Two-channel heterogeneous time sequence causal analysis method based on characterization embedding and time-delay pairing

Abstract

The invention relates to the technical field of data analysis and artificial intelligence, in particular to a two-channel heterogeneous time sequence causal analysis method based on characterization embedding and time delay pairing, which comprises the following steps: collecting state data of a power battery to obtain multi-scale features; performing iterative optimization based on the multi-scale features to minimize a loss function to obtain a first edge type confidence probability of the power battery state causality graph; determining a target-explanatory variable pair, and randomly extracting data from the multi-scale features for multiple times to form a sampling set; performing conditional independence test on the target-explanatory variable pair based on a sampling set to obtain a second edge type confidence probability of the power battery state causality graph; based on the first and second edge type confidence probabilities of the power battery state causal diagram, obtaining a fusion causal diagram and carrying out acyclic processing to obtain an acyclic power battery state causal diagram as an analysis result; according to the method, the causal discovery robustness, interpretability and efficiency can be enhanced.

Description

Technical Field

[0001] The present invention relates to the field of data analysis and artificial intelligence technology, and in particular to a dual-pathway heterogeneous temporal series causal analysis method based on representation embedding and time-delay pairing. Background Art

[0002] With the development of industries like the Industrial Internet of Things and advances in digital sensing technology, multimodal time-series data is rapidly increasing. Data correlation alone cannot effectively guide decision-making. Causal analysis, with its powerful ability to reveal causal information in data, has become a key technology for scenarios such as fault tracing and risk early warning.

[0003] Existing causal discovery techniques primarily include constraint-based methods, such as PCMCI, LPCMCI, and CD-NOD. These algorithms rely on the Markov hypothesis and the causal fidelity hypothesis, iteratively delete and direct edges through conditional independence tests, and can output equivalence class structures when variable types are unknown. Functional causal or scoring methods, such as VarLiNGAM and TiMINo, fit data to structural equations and search for optimal graphs using likelihood or information criteria, directly generating directed acyclic graphs. Differentiable optimization methods, such as DYNOTEARS and NTS-NOTEARS, transform combinatorial optimization into smooth optimization by continuousizing the adjacency matrix and introducing acyclic constraints, supporting end-to-end training. Granger causality methods determine time series causality based on prediction gain. They are easy to use but essentially measure predictability. Information theory-based methods, such as TE, STE, PTE, and DTE, can capture nonlinear information flows without explicit models but require large high-dimensional estimation samples. There are also emerging methods, such as cross-environment invariance and reinforcement learning intervention, but they generally still use a single metric space.

[0004] However, the above methods usually only model within a single metric space (continuous or discrete), and it is difficult to simultaneously process heterogeneous information such as counts, categories, and intensities. Specifically, they are manifested as follows: 1) the data must first be discretized or continuous, resulting in varying degrees of information loss; 2) the test error or search space grows exponentially with the dimension, and the accuracy decreases when the sample is insufficient; 3) there is a lack of a unified confidence assessment mechanism, making it difficult to quantify inference uncertainty. Summary of the Invention

[0005] In view of the above problems, the present invention provides a dual-path heterogeneous time series causal analysis method based on representation embedding and time lag pairing, which solves the technical problems in the existing technology that it is difficult to simultaneously retain continuous intensity information and discrete confidence information in multi-source heterogeneous time series data, and the inference efficiency and quantifiable uncertainty assessment are insufficient.

[0006] The present invention provides a dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing, comprising the following steps:

[0007] Step S1: collecting power battery status data at fixed time intervals, aggregating and splicing the collected power battery status data at each time point in multiple time windows to obtain multi-scale features;

[0008] Step S2: determining a loss function, performing iterative optimization based on the multi-scale features to minimize the loss function, and obtaining a confidence probability of a first edge type of a power battery state causal graph;

[0009] The nodes of the power battery state causal graph are used to represent variables in the power battery state data, and the edge connecting two nodes is used to represent the causal relationship type of the two variable values;

[0010] Step S3: determining the target variable and explanatory variable in the power battery status data to form a target-explanatory variable pair, randomly extracting data from the multi-scale features multiple times to form a sampling set; performing a conditional independence test on the target-explanatory variable pair based on the sampling set to obtain a second edge type confidence probability of the power battery status causal graph;

[0011] Step S4: obtaining a fused causal graph based on the confidence probabilities of the first and second edge types of the power battery state causal graph;

[0012] Step S5: perform acyclic processing on the fused causal graph to obtain an acyclic power battery state causal graph as an analysis result.

[0013] Preferably, the variables in the power battery status data in step S2 include: average cell voltage, maximum cell temperature, battery pack charge and discharge current, state of charge, battery health status, insulation resistance, ambient temperature and rate of change of state of charge; the causal relationship types in step S2 include: forward, reverse and none.

[0014] Preferably, the steps of aggregating and splicing multiple time windows in step S1 specifically include:

[0015] For each time window, a weighted calculation is performed on the historical values ​​of each variable at each time point in the time window to form the local features of the variable in the time window;

[0016] The window features of all variables and all window scales are concatenated to obtain the multi-scale features.

[0017] Preferably, step S2 specifically includes:

[0018] Step S2-1, determining the loss function for iterative optimization, and setting the initial value of the edge type prior probability of the power battery state causal graph to be uniformly distributed;

[0019] Step S2-2: randomly extract data from the multi-scale features multiple times, and obtain an inferred edge type prior probability based on the frequency of the edge type in the extracted data; iteratively update the edge type prior probability and calculate a loss function; when the loss function is less than a preset threshold, the iteration is completed;

[0020] Step S2-3: For multiple random sampling data in the latest round of iteration, count the frequency of each edge type in the causal graph Determine the confidence probability of the first edge type as Among them, K is the number of times the data is randomly sampled. Indicates that the obedience parameter is Beta distribution.

[0021] Preferably, the loss function in step S2-1 is:

[0022] L=L ELBO +λ1L stable +λ2L struct +λ3L balance

[0023] Or L=L ELBO +λ1L sparse +λ2L balance +λ3L prior

[0024] Among them, L is the loss function, L ELBO is the likelihood loss, L stable is the edge type probability loss, L struct is the structural constraint loss, L balance is the structural balance loss, L sparse is the sparse regularization loss, L prior is the prior consistency loss, λ1, λ2 and λ3 are weighting coefficients;

[0025] The expression for iteratively updating the edge type prior probability in step S2-2 is:

[0026]

[0027] in, is the prior probability of edge type inferred by the t+1 round model, is the prior probability of edge type inferred by the t-round model, It represents the prior probability that the variable pair e output in the tth iteration belongs to type λ, and ξ is the smoothing coefficient.

[0028] Preferably, step S3 specifically includes:

[0029] Step S3-1, classifying the variables in the multi-scale features into target variables and explanatory variables;

[0030] Step S3-2, randomly sampling data from the multi-scale features multiple times with replacement, determining the type of causal relationship for each pair of variables in each sampling data and recording;

[0031] Step S3-3, counting the frequency of occurrence of each edge type in the causal graph in the recorded content The second edge type confidence probability is determined as Where B represents the total number of random sampling with replacement, represents a Beta distribution with parameters .

[0032] Preferably, in step S3-1, the state of charge and the state of health of the battery are taken as target variables, and the average voltage of the monomer, the maximum temperature of the monomer, the charge and discharge current of the battery pack, the insulation resistance, the ambient temperature and the rate of change of the state of charge are taken as explanatory variables.

[0033] Preferably, step S4 specifically includes:

[0034] Step S4-1, calculate the fusion probability according to the following expression:

[0035]

[0036] Wherein, Table Fusion probability, γ is the fusion weight;

[0037] Step S4-2, for each edge e, according to the fusion probability Select the edge type with the highest confidence as the edge type of e, and get the fused causal graph Α * .

[0038] Preferably, step S5 specifically includes:

[0039] Step S5-1, sort all directed edges according to From high to low, and record the sequence;

[0040] Step S5-2, define the initial directed edge set E * According to the recorded sequence, and initialize the causal network G as a network without edges;

[0041] Step S5-3, add the edges in E * To the current causal network, and use topological sorting to determine whether there is a loop in the current causal network; if a loop is found in the network, remove the edge with the lowest confidence from the edge set that generates the loop;

[0042] Step S5-4, return to step S5-3 until there is no loop in the network, and get the edge set E' DAGcorresponding to the acyclic, directed causal network with the highest total confidence;

[0043] Step S5-5, determining the acyclic directed edge set E' according to the acyclic directed edge set E DAG determining the acyclic dynamic battery state causal graph A *,DAG .

[0044] Compared with the prior art, the present application has at least the following beneficial effects:

[0045] (1) The present application retains the original information of the dynamic battery state data in continuous and discrete metric spaces at the same time, without the need for complex preprocessing steps such as discretization and normalization of the data, thereby maximizing the avoidance of information loss introduced by preprocessing.

[0046] (2) The present application sets the prior distribution of edge type in causal structure learning, and continuously performs dynamic smoothing and accumulation of edge type prior probability using historical model iterative inference results, so that the model can quickly converge with less data samples and computational cost. At the same time, combining structure constraints and sparse regular loss terms, the redundant edges are effectively reduced, and the overall efficiency and scalability of causal inference are improved.

[0047] (3) The present application adopts a double-path fusion inference strategy based on Bayesian confidence bootstrap sampling, obtains confidence probabilities in the representation embedding and time-lag pairing paths respectively through multiple sampling with replacement and conditional independence test, finally fuses them as the causal inference result, and performs beta distribution modeling and evaluation on the result, which can enhance the robustness and interpretability of causal discovery. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings are for the purpose of illustrating preferred embodiments of the present application and are not to be construed as limiting the present application.

[0049] Figure 1 The flowchart of the double-path heterogeneous time series causal analysis method based on representation embedding and time-lag pairing provided by the present application.

[0050] Figure 2 The first part of the detailed flowchart of the double-path heterogeneous time series causal analysis method based on representation embedding and time-lag pairing provided by the present application.

[0051] Figure 3 The second part of the detailed flowchart of the double-path heterogeneous time series causal analysis method based on representation embedding and time-lag pairing provided by the present application.

[0052] Figure 4 The first part of the principle diagram of the double-path heterogeneous time series causal analysis method based on representation embedding and time-lag pairing provided by the present application.

[0053] Figure 5This is the second part of the principle diagram of the dual-pathway heterogeneous temporal causal analysis method based on representation embedding and time delay pairing provided by the present invention.

[0054] Figure 6 This is the third part of the principle diagram of the dual-pathway heterogeneous temporal causal analysis method based on representation embedding and time delay pairing provided by the present invention.

[0055] Figure 7 This is the fourth part of the principle diagram of the dual-pathway heterogeneous temporal causal analysis method based on representation embedding and time delay pairing provided by the present invention.

[0056] Figure 8 The present invention provides a causal graph obtained by multi-type Bayesian confidence bootstrap sampling.

[0057] Figure 9 This is a causal graph obtained by the structural prior adaptive multi-type amortized causal discovery provided by the present invention.

[0058] Figure 10 This is the causal graph obtained by aggregating heterogeneous dual-source results provided by the present invention.

[0059] Figure 11 This is the causal graph after the acyclic processing provided by the present invention. DETAILED DESCRIPTION

[0060] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0061] In response to the problems that existing technologies have difficulty in simultaneously retaining continuous intensity information and discrete confidence information in multi-source heterogeneous time series data, and that inference efficiency and quantifiable uncertainty assessment are insufficient, the present invention provides a dual-pathway heterogeneous time series causal analysis scheme based on representation embedding and time lag pairing. It can efficiently integrate multiple types of features, quantify the confidence of causal structures, and support end-to-end inference in large-scale scenarios.

[0062] A specific embodiment of the present invention is Figure 1 As shown, a dual-pathway heterogeneous temporal causal analysis method based on representation embedding and time lag pairing is disclosed, comprising the following steps:

[0063] Step S1: collecting power battery status data at fixed time intervals, aggregating and splicing the collected power battery status data at each time point in multiple time windows to obtain multi-scale features;

[0064] In this step, the state of the new energy vehicle power battery is first set. The power battery state has multiple variables, such as the voltage, current, temperature, and health status of the power battery.

[0065] Collect N-dimensional time series data D = {w t |t=0,1,...,T-1} (at each time point, the observation results of all variables are listed in order), where, Represents the observation vector at each time point, that is, the joint observation value of all variables at time point t, where Represents the first variable at time t, and so on until N is the number of variables, and T represents the total sampling length of the time series, that is, the total number of time points / data points.

[0066] Afterwards, the power battery status data at each time point is subjected to multi-scale adaptive windowing and dynamic window sliding summation.

[0067] Set the multiple time scale set S = {s1,s2,...,s L} and the maximum time lag τ max Among them, S represents the set of all window lengths used for time series feature extraction, s l is the window length of the lth scale, and L is the total number of scales. Each scale can be specified based on the dynamic characteristics of the variable or experience, aiming to capture signal changes and causal effects at different time granularities.

[0068] For the i-th variable, at each window size s l (s l ∈S), at each time point t, using the most recent s l The historical values ​​are weighted to form the local features at the window scale, and the expression is:

[0069]

[0070] in, It represents the characteristic value of the i-th variable after aggregation with the l-th scale window at time t, and is a component of the multi-scale, multi-variable combination feature; represents the observed value of the i-th variable at the k-th time point before the current time point t; α i,l,k is the dynamic weight coefficient of the i-th variable in the l-th scale window at k time points away from the end of the window. It can be designed as a customized weight according to actual needs or an adaptive weight can be obtained through data-driven model learning to make the window's sensitivity to new and old information controllable.

[0071] The window features of all variables and all window scales are spliced ​​to obtain the multi-scale combined feature vector at time point t: where z t The feature concatenation vector for all variables and all scales at time point t is used for the subsequent causal inference entry. t Arrange them in sequence to obtain the processed multi-scale feature sample set: Z = {z t |t≥max(S)}, where Z represents the entire sample set after multi-scale windowing and sliding weighting. This multi-scale feature effectively reflects the dynamic activities and interactive characteristics of multiple variables at different time scales and time lags. max(S) represents the maximum window length in the set of window lengths.

[0072] Step S2: determining a loss function, performing iterative optimization based on the multi-scale features to minimize the loss function, and obtaining a confidence probability of a first edge type of a power battery state causal graph;

[0073] The nodes of the power battery state causal graph are used to represent variables in the power battery state data, and the edge connecting two nodes is used to represent the causal relationship between the two variable values.

[0074] The present invention models the variables in the power battery status data as nodes in a causal graph, and the edge between two nodes is used to represent the causal relationship between the two variable values. Unlike traditional binary (yes / no) edge modeling, the present invention provides a modeling method for the multivariate probability distribution of edge types, determining the type of each edge as a polynomial distribution including types such as "forward, reverse, none", thereby improving the expressiveness of structural modeling. For the node pair (i, j), the edge type λ∈∧, where ∧ is the set of all types, such as forward, reverse, none, etc., ∧ can flexibly define the boundary space.

[0075] In this step, set is the prior probability that edge e belongs to type λ at time point t, satisfying The present invention sets the initial value of the prior probability at the initial moment For uniform distribution,

[0076] Establish a probability model q θ (A|Z), where A is a causal graph, represented by an N×N adjacency matrix, and the values ​​of the elements in the matrix represent the probability of each edge belonging to a certain type; q θ (A|Z) represents the posterior probability distribution of the type of each edge in the causal graph A under the condition of multi-scale feature Z.

[0077] The loss function optimized by this invention uses a multi-objective regularization approach to improve structural robustness and interpretability. The loss function covers structural robustness (encouraging consistent edge frequencies or high / low confidence differentiation), graph coefficient regularization (encouraging overall structural coefficients or stratification), and edge frequency balance (avoiding excessive bias of certain types of edges). The total structural loss function is constructed as:

[0078] L=L ELBO +λ1L stable +λ2L struct +λ3L balance

[0079] Or L=L ELBO +λ1L sparse +λ2L balance +λ3L prior

[0080] Among them, L is the loss function, L ELBO is the likelihood loss, L stable is the edge type probability loss, L struct is the structural constraint loss, L balance is the structural balance loss, L sparse is the sparse regularization loss, L prior is the prior consistency loss, λ1, λ2, and λ3 are the weighted coefficients of the three types of auxiliary constraint target losses, which control the influence of different constraint losses on the total constraint loss. The specific expressions of each loss are as follows:

[0081]

[0082] L sparse =∑ (i,j) ||p ij ||1

[0083] L prior =||q θ (Α)-p t (Α)||2

[0084] in, Represents the structural distribution q θ The expectation of logp(Z|A) represents the log likelihood of the multi-scale feature Z in the causal graph A, q θ (A) represents the posterior probability distribution of the causal graph A, p(A) represents the prior probability distribution of the causal graph A, KL(·) represents the Kullback-Leibler divergence, p e,λ represents the prior probability that edge e belongs to type λ; α1 is the sparsity weight, α2 is the layer penalty weight, μ is the layer number coefficient; |E| is the total number of edges, ||p ij||1 is the L1 norm of the edge type probability between the nodes corresponding to the i-th and j-th variables; p t (Α) represents the prior probability distribution of the causal graph A at time point t, and ||·||2 represents the calculation of the L2 norm.

[0085] It means traversing the entire graph and summing the node pairs e=(i,j) in the structure graph; It means that for each edge e, the probability of all edge types λ other than the “no connection” type is accumulated.

[0086] The above loss function is described as follows:

[0087] Likelihood loss L ELBO Used to measure the distance between the data distribution generated by the model and the true distribution; edge type probability L stable The loss encourages the structural inference output to be more stable, penalizes the high entropy of the edge type probability distribution, and makes the model tend to output "more confident" edge type probabilities, helping the model to converge stably; the structural constraint loss L struct The model is encouraged to learn a structure that is as simple and sparse (with as few edges as possible) to avoid over-density of the causal network. The coefficients of the layers are often prior knowledge, manually specified, or automatically inferred by the network to prevent the causal network results from being too complex. balance is the type distribution balance constraint loss, which measures whether the edge type distribution is balanced; |E| is the total number of edge sets, that is, the number of all edges that can be taken, which is N(N-1) for directed and N(N-1) for undirected. represents the average proportion of edge type λ in the entire structure, that is, the normalized probability, It constitutes an entropy to measure the uniformity of category probability (type distribution). If the distribution of all types is extremely skewed (for example, only one edge type appears and the others are zero), the entropy value is the smallest; if the distribution of different types is relatively balanced, the entropy value is the largest, indicating that the model output structure types are rich, which is conducive to generalization and stability.

[0088] The present invention performs iterative optimization based on these multi-scale features to minimize the loss function. This iterative process utilizes an automatic evolution mechanism for the structural prior. Prior weights are adjusted as the model iterates and the sampling distribution changes, allowing the prior to reflect data-driven structural evolution, achieving a transition from a "static" structural prior to a "dynamically generative" one.

[0089] The initial value of the prior probability at the initial time To achieve uniform distribution, data is randomly sampled multiple times from the multi-scale features in each iteration, mini-batch / B-step sampling is performed, the structural tendency of the edge frequency distribution obtained by sampling is estimated, and the prior probability distribution is dynamically corrected. The expression is:

[0090]

[0091] in, is the prior probability of edge type inferred by the t+1 round model, is the prior probability of edge type inferred by the t-round model, It represents the prior probability that the variable pair e output in the tth iteration belongs to type λ, ξ is the smoothing coefficient, and its value range is ξ∈[0,1].

[0092] ξ determines the prior probability of this iteration and the current observed edge distribution The weighted fusion ratio of ξ is ξ. If ξ is close to 1, the prior is dominant and the weight is conservative; if ξ is small, the new sampling distribution has a greater impact and the weight adaptation is faster.

[0093] After each round of iteration, the edge type prior probability is updated and the loss function is calculated. After multiple iterations, the iteration is completed when the loss function is less than the preset threshold.

[0094] During the latest iteration, data is randomly extracted from the multi-scale features multiple times, and the frequency of the occurrence of the e-th edge belonging to type λ is counted. based on The number of occurrences of each edge sampling is parameterized as Beta distribution, and the confidence probability of the first edge type of the power battery state causal graph is finally obtained as in, represents the Bayesian posterior probability distribution of edge e being of type λ, Indicates that the obedience parameter is Beta distribution, K is the number of times the data is randomly sampled, and the 2.5% and 97.5% quantiles of the Beta distribution are taken as the confidence intervals.

[0095] In some embodiments, for each edge, by comparing The edge type with the largest number of occurrences, that is, the largest sampling ratio, is selected as the final judgment type to form a causal graph corresponding to the confidence probability of the first edge type.

[0096] Step S3: determining the target variable and explanatory variable in the power battery status data to form a target-explanatory variable pair, randomly extracting data from the multi-scale features multiple times to form a sampling set; performing a conditional independence test on the target-explanatory variable pair based on the sampling set to obtain a second edge type confidence probability of the power battery status causal graph;

[0097] In the previous step, multi-scale features Z have been extracted using weighted windows of different time scales.

[0098] In some embodiments, the variables in the multi-scale features may be classified into target variables and explanatory variables, respectively. The explanatory variables include the candidate dependent variables and the conditional variable group.

[0099] In order to avoid high correlation between adjacent samples, in this step, the window step size is set to sample the multi-scale feature Z in equal time windows to obtain window samples. Among them, X, Y, and Z represent different sets of variable nodes, X represents the target variable, Y represents the candidate dependent variable, and Z represents the conditional variable group. Indicates that at index s, for the variable combination (X, Y, Z) and its corresponding window observation samples, it contains The sample vector after splicing three groups of observations with different variables and lags, where τ X , τ Y , τ Z Respectively represent the time lag order corresponding to variables X, Y, and Z, that is, the historical step length referenced by each variable

[0100] Get window sample set Among them, I s Represents a set of sampling indices that can be used for causal analysis, defined as I s ={τ max ,...,T-1}, where τ max is the maximum lag order of all involved variables.

[0101] Then, sample reorganization and causal inference are performed: multi-scale feature Z, repeated B = 100 times with replacement sampling, to form a sampling set Where b = 1,...,B. For each sample set, a conditional independence test is performed. For each pair of variables (X, Y), the presence of a causal effect is analyzed, and the type of edge is recorded: positive, negative, or no edge. In this paper, the P-value is a probability indicator obtained from a statistical hypothesis test, indicating the probability of observing the current or more extreme results under the null hypothesis of no causal effect. The P-value is used to determine the significance of the causal or association relationship between the variable pairs (X, Y).

[0102] Count the number of times edge e type λ appears to estimate the probability, which is recorded as Where λ∈∧, where ∧ is the set of all allowed edge types, including directed, reverse, undirected, no edge, etc., and e represents the edge connecting the variable pair. The confidence probability of the second edge type is modeled using the Beta distribution: Where B represents the total number of samplings with replacement, Indicates that the obedience parameter is Beta distribution.

[0103] In some embodiments, a corresponding causal graph may be obtained according to the second edge type confidence probability.

[0104] Step S4: obtaining a fused causal graph based on the confidence probabilities of the first and second edge types of the power battery state causal graph;

[0105] The present invention adopts a dynamic weight integration method to simultaneously run step S3 and step S4 on the same data to obtain the confidence probabilities of the first and second edge types. and The fusion probability is:

[0106]

[0107] in, represents the fusion probability, γ is the fusion weight, which can be adapted to different application scenarios and flexibly set according to historical accuracy or the width of the Bayesian interval.

[0108] It represents the fused confidence probability of each edge e and all possible types λ, which is ultimately used to select the type of each edge.

[0109] For each edge e, according to the fusion probability Select the edge type with the highest confidence (probability): Get the final fusion causal graph A * (The form can be a causal graph structure composed of the best types of all edges, which can be an adjacency matrix, an edge list, etc.), the probability distribution of each edge type (i.e. complete distribution information to support probabilistic interpretation and downstream analysis), marginal confidence intervals (which can be used to provide a basis for risk assessment and high-confidence analysis), and multi-scale feature indexing (which associates and records the contributions of original variables and windows of different time scales to facilitate traceability analysis and explainable modeling).

[0110] Step S5: perform acyclic processing on the fused causal graph to obtain an acyclic power battery state causal graph as an analysis result.

[0111] In this step, first, all directed edges are sorted according to the confidence Sort from high to low and record the sequence; perform iterative edge deletion and loop elimination. The steps include:

[0112] (1) Define the initial directed edge set E according to the recorded sequence * Initialize the causal network G as a network without edges and loop through the following steps until the network is acyclic:

[0113] (2) E *The edges in are added to the current causal network. After each new edge is introduced, topological sorting is used to determine whether there is a loop in the current causal network. If a loop is found in the network, the edge with the lowest confidence is removed first in the edge set that generates the loop.

[0114] (3) Return to step (2) until there is no loop in the network, and the obtained edge set E' DAG That is, it corresponds to a directed causal network without cycles and with the highest total confidence.

[0115] Finally, according to the acyclic directed edge set E' DAG Determine the acyclic power battery state causal graph A *,DAG .

[0116] This acyclic processing method fully utilizes the fusion confidence in the previous steps Prioritize the causal connections with high confidence and eliminate the redundant loops caused by insufficient confidence. After the algorithm terminates, the causal structure matrix A is obtained. *,DAG It has natural directed acyclicity, and the processing method is automatically iterative and does not require human intervention.

[0117] The acyclic power battery state causal graph finally obtained by the present invention not only reveals the causal influence path between key state variables such as SOC, SOH, internal resistance, temperature, etc., but also quantifies the Bayesian confidence for each causal edge. Starting from this causal graph, the following scenarios can be directly applied: (1) The causal graph can be converted into a noisy causal equation model, and the residuals of each equation can be weighted by the previously calculated edge confidence, which can accurately fit the nonlinear mapping between variables and make a strict assessment of the confidence interval of the predicted output; (2) The causal structural equation model can be further used for intervention analysis, such as the impact of adjusting the charging current to a certain level on SOH. The posterior distribution of SOH or remaining life (RUL) can be quickly obtained through Monte Carlo or analytical analysis, providing a quantitative basis for operation strategy optimization and remaining life management; (3) The causal structural equation is embedded into the model predictive control framework, with the SOH degradation rate or temperature fluctuation amplitude as the objective function, and the control outputs such as charge and discharge current and temperature management are collaboratively optimized to achieve a multi-objective balance of performance, safety, and life; (4) Based on the posterior distribution of RUL and the confidence of the causal edge, dynamic maintenance and replacement timing are formulated to optimize the maintenance cycle, reduce maintenance costs, and ensure the operational reliability of the battery management system; (5) In the battery management system interface, the causal structural equation and confidence interval are updated regularly or in real time based on the confidence of the causal edge and new data, so as to realize the adaptive evolution of the model and intuitive decision support for operation and maintenance personnel.

[0118] In order to illustrate the effectiveness of the method proposed in the present invention, the above technical solution of the present invention is described in detail below through a specific embodiment.

[0119] Example 1

[0120] This embodiment takes a new energy power battery pack as an example to illustrate the implementation process of the present invention for causal structure inference of battery multi-source time series monitoring variables. Figure 2 、 Figure 3 The specific steps are as follows:

[0121] 1. First Path: Structural Prior Adaptive Multi-Type Amortized Causal Discovery

[0122] A group of new energy vehicles (a total of 20 experimental vehicles) were selected and the power battery data were continuously sampled for one month. The sampling period was Δt = 5 minutes, the number of sampling points per day was 288 points, and the total number of sampling points per month was T = 8640. The total number of variables was defined as N = 8, and each observation variable was defined as and in Indicates the average voltage of a cell (CellVolt, unit: V), Indicates the maximum temperature of the monomer (CellTemp, unit: ℃), Indicates the battery pack charge and discharge current (Current, unit: A), Indicates the state of charge (SOC, unit: %), Indicates the battery health status (SOH, unit: %), InsRes represents insulation resistance (unit: KΩ). Indicates the ambient temperature (EnvTemp, unit: °C), Indicates the rate of change of state of charge (DeltaSOC, unit: % / h).

[0123] The present invention selects the above parameters as variables to comprehensively reflect the operating status and safety of the battery system. Among them, voltage and current directly reflect the battery's operating status, temperature (including maximum temperature and ambient temperature) helps assess thermal management and safety risks, SOC and SOH describe the battery's remaining charge and overall health, respectively, insulation resistance reflects the safety of the battery system, and the state of charge change rate reveals the battery's dynamic performance under different operating conditions. The comprehensive utilization of these characteristics facilitates more accurate and comprehensive cause-and-effect diagram analysis.

[0124] All variables constitute the observation vector at each moment Where t = 1, 2, ..., T. Furthermore, each vehicle is stored separately, and all observation data form a three-dimensional tensor Where v represents the vehicle number, V = 20. The multi-source time series composed of sampled power battery data is as follows Figure 4 shown.

[0125] Three window scales are selected for multi-time scale feature extraction: s1 = 6 (30 minutes), s2 = 24 (2 hours), s3 = 72 (6 hours), and the window scale set is denoted as S = {s1, s2, s3}, and the number of windows is L = 3. For each variable i (i = 1, 2, ..., 8), each time t, and each scale s l , calculate the most recent s l Weighted window characteristics of sampling points: in, represents exponential decay weight; is the observed value of variable i at time tk; the window features of each variable, each scale, and each moment are obtained, totaling N×L=24 feature dimensions. The overall window feature at each moment is The feature set Z = {z t |t=max(S),...,T}, where the total number of rows in Z is T-max(S)+1.

[0126] For all pairs of variables, three types of edges, forward, reverse, and none, are assigned uniform priors: in Denotes the structural prior probability that the variable pair e belongs to type λ in round t. Establish the probability model q θ (A|Z), where A is the causal structure, that is, an N×N adjacency matrix, each edge is one of the three categories; Z is a multi-scale windowed feature matrix, which is the feature set of all observed samples. The loss function has multiple parts, including the likelihood loss L ELBO , sparse regularization L sparse =∑ (i,j) ||p ij ||1. Structural balance regularization (encourages uniform probability distribution of the three categories) L balance , prior consistency loss L prior =||q θ (Α)-p t (Α)||2, the total loss is L=L ELBO +λ1L sparse +λ2L balance +λ3L prior , where λ1, λ2 and λ3 are hyperparameters, p ij represents the element in the adjacency matrix, which represents the probability of the causal edge type between the i and j variable pairs, q θ (Α) represents the structural distribution inferred by the structural model, p t (A) represents the overall distribution of the structural prior matrix at round t. The optimizer uses the Adam optimizer with a learning rate of 0.001 and a maximum number of iterations T = 20. After each iteration, the structural prior is updated: in is the edge type probability inferred by the t-round model, represents the probability that the variable pair e output in the tth iteration belongs to type λ, ξ is the smoothing coefficient, and the value is 0.7. Each round K = 50 structural samplings are performed, and the frequency of all variable pairs and types is counted. Finally, according to the bootstrap path equation, the frequency of the occurrence of the eth variable pair belonging to type λ is counted Then the edge type confidence probability of the structural prior adaptive multi-type amortized causal discovery Beta distribution is Taking the 2.5% and 97.5% quantiles as the confidence interval, the results are as follows Figure 9 As shown in the figure, we can see that compared with path 1, the causal flow direction of w3→w2→w5 is retained. At the same time, the network structure has changed, and there are more direct connections between variables (such as w6→w5, w8→w4, ), the distribution and weight of edges have also changed, indicating that structural prior adaptive multi-type amortized causal discovery is a supplement to causal relationships.

[0127] The overall process diagram of structural prior adaptive multi-type amortized causal discovery is as follows Figure 5 shown.

[0128] 2. Second Pass: Multi-Type Bayesian Confidence Bootstrap Sampling

[0129] Assuming the window step size is Δ=3, and the sampling period is Δt=5 minutes, that is, sampling once every 15 minutes to avoid excessive sample correlation, the total number of observed samples is Each set of samples records the time, vehicle ID, and the corresponding 24-dimensional features. Then, a set of "target-explanatory" variable pairs is established, focusing on the causal target variable Y ( to ), construct all possible one-to-one variable pairs (X i ,Y j ), a total of N pair =(N-1)×2=14 pairs, where i=1,2,...,8 represents the subscript of the independent variable (explanatory variable), j=4,5 represents the subscript of the dependent variable (target variable), and except for the target variable itself, the remaining 7 variables can be used as explanatory variables X i .

[0130] For all observed samples Z, repeat B = 100 times with replacement sampling to form a sampling set Where b=1,...,B. On each sampling set, the conditional independence test is used, that is, for each pair of variables (X i ,Y j ), analyze X i Y j Is there a causal effect and record the type of edge: positive (X i →Y j), Reverse(X i ←Y j ) or no boundary. In the present invention, the P value is a probability index obtained from a statistical hypothesis test, which indicates the probability of observing the current and more extreme results under the assumption that there is no causal effect. The P value is used to judge the effect of the variable on (X i ,Y j The threshold is set to 0.05, that is, when the P value is less than 0.05, it is considered that there is a statistically significant relationship, and the specific method and relationship type can be further determined by the causal inference algorithm; if the P value is greater than or equal to 0.05, it is considered that there is no significant relationship and it is not recorded as a valid causal edge. is the number of times edge type λ appears, where λ∈{forward, reverse, none}, e represents a variable pair, and MTBBCS represents the multi-type Bayesian confidence bootstrap sampling method. The confidence probability estimate is calculated as follows: Model confidence intervals using the Beta distribution: Take the 2.5% and 97.5% quantiles as the upper and lower bounds of the interval. Record the three edge type confidence probabilities and their Bayesian confidence intervals for each variable pair. The results are as follows Figure 8 As shown in the figure, we can see that there is a causal relationship between w3, w2, w4 and w5, such as current affects temperature and then affects the battery health state SOH (w3→w2→w5); at the same time, some edges have higher weights, indicating that the causal direction is more reliable.

[0131] The overall process diagram of multi-type Bayesian confidence bootstrap sampling is as follows Figure 6 shown.

[0132] 3. Heterogeneous Dual-Source Result Aggregation and Causal Graph Generation

[0133] For each variable pair e and type λ, let the confidence probabilities of the two paths be and Its dynamic integration probability is: Among them, γ is the fusion weight, which is set to 0.5. It can be adapted to different application scenarios and flexibly set according to historical accuracy or the width of the Bayesian interval. The confidence interval of the two paths is taken as the weighted average output intersection. Finally, the edge type with the highest confidence for each variable is output, and its confidence probability lower bound is greater than 0.5 to output "high reliability causality". The remaining relationship types are recorded as "uncertain" or "no edge". The results are as follows Figure 10 As shown in the figure, we can see that the information and weights of path one and path two are integrated. Figure 10It includes more possible causal paths, the weights of some edges are improved, and the main causal directions (such as w3→w2→w5, w6→w5) are clearer, reflecting the true logic of multiple causality in complex systems. However, due to the existence of loops in path 2, there are still loops in the overall structure after fusion. Therefore, it cannot be directly used as the final diagram of causal intervention.

[0134] The overall process diagram of heterogeneous dual-source result aggregation and causal graph generation is as follows: Figure 7 shown.

[0135] 4. Implementation of acyclic constraints

[0136] Finally, the causal network G'=(V',E') obtained by integrated inference is acyclic, where V' represents the node set in the inferred causal network graph and E' represents the edge set in the inferred causal network. Sort from high to low and initialize the candidate network The second step is to try to add edges to the current network in sequence. After each new edge is introduced, a topological sort is used to determine whether there is a loop in G'0. The third step is to remove the edge with the lowest confidence in the edge set that generates the loop, and to repeatedly detect and remove it until there is no directed loop in the network. The fourth step is to remove the edge set E' after the above process. DAG That is, a directed causal network without cycles and with the highest total confidence. DAG =(V',E' DAG ) can ensure the structural correctness of causal inference and the consistency of physical explanation. The results are as follows Figure 11 As shown in the figure, the final acyclic constraint implementation systematically removes causal loops, forming a completely directed acyclic causal network. In this diagram, all edges clearly point to downstream variables (source variables such as w1, w2, w3, and w8 each influence w4 and w5 through several paths), and there are no closed loops. This clearly indicates which variables ultimately influence SOC and SOH. Each causal path can be traced back to its source, and all causal influences converge only on the output variable, without any closed loops or interference. Under this structure, causal results can be easily used for reliable causal analysis and intervention prediction, such as how regulating w3 affects w4 or w5, ensuring the validity and interpretability of causal inference. In this case, it is also shown that w4 is mainly affected by CellVolt and DeltaSOC, while w5 is mainly directly determined by CellTemp and InsRes, and is indirectly affected by variables such as Current, SOC, CellVolt, and DeltaSOC.

[0137] Although the specific embodiments of the present invention depict various actions or steps in a specific order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in a sequential order, or requiring that all illustrated actions or steps should be performed to obtain the desired results. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although a number of specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of this disclosure. Certain features described in the context of separate embodiments can also be implemented in combination in a single implementation. Conversely, the various features described in the context of a single implementation can also be implemented in multiple implementations individually or in any suitable sub-combination. The above is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the scope of protection of the present invention.

[0138] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A dual-pathway heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing, characterized by: The following steps are involved: Step S1: collecting power battery status data at fixed time intervals, aggregating and splicing the collected power battery status data at each time point in multiple time windows to obtain multi-scale features; Step S2: determining a loss function, performing iterative optimization based on the multi-scale features to minimize the loss function, and obtaining a confidence probability of a first edge type of a power battery state causal graph; The nodes of the power battery state causal graph are used to represent variables in the power battery state data, and the edge connecting two nodes is used to represent the causal relationship type of the two variable values; Step S3: determining the target variable and explanatory variable in the power battery status data to form a target-explanatory variable pair, randomly extracting data from the multi-scale features multiple times to form a sampling set; performing a conditional independence test on the target-explanatory variable pair based on the sampling set to obtain a second edge type confidence probability of the power battery status causal graph; Step S4: obtaining a fused causal graph based on the confidence probabilities of the first and second edge types of the power battery state causal graph; Step S5: perform acyclic processing on the fused causal graph to obtain an acyclic power battery state causal graph as an analysis result.

2. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 1 is characterized in that: The variables in the power battery status data in step S2 include: average cell voltage, maximum cell temperature, battery pack charge and discharge current, state of charge, battery health status, insulation resistance, ambient temperature and rate of change of state of charge; the causal relationship types in step S2 include: forward, reverse and none.

3. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 2 is characterized in that: The steps of aggregating and splicing multiple time windows in step S1 specifically include: For each time window, a weighted calculation is performed on the historical values ​​of each variable at each time point in the time window to form the local features of the variable in the time window; The window features of all variables and all window scales are concatenated to obtain the multi-scale features.

4. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 3 is characterized in that: Step S2 specifically includes: Step S2-1, determining the loss function for iterative optimization, and setting the initial value of the edge type prior probability of the power battery state causal graph to be uniformly distributed; Step S2-2: randomly extract data from the multi-scale features multiple times, and obtain an inferred edge type prior probability based on the frequency of the edge type in the extracted data; iteratively update the edge type prior probability and calculate a loss function; when the loss function is less than a preset threshold, the iteration is completed; Step S2-3: For multiple random sampling data in the latest round of iteration, count the frequency of each edge type in the causal graph Determine the confidence probability of the first edge type as Among them, K is the number of times the data is randomly sampled. Indicates that the obedience parameter is Beta distribution.

5. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 4 is characterized in that: The loss function in step S2-1 is: L=L ELBO +λ1L stable +λ2L struct +λ3L balance or L = L ELBO + λ1L sparse + λ2L balance + λ3L prior Among them, L is the loss function, L ELBO is the likelihood loss, L stable is the edge type probability loss, L struct is the structural constraint loss, L balance is the structural balance loss, L sparse is the sparse regularization loss, L prior is the prior consistency loss, λ1, λ2 and λ3 are weighting coefficients; The expression for iteratively updating the edge type prior probability in step S2-2 is: in, is the prior probability of edge type inferred by the t+1 round model, is the prior probability of edge type inferred by the t-round model, It represents the prior probability that the variable pair e output in the tth iteration belongs to type λ, and ξ is the smoothing coefficient.

6. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 5 is characterized in that: Step S3 specifically includes: Step S3-1, classifying the variables in the multi-scale features into target variables and explanatory variables; Step S3-2: randomly extract data with replacement from the multi-scale features multiple times, and determine and record the causal relationship type for each pair of variables in each extracted data; Step S3-3: Count the frequency of occurrence of each edge type in the causal graph in the recorded content Determine the confidence probability of the second edge type as Among them, B represents the total number of random draws with replacement, Indicates that the obedience parameter is Beta distribution.

7. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 6 is characterized in that: In step S3-1, the state of charge and battery health status are used as target variables, and the average cell voltage, the maximum cell temperature, the battery pack charge and discharge current, the insulation resistance, the ambient temperature and the rate of change of the state of charge are used as explanatory variables.

8. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 7 is characterized in that: Step S4 specifically includes: Step S4-1: Calculate the fusion probability according to the following expression: in, represents the fusion probability, γ is the fusion weight; Step S4-2: For each edge e, according to the fusion probability Select the edge type with the largest confidence as the edge type of e, and obtain the fused causal graph A * .

9. The dual-path heterogeneous temporal causal analysis method based on representation embedding and time-delay pairing according to claim 8, characterized in that: Step S5 specifically includes: Step S5-1, all directed edges are Sort from high to low and record the sequence; Step S5-2: Define the initial directed edge set E according to the recorded sequence * And initialize the causal network G as a network without edges; Step S5-3: E * The edges in are added to the current causal network, and topological sorting is used to determine whether there is a loop in the current causal network. If a loop is found in the network, the edge with the lowest confidence is removed first in the edge set that generates the loop. Step S5-4, return to step S5-3 until there is no loop in the network, and the edge set E' is obtained DAG That is, it corresponds to the directed causal network without circularity and with the highest total confidence; Step S5-5, according to the acyclic directed edge set E' DAG Determine the acyclic power battery state causal graph A *,DAG .

Images