Hybrid intelligent-based parameter optimization method for power battery recycling material purification

Abstract

This invention provides a method for optimizing purification parameters of recycled power battery materials based on hybrid intelligence, relating to the field of hybrid intelligence technology. The method includes acquiring multimodal degradation characterization data of power batteries, constructing a multi-layered heterogeneous battery material knowledge graph and establishing cross-layer causal reasoning paths; then, performing constraint propagation reasoning based on the graph's state characterization, generating recycling candidate solutions through reverse tracing and forward deduction; next, performing multi-objective optimization by combining material value retention rate, process energy consumption indicators, and environmental load indicators; and finally, triggering structural correction of the knowledge graph through performance deviation analysis to achieve self-evolutionary updates. This invention improves the optimization efficiency of purification parameters for recycled power battery materials.

Description

TECHNICAL FIELD

[0001] The present application relates to hybrid intelligent technology, and in particular to a hybrid intelligent-based power battery recycling material purification parameter optimization method. BACKGROUND

[0002] With the rapid development of new energy vehicle industry, recycling and utilization of power batteries have become an important issue of resource circulation and environmental protection. The existing power battery recycling technology mainly relies on traditional physical disassembly and chemical extraction methods, and there are many technical defects in the optimization of material purification parameters.

[0003] In the prior art, power battery recycling material purification mainly adopts an empirical parameter setting method, lacking in-depth understanding of the complex degradation mechanism of battery materials. The traditional method usually only considers single-dimensional material property analysis, and cannot effectively integrate the complex correlation between the physical degradation state of the battery, the chemical component change and the regeneration process parameters, resulting in low recycling efficiency and serious resource waste.

[0004] The existing parameter optimization methods mostly use static linear optimization models, lacking comprehensive analysis ability of multi-modal data in the material degradation process. These methods cannot accurately identify the failure mode of key components, and it is also difficult to predict the performance recovery effect under different process conditions, so that the selection of process parameters in the recycling process lacks scientific basis.

[0005] The prior art lacks effective knowledge accumulation and self-adaptive learning mechanism, and the traditional parameter optimization method cannot be dynamically adjusted according to the actual recycling effect, and cannot form a closed-loop optimization system for continuous improvement. When facing different types and different degradation degrees of power batteries, the existing method often needs to re-do a large number of experimental verification, which is low in efficiency and high in cost.

[0006] The prior art also has the problem of insufficient multi-objective optimization capability. In the actual power battery recycling process, multiple mutually restrictive objectives such as material value retention rate, process energy consumption control and environmental load minimization need to be considered, but the existing method lacks effective multi-objective collaborative optimization mechanism, and can only achieve local optimization on a single objective, making it difficult to achieve overall optimal recycling effect.

[0007] It is urgent to develop a power battery recycling material purification parameter optimization method that can comprehensively consider multi-level information of battery materials, has intelligent reasoning ability and self-adaptive learning mechanism, to solve the deficiencies of the prior art in recycling efficiency, resource utilization rate and environmental friendliness, and provide technical support for efficient recycling of power batteries. SUMMARY

[0008] The embodiment of the application provides a power battery recycling material purification parameter optimization method based on hybrid intelligence, which can solve the problems in the prior art.

[0009] In a first aspect, the embodiment of the application provides a power battery recycling material purification parameter optimization method based on hybrid intelligence, which comprises the following steps:

[0010] Obtain multi-modal degradation characterization data of a power battery to be diagnosed, construct a multi-layer heterogeneous battery material knowledge graph comprising a material composition layer, a degradation mechanism layer and a regeneration process layer, establish a cross-layer causal reasoning path between the layers, map the multi-modal degradation characterization data to the corresponding hierarchical nodes, and form a graph state representation that integrates physical degradation state and chemical component change;

[0011] Perform constraint propagation reasoning based on the graph state representation, integrate material degradation irreversibility constraints and regeneration process cost-efficiency boundary constraints on the cross-layer causal reasoning path, identify important component failure modes through reverse tracing, and predict performance recovery intervals through forward deduction to generate multi-granularity recycling candidate schemes containing confidence quantification;

[0012] According to the material value retention rate, process energy consumption index and environmental load index of each scheme in the multi-granularity recycling candidate scheme, and in combination with the material residual electrochemical activity distribution in the graph state representation, a recycling execution scheme is generated through multi-objective collaborative optimization;

[0013] According to the recycling execution scheme, perform deviation analysis on the measured data of the regenerated material performance after execution and the performance recovery interval, trigger knowledge graph structural modification when the deviation exceeds the tolerance range, and realize knowledge graph self-evolution update by adding or deleting causal relationship edges of the cross-layer causal reasoning path and adjusting constraint weights.

[0014] Construct a multi-layer heterogeneous battery material knowledge graph comprising a material composition layer, a degradation mechanism layer and a regeneration process layer, and establish a cross-layer causal reasoning path between the layers, map the multi-modal degradation characterization data to the corresponding hierarchical nodes, and form a graph state representation that integrates physical degradation state and chemical component change.

[0015] Construct a multi-layer heterogeneous battery material knowledge graph comprising a material composition layer, a degradation mechanism layer and a regeneration process layer, decouple the multi-modal degradation characterization data in the feature space, extract the invariance component reflecting the intrinsic properties of the material and the evolution component reflecting the degradation process, map the invariance component to the corresponding nodes of the material composition layer, and map the evolution component to the corresponding nodes of the degradation mechanism layer through impedance spectrum mode matching, morphology degradation mode identification and decay rate segmentation analysis;

[0016] establish a material component dependent causal relationship edge between the material composition layer and the degradation mechanism layer, a process response dependent causal relationship edge between the degradation mechanism layer and the regeneration process layer, and a cross-layer causal reasoning path composed of the material component dependent causal relationship edge and the process response dependent causal relationship edge;

[0017] performing backward propagation to trace the responsibility material component based on the material component dependent causal relationship edge in the cross-layer causal reasoning path, performing forward propagation to predict the process repair feasibility based on the process response dependent causal relationship edge, and generating a graph state representation.

[0018] performing constraint propagation reasoning based on the graph state representation, fusing a material degradation irreversibility constraint and a regeneration process cost-efficiency boundary constraint on the cross-layer causal reasoning path, identifying an important component failure mode through backward tracing and predicting a performance recovery interval through forward deduction, and generating a multi-granularity recycling candidate solution including a confidence quantization.

[0019] constructing a constraint propagation network based on the graph state representation, mapping the material degradation irreversibility constraint to a state transition probability threshold of a material composition layer node through a Monte Carlo tree search, and mapping the regeneration process cost-efficiency boundary constraint to an upper limit of resource consumption of a regeneration process layer node; inputting the state transition probability threshold into a conditional random field model to construct a constraint propagation probability distribution, and establishing a constraint propagation function on the cross-layer causal reasoning path based on the constraint propagation probability distribution;

[0020] performing bidirectional reasoning calculation in the constraint propagation network, calculating a contribution degree distribution of material components to failure modes and quantifying an importance degree through backward tracing based on the constraint propagation function using a Monte Carlo tree search method, and calculating a performance recovery degree under different process combinations and determining a performance recovery interval through forward deduction; inputting the contribution degree distribution and the confidence of the performance recovery interval output by the constraint propagation function into the conditional random field model for probability reasoning, and generating a multi-granularity recycling candidate solution based on a reasoning result.

[0021] calculating a contribution degree distribution of material components to failure modes and quantifying an importance degree through backward tracing based on the constraint propagation function using a Monte Carlo tree search method, and calculating a performance recovery degree under different process combinations and determining a performance recovery interval through forward deduction including:

[0022] construct a tree-shaped search space based on the constraint propagation function, take a material component node as a root node of the search tree, take a failure mode node as an intermediate node of the search tree, calculate a confidence probability of each search path by using a Monte Carlo tree search method through depth-first traversal, generate a contribution degree distribution of the material component to the failure mode according to the confidence probability, determine the importance of the component based on the contribution degree distribution, and mark the component with the highest contribution degree as a failure component;

[0023] map the failure component to the tree-shaped search space, calculate a corresponding relationship between a process parameter and a performance index based on the constraint propagation function, calculate an influence degree of the process parameter on performance recovery by using forward propagation, and determine a performance recovery interval according to the influence degree.

[0024] generate a recycling execution scheme by multi-objective collaborative optimization according to the material value retention rate, the process energy consumption index, and the environmental load index of each scheme in the multi-granularity recycling candidate scheme, and in combination with the material residual electrochemical activity distribution in the atlas state representation, including:

[0025] obtain the material value retention rate, the process energy consumption index, and the environmental load index of each scheme in the multi-granularity recycling candidate scheme, and obtain the material residual electrochemical activity distribution in the atlas state representation, and construct a multi-objective optimization space by taking the material residual electrochemical activity distribution as a constraint condition;

[0026] construct a solution space tree based on the multi-objective optimization space by using a beam search method, take the material value retention rate as a value evaluation function input item, take the process energy consumption index and the environmental load index as cost evaluation function input items, generate a target optimization function according to a weighted combination of the value evaluation function input item and the cost evaluation function input item, perform branch and bound search on the solution space tree based on the target optimization function, and calculate a comprehensive score of each candidate scheme;

[0027] input the candidate scheme with the highest comprehensive score into a greedy search algorithm for local optimization, construct a Pareto front in the multi-objective optimization space by iteratively updating a weight coefficient of the target optimization function, determine an optimal weight combination based on the Pareto front, and generate a recycling execution scheme according to the optimal weight combination.

[0028] input the candidate scheme with the highest comprehensive score into a greedy search algorithm for local optimization, construct a Pareto front in the multi-objective optimization space by iteratively updating a weight coefficient of the target optimization function, including:

[0029] The candidate solution with the highest comprehensive score is set as the initial optimization node. An optimization feature vector is constructed based on the initial optimization node. A greedy search algorithm is used to determine the local search space, and iterative calculations are performed in the local search space. The convergence direction of each iteration is used as the basis for updating the weights of the target optimization function. The weight coefficients are dynamically adjusted based on the weight update basis, and the product of the weight coefficients and the optimization feature vector is used as the optimization direction of the target optimization function.

[0030] Sampling is performed in the multi-objective optimization space according to the optimization direction, and the optimal solution set is constructed based on the sampling results. The distribution curve of the non-dominated solutions in the optimal solution set is taken as the Pareto front.

[0031] Based on the recycling implementation plan, a deviation analysis is performed between the measured performance data of the recycled materials after implementation and the performance recovery range. When the deviation exceeds the tolerance range, a knowledge graph structural correction is triggered, including:

[0032] A performance deviation assessment matrix is ​​constructed, and the difference between the measured performance data of recycled materials and the performance recovery interval is used as the assessment parameter. The assessment parameter is input into the performance deviation assessment matrix to calculate the deviation feature distribution. The Monte Carlo method is used to perform probability sampling on the deviation feature distribution. Based on the result of the probability sampling, the deviation confidence interval is determined, and the deviation confidence interval is compared with the preset tolerance range.

[0033] When the deviation confidence interval exceeds the preset tolerance range, a knowledge graph correction space is constructed based on the beam search method. The eigenvalues ​​of the performance deviation evaluation matrix are used as search weights to calculate the structural adjustment direction in the knowledge graph correction space. Based on the structural adjustment direction, iterative optimization is performed on the knowledge graph correction space to determine the optimal correction scheme for the graph structure and to perform knowledge graph structural correction.

[0034] A second aspect of the present invention provides an electronic device, comprising:

[0035] processor;

[0036] Memory used to store processor-executable instructions;

[0037] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0038] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0039] The beneficial effects of this application are as follows:

[0040] This invention constructs a multi-layered heterogeneous battery material knowledge graph comprising a material composition layer, a degradation mechanism layer, and a regeneration process layer, and establishes cross-layer causal reasoning paths between each layer. This enables the mapping of multimodal degradation characterization data to corresponding level nodes, forming a graph-based state characterization that integrates physical degradation state and chemical composition changes. This achieves a deeper understanding of the degradation mechanism of power batteries, solves the technical problem of traditional methods being unable to correlate battery degradation characterization with microscopic material changes, and improves the accuracy and comprehensiveness of degradation state assessment.

[0041] This invention integrates the constraints of irreversible material degradation and the cost-effectiveness boundary constraints of regeneration processes along a cross-layer causal reasoning path. It adopts a reverse tracing approach to identify failure modes of important components and a forward deduction approach to predict performance recovery intervals, generating multi-granularity recycling candidate schemes with confidence quantification. This approach can simultaneously consider material value retention rate, process energy consumption indicators, and environmental impact indicators. Through multi-objective collaborative optimization, it generates the optimal recycling execution scheme, effectively improving the scientific and economic efficiency of optimizing purification parameters for power battery recycled materials and reducing the energy consumption and environmental impact of the regeneration process.

[0042] This invention analyzes the discrepancy between the measured performance data of recycled materials after execution and the predicted performance recovery range. When the discrepancy exceeds the tolerance range, it automatically triggers the structural correction of the knowledge graph. By adding or deleting causal relationship edges in cross-layer causal reasoning paths and adjusting constraint weights, the knowledge graph is self-evolved and updated, enabling the system to have continuous learning and self-improvement capabilities. As recycling experience accumulates, the prediction accuracy and optimization effect can be continuously improved, enhancing the adaptability and robustness of the method. Attached Figure Description

[0043] Figure 1 This is a schematic flowchart of the method for optimizing purification parameters of recycled power battery materials based on hybrid intelligence, according to an embodiment of the present invention.

[0044] Figure 2 This is a flowchart illustrating the research on the failure mechanism and performance recovery of power battery material components in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0047] Figure 1 This is a schematic flowchart of the method for optimizing purification parameters of recycled power battery materials based on hybrid intelligence, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0048] Multimodal degradation characterization data of the power battery to be diagnosed is acquired, a multi-layer heterogeneous battery material knowledge graph including material composition layer, degradation mechanism layer and regeneration process layer is constructed, and cross-layer causal reasoning paths are established between each layer. The multimodal degradation characterization data is mapped to the corresponding level nodes to form a graph state characterization that integrates physical degradation state and chemical composition changes.

[0049] Based on the state characterization of the graph, constraint propagation reasoning is performed. The constraints of irreversible material degradation and cost-effectiveness boundary constraints of recycling process are integrated on the cross-layer causal reasoning path. The failure modes of important components are identified by reverse tracing and the performance recovery range is predicted by forward deduction, generating multi-granularity recycling candidate schemes with confidence quantification.

[0050] Based on the material value retention rate, process energy consumption index, and environmental load index of each scheme in the multi-granularity recycling candidate schemes, and combined with the distribution of residual electrochemical activity of materials in the spectrum state characterization, a recycling implementation scheme is generated through multi-objective collaborative optimization.

[0051] According to the recycling implementation plan, the measured performance data of the recycled materials after the implementation are compared with the performance recovery range. When the deviation exceeds the tolerance range, the knowledge graph structure is modified. The knowledge graph is updated by adding or deleting causal relationship edges of the cross-layer causal reasoning path and adjusting the constraint weights.

[0052] In one optional implementation, a multi-layered heterogeneous battery material knowledge graph is constructed, comprising a material composition layer, a degradation mechanism layer, and a regeneration process layer. Cross-layer causal reasoning paths are established between each layer, mapping the multi-modal degradation characterization data to corresponding level nodes to form a graph state characterization that integrates physical degradation state and chemical composition changes.

[0053] A multi-layer heterogeneous battery material knowledge graph containing a material composition layer, a degradation mechanism layer, and a regeneration process layer is constructed. The feature space of the multimodal degradation characterization data is decoupled, and the invariant components reflecting the intrinsic properties of the material and the evolutionary components reflecting the degradation process are extracted. The invariant components are mapped to the corresponding nodes of the material composition layer, and the evolutionary components are mapped to the corresponding nodes of the degradation mechanism layer through impedance spectrum mode matching, morphological degradation mode identification, and attenuation rate segmentation analysis.

[0054] A material component-dependent causal relationship edge is established between the material composition layer and the degradation mechanism layer, and a process-responsive causal relationship edge is established between the degradation mechanism layer and the regeneration process layer. The material component-dependent causal relationship edge and the process-responsive causal relationship edge constitute a cross-layer causal reasoning path.

[0055] Based on the material component-dependent causal relationship in the cross-layer causal reasoning path, reverse propagation is performed to trace the responsible material components, and based on the process response-type causal relationship, forward propagation is performed to predict the feasibility of process repair, generating a graph state representation.

[0056] The construction process of the multi-layer heterogeneous knowledge graph of power batteries is based on the systematic processing of input multimodal degradation characterization data. The multimodal degradation characterization data includes five main information sources: electrochemical impedance spectroscopy data, scanning electron microscopy image data, X-ray diffraction pattern data, capacity decay curve data, and thermal analysis data. These data first undergo standardization preprocessing, transforming data with different dimensions into a unified numerical range of 0 to 1, eliminating the influence of dimensional differences on subsequent analysis.

[0057] Feature space decoupling is achieved using a multidimensional data dimensionality reduction technique based on principal component analysis. All input data are constructed into a high-dimensional feature matrix, with the matrix dimension being the number of samples multiplied by the number of features. The number of samples is determined based on the frequency of battery testing, and the number of features consists of the total number of effective measurement parameters for each data type. The covariance matrix of the feature matrix is ​​calculated, and its eigenvalues ​​and eigenvectors are solved. The principal components are then sorted in descending order of eigenvalues ​​to obtain the principal component sequence. The cumulative variance contribution rate is calculated to determine the number of principal components to retain; principal component selection stops when the cumulative contribution rate reaches 95%.

[0058] Invariant component extraction is achieved by retaining the first 60% of the principal components. The eigenvectors corresponding to these principal components reflect the intrinsic properties of the material. Invariant components include stable material characteristics such as lattice constant, elemental molar ratio, crystal symmetry, and electronic structure parameters. Evolutionary components consist of the remaining principal components and contain dynamic evolution information such as capacity decay slope, impedance growth trend, active material loss rate, and changes in structural order. The accuracy of component separation is verified through reconstruction error, which is controlled within 3% of the variance of the original data.

[0059] The material composition layer nodes are constructed using a hierarchical organizational structure based on the periodic law of chemistry. The first layer consists of element category nodes, including basic classifications such as alkali metals, transition metals, and nonmetals. The second layer consists of specific element nodes, such as lithium, iron, phosphorus, and oxygen, which are key elements constituting battery materials. The third layer consists of compound nodes, including composite material components such as lithium iron phosphate, graphite, and electrolytes. Each node has attribute fields containing basic attributes such as element symbol, atomic number, electronegativity, ionic radius, and standard electrode potential, as well as compositional proportion attributes such as mole fraction, mass fraction, and volume fraction.

[0060] The invariant component mapping process is achieved through similarity matching between feature vectors and predefined material templates. The material template library contains standard feature vectors of common battery materials, covering various components such as cathode materials, anode materials, electrolytes, conductive agents, and binders. Similarity calculation uses a cosine similarity algorithm to calculate the cosine value of the angle between the input invariant component vector and the template vector. When the cosine similarity exceeds a threshold of 0.85, a successful match is confirmed, and the corresponding invariant component is assigned to the matched material composition layer node. Node attribute updates use an exponentially weighted moving average method, with a weight of 0.3 for new data and 0.7 for historical data.

[0061] The degradation mechanism layer nodes are constructed based on a classification system of battery degradation physicochemical mechanisms. The main degradation mechanisms include five categories: active material loss, conductive network degradation, electrolyte decomposition, increased side reactions, and ion transport blockage. Each degradation mechanism is further subdivided into specific sub-mechanism nodes; for example, active material loss includes sub-mechanisms such as particle breakage, phase transition, and dissolution. Node attributes include quantitative parameters such as degradation degree, probability of occurrence, scope of influence, and time characteristics.

[0062] An impedance spectrum pattern matching algorithm processes electrochemical impedance spectroscopy data, extracting the real and imaginary parts of impedance at characteristic frequency points. The frequency scan range is set from 0.01 Hz to 100,000 Hz, using logarithmic interval sampling, with 10 sampling points per ten-fold range at each frequency. The impedance data undergoes smoothing filtering to eliminate measurement noise, with a filtering window width of three adjacent frequency points. A standard degradation mode library contains impedance spectrum features corresponding to typical degradation mechanisms. Matching is performed by calculating the Euclidean distance between the input spectrum and the standard mode. The distance threshold is set to 0.15. Upon successful matching, the evolutionary component is assigned to the corresponding degradation mechanism node.

[0063] Morphology degradation pattern recognition utilizes digital image processing techniques to analyze scanning electron microscope (SEM) images. Image preprocessing includes grayscale conversion, histogram equalization, and Gaussian filtering for noise reduction. Feature extraction employs a local binary pattern operator, calculating binary pattern codes within an 8-pixel multi-8-pixel local window. The window slides across the image with a 4-pixel stride to extract the texture feature vector of the entire image. Degradation pattern classification uses a support vector machine (SVM) classifier. The training data includes five typical morphology degradation patterns: particle fragmentation, surface passivation, structural collapse, increased porosity, and crack propagation, with 500 labeled images for each pattern. The classifier achieved an accuracy of 92% and a recall of 89%.

[0064] Segmented degradation rate analysis divides the battery's life-cycle capacity degradation curve into different stages. By calculating the first and second derivatives of the capacity degradation rate, inflection points in the degradation mode are identified. A segmentation point is confirmed when the absolute value of the second derivative exceeds a threshold of 0.05. A typical segment includes three stages: the initial break-in period, the mid-term linear degradation period, and the late-term accelerated degradation period. Within each segment, a least-squares method is used for linear fitting to obtain the corresponding degradation rate parameters. The goodness of fit is evaluated using the correlation coefficient, which is required to be greater than 0.95.

[0065] The material composition-dependent causal relationship edge is constructed based on materials science theory and experimental data verification. The causal relationship edge connects relevant nodes between the material composition layer and the degradation mechanism layer. Edge attributes include parameters such as causal strength, confidence level, and time delay. Causal strength is calculated through statistical analysis based on the correlation data between material composition and degradation modes from a large number of battery samples. The confidence level is determined using the significance test results of the Pearson correlation coefficient, with a significance level set at 0.05. The time delay parameter reflects the lag effect of the causal action; the time offset corresponding to the maximum correlation coefficient is determined through cross-correlation analysis.

[0066] A process-responsive causal relationship edge connects the degradation mechanism layer and the regeneration process layer, established based on the physicochemical principles of the regeneration process. The regeneration process layer includes major regeneration methods such as heat treatment, chemical leaching, electrochemical repair, and mechanical treatment. Each process includes key parameter nodes such as temperature, pressure, time, concentration, and current density. The weights of the causal relationship edges are determined through sensitivity analysis. A single-factor perturbation method is used, perturbing the process parameters by ±10% from their standard values ​​to observe the impact on the regeneration effect.

[0067] The backpropagation tracing algorithm starts from anomalous nodes in the degradation mechanism layer. Anomalous nodes are identified by their deviation from the normal baseline; nodes with a deviation exceeding two standard deviations are marked as anomalous. The tracing process employs a breadth-first search algorithm, propagating along material composition-dependent causal relationships towards the material composition layer. Path weights are calculated considering the strength and confidence of the causal relationship edges, using a weighted geometric average method. When multiple paths exist, the path with the highest weight is selected as the primary tracing path. The tracing result is represented by a degree of responsibility, ranging from 0 to 1, indicating the degree of responsibility of the corresponding material composition for the degradation mechanism.

[0068] The forward propagation prediction algorithm starts from the current state of the material composition layer and predicts the feasibility of regeneration processes. The prediction process employs a probabilistic inference method based on Bayesian networks, considering the uncertainty and nonlinearity of causal relationships. Network parameters are trained using historical regeneration process data; the training set contains 3000 valid regeneration experiment records, and the validation set contains 500 independent test data sets. The prediction results include information such as the probability of process feasibility, expected repair effect, and confidence interval.

[0069] The graph state representation generates integrated calculation results combining reverse tracing and forward prediction. Node state vectors are represented by 64-dimensional real-valued vectors, containing information such as material property encoding, degradation quantification, and process response prediction. Vector encoding employs a combination of one-hot encoding and numerical normalization to ensure effective fusion of different types of information. The edge weight matrix is ​​stored in a sparse matrix format, with non-zero elements accounting for approximately 15%, effectively saving storage space and computational resources.

[0070] In one optional implementation, constraint propagation reasoning is performed based on the spectrum state characterization. The constraint of material degradation irreversibility and the cost-effectiveness boundary constraint of the recycling process are integrated along the cross-layer causal reasoning path. By identifying failure modes of important components through reverse tracing and predicting performance recovery intervals through forward deduction, multi-granularity recycling candidate schemes containing confidence quantification are generated, including:

[0071] Based on the graph state representation, a constraint propagation network is constructed. The irreversible material degradation constraint is mapped to the state transition probability threshold of the material composition layer node through Monte Carlo tree search. The cost-efficiency boundary constraint of the regeneration process is mapped to the resource consumption upper limit of the regeneration process layer node. The state transition probability threshold is input into the conditional random field model to construct the constraint propagation probability distribution. Based on the constraint propagation probability distribution, a constraint propagation function is established on the cross-layer causal reasoning path.

[0072] Bidirectional inference calculations are performed in the constraint propagation network. Based on the constraint propagation function, the Monte Carlo tree search method is used to calculate the contribution distribution of material components to failure modes through reverse sourcing and quantify their importance. The performance recovery degree under different process combinations is calculated through forward deduction and the performance recovery interval is determined. The contribution distribution output by the constraint propagation function and the confidence of the performance recovery interval are input into the conditional random field model for probabilistic inference. Based on the inference results, multi-granularity recycling candidate schemes are generated.

[0073] The constraint propagation network is constructed based on the initialization of the input graph state representation, which comprises two core data structures: a node state vector matrix and an edge weight matrix. The node state vector matrix has a dimension equal to the number of nodes multiplied by 64, and each node's state vector contains encoded information such as material properties, degradation degree, and process response. The edge weight matrix uses a sparse storage format and includes attribute parameters such as causal relationship strength, confidence level, and time delay. The constraint propagation network then constructs a directed graph structure based on this, where nodes retain their original attributes, and edges are augmented with constraint propagation function attributes.

[0074] The irreversible material degradation constraint mapping employs a Monte Carlo tree search algorithm to calculate the state transition probability threshold. The Monte Carlo tree search module comprises four core steps: selection, expansion, simulation, and propagation. The selection step uses the UCB1 strategy, with a tradeoff coefficient of 1.414 balancing exploration and utilization. The expansion step randomly samples the sub-state space of the current node, with a sampling count of 100. The simulation step executes a random strategy to reach the termination state, with a simulation depth limited to 20 layers. The propagation step updates the node statistics by propagating the evaluation value of the termination state upwards along the path.

[0075] The state transition probability threshold is determined based on the physicochemical laws of material degradation. For irreversible degradation processes such as crystal structure destruction and dissolution of active substances, the state transition probability threshold is set to 0.1, indicating that transitions with a probability of less than 10% are constrained and prevented. For reversible degradation processes such as surface passivation and pore blockage, the threshold is set to 0.8, allowing for higher state transition probabilities. The Monte Carlo tree search calculates the state transition probability distribution of each material composition layer node through 10,000 iterations, and the quantile of 0.05 is taken as the irreversibility constraint threshold.

[0076] The cost-efficiency boundary constraints of the recycling process are mapped to the resource consumption caps for each node in the recycling process layer. Resource consumption includes quantitative indicators across four dimensions: energy consumption, time, material cost, and environmental impact. The energy consumption cap is determined based on a baseline value for energy consumption per unit mass of battery material: 5 kWh per kilogram for thermal treatment and 3 kWh per kilogram for chemical leaching. The time cap is set based on process efficiency requirements, with continuous processing time not exceeding 8 hours and total intermittent processing time not exceeding 72 hours. The material cost cap is set at 30% of the value of the recycled materials, and the environmental impact cap is expressed in carbon emission equivalents, set at 2 kg of CO2 equivalent per kilogram of material.

[0077] A Conditional Random Field (CRF) model is constructed for calculating the probability distribution of constraint propagation. The CRF model employs a linear chain structure, where the observed sequence represents the node states along the cross-layer causal inference path, and the hidden sequence represents the constraint propagation states. Feature functions include two types: node feature functions and edge feature functions. Node feature functions are constructed based on linear combinations of node state vectors, with weight parameters obtained through maximum likelihood estimation. Edge feature functions are constructed based on the interaction terms of adjacent node states, considering the local dependencies in constraint propagation. Model parameters are optimized using a quasi-Newton method, with a convergence accuracy set to 0.001.

[0078] The constraint propagation probability distribution is calculated using a forward-backward algorithm. Forward variable calculation starts from the starting node of the path and recursively calculates the cumulative probability of reaching each node. Backward variable calculation starts from the ending node of the path and recursively calculates the cumulative probability of leaving each node. The constraint propagation probability is the normalized product of the forward and backward variables. Logarithmic space operations are used during the calculation to avoid numerical underflow, and the precision is maintained within the range of double-precision floating-point numbers.

[0079] The constraint propagation function is established on the cross-layer causal reasoning path based on the constraint propagation probability distribution. The function is a piecewise linear function, with the piecewise points determined by the quantiles of the probability distribution. Probability values ​​from 0 to 0.2 correspond to a constraint strength of 1.0, indicating a strong constraint. Probability values ​​from 0.2 to 0.8 correspond to a linearly decreasing constraint strength from 1.0 to 0.2. Probability values ​​from 0.8 to 1.0 correspond to a constraint strength of 0.2, indicating a weak constraint. The input to the constraint propagation function is the node state vector, and the output is the constrained state vector and the propagation probability.

[0080] Bidirectional inference computation simultaneously performs backward sourcing and forward deduction processes within the constraint propagation network. Backward sourcing begins at the failure mode node in the degradation mechanism layer and propagates along material composition-dependent causal relationships to the material composition layer. Forward deduction begins at the current state of the material composition layer and propagates along process response-dependent causal relationships to the regeneration process layer. The two inference computations employ a parallel processing architecture to avoid mutual interference.

[0081] Reverse tracing employs the Monte Carlo tree search method to calculate the contribution distribution of material components to failure modes. The search process begins by constructing a search tree from the failure mode nodes, with each node representing a contribution path. An ε-greedy algorithm is used as the search strategy, with ε set to 0.1 to ensure diversity in the exploration. The search depth is set to the maximum length of the cross-layer causal reasoning path plus 2. Each search iteration includes four steps: path selection, node expansion, random simulation, and result feedback. Contribution calculation is based on the cumulative product of the weights of causal relationship edges along the path, considering a path length decay factor of 0.9.

[0082] Importance is quantified by calculating the statistical characteristics of the contribution distribution. The importance index includes four statistical measures: mean contribution, variance, skewness, and kurtosis. The mean contribution reflects the average influence of the material component, the variance reflects the stability of the influence, the skewness reflects the symmetry of the distribution, and the kurtosis reflects the frequency of extreme values. The comprehensive importance score is calculated using a weighted summation method, with weights set at 0.4, 0.3, 0.2, and 0.1. The score results are normalized to the interval between 0 and 1, and material components with scores greater than 0.7 are marked as important components.

[0083] Forward deductive calculations employ dynamic programming algorithms to predict the performance recovery rate under different process combinations. The process combination space is constructed based on the process nodes of the regeneration process layer, and each process includes discretized options for parameter settings. The heat treatment process offers four temperature options: 300, 400, 500, and 600 degrees Celsius, and four time options: 1, 2, 4, and 8 hours. The chemical leaching process offers four concentration options: 0.1, 0.5, 1.0, and 2.0 mol / L. The dynamic programming state is defined as the material state after process execution, and the state transition function is calculated based on the weights of the causal edges in the process response.

[0084] The performance recovery rate is calculated based on the change in the material state vector. The recovery rate indicators include three dimensions: capacity recovery rate, impedance recovery rate, and cycle life recovery rate. Capacity recovery rate is defined as the ratio of the treated capacity to the initial capacity; impedance recovery rate is defined as the ratio of the initial impedance to the treated impedance; and cycle life recovery rate is estimated based on extrapolated results from accelerated aging tests. The overall recovery rate is calculated using a weighted geometric mean with weights of 0.5, 0.3, and 0.2.

[0085] The performance recovery interval was determined based on uncertainty analysis, which identified three sources of uncertainty: measurement error, model error, and parameter uncertainty. Measurement error was based on sensor accuracy specifications: ±1% for capacity testing and ±2% for impedance testing. Model error was evaluated through cross-validation, with the average relative error controlled within 5%. Parameter uncertainty was quantified using the Monte Carlo method, with parameter perturbation range set at ±10% of the standard value. Confidence interval calculation employed a guided approach, with 1000 resampling iterations and a confidence level of 95%.

[0086] The conditional random field model uses probabilistic inference to process the contribution distribution and confidence levels of the performance recovery interval from the constraint propagation function output. Input data includes the probability density function parameters of the contribution distribution and the confidence levels of the performance recovery interval. Probabilistic inference employs variational inference methods for approximate solution, with the iterative convergence criterion being that the parameter variation amplitude is less than 0.001. The inference results include the probability distribution and confidence assessment of multi-granularity recycling candidate solutions.

[0087] The generation of multi-granularity recycling candidate solutions involves screening and ranking solutions based on probabilistic reasoning results. The granularity levels include three levels: material-level, component-level, and battery-level. Material-level solutions focus on recycling individual material components, component-level solutions handle the overall processing of functional components such as electrodes, separators, and electrolytes, and battery-level solutions comprehensively process the entire battery. Solution selection is based on both technical and economic feasibility criteria. Technical feasibility requires a performance recovery rate greater than 80%, and economic feasibility requires processing costs to be less than 50% of the material value.

[0088] In one optional implementation, based on the constraint propagation function, a Monte Carlo tree search method is used to calculate the contribution distribution of material components to failure modes through reverse sourcing and quantify their importance. The performance recovery degree under different process combinations is calculated through forward deduction, and the performance recovery range is determined, including:

[0089] A tree-shaped search space is constructed based on the constraint propagation function. Material component nodes are used as the root nodes of the search tree, and failure mode nodes are used as the intermediate nodes of the search tree. The Monte Carlo tree search method is used to calculate the confidence probability of each search path through depth-first traversal. The contribution distribution of material components to failure modes is generated based on the confidence probability. The importance of components is determined based on the contribution distribution, and the component with the highest contribution is marked as the failure component.

[0090] The failed components are mapped to the tree search space. The correspondence between process parameters and performance indicators is calculated based on the constraint propagation function. Forward propagation is used to calculate the degree of influence of the process parameters on performance recovery. The performance recovery range is determined based on the degree of influence.

[0091] like Figure 2 As shown, the method includes:

[0092] When constructing a tree-structured search space, it is necessary to establish the hierarchical relationships between nodes and the constraint propagation path. Taking aluminum alloy die castings as an example, the composition ratio of the aluminum-silicon-magnesium ternary alloy system is set as the root node layer. This includes nodes with silicon content set to a concentration range of 7.5% to 9.5%, nodes with magnesium content set to a concentration range of 0.3% to 0.6%, and nodes with iron impurity content set to a control range of less than 0.8%. Above the root node, an intermediate node layer of failure modes is established, containing three types of failure characteristics: porosity defect nodes, hot crack nodes, and microstructure segregation nodes. Each material component node establishes a directed connection edge with the failure mode node through a constraint propagation function. The weight parameters of the connection edges are initialized using statistical frequencies from the historical failure database. The connection weight between the porosity defect node and the silicon content node is set to 0.62, the connection weight with the magnesium content node is set to 0.28, and the connection weight with the iron impurity content node is set to 0.73. The connection weights between the hot crack node and each component node are set to 0.81, 0.54, and 0.39, respectively.

[0093] When performing a Monte Carlo tree search, a depth-first traversal strategy is used to expand the search path starting from the root node. For the silicon content node, the current detection value is set to 9.2%, exceeding the standard specification limit of 8.5%. Starting from this node, the search proceeds upwards along the connecting edges to the porosity defect node. The confidence probability of this path needs to consider both the weight parameter and the current deviation level. The silicon content deviation is 0.7%, and the deviation rate relative to the standard limit is 8.2%. Multiplying this deviation rate by the connecting weight 0.62 yields a path confidence of 0.051. Continuing the traversal from the silicon content node to the hot crack node, the path confidence is calculated using the weight parameter 0.81 and the same deviation rate, resulting in a path confidence of 0.066. The same operation is performed on the magnesium content node; the current detection value is 0.52%, within the specification range, and the deviation rate is set to zero, resulting in a path confidence of zero for all nodes. The iron impurity content was 0.92%, exceeding the control threshold of 0.8%, with a deviation rate of 15%. The path confidence of the porosity defect node was calculated to be 0.110, and the path confidence of the hot crack node was calculated to be 0.059.

[0094] After traversing all paths from root nodes to intermediate nodes, the cumulative confidence probability of each failure mode node is calculated. The cumulative confidence probability of the porosity defect node is 0.161, the cumulative confidence probability of the hot crack node is 0.125, and the cumulative confidence probability of the microstructure segregation node is 0.043. Based on the cumulative confidence probability, a contribution distribution vector is generated, and the confidence probabilities of each failure mode are normalized. The normalized contribution of porosity defects is 48.9%, the normalized contribution of hot cracks is 38.0%, and the normalized contribution of microstructure segregation is 13.1%. Tracing back to the material composition layer, the contribution of each component to the dominant failure mode is calculated. For porosity defects, the failure mode with the highest contribution, the path confidence of iron impurity content is 0.110, accounting for 68.3% of the total confidence of porosity defects, and the path confidence of silicon content is 0.051, accounting for 31.7%. Iron impurity content is marked as the first failure component, and silicon content is marked as the second failure component.

[0095] After mapping the failed components back to the tree-shaped search space, a forward propagation path from the failed components to the process parameter layer needs to be constructed. To address the issue of excessive iron impurity content, three process control points are established at the process parameter layer: smelting temperature, refining time, and iron removal agent dosage. The constraint propagation function defines the correspondence between iron impurity content and each process parameter: every 10°C increase in smelting temperature reduces iron solubility by 0.03%, every 5-minute extension of refining time reduces residual iron by 0.02%, and every 0.1% increase in iron removal agent dosage reduces iron content by 0.05%. Currently, the iron impurity content needs to be reduced from 0.92% to 0.75% to meet the control requirements, requiring a reduction of 0.17%.

[0096] When calculating the impact of process parameters on performance recovery using forward propagation, the current melting temperature is set at 730 degrees Celsius; adjusting it to 760 degrees Celsius achieves a 0.09% reduction in iron content. The current refining time is 15 minutes; extending it to 30 minutes achieves a 0.06% reduction. The current iron removal agent dosage is 0.3%; increasing it to 0.6% achieves a 0.15% reduction. The degree of influence of the three process parameters, ranked by the magnitude of reduction, is as follows: iron removal agent dosage contributes 88.2%, melting temperature contributes 52.9%, and refining time contributes 35.3%. Due to the interaction between the process parameters, the actual combined effect needs to be corrected by constraining the coupling term of the propagation function. Increasing the melting temperature reduces the efficiency of the iron removal agent; the coupling coefficient is set to -0.15. Extending the refining time enhances the diffusion effect of the iron removal agent; the coupling coefficient is set to +0.08.

[0097] When calculating the performance recovery of different process combinations, Scheme 1, which uses a single increase in the amount of iron remover to 0.6%, was expected to reduce the iron content by 0.15%. After correction, the actual reduction was 0.14%, resulting in a final iron content of 0.78%, which did not meet the control requirements. Scheme 2, combining an iron remover addition of 0.5% and a melting temperature of 750°C, theoretically reduced the iron content by 0.19%. After correction considering the negative coupling effect, the actual reduction was 0.16%, resulting in a final iron content of 0.76%, close to the control threshold. Scheme 3, combining an iron remover addition of 0.5%, a melting temperature of 740°C, and a refining time of 25 minutes, theoretically reduced the iron content by 0.23%. After correction considering the overall coupling effect, the actual reduction was 0.20%, resulting in a final iron content of 0.72%, which met the control requirements. Based on the calculation results of the three schemes, the performance recovery range was determined to be a reduction in iron content of 0.14% to 0.20%, corresponding to a final iron content range of 0.72% to 0.78%. The lower boundary of this range corresponds to the most conservative single-parameter adjustment strategy, while the upper boundary corresponds to the multi-parameter collaborative optimization strategy. In actual production, a suitable process combination scheme is selected within this range based on cost constraints and time constraints.

[0098] In one optional implementation, based on the material value retention rate, process energy consumption index, and environmental load index of each scheme in the multi-granularity recycling candidate schemes, and combined with the distribution of residual electrochemical activity of materials in the spectral state characterization, a recycling implementation scheme is generated through multi-objective collaborative optimization, including:

[0099] The material value retention rate, process energy consumption index and environmental load index of each scheme in the multi-granularity recycling candidate scheme are obtained. The distribution of residual electrochemical activity of materials in the spectral state characterization is obtained. The distribution of residual electrochemical activity of materials is used as a constraint to construct a multi-objective optimization space.

[0100] Based on the multi-objective optimization space, a solution space tree is constructed using a bundle search method. The material value retention rate is used as the input term of the value evaluation function, and the process energy consumption index and the environmental load index are used as the input terms of the cost evaluation function. A target optimization function is generated based on the weighted combination of the value evaluation function input term and the cost evaluation function input term. Based on the target optimization function, a branch and bound search is performed on the solution space tree to calculate the comprehensive score of each candidate solution.

[0101] The candidate solution with the highest comprehensive score is input into a greedy search algorithm for local optimization. The weight coefficients of the objective optimization function are updated iteratively to construct a Pareto front in the multi-objective optimization space. The optimal weight combination is determined based on the Pareto front, and a cyclical execution solution is generated according to the optimal weight combination.

[0102] The multi-granularity recycling candidate scheme parameter acquisition module is responsible for collecting key performance indicator data for each scheme. The material value retention rate is calculated by comparing the composition of raw materials with that of recycled products, with the value ranging from 0 to 1 as a floating-point number, and the precision maintained to three decimal places. Process energy consumption indicators are quantified in kilowatt-hours per kilogram, including three sub-items: thermal treatment energy consumption, mechanical treatment energy consumption, and chemical reaction energy consumption. Environmental impact indicators are quantified using a life cycle assessment method, expressed in carbon dioxide equivalents, covering the environmental impact of the entire process from raw material acquisition, process treatment, to waste discharge. The data interface is defined in JSON format, including fields such as scheme identifier, indicator name, value, unit, and timestamp.

[0103] In the spectral characterization, the distribution of residual electrochemical activity of materials is obtained through state vector analysis. The electrochemical activity distribution data structure is a 64-dimensional vector, with each dimension corresponding to a specific activity characteristic parameter. These activity characteristic parameters include key indicators such as ion diffusion coefficient, electronic conductivity, interfacial reactivity, and structural stability. The data acquisition interface extracts activity-related vector components through slicing operations on the state vector matrix, with the index range from dimension 40 to 64. The activity distribution data is normalized to ensure that the values ​​are within the range of 0 to 1; the normalization method employs a maximum-minimum scaling strategy.

[0104] The multi-objective optimization space is constructed based on constraints set according to the distribution of residual electrochemical activity of materials. These constraints are in the form of inequalities, requiring each component of the activity distribution vector to be greater than a set activity threshold. The activity threshold is dynamically set according to the material type: 0.6 for lithium-ion battery cathode materials, 0.7 for anode materials, and 0.5 for electrolyte materials. The optimization space dimension equals the number of candidate solutions multiplied by the number of evaluation indicators, forming a high-dimensional decision variable space. The space boundaries are determined by the physical constraints of each indicator: the material value retention rate boundary is 0 to 1, the process energy consumption boundary is 0 to 50 kWh / kg, and the environmental load boundary is 0 to 10 kg CO2 equivalent / kg.

[0105] The bundle search method constructs the solution space tree using a breadth-first expansion strategy. The root node of the solution space tree represents the initial state, and each child node corresponds to a candidate solution selection decision. The bundle width parameter is set to the square root of the total number of candidate solutions, rounded up, to ensure a balance between search sufficiency and computational efficiency. The tree depth is equal to the number of decision levels, typically set to 3 levels, corresponding to material-level, component-level, and battery-level recycling decisions, respectively. During node expansion, attributes such as node evaluation value, path information, and constraint satisfaction status are maintained. The node elimination strategy is based on evaluation value sorting, retaining the bundle width nodes with the highest evaluation values ​​for further expansion.

[0106] The value assessment function is constructed based on a linear combination of the material value retention rate input. The function form is a weighted sum of the value retention rates of each option. The weight coefficients are initially set to a uniform distribution, meaning each option has a weight of 1 divided by the total number of options. The output range of the value assessment function is 0 to 1; a larger value indicates a better overall value retention effect. The function calculation process includes steps such as input validation, weight normalization, weighted summation, and result validation. Input validation ensures that the value retention rate is within a reasonable range, and weight normalization ensures that the sum of the weight coefficients is 1.

[0107] The cost assessment function is constructed based on a composite calculation of process energy consumption and environmental load indicators. The function uses a weighted geometric average, with the energy consumption indicator weighted at 0.6 and the environmental load indicator weighted at 0.4. The indicator values ​​are standardized to a dimensionless form, and the standardized benchmark values ​​are determined based on industry averages. The standardized benchmark value for energy consumption is set at 10 kWh per kilogram, and the benchmark value for environmental load is set at 3 kg CO2 equivalent per kilogram. The cost assessment function outputs a standardized cost value from 0 to 1, with smaller values ​​indicating lower costs.

[0108] The objective optimization function is generated by a weighted combination of the value evaluation function and the cost evaluation function. The combination employs a dual-objective balance strategy of maximizing value and minimizing cost. The initial weight of the value term is set to 0.7, and the weight of the cost term is set to 0.3. The dynamic adjustment mechanism of the weight coefficients is determined based on the convergence performance during the iteration process; the weight of the value term is increased when the convergence speed is fast, and the weight of the cost term is increased when the convergence is slow. The numerical range of the objective optimization function is 0 to 1, and the calculation precision is maintained to 5 decimal places. The function gradient is calculated using a numerical differentiation method with a step size set to 0.001.

[0109] Branch and bound search performs a systematic traversal of the solution space tree based on the objective function. The search process maintains key state information such as the current optimal solution, upper bound, and lower bound. The upper bound is obtained through linear programming with relaxed constraints, and the lower bound is determined by the objective function value of the current optimal solution. The branching strategy prioritizes the most promising node, expanding nodes with the largest upper bounds first. The bounding strategy determines pruning by comparing a node's upper bound with its current lower bound; nodes with upper bounds smaller than lower bounds are pruned. The search terminates when all nodes have been processed or a preset time limit is reached.

[0110] The comprehensive score calculation for candidate solutions is based on the results of branch-and-bound search. The score calculation includes three dimensions: objective function value, constraint satisfaction, and stability index. The objective function value is directly taken from the output of the optimization function; constraint satisfaction is quantified through a penalty term for breach of contract; and the stability index is evaluated based on the magnitude of function value change under parameter perturbations. The comprehensive score is calculated using a weighted summation method, with weights set to 0.6, 0.3, and 0.1 respectively. After ranking the scores, the highest-scoring solution is selected as the starting point for the greedy search.

[0111] The greedy search algorithm optimizes the local solution by finding the candidate with the highest overall score. It employs a hill-climbing strategy, starting from the current solution and searching for a better solution within its neighborhood. The neighborhood is defined as the set of solutions in the parameter space whose Euclidean distance is less than 0.1. The search step size uses an adaptive adjustment strategy: it increases by 10% for continuous improvement and decreases by 20% for no improvement. The lower limit for the step size is set to 0.001, and the upper limit is set to 0.5. The local optimization terminates when there is no improvement after 10 consecutive iterations or when the maximum number of iterations (500) is reached.

[0112] The gradient ascent method is used for iterative updates of the objective function weights. The weight update direction is calculated based on the partial derivatives of the objective function with respect to the weights, and the learning rate is set to 0.01. The weights are constrained to be non-negative and sum to 1, and this constraint is ensured through projection operations. During iteration, the weight values, objective function values, and convergence indices are recorded for each update. Convergence is determined based on the objective function value changing by less than 0.001 over five consecutive iterations. A momentum mechanism is used to accelerate convergence during the weight update process, with a momentum coefficient set to 0.9.

[0113] The Pareto front is constructed in the multi-objective optimization space using non-dominated sorting. The front contains the set of all non-dominated solutions, defined as a solution for which no other solution is inferior in all objectives and strictly superior in at least one objective. The front is constructed using a fast non-dominated sorting algorithm with a time complexity of O(MN). 2 ), where M is the number of objectives and N is the number of solutions. The front density assessment uses the crowding distance index, calculated based on the Euclidean distance between adjacent solutions in the objective space. The front maintenance strategy employs an elite retention mechanism, preserving all solutions on the front for subsequent analysis.

[0114] The optimal weight combination is determined using multi-criteria decision analysis based on the Pareto front. The decision method employs a strategy of minimizing the distance to the ideal point, defined as the optimal combination of objectives. Distance calculation uses weighted Euclidean distance, with weights set based on decision-maker preferences. Preference weights are determined using the analytic hierarchy process (AHP): value preservation preference is weighted at 0.5, cost control preference at 0.3, and environmental friendliness preference at 0.2. The optimal weight combination corresponds to the solution on the frontier closest to the ideal point.

[0115] The recycling implementation plan generates specific parameter configurations based on the optimal weight combination. The plan includes details such as material processing technology selection, process parameter settings, quality control standards, and environmental monitoring requirements. Process selection prioritizes the processes based on the weight combination results, choosing the process combination with the highest overall score. Parameter settings employ a weighted interpolation method, determining specific values ​​according to weight proportions within the feasible parameter range. Quality control standards are set based on the material value retention rate target, and environmental monitoring requirements are determined based on environmental load index limits.

[0116] In one optional implementation, the candidate solution with the highest comprehensive score is input into a greedy search algorithm for local optimization. The Pareto front is constructed in the multi-objective optimization space by iteratively updating the weight coefficients of the objective optimization function, including:

[0117] The candidate solution with the highest comprehensive score is set as the initial optimization node. An optimization feature vector is constructed based on the initial optimization node. A greedy search algorithm is used to determine the local search space, and iterative calculations are performed in the local search space. The convergence direction of each iteration is used as the basis for updating the weights of the target optimization function. The weight coefficients are dynamically adjusted based on the weight update basis, and the product of the weight coefficients and the optimization feature vector is used as the optimization direction of the target optimization function.

[0118] Sampling is performed in the multi-objective optimization space according to the optimization direction, and the optimal solution set is constructed based on the sampling results. The distribution curve of the non-dominated solutions in the optimal solution set is taken as the Pareto front.

[0119] The candidate solution with the highest overall score is selected as the initial optimization node, which contains all the key parameter information of the current solution. Feature extraction is performed on this initial optimization node to construct an optimization feature vector. Specifically, assuming the initial optimization node contains five key parameters: execution efficiency (0.85), resource consumption (0.62), stability (0.91), response time (0.78), and throughput (0.69), these five parameter values ​​are processed according to a preset normalization rule to form a standardized optimization feature vector. Each dimension of this vector corresponds to the normalized value of the aforementioned five parameters.

[0120] Based on the optimized feature vector, the range of the local search space is determined. Taking the initial optimized node as the center, a perturbation range is set for each feature dimension. The perturbation range for the execution efficiency dimension is set to ±0.1, for the resource consumption dimension ±0.08, for the stability dimension ±0.05, for the response time dimension ±0.12, and for the throughput dimension ±0.09. The hyperrectangular space formed within these perturbation ranges constitutes the local search space, which includes the neighborhood solutions surrounding the initial node.

[0121] The iterative computation process of the greedy search algorithm is performed in the local search space, starting from the initial optimization node and gradually adjusting each feature dimension. The execution efficiency parameter is adjusted in increments of 0.02, and the adjusted objective function value is calculated. When the execution efficiency parameter is adjusted to 0.87, the objective function value increases from the initial 1.35 to 1.42; this adjustment direction is recorded as a valid improvement direction. The resource consumption parameter is adjusted in increments of 0.015; when this parameter decreases to 0.605, the objective function value further increases to 1.48, which is also recorded as a valid improvement direction.

[0122] After each iteration, the trend of the objective optimization function value is analyzed to determine the convergence direction, and the contribution of adjustments to each feature dimension to the improvement of the objective optimization function is statistically analyzed. In the first iteration, the contribution of the execution efficiency dimension was 0.32, the resource consumption dimension was 0.28, the stability dimension was 0.15, the response time dimension was 0.18, and the throughput dimension was 0.07. These contributions are used as the basis for weight updates, reflecting the importance of each optimization objective in the current iteration stage.

[0123] The weight coefficients of each sub-objective in the objective optimization function are dynamically adjusted. Initially, the weight coefficients are set to 0.2 for execution efficiency, 0.2 for resource consumption, 0.2 for stability, 0.2 for response time, and 0.2 for throughput, exhibiting a uniform distribution. Based on the contribution obtained in the first iteration, the weight coefficient for execution efficiency is adjusted to 0.25, the weight coefficient for resource consumption to 0.23, the weight coefficient for stability to 0.18, the weight coefficient for response time to 0.21, and the weight coefficient for throughput to 0.13. The sum of the weight coefficients is kept at 1 to ensure the numerical stability of the optimization process.

[0124] The updated weight coefficients are multiplied by the components of each dimension of the optimized feature vector to obtain a weighted feature vector. In this weighted feature vector, the weighting value for execution efficiency is 0.2125, for resource consumption it is 0.1426, for stability it is 0.1638, for response time it is 0.1638, and for throughput it is 0.0897. These weighted values ​​are combined to form the optimization direction of the objective function, indicating the direction vector to be moved in the multi-objective optimization space.

[0125] Sampling operations are performed in the multi-objective optimization space along the determined optimization direction. Starting from the current optimization node, a sampling point is set at intervals of 0.05 along the optimization direction, for a total of 20 sampling points. Multiple objective function values ​​are calculated for each sampling point. For example, at sampling point number 7, the objective function values ​​are: execution efficiency 0.89, resource consumption 0.58, stability 0.93, response time 0.81, and throughput 0.72. At sampling point number 12, the objective function values ​​are: execution efficiency 0.91, resource consumption 0.55, stability 0.92, response time 0.84, and throughput 0.75.

[0126] Dominance relationships were determined for all sampled points, and non-dominated solutions were selected. Sampled point 7 was compared with sampled point 12. Sampled point 12 was superior to or equal to sampled point 7 in all four objectives: execution efficiency, resource consumption, response time, and throughput. Therefore, sampled point 7 was dominated by sampled point 12 and was not included in the non-dominated solution set. Sampled point 12 was compared with sampled point 15. Sampled point 15 was superior to sampled point 12 in the stability objective but inferior to sampled point 12 in the execution efficiency objective. Since they were not mutually dominant, both were included in the non-dominated solution set. After complete pairwise comparisons, 8 non-dominated solutions were selected from the 20 sampled points.

[0127] The selected non-dominated solutions are sorted according to specific objective dimensions to construct an optimal solution set. Execution efficiency is chosen as the primary sorting dimension, and the eight non-dominated solutions are arranged from lowest to highest execution efficiency, resulting in an execution efficiency sequence of 0.88, 0.90, 0.91, 0.92, 0.93, 0.94, 0.95, and 0.96. Simultaneously, the values ​​of each non-dominated solution in other objective dimensions are recorded, forming a complete multi-dimensional solution vector set. The distribution of these non-dominated solutions in the multi-objective optimization space constitutes a continuous curve, which is the Pareto front, representing the trade-offs and optimal equilibrium states among the optimization objectives.

[0128] In one optional implementation, a deviation analysis is performed between the measured performance data of the recycled material after the recycling implementation plan and the performance recovery range. When the deviation exceeds the tolerance range, a knowledge graph structural correction is triggered, including:

[0129] A performance deviation assessment matrix is ​​constructed, and the difference between the measured performance data of recycled materials and the performance recovery interval is used as the assessment parameter. The assessment parameter is input into the performance deviation assessment matrix to calculate the deviation feature distribution. The Monte Carlo method is used to perform probability sampling on the deviation feature distribution. Based on the result of the probability sampling, the deviation confidence interval is determined, and the deviation confidence interval is compared with the preset tolerance range.

[0130] When the deviation confidence interval exceeds the preset tolerance range, a knowledge graph correction space is constructed based on the beam search method. The eigenvalues ​​of the performance deviation evaluation matrix are used as search weights to calculate the structural adjustment direction in the knowledge graph correction space. Based on the structural adjustment direction, iterative optimization is performed on the knowledge graph correction space to determine the optimal correction scheme for the graph structure and to perform knowledge graph structural correction.

[0131] After the recycling program for recycled materials is implemented, a systematic deviation analysis of the actual performance of the recycled materials is required. A three-dimensional performance deviation evaluation matrix is ​​established. The rows of this matrix correspond to various performance indicators of the recycled materials, including tensile strength, flexural modulus, impact toughness, and other physical and mechanical properties. The columns correspond to different test batches, and the depth corresponds to different test locations. The difference between the measured data of each recycled material performance and the corresponding performance recovery range is calculated to obtain the evaluation parameter. For example, if the measured tensile strength of a batch of recycled plastic is 42 MPa, while the preset performance recovery range is 45 to 50 MPa, then the evaluation parameter for this item is -3 MPa. The evaluation parameters for all performance indicators are normalized according to their physical meaning, converted into dimensionless deviation coefficients, and filled into the corresponding positions in the performance deviation evaluation matrix.

[0132] The performance deviation evaluation matrix is ​​subjected to eigenvalue decomposition to extract the main deviation patterns. The distribution of data points in the matrix reflects the deviation characteristics of recycled material performance. Statistical analysis identifies the central tendency and dispersion of the deviations. The system records the value and frequency of each element in the evaluation matrix to construct a deviation characteristic distribution model. This distribution model is represented by a multi-dimensional vector, where each component of the vector represents the deviation intensity of a specific performance index under specific conditions.

[0133] Monte Carlo random sampling technique was used to perform probabilistic analysis on the deviation characteristic distribution. The sampling frequency was set to 10,000 times, with each sample randomly selecting a set of performance deviation data from the deviation characteristic distribution. During the sampling process, the system determined the probability of each value being selected based on the probability density function of the deviation characteristic distribution, ensuring that the sampling results accurately reflect the statistical characteristics of the deviation distribution. For the aforementioned recycled plastic case, the system recorded all values ​​of tensile strength deviation in 10,000 samples, forming a sampling set containing 10,000 data points. This sampling set was sorted, and percentile values ​​were calculated. The values ​​corresponding to the 2.5th and 97.5th percentiles were extracted as the boundaries of the deviation confidence interval. For example, if the sampling results show that the 2.5th percentile of the tensile strength deviation is -4.2 MPa and the 97.5th percentile is -1.8 MPa, then the deviation confidence interval for this performance index is -4.2 to -1.8 MPa.

[0134] The preset tolerance range is determined based on the application scenario and quality standards of the recycled materials. For recycled materials used in structural applications, the tolerance range is set to -5% to +5% of the lower limit of the performance recovery interval. The calculated deviation confidence interval is compared with the preset tolerance range to determine whether the deviation exceeds the acceptable range. In the aforementioned case, if the preset tolerance range is -2.25 to +2.5 MPa, while the lower bound of the actual deviation confidence interval is -4.2 MPa, then the deviation is determined to exceed the tolerance range, requiring the triggering of the knowledge graph's structural correction mechanism.

[0135] The knowledge graph correction space is constructed based on a bundle search algorithm framework. Nodes and edges in the knowledge graph are used as search objects. Each node represents a processing unit or material state in the recycling process, and each edge represents a transformation relationship or causal association between units. The bundle search method sets the bundle width parameter to 5, meaning that the top 5 candidate correction schemes with the highest scores are retained at each search level. The search weights are calculated based on the eigenvalues ​​of the performance deviation evaluation matrix. Larger eigenvalues ​​indicate a more significant impact of the deviation pattern on overall performance, resulting in higher search weights. The system performs eigenvalue decomposition on the performance deviation evaluation matrix, obtaining several eigenvalues ​​and their corresponding eigenvectors. The three largest eigenvalues ​​are used as the main search weights, assigned weight coefficients of 0.5, 0.3, and 0.2, respectively.

[0136] Calculating the direction of structural adjustment in the knowledge graph correction space involves evaluating the attributes of graph nodes and edges, traversing all nodes related to performance deviations in the knowledge graph, and analyzing the correlation between node attributes and deviation characteristics. For example, in the knowledge graph of recycled plastics, the particle size control node, melt temperature setting node, and additive ratio node are all related to the final tensile strength. The system calculates the sensitivity index of each node, which is determined by the ratio of the change in node attributes to the change in performance deviation. For the particle size control node, if adjusting the particle size from 5 mm to 3 mm improves the tensile strength deviation from -3 MPa to -1.5 MPa, then the sensitivity index of this node is 0.75 MPa per millimeter. The system sorts the sensitivity indices of all relevant nodes, selects the node with the highest sensitivity as the priority adjustment target, and determines the main direction of structural adjustment.

[0137] Based on the determined structural adjustment direction, an iterative optimization process is performed on the knowledge graph correction space. The maximum number of iterations is set to 50, and the convergence criterion is that the deviation improvement is less than 0.1% in three consecutive iterations. In each iteration, the system selects candidate schemes within the current bundle width and performs local corrections on the knowledge graph, including adjusting node attribute values, adding or deleting edge connections, and modifying edge weight coefficients. The corrected knowledge graph generates a new recycling execution scheme, and the system simulates the execution result of this scheme to predict the performance of the recycled material. By comparing the degree of matching between the predicted performance and the target performance recovery range, the score of the correction scheme is calculated. The score comprehensively considers the degree of reduction in performance deviation, the complexity of the correction operation, and the feasibility of the scheme. In the first round of iterations, the correction scheme for the particle size control node adjusts the particle size from 5 mm to 3 mm, and the score of this scheme is 0.85; the correction scheme for the melting temperature setting node adjusts the temperature from 190 degrees Celsius to 200 degrees Celsius, and the score is 0.78. The system retains the five schemes with the highest scores for the next round of iterations.

[0138] The iterative optimization process also needs to consider the synergistic effect of joint adjustments to multiple nodes, establish an interaction model between nodes, and evaluate the cumulative effect of different node correction operations. For example, when adjusting crushing particle size and melting temperature simultaneously, their effects on improving tensile strength may exhibit positive synergy or negative interference. The system quantifies this interaction effect through simulation calculations or historical data backtesting and incorporates it into the scoring calculation of the correction scheme. After multiple rounds of iteration, the system identifies a set of optimal node adjustment parameter combinations that reduce the predicted performance deviation of recycled materials to within the tolerance range while maintaining the economy and feasibility of the correction operation.

[0139] After determining the optimal correction scheme for the knowledge graph structure, a detailed correction instruction set is generated. This set includes the knowledge graph node identifiers to be modified, the corrected attribute values, the edge relationships to be adjusted, and the corresponding weight parameters. The system executes the knowledge graph structural correction, applying the correction instructions to the actual knowledge graph database and updating the storage information of relevant nodes and edges. After correction, the system regenerates the recycling execution scheme based on the updated knowledge graph and triggers a new round of material processing to ensure that the performance of subsequent recycled materials meets the expected recovery range requirements.

[0140] A second aspect of the present invention provides an electronic device, comprising:

[0141] processor;

[0142] Memory used to store processor-executable instructions;

[0143] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0144] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0145] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing purification parameters of recycled power battery materials based on hybrid intelligence, characterized in that, include: Multimodal degradation characterization data of the power battery to be diagnosed is acquired, a multi-layer heterogeneous battery material knowledge graph including material composition layer, degradation mechanism layer and regeneration process layer is constructed, and cross-layer causal reasoning paths are established between each layer. The multimodal degradation characterization data is mapped to the corresponding level nodes to form a graph state characterization that integrates physical degradation state and chemical composition changes. Based on the aforementioned graph state characterization, constraint propagation reasoning is performed. The constraints of material degradation irreversibility and recycling process cost-effectiveness boundary constraints are integrated along the cross-layer causal reasoning path. Failure modes of key components are identified through reverse tracing, and performance recovery intervals are predicted through forward deduction. This generates multi-granularity recycling candidate schemes containing confidence quantification, including: Based on the graph state representation, a constraint propagation network is constructed. The irreversible material degradation constraint is mapped to the state transition probability threshold of the material composition layer node through Monte Carlo tree search. The cost-efficiency boundary constraint of the regeneration process is mapped to the resource consumption upper limit of the regeneration process layer node. The state transition probability threshold is input into the conditional random field model to construct the constraint propagation probability distribution. Based on the constraint propagation probability distribution, a constraint propagation function is established on the cross-layer causal reasoning path. Bidirectional inference calculations are performed in the constraint propagation network. Based on the constraint propagation function, a Monte Carlo tree search method is used to calculate the contribution distribution of material components to failure modes through reverse sourcing and quantify their importance. The performance recovery degree under different process combinations is calculated through forward deduction, and the performance recovery interval is determined. The contribution distribution output by the constraint propagation function and the confidence level of the performance recovery interval are input into the conditional random field model for probabilistic inference. Based on the inference results, multi-granularity recycling candidate schemes are generated. Based on the material value retention rate, process energy consumption index, and environmental load index of each scheme in the multi-granularity recycling candidate schemes, and combined with the distribution of residual electrochemical activity of materials in the spectrum state characterization, a recycling implementation scheme is generated through multi-objective collaborative optimization. According to the recycling implementation plan, the measured performance data of the recycled materials after the implementation are compared with the performance recovery range. When the deviation exceeds the tolerance range, the knowledge graph structure is modified. The knowledge graph is updated by adding or deleting causal relationship edges of the cross-layer causal reasoning path and adjusting the constraint weights.

2. The method according to claim 1, characterized in that, A multi-layered heterogeneous battery material knowledge graph is constructed, comprising a material composition layer, a degradation mechanism layer, and a regeneration process layer. Cross-layer causal reasoning paths are established between each layer, mapping the multi-modal degradation characterization data to corresponding level nodes to form a graph-based state characterization that integrates physical degradation state and chemical composition changes. A multi-layer heterogeneous battery material knowledge graph containing a material composition layer, a degradation mechanism layer, and a regeneration process layer is constructed. The feature space of the multimodal degradation characterization data is decoupled, and the invariant components reflecting the intrinsic properties of the material and the evolutionary components reflecting the degradation process are extracted. The invariant components are mapped to the corresponding nodes of the material composition layer, and the evolutionary components are mapped to the corresponding nodes of the degradation mechanism layer through impedance spectrum mode matching, morphological degradation mode identification, and attenuation rate segmentation analysis. A material component-dependent causal relationship edge is established between the material composition layer and the degradation mechanism layer, and a process-responsive causal relationship edge is established between the degradation mechanism layer and the regeneration process layer. The material component-dependent causal relationship edge and the process-responsive causal relationship edge constitute a cross-layer causal reasoning path. Based on the material component-dependent causal relationship in the cross-layer causal reasoning path, reverse propagation is performed to trace the responsible material components, and based on the process response-type causal relationship, forward propagation is performed to predict the feasibility of process repair, generating a graph state representation.

3. The method according to claim 1, characterized in that, Based on the constraint propagation function, the Monte Carlo tree search method is used to calculate the contribution distribution of material components to failure modes through reverse sourcing and quantify their importance. The performance recovery degree under different process combinations is calculated through forward deduction, and the performance recovery range is determined, including: A tree-shaped search space is constructed based on the constraint propagation function. Material component nodes are used as the root nodes of the search tree, and failure mode nodes are used as the intermediate nodes of the search tree. The Monte Carlo tree search method is used to calculate the confidence probability of each search path through depth-first traversal. The contribution distribution of material components to failure modes is generated based on the confidence probability. The importance of components is determined based on the contribution distribution, and the component with the highest contribution is marked as the failure component. The failed components are mapped to the tree search space. The correspondence between process parameters and performance indicators is calculated based on the constraint propagation function. Forward propagation is used to calculate the degree of influence of the process parameters on performance recovery. The performance recovery range is determined based on the degree of influence.

4. The method according to claim 1, characterized in that, Based on the material value retention rate, process energy consumption index, and environmental load index of each scheme in the multi-granularity recycling candidate schemes, and combined with the distribution of residual electrochemical activity of materials in the spectral state characterization, a recycling implementation scheme is generated through multi-objective collaborative optimization, including: The material value retention rate, process energy consumption index and environmental load index of each scheme in the multi-granularity recycling candidate scheme are obtained. The distribution of residual electrochemical activity of materials in the spectral state characterization is obtained. The distribution of residual electrochemical activity of materials is used as a constraint to construct a multi-objective optimization space. Based on the multi-objective optimization space, a solution space tree is constructed using a bundle search method. The material value retention rate is used as the input term of the value evaluation function, and the process energy consumption index and the environmental load index are used as the input terms of the cost evaluation function. A target optimization function is generated based on the weighted combination of the value evaluation function input term and the cost evaluation function input term. Based on the target optimization function, a branch and bound search is performed on the solution space tree to calculate the comprehensive score of each candidate solution. The candidate solution with the highest comprehensive score is input into a greedy search algorithm for local optimization. The weight coefficients of the objective optimization function are updated iteratively to construct a Pareto front in the multi-objective optimization space. The optimal weight combination is determined based on the Pareto front, and a cyclical execution solution is generated based on the optimal weight combination.

5. The method according to claim 4, characterized in that, The candidate solution with the highest comprehensive score is input into a greedy search algorithm for local optimization. By iteratively updating the weight coefficients of the objective optimization function, a Pareto front is constructed in the multi-objective optimization space, including: The candidate solution with the highest comprehensive score is set as the initial optimization node. An optimization feature vector is constructed based on the initial optimization node. A greedy search algorithm is used to determine the local search space, and iterative calculations are performed in the local search space. The convergence direction of each iteration is used as the basis for updating the weights of the target optimization function. The weight coefficients are dynamically adjusted based on the weight update basis, and the product of the weight coefficients and the optimization feature vector is used as the optimization direction of the target optimization function. Sampling is performed in the multi-objective optimization space according to the optimization direction, and the optimal solution set is constructed based on the sampling results. The distribution curve of the non-dominated solutions in the optimal solution set is taken as the Pareto front.

6. The method according to claim 1, characterized in that, Based on the recycling implementation plan, a deviation analysis is performed between the measured performance data of the recycled materials after implementation and the performance recovery range. When the deviation exceeds the tolerance range, a knowledge graph structural correction is triggered, including: A performance deviation assessment matrix is ​​constructed, and the difference between the measured performance data of recycled materials and the performance recovery interval is used as the assessment parameter. The assessment parameter is input into the performance deviation assessment matrix to calculate the deviation feature distribution. The Monte Carlo method is used to perform probability sampling on the deviation feature distribution. Based on the result of the probability sampling, the deviation confidence interval is determined, and the deviation confidence interval is compared with the preset tolerance range. When the deviation confidence interval exceeds the preset tolerance range, a knowledge graph correction space is constructed based on the beam search method. The eigenvalues ​​of the performance deviation evaluation matrix are used as search weights to calculate the structural adjustment direction in the knowledge graph correction space. Based on the structural adjustment direction, iterative optimization is performed on the knowledge graph correction space to determine the optimal correction scheme for the graph structure and to perform knowledge graph structural correction.

7. A hybrid intelligent power battery recycled material purification parameter optimization system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to acquire multimodal degradation characterization data of the power battery to be diagnosed, construct a multi-layer heterogeneous battery material knowledge graph including a material composition layer, a degradation mechanism layer, and a regeneration process layer, and establish cross-layer causal reasoning paths between each layer to map the multimodal degradation characterization data to the corresponding level nodes, forming a graph state characterization that integrates physical degradation state and chemical composition changes; The second unit is used to perform constraint propagation reasoning based on the spectrum state characterization, integrate the material degradation irreversibility constraint and the cost-effectiveness boundary constraint of the recycling process on the cross-layer causal reasoning path, identify the failure modes of important components through reverse tracing and predict the performance recovery range through forward deduction, and generate multi-granularity recycling candidate schemes containing confidence quantification; The third unit is used to generate a recycling implementation plan through multi-objective collaborative optimization based on the material value retention rate, process energy consumption index, and environmental load index of each scheme in the multi-granularity recycling candidate schemes, combined with the distribution of residual electrochemical activity of materials in the spectrum state characterization; The fourth unit is used to perform deviation analysis between the measured performance data of the recycled materials after the execution of the recycling plan and the performance recovery range. When the deviation exceeds the tolerance range, the knowledge graph structure is modified. The knowledge graph is updated by adding or deleting causal relationship edges of the cross-layer causal reasoning path and adjusting the constraint weights.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images