An online monitoring method and system based on battery state of health

Abstract

The present application relates to the technical field of battery management, aiming at the problem that the existing battery monitoring static model is difficult to adapt to aging drift, resulting in maintenance lag, applicable to the scene of new energy vehicle power battery, specifically an online monitoring method and system based on battery health state, comprising: obtaining the voltage, current, temperature and surface strain change curve in the charging and discharging cycle of the target battery; time synchronization calibration and feature synchronization extraction are performed on the four types of curves to generate a battery state feature vector set; the set is input into an improved convolutional neural network architecture to output a battery health state evaluation score and a capacity attenuation mode identifier; the score, identifier and multi-dimensional operation data are input into an online model corrector to dynamically correct the evaluation network; according to the attenuation mode identifier, the maintenance strategy library is searched, and the evaluation score is combined to generate a maintenance scheme containing operation and priority. The method improves the monitoring accuracy and maintenance pertinence through multi-dimensional data fusion and dynamic model correction.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and in particular to an online monitoring method and system based on battery health status. Background Technology

[0002] Current battery health monitoring methods primarily employ electrochemical impedance spectroscopy or threshold alarms based on voltage, current, and temperature, relying on the trends of single or limited parameter changes to infer health status. These methods fail to incorporate real-time data on battery physical deformation, and the feature extraction process lacks time alignment processing for multi-source heterogeneous data. This results in the generated evaluation vectors failing to comprehensively map the correlation between the battery's internal structural evolution and performance degradation. Furthermore, once the static evaluation model is trained, its parameters are fixed, making it unable to adapt to the data distribution drift caused by material aging and internal resistance growth throughout the battery's lifespan. Consequently, prediction biases continue to widen over long-term operation.

[0003] Existing solutions can only output a single health index, failing to distinguish specific capacity decay patterns, leading to a disconnect between maintenance strategies and actual battery failure mechanisms. The multi-dimensional data acquisition process suffers from timestamp asynchrony issues; the sampling frequencies for voltage and current differ significantly from those for temperature and strain, and the timing correspondence is ignored during feature extraction, resulting in the loss of crucial information. Model updates rely on manual offline retraining, making it impossible to achieve autonomous evolution using online data and struggling to cope with sudden state changes under complex operating conditions.

[0004] To address the issues of insufficient multiphysics data fusion and inadequate model dynamic adaptability, it is necessary to construct a synchronous acquisition system covering deformation features and establish a closed-loop feedback mechanism for evaluation results. By introducing high-dimensional feature synchronous extraction and online correction logic, the problems of monitoring distortion and maintenance lag caused by the single data dimension and rigid model in traditional technologies can be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies by proposing an online monitoring method and system based on battery health status.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an online monitoring method based on battery health status, comprising:

[0007] Acquire multi-dimensional operational data of the target battery during charge-discharge cycles, including voltage change curves, current change curves, temperature change curves, and surface strain change curves.

[0008] Time synchronization calibration and feature extraction are performed on the voltage change curve, current change curve, temperature change curve and surface strain change curve to generate a set of battery state feature vectors.

[0009] The set of battery state feature vectors is input into the trained battery health status assessment network for processing. The battery health status assessment network adopts an improved convolutional neural network architecture and outputs a battery health status assessment score and a capacity decay mode identifier.

[0010] The battery health status assessment score, capacity decay mode identifier, and multi-dimensional operating data are input into the online model corrector to dynamically correct the battery health status assessment network.

[0011] Based on the capacity decay mode identifier, the corresponding recommended maintenance strategy is retrieved from a preset library of multiple maintenance strategies, and combined with the battery health status assessment score, a battery health maintenance plan containing specific maintenance operations and execution priorities is generated.

[0012] As a further aspect of the present invention, the voltage change curve, current change curve, temperature change curve, and surface strain change curve are subjected to time synchronization calibration and feature synchronization extraction to generate a battery state feature vector set, including:

[0013] A unified timestamp reference is configured for the voltage change curve, current change curve, temperature change curve and surface strain change curve. The start time of the charge and discharge cycle is taken as the zero point, and the time axis of all curves is aligned to eliminate the phase difference caused by the asynchronous sampling time of different sensors.

[0014] On each curve after time axis alignment, instantaneous observations are extracted at fixed sampling time intervals to form voltage, current, temperature and strain sequences.

[0015] Key charge and discharge feature points are simultaneously identified and located on the voltage sequence, current sequence, temperature sequence, and strain sequence. These key charge and discharge feature points include the constant current charging to constant voltage charging point, the charging termination voltage point, and the discharging termination voltage point.

[0016] Taking each key charging and discharging characteristic point as the center, expand the data window to the front and back with a preset width respectively. Extract the average voltage and voltage range in the data window of the voltage change curve, extract the average current and current ripple coefficient in the data window of the current change curve, extract the temperature rise rate in the data window of the temperature change curve, and extract the cumulative strain in the data window of the surface strain change curve.

[0017] After normalizing the average voltage, voltage range, average current, current ripple coefficient, temperature rise rate, and strain accumulation corresponding to the same key charge and discharge characteristic point, they are merged into a multi-dimensional feature vector. The multi-dimensional feature vectors corresponding to all key charge and discharge characteristic points are arranged in chronological order to form the battery state feature vector set.

[0018] As a further aspect of the present invention, the synchronous identification and positioning of key charging and discharging feature points includes:

[0019] On the current change curve, locate the inflection point where the current value starts to decrease from the constant current value, mark the inflection point as the point where constant current charging turns into constant voltage charging, and record the timestamp corresponding to the inflection point and the voltage value at this time.

[0020] On the voltage change curve, locate the moment when the voltage value first reaches the preset charging cutoff voltage threshold, mark the moment as the charging termination voltage point, and record the current value at that moment.

[0021] On the voltage change curve, the moment when the voltage value first drops to the preset discharge cutoff voltage threshold is marked as the discharge termination voltage point, and the current value at that moment is recorded.

[0022] When the slope of the temperature change curve changes during the charging process, the change point is marked as a temperature anomaly inflection point, and its timestamp and temperature value are recorded.

[0023] All identified key charging and discharging feature points, including the constant current charging to constant voltage charging point, the charging termination voltage point, the discharging termination voltage point, and the temperature anomaly inflection point, are sorted according to the time stamp order to form a feature point sequence with time sequence and type labels.

[0024] As a further aspect of the present invention, the battery state feature vector set is input into a trained battery health status assessment network for processing. The battery health status assessment network employs an improved convolutional neural network architecture, including:

[0025] The battery health status assessment network includes a feature encoder, a temporal correlation analyzer, and a health status decoder.

[0026] The feature encoder performs point-by-point feature transformation and dimensionality upscaling on the input set of battery state feature vectors to extract a high-dimensional abstract representation of each feature point.

[0027] The temporal correlation analyzer is based on an improved convolutional neural network architecture to process the high-dimensional abstract representation of the sequence. The improved convolutional neural network architecture is based on the causal dependency of the battery state feature vector in the time dimension and adopts a convolutional kernel structure with causal constraints to ensure that the network does not use feature information from future times when processing features at the current time.

[0028] The time-series correlation analyzer outputs a context encoding vector that characterizes the overall state evolution trend of the battery;

[0029] The health status decoder outputs a scalar battery health status assessment score and a vector capacity decay pattern identifier based on the context encoding vector. The capacity decay pattern identifier is used to indicate the main cause category of capacity decay.

[0030] As a further aspect of the present invention, the working principle of the improved convolutional neural network architecture includes:

[0031] Construct a set of one-dimensional causal convolution kernels, wherein the convolution calculation range of the causal convolution kernels is strictly limited to the feature data of the current time and the historical time, and the feature data of the future time do not participate in the output calculation of the current time;

[0032] In the time-series correlation analyzer, multiple causal convolutional layers composed of the causal convolutional kernels are stacked, and the output of each causal convolutional layer depends only on the data of the current and previous time steps in its input;

[0033] An extended causal convolutional structure with an adjustable dilation factor is introduced between adjacent causal convolutional layers. By adjusting the dilation factor, higher-level causal convolutional layers can capture dependencies over a longer temporal range with an exponentially expanded receptive field, while maintaining causal constraints.

[0034] After each causal convolution operation, a gated activation mechanism is introduced. This mechanism uses a gate value generated by a sigmoid function to control the proportion of feature information passed to the next layer, thereby enabling the filtering of features of different importance.

[0035] At the end of the temporal correlation analyzer, global temporal pooling is performed on the feature sequences after multi-layer causal convolution and gated activation processing, aggregating the variable-length feature sequences into a fixed-length context encoding vector.

[0036] As a further aspect of the present invention, the battery health status assessment score, capacity decay mode identifier, and multi-dimensional operating data are input together into an online model corrector to dynamically correct the battery health status assessment network, including:

[0037] Collect historical multi-dimensional operating data of the target battery within a predetermined time window, as well as historical battery health status assessment scores and historical capacity degradation pattern identifiers output by the battery health status assessment network within the corresponding time window.

[0038] When a complete offline battery capacity calibration result is obtained, the offline battery capacity calibration result is used as the true value reference of the battery health status at the current moment.

[0039] The true reference is compared with the battery health status assessment score output by the battery health status assessment network at the current time, and the assessment error is calculated.

[0040] Based on the magnitude and direction of the assessment error, and the evolution trend of historical capacity decay pattern identifiers, a fine-tuning gradient for the internal parameters of the battery health status assessment network is generated.

[0041] A conservative online learning strategy is adopted, in which the fine-tuning gradient is superimposed and updated on the trainable parameters of the battery health status assessment network at a learning rate less than a preset upper limit, so as to achieve gradual dynamic correction of the model and avoid drastic fluctuations in model performance due to excessively large single updates.

[0042] As a further aspect of the present invention, the conservative online learning strategy includes:

[0043] Maintain a model parameter update cache for the online model corrector to temporarily store the calculated fine-tuned gradients;

[0044] Set a gradient accumulation period. Within the gradient accumulation period, perform a moving average calculation on the fine-tuned gradients obtained from multiple consecutive evaluations in the update buffer to obtain the smoothed average fine-tuned gradient.

[0045] A decay function is set for the learning rate of the model, which is related to the number of battery cycles. As the number of cycles increases, the learning rate gradually decreases.

[0046] Each time a parameter update is performed, the smoothed average fine-tuning gradient is multiplied by the learning rate at the current time to obtain the final actual update amount applied to the model parameters;

[0047] The actual update amount is applied to the trainable parameters of the battery health status assessment network to complete an online correction, and the model parameter update cache is cleared to prepare for the next gradient accumulation cycle.

[0048] As a further aspect of the present invention, the capacity decay mode identifier is used to indicate the main cause category of capacity decay, and its generation process includes:

[0049] The health status decoder includes a multi-branch classifier, with each branch corresponding to a preset capacity decay cause category. The cause categories include active lithium ion loss, electrode material structure degradation, excessive growth of solid electrolyte interface film, and electrolyte drying.

[0050] The context encoding vector is input to the multi-branch classifier, and a class activation value is calculated on each classification branch;

[0051] The activation values ​​of each branch are normalized and converted into a probability distribution. The probability value of each branch represents the confidence that the corresponding cause category is the main cause of the current battery capacity degradation.

[0052] The normalized probability distribution vector is output as the capacity decay mode identifier. The capacity decay mode identifier is a multi-dimensional vector whose dimension is equal to the preset number of cause categories. The value of each element in the vector represents the probability of the corresponding category.

[0053] As a further aspect of the present invention, a corresponding recommended maintenance strategy is retrieved from a preset database of multiple maintenance strategies based on the capacity decay mode identifier, and combined with the battery health status assessment score, a battery health maintenance plan including specific maintenance operations and execution priorities is generated, including:

[0054] The capacity decay mode identifier is analyzed to determine the cause category with the highest probability value as the dominant decay mode;

[0055] Using the dominant decay mode as the index key, a search is performed in a preset maintenance strategy library, which stores a list of typical maintenance operations corresponding to different cause categories.

[0056] From the retrieved list of typical maintenance operations, select maintenance operation suggestions that match the current battery type and usage environment to form a basic maintenance operation set;

[0057] Based on the numerical range of the battery health status assessment score, each operation in the basic maintenance operation set is assigned an urgency weight; the lower the assessment score, the higher the urgency weight.

[0058] By combining the estimated time and resource consumption of each maintenance operation with the urgency weight, a comprehensive execution priority score for each operation is calculated.

[0059] Based on the comprehensive execution priority score from high to low, the operations in the basic maintenance operation set are sorted, and a suggested execution time window for each operation is attached to generate a structured battery health maintenance plan document.

[0060] Assign an urgency weight to each operation in the basic maintenance operation set, specifically including:

[0061] The range of battery health status assessment scores is divided into multiple consecutive level intervals, with each level interval corresponding to a standard urgency weight base.

[0062] Determine which grade range the current battery health status assessment score falls into, and obtain the standard urgency weight base corresponding to the range;

[0063] For each operation in the basic maintenance operation set, determine whether it is used to alleviate or repair the dominant decay mode. If so, multiply the standard urgency weight base corresponding to the operation by a mode enhancement coefficient greater than one.

[0064] Considering the current operating temperature of the battery, if the ambient temperature is higher or lower than the normal operating temperature range, the urgency weight of all operations is multiplied by an environmental degradation coefficient.

[0065] The weight values ​​adjusted by the mode enhancement coefficient and the environmental degradation coefficient will be used as the final urgency weights for basic maintenance operations.

[0066] As a further aspect of the present invention, the present invention also includes an online monitoring system based on battery health status, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the online monitoring method based on battery health status as described above.

[0067] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0068] The surface strain change curves are synchronized with the voltage, current, and temperature change curves at the nanosecond level. Waveform segments of each curve within the same charge-discharge range are extracted using a sliding window, and wavelet packet transform is employed to synchronously extract frequency and time domain features, eliminating phase deviations caused by multi-sensor sampling delays. The synchronized four-dimensional feature vector fully preserves the information on the coupling effect of battery electrochemical reaction and mechanical stress, enabling the feature set to simultaneously include the quantitative relationship between polarization voltage, heat generation rate, and electrode expansion and contraction. This overcomes the deficiency of single electrical parameters being easily affected by load fluctuations and improves the sensitivity of the features in characterizing early faults such as microcracks and lithium plating.

[0069] The improved convolutional neural network architecture employs a convolutional kernel structure with causal constraints, and fully connected layers output health status scores and decay mode classification labels in parallel. An online model corrector receives real-time evaluation results and raw running data, calculates the residual between predicted and measured values, and dynamically adjusts network weights through gradient backpropagation, enabling continuous iteration of model parameters without manual annotation. When matching decay mode identifiers to the maintenance strategy library, priorities are assigned based on health status scores, triggering differentiated equalization charging, temperature control adjustment, or static maintenance commands for different decay mechanisms such as active material shedding and electrolyte drying. Attached Figure Description

[0070] Figure 1 This is a flowchart of an online monitoring method based on battery health status according to the present invention;

[0071] Figure 2 A flowchart for generating a set of feature vectors for time synchronization calibration and feature synchronization extraction;

[0072] Figure 3A flowchart of the network architecture for battery health status assessment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0075] See Figure 1 This invention provides an online monitoring method based on battery health status. The method includes: acquiring multi-dimensional operational data of a target battery during charge-discharge cycles, including voltage change curves, current change curves, temperature change curves, and surface strain change curves; performing time synchronization calibration and feature extraction on the acquired voltage change curves, current change curves, temperature change curves, and surface strain change curves to generate a battery state feature vector set; inputting the generated battery state feature vector set into a pre-trained battery health status assessment network for processing, which employs an improved convolutional neural network architecture, and outputting a battery health status assessment score and a capacity decay mode identifier after processing; inputting the battery health status assessment score, capacity decay mode identifier, and the original multi-dimensional operational data into an online model corrector, which uses this information to dynamically correct the battery health status assessment network to maintain the accuracy of the assessment; and retrieving corresponding recommended maintenance strategies from a pre-set library of multiple maintenance strategies based on the capacity decay mode identifier output by the network, and combining this with the battery health status assessment score to generate a battery health maintenance plan that includes specific maintenance operations and execution priorities.

[0076] In one embodiment of the present invention, see [reference] Figure 2A unified timestamp reference was configured for the voltage, current, temperature, and surface strain curves. Using the start of the charge / discharge cycle as zero, all curves were time-axis aligned, eliminating phase differences caused by asynchronous sampling times from different sensors. On each time-axis aligned curve, instantaneous observations were extracted at fixed sampling intervals to form voltage, current, temperature, and strain sequences. Key charge / discharge feature points were simultaneously identified and located on these sequences, including the point where constant current charging transitions to constant voltage charging, the charging termination voltage, and the discharging termination voltage. The identification and positioning process is as follows: On the current change curve, the inflection point where the positioning current value starts to decrease from a constant current value is marked as the constant current charging to constant voltage charging point, and the corresponding timestamp and voltage value at this time are recorded; On the voltage change curve, the moment when the positioning voltage value first reaches the preset charging cutoff voltage threshold is marked as the charging termination voltage point, and the current value at this time is recorded; On the voltage change curve, the moment when the positioning voltage value first drops to the preset discharging cutoff voltage threshold is marked as the discharging termination voltage point, and the current value at this time is recorded; When the temperature change curve shows a change in slope during charging, this change point is marked as a temperature anomaly inflection point, and its timestamp and temperature value are recorded. All identified key charging and discharging feature points, including the constant current charging to constant voltage charging point, the charging termination voltage point, the discharging termination voltage point, and the temperature anomaly inflection point, are sorted in chronological order according to their timestamps to form a feature point sequence with time sequence and type labels. Centered on each key charge / discharge feature point, a data window of preset width is expanded forward and backward. Within the data window of the voltage change curve, the mean voltage and voltage range are extracted; within the data window of the current change curve, the average current and current ripple coefficient are extracted; within the data window of the temperature change curve, the temperature rise rate is extracted; and within the data window of the surface strain change curve, the cumulative strain is extracted. The mean voltage, voltage range, average current, current ripple coefficient, temperature rise rate, and cumulative strain corresponding to the same key charge / discharge feature point are normalized and merged into a multi-dimensional feature vector. All multi-dimensional feature vectors corresponding to key charge / discharge feature points are arranged in chronological order to form a set of battery state feature vectors.

[0077] In practical implementation, multi-dimensional operational data of the target battery during charge-discharge cycles are acquired. This multi-dimensional operational data includes voltage change curves, current change curves, temperature change curves, and surface strain change curves. A unified timestamp reference is configured for the voltage change curves, current change curves, temperature change curves, and surface strain change curves. All curves are aligned on the time axis with the start time of the charge-discharge cycle as the zero point. For example, in the charging cycle of a lithium-ion battery, the start time is defined as the moment when the charging current starts to rise from zero. The sampling frequency of the voltage sensor is 100 Hz, while the sampling frequency of the temperature sensor is 10 Hz. The temperature change curve is resampled to 100 Hz and aligned on the time axis with the voltage change curve using a linear interpolation algorithm. The current change curve and surface strain change curve are also aligned in the same way. Instantaneous observations are extracted from each curve after time axis alignment at fixed sampling intervals to form voltage, current, temperature, and strain sequences. The fixed sampling interval is set to 0.01 seconds, and data points are uniformly extracted from the aligned curves. The voltage sequence contains the voltage value at each time point in millivolts (mV), the current sequence contains the current value at each time point in milliamperes (mA), the temperature sequence contains the temperature value at each time point in degrees Celsius (°C), and the strain sequence contains the strain value at each time point in microstrain. Key charge / discharge characteristic points are simultaneously identified and located on the voltage, current, temperature, and strain sequences. These key charge / discharge characteristic points include the constant current charging to constant voltage charging point, the charging termination voltage point, and the discharging termination voltage point. The inflection point where the current value begins to decrease from a constant current value is located on the current change curve. This inflection point is marked as the constant current charging to constant voltage charging point, and the corresponding timestamp and voltage value are recorded. For example, during the constant current charging stage, when the current is maintained at 1000 mA, the first drop to 980 mA is identified as an inflection point, and the timestamp and inflection point voltage value of 3.9 V are recorded. The moment when the voltage value first reaches the preset charging cutoff voltage threshold on the voltage change curve is marked as the charging termination voltage point, and the current value at that moment is recorded. The preset charging cutoff voltage threshold is set to 4.2 volts for lithium-ion batteries. When the voltage value on the voltage change curve rises from below 4.2 volts to 4.2 volts, the timestamp of the first attainment of 4.2 volts and the corresponding current value of 500 mA are recorded. The moment when the voltage value first drops to the preset discharging cutoff voltage threshold on the voltage change curve is marked as the discharging termination voltage point, and the current value at that moment is recorded. The preset discharging cutoff voltage threshold is set to 3.0 volts. When the voltage value on the voltage change curve drops from above 3.0 volts to 3.0 volts, the timestamp of the first drop to 3.0 volts and the current value of -1000 mA are recorded.When the temperature change curve shows a change in slope during charging, this point is marked as a temperature anomaly inflection point, and its timestamp and temperature value are recorded. The point of change in slope of the temperature change curve is determined by calculating the derivative of temperature with respect to time and finding the local extremum of the derivative value. For example, when the temperature rise rate changes from 0.1 degrees Celsius per second to 0.3 degrees Celsius per second during the middle of charging, the timestamp and temperature value of 45 degrees Celsius are recorded. All identified key charging and discharging feature points, including the constant current charging to constant voltage charging point, the charging termination voltage point, the discharging termination voltage point, and the temperature anomaly inflection point, are sorted in chronological order according to their timestamps to form a feature point sequence with time sequence and type labels. The feature point sequence is stored in array form, with each element containing a timestamp, feature point type, and associated observation value such as voltage value, current value, or temperature value.

[0078] In some embodiments, time axis alignment employs a synchronization method based on a reference clock. The battery management system sends synchronization pulse signals to all sensors. Each sensor records data starting from the synchronization pulse and includes a high-precision timestamp. During subsequent processing, the multi-dimensional operational data is aligned to a unified time axis based on the timestamps. A data window of a preset width is extended forward and backward from each key charge / discharge feature point. Within the data window of the voltage change curve, the average voltage and voltage range are extracted; within the data window of the current change curve, the average current and current ripple coefficient are extracted; within the data window of the temperature change curve, the temperature rise rate is extracted; and within the data window of the surface strain change curve, the cumulative strain is extracted. The preset width data window is set to a time interval extending 2 seconds before and after the feature point's timestamp, containing 200 data points corresponding to a sampling frequency of 100 Hz. In a specific implementation, the average voltage is calculated as the arithmetic mean of all voltage values ​​within the data window, and the voltage range is calculated as the difference between the maximum and minimum voltage values ​​within the data window. For example, for the constant current charging to constant voltage charging point, the average voltage within the data window is 3.92 volts, and the voltage range is 0.08 volts. The current ripple factor is calculated using the following formula:

[0079] ;

[0080] in: Represents the current ripple factor. This represents the standard deviation of the current values ​​within the data window. This represents the average and standard deviation of the current values ​​within the data window. The average value measures the amplitude of current fluctuations. The current ripple coefficient, a dimensionless coefficient, represents the current level and reflects the relative intensity of current fluctuations. The temperature rise rate is calculated as the ratio of temperature change to time change within the data window, i.e., the first derivative of temperature with respect to time is estimated using the central difference method. For example, if the data window starts at 25°C and ends at 30°C with a time span of 4 seconds, the temperature rise rate is 1.25°C per second. The cumulative strain is calculated as the area integral under the surface strain change curve within the data window, approximated using the trapezoidal numerical integration method. Integrating the strain sequence with the time sequence yields the microstrain per second. In some embodiments, the width of the data window can be adjusted according to the battery type and application scenario. For high-rate charge / discharge batteries, the data window width is set to 1 second to reduce the impact of environmental noise. Optionally, the feature extraction process is executed in real time on the embedded processor. Feature calculation of the data window is triggered immediately after each key charge / discharge feature point is identified to reduce storage overhead. The average voltage, voltage range, average current, current ripple coefficient, temperature rise rate, and strain accumulation corresponding to the same key charge / discharge feature point are normalized and then merged into a multi-dimensional feature vector. The normalization process uses a min-max normalization method, obtaining the minimum and maximum values ​​of each feature from historical data and scaling the feature values ​​to the [0,1] interval. For example, if the historical range of the average voltage is 3.0V to 4.2V, then 3.92V is normalized to 0.767. This normalization eliminates the influence of different feature dimensions, making subsequent network processing more stable. The multi-dimensional feature vectors corresponding to all key charge / discharge feature points are arranged in chronological order to form a battery state feature vector set. This set is represented as a two-dimensional matrix, with each row corresponding to a feature point. The row order is arranged in ascending order of the feature point's timestamp. For example, if five key feature points are identified in one charge / discharge cycle, the matrix will have 5 rows and 6 columns, corresponding to six extracted features. In practice, the battery state feature vector set is transmitted to the battery health status assessment network via a data bus as input. Optionally, the feature extraction process is implemented in the microcontroller of the battery management system, using a fixed-point arithmetic library for numerical computation to improve efficiency. It can be understood that the generation of the battery state feature vector set can be performed offline or online. When executed online, it is synchronized with the battery charge-discharge cycle, and an updated battery state feature vector set is output after each charge-discharge cycle.

[0081] In one embodiment of the present invention, see [reference] Figure 3The battery health status assessment network comprises a feature encoder, a temporal correlation analyzer, and a health status decoder. The feature encoder performs point-by-point feature transformation and dimensionality upscaling on the input battery state feature vector set, extracting a high-dimensional abstract representation of each feature point. The temporal correlation analyzer, based on an improved convolutional neural network architecture, processes the sequence of high-dimensional abstract representations. This improved architecture, based on the causal dependencies between battery state feature vectors in the temporal dimension, employs a convolutional kernel structure with causal constraints. This structure ensures that the network does not use feature information from future times when processing features at the current moment. The temporal correlation analyzer outputs a context-encoded vector representing the overall evolution trend of the battery's state. Based on the context-encoded vector, the health status decoder outputs a scalar battery health status assessment score and a vector-based capacity decay pattern identifier. The capacity decay pattern identifier indicates the main causal category of capacity decay. Its generation process is as follows: the health status decoder includes a multi-branch classifier, with each branch corresponding to a preset capacity decay causal category. These causal categories include active lithium ion loss, electrode material structural degradation, excessive growth of the solid electrolyte interface film, and electrolyte drying. The context encoding vector is input to a multi-branch classifier, where a class activation value is calculated for each classification branch. The class activation values ​​for each branch are normalized and converted into a probability distribution. The probability value for each branch represents the confidence that the corresponding cause category is the primary reason for the current battery capacity degradation. The normalized probability distribution vector is output as a capacity degradation pattern identifier, which is a multi-dimensional vector with a dimension equal to the preset number of cause categories. Each element in the vector represents the probability of the corresponding category.

[0082] In its implementation, the battery health status assessment network comprises a feature encoder, a time-series correlation analyzer, and a health status decoder. The network receives an input sequence consisting of a set of battery state feature vectors from multiple charge-discharge cycles. The input sequence has dimensions [T, D], where T represents the number of time steps corresponding to the total number of identified key charge-discharge feature points, and D represents the dimension of each feature point corresponding to the number of extracted features. The feature encoder performs point-by-point feature transformation and dimensionality upscaling on the input battery state feature vector set, extracting a high-dimensional abstract representation for each feature point. The feature encoder consists of two fully connected neural networks. The first fully connected layer maps the input dimension D to the hidden dimension H1, and the second fully connected layer maps the hidden dimension H1 to the high-dimensional abstract representation dimension H2. For each time step feature vector in the input sequence, the feature encoder independently transforms it to generate a high-dimensional abstract representation of dimension H2, thus converting the input sequence [T, D] into a high-dimensional abstract representation sequence [T, H2]. The temporal correlation analyzer, based on an improved convolutional neural network architecture, processes high-dimensional abstract representation sequences. This improved architecture leverages the causal dependencies between battery state feature vectors over time, employing a causally constrained convolutional kernel structure. This ensures the network doesn't use feature information from future times when processing current features. The analyzer outputs a context-encoded vector representing the overall battery state evolution trend. This vector is a fixed-length vector with dimension C. The health state decoder, based on the context-encoded vector, outputs a scalar battery health state assessment score and a vector-based capacity decay pattern identifier. In some embodiments, the feature encoder uses the ReLU activation function, and the high-dimensional abstract representation dimension H2 is set to 128. Optionally, the feature encoder weights are shared across all time steps to ensure consistent transformation for each feature point. Internally, the temporal correlation analyzer contains multiple causal convolutional layers and gated activation units, with the context encoding vector dimension C set to 256. It is understandable that the high-dimensional abstract representation sequence captures the local deep features of each key feature point, while the context encoding vector integrates the global temporal information of the entire sequence.

[0083] In practical implementation, the capacity decay pattern identifier is used to indicate the main causes of capacity decay. The generation process of the capacity decay pattern identifier includes a health status decoder containing a multi-branch classifier, where each branch corresponds to a preset capacity decay cause category. The preset capacity decay cause categories include active lithium ion loss, electrode material structure degradation, excessive growth of the solid electrolyte interface film, and electrolyte drying. The context encoding vector is input to the multi-branch classifier, which consists of multiple parallel fully connected layers. Each fully connected layer corresponds to a classification branch, and a class activation value is calculated on each classification branch. For the j-th classification branch, its class activation value is... From the formula The calculation shows that, among which This represents the input context encoding vector. This represents the weight vector corresponding to the j-th branch. This represents the bias term corresponding to the j-th branch. This represents the class activation value of the j-th branch. The class activation values ​​of each branch are normalized and converted into a probability distribution. The normalization process uses the softmax function. The probability value of the j-th branch is... The calculation formula is:

[0084] ;

[0085] in: This represents the probability of the j-th cause category. This represents the category activation value of the j-th branch. This represents the category activation value of the k-th branch, where N represents the preset number of trigger categories and probability values. This represents the confidence level that the corresponding cause category is the main reason for the current battery capacity degradation. The normalized probability distribution vector is output as the capacity degradation mode identifier. The capacity degradation mode identifier is a multi-dimensional vector, the dimension of which is equal to the preset number of cause categories N. The value of each element in the vector represents the probability of the corresponding category. For example, if there are 4 preset categories, the capacity degradation mode identifier is [P1, P2, P3, P4], where P1 corresponds to the confidence level of active lithium ion loss, P2 corresponds to the confidence level of electrode material structure degradation, P3 corresponds to the confidence level of excessive growth of the solid electrolyte interface film, and P4 corresponds to the confidence level of electrolyte drying. In some embodiments, the branch in the health status decoder that outputs the battery health status assessment score in parallel is an independent regressor. The regressor receives the same context encoding vector and outputs a scalar value between 0 and 1 through a fully connected layer, where 0 represents complete battery failure and 1 represents a battery in excellent health. Optionally, the output layer of the multi-branch classifier does not use a bias term to simplify the model. It is understandable that the capacity degradation mode identifier provides interpretable information about battery aging, while the battery health status assessment score provides a quantitative indicator of health. Together, they constitute a comprehensive assessment output of battery health status.

[0086] In one embodiment of the invention, the improved convolutional neural network architecture operates by constructing a set of one-dimensional causal convolutional kernels. The convolutional computation of these kernels is strictly limited to the feature data at the current and historical time steps; feature data from future time steps are not included in the output computation at the current time step. Multiple causal convolutional layers composed of causal kernels are stacked in the temporal correlation analyzer. The output of each causal convolutional layer depends only on the data from the current and previous time steps in its input. An adjustable dilation factor dilation factor dilation factor dilation is introduced between adjacent causal convolutional layers. By adjusting the dilation factor, higher-level causal convolutional layers capture dependencies over a longer temporal range with an exponentially expanded receptive field, while maintaining causal constraints. After each causal convolutional operation, a gated activation mechanism is introduced. This mechanism uses a sigmoid function to control the proportion of feature information passed to the next layer, enabling the filtering of features of different importance. At the end of the temporal correlation analyzer, global temporal pooling is performed on the feature sequences processed by multiple causal convolutions and gated activation, aggregating the variable-length feature sequences into a fixed-length context encoding vector.

[0087] In practical implementation, the working principle of the improved convolutional neural network architecture includes constructing a set of one-dimensional causal convolutional kernels. The convolutional computation of these kernels is strictly limited to the feature data at the current and historical time steps; feature data from future time steps are not included in the output computation at the current time step. For a one-dimensional causal convolutional kernel of length L, its output at time step t... Only depends on the input sequence from time step Time to step The data, mathematically represented as:

[0088] ;

[0089] in: This represents the output value at time step t. This represents the weight parameters of the one-dimensional causal convolution kernel at position i. Indicates the input sequence at time step The eigenvalues ​​are denoted as . In the temporal correlation analyzer, multiple one-dimensional causal convolutional layers composed of causal convolutional kernels are stacked. The output of each one-dimensional causal convolutional layer depends only on the data from the current and previous time steps in its input. For example, the first one-dimensional causal convolutional layer processes the original high-dimensional abstract representation sequence, and each point of its output sequence is calculated based on the current and historical high-dimensional abstract representations. The second one-dimensional causal convolutional layer takes the output sequence of the first layer as input, and its output further depends on a longer historical context. An dilated causal convolutional structure with an adjustable dilation factor is introduced between adjacent one-dimensional causal convolutional layers. By adjusting the dilation factor, higher-level one-dimensional causal convolutional layers capture dependencies over a longer temporal range with an exponentially expanded receptive field, while maintaining causal constraints. When calculating the output, the input points of the convolutional kernels of the dilated causal convolution have a gap, the size of which is determined by the dilation factor. The output of a one-dimensional causal convolution with a dilation factor of d at time step t is calculated as . Refer to Table 1, which shows a configuration example of a temporal correlation analyzer with four layers of dilated causal convolutions.

[0090] Table 1: Temporal Correlation Analyzer Configuration Table for Four-Layer Dilated Causal Convolution

[0091] In practice, the receptive field size represents the historical length of the input sequence that the network output can "see." The receptive field size R is calculated using the formula:

[0092] ;

[0093] in: This represents the size of the convolutional kernel in the l-th layer. Let represent the expansion factor of layer l, and N represent the total number of layers. According to the table configuration, the receptive field size of the fourth layer output is 31 time steps. After each one-dimensional causal convolution operation, a gated activation mechanism is introduced. This mechanism uses a sigmoid function to generate a gate value to control the proportion of feature information passed to the next layer, thus filtering features of different importance. This is applied to the output feature mapping of the one-dimensional causal convolution layer. The calculation process for gating activation is as follows:

[0094] ;

[0095] in: This represents the convolution operation. This represents the sigmoid function. and These are the weights and biases of the gated convolutional layers. and These are the weights and biases of the output transform layer. This represents element-wise multiplication, with a gate value ranging from 0 to 1, used to modulate the intensity of the output feature. In some embodiments, the expansion factor is set according to the expected length and dependency range of the input sequence; for long-cycle battery data, the higher-level expansion factor can be set to 16 or 32 to obtain a larger receptive field. Optionally, the output transformation layer in the gated activation mechanism can be omitted, simplifying it to... In the form of a time-series correlation analyzer, global temporal pooling is performed on the feature sequence after multiple layers of one-dimensional causal convolution and gated activation processing. This aggregates the variable-length feature sequence into a fixed-length context encoding vector. The global temporal pooling operation takes the maximum or average value of the feature sequence in the time dimension. For example, for a feature sequence of shape [T, F] (T is the number of time steps, F is the number of feature channels), max pooling along the time dimension T yields a context encoding vector of dimension [F]. In some embodiments, global temporal pooling employs an attention pooling mechanism, calculating the weight of each feature at each time step through a learnable query vector and performing a weighted sum. It can be understood that the one-dimensional convolutional kernel with causal constraints ensures the temporal causality of feature extraction, the expanded structure extends the receptive field, and the gating mechanism enhances the network's ability to filter important features. Optionally, the padding strategy of the one-dimensional causal convolutional layer adopts left padding, that is, padding the beginning of the input sequence with zeros to ensure that the length of the input sequence and the output sequence is consistent in the time dimension. It is understandable that the context encoding vector integrates the temporal features extracted from the entire input sequence under causal constraints, providing an information basis for the subsequent decoding process.

[0096] In one embodiment of the present invention, historical multi-dimensional operational data of the target battery within a predetermined time window are collected, along with historical battery health status assessment scores and historical capacity decay pattern identifiers output by the battery health status assessment network within the corresponding time window. When a complete offline capacity calibration result of the battery is obtained, this offline capacity calibration result is used as the true value reference for the battery health status at the current moment. The true value reference is compared with the battery health status assessment score output by the battery health status assessment network at the current moment to calculate the assessment error. Based on the magnitude and direction of the assessment error and the evolution trend of the historical capacity decay pattern identifiers, a fine-tuning gradient for the internal parameters of the battery health status assessment network is generated. A conservative online learning strategy is adopted, in which the fine-tuning gradient is superimposed and updated onto the trainable parameters of the battery health status assessment network at a learning rate less than a preset upper limit, thereby achieving progressive dynamic correction of the model. This strategy also avoids drastic fluctuations in model performance due to excessively large single updates. The specific implementation of this conservative online learning strategy includes: maintaining a model parameter update cache for the online model corrector to temporarily store the calculated fine-tuning gradient. A gradient accumulation period is set. Within this period, the fine-tuned gradients obtained from multiple consecutive evaluations are calculated using a moving average within the update buffer to obtain the smoothed average fine-tuned gradient. A decay function related to the number of battery cycles is set for the model's learning rate, which gradually decreases as the number of cycles increases. At each parameter update, the smoothed average fine-tuned gradient is multiplied by the current learning rate to obtain the final actual update amount applied to the model parameters. This actual update amount is applied to the trainable parameters of the battery health assessment network to complete an online calibration, and the model parameter update buffer is cleared to prepare for the next gradient accumulation period.

[0097] In practice, historical multi-dimensional operational data of the target battery within a predetermined time window is collected, along with historical battery health status assessment scores and historical capacity decay pattern identifiers output by the battery health status assessment network within the corresponding time window. The predetermined time window is set to the most recent 100 complete charge-discharge cycles. The historical multi-dimensional operational data includes voltage change curves, current change curves, temperature change curves, and surface strain change curves for each cycle. The historical battery health status assessment score is a scalar value output by the battery health status assessment network after each assessment of these 100 cycles, and the historical capacity decay pattern identifier is a corresponding multi-dimensional vector. This data is stored in a circular buffer. When a complete offline battery capacity calibration result is obtained, the offline battery capacity calibration result is used as the true reference for the current battery health status. The offline battery capacity calibration measures the actual capacity by fully discharging the battery to the cutoff voltage under laboratory conditions with a constant current. For example, if the measured actual capacity of the current battery is 85% of the rated capacity, the true reference is set to 0.85. The true reference is compared with the battery health status assessment score output by the battery health status assessment network at the current moment to calculate the assessment error. The calculation method is as follows:

[0098] ;

[0099] in: This represents the true reference value obtained from the battery offline capacity calibration results. This represents the battery health status assessment score output by the battery health status assessment network at the current moment. For example, if the true reference is 0.85 and the network output score is 0.82, then the assessment error is... The value is +0.03. Based on the magnitude and direction of the evaluation error, and the evolution trend of historical capacity decay pattern identifiers, a fine-tuning gradient for the internal parameters of the battery health status evaluation network is generated. The calculation of the fine-tuning gradient combines the partial derivatives of the mean squared error loss function of the evaluation error with respect to the network parameters, and is weighted with reference to the stability of the dominant pattern in the historical capacity decay pattern identifiers. If the historical capacity decay pattern identifiers remain consistent within the recent window, the magnitude of the fine-tuning gradient is appropriately reduced. A conservative online learning strategy is adopted, whereby the fine-tuning gradient is superimposed on the trainable parameters of the battery health status evaluation network with a learning rate less than a preset upper limit. The preset upper limit learning rate is set to 0.001 to achieve gradual dynamic correction of the model, while avoiding drastic fluctuations in model performance due to excessively large single updates.

[0100] In practice, the conservative online learning strategy involves maintaining a model parameter update cache for the online model adjuster to temporarily store the calculated fine-tuned gradients. This cache is a first-in, first-out (FIFO) queue data structure, with its capacity corresponding to the gradient accumulation period. A gradient accumulation period is set, and within this period, the fine-tuned gradients obtained from multiple consecutive evaluations are averaged within the update cache to obtain a smoothed average fine-tuned gradient. The gradient accumulation period is set to five battery offline capacity calibration events, meaning that an average update is performed after accumulating five fine-tuned gradients. The moving average calculation uses a weighted average, with the most recent fine-tuned gradient having the highest weight. See Table 2 for the calculation of the moving average of the fine-tuned gradients within one gradient accumulation period.

[0101] Table 2: Calculation of Moving Average of Fine-tuned Gradient within One Gradient Accumulation Period

[0102] In practical implementation, based on the fine-tuning gradients G1 to G5 in the table above, the smoothed average fine-tuning gradient is... Calculated as A decay function is set for the model's learning rate, which is related to the number of battery cycles. As the number of cycles increases, the learning rate gradually decreases. The learning rate decay function is defined as follows:

[0103] ;

[0104] in: This represents the learning rate of the battery in the nth cycle. This indicates that the initial learning rate is set to 0.001. This indicates that the decay coefficient is set to 0.0001, and n represents the number of battery cycles. During each parameter update, the smoothed average fine-tuned gradient is multiplied by the current learning rate to obtain the final actual update amount applied to the model parameters. The result is calculated using the formula:

[0105] ;

[0106] in: Indicates the actual update volume. This represents the learning rate during the nth iteration. This represents the smoothed average fine-tuned gradient. The actual update is applied to the trainable parameters of the battery health assessment network to complete an online correction. The trainable parameters include convolutional kernel weights, fully connected layer weights, and bias terms. The update operation is as follows:

[0107] ;

[0108] in: This indicates the parameters before the update. This represents the updated parameters. After updating the parameters, the model parameter update cache is cleared to prepare for the next gradient accumulation cycle. In some embodiments, the gradient accumulation cycle can be adjusted according to the stability of the battery type and application scenario. For batteries with drastic changes in operating conditions, the gradient accumulation cycle can be shortened to 3 times. Optionally, the moving average calculation can use a simple arithmetic average. It can be understood that the model parameter update cache and gradient accumulation cycle buffer the fluctuations of a single evaluation, making the update smoother. In some embodiments, the learning rate decay function can also use a piecewise constant or inverse time decay form. Optionally, the actual update amount can be gradient clipped before application to prevent the update step size from being too large in extreme cases. It can be understood that the conservative online learning strategy realizes the slow drift of the battery health status evaluation network parameters, enabling the network to adapt to the long-term slow changes in battery characteristics while maintaining robustness to short-term fluctuations.

[0109] In one embodiment of the present invention, a recommended maintenance strategy is retrieved from a preset database of multiple maintenance strategies based on the capacity decay mode identifier, and combined with the battery health status assessment score to generate a battery health maintenance plan that includes specific maintenance operations and execution priorities. This process includes: parsing the capacity decay mode identifier to determine the cause category with the highest probability value as the dominant decay mode; using the dominant decay mode as the index key to search the preset maintenance strategy database, which stores a list of typical maintenance operations corresponding to different cause categories; selecting maintenance operation suggestions that match the current battery type and usage environment from the retrieved list of typical maintenance operations to form a basic maintenance operation set; assigning an urgency weight to each operation in the basic maintenance operation set based on the numerical range of the battery health status assessment score, with a higher urgency weight for lower assessment scores; and dividing the battery health status assessment score range into multiple consecutive level intervals, each level interval corresponding to a standard urgency weight base; determining which level interval the current battery health status assessment score falls into and obtaining the standard urgency weight base corresponding to that interval. For each operation in the basic maintenance operation set, determine whether it is used to mitigate or repair the dominant degradation mode. If so, multiply the standard urgency weight base of the operation by a mode enhancement coefficient greater than one. Consider the current operating temperature of the battery. If the ambient temperature is higher or lower than the normal operating temperature range, multiply the urgency weight base of all operations by an environmental degradation coefficient. The weight value adjusted by the mode enhancement coefficient and the environmental degradation coefficient is used as the final urgency weight of the basic maintenance operation. Combining the estimated time and resource consumption of each maintenance operation and the assigned urgency weight, calculate the comprehensive execution priority score for each operation. Sort the operations in the basic maintenance operation set according to the comprehensive execution priority score from high to low, and attach a suggested execution time window for each operation to generate a structured battery health maintenance plan document.

[0110] In practice, based on the capacity decay mode identifier, the corresponding recommended maintenance strategy is retrieved from a pre-defined database of multiple maintenance strategies. Combined with the battery health status assessment score, a battery health maintenance plan is generated, including specific maintenance operations and their priorities. The capacity decay mode identifier is parsed to determine the category of the cause with the highest probability value as the dominant decay mode. For example, if the capacity decay mode identifier vector is [0.1, 0.7, 0.15, 0.05], corresponding to the probabilities of active lithium ion loss, electrode material structure degradation, excessive growth of the solid electrolyte interface film, and electrolyte drying, then the dominant decay mode is "electrode material structure degradation." Using the dominant decay mode as the index key, a search is performed in the pre-defined maintenance strategy database. This database stores a list of typical maintenance operations corresponding to different cause categories. The database is stored in key-value pairs, where the key is the cause category and the value is a list of recommended maintenance operations for that category. From the retrieved list of typical maintenance operations, maintenance operation suggestions that match the current battery type and usage environment are selected to form a basic maintenance operation set. For example, if the battery type is a lithium-ion ternary battery and the usage environment is an electric vehicle, the retrieved list of typical maintenance operations for "electrode material structure degradation" includes "performing a deep charge-discharge calibration cycle within a specific temperature window," "letting the battery rest in a constant temperature environment for 24 hours," and "performing a pulse charge repair procedure." "Performing a pulse charge repair procedure" and "performing a deep charge-discharge calibration cycle within a specific temperature window," which are suitable for the vehicle environment, are selected as the basic maintenance operation set. Based on the numerical range of the battery health status assessment score, each operation in the basic maintenance operation set is assigned an urgency weight; the lower the assessment score, the higher the urgency weight.

[0111] The specific method for assigning urgency weights is as follows: The range of battery health status assessment scores is divided into multiple consecutive level intervals. Each level interval corresponds to a standard urgency weight base. The level intervals are divided as follows: [0.9, 1.0] corresponds to a base of 1.0, [0.8, 0.9) corresponds to a base of 1.5, [0.7, 0.8) corresponds to a base of 2.0, [0.6, 0.7) corresponds to a base of 2.5, and [0.0, 0.6) corresponds to a base of 3.0. The specific level interval into which the current battery health status assessment score falls is determined, and the standard urgency weight base corresponding to that interval is obtained. For example, if the current battery health status assessment score is 0.82, falling into the [0.8, 0.9) interval, the standard urgency weight base is 1.5. For each operation in the basic maintenance operation set, it is determined whether it is used to alleviate or repair the dominant degradation mode. If so, the standard urgency weight base corresponding to the operation is multiplied by a mode enhancement coefficient greater than one. The mode enhancement coefficient is set to 1.2. For example, "execute pulse charging repair procedure" and "execute deep charge-discharge calibration cycle at a specific temperature window" in the basic maintenance operation set are both used to alleviate "electrode material structure degradation", so both are multiplied by 1.2 to get 1.8. Considering the current operating temperature of the battery, if the ambient temperature is higher or lower than the normal operating temperature range, the urgency weight base of all operations is multiplied by an environmental degradation coefficient. The normal operating temperature range is defined as 15 degrees Celsius to 35 degrees Celsius. The environmental degradation coefficient is set to 1.1 when the temperature is higher than 35 degrees Celsius and 1.05 when the temperature is lower than 15 degrees Celsius. For example, if the current ambient temperature is 40 degrees Celsius, the urgency weight base of all operations needs to be multiplied by 1.1, and the weights of the aforementioned two operations are adjusted from 1.8 to 1.98. The weighted values, adjusted for mode enhancement and environmental degradation coefficients, are used as the final urgency weights for basic maintenance operations. In some embodiments, the division of the level intervals and the corresponding base values ​​can be adjusted according to the actual application strategy, and the specific values ​​of the mode enhancement and environmental degradation coefficients can also be configured. Optionally, the calculation of the urgency weights can be completed in real time in the edge computing module of the battery management system.

[0112] By combining the estimated time and resource consumption scores of each maintenance operation, along with their assigned urgency weights, a comprehensive execution priority score is calculated for each operation. Both the estimated time and resource consumption scores are dimensionless values ​​normalized based on historical experience; lower scores indicate shorter execution time or lower resource consumption. The calculation formula is:

[0113] ;

[0114] in: This represents the overall execution priority score of the i-th operation. This represents the urgency weight assigned to the i-th operation. This represents the estimated time score for the i-th operation. This represents the estimated resource consumption score for the i-th operation. The formula calculates priority by multiplying an urgency weight by an adjustment factor based on time and resource consumption. The adjustment factor ensures that the lower the time and resource consumption score (i.e., the easier the operation is to execute), the higher the priority score. The higher the value, for example, the urgency weight of "performing a pulse charging repair procedure" is 1.98, the estimated time score is 0.3, and the estimated resource consumption score is 0.4, resulting in a comprehensive execution priority score of 1.98*(1+1 / (0.3+0.4))≈4.80. Operations in the basic maintenance operation set are sorted according to their comprehensive execution priority scores from highest to lowest, and a suggested execution time window is attached for each operation, generating a structured battery health maintenance plan document. The suggested execution time window is set based on idle periods in the vehicle usage log; for example, "performing a pulse charging repair procedure" is suggested to be performed between 2 AM and 4 AM the following day when the vehicle is parked. The structured document encapsulates the operation list, priority score, and suggested execution time window in JSON or XML format. In some embodiments, the estimated time score and resource consumption score can be mapped to scalar values ​​between 0 and 1 based on a preset lookup table. Optionally, the maintenance strategy library can be stored on a cloud server and retrieved and updated via a network interface. It can be understood that the comprehensive execution priority score quantifies the balance between the urgency and execution cost of maintenance operations. In practice, the basic maintenance operation set and the calculated comprehensive execution priority score can be displayed through an onboard human-machine interface or transmitted to a remote monitoring center via a wireless network. Optionally, after the basic maintenance operation set is completed, its execution record and battery status feedback can be used to update the operation effectiveness data in the maintenance strategy library.

[0115] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An online monitoring method based on battery health status, characterized in that, include: Acquire multi-dimensional operational data of the target battery during charge-discharge cycles, including voltage change curves, current change curves, temperature change curves, and surface strain change curves. The voltage change curve, current change curve, temperature change curve, and surface strain change curve are time-synchronized and feature-synchronized extracted to generate a set of battery state feature vectors. The feature-synchronized extraction includes synchronously identifying key charge and discharge feature points and extracting multi-dimensional time-series features, including strain accumulation, based on the key charge and discharge feature points. The set of battery state feature vectors is input into the trained battery health status assessment network for processing, and the battery health status assessment score and capacity decay pattern identifier in vector form are output to indicate the dominant cause category of capacity decay; the battery health status assessment network adopts an improved convolutional neural network architecture with causal constraints. The battery health status assessment score, capacity decay mode identifier, and obtained offline capacity calibration results are input into the online model corrector to dynamically correct the battery health status assessment network. Based on the capacity decay mode identifier, the corresponding recommended maintenance strategy is retrieved from a preset library of multiple maintenance strategies. Combined with the battery health status assessment score, a battery health maintenance plan is generated that includes specific maintenance operations and execution priorities. When generating the maintenance plan, maintenance operations targeting the dominant cause category are assigned higher execution priority weights.

2. The online monitoring method based on battery health status according to claim 1, characterized in that, The voltage change curve, current change curve, temperature change curve, and surface strain change curve are time-synchronized and feature-synchronized extracted to generate a set of battery state feature vectors, including: A unified timestamp reference is configured for the voltage change curve, current change curve, temperature change curve and surface strain change curve. The start time of the charge and discharge cycle is taken as the zero point, and the time axis of all curves is aligned to eliminate the phase difference caused by the asynchronous sampling time of different sensors. On each curve after time axis alignment, instantaneous observations are extracted at fixed sampling time intervals to form voltage, current, temperature and strain sequences. Key charge and discharge feature points are simultaneously identified and located on the voltage sequence, current sequence, temperature sequence, and strain sequence. These key charge and discharge feature points include the constant current charging to constant voltage charging point, the charging termination voltage point, and the discharging termination voltage point. Taking each key charging and discharging characteristic point as the center, expand the data window to the front and back with a preset width respectively. Extract the average voltage and voltage range in the data window of the voltage change curve, extract the average current and current ripple coefficient in the data window of the current change curve, extract the temperature rise rate in the data window of the temperature change curve, and extract the cumulative strain in the data window of the surface strain change curve. After normalizing the average voltage, voltage range, average current, current ripple coefficient, temperature rise rate, and strain accumulation corresponding to the same key charge and discharge characteristic point, they are merged into a multi-dimensional feature vector. The multi-dimensional feature vectors corresponding to all key charge and discharge characteristic points are arranged in chronological order to form the battery state feature vector set.

3. The online monitoring method based on battery health status according to claim 2, characterized in that, The synchronous identification and location of key charging and discharging feature points includes: On the current change curve, locate the inflection point where the current value starts to decrease from the constant current value, mark the inflection point as the point where constant current charging turns into constant voltage charging, and record the timestamp corresponding to the inflection point and the voltage value at this time. On the voltage change curve, locate the moment when the voltage value first reaches the preset charging cutoff voltage threshold, mark the moment as the charging termination voltage point, and record the current value at that moment. On the voltage change curve, the moment when the voltage value first drops to the preset discharge cutoff voltage threshold is marked as the discharge termination voltage point, and the current value at that moment is recorded. When the slope of the temperature change curve changes during the charging process, the change point is marked as a temperature anomaly inflection point, and its timestamp and temperature value are recorded. All identified key charging and discharging feature points, including the constant current charging to constant voltage charging point, the charging termination voltage point, the discharging termination voltage point, and the temperature anomaly inflection point, are sorted according to the time stamp order to form a feature point sequence with time sequence and type labels.

4. The online monitoring method based on battery health status according to claim 1, characterized in that, The battery health status assessment network includes a feature encoder, a temporal correlation analyzer, and a health status decoder. The feature encoder performs point-by-point feature transformation and dimensionality upscaling on the input set of battery state feature vectors to extract a high-dimensional abstract representation of each feature point. The temporal correlation analyzer is based on an improved convolutional neural network architecture to process the high-dimensional abstract representation of the sequence. The improved convolutional neural network architecture is based on the causal dependency of the battery state feature vector in the time dimension and adopts a convolutional kernel structure with causal constraints to ensure that the network does not use feature information from future times when processing features at the current time. The time-series correlation analyzer outputs a context encoding vector that characterizes the overall state evolution trend of the battery; The health status decoder outputs a scalar battery health status assessment score and a vector capacity decay pattern identifier based on the context encoding vector. The capacity decay pattern identifier is used to indicate the main cause category of capacity decay.

5. The online monitoring method based on battery health status according to claim 4, characterized in that, The improved convolutional neural network architecture includes: Construct a set of one-dimensional causal convolution kernels. The convolution calculation range of the causal convolution kernels is strictly limited to the feature data of the current time and the historical time. The feature data of the future time do not participate in the output calculation of the current time. In the time-series correlation analyzer, multiple causal convolutional layers composed of the causal convolutional kernels are stacked, and the output of each causal convolutional layer depends only on the data of the current and previous time steps in its input; An extended causal convolutional structure with an adjustable dilation factor is introduced between adjacent causal convolutional layers. By adjusting the dilation factor, higher-level causal convolutional layers can capture dependencies over a longer temporal range with an exponentially expanded receptive field, while maintaining causal constraints. After each causal convolution operation, a gated activation mechanism is introduced. This mechanism uses a gate value generated by a sigmoid function to control the proportion of feature information passed to the next layer, thereby enabling the filtering of features of different importance. At the end of the temporal correlation analyzer, global temporal pooling is performed on the feature sequences after multi-layer causal convolution and gated activation processing, aggregating the variable-length feature sequences into a fixed-length context encoding vector.

6. The online monitoring method based on battery health status according to claim 1, characterized in that, The battery health status assessment score, capacity degradation mode identifier, and multi-dimensional operational data are input into an online model corrector to dynamically correct the battery health status assessment network, including: Collect historical multi-dimensional operating data of the target battery within a predetermined time window, as well as historical battery health status assessment scores and historical capacity degradation pattern identifiers output by the battery health status assessment network within the corresponding time window. When a complete offline battery capacity calibration result is obtained, the offline battery capacity calibration result is used as the true value reference of the battery health status at the current moment. The true reference is compared with the battery health status assessment score output by the battery health status assessment network at the current time, and the assessment error is calculated. Based on the magnitude and direction of the assessment error, and the evolution trend of historical capacity decay pattern identifiers, a fine-tuning gradient for the internal parameters of the battery health status assessment network is generated. A conservative online learning strategy is adopted, in which the fine-tuning gradient is superimposed and updated onto the trainable parameters of the battery health status assessment network at a learning rate less than a preset upper limit, thereby achieving progressive dynamic correction of the model.

7. The online monitoring method based on battery health status according to claim 6, characterized in that, The conservative online learning strategies include: Maintain a model parameter update cache for the online model corrector to temporarily store the calculated fine-tuned gradients; Set a gradient accumulation period. Within the gradient accumulation period, perform a moving average calculation on the fine-tuned gradients obtained from multiple consecutive evaluations in the update buffer to obtain the smoothed average fine-tuned gradient. A decay function is set for the learning rate of the model, which is related to the number of battery cycles. As the number of cycles increases, the learning rate gradually decreases. Each time a parameter update is performed, the smoothed average fine-tuning gradient is multiplied by the learning rate at the current time to obtain the final actual update amount applied to the model parameters; The actual update amount is applied to the trainable parameters of the battery health status assessment network to complete an online correction, and the model parameter update cache is cleared to prepare for the next gradient accumulation cycle.

8. The online monitoring method based on battery health status according to claim 4, characterized in that, The capacity decay pattern identifier is used to indicate the main cause category of capacity decay, and its generation process includes: The health status decoder includes a multi-branch classifier, with each branch corresponding to a preset category of capacity decay causes. These causes include loss of active lithium ions, degradation of electrode material structure, excessive growth of solid electrolyte interface film, and electrolyte drying. The context encoding vector is input to the multi-branch classifier, and a class activation value is calculated on each classification branch; The activation values ​​of each branch are normalized and converted into a probability distribution. The probability value of each branch represents the confidence that the corresponding cause category is the main cause of the current battery capacity degradation. The normalized probability distribution vector is output as the capacity decay mode identifier. The capacity decay mode identifier is a multi-dimensional vector whose dimension is equal to the preset number of cause categories. The value of each element in the vector represents the probability of the corresponding cause category.

9. The online monitoring method based on battery health status according to claim 1, characterized in that, Based on the capacity decay mode identifier, a corresponding recommended maintenance strategy is retrieved from a preset library of multiple maintenance strategies. Combined with the battery health status assessment score, a battery health maintenance plan is generated, including specific maintenance operations and execution priorities, comprising: The capacity decay mode identifier is analyzed to determine the cause category with the highest probability value as the dominant decay mode; Using the dominant decay mode as the index key, a search is performed in a preset maintenance strategy library, which stores a list of typical maintenance operations corresponding to different cause categories. From the retrieved list of typical maintenance operations, select maintenance operation suggestions that match the current battery type and usage environment to form a basic maintenance operation set; Based on the numerical range of the battery health status assessment score, each operation in the basic maintenance operation set is assigned an urgency weight; the lower the assessment score, the higher the urgency weight. By combining the estimated time and resource consumption of each maintenance operation with the urgency weight, a comprehensive execution priority score for each operation is calculated. Based on the comprehensive execution priority score from high to low, the operations in the basic maintenance operation set are sorted, and a suggested execution time window for each operation is attached to generate a structured battery health maintenance plan document. Assign an urgency weight to each operation in the basic maintenance operation set, specifically including: The range of battery health status assessment scores is divided into multiple consecutive level intervals, with each level interval corresponding to a standard urgency weight base. Determine which grade range the current battery health status assessment score falls into, and obtain the standard urgency weight base corresponding to the range; For each operation in the basic maintenance operation set, determine whether it is used to alleviate or repair the dominant decay mode. If so, multiply the standard urgency weight base corresponding to the operation by a mode enhancement coefficient greater than one. Considering the current operating temperature of the battery, if the ambient temperature is higher or lower than the normal operating temperature range, the urgency weight of all operations is multiplied by an environmental degradation coefficient. The weight values ​​adjusted by the mode enhancement coefficient and the environmental degradation coefficient will be used as the final urgency weights for basic maintenance operations.

10. An online monitoring system based on battery health status, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the online monitoring method based on battery health status as described in any one of claims 1 to 9.

Images