A marine lithium battery BMS control method and system based on deep learning

Abstract

The present invention relates to the technical field of lithium battery management systems, and discloses a marine lithium battery BMS control method and system based on deep learning. The marine lithium battery BMS control method based on deep learning comprises: constructing a three-layer causal structure of an environment layer, a system layer, and a fault layer; learning a stable causal structure using an invariant risk minimization algorithm with multi-environment data and gradient penalty; evaluating the strength of causal edges based on mutual information theory and extracting a stable causal skeleton; constructing a cross-environment enhanced data set through counterfactual data generation technology; and constructing a meta-learning framework to enable the model to quickly adapt to new environments. The present invention solves the generalization problem of fault diagnosis for marine lithium batteries in different sea environments around the world, improves the accuracy of fault root cause identification, reduces the misdiagnosis rate, and achieves rapid adaptation to new environments.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery management systems, and more specifically, to a marine lithium battery BMS control method and system based on deep learning. Background Art

[0002] Deployment of marine lithium-ion battery management systems (BMS) in diverse ocean regions around the world presents significant challenges. The same battery system, operating in diverse waters—such as Arctic, tropical, and temperate zones—is affected by multiple factors, including sea conditions, climate, and mission missions, resulting in varying operating characteristics and failure modes. These differences not only impact the basic battery characteristics but also affect the accuracy and reliability of fault diagnosis.

[0003] Traditional marine lithium battery fault diagnosis methods rely primarily on correlation analysis or simple rule-based reasoning, making it difficult to effectively distinguish between the true cause of a fault and its symptoms. Diagnostic accuracy decreases significantly when multiple factors, including environmental conditions, operating modes, and maintenance history, are combined. For example, in tropical high-temperature environments, battery derating protection mechanisms may be misdiagnosed as cell failures; while in polar low-temperature environments, a normal decrease in charge and discharge efficiency may be mistakenly identified as accelerated aging. This environmental dependency results in inconsistent performance of the same diagnostic algorithm in different deployment environments, severely impacting the reliability of the battery management system and the accuracy of maintenance decisions.

[0004] In the existing technology, there is still a lack of methods that can effectively solve the generalization problem of marine lithium battery fault diagnosis in different sea environments, especially in identifying the root cause of the fault and quickly adapting to the new environment. Summary of the Invention

[0005] The present invention provides a deep learning-based marine lithium battery BMS control method and system to solve the technical problem of the generalization difficulty of fault diagnosis of marine lithium batteries in different environments in the related art.

[0006] The present invention provides a deep learning-based marine lithium battery management system control method, comprising the following steps:

[0007] Construct a three-layer causal structure including the environment layer, system layer and fault layer to form a hierarchical causal diagram;

[0008] Based on the hierarchical causal graph, we use the invariant risk minimization algorithm with multi-environment data and gradient penalty to learn stable causal structures.

[0009] For the learned stable causal structure, the strength of the causal edge is evaluated based on mutual information theory, and a stable causal skeleton is extracted;

[0010] Based on the extracted stable causal skeleton, we construct an enhanced dataset across environments through counterfactual data generation technology;

[0011] By utilizing the constructed cross-environment enhanced dataset, a meta-learning framework is constructed to enable the model to quickly adapt to the new environment.

[0012] In a preferred embodiment, the step of constructing a three-layer causal structure including an environment layer, a system layer, and a fault layer includes:

[0013] Construct an environmental layer variable set to collect variables related to the ship's operating environment;

[0014] Construct a system-level variable set to collect operating parameters of the marine lithium battery system;

[0015] Construct a set of fault layer variables to define the possible fault types of marine lithium batteries;

[0016] Based on domain knowledge and historical data, causal relationships between variables are established to form a hierarchical causal diagram.

[0017] In a preferred embodiment, the environmental layer variables include sea area type, sea state, climate conditions, navigation mission and ship load status.

[0018] In a preferred embodiment, the system-level variables include battery pack voltage, cell voltage, charge and discharge current, surface temperature, internal temperature, state of charge, and state of health.

[0019] In a preferred embodiment, the fault layer variables include overcharge, overdischarge, internal short circuit, external short circuit, thermal runaway, insulation failure and balance circuit failure.

[0020] In a preferred embodiment, the step of learning a stable causal structure includes:

[0021] Construct multi-environment datasets;

[0022] Construct a deep learning model in which a subset of parameters related to system-level variables are represented as system-level parameters;

[0023] Design a gradient-penalty-based invariant risk minimization learning objective;

[0024] Use the optimization algorithm to optimize the objective function and obtain the model parameters that meet the invariance.

[0025] In a preferred embodiment, the step of evaluating the strength of the causal edge comprises:

[0026] Calculate mutual information between variables;

[0027] Calculate conditional mutual information;

[0028] Compute the strength of causal edges;

[0029] Based on the causal strength threshold, causal edges with strength greater than the threshold are extracted to form a core causal skeleton.

[0030] In a preferred embodiment, the step of constructing a cross-environment enhanced dataset includes:

[0031] Construct an intervention model based on the core causal framework;

[0032] Construct a counterfactual data generator to generate counterfactual samples;

[0033] For each environment, select representative samples and generate a set of counterfactual samples;

[0034] The original dataset is merged with the counterfactual dataset to form an augmented dataset.

[0035] In a preferred embodiment, the step of constructing a meta-learning framework includes:

[0036] Treat fault diagnosis in each environment as a task and define a set of meta-learning tasks;

[0037] Based on the trained model parameters, they are used as the initial parameters for meta-learning;

[0038] For new environments, use a small amount of support set data to update parameters through gradient descent;

[0039] The adapted model performance is evaluated based on the query set and the initial parameters are updated through meta-optimization.

[0040] In a preferred embodiment, a deep learning-based marine lithium battery BMS control system is used to execute a deep learning-based marine lithium battery BMS control method, including:

[0041] A multi-level causal structure building module is used to construct a three-level causal structure consisting of the environment layer, the system layer, and the fault layer;

[0042] Invariant causal structure learning module, used to learn stable causal structures;

[0043] Causal strength assessment and core skeleton extraction module, used to assess the strength of causal edges and extract stable causal skeletons;

[0044] Multi-source environment data enhancement module, used to build enhanced datasets across environments;

[0045] The meta-learning fast adaptation module is used to enable the model to quickly adapt to new environments.

[0046] The beneficial effects of the present invention are:

[0047] Through hierarchical causal structure and invariant causal learning, the accuracy of fault root cause identification is improved and the misdiagnosis rate is reduced;

[0048] Leverage multi-source environment data augmentation and meta-learning for rapid adaptation;

[0049] Simplified reasoning based on a core causal framework shortens diagnostic time;

[0050] Optimize maintenance strategies through causal reasoning to reduce maintenance costs;

[0051] The stable performance of the same model in different sea areas, seasons and operating modes is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a deep learning-based marine lithium battery BMS control method of the present invention;

[0053] Figure 2 is a detailed flow chart of forming a hierarchical cause-effect graph of the present invention;

[0054] Figure 3 It is a detailed flow chart of the present invention for learning a stable causal structure;

[0055] Figure 4 is a detailed flow chart of the present invention for extracting a stable causal skeleton;

[0056] Figure 5 is a detailed flow chart of the present invention for constructing a cross-environment enhanced dataset;

[0057] Figure 6 It is a detailed flowchart of constructing the meta-learning framework of the present invention. DETAILED DESCRIPTION

[0058] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0059] At least one embodiment of the present invention discloses a marine lithium battery BMS control method based on deep learning, such as Figures 1 to 6 As shown, the following steps are included:

[0060] Step 1: Construct a three-layer causal structure including the environment layer, system layer, and fault layer to form a hierarchical causal graph;

[0061] The specific steps include:

[0062] Step 1.1, construction of environment layer variable set;

[0063] Collect variables related to the ship's operating environment, including sea area type, sea state, climate conditions, navigation tasks, and ship load status, to form an environmental layer variable set ,in, 、 、 Respectively represent 、 、 environment variables, Indicates the total number of environment variables.

[0064] In the marine lithium battery management system, these environmental variables are obtained through the ship's navigation status monitoring system, meteorological system and mission management system, and transmitted to the data preprocessing module in real time.

[0065] Step 1.2, construction of system-level variable set;

[0066] Collect the operating parameters of the marine lithium battery system, including battery pack voltage, single cell voltage, charge and discharge current, surface temperature, internal temperature, state of charge (SOC), state of health (SOH) and other variables to form a system-level variable set ,in, 、 、 Respectively represent 、 、 system variables, Indicates the total number of system variables.

[0067] These variables are captured by the sensor network of the lithium battery management system and transmitted to the data acquisition module via the CAN bus or other ship communication protocols.

[0068] Step 1.3, construction of the fault layer variable set;

[0069] Define possible fault types of marine lithium batteries, including overcharge, overdischarge, internal short circuit, external short circuit, thermal runaway, insulation failure, balance circuit failure and other variables to form a fault layer variable set ,in, 、 、 Respectively represent 、 、 Fault type variables, Indicates the total number of fault types.

[0070] This application subdivides fault layer variables according to severity and impact range. For example, thermal runaway faults are subdivided into three stages: early temperature anomaly, hotspot formation, and heat diffusion, to facilitate accurate location of the root cause of the fault.

[0071] Step 1.4, construction of hierarchical causal diagram;

[0072] Based on domain knowledge and historical data, causal relationships between variables are established to form a hierarchical causal graph:

[0073] ;

[0074] in, represents a hierarchical causal graph, Represents a set of directed edges between variables, representing causal relationships, Represents a collection of environment layer variables, system layer variables, and fault layer variables.

[0075] Each edge Representing variables to variables The causal impact includes the causal relationship from environment to system, system to fault, environment to fault, and between variables within a layer.

[0076] In the specific implementation, this application uses an adjacency matrix or an adjacency list to represent the causal graph structure, supporting efficient graph operations and traversal.

[0077] Step 2: Based on the hierarchical causal graph, we use the invariant risk minimization algorithm with multi-environment data and gradient penalty to learn a stable causal structure.

[0078] The specific steps include:

[0079] Step 2.1, multi-environment dataset construction;

[0080] Collect battery operation data from ships in different sea areas, seasons, and mission types to form a multi-environmental data set ,in 、 、 Respectively represent 、 、 Data sets in this environment, Indicates the total number of environments.

[0081] For the application scenarios of marine lithium batteries, this application collects data from three typical environments: Arctic routes, equatorial routes, and temperate routes. Each environment contains at least 500 hours of continuous operation data, covering normal operating conditions and various fault conditions.

[0082] Step 2.2, system layer parameter representation;

[0083] Construct a deep learning model where the subset of parameters related to system-level variables is represented as , which is the parameter part of the model responsible for processing system-level variables.

[0084] This application uses a hybrid architecture of a multi-layer perceptron (MLP) and a long short-term memory (LSTM). The MLP processes static features, while the LSTM captures temporal features. System-level parameters, including network weights and biases related to battery status, are optimized using a backpropagation algorithm.

[0085] Step 2.3, construct the objective function of constant risk minimization;

[0086] Design a gradient-penalty-based constant risk minimization learning objective:

[0087] ;

[0088] in, For all environments The loss function Sum and find the model parameters that minimize this total loss ; Represents the model in the environment The loss function under Represents the gradient of the loss function with respect to the system layer parameters; is the gradient penalty weight coefficient, which is used to balance the empirical risk and invariance constraints; Represents all parameters of the model; Represents the set of all environments; represents the sum over all environments; In the application of marine lithium battery fault diagnosis, this application determines the optimal A value of 0.1 effectively improves the generalization ability while ensuring the accuracy of the model.

[0089] Step 2.4, invariant causal structure optimization;

[0090] Use optimization algorithms such as stochastic gradient descent to optimize the invariant risk minimization objective function and obtain model parameters that meet the invariance Furthermore, by penalizing risk gradient variations across environments, the model automatically ignores environment-specific spurious correlations and preserves the causal structure that remains constant across environments. During optimization, this application uses the Adam optimizer, sets an initial learning rate of 0.001, and combines this with a learning rate decay strategy to achieve model convergence within 40 iterations.

[0091] Step 3: Based on the mutual information theory, the strength of the causal edge is evaluated for the learned stable causal structure, and a stable causal skeleton is extracted;

[0092] The specific steps include:

[0093] Step 3.1, calculation of mutual information between variables;

[0094] For hierarchical causal diagrams Each pair of variables in , calculate their mutual information :

[0095] ;

[0096] in, Representing variables and Mutual information between and Represent variables respectively and The specific value of and Respectively represent variables and The sum of all possible values; Representing variables The value is And variable The value is The joint probability distribution of and Represent variables respectively The value is and variables The value is The marginal probability distribution of Represents the natural logarithm function.

[0097] Mutual information measures the interdependence between two variables, with larger values ​​indicating stronger dependence.

[0098] In practical applications, this application uses the k-nearest neighbor estimation method to calculate the mutual information of high-dimensional continuous variables. This method shows good stability when the number of samples is limited.

[0099] For battery temperature and SOC variables, this application observed that they exhibited high mutual information values ​​in different environments, indicating that the impact of temperature on battery status is a key factor that remains unchanged in the environment.

[0100] Step 3.2, conditional mutual information calculation;

[0101] For each pair of variables and its parent node set , calculate the conditional mutual information :

[0102] ;

[0103] in, Indicates that in a given variable Under the condition of and The conditional mutual information between 、 and Represent variables respectively 、 and The specific value of 、 and Respectively represent variables 、 and The sum of all possible values; Representing variables The value is ,variable The value is And variable The value is The joint probability distribution of Indicates in variables The value is Under the condition of The value is And variable The value is The conditional probability of and Represented in variables The value is Under the condition of The value is and variables The value is The conditional probability of represents the natural logarithm function; Representing variables The parent node set of Representing variables Exclude variables from the parent node collection After the collection.

[0104] Conditional mutual information measures the value of In this case, the variable and interdependence between them.

[0105] This application applies conditional mutual information analysis in thermal runaway fault diagnosis and finds that battery thermal runaway fault is directly correlated with internal temperature, while the correlation with ambient temperature is significantly weakened after controlling the internal temperature, indicating that internal temperature is the direct cause of thermal runaway, while ambient temperature is an indirect influencing factor.

[0106] Step 3.3, causal strength calculation;

[0107] Using mutual information and conditional mutual information, calculate causal edges Strength:

[0108] ;

[0109] in, Indicates that the variable to variables The strength of the causal edge; Representing variables and The mutual information between them is used to quantify the degree of mutual dependence between two variables; Indicates in variables All parent nodes (except Take the maximum value among the two; Representing variables The set of all parent nodes of ; Representing variables Exclude variables from the parent node collection The collection after Indicates that in a given variable under conditions and The conditional mutual information between .

[0110] This formula quantifies the variables For variables The direct causal influence strength is calculated by subtracting the mutual information conditional on other parent nodes to eliminate the influence of indirect associations.

[0111] The causal strength calculation module implemented in this application can automatically identify key causal relationships. For example, in the low-temperature Arctic environment, the causal strength between the battery internal resistance and SOC is significantly higher than that in the normal temperature environment, indicating that the battery internal resistance under low-temperature conditions is a key factor affecting the battery status and should be monitored closely.

[0112] Step 3.4, core causal skeleton extraction;

[0113] Based on causal strength threshold , extract causal edges with strength greater than the threshold to form a core causal skeleton:

[0114] ;

[0115] ;

[0116] in, Represents the causal strength threshold, which is used to screen important causal relationships. Only causal edges with strength exceeding this threshold will be retained; Represents the core causal skeleton, which is a simplified version of the original causal diagram, containing only the most important causal relationships; Represents a variable set, including all variables in the environment layer, system layer, and fault layer; Represents the core causal edge set, which is the original edge set A subset containing only those with causal strength greater than the threshold edge; Indicates that the variable to variables The directed edges of ,represent causal relationships; Indicates that the variable to variables causal strength.

[0117] Therefore, the core causal skeleton contains causal relationships that are stable across environments, providing a structural basis for environmentally robust fault diagnosis.

[0118] In the marine lithium battery management system, this application determines the optimal The threshold is dynamically adjusted according to the complexity of the environment. A higher threshold is used in a complex environment to retain the most stable causal relationship, and a lower threshold is used in a simple environment to retain more information.

[0119] Step 4: Based on the extracted stable causal skeleton, we construct an enhanced dataset across environments through counterfactual data generation technology.

[0120] The specific steps include:

[0121] Step 4.1, construction of intervention model;

[0122] Based on the core causal skeleton extracted in step 3 , build the intervention model:

[0123] ;

[0124] in, represents the target variable; represents the intervention variable; represents the intervention value; Represents the intervention operator, which means the variable Perform human intervention and set to value The operation is different from simple conditional probability; Indicates that an intervention operation is being performed After that, the variable The probability distribution of is used to evaluate the effect of the intervention.

[0125] In a marine lithium battery management system, the intervention model implemented in this application can simulate the impact of different environmental conditions on battery performance. For example, by simulating the intervention of "disconnecting the causal edge from ambient temperature to SOC", the direct impact of temperature changes on the battery status is evaluated, providing a reliable counterfactual basis for fault diagnosis.

[0126] Step 4.2, counterfactual data generator implementation;

[0127] Building a counterfactual data generator , generate counterfactual samples ,in is the original data, For environment variables, For intervention indicators, specify the intervention variable and target value, Represents a counterfactual data generator function, which is used to generate counterfactual samples that meet specific intervention conditions. represents the generated counterfactual samples, which represent simulated data under specific intervention conditions.

[0128] The counterfactual data generator of this application adopts a variational autoencoder (VAE) architecture, which includes three parts: encoder, latent space operation and decoder.

[0129] The encoder maps the raw data into a latent space that adjusts the variable distribution according to the intervention indicator, and the decoder generates counterfactual samples that comply with the physical constraints.

[0130] For marine lithium batteries, this application specifically adds electrochemical consistency constraints to ensure that the generated counterfactual samples conform to the basic physical and chemical laws of batteries.

[0131] Step 4.3, multi-environment counterfactual sample generation;

[0132] For each environment , select representative samples and generate a counterfactual sample set:

[0133] ;

[0134] in, Indicates a specific environment; Represents the set of all environments; Represents the environment The counterfactual sample set under ; Indicates that the environment The next generated Counterfactual input samples, including various characteristic parameters of the battery system; Indicates that the environment The next generated The labels corresponding to the counterfactual samples; Represents the environment The total number of counterfactual samples generated under

[0135] The multi-environment counterfactual sample generation scheme designed in this application pays special attention to rare cross-environment failure scenarios. For example, for the rare high-temperature runaway failures in polar environments, by performing intervention transformations based on normal temperature environment data, high-temperature runaway samples that conform to polar characteristics are generated to make up for the lack of actual data.

[0136] Step 4.4, enhanced dataset construction;

[0137] The original data set With the counterfactual dataset Merge to form an enhanced dataset , used for model training to improve robustness to environmental changes.

[0138] in, represents the original dataset; Indicates a specific environment The counterfactual dataset generated below; represents the counterfactual dataset of all environments merged, i.e. , contains counterfactual samples generated in all environments; The final enhanced dataset is the union of the original dataset and the counterfactual dataset. Represents the set of all environments; Represents a set union operation.

[0139] The example of enhancing the dataset construction implemented in this application on ships on the Arctic route shows that by introducing counterfactual samples, the fault detection accuracy under extreme low temperature conditions is improved from the original 78% to 92%. In particular, for the short-circuit fault type with sparse data, the recognition accuracy is improved by more than 30%.

[0140] Step 5: Use the constructed cross-environment enhanced dataset to build a meta-learning framework to enable the model to quickly adapt to the new environment;

[0141] The specific steps include:

[0142] Step 5.1, meta-learning task definition;

[0143] Each environment The following fault diagnosis is considered as a task , define the set of meta-learning tasks , where each task contains a support set and queryset , 、 、 Represents the first 、 、 Specific tasks, Indicates the total number of tasks.

[0144] In the case of marine lithium battery systems, this application defines fault diagnosis problems in different environments, such as polar, equatorial, and temperate routes, as independent tasks. Using a meta-learning framework, this application learns the common knowledge between tasks while preserving the characteristics of each task. For each task, this application selects 100 samples as the support set for within-task adaptation and 1000 samples as the query set for evaluating the adaptation effect.

[0145] Step 5.2, model initialization;

[0146] Model parameters trained based on steps 2 and 4 , as the initial parameters for meta-learning, the parameters already contain causal structure information that is invariant across environments.

[0147] The model initialization strategy designed in this application incorporates the concept of transfer learning, leveraging knowledge extracted from invariant causal structures as prior information for model initialization to accelerate the meta-learning process. Experiments show that compared to random initialization, this initialization strategy based on invariant causal structures can shorten meta-learning convergence time by 50% and achieve superior final performance.

[0148] Step 5.3, inner loop adaptation;

[0149] For new environments , using a small amount of support data , update the parameters by gradient descent:

[0150] ;

[0151] in, Indicates a new environment; Indicates a new environment The support set data under is a small set of samples used for rapid model adaptation; represents the initial parameters of the model; Indicates the updated parameters after the inner loop adaptation, which have been adapted to the new environment Adjustments were made; Represents the inner loop learning rate, controls the step size of parameter updates, and affects the speed and stability of model adaptation; Indicates about parameters Gradient operator, used to calculate the partial derivative of the loss function with respect to the parameters; represents the meta-learning loss function.

[0152] In practical applications of marine lithium battery systems, the internal loop adaptation mechanism implemented in this application can rapidly adjust model parameters to adapt to new environmental characteristics using only 5 to 10 new environmental samples. For example, when a ship enters tropical waters from temperate waters, the system only needs to collect 10 minutes of battery operating data to complete model adaptation, significantly improving the system's environmental adaptability.

[0153] Step 5.4, outer loop optimization;

[0154] Based on query set Evaluate the performance of the adapted model and update the initial parameters through meta-optimization :

[0155] ;

[0156] in, Indicates a new environment The query set data below is used to evaluate the performance of the model after adaptation; represents the initial parameters of the meta-learning model; Represents the model parameters after inner loop adaptation; Represents the outer loop learning rate, which controls the step size of the outer loop optimization; Indicates about parameters The gradient operator is used to calculate the gradient of the loss function with respect to the parameters; represents the meta-learning loss function used to evaluate the performance of the model on the query set; Indicates a parameter update operation, which assigns the calculation result on the right to the variable on the left.

[0157] By repeatedly iterating inner-loop adaptation and outer-loop optimization, the model's ability to quickly adapt to new environments is improved.

[0158] This application uses a second-order optimization algorithm in the outer loop optimization phase, taking into account the gradient effect of parameter adaptation on the initial parameters, thereby improving the stability of meta-learning. In an actual deployment case on a large container ship, after completing data collection and meta-learning training on three different routes, the system was able to use only 10% of the data when entering a new route, rapidly improving fault diagnosis accuracy to over 95%, comparable to expert performance.

[0159] Application examples of this implementation:

[0160] The implementation methods of this application have been tested in actual applications on a number of large commercial ships. A typical application example is introduced below.

[0161] A large container ship, primarily engaged in ocean shipping from Asia to Europe, faced the following challenges with its lithium battery management system:

[0162] The route spans tropical, temperate and subarctic waters, with ambient temperatures ranging from -10°C to 40°C and sea conditions ranging from calm waters to force 9 winds and waves;

[0163] The battery system needs to operate stably under different ship load conditions (low-speed sailing, full-speed sailing, and port standby);

[0164] The fault diagnosis accuracy of traditional BMS systems is less than 70%, and false alarms are frequent, especially during cross-sea navigation.

[0165] Faults manifest differently in different sea areas, making it difficult for maintenance personnel to determine the root cause of the fault.

[0166] Implementation process example:

[0167] Multi-level causal structure construction: Based on the existing sensor network of the ship, 38 environmental layer variables, 52 system layer variables and 24 fault layer variables were collected.

[0168] Environmental layer variables include location information (latitude and longitude), sea conditions (wind force, wave height), ambient temperature and humidity, ship load and other data;

[0169] System-level variables include total battery pack voltage, 96 cell voltages, charge and discharge currents, temperatures at each point, SOC, SOH, etc.

[0170] Fault layer variables include various fault types and their subtypes such as overcharge, overdischarge, internal short circuit, external short circuit, and thermal runaway.

[0171] By analyzing the ship's operation logs and fault records over the past three years and combining them with expert knowledge, a hierarchical causal graph with 114 nodes and 376 edges was initially constructed, establishing a complex association network between environmental conditions, battery status, and fault types.

[0172] Invariant causal structure learning: We collected battery operation data from three typical routes (Asia to Europe North, Asia to Europe South, and Asia to the Middle East). Each route contains approximately 750 hours of continuous recording, covering various environmental conditions.

[0173] The causal structure that is stable across environments is extracted through the invariant risk minimization algorithm, and the gradient penalty weight coefficient λ is set to 0.1.

[0174] After 40 iterations of training, the model successfully identified the causal relationships of battery failures that remained stable under environmental changes.

[0175] Causal Strength Assessment and Core Skeleton Extraction: Using mutual information theory to assess causal edge strength, we identified 46 key causal edges with significant strength and stability across different environments. These causal edges constitute the core causal skeleton, representing only 12.2% of the original causal edges but containing over 85% of the critical diagnostic information. For example, we found that the causal relationship strength between abnormal internal battery temperature and thermal runaway failure remained above 0.72 across all environments, making it one of the most stable diagnostic indicators.

[0176] Multi-source environmental data enhancement: For rare high-temperature failures on Northern European routes and rare low-temperature failures on Middle Eastern routes, approximately 2,000 high-quality virtual samples were synthesized using counterfactual data generation technology. These samples maintain the rationality of battery physical properties while simulating possible failure modes in extreme environments, effectively expanding the model's training data.

[0177] Meta-learning Rapid Adaptation Mechanism: To address situations where ships enter new routes, a meta-learning framework was built to enable rapid adaptation to new environments. During its initial entry into the Baltic Sea route, the system collected only 12 hours of operational data (approximately 240 samples). Through meta-learning, the system quickly adapted to the characteristics of the new environment, increasing fault diagnosis accuracy from an initial 76% to 94%.

[0178] Technical effect verification:

[0179] Improved diagnostic accuracy: The proposed method significantly improved the accuracy of fault root cause identification in various route tests. The following is a comparison with traditional diagnostic methods:

[0180]

[0181] Improved environmental adaptability: The method of this application has significantly improved its adaptability to different environmental changes, especially when encountering extreme environmental changes. The following are the test results of the adaptability to environmental changes:

[0182]

[0183] This application example fully demonstrates the effectiveness and superiority of the deep learning-based marine lithium-ion battery management system control method proposed in this application in actual ship operating environments. In particular, in the complex and changing marine environment, this method can accurately identify the root cause of faults and quickly adapt to environmental changes, thereby improving the safety and reliability of the ship's battery system.

[0184] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A marine lithium battery BMS control method based on deep learning, characterized in that: The following steps are involved: A three-layer causal structure consisting of an environmental layer, a system layer, and a fault layer is constructed to form a hierarchical causal graph. The environmental layer variables include sea area type, sea state, climate conditions, navigation mission, and ship load status; the system layer variables include battery pack voltage, cell voltage, charge and discharge current, surface temperature, internal temperature, state of charge, and health status; and the fault layer variables include overcharge, overdischarge, internal short circuit, external short circuit, thermal runaway, insulation failure, and balancing circuit failure. Based on the hierarchical causal graph, the invariant risk minimization algorithm using multi-environment data and gradient penalty is used to learn a stable causal structure. The invariant risk minimization objective function is: ； in, For all environments The loss function Sum and find the model parameters that minimize this total loss ; Represents the model in the environment The loss function under Represents the gradient of the loss function with respect to the system layer parameters; is the gradient penalty weight coefficient, which is used to balance the empirical risk and invariance constraints; Represents all parameters of the model; Represents the set of all environments; represents the sum over all environments; Represents the square of the norm, which is used to measure the magnitude of the gradient; For the stable causal structure learned, based on the mutual information theory, the strength of the causal edge is evaluated, the stable causal skeleton is extracted, and the causal edge is calculated. The strength is: ； in, Indicates that the variable to variables The strength of the causal edge; Representing variables and The mutual information between them is used to quantify the degree of mutual dependence between two variables; Indicates in variables Take the maximum value among all parent nodes; Representing variables The set of all parent nodes of ; Representing variables Exclude variables from the parent node collection The collection after Indicates that in a given variable under conditions and The conditional mutual information between Based on the extracted stable causal skeleton, we construct an enhanced dataset across environments through counterfactual data generation technology, counterfactual data generator Generating counterfactual examples ,in is the original data, For environment variables, For intervention indicators, specify the intervention variable and target value, represents the counterfactual data generator function, represents the generated counterfactual sample; By using the constructed cross-environment enhanced dataset, a meta-learning framework is constructed to enable the model to quickly adapt to the new environment. The inner loop adaptation updates the parameters through gradient descent: ； in, Indicates a new environment; Indicates a new environment The support set data under ; represents the initial parameters of the model; Indicates the updated parameters after the inner loop adaptation, which have been adapted to the new environment Adjustments were made; Represents the inner loop learning rate, controls the step size of parameter updates, and affects the speed and stability of model adaptation; Indicates about parameters Gradient operator, used to calculate the partial derivative of the loss function with respect to the parameters; represents the meta-learning loss function; The outer loop optimization updates the initial parameters through meta-optimization: ； in, Indicates a new environment The query set data below; represents the initial parameters of the meta-learning model; Represents the model parameters after inner loop adaptation; Represents the outer loop learning rate, which controls the step size of the outer loop optimization; Indicates about parameters The gradient operator is used to calculate the gradient of the loss function with respect to the parameters; represents the meta-learning loss function used to evaluate the performance of the model on the query set; Indicates a parameter update operation, which assigns the calculation result on the right to the variable on the left.

2. A marine lithium battery BMS control method based on deep learning according to claim 1, characterized in that: The steps of constructing a three-layer causal structure including an environment layer, a system layer, and a fault layer include: Construct an environmental layer variable set to collect variables related to the ship's operating environment; Construct a system-level variable set to collect operating parameters of the marine lithium battery system; Construct a set of fault layer variables to define the possible fault types of marine lithium batteries; Based on domain knowledge and historical data, causal relationships between variables are established to form a hierarchical causal diagram.

3. A marine lithium battery BMS control method based on deep learning according to claim 1, characterized in that: The steps of learning a stable causal structure include: Construct multi-environment datasets; Construct a deep learning model in which a subset of parameters related to system-level variables are represented as system-level parameters; Design a gradient-penalty-based invariant risk minimization learning objective; Use the optimization algorithm to optimize the objective function and obtain the model parameters that meet the invariance.

4. A marine lithium battery BMS control method based on deep learning according to claim 1, characterized in that: The step of evaluating the strength of the causal edge comprises: Calculate mutual information between variables; Calculate conditional mutual information; Compute the strength of causal edges; Based on the causal strength threshold, causal edges with strength greater than the threshold are extracted to form a core causal skeleton.

5. A marine lithium battery BMS control method based on deep learning according to claim 1, characterized in that: The steps of constructing a cross-environment enhanced dataset include: Construct an intervention model based on the core causal framework; Construct a counterfactual data generator to generate counterfactual samples; For each environment, select representative samples and generate a set of counterfactual samples; The original dataset is merged with the counterfactual dataset to form an augmented dataset.

6. A marine lithium battery BMS control method based on deep learning according to claim 1, characterized in that: The steps of constructing the meta-learning framework include: Treat fault diagnosis in each environment as a task and define a set of meta-learning tasks; Based on the trained model parameters, they are used as the initial parameters for meta-learning; For new environments, use a small amount of support set data to update parameters through gradient descent; The adapted model performance is evaluated based on the query set and the initial parameters are updated through meta-optimization.

7. A deep learning-based marine lithium battery BMS control system, used to execute a deep learning-based marine lithium battery BMS control method according to any one of claims 1 to 6, characterized in that: include: A multi-level causal structure building module is used to construct a three-level causal structure consisting of the environment layer, the system layer, and the fault layer; Invariant causal structure learning module, used to learn stable causal structures; Causal strength assessment and core skeleton extraction module, used to assess the strength of causal edges and extract stable causal skeletons; Multi-source environment data enhancement module, used to build enhanced datasets across environments; The meta-learning fast adaptation module is used to enable the model to quickly adapt to new environments.

Images