Box-type power supply battery management system for ship

Abstract

The invention provides a marine box type power supply battery management system, which belongs to the technical field of marine power supplies, and comprises a control chip, a battery pack management unit (BMU), a battery cluster management unit (BCMU), a battery stack management unit (BSMU) and a related monitoring device, multi-dimensional battery state data is acquired by adopting a time window segmented acquisition method, a battery working state feature vector is extracted by utilizing a battery state evaluation function, battery parameters are calculated based on a depth time sequence analysis model, error compensation parameters are generated, and battery output errors are corrected in real time through a voltage and temperature combined compensation mechanism. And calling an equalization strategy optimization function to determine a multi-stage equalization control scheme, and predicting and compensating an error generated in an equalization operation instruction sequence execution process by using a pre-trained battery parameter error model, thereby solving the core technical problem that the marine box-type power supply battery management system cannot adapt to battery parameter changes in a complex water area environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine power supplies, and in particular relates to a marine box-type power supply battery management system. Background Art

[0002] Marine box-type power battery management systems are widely used in various ship power systems. Traditional technologies primarily rely on voltage monitoring, temperature monitoring, and ampere-hour integration to calculate the remaining battery charge and health status. These systems also maintain battery pack consistency through passive or simple active balancing. These systems typically employ single-parameter monitoring and fixed algorithm models for battery status assessment and management.

[0003] However, traditional technologies exhibit significant limitations in complex aquatic environments. Due to the large temperature fluctuations, high humidity, and frequent load changes found in some aquatic environments, single-parameter monitoring and fixed algorithms struggle to accurately reflect the battery's true state. The linear models and empirical formulas used in traditional systems lack adaptability across diverse operating conditions, leading to accumulated errors in battery parameter calculations and inefficient balancing strategies, making them unable to meet the high reliability requirements of long-duration endurance.

[0004] Especially in aquatic environments, battery parameters exhibit nonlinear variations due to multiple factors. Conventional technologies struggle to adapt to these complex variations, resulting in inaccurate battery status assessments. This, in turn, impacts balancing control and charge / discharge management, reducing battery life and system reliability. In other words, existing technologies suffer from the inability of marine box-type power battery management systems to adapt to the changing battery parameters in complex aquatic environments, leading to inaccurate battery status assessments. Summary of the Invention

[0005] In view of this, the present invention provides a marine box-type power battery management system, which can solve the technical problem in the prior art that the marine box-type power battery management system cannot adapt to changes in battery parameters in complex water environments, resulting in inaccurate battery status assessment.

[0006] The present invention is implemented as follows: The present invention provides a marine box-type power battery management system, including a control chip, a battery pack management unit BMU, a battery cluster management unit BCMU, a battery stack management unit BSMU and related monitoring devices. The control chip is provided with a battery management control module, adopts a time window segmented acquisition method to obtain multi-dimensional battery status data, uses a battery status evaluation function to extract the battery working status feature vector, calculates battery parameters based on a deep time series analysis model and generates error compensation parameters, corrects battery output errors in real time through a voltage-temperature joint compensation mechanism, calls a balancing strategy optimization function to determine a multi-level balancing control scheme, and uses a pre-trained battery parameter error model to predict and compensate for errors generated during the execution of a balancing operation instruction sequence.

[0007] Among them, the battery management control module is used to perform the following steps: adopting a time window segmented acquisition method to obtain multi-dimensional battery status data, using a battery status evaluation function to obtain a battery working status feature vector, calculating battery parameters based on a deep time series analysis model, correcting battery parameters through a voltage-temperature joint compensation mechanism, calling a balancing strategy optimization function to determine a balancing control scheme, using a pre-trained battery parameter error model for prediction and compensation, and updating the pre-trained battery parameter error model according to a battery abnormality evaluation function.

[0008] Among them, the multi-dimensional battery status data includes voltage time series data, temperature time series data, current time series data and internal resistance time series data, which are obtained using the relevant monitoring device. Specifically, the voltage time series data is provided by the voltage monitoring device, the temperature time series data is provided by the temperature monitoring device, the current time series data is provided by the current metering device, and the internal resistance time series data is calculated by dividing the voltage time series data by the current time series data.

[0009] The battery operating state feature vector includes battery health features, temperature distribution features, voltage consistency features, and load response features. The battery operating state feature vector is used for subsequent battery parameter calculation and error compensation.

[0010] Among them, the battery status evaluation function extracts the fluctuation spectrum characteristics of voltage time series data through Fourier transform, evaluates the consistency level of the battery pack through entropy calculation, identifies potential hotspots through temperature gradient analysis, and analyzes the dynamic performance of the battery through load step response.

[0011] The corrected battery parameters include a corrected battery actual remaining capacity SOC value and a corrected state of health SOH value, and the corrected battery parameters are used for balancing strategy optimization and charge and discharge control.

[0012] Among them, the balancing strategy optimization function is used to determine the optimal balancing control strategy according to the battery status. The input includes the voltage value of each battery cell, the corrected actual remaining battery SOC value, the temperature distribution status of the battery pack, the current charge and discharge status, and the ambient temperature value. The output is a balancing operation instruction sequence including the balancing target battery cell identifier, balancing current size, balancing duration, and balancing priority.

[0013] The balancing strategy optimization function takes into account both balancing speed and balancing efficiency through a multi-objective optimization algorithm, calculates the optimal balancing path based on the voltage difference matrix, and adjusts the balancing intensity in combination with the temperature influencing factor.

[0014] Among them, the battery abnormality evaluation function is used to calculate the abnormality degree of the battery operating state. The input includes the battery working state feature vector, the historical battery working state feature vector database, the corrected battery parameters, the voltage fluctuation rate and the temperature distribution uniformity. The output is the battery abnormality index.

[0015] Among them, the battery abnormality evaluation function determines the degree of abnormality by calculating the Mahalanobis distance between the current battery working state feature vector and the standard working condition feature vector in the historical battery working state feature vector database, and performs a comprehensive score based on the voltage fluctuation rate and temperature distribution uniformity to generate the final battery abnormality index.

[0016] The pre-trained battery parameter error model is a deep residual attention network model. The specific structure of the deep residual attention network model is a deep neural network consisting of an input layer, multiple residual-connected convolutional layers, a multi-head self-attention mechanism layer, a fully connected layer, and an output layer. The number of multi-head attention mechanism heads in the deep residual attention network model is determined based on the number of battery cells, the sampling frequency, and the data dimension. The specific calculation formula is that the number of attention heads is equal to the smaller value of the number of battery cells divided by 8, rounded up, or the square root of the data dimension, rounded down.

[0017] The steps for establishing a training dataset for the pre-trained battery parameter error model include collecting battery system operating data over a long period of time, classifying the data according to different operating conditions, using a sliding window method to divide the time series data into fixed-length sample sequences, calculating the error as the model training target, and generating more diverse training samples through data enhancement technology. The steps for training the pre-trained battery parameter error model include initializing network parameters, setting the batch size, using an optimizer and setting the initial learning rate, using a cosine annealing learning rate scheduling strategy, using mean square error as the loss function for model training, regularly evaluating model performance on a validation set, and using an early stopping strategy to select the optimal model parameters.

[0018] The system adjusts the parameters of the multi-head attention mechanism in the pre-trained battery parameter error model in real time according to the battery abnormality index. The adjustment method is to dynamically adjust the attention weight distribution based on the battery abnormality index, adjust the attention intensity of the corresponding channel according to the degradation degree of different battery cells, and update the model parameters through incremental learning.

[0019] The time window segmented acquisition method refers to an acquisition method that divides continuously acquired data into multiple data windows according to a fixed time period for processing. The window length is determined to be 10 seconds to 60 seconds based on the battery response characteristics.

[0020] The present invention adopts a time window segmented acquisition method to obtain multi-dimensional battery status data, and combines a deep time series analysis model with a voltage and temperature joint compensation mechanism to achieve high-precision calculation of battery parameters and real-time error correction, greatly improving the accuracy of battery status assessment. By introducing a battery status assessment function to extract key feature vectors, the system can comprehensively capture the changing laws of the battery working state, and combine the deep residual attention network model for parameter prediction and compensation, effectively solving the problem of large parameter calculation deviations in traditional technologies under complex water environments. In particular, the synergistic effect of the battery anomaly assessment function and the balancing strategy optimization function enables the system to dynamically adjust the balancing strategy according to the real-time status, significantly improving the balancing efficiency and battery pack consistency. The present invention successfully solves the core technical problem that the ship's box-type power battery management system cannot adapt to changes in battery parameters in complex water environments through a method that combines multi-dimensional data analysis with deep learning, providing a more reliable, efficient and adaptable battery management solution for ship power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the method of the present invention.

[0022] Figure 2 This is a layered architecture diagram of the marine box-type power battery management system in Example 2.

[0023] Figure 3 This is a structural diagram of the battery module and bidirectional active balancing circuit in Example 2. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] The present invention provides a marine box-type power supply battery management system, comprising a control chip, a battery pack management unit BMU, a battery cluster management unit BCMU, a battery stack management unit BSMU, a CAN communication device, an Ethernet communication device, a battery balancing device, a temperature monitoring device, a voltage monitoring device, a relay control device, a current metering device, a remote communication device and a power management device, wherein the control chip is electrically connected to the BMU, the BCMU, the BSMU, the CAN communication device, the Ethernet communication device, the battery balancing device, the temperature monitoring device, the voltage monitoring device, the relay control device, the current metering device, the remote communication device and the power management device respectively, a battery management control module is provided in the control chip, the BMU is used to collect cell voltage and temperature data and perform battery balancing operations, and the BCMU is used to collect cell voltage and temperature data and perform battery balancing operations. The system collects lower-level BMU data and calculates battery cluster parameters. The BSMU is used to manage the entire battery system and communicate with the remote management platform. The CAN communication device is used for communication between the BMU and the BCMU. The Ethernet communication device is used for communication between the BCMU and the BSMU. The battery balancing device is used to implement bidirectional active balancing control of the battery. The temperature monitoring device is used to collect the battery pack temperature and ambient temperature. The voltage monitoring device is used to collect the battery cell voltage and total voltage. The relay control device is used to control the charge and discharge circuit. The current metering device is used to measure the battery charge and discharge current. The remote communication device is used for communication between the system and the remote monitoring platform. The power management device is used to monitor the battery voltage and control the system power supply. The sampling frequency of the voltage monitoring device is 5 times per second, the sampling frequency of the temperature monitoring device is 1 time per minute, and the sampling frequency of the current metering device is 1 time per second.

[0026] like Figure 1 As shown, the battery management control module is used to perform the following steps:

[0027] S01. Acquire multi-dimensional battery status data using a time window segmented acquisition method, wherein the multi-dimensional battery status data includes voltage time series data, temperature time series data, current time series data, and internal resistance time series data, wherein the voltage time series data is provided by the voltage monitoring device, the temperature time series data is provided by the temperature monitoring device, the current time series data is provided by the current metering device, and the internal resistance time series data is calculated by dividing the voltage time series data by the current time series data;

[0028] S02. Using a battery state evaluation function to perform feature extraction on the multidimensional battery state data to obtain a battery operating state feature vector, wherein the battery operating state feature vector includes battery health features, temperature distribution features, voltage consistency features, and load response features. The battery operating state feature vector is used for subsequent battery parameter calculation and error compensation;

[0029] S03. Calculating the actual remaining capacity (SOC) and state of health (SOH) of the battery based on a deep time series analysis model, and generating a predicted battery output error compensation parameter, wherein the predicted battery output error compensation parameter is used to correct the actual remaining capacity (SOC) and state of health (SOH) of the battery;

[0030] S04. Correcting the battery output error in real time through a voltage-temperature combined compensation mechanism, and outputting corrected battery parameters. The corrected battery parameters include a corrected battery actual remaining capacity (SOC) value and a corrected state of health (SOH) value. The corrected battery parameters are used for balancing strategy optimization and charge and discharge control.

[0031] S05. Calling a balancing strategy optimization function to process the corrected battery parameters, determining a multi-level balancing control scheme, and generating a balancing operation instruction sequence, which is sent to the battery balancing device to perform a balancing operation;

[0032] S06. Predict and compensate errors generated during the execution of the balancing operation instruction sequence using a pre-trained battery parameter error model to generate a balancing correction instruction, which is sent to the battery balancing device for correction.

[0033] S07. Calculate a battery abnormality index according to a battery abnormality evaluation function, and update the weight of the pre-trained battery parameter error model in real time based on the battery abnormality index and the balancing operation result to optimize the accuracy of the pre-trained battery parameter error model. The updated pre-trained battery parameter error model is used for error prediction and compensation in the next cycle.

[0034] Among them, the battery status evaluation function is used to extract key battery working status characteristics from the time window data. The input includes the time window length, the voltage time series data, the temperature time series data, the current time series data and the internal resistance time series data. The output is the battery working status feature vector including the battery health characteristics, the temperature distribution characteristics, the voltage consistency characteristics and the load response characteristics; the battery status evaluation function extracts the fluctuation spectrum characteristics of the voltage time series data through Fourier transform, evaluates the consistency level of the battery pack through entropy calculation, identifies potential hotspots through temperature gradient analysis, and analyzes the battery dynamic performance through load step response.

[0035] The balancing strategy optimization function is used to determine the optimal balancing control strategy according to the battery status. The input includes the voltage value of each battery cell in the voltage time series data, the actual remaining battery capacity SOC value after correction, the battery pack temperature distribution state formed by the temperature time series data, the current charge and discharge state represented by the current time series data, and the ambient temperature value in the temperature time series data. The output is the balancing operation instruction sequence including the balancing target battery cell identifier, balancing current size, balancing duration and balancing priority. The balancing strategy optimization function takes into account both balancing speed and balancing efficiency through a multi-objective optimization algorithm, calculates the optimal balancing path based on the voltage difference matrix, and adjusts the balancing intensity in combination with the temperature influencing factor.

[0036] The battery abnormality evaluation function is used to calculate the abnormality degree of the battery operating state. The input includes the battery working state feature vector, the historical battery working state feature vector database, the corrected battery parameters, the voltage fluctuation rate in the voltage time series data, and the temperature distribution uniformity in the temperature time series data. The output is the battery abnormality index. The battery abnormality evaluation function determines the abnormality degree by calculating the Mahalanobis distance between the current battery working state feature vector and the standard operating condition feature vector in the historical battery working state feature vector database, and performs a comprehensive score based on the voltage fluctuation rate and temperature distribution uniformity to generate a final battery abnormality index.

[0037] The pre-trained battery parameter error model is a deep residual attention network model. The specific structure of the deep residual attention network model is a deep neural network consisting of an input layer, a plurality of residual-connected convolutional layers, a multi-head self-attention mechanism layer, a fully connected layer and an output layer. The input layer receives the multi-dimensional battery status data in the time window, the convolutional layer extracts the timing features, the multi-head self-attention mechanism layer captures the long-term dependency between the data, the fully connected layer fuses the features to generate the prediction results, and the residual connection structure ensures the stability of the deep network training; the number of multi-head attention mechanism heads in the deep residual attention network model is determined according to the number of battery cells, the sampling frequency and the data dimension. The specific calculation formula is that the number of attention heads is equal to the smaller value of the number of battery cells divided by 8 rounded up and the square root of the data dimension rounded down.

[0038] The steps of establishing a training data set for the pre-trained battery parameter error model specifically include collecting battery system operation data over a long period of time, including the voltage time series data, the temperature time series data, the current time series data, and the battery's actual remaining power SOC value annotation data, classifying the data according to different working conditions, including steady-state discharge conditions, pulse discharge conditions, steady-state charging conditions, and balancing conditions, using a sliding window method to divide the time series data into sample sequences of fixed length, calculating the error between the theoretical value and the measured value of the battery parameter in each sample sequence as the model training target, generating more diverse training samples through data enhancement technology, including adding Gaussian noise, time scale stretching, and random masking, and finally dividing all samples into training set, validation set, and test set ratios of 7:1:2.

[0039] The steps of training the pre-trained battery parameter error model specifically include initializing the network parameters using the He initialization method, setting the batch size to 128 samples per batch, using the Adam optimizer and setting the initial learning rate to 0.001, adopting the cosine annealing learning rate scheduling strategy in the training process, and using the mean square error as the loss function for model training. The model performance is evaluated on the validation set every 10 training cycles. When the validation set loss does not decrease for 5 consecutive cycles, the early stopping strategy is adopted, and finally the model parameters with the best validation set performance are selected as the pre-training model.

[0040] The system adjusts the multi-head attention mechanism parameters in the pre-trained battery parameter error model in real time according to the battery abnormality index. The specific adjustment method is to dynamically adjust the attention weight distribution based on the battery abnormality index so that the model pays more attention to abnormal points. At the same time, it adjusts the attention intensity of the corresponding channel according to the degradation degree of different battery cells, and updates the model parameters through incremental learning to ensure model adaptability.

[0041] The time window segmented acquisition method refers to an acquisition method that divides continuously acquired data into multiple data windows according to a fixed time period for processing, and the window length is determined to be 10 seconds to 60 seconds based on the battery response characteristics.

[0042] The battery health characteristics refer to a set of characteristic parameters that characterize the battery's health status, including the internal resistance growth rate, capacity decay rate, and charge / discharge efficiency change rate. The temperature distribution characteristics refer to a set of characteristic parameters that characterize the battery's temperature distribution, including the highest temperature point, lowest temperature point, temperature mean, and temperature variance. The voltage consistency characteristics refer to a set of characteristic parameters that characterize the degree of voltage consistency among cells within a battery pack, including the voltage range, voltage standard deviation, and voltage distribution entropy. The load response characteristics refer to a set of characteristic parameters that characterize the battery's ability to respond to load changes, including voltage response rate, current tracking accuracy, and power fluctuation suppression capability. The battery's actual remaining capacity (SOC) value is the percentage of the battery's current remaining capacity to its rated capacity, calculated using the combined voltage method and the ampere-hour integration method. The state of health (SOH) value is the percentage of the battery's current capacity to its initial capacity, characterizing the battery's aging. The voltage fluctuation rate refers to the ratio of the battery cell voltage change per unit time to the average voltage, characterizing voltage stability. The temperature distribution uniformity refers to the degree of temperature consistency among the various temperature measurement points within the battery pack, calculated using the temperature standard deviation and temperature range.

[0043] The specific implementation of the above steps is described in detail below.

[0044] The specific implementation method of step S01 is to obtain multi-dimensional battery status data using a time window segmented acquisition method. This step is achieved by the following method: first, a fixed time window length is set to 30 seconds, and the system obtains voltage time series data by sampling 5 times per second through a voltage monitoring device, obtains temperature time series data by sampling once per minute through a temperature monitoring device, and obtains current time series data by sampling once per second through a current metering device. For the acquisition of internal resistance time series data, the system performs local average filtering on the collected voltage time series data to remove high-frequency noise, and applies median filtering to the current time series data to remove mutation interference, and then divides the processed voltage time series data by the processed current time series data to calculate the internal resistance time series data. The system uses a data synchronization mechanism to ensure that data with different sampling frequencies are aligned, and supplements the low-frequency sampling data through an interpolation algorithm, and finally forms a multi-dimensional battery status data matrix with a unified time scale. The main purpose of this step is to obtain multi-dimensional time series data that comprehensively reflects the working status of the battery, laying a data foundation for subsequent battery feature extraction and status evaluation.

[0045] The specific implementation of step S02 is to use the battery status evaluation function to extract features from the multi-dimensional battery status data. The function first applies a fast Fourier transform to the voltage time series data, extracts the spectral features in the 0.01Hz to 0.5Hz frequency band, and calculates the spectral energy distribution as the battery dynamic response feature; then calculates the spatial distribution features of the temperature time series data, and identifies potential hot spots through a temperature gradient analysis algorithm. The temperature gradient threshold is set to 2.5℃ / cm2 , exceeding this threshold is determined to be a hotspot; then Shannon entropy is applied to calculate the entropy value of the battery pack voltage data distribution to evaluate the battery consistency level. The entropy value threshold is 1.8. A value lower than this indicates good consistency; at the same time, the battery load response characteristics are extracted, and the voltage response rate is calculated by analyzing the slope of the voltage response curve at the moment of current mutation. The response time threshold is set to 50ms. For the battery health characteristics, the internal resistance growth rate is calculated by comparing the current internal resistance with the benchmark internal resistance to calculate the growth percentage. The internal resistance growth rate threshold is 20%. Exceeding this value indicates that the battery health has deteriorated. The system combines all the extracted features to form a battery working status feature vector with a length of 64, which is used for subsequent battery parameter calculation and error compensation. The main purpose of this step is to extract key features from massive time series data, reduce computational complexity while retaining key information on battery status.

[0046] Step S03 is implemented by calculating the battery's actual remaining charge (SOC) and state of health (SOH) values based on a deep time series analysis model. This model utilizes a long short-term memory (LSTM) network architecture, comprising three layers of LSTM units, each containing 128 hidden nodes. The input is the sequence of battery operating state feature vectors obtained in step S02. The model is divided into two parallel branches: one for SOC estimation and the other for SOH estimation. The SOC estimation branch connects to a two-layer fully connected network with 64 and 1 nodes, respectively, at its terminal. It uses a ReLU activation function and outputs a normalized SOC value. The SOH estimation branch also connects to a two-layer fully connected network with the same number of nodes, but uses a Sigmoid activation function to output the SOH value. The model also outputs predicted battery output error compensation parameters, including SOC correction coefficients and SOH correction coefficients, which are learned based on deviations from the actual observed value. The initial SOC estimation utilizes a weighted fusion of the open-circuit voltage method and the ampere-hour integration method. The weighting is dynamically adjusted to 0.7 for SOC < 20% or SOC > 80%, and 0.3 for all other conditions. The main purpose of this step is to use deep learning technology to achieve accurate estimation of key battery parameters and generate error compensation parameters to improve estimation accuracy.

[0047] Step S04 specifically implements real-time correction of battery output errors through a combined voltage-temperature compensation mechanism. This mechanism first constructs a temperature compensation function, modeling the relationship between temperature and battery internal resistance based on the Arrhenius equation. Within the 0°C to 45°C range, each 10°C temperature difference results in approximately a 15% change in internal resistance. Next, a voltage compensation function is established based on a battery polarization effect model, taking into account the voltage response characteristics at different depths of discharge. An exponential compensation coefficient is employed to enhance the correction effect, particularly in the SOC range below 20%. The system inputs the SOC and SOH values calculated in step S03, along with the predicted battery output error compensation parameters, into the compensation module. Using bilinear interpolation, the system searches a two-dimensional temperature-voltage compensation table for the most appropriate correction coefficient. The correction coefficient is then applied to the original estimated values to generate a corrected battery actual remaining capacity (SOC) value and a corrected state-of-health (SOH) value. The compensation strength dynamically adjusts based on the battery's operating state. Under high-current discharge conditions (current greater than 0.5C), the compensation coefficient increases by 30%, while in abnormal temperature ranges (below 5°C or above 40°C), the compensation coefficient increases by 50%. The main purpose of this step is to consider the interaction between temperature and voltage, perform multi-dimensional error compensation on battery parameter estimation, and improve the calculation accuracy of the system.

[0048] The specific implementation of step S05 involves calling a balancing strategy optimization function to process the corrected battery parameters and determine a multi-level balancing control scheme. This function uses a decision-making method that combines the analytic hierarchy process (AHP) with fuzzy comprehensive evaluation. First, a battery state scoring matrix is constructed based on the corrected SOC values and cell voltages. The voltage difference threshold is set to 50mV, exceeding which balancing is considered necessary. The balancing intensity is then adjusted based on the temperature distribution. The temperature difference threshold is set to 5°C; for cells with high temperatures exceeding this threshold, the balancing current is reduced. A multi-objective optimization problem is then constructed. The objective function includes minimizing balancing time and minimizing energy loss. Constraints include an upper limit of 0.2C for balancing current, a lower limit of 10 minutes for balancing time, and an upper limit of 40°C for temperature. The system uses a non-dominated sorting genetic algorithm (NSGA-II) to solve this multi-objective optimization problem. The population size is set to 100, the number of generations is set to 50, the crossover probability is set to 0.85, and the mutation probability is set to 0.1. The Pareto optimal solution set is ultimately obtained. The system selects the balancing strategy that best suits the current operating conditions from the optimal solution set and generates a balancing operation instruction sequence that includes the identification of the target battery cell, the balancing current (ranging from 0.05C to 0.2C), the balancing duration (ranging from 10 minutes to 120 minutes), and the balancing priority (levels 1 to 5). The main purpose of this step is to make intelligent balancing decisions based on the battery status, balancing balancing speed and energy efficiency, and extending the battery system life.

[0049] The specific implementation of step S06 utilizes a pre-trained battery parameter error model to predict and compensate for errors generated during the execution of the balancing instruction sequence. This model, based on a deep residual attention network structure, monitors battery responses in real time during the balancing process. First, battery state change data is acquired before and after the balancing operation, including voltage distribution, SOC difference, and temperature change before and after balancing. This data is then compared with the ideal balancing effect predicted by the model to calculate an error vector. The system applies a Kalman filter algorithm to the error vector for noise suppression. The filtered error is input into the multi-head self-attention layer of the pre-trained model. The attention weight threshold is set to 0.6; abnormal features exceeding this threshold receive a higher attention weight. The model outputs balancing error compensation parameters, including a balancing time adjustment factor (ranging from 0.8 to 1.2), a balancing current adjustment factor (ranging from 0.7 to 1.3), and a balancing target adjustment parameter. Based on these parameters, the system generates balancing correction instructions, such as extending the balancing time, adjusting the balancing current, or changing the balancing target cell, and sends them to the battery balancing device for correction. The error threshold is set to 5% of the voltage correction target; exceeding this threshold triggers a forced correction mechanism. The main purpose of this step is to make up for the deviation between theoretical equilibrium and actual effect through the prediction model and realize dynamic optimization of the equilibrium process.

[0050] The specific implementation of step S07 involves calculating a battery abnormality index based on a battery abnormality evaluation function and updating the weights of the pre-trained battery parameter error model in real time based on this index and the results of the balancing operation. This function first extracts a set of standard operating condition feature vectors from a database of historical battery operating state feature vectors as a reference baseline. It then calculates the Mahalanobis distance between the current battery operating state feature vector and the reference baseline, while also considering two auxiliary indicators: voltage fluctuation rate and temperature distribution uniformity. The voltage fluctuation threshold is set at 3% / min, and temperature distribution uniformity is represented by the temperature standard deviation, with a threshold set at 3°C. The system combines the Mahalanobis distance and the auxiliary indicators using a weighted average method to generate a comprehensive battery abnormality index. The abnormality index ranges from 0 to 1, with a threshold set at 0.7; exceeding this value indicates an abnormal state. Based on the abnormality index, the system updates the weights of the pre-trained battery parameter error model using stochastic gradient descent. The learning rate is initially set at 0.001. When the abnormality index exceeds 0.7, the learning rate is automatically increased to 0.005 to accelerate model adjustment. The system uses gradient clipping to limit the range of weight updates, with a clipping threshold of 5.0 to prevent over-adjustment of the model. The update primarily targets the weights of the attention heads associated with detected anomalies in the multi-head self-attention layer, allowing the model to focus more on anomalous patterns. The system uses an annealing strategy to periodically reduce the learning rate, with a learning rate decay factor of 0.9 after every 500 weight updates to ensure the stability of model updates. The primary purpose of this step is to optimize the performance of the prediction model through a real-time feedback mechanism and improve its adaptability to abnormal conditions.

[0051] Regarding the hardware implementation, the system adopts a layered distributed architecture. The control chip utilizes a 32-bit ARM Cortex-M4 series processor with a main frequency of 120MHz, built-in 512KB Flash memory and 128KB RAM, and an integrated hardware floating-point unit to accelerate complex calculations. Each BMU module manages up to 16 battery cells. It uses a high-precision AD7280 analog-to-digital converter for voltage measurement, achieving an accuracy of ±0.25mV, and an NTC thermistor combined with a low-noise signal conditioning circuit for temperature measurement, achieving an accuracy of ±0.5°C. The BCMU module utilizes an STM32F407 processor, managing up to eight subordinate BMUs. An integrated CAN controller enables communication with the BMUs at a rate of 500kbps. The BSMU module utilizes a higher-performance embedded computing platform, equipped with an ARM Cortex-A72 quad-core processor running at a main frequency of 1.5GHz and 8GB of internal storage for running deep learning models and storing historical data. The battery balancing device utilizes switched capacitor technology for bidirectional active balancing, with a maximum balancing current of 2A and a balancing efficiency of up to 95%. The current metering device utilizes a Hall-effect current sensor and a gain-adjustable amplifier, offering a measurement range of -100A to 100A with an accuracy of 0.5% of full scale. The Ethernet communication device supports 100Mbps communication speeds and utilizes an industrial-grade protection design to ensure reliable communication. The remote communication device integrates a 4G module and a satellite communication backup system, supporting multi-path redundant communication. The power management device utilizes a low-power design, with static power consumption of less than 10mW and supports a battery voltage range of 9V to 36V.

[0052] Regarding the detailed structure of the battery parameter error model, it uses a deep residual attention network architecture, consisting of an input layer, four residual-connected convolutional blocks, two multi-head self-attention layers, two fully connected layers, and an output layer. The input layer accepts a three-dimensional tensor of shape [batch size × time window length × feature dimension]. Each residual convolutional block consists of two one-dimensional convolutional layers with kernel sizes of 3 and 5, respectively. The number of channels increases layer by layer to 64, 128, 256, and 512. Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The residual connections use 1×1 convolutions to achieve dimensionality matching. The first multi-head self-attention layer has 4 heads, and the second has 8, with an attention dimension of 64. The residual connection structure and layer normalization ensure stable gradient propagation. The fully connected layers have 256 and 128 nodes, respectively. Dropout is used to prevent overfitting, with a dropout rate of 0.3. The output layer has the same number of nodes as the predicted parameter dimensions and provides predicted values for the error compensation parameters. The steps for establishing the training data set include: collecting at least 6 months of historical battery operation data, covering various operating conditions such as charging, discharging, static, and balancing; dividing the data into standard operating condition categories based on factors such as charge and discharge current, duration, and ambient temperature, with each category containing at least 500 samples; using a 60-second sliding window with a step size of 10 seconds to slice the time series data into fixed-length samples; calculating the error between the theoretical SOC value and the measured SOC value in each sample as the model learning target; expanding the training set by adding Gaussian noise (standard deviation is 3% of the original signal), time scale stretching (factor range is 0.8 to 1.2), random masking (maximum masking ratio is 20%) and other data enhancement techniques; finally dividing the training set, validation set, and test set in a ratio of 7:1:2 to ensure the generalization ability of the model.

[0053] Specifically, the principle of the present invention is: the core technical principle of the present invention lies in constructing an adaptive battery management framework based on multi-dimensional data analysis and deep learning, which realizes high-precision evaluation and control of battery status through time window segmented acquisition, feature extraction, deep model prediction and real-time compensation.

[0054] First, the system uses a time window segmented acquisition method to acquire multi-dimensional battery status data such as voltage, temperature, current, and internal resistance to form a time series data stream. Unlike traditional single-point data acquisition, this method can capture the dynamic change characteristics of the battery and provide a more comprehensive data foundation for subsequent analysis. The battery status evaluation function is used to extract features from these time series data, including Fourier transform to extract voltage fluctuation spectrum features and entropy calculation to evaluate battery pack consistency. The system obtains battery operating status feature vectors including battery health features, temperature distribution features, voltage consistency features, and load response features. These features can fully reflect the working status of the battery.

[0055] Secondly, the system implements battery parameter prediction and error compensation based on a deep residual attention network model. This model captures long-term dependencies between data using a multi-head self-attention mechanism. Combined with a residual connection structure to ensure deep network training stability, it can learn battery parameter variation patterns from complex time series data. The model is trained using a rich dataset encompassing diverse operating conditions. The error between theoretical and measured battery parameter values is used as a training target, resulting in excellent generalization and accuracy.

[0056] In addition, the system incorporates a combined voltage and temperature compensation mechanism to correct battery output errors in real time. A multi-level balancing control scheme is dynamically determined using a balancing strategy optimization function. A battery anomaly assessment function is used to assess the degree of battery status anomalies in real time, further optimizing the pre-trained model. This closed-loop control mechanism enables the system to continuously adapt to environmental and load changes, maintaining highly accurate battery status assessment and management.

[0057] The organic combination of these technical principles enables the present invention to accurately adapt to changes in battery parameters in complex water environments, solves the problem of inaccurate battery status assessment faced by traditional technologies, and achieves a technological breakthrough in the marine box-type power battery management system.

[0058] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0059] The specific implementation of step S01 is to use a time window segmented acquisition method to obtain multi-dimensional battery status data. The method first sets a fixed time window length of T w The value range is 10 to 60 seconds, preferably 30 seconds. The system uses a voltage monitoring device to sample at a frequency of f v =5Hz to obtain the voltage time series data matrix V, and the temperature monitoring device is used to sample at a frequency of Get the temperature time series data matrix T, and use the current metering device to sample at a frequency f i =1Hz to obtain the current time series data matrix I. For the acquisition of internal resistance time series data, the system performs local average filtering on the collected voltage time series data to remove high-frequency noise. The filtered voltage data V f The calculation formula is:

[0060]

[0061] Where V f (k) is the voltage value of the Kth sampling point after filtering; V(k+j) is the voltage value of the k+jth sampling point of the original sampling; n is the half-width of the sliding average window, which is 2.

[0062] At the same time, median filtering is applied to the current time series data to remove sudden interference. The filtered current data I f The calculation formula is:

[0063] I f (k)=median{I(km), I(k-m+1),..., I(k+m)};

[0064] Where, I f (k) is the current value of the kth sampling point after filtering; I(km) to I(k+m) are the current values of the original sampling points from km to k+m; median{} represents the median operation; m is the half-width of the median filter window, which is 1.

[0065] Then, the processed voltage time series data is divided by the processed current time series data to calculate the internal resistance time series data matrix R. The calculation formula is:

[0066]

[0067] Where R(k) is the estimated internal resistance at the kth sampling point, in ohms (Ω).

[0068] The system uses a data synchronization mechanism to ensure that data with different sampling frequencies are aligned, and uses a linear interpolation algorithm to supplement low-frequency sampling data. The interpolation formula is:

[0069]

[0070] Where, T interp (t) is the interpolated temperature at time t; T(t i ) and T(t i+1 ) are respectively the time t i and t i+1 Sampling temperature value; t i ≤t <t i+1 .

[0071] Finally, a multi-dimensional battery state data matrix X with a unified time scale is formed, which is expressed as:

[0072] X=[V f , T interp , I f , R];

[0073] Where X is the multidimensional battery status data matrix; V f is the voltage data vector after filtering; T interp is the interpolated temperature data vector; I fis the filtered current data vector; R is the calculated internal resistance data vector. The main purpose of this step is to obtain multi-dimensional time series data that fully reflects the battery's operating status, laying the data foundation for subsequent battery feature extraction and status assessment.

[0074] The specific implementation of step S02 is to use the battery status evaluation function to extract features from the multi-dimensional battery status data to obtain the battery operating status feature vector. The function first applies a fast Fourier transform to the voltage time series data to extract the spectrum features. The voltage spectrum transformation formula is:

[0075]

[0076] Where, F v (f) is the frequency spectrum of voltage time series data; V f (n) is the filtered voltage sample value; N is the number of sampling points; f is the frequency; j is the imaginary unit; and e is the base of the natural logarithm, approximately 2.71828. The system extracts the spectral energy distribution within the 0.01Hz to 0.5Hz frequency band as the battery's dynamic response signature.

[0077] Then, the spatial distribution characteristics of the temperature time series data are calculated, and potential hot spots are identified through the temperature gradient analysis algorithm. The temperature gradient calculation formula is:

[0078]

[0079] Where, is the temperature gradient vector at point (x, y); T(x, y) is the temperature value at point (x, y); and Represent the partial derivatives of temperature in the x and y directions respectively. The temperature gradient modulus calculation formula is:

[0080]

[0081] The temperature gradient threshold is set to G T =2.5℃ / cm 2 ,when When , the point (x, y) is determined to be a hotspot.

[0082] Then, Shannon entropy is applied to calculate the entropy value of the battery pack voltage data distribution to evaluate the battery consistency level. The voltage distribution entropy value calculation formula is:

[0083]

[0084] Where H(V) is the entropy of the battery pack voltage distribution; n is the number of battery cells; v i is the voltage value of the ith battery cell; p(v i ) is the voltage value v iThe probability density of is obtained by kernel density estimation method. The entropy threshold is H thre =1.8, a value lower than this indicates good consistency.

[0085] At the same time, the battery load response characteristics are extracted and the voltage response rate is calculated by analyzing the slope of the voltage response curve at the moment of current mutation. The voltage response rate calculation formula is:

[0086]

[0087] Where S v is the voltage response rate, in V / s; V(t1) and V(t2) are the voltage values at time t1 and t2 respectively; t1 is the moment when the current mutation begins, and t2 is the moment when the voltage reaches 90% of the stable value. The response time threshold is set to T resp =50ms. If the value exceeds this value, the response speed is slow.

[0088] Based on the battery health characteristics, the internal resistance growth rate is calculated by comparing the current internal resistance with the baseline internal resistance. The internal resistance growth rate calculation formula is:

[0089]

[0090] Where R growth is the internal resistance growth rate; R current is the internal resistance value currently measured; R base The internal resistance value of the battery when it leaves the factory. The internal resistance growth rate threshold is R thre = 20%, exceeding this value indicates that the battery health is deteriorating.

[0091] The system combines all the extracted features to form a battery working status feature vector F with a length of 64. state , expressed as:

[0092] F state =[F health , F temp , F volt , F resp ];

[0093] Where, F health is the battery health feature vector, with a length of 16; F temp is the temperature distribution characteristic subvector, with a length of 16; F volt is the voltage consistency feature vector, with a length of 16; F resp is the load response feature subvector with a length of 16. The main purpose of this step is to extract key features from massive time series data, reduce computational complexity, and retain key information about the battery status.

[0094] The specific implementation of step S03 is to calculate the battery's actual remaining capacity (SOC) value and state of health (SOH) value based on a deep time series analysis model. This model uses a long short-term memory (LSTM) network structure, consisting of three layers of LSTM units, each containing 128 hidden nodes. The input is the battery operating state feature vector sequence obtained in step S02. The core calculation formula of the LSTM unit is as follows:

[0095] Input gate i t Calculation formula:

[0096] i t =σ(W i ·[h t-1 , x t ]+b i );

[0097] Forget Gate f t Calculation formula:

[0098] f t =σ(W f ·[h t-1 , x t ]+b f );

[0099] Output gate o t Calculation formula:

[0100] o t =σ(W o ·[h t-1 , x t ]+b o );

[0101] Candidate memory units Calculation formula:

[0102]

[0103] Memory unit C t Update formula:

[0104]

[0105] Hidden state h t Calculation formula:

[0106] h t =o t ⊙tanh(C t );

[0107] Where i t 、f t 、o t are the activation values of the input gate, forget gate, and output gate respectively; is the candidate memory cell state; C t is the current memory unit state; h t is the current hidden state; x t is the input feature vector at the current moment; h t-1 is the hidden state at the previous moment; C t-1 is the memory unit state at the previous moment; W i 、W f 、W o 、W c is the corresponding weight matrix; b i 、b f 、b o 、b c is the corresponding bias vector; σ is the sigmoid activation function; tanh is the hyperbolic tangent activation function; ⊙ represents the element-wise multiplication operation.

[0108] The model is divided into two parallel branches, one for SOC estimation and the other for SOH estimation. The SOC estimation branch is connected to a two-layer fully connected network at the end, with 64 and 1 nodes respectively. It uses the ReLU activation function and outputs the normalized SOC value. The calculation formula of the fully connected layer is:

[0109] z l+1 =ReLU(W l ·z l +b l );

[0110] Where z l+1 is the output of the l+1th layer; z l is the output of the lth layer; W l is the weight matrix of the lth layer; b l is the bias vector of the lth layer; ReLU is the linear rectification function, defined as ReLU(x)=max(0,x).

[0111] The SOH estimation branch also connects to a two-layer fully connected network with the same number of nodes, but uses the Sigmoid activation function to output the SOH value. The Sigmoid function formula is:

[0112]

[0113] Where x is the input value and e is the base of the natural logarithm.

[0114] The model also outputs the predicted battery output error compensation parameters, including the SOC correction coefficient α SOC and SOH correction factor α SOH , which is obtained by learning the deviation from the actual observation value. The error loss function is defined as:

[0115] L=λ1·MSE(SOC pred , SOC true )+λ2·MSE(SOH pred , SOH true );

[0116] Where L is the total loss; MSE is the mean square error function; SOC pred and SOC true They are the predicted SOC value and the real labeled SOC value; SOH pred and SOH true are the predicted SOH value and the true labeled SOH value respectively; λ1 and λ2 are weight coefficients, both of which are 0.5.

[0117] The initial SOC estimation uses a weighted fusion of the voltage open circuit method and the ampere-hour integration method, with the weights adjusted dynamically. The fused SOC calculation formula is:

[0118] SOC fused =w ocv ·SOC ocv +(1-w ocv )·SOC ah ;

[0119] Where, SOC fused is the estimated SOC value after fusion; SOC ocv is the SOC value estimated based on the open circuit voltage method; SOC ah is the SOC value estimated based on the ampere-hour integration method; w ocv The voltage open-circuit method weight is 0.7 when SOC < 20% or SOC > 80%, and 0.3 in other states. The main purpose of this step is to use deep learning technology to accurately estimate key battery parameters and generate error compensation parameters to improve estimation accuracy.

[0120] The specific implementation of step S04 is to correct the battery output error in real time through a voltage-temperature combined compensation mechanism. This mechanism first constructs a temperature compensation function and establishes a temperature-battery internal resistance relationship model based on the Arrhenius equation. The temperature compensation coefficient is calculated as follows:

[0121]

[0122] Where k T is the temperature compensation coefficient; E a is the activation energy, with a typical value of 15000-25000 J / mol; R is the gas constant, 8.314 J / (mol·K); T ref is the reference temperature, which is 298.15K (25℃); T act is the actual operating temperature in K.

[0123] Then, a voltage compensation function is established based on the battery polarization effect model, taking into account the voltage response characteristics at different discharge depths. The voltage compensation coefficient calculation formula is:

[0124] k V =1+β·exp(-γ·SOC);

[0125] Where k V is the voltage compensation coefficient; β is the basic compensation coefficient, ranging from 0.05 to 0.15; γ is the exponential adjustment parameter, ranging from 3 to 5; SOC is the current remaining battery capacity percentage.

[0126] The system inputs the SOC value and SOH value calculated in step S03 and the predicted battery output error compensation parameter into the compensation module, and uses bilinear interpolation to find the most suitable correction coefficient in the two-dimensional temperature voltage compensation table. The bilinear interpolation calculation formula is:

[0127]

[0128] Where f(x, y) is the interpolation result at point (x, y); Q 11 , Q 21 , Q 12 , Q 22 are four adjacent sampling points; f(Q ij ) is point Q ij The function value at ; (x1, y1) and (x2, y2) are the diagonal coordinates of the rectangular area containing (x, y).

[0129] Apply the correction coefficient to adjust the original estimated value to generate the corrected battery actual remaining capacity SOC value and the corrected health state SOH value. The correction formula is:

[0130] SOC corr =SOC pred ·k T ·k V α SOC ;

[0131] SOH corr =SOH pred ·k T α SOH ;

[0132] Where, SOC corr is the corrected SOC value; SOC pred SOC value predicted by the model; SOH corr is the corrected SOH value; SOH pred is the SOH value predicted by the model; k Tis the temperature compensation coefficient; k V is the voltage compensation coefficient; α SOC and α SOH are the SOC and SOH error compensation parameters output by the model respectively.

[0133] The compensation intensity is dynamically adjusted according to the battery working status. The adjustment coefficient is calculated as follows:

[0134] α adj =1+0.3·δ(I>0.5C)+0.5·δ(T<5℃∨T>40℃);

[0135] Where, α adj is the compensation intensity adjustment coefficient; δ() is an indicator function, which takes the value 1 when the conditions in the brackets are met and 0 otherwise; I is the current value; C is the rated battery capacity; and T is the battery temperature. The main purpose of this step is to consider the interaction between temperature and voltage, perform multi-dimensional error compensation on battery parameter estimation, and improve system calculation accuracy.

[0136] The specific implementation of step S05 is to call the balancing strategy optimization function to process the corrected battery parameters and determine the multi-level balancing control scheme. This function uses a decision-making method that combines the hierarchical analysis method with fuzzy comprehensive evaluation. First, the battery state score matrix M is constructed based on the corrected SOC value and cell voltage. The calculation formula is:

[0137]

[0138] Where M ij is the imbalance score between the i-th battery cell and the j-th battery cell; V i and V j are the voltage values of the i-th and j-th battery cells respectively; V max To design the maximum single cell voltage difference, the typical value is 0.3V; SOC i and SOC j are the SOC values of the i-th and j-th battery cells respectively; SOC max To design the maximum SOC difference, the typical value is 0.1; w1 and w2 are weight coefficients, which are 0.6 and 0.4 respectively. The voltage difference threshold is set to ΔV thre =50mV, when |V i -V j |>ΔV thre When determining the need for balance.

[0139] Then adjust the equilibrium intensity according to the temperature distribution state. The temperature adjustment coefficient calculation formula is:

[0140]

[0141] Where kT，i is the temperature adjustment coefficient of the i-th battery cell; T i is the temperature of the i-th battery cell; T avg is the average temperature of the battery pack; ΔT thre The temperature difference threshold is 5°C. If the high-temperature monomer exceeds this threshold, the balancing current will be reduced.

[0142] Next, we construct a multi-objective optimization problem. The objective function includes two dimensions: minimizing the equilibrium time and minimizing the energy loss. The objective function is expressed as:

[0143]

[0144] Where f1(x) is the equilibrium time objective function; f2(x) is the energy loss objective function; T balance The total time required to complete equilibrium; E loss is the energy loss during the balancing process; C is the rated capacity of the battery; I bal，ij is the equilibrium current from the i-th monomer to the j-th monomer; R bal is the equivalent resistance of the balancing circuit; t bal，ij is the equilibrium duration between monomers i and j.

[0145] Constraints include:

[0146] 0.05C≤I bal，ij ≤0.2C, the upper limit of the balancing current is 0.2C rate;

[0147] t bal，ij ≥10min, the lower limit of the equilibrium time is 10 minutes;

[0148] T i ≤40℃, the upper temperature limit is 40℃.

[0149] The system uses the non-dominated sorting genetic algorithm (NSGA-II) to solve the multi-objective optimization problem. The population size is set to 100, the evolutionary generations are 50, the crossover probability is 0.85, and the mutation probability is 0.1. Finally, the Pareto optimal solution set P is obtained. * .

[0150] The system selects the most suitable equilibrium strategy for the current working conditions from the optimal solution set. The selection criteria are:

[0151]

[0152] Where x * is the selected equilibrium strategy; P * is the Pareto optimal solution set; and f1(x) and f2(x) in P *The maximum value in w t and w e They are the weight coefficients of time and energy respectively. In the case of urgent need for balance, w t =0.7, w e =0.3, under normal equilibrium conditions w t =0.5, w e =0.5.

[0153] The system generates a balancing operation instruction sequence including the balancing target battery cell ID, balancing current size (ranging from 0.05C to 0.2C), balancing duration (ranging from 10 minutes to 120 minutes), and balancing priority (levels 1 to 5). The balancing priority calculation formula is:

[0154]

[0155] Where, P pri，ij is the priority of the balancing operation between monomers i and j; M ij score the imbalance; Indicates a round-up operation. The main purpose of this step is to implement intelligent balancing decisions based on battery status, balance balancing speed and energy efficiency, and extend the battery system life.

[0156] The specific implementation of step S06 is to use a pre-trained battery parameter error model to predict and compensate for the errors generated during the execution of the balancing operation instruction sequence. This model is based on a deep residual attention network structure and monitors the battery response in real time during the balancing process. First, the battery state change data before and after the balancing operation is obtained, including the voltage distribution before and after balancing, SOC difference, and temperature change. The theoretical calculation formula for balancing effect is:

[0157]

[0158] Where ΔV theo，i is the theoretical voltage change of the i-th battery cell; I bal，i is the balancing current of the i-th battery cell; t bal，i is the equilibrium duration; C i is the monomer capacity; It is the derivative of voltage with respect to SOC, indicating the rate of change of voltage at the battery operating point.

[0159] These data are then compared with the ideal equilibrium effect predicted by the model to calculate the error vector E, which is calculated as follows:

[0160] E=[ΔV real，1 -ΔV theo，1 , ΔV real，2 -ΔV theo，2 , ..., ΔVreal，n -ΔV theo，n ];

[0161] Where, E is the error vector of the equalization effect; ΔV real，i is the actual measured voltage change of the i-th battery cell; ΔV theo，i is the theoretically calculated voltage change; n is the number of battery cells.

[0162] The system applies the Kalman filter algorithm to the error vector to suppress noise. The state equation and observation equation of the Kalman filter are:

[0163] x k =Ax k-1 +w k-1 ;

[0164] z k =Hx k +v k ;

[0165] Where x k is the state vector, which represents the error vector after filtering; z k is the observation vector, which represents the error vector of the original calculation; A is the state transfer matrix; H is the observation matrix; w k is process noise, with mean 0 and covariance Q; v k is the observation noise with mean 0 and covariance R.

[0166] The prediction step and update step calculation formula of Kalman filter are:

[0167] Prediction steps:

[0168]

[0169] Update steps:

[0170]

[0171] Where, is the prior state estimate; is the prior estimation error covariance; K k is the Kalman gain; is the posterior state estimate; P k is the posterior estimation error covariance; I is the identity matrix.

[0172] The filtered error vector is input into the multi-head self-attention layer of the pre-trained model, and the attention weight threshold is set to θ attn =0.6. Abnormal features exceeding this value will receive higher attention weights. The calculation formula of the multi-head self-attention mechanism is:

[0173]

[0174] Where Q, K, and K are query matrix, key matrix, and value matrix respectively; d k is the dimension of the key vector; softmax is the normalized exponential function. The calculation formula of the multi-head attention mechanism is:

[0175] MultiHead(Q,K,V)=Concat(head1,head2,...,head h )W O ;

[0176] head i =Attention(QW i Q , KW i K , VW i V );

[0177] Where MultiHead is the multi-head attention output; head i is the output of the i-th attention head; h is the number of attention heads; W i Q 、W i K 、W i V are the query, key, and value projection matrices of the i-th attention head respectively; W O is the output projection matrix; Concat is the splicing operation.

[0178] The model outputs equalization error compensation parameters, including the equalization time adjustment coefficient α t , Balanced current adjustment coefficient α I and the equilibrium target adjustment parameter δ V The calculation formula of the adjusted equalization parameters is:

[0179] t bal，adj =t bal α t ;

[0180] I bal，adj =I bal α I ;

[0181] V target，adj =V target +δ V ;

[0182] Where, t bal，adj is the adjusted equilibrium time, the value range is 0.8t bal Up to 1.2tbal ;I bal，adj is the adjusted balancing current, with a value range of 0.7I bal to 1.3I bal ; V target，adj is the adjusted balanced target voltage; t bal , I bal 、V target are the original balancing time, current, and target voltage, respectively.

[0183] The system generates balancing correction instructions based on these parameters, such as extending the balancing time, adjusting the balancing current, or changing the balancing target cell, and sends them to the battery balancing device for correction. The error threshold is set to ε of the voltage correction target. thre =5%, when The main purpose of this step is to compensate for the deviation between theoretical equilibrium and actual effect through the prediction model and realize dynamic optimization of the equilibrium process.

[0184] The specific implementation of step S07 is to calculate the battery abnormality index according to the battery abnormality evaluation function, and update the pre-trained battery parameter error model weight in real time based on the index and the equalization operation result. The function first extracts the standard working condition feature vector set F from the historical battery working state feature vector database. std As a reference benchmark, the current battery working state feature vector F is then calculated. current The Mahalanobis distance from the reference is calculated as:

[0185]

[0186] Where D M is the Mahalanobis distance; F current is the current working state feature vector; is the mean of the characteristic vector of the standard working condition; is the covariance matrix of the eigenvector of the standard operating condition; T represents the matrix transpose.

[0187] Taking into account the two auxiliary indicators of voltage fluctuation rate and temperature distribution uniformity, the voltage fluctuation rate calculation formula is:

[0188]

[0189] Where R V is the voltage fluctuation rate; T is the time window length in minutes; V(t) is the voltage value at time t; is the average voltage in the window; Indicates the maximum absolute value of the voltage deviation from the average value within the window. The voltage fluctuation rate threshold is set to R V，thre =3% / min.

[0190] The temperature distribution uniformity is expressed by the temperature standard deviation, and the calculation formula is:

[0191]

[0192] Where σ T is the temperature standard deviation; n is the number of temperature measurement points; T i is the temperature of the i-th measurement point; is the average temperature. The temperature standard deviation threshold is set to σ T，thre =3℃.

[0193] The system combines Mahalanobis distance and auxiliary indicators through weighted average method to generate comprehensive battery abnormality index A batt , the calculation formula is:

[0194]

[0195] Where A batt D is the battery abnormality index, ranging from 0 to 1; M，max is the upper limit of the normal Mahalanobis distance, which is 10; w1, w2, and w3 are weight coefficients, which are 0.5, 0.3, and 0.2 respectively. The abnormality index threshold is set to A thre =0.7, exceeding this value is considered an abnormal state.

[0196] Based on the abnormality index, the system uses the stochastic gradient descent method to update the weights of the pre-trained battery parameter error model. The weight update formula is:

[0197]

[0198] Where W t+1 is the updated weight; W t is the current weight; η is the learning rate; is the gradient of the loss function with respect to the weight. The formula for dynamic adjustment of learning rate is:

[0199] η=η0·(1+α·δ(A batt >A thre ));

[0200] Where η is the actual learning rate used; η0 is the basic learning rate, with an initial value of 0.001; α is the learning rate gain coefficient, with a value of 4; δ() is the indicator function. batt >A thre The value is 1 when , otherwise it is 0.

[0201] The system uses gradient clipping technology to limit the weight update amplitude. The clipping formula is:

[0202]

[0203] Where, is the clipped gradient; G thre is the gradient clipping threshold, which is 5.0; is the L2 norm of the gradient.

[0204] The update mainly targets the attention head weights related to the detected anomalies in the multi-head self-attention layer. The attention head weight update strategy is:

[0205] ΔW attn，i =β i ΔW attn ;

[0206]

[0207] In the formula, ΔW attn，i is the update amount of the i-th attention head weight; ΔW attn is the basic update amount of all attention layer weights; β i is the weight distribution coefficient of the i-th attention head; s i is the abnormal correlation score of the ith attention head; τ is the temperature parameter that controls the smoothness of weight distribution and its value is 0.5; h is the total number of attention heads.

[0208] The system uses an annealing strategy to periodically reduce the learning rate. The learning rate attenuation formula is:

[0209]

[0210] Where, is the basic learning rate after the k+1th annealing; is the base learning rate after the kth annealing; γ is the learning rate decay factor, set to 0.9. Learning rate annealing is performed after every 500 weight updates. The main purpose of this step is to optimize the performance of the prediction model through a real-time feedback mechanism and improve the model's adaptability to abnormal conditions.

[0211] Regarding the hardware implementation, the system adopts a layered distributed architecture. The control chip utilizes a 32-bit ARM Cortex-M4 series processor with a main frequency of 120MHz, built-in 512KB Flash memory and 128KB RAM, and an integrated hardware floating-point unit to accelerate complex calculations. Each BMU module manages up to 16 battery cells. It uses a high-precision AD7280 analog-to-digital converter for voltage measurement, achieving an accuracy of ±0.25mV, and an NTC thermistor combined with a low-noise signal conditioning circuit for temperature measurement, achieving an accuracy of ±0.5°C. The BCMU module utilizes an STM32F407 processor, managing up to eight subordinate BMUs. An integrated CAN controller enables communication with the BMUs at a rate of 500kbps. The BSMU module utilizes a higher-performance embedded computing platform, equipped with an ARM Cortex-A72 quad-core processor running at a main frequency of 1.5GHz and 8GB of internal storage for running deep learning models and storing historical data. The battery balancing device utilizes switched capacitor technology for bidirectional active balancing, with a maximum balancing current of 2A and a balancing efficiency of up to 95%. The current metering device utilizes a Hall-effect current sensor and a gain-adjustable amplifier, offering a measurement range of -100A to 100A with an accuracy of 0.5% of full scale. The Ethernet communication device supports 100Mbps communication speeds and utilizes an industrial-grade protection design to ensure reliable communication. The remote communication device integrates a 4G module and a satellite communication backup system, supporting multi-path redundant communication. The power management device utilizes a low-power design, with static power consumption of less than 10mW and supports a battery voltage range of 9V to 36V.

[0212] Regarding the detailed structure of the battery parameter error model, the model adopts a deep residual attention network architecture, which includes an input layer, four residual-connected convolution blocks, two multi-head self-attention mechanism layers, two fully connected layers, and an output layer. The input layer receives a three-dimensional tensor of shape [B×T×F], where B is the batch size, T is the time window length, and F is the feature dimension. Each residual convolution block consists of two one-dimensional convolution layers with convolution kernel sizes of 3 and 5 respectively. The number of channels increases layer by layer to 64, 128, 256, and 512. Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The residual connection uses 1×1 convolution to achieve dimensional matching. The mathematical expression of the residual block is:

[0213]

[0214] X l+1 =ReLU(Y l );

[0215] Where, X l is the input of the lth residual block; Y l is the result after residual connection; Xl+1 is the output after activation; is the residual function, which consists of two convolutional layers; W l is the weight parameter of the lth residual block.

[0216] The first multi-head self-attention layer has 4 heads, the second has 8 heads, and the attention dimension is 64. The residual connection structure and layer normalization ensure stable gradient propagation. The layer normalization formula is:

[0217]

[0218] Where γ and β are learnable scaling and bias parameters; μ and σ are the mean and standard deviation of the input x, respectively; ∈ is a small constant to prevent division by zero, and its value is 10 -5 ; ⊙ represents element-wise product.

[0219] The number of nodes in the fully connected layer is 256 and 128 respectively. Dropout technology is used to prevent overfitting, and the dropout rate is 0.3. The dropout operation can be expressed as:

[0220]

[0221] Where x is the input; y is the output; p is the Dropout rate, which is 0.3.

[0222] The number of nodes in the output layer is consistent with the dimension of the prediction parameters, providing prediction values for the error compensation parameters. The loss function of the model is a combination of mean square error and L2 regularization term, expressed as:

[0223] L total =L MSE +λ·L reg ;

[0224]

[0225] L reg =∑ w∈W w 2 ;

[0226] Where, L total is the total loss; L MSE is the mean square error loss; L reg is the L2 regularization term; λ is the regularization coefficient, which is 10 -4 ; N is the number of samples; y i is the true label; is the model prediction value; W is the set of all weight parameters of the model.

[0227] The steps for establishing the training data set include: collecting at least 6 months of historical battery operation data, covering various operating conditions such as charging, discharging, static, and balancing; dividing the data into standard operating condition categories based on factors such as charge and discharge current, duration, and ambient temperature, with each category containing at least 500 samples; using a 60-second sliding window with a step size of 10 seconds to slice the time series data into fixed-length samples; calculating the error between the theoretical SOC value and the measured SOC value in each sample as the model learning target; expanding the training set by adding Gaussian noise (standard deviation is 3% of the original signal), time scale stretching (factor range is 0.8 to 1.2), random masking (maximum masking ratio is 20%) and other data enhancement techniques; finally dividing the training set, validation set, and test set in a ratio of 7:1:2 to ensure the generalization ability of the model.

[0228] In order to better understand and implement the present invention, the following provides a specific application scenario of the present invention, Example 2: A research team implemented a marine box-type power battery management system on a large ocean-going vessel, such as Figure 2-3 As shown in the figure, the system adopts a hierarchical distributed architecture, including a top-level battery stack management unit BSMU, four battery cluster management units BCMU, and 16 battery module management units BMU under each BCMU. The system manages a total of 1024 lithium-ion battery cells, each BMU manages 16 battery cells, and each BCMU manages 16 BMUs. The BSMU communicates with the BCMU via Ethernet, and the BCMU communicates with the BMU via the CAN bus. The ship is usually powered by a diesel generator and switches to battery power when it is docked and sailing in special sea areas to reduce pollution emissions. During the implementation process, the researchers first configured and initialized the battery management system, as shown in Table 1:

[0229] Table 1 Battery management system initialization parameters

[0230] Parameter name Parameter value unit illustrate Voltage sampling frequency 5 Hz 5 samples per second Temperature sampling frequency 1 / 60 Hz 1 sample per minute Current sampling frequency 1 Hz 1 sample per second Time window length 30 s Data processing window Sliding average window half-width 2 - Voltage filter parameters Median filter window half-width 1 - Current filter parameters CAN communication rate 500 kbps Between BMU and BCMU Ethernet communication rate 100 Mbps Between BCMU and BSMU Activation energy 20000 J / mol Temperature compensation parameters Voltage compensation basic coefficient 0.08 - Voltage compensation parameters Voltage difference threshold 50 mV Balance trigger threshold Temperature difference threshold 5 ℃ Temperature impact threshold Balancing current upper limit 0.2C - C is the rated capacity Minimum balancing time 10 min Equilibrium Sustained Lower Limit

[0231] After the system was operational, researchers monitored the battery balancing performance after multiple charge and discharge cycles. During a typical balancing operation, the system identified significant consistency differences in the battery cells managed by the seventh BMU in the second BCMU group. The battery status data collected by the system is shown in Table 2:

[0232] Table 2 Battery status parameters before balancing

[0233]

[0234]

[0235] Based on this data, the system calculated a voltage range of 0.095V, a SOC difference of 13.3%, and a temperature difference of 2.2°C. The battery voltage distribution entropy was calculated to be 1.42, below the threshold of 1.8, indicating a significant consistency issue. Using a deep time series analysis model, the system generated a balanced operation instruction sequence, as shown in Table 3:

[0236] Table 3 Equalization operation instruction sequence

[0237] Balanced Object Balancing current (mA) Equilibrium time (min) Balance priority Predicted voltage change (mV) B2-7-5→B2-7-11 640 42 1 28 B2-7-14→B2-7-11 680 45 1 32 B2-7-8→B2-7-6 320 25 2 12 B2-7-1→B2-7-15 280 22 3 10 B2-7-9→B2-7-4 240 18 4 8

[0238] During the balancing operation, the pre-trained battery parameter error model monitored the operation's effectiveness in real time. The model detected a 7.5% deviation between the actual voltage change and the predicted value for the first group of balanced objects, exceeding the 5% error threshold. Therefore, the system generated a correction instruction to increase the balancing time from 42 minutes to 51 minutes, ensuring that the balancing effect met the expected target. After the balancing operation was completed, the system collected battery status data again, and the results are shown in Table 4:

[0239] Table 4 Battery status parameters after equalization

[0240] Battery number Voltage (V) SOC (%) Temperature (℃) Internal resistance (mΩ) B2-7-1 3.645 77.6 25.4 42.6 B2-7-2 3.648 77.9 25.6 43.1 B2-7-3 3.635 75.8 25.3 44.2 B2-7-4 3.636 76.0 25.7 43.9 B2-7-5 3.657 78.9 26.1 41.8 B2-7-6 3.632 75.3 25.2 45.2 B2-7-7 3.628 74.6 25.3 44.6 B2-7-8 3.643 77.3 25.5 42.8 B2-7-9 3.641 76.8 25.6 43.5 B2-7-10 3.638 76.3 25.4 43.9 B2-7-11 3.619 73.3 25.2 46.5 B2-7-12 3.639 76.5 25.5 43.6 B2-7-13 3.641 76.9 25.6 43.2 B2-7-14 3.658 79.0 25.9 41.5 B2-7-15 3.633 75.4 25.3 44.9 B2-7-16 3.630 75.0 25.4 44.1

[0241] After balancing, the battery pack's voltage range decreased from 0.095V to 0.039V, the SOC variation decreased from 13.3% to 5.7%, and the temperature variation decreased from 2.2°C to 0.9°C. The battery abnormality index decreased from 0.74 to 0.32, below the threshold of 0.7, indicating that the system has returned to normal.

[0242] Based on the data collected by the system, the researchers also calculated the battery management performance indicators of the ship during the six-month operation period, as shown in Table 5:

[0243] Table 5 Statistics of battery management system performance indicators

[0244]

[0245]

[0246] Traditional marine battery management systems primarily employ simple voltage detection and fixed-threshold balancing strategies, lacking comprehensive analysis and dynamic balancing control of multidimensional battery states. This approach can easily lead to inaccurate battery state estimation, low balancing efficiency, and poor battery pack consistency in complex water environments and changing load conditions, ultimately impacting battery life and system reliability. In contrast, the marine box-type power battery management system implemented in the present invention collects multidimensional battery state data in time window segments, applies a deep time series analysis model to accurately calculate battery SOC and SOH values, corrects output errors in real time through a combined voltage and temperature compensation mechanism, and implements intelligent balancing decisions based on a multi-objective optimization algorithm. The system also uses a pre-trained battery parameter error model to provide real-time compensation for the balancing process and dynamically updates the model based on battery anomaly indicators. These innovative technologies significantly improve battery state estimation accuracy, reduce the frequency of balancing operations, improve battery pack consistency and temperature uniformity, extend battery life, and enhance the reliability and economy of the ship's power system.

[0247] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 6, 7, 8 and 9 below.

[0248] Table 6 Variable Explanation Table (Part 1)

[0249]

[0250] Table 7 Variable Explanation Table (Part 2)

[0251]

[0252]

[0253] Table 8 Variable Explanation Table (Part 3)

[0254]

[0255] Table 9 Variable Explanation Table (Part 4)

[0256]

[0257] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A marine box-type power battery management system, characterized in that: It includes a control chip, a battery pack management unit BMU, a battery cluster management unit BCMU, a battery stack management unit BSMU and related monitoring devices. The control chip is equipped with a battery management control module, which adopts a time window segmented acquisition method to obtain multi-dimensional battery status data, and uses a battery status evaluation function to extract the battery working status feature vector. Based on the deep time series analysis model, the battery parameters are calculated and the error compensation parameters are generated. The battery output error is corrected in real time through the voltage-temperature joint compensation mechanism, and the balancing strategy optimization function is called to determine the multi-level balancing control scheme. The pre-trained battery parameter error model is used to predict and compensate for the errors generated during the execution of the balancing operation instruction sequence.

2. The marine box-type power supply battery management system according to claim 1, characterized in that: The battery management control module is used to perform the following steps: using a time window segmented acquisition method to obtain multidimensional battery status data, using a battery status evaluation function to obtain a battery operating status feature vector, calculating battery parameters based on a deep time series analysis model, correcting battery parameters through a voltage-temperature joint compensation mechanism, calling a balancing strategy optimization function to determine a balancing control scheme, using a pre-trained battery parameter error model for prediction and compensation, and updating the pre-trained battery parameter error model according to a battery abnormality evaluation function.

3. The marine box-type power supply battery management system according to claim 2, characterized in that: The multi-dimensional battery status data includes voltage time series data, temperature time series data, current time series data and internal resistance time series data, which are obtained using the relevant monitoring device. Specifically, the voltage time series data is provided by the voltage monitoring device, the temperature time series data is provided by the temperature monitoring device, the current time series data is provided by the current metering device, and the internal resistance time series data is calculated by dividing the voltage time series data by the current time series data.

4. The marine box-type power supply battery management system according to claim 3, characterized in that: The battery operating state feature vector includes battery health features, temperature distribution features, voltage consistency features, and load response features. The battery operating state feature vector is used for subsequent battery parameter calculation and error compensation.

5. The marine box-type power supply battery management system according to claim 4, characterized in that: The battery state evaluation function extracts the fluctuation spectrum characteristics of voltage time series data through Fourier transform, evaluates the consistency level of the battery pack through entropy calculation, identifies potential hot spots through temperature gradient analysis, and analyzes the dynamic performance of the battery through load step response.

6. The marine box-type power supply battery management system according to claim 5, characterized in that: The corrected battery parameters include a corrected battery actual remaining capacity SOC value and a corrected state of health SOH value, and the corrected battery parameters are used for balancing strategy optimization and charge and discharge control.

7. The marine box-type power supply battery management system according to claim 6, characterized in that: The balancing strategy optimization function is used to determine the optimal balancing control strategy based on the battery status. The input includes the voltage value of each battery cell, the corrected actual remaining battery capacity SOC value, the temperature distribution status of the battery pack, the current charge and discharge status, and the ambient temperature value. The output is a balancing operation instruction sequence including the balancing target battery cell identifier, balancing current size, balancing duration, and balancing priority.

8. The marine box-type power supply battery management system according to claim 7, characterized in that: The balancing strategy optimization function takes into account both balancing speed and balancing efficiency through a multi-objective optimization algorithm, calculates the optimal balancing path based on the voltage difference matrix, and adjusts the balancing intensity in combination with the temperature influence factor.

9. The marine box-type power supply battery management system according to claim 8, characterized in that: The battery abnormality evaluation function is used to calculate the abnormality of the battery operating state. The input includes the battery working state feature vector, the historical battery working state feature vector database, the corrected battery parameters, the voltage fluctuation rate and the temperature distribution uniformity. The output is the battery abnormality index.

10. The marine box-type power supply battery management system according to claim 9, characterized in that: The battery abnormality evaluation function determines the degree of abnormality by calculating the Mahalanobis distance between the current battery operating state feature vector and the standard operating condition feature vector in the historical battery operating state feature vector database, and performs a comprehensive score based on the voltage fluctuation rate and temperature distribution uniformity to generate the final battery abnormality index.

Images