on the quantitative impact assessment of each operating factor on capacitor life provided by the causal contribution calculation module, the present application can generate targeted maintenance strategies. These maintenance strategies not only consider "what" needs to be maintained, but also explicitly "why" maintenance is needed and "how" to maintain it most effectively, making maintenance decisions have scientific basis and interpretability. Through the counterfactual scenario simulator, the system can predict the impact of different operating conditions and intervention measures on capacitor life without the need to actually perform potentially risky experiments.

[0019] The balance between resource optimization and reliability improvement is achieved by the cost-benefit evaluation module, which finds the best balance point between maintenance cost control and PCS system reliability improvement, avoids resource waste caused by excessive maintenance, and prevents PCS system failure risk caused by insufficient maintenance. The health index prediction model based on causality is significantly better than the traditional method, which can more accurately reflect the actual health status of the capacitor, thereby realizing the establishment of the capacitor balancing mechanism, prolonging the service life of the capacitor, improving the overall reliability of the system, and optimizing the allocation efficiency of maintenance resources. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a module diagram of a bridging capacitor balancing type PCS converter of the present application; Figure 2 is a line chart of the trend of the multi-scale entropy feature with the change of the capacitor usage time of the present application; Figure 3 is a Sankey diagram of the contribution degree analysis of the capacitor degradation key causal path of the present application; Figure 4 is a column chart of the effect comparison of the counterfactual intervention strategy of the present application; Figure 5 is a scatter matrix diagram of the comparison of the causal reasoning and correlation analysis of the present application; Figure 6 is a column chart of the comparison of the maintenance decision based on causality and the traditional method of the present application; Figure 7 is a line chart of the time series performance evaluation of the capacitor health index prediction model of the present application. DETAILED DESCRIPTION

[0021] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described in some examples can be combined in other examples.

[0022] At least one embodiment of the present application discloses a bridging capacitor balancing type PCS converter, as shown in Figure 1 comprises: A data acquisition module for acquiring PCS system multi-modal sensor data, including electrical, temperature, vibration and capacitor parameter data, and calculating multi-scale entropy features; A variety of types of sensors are deployed at key positions of the PCS system, including but not limited to: Electrical parameter sensors: collect electrical parameters such as voltage, current, power, etc.; Temperature sensor: Monitor the temperature of the across capacitor and the surrounding environment; Vibration sensor: Collect vibration information during system operation; Capacitor parameter sensor: Monitor the equivalent series resistance (ESR), capacitance, leakage current, and other parameters of the capacitor.

[0023] Preprocess the collected raw data, including: Data cleaning: Remove noise and outliers; Data standardization: Unify different dimensional data to the same scale; Time synchronization: Ensure the time consistency of different sensor data.

[0024] Further, for different types of sensor data, specific data preprocessing methods are used: Electrical parameter data: Use Z-score standardization method to process voltage, current and other high dynamic range data: ; Where z represents the normalized value; x represents the original data value; μ represents the average value of the data set; σ represents the standard deviation of the data set; Temperature data: Use min-max normalization method to map temperature values to the [0, 1] interval: ; Where, represents the normalized temperature value; represents the original temperature value; represents the minimum value in the temperature data set; represents the maximum value in the temperature data set; Vibration spectrum data: Apply logarithmic transformation to reduce the influence of high amplitude frequency components, and use wavelet transform to eliminate background noise; Discrete state parameters: Use One-Hot Encoding to convert classification data such as fault codes and working modes into numerical representation.

[0025] Calculate multi-scale entropy features: Calculate the sample entropy (Sample Entropy) for each type of sensor data, input the time series data, and output the entropy value representing the complexity of the signal; Calculate the entropy value at multiple time scales to form a multi-scale entropy curve; Extract feature parameters such as entropy value mean, entropy value slope, and entropy value fluctuation from the multi-scale entropy curve.

[0026] For example, Figure 2As shown, the figure shows the trend of the change of the entropy value of the capacitor at different scales with the use time. The figure intuitively shows that as the capacitor ages, the signal complexity (entropy value) gradually decreases, providing an important basis for capacitor health state evaluation.

[0027] The feature extraction module performs time correlation analysis based on the multi-scale entropy features, and constructs an initial time causal graph model. The dynamic Bayesian network framework is adopted, the input is multi-modal sensor data and multi-scale entropy features, and the output is the network structure describing the time dependence between variables.

[0028] Further, before constructing the time causal graph model, the input feature data is preprocessed: Time series stationarization: apply difference transformation or Hodrick-Prescott filter to remove time series trend, so that the data meet the stationarity assumption of causal discovery algorithm; Variable type processing: for mixed type variables (continuous, discrete, binary), adopt adaptive coding strategy, including binning discretization of continuous variables and ordered encoding of discrete variables; Time delay data generation: based on domain knowledge, construct variable copies containing different time delays to capture the dynamic causal relationship between variables; Missing value processing: use multiple imputation and expectation maximization algorithm to process missing values in time series data, and maintain the time structure characteristics of the data.

[0029] Further, the dynamic Bayesian network framework introduces a time lag matrix which extends the traditional Bayesian network. The time lag matrix is specifically used to quantify the time dependence between different variables, where the matrix element represents the time impact delay of variable on variable , so that the dynamic Bayesian network framework can capture the non-synchronous causal effect in the capacitor degradation process.

[0030] Perform conditional independence test, input is the time series data of variable pairs, output is the conditional independence judgment result between variables.

[0031] Establish time causal graph structure, input is the conditional independence test result and time lag information, output is the directed acyclic graph representing the causal relationship.

[0032] As shown in Figure 3 , the causal paths and their contributions between the factors affecting the life of the capacitor are shown. Visualization shows how factors such as temperature fluctuations and voltage ripple affect capacitor life through different paths, which helps to identify key influencing factors and intervention points.

[0033] Structure optimization module, using a hybrid structure learning algorithm, combining conditional independence test and Bayesian scoring, optimizes the structure of causal graph; Using a hybrid structure learning algorithm, the input is multi-modal sensor data and conditional independence test results, and the output is the optimized structure of the causal graph.

[0034] Further, the hybrid structure learning algorithm combines the advantages of constraint-based algorithms and scoring-based algorithms, and is implemented through a three-stage iterative optimization process: the first stage constructs a skeleton structure based on conditional independence test; The second stage applies the direction determination rule to determine the direction of the edge; The third stage uses the Bayesian scoring function to evaluate the overall structure and fine-tune, while introducing sparsity constraints to control the complexity of the hybrid structure learning algorithm to prevent overfitting and improve the accuracy of structure learning.

[0035] Performing a learning process, the input is an initial graph structure and training data, and the output is an updated graph structure.

[0036] Applying regularization techniques, the input is a candidate graph structure, and the output is the optimal model structure after complexity control.

[0037] As shown in Figure 4 , the life extension effect and intervention cost of different intervention strategies are compared. By visualizing the effect and cost of different intervention strategies, a scientific basis is provided for maintenance decision-making.

[0038] Counterfactual reasoning module, establish structural causal model and scenario simulator, realize counterfactual query and analysis; Establish a structural causal model, the input is a time series causal graph, and the output is a mathematical model containing the functional relationship between variables.

[0039] Further, the structural causal model clearly distinguishes between endogenous variables and exogenous variables, and establishes a functional relationship equation for each endogenous variable: ; Among them, represents the i-th endogenous variable, i.e. the variable that can be affected by other variables within the system; represents the functional relationship of the i-th variable; represents the parent node set of variable , i.e. the set of all variables that directly affect ; represents the exogenous noise variable, i.e. random disturbance from outside the system.

[0040] Design counterfactual query mechanism, input is the hypothetical intervention condition and the current system state, output is the expected result after intervention.

[0041] Further, the counterfactual query mechanism is based on Pearl's three-step method of causal reasoning: first, abduction step, updating the posterior distribution of exogenous variables according to the observed data; then, action step, applying the hypothetical intervention in the structural causal model ; finally, the prediction step, calculating the intervention effect through the modified structural causal model, thus realizing the accurate reasoning of the "hypothetical situation", breaking through the limitations of traditional correlation analysis. represents the intervention operator, which is used to represent the mathematical form of forcing the variable X to be set to the value x. The structural causal model realizes the accurate intervention simulation capability of the capacitor operating environment by introducing the mathematical formalization representation of the intervention operator .

[0042] The counterfactual scenario simulator is constructed, the input is the intervention strategy and the system model, and the output is the simulated intervention effect analysis.

[0043] As shown in Figure 5 , the correlation coefficients and causal impact indexes of various influencing factors are compared.

[0044] The causal evaluation module calculates the average causal effect and the comprehensive causal impact index to quantify the influence of each factor on the capacitor life; The causal contribution quantification index is defined as: Average causal effect (ACE): ; Where, represents the cause variable; represents the result variable; represents the value of the variable after intervention; represents the baseline value of the variable before intervention; represents the intervention operation of forcing the variable to be set to the value ; , respectively represent the expected value after executing the intervention and ; represents the average causal effect size of to .

[0045] Conditional average causal effect (CACE): ; Where, represents the cause variable; represents the result variable; ​Represents a condition variable; Variables after intervention The value; Variables before intervention The baseline value; Representing condition variables Specific values ​​that can be taken; Indicates the variable Forced to be set to value Intervention procedures; , These respectively indicate the implementation of intervention. and And conditions hour Expected value; Indicates the condition Down right The average causal effect size.

[0046] Causal Impact Index (CII): A comprehensive index that takes into account both direct and indirect effects.

[0047] Furthermore, the causal contribution quantification index needs to undergo data standardization before application to ensure that variables of different dimensions and scales can be effectively compared: For continuous variables (such as voltage, temperature, etc.): a robust MinMax scaling method is used to map the data to the [0, 1] interval, and the minimum and maximum values ​​are replaced with the 5% and 95% quantiles to mitigate the impact of outliers; Time series variables: Apply time series-specific Z-normalization methods to normalize data while preserving temporal characteristics; Dual causal effect: Using contrastive coding on intervention variables makes counterfactual comparisons more interpretable.

[0048] Furthermore, the Causal Influence Index (CII) is defined using a new mathematical formula: ; in, Indicates from variable to variable The causal influence index; Describing the cause variable Represents the outcome variable; This indicates the direct effect after normalization. This indicates the indirect effect after normalization. This indicates the normalized moderating effect; The weighting coefficients representing direct effects; weight coefficient representing indirect effect; weight coefficient representing regulatory effect; constraint condition representing the sum of three weight coefficients as 1.

[0049] The weight coefficients are optimized by maximizing an objective function related to the actual observed system degradation rate: wherein, represents a correlation function; represents an actual observed capacitance degradation rate; represents a maximization operation; represents a causal influence index from variable to variable ; weight coefficient representing direct effect; weight coefficient representing indirect effect; weight coefficient representing regulatory effect.

[0050] The causal influence index (CII) solves the technical problem that traditional single causal effect evaluation methods are difficult to quantify the influence of complex causal networks by comprehensively considering the contribution of multiple causal paths.

[0051] Specifically, each inner function is defined as follows: : direct effect function, calculating the direct causal influence of variable on variable , specifically represented as the average effect change amount generated by the direct path influence , which can be calculated by the formula: wherein, represents the direct effect value from variable to variable Y; represents a cause variable; represents a result variable; represents the value of variable after intervention; represents the baseline value of variable before intervention; represents the set of intermediate variables between and ; represents the specific value of the set of intermediate variables ; represents an expectation operator; represents an intervention operator; represents the specific value of the set of intermediate variables​​ post intervention expectation post intervention

[0052] : indirect effect function, computes the indirect effect of variable on through mediator variable, denoted as: ; where, denotes the indirect effect value from variable to variable ; denotes the cause variable; denotes the outcome variable; denotes the value of variable post intervention ; denotes the value of mediator variable when ; denotes the expectation operator; denotes the intervention operator; post intervention ; denotes the expectation value post intervention

[0053] : moderation effect function, quantifies the moderating effect of other variables on the strength of causal relationship, computed as: ; where, denotes the moderation effect value from variable to variable ; denotes the cause variable; denotes the outcome variable; denotes the moderator variable, used to moderate the strength of effect of on ; denotes the value of variable post intervention ; denotes the baseline value of variable pre intervention ; ​denotes the desired operator; denotes the intervention operator; denotes the desired value when intervention is performed; denotes the desired value when intervention is performed; denotes the desired value when intervention is performed; denotes the desired value when intervention is performed; denotes the desired value when intervention is performed; denotes the desired value when intervention is performed. , and : normalized direct effect, indirect effect and mediation effect, respectively, calculated by Min-Max normalization method to ensure dimensional consistency and additivity: ; wherein, denotes the normalized effect value; denotes the original effect value; denotes the minimum value of the effect; denotes the maximum value of the effect; denotes the cause variable; denotes the result variable; denotes the normalized direct effect value, with a value range of [0, 1]; denotes the normalized indirect effect value, with a value range of [0, 1]; denotes the normalized mediation effect value, with a value range of [0, 1].

[0054] Calculate the causal contribution of each operating factor to the capacitance life, input the counterfactual query result, and output the causal effect size and direction of each factor.

[0055] Build a causal impact ranking, input the causal contribution of each factor, and output a list of key factors that affect the capacitance life.

[0056] As Figure 6 shown, compare the performance of the embodiment and the traditional method in terms of prediction accuracy, capacitance life and other indicators.

[0057] Health assessment module, based on causal contribution, combined with nonlinear conversion and time decay effect, to build a system health index; Design health index based on causal structure, input causal contribution of key factors and capacitance history parameters, output health status indicators.

[0058] Further, before building the capacitor health index prediction model, the input parameters need to be uniformly pre-processed: Time window segmentation: divide the historical time series data into fixed length time windows, and extract features using sliding window technology; Multi-scale feature extraction: calculate statistical features (mean, standard deviation, kurtosis, skewness) and frequency domain features (power spectral density, harmonic distortion) for each time window; Feature selection and dimensionality reduction: combine causal structure information, apply principal component analysis (PCA) and recursive feature elimination (RFE) methods to retain a subset of features highly correlated with health status; Imbalanced data processing: to address the problem of few fault samples, use the synthetic minority over-sampling technique (SMOTE) to balance the data distribution.

[0059] Further, the capacitor health index is calculated by a new combination function: ; Where, represents the capacitor health index (HealthIndex), with a value range of , represents the complete failure state, represents the complete health state; represents the total number of factors affecting capacitor health; represents the causal contribution weight of the th factor; represents the actual measurement value of the th factor; represents the time variable; represents the nonlinear conversion function of the th factor, used to convert the original measurement value to a standardized health contribution value; represents a dynamic adjustment function considering time decay effect.

[0060] The calculation method of weight is: ; Ensure ; Where, represents the weight of the th factor; represents the causal impact index of the th factor; represents the total number of influencing factors; represents the sum of causal impact indices of all factors; The constraint condition that the sum of all weights is 1 The capacitance health index makes full use of causal structure information, and is more accurate in reflecting the actual health state of the capacitance than traditional health indexes based only on statistical correlation.

[0061] Specifically, each inner layer function is defined as follows: : factor The nonlinear conversion function of the factor converts the original monitoring parameter into a standardized health contribution value, and adopts an S-shaped function: ; wherein, represents the nonlinear conversion function value of the th factor; represents the original monitoring parameter value of the th factor; represents the adjustment parameter of the th factor; represents the critical threshold value of the th factor; represents the base parameter of the natural logarithm determined by fitting historical data.

[0062] : dynamic adjustment function of time decay effect, considering the decay of the influence of parameter historical value on the current health state, defined as: ; wherein, represents the time decay effect function value of the th factor; represents the time variable; represents the decay coefficient of the th factor; represents the time interval; represents the base of the natural logarithm.

[0063] The decay coefficient is determined by maximizing the correlation between the health index and the actually observed degradation degree: ; wherein, represents the decay coefficient of the th factor; represents the correlation function; represents the health index calculated using the decay coefficient ; represents the actually observed degradation degree of the capacitance; represents the value of the th factor corresponding to the maximum value of the correlation function, The range of values ​​is .

[0064] An ensemble learning approach is used to construct a prediction model. The input consists of multimodal sensor data and causal features, and the output consists of health status prediction results at different time scales.

[0065] This allows for model training and optimization, with historical data as input and a validated optimized prediction model as output.

[0066] like Figure 7 As shown, this illustrates the comparison between the predicted and actual values ​​of the capacitance health index over time.

[0067] The maintenance decision module generates the optimal maintenance strategy and evaluates cost-effectiveness under multiple constraints based on the health index. Construct a maintenance strategy generator, taking capacitor health index prediction and system operation plan as inputs, and outputting maintenance decision schemes.

[0068] Furthermore, the multi-source heterogeneous decision data is preprocessed before the maintenance strategy is generated: Data alignment and fusion: Aligning data at different time granularities (real-time monitoring data, historical fault records, maintenance plans) to a unified time scale through interpolation and aggregation methods; Cost standardization: Convert maintenance costs from different dimensions (equipment replacement costs, labor costs, downtime losses) into a unified economic value unit and use a time value-based discounting method; Maintenance timing coding: Convert feasible maintenance time windows into discrete decision variables and represent the combinations of selectable maintenance activities using binary coding; Reliability index transformation: The probabilistic reliability index is transformed into a linearly additive utility function through logarithmic transformation, which is convenient for application in multi-objective optimization.

[0069] Furthermore, the maintenance strategy generator, based on a multi-objective optimization framework, formalizes the maintenance problem as follows: ; in, This indicates a maintenance strategy that needs to be optimized. Represents the objective function vector for multi-objective optimization; , , They represent the 1st and 2nd respectively. , Objective functions, such as cost function, reliability function, lifetime function, etc.; This represents the total number of objective functions.

[0070] The constraints of this multi-objective optimization problem include: Budget constraint: ; where, denotes the total maintenance cost of maintenance policy s; denotes the maximum maintenance budget allowed by the system.

[0071] Reliability constraint: ; where, denotes the system reliability indicator under maintenance policy s; denotes the minimum reliability threshold required by the system.

[0072] Time window constraint: ; where, denotes the specific maintenance execution time point; denotes the set of all feasible maintenance time windows.

[0073] By introducing constraint conditions based on causal knowledge and a Pareto optimal solution screening mechanism, the maintenance policy generator can find the optimal maintenance policy that balances various objectives under complex conditions.

[0074] Specifically, each inner function is defined as follows: : Maintenance cost function, calculates the total cost of policy , including equipment replacement cost, labor cost, and downtime loss, denoted as: ; where, denotes the maintenance cost function; denotes the maintenance policy; denotes the equipment replacement cost under policy ; denotes the labor cost under policy ; denotes the downtime loss cost under policy ; : System reliability function, evaluates the overall system reliability after implementing policy , defined as: ; where, denotes the system reliability function; denotes the maintenance policy; denotes the total number of system components; denotes the component number; denotes the reliability of the th component under policy ; represents a multiplication operator.

[0075] : device remaining life function, prediction strategy Expected remaining useful life of the device after implementation, calculated based on the causal model: ; wherein, represents a device remaining life function; represents a maintenance strategy; represents an initial design life; represents a time variable; represents a prediction time range; represents a time-dependent degradation rate under the strategy ; represents an integral operator.

[0076] Generate maintenance decisions, input is the health index prediction result and the causal analysis result, output is the specific maintenance time window and measures.

[0077] Perform cost-benefit analysis, input is the maintenance strategy scheme, output is the cost-benefit evaluation result of each scheme.

[0078] Application examples of the present embodiment: Application scenario: large-scale photovoltaic power station PCS system; The present embodiment is applied in a 100 MW-level photovoltaic power station energy storage PCS system. The PCS system of the power station contains 50 groups of 20 kW power modules in parallel, each of which is configured with multiple groups of cross-capacitors. The system environmental characteristics include: temperature variation range -10℃ to 45℃, relative humidity 30% to 85%, altitude 1200 meters, and large daily load fluctuation.

[0079] Implementation process instance: Data acquisition and preprocessing instance: In the PCS system of the photovoltaic power station, 126 sensor points are deployed, including 42 electrical parameter sensors, 36 temperature sensors, 24 vibration sensors, and 24 capacitor parameter sensors. The system has been running continuously for 6 months, and about 2.1 TB of data has been collected. The multi-modal sensor deployment and sampling configuration are shown in Table 1: Table 1: Multi-modal sensor deployment and sampling configuration The multi-scale entropy feature calculation results of the key capacitor groups show that as the capacitor aging degree deepens, the signal complexity significantly decreases, which is manifested as the overall downward shift of the multi-scale entropy curve. The multi-scale entropy feature calculation results of typical cross-capacitors are shown in Table 2: Table 2: Results of multi-scale entropy feature calculation of typical bridging capacitors Causal graph construction and causal contribution analysis example: The time-series causal graph constructed by the hybrid structure learning algorithm reveals the complex causal network of capacitor degradation, identifying 36 causal edges and 12 key nodes. The algorithm determines the skeleton structure (undirected graph) in the initial stage, then determines the direction of the edges through the direction determination rule and Bayesian score, and finally obtains a stable causal structure. The key causal paths of capacitor degradation and their contribution degrees are shown in Table 3: Table 3: Key causal paths of capacitor degradation and their contribution degrees; Counterfactual reasoning framework example: Based on the constructed time-series causal graph, the structural causal model is applied to model the capacitor degradation process, clearly distinguishing between endogenous and exogenous variables. For each key factor, an intervention hypothesis is made, and the counterfactual query mechanism is used to evaluate its impact on capacitor life. The results of counterfactual intervention effect analysis are shown in Table 4: Table 4: Results of counterfactual intervention effect analysis; The counterfactual scenario simulator is realized through Pearl's three-step causal reasoning method. First, the posterior distribution of the exogenous variable is updated based on the observed data, then the hypothetical intervention is applied, and finally the intervention effect is calculated through the modified structural model, realizing the accurate evaluation of the effect of different intervention measures, breaking through the limitations of traditional correlation analysis.

[0080] Health index prediction and maintenance strategy generation example The health index model constructed based on causal contribution weight is significantly better than traditional methods. The model applies a nonlinear transformation function and a time decay adjustment function to each key factor, and calculates the health index according to the formula , where the weight wi is determined by normalizing the causal influence index (CII). Through 10-fold cross-validation, the prediction accuracy reaches 92.7%, which is 17.3 percentage points higher than traditional correlation-based methods. The cost-benefit analysis of different maintenance strategy schemes is shown in Table 5: Table 5: Cost-benefit analysis of different maintenance strategy schemes; Technical effect verification: Deep causal mechanism reveals: By the causal inference technology of the embodiment, the real causal mechanism of the capacitor degradation is successfully identified, not only the temperature fluctuation and voltage ripple are found to be the main factors affecting the capacitor life, but also the mechanism of how they affect the capacitor through different causal paths is revealed. The pseudo-correlation relationship (such as environmental humidity and capacitor life) which is difficult to distinguish by traditional correlation analysis is correctly identified. The comparison of traditional correlation analysis and causal inference method is shown in Table 6: Table 6: Comparison of traditional correlation analysis and causal inference method Accurate maintenance decision support The causal-based maintenance strategy generated by the embodiment significantly reduces the maintenance cost and improves the system reliability compared with the traditional method. In a 6-month operation cycle, the system availability is improved by 3.2 percentage points, and the maintenance cost is reduced by 32.6%. The comparison of the effect of causal-based maintenance decision and traditional method is shown in Table 7: Table 7: Comparison of the effect of causal-based maintenance decision and traditional method The above describes the embodiments of the present application, but the embodiments are not limited to the specific embodiments described above, and the specific embodiments described above are only illustrative and not limiting, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A PCS converter with a cross capacitance balancing, characterized in that, The method comprises the following steps: A data acquisition module is used to collect multi-modal sensor data of the PCS system, including electrical, temperature, vibration, and capacitance parameter data, and calculate multi-scale entropy features; A feature extraction module is used to perform time correlation analysis based on the multi-scale entropy features, and construct an initial time causal graph model; A structure optimization module is used to optimize the causal graph structure by using a hybrid structure learning algorithm combined with conditional independence testing and Bayesian scoring; An counterfactual reasoning module is used to establish a structural causal model and a scenario simulator to realize counterfactual queries and analysis; A causal assessment module is used to calculate average causal effects and comprehensive causal impact indexes to quantify the influence of each factor on the capacitance life; A health assessment module is used to construct a system health index based on causal contribution degrees, combined with nonlinear conversion and time decay effects; A maintenance decision module is used to generate optimal maintenance strategies and evaluate cost-effectiveness under multiple constraint conditions according to the health index.

2. The PCS converter of claim 1, wherein, The multi-modal sensor data collected by the data acquisition module includes electrical parameter data, temperature data, vibration data, and capacitance parameter data. The multi-scale entropy features are obtained by calculating sample entropy at different time scales.

3. The PCS converter of claim 1, wherein, The hybrid structure learning algorithm in the structure optimization module includes: Constructing a skeleton structure based on conditional independence testing; Determining the direction of the edge using direction determination rules; Using the Bayesian scoring function to evaluate the overall structure and fine-tune it.

4. The PCS converter of claim 1, wherein, The counterfactual reasoning module includes: Establishing a structural causal model; Designing a counterfactual query mechanism; Building a counterfactual scenario simulator, where the counterfactual query mechanism is based on the three-step method of causal reasoning.

5. The PCS converter of claim 1, wherein, The causal assessment module quantifies causal contribution degrees through the following indicators: average causal effects, conditional average causal effects, and causal impact indexes, which consider direct effects, indirect effects, and mediating effects.

6. The PCS converter of claim 5, wherein, The causal impact index is calculated by the weighted combination of normalized direct effects, indirect effects, and mediating effects. The weight coefficients are optimized by maximizing the objective function related to the actual observed system degradation rate.

7. The PCS converter of claim 1, wherein, The health assessment module calculates the health index by combining the causal contribution degree weight of each factor, the nonlinear conversion function, and the dynamic adjustment function of the time decay effect.

8. The PCS converter of claim 1, wherein, The maintenance decision module generates maintenance strategies based on a multi-objective optimization framework, considering maintenance cost, system reliability, and equipment remaining life, and is subject to budget constraints, reliability constraints, and time window constraints.

9. The PCS converter of claim 8, wherein, The cost-effectiveness evaluation of the maintenance decision module includes calculating the maintenance cost function, the system reliability function, and the equipment remaining life function, and finding the optimal maintenance strategy that balances various objectives through the Pareto optimal solution screening mechanism.

10. A computer-readable storage medium, characterized in that, The computer readable instructions stored therein can run a cross-capacitance balanced PCS converter as claimed in any one of claims 1-9 when read by a computer.

Images