A method and device for predicting the remaining useful life of a lithium battery

Abstract

The application relates to a lithium battery residual service life prediction method and device. The lithium battery residual service life prediction method comprises the following steps: characteristic expansion is carried out on the voltage, current and surface temperature of a battery to obtain electric-thermal coupling data of the battery, and cycle position coding preprocessing is carried out on the electric-thermal coupling data to obtain an electric-thermal coupling embedded vector of the battery; then, a double-channel attention feature extraction is carried out on the electric-thermal coupling embedded vector of the battery through a backbone network based on double-channel attention fusion to obtain a double-channel fusion feature tensor of the battery; finally, battery life prediction is carried out on the double-channel fusion feature tensor of the battery to obtain a residual life prediction value of the current battery. The lithium battery residual service life prediction method can significantly improve the accuracy and real-time performance of battery residual life prediction.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery management technology, and in particular to a method, apparatus, system, electronic device, and computer storage medium for predicting the remaining lifespan of a lithium battery. Background Technology

[0002] Lithium-ion batteries, also known as lithium-ion batteries, are currently important energy storage devices widely used in mobile phones, laptops, electric vehicles, and energy storage power stations. As these applications continue to expand, the safety and reliability of lithium-ion batteries have become increasingly prominent. To optimize the safety of battery management systems, ensure the stability of battery operation, and extend battery life, scientific prediction of lithium-ion battery life is necessary. The main tasks of this scientific prediction include assessing the State of Health (SOH) and predicting the Remaining Useful Life (RUL). Through joint assessment of SOH and RUL, the battery management system can comprehensively reflect the degree of battery degradation and its health status, thus providing crucial information for battery safety control, maintenance strategies, and lifespan management.

[0003] Traditional battery state health assessments and remaining battery life predictions typically employ physicochemical modeling methods and data-driven machine learning methods. Physicochemical models primarily include equivalent circuit models and single-particle models. The equivalent circuit model approximates the battery's electrochemical characteristics using circuit elements such as resistors and capacitors, thus describing the battery's charge and discharge behavior. The single-particle model characterizes the battery's electrochemical kinetics based on the transport and reaction mechanisms of lithium ions in a single electrode particle. However, while physicochemical modeling methods offer good interpretability, their parameter calibration process is complex, often requiring experimental measurements and calibrations for each individual battery. This process is time-consuming and costly, hindering large-scale application.

[0004] Purely data-driven machine learning methods, such as Gaussian process regression, random forests, and long short-term memory networks (LSTM), can train models directly on large amounts of experimental data without introducing physical parameters, achieving high-precision predictions of SOH and RUL. However, these methods are highly dependent on the quantity and quality of training data, are easily affected by experimental noise, and lack generalization ability in prediction tasks spanning multiple cycles and lifetimes.

[0005] Based on this, existing technologies introduce physical constraints and attention mechanisms into deep learning models to fuse multi-channel information such as battery current and voltage, thereby improving the accuracy of SOH assessment or RUL prediction. This is mainly achieved by stacking multi-layer deep learning network structures to learn temporal dependencies across cycles and lifetime stages, and by utilizing attention mechanisms (weighting) to enhance model interpretability, making the capture of battery degradation trends more accurate. However, existing methods generally use simple weighting based on or directly employing self-attention during channel fusion, without deeply exploring electro-thermal coupling (the relationship between charging voltage, current, and battery temperature) and micro-degradation characteristics, resulting in insufficient information utilization. Simultaneously, the stacking of multi-layer network structures leads to model bloat and computational redundancy, resulting in high inference latency, making it difficult to meet the needs of battery management systems for online diagnosis and rapid decision-making.

[0006] In summary, existing technologies suffer from technical problems such as coarse multi-channel fusion strategies, excessive model complexity, and high inference latency when performing SOH assessments or RUL predictions. Summary of the Invention

[0007] Therefore, the purpose of this invention is to provide a method for predicting the remaining lifespan of a lithium battery.

[0008] A method for predicting the remaining lifespan of a lithium battery includes the following steps:

[0009] S1: Feature extension of the battery's voltage, current, and surface temperature to obtain the battery's electrothermal coupling data;

[0010] The feature is extended to obtain the instantaneous charge and discharge capacity by integrating the battery current, and to perform first-order and second-order finite difference calculations on the battery voltage and corresponding capacity respectively using the finite difference method to obtain the first-order and second-order derivatives of the voltage-capacity relationship. The battery current, voltage, temperature, instantaneous charge and discharge capacity, and the first-order and second-order derivatives of the voltage-capacity relationship are combined to obtain the battery's electrothermal coupling data.

[0011] S2: Perform cyclic position encoding preprocessing on the battery's electrothermal coupling data to obtain the battery's electrothermal coupling embedding vector;

[0012] The cyclic position encoding preprocessing includes the following sub-steps:

[0013] First, the electrothermal coupling data of the battery is subjected to linear transformation and layer normalization operations in sequence to obtain the normalized embedded representation;

[0014] Simultaneously, cyclic position encoding is performed on each sampling moment of the battery's electrothermal coupling data to obtain the cyclic position encoding vector corresponding to each sampling moment;

[0015] Finally, the normalized embedding representation is fused with the cyclic position encoding vector to obtain the battery's electrothermal coupling embedding vector;

[0016] S3: Perform several dual-channel attention feature extractions on the battery's electrothermal coupling embedding vector to obtain the battery's dual-channel fusion feature tensor.

[0017] S4: Perform battery lifetime prediction on the dual-channel fused feature tensor of the battery to obtain the current remaining battery lifetime prediction value.

[0018] The lithium battery remaining life prediction method of the present invention, compared with the prior art, effectively improves the fusion efficiency of multi-channel information by expanding the electrothermal coupling features of current, voltage and surface temperature, combining cyclic position encoding preprocessing and multiple dual-channel attention feature extractions, and performing battery life prediction on this basis. This significantly improves the accuracy and real-time performance of battery remaining life prediction.

[0019] In addition, this invention explicitly incorporates the charge-discharge cycle number into the feature embedding through cyclic position encoding, enabling the features of different cycle stages to be separated in the encoding space. This avoids the feature aliasing problem caused by the superposition of similar charge-discharge rate segments in cross-cycle prediction tasks, effectively enhancing the model's ability to characterize the lifetime evolution law and significantly improving the cross-cycle generalization performance of battery remaining life prediction.

[0020] Furthermore, the dual-channel attention feature extraction is performed by a lightweight dual-channel attention module, and each layer of the lightweight dual-channel attention module includes a cross-channel attention submodule and a local convolutional feedforward submodule;

[0021] The cross-channel attention submodule is used to perform layer normalization on the current dual-channel fusion feature tensor of the battery to obtain the normalized dual-channel fusion feature tensor.

[0022] If the current dual-channel fusion feature tensor of the battery is the first layer, then it is the battery's electrothermal coupling embedding vector.

[0023] Next, the embedding representations belonging to the voltage channel in the normalized dual-channel fused feature tensor are queried and mapped to obtain the query matrix;

[0024] Meanwhile, the embedding representations belonging to the current channel and the surface temperature channel in the normalized dual-channel fusion feature tensor are concatenated, and the concatenated embedding representations are mapped to the key matrix and the value matrix, respectively.

[0025] Then, based on the similarity weights between the query matrix and the key matrix, the value matrix is ​​weighted and summed to obtain the cross-channel fused attention representation. The cross-channel fused attention representations corresponding to all attention heads are then concatenated to obtain the attention weight vector.

[0026] Finally, the current attention weight vector is fused with the current dual-channel fusion feature tensor of the battery to obtain the current cross-channel fusion representation.

[0027] Accordingly, this invention uses voltage channel embedding representation as the query vector and current and temperature channel embedding representation as keys and values ​​for cross-attention calculation, thereby embodying the physical mechanism assumption of "potential first". This allows the voltage signal to play a dominant role in the fusion process. At the same time, by combining the mechanism supplementary information of current and temperature, it can highlight the correspondence between voltage curve degradation signs and operating conditions, thereby effectively improving the feature interaction efficiency of the model and the interpretability of the prediction results.

[0028] Furthermore, the local convolutional feedforward submodule is used to perform layer normalization on the cross-channel fusion representation to stabilize the distribution of different channels and obtain the normalized cross-channel fusion representation.

[0029] Next, the normalized cross-channel fusion representation is subjected to depthwise convolution and linear transformation to obtain local convolutional features;

[0030] Simultaneously, point convolution is performed on the cross-channel fusion representation to obtain point convolution features;

[0031] Then, the local convolutional features are multiplied element-wise with the point convolutional features to obtain the local feature vector of charging and discharging.

[0032] Finally, the current cross-channel fusion representation is fused with the current charge / discharge local feature vectors to obtain the dual-channel fusion feature tensor of the battery in the next layer.

[0033] Accordingly, this invention introduces a local convolutional feedforward submodule based on depthwise separable convolution and point convolution by making lightweight adjustments to the existing feedforward network structure (FFN). This enables efficient capture of local degradation features of adjacent sampling points during the charging and discharging process. Furthermore, by multiplying elements one by one, complementary fusion of local and global features is achieved. This significantly reduces the number of model parameters and computational overhead while enhancing the model's ability to model fine-grained degradation features such as micro-discharge plateaus and temperature fluctuations. Consequently, it effectively improves the real-time performance and accuracy of battery life prediction.

[0034] A lithium battery remaining life prediction device includes a charge / discharge data feature expansion unit, a cycle position encoding preprocessing unit, a backbone network based on dual-channel attention fusion, and a battery life prediction unit.

[0035] The charge / discharge data feature extension unit is used to extend the voltage, current and surface temperature of the battery to obtain the battery's electrothermal coupling data.

[0036] The feature is extended to obtain the instantaneous charge and discharge capacity by integrating the battery current, and to perform first-order and second-order finite difference calculations on the battery voltage and corresponding capacity respectively using the finite difference method to obtain the first-order and second-order derivatives of the voltage-capacity relationship. The battery current, voltage, temperature, instantaneous charge and discharge capacity, and the first-order and second-order derivatives of the voltage-capacity relationship are combined to obtain the battery's electrothermal coupling data.

[0037] The cyclic position encoding preprocessing unit is used to perform cyclic position encoding preprocessing on the electrothermal coupling data of the battery to obtain the electrothermal coupling embedding vector of the battery.

[0038] The backbone network based on dual-channel attention fusion is used to perform several dual-channel attention feature extractions on the battery's electrothermal coupling embedding vector to obtain the battery's dual-channel fusion feature tensor.

[0039] The battery life prediction unit is used to predict the battery life of the battery by performing battery life prediction on the dual-channel fused feature tensor, and to obtain the predicted value of the remaining life of the current battery.

[0040] Furthermore, the lithium battery remaining life prediction device also includes a data acquisition module, which is equipped with a voltage sensor, a current sensor and a temperature sensor;

[0041] The voltage sensor is used to collect the battery voltage, obtain the battery voltage data, and transmit it to the charge / discharge data feature extension unit.

[0042] The current sensor is used to collect the battery current, obtain the battery current data, and transmit it to the charge / discharge data feature extension unit.

[0043] The temperature sensor is used to collect the surface temperature of the battery, obtain the surface temperature data of the battery, and transmit it to the charge / discharge data feature extension unit.

[0044] A lithium battery remaining life prediction system includes a device to be tested and the lithium battery remaining life prediction device described above.

[0045] The device under test is equipped with a lithium battery and a data acquisition unit connected to the lithium battery;

[0046] The data acquisition unit connected to the lithium battery is used to acquire voltage, current and surface temperature data during the charging and discharging process of the lithium battery, and transmit the voltage, current and surface temperature data to the lithium battery remaining life prediction device via wired or wireless means to predict the remaining life of the lithium battery.

[0047] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0048] Figure 1 This is a simplified structural diagram of the lithium battery remaining life prediction device described in this invention;

[0049] Figure 2 This is a simplified flowchart illustrating the lithium battery remaining lifespan prediction method described in this invention.

[0050] Figure 3 A comparative diagram showing the experimental results of predicting the remaining battery life for four sets of battery data. Detailed Implementation

[0051] To address the technical problems of coarse multi-channel fusion strategies, excessive model complexity, and high inference latency in existing technologies for SOH assessment or RUL prediction, this invention expands the features of battery voltage, current, and surface temperature to obtain battery electrothermal coupling data, and performs cyclic position encoding preprocessing to obtain the battery's electrothermal coupling embedding vector. Next, a backbone network based on dual-channel attention fusion is used to perform several dual-channel attention feature extractions on the battery's electrothermal coupling embedding vector to obtain the battery's dual-channel fused feature tensor. Finally, the battery's dual-channel fused feature tensor is used to predict battery lifetime, yielding the current predicted remaining battery lifetime.

[0052] Accordingly, this invention effectively improves the characterization capability and fusion efficiency of multi-channel battery information through electrothermal coupling feature extension, cyclic position encoding preprocessing, and backbone network design based on dual-channel attention fusion, thereby significantly improving the accuracy and real-time performance of battery life prediction.

[0053] Based on the above design, this invention proposes a method for predicting the remaining lifespan of a lithium battery, and based on this method, proposes a device for predicting the remaining lifespan of a lithium battery.

[0054] A lithium battery remaining life prediction system includes a device to be tested and the lithium battery remaining life prediction device.

[0055] The device under test is equipped with a lithium battery and a data acquisition unit connected to the lithium battery.

[0056] The device to be tested can be a device or system equipped with a lithium battery, such as an electric vehicle, an energy storage power station, a power tool, a drone, an electric bicycle, or a portable electronic device. It is used to charge and discharge the lithium battery under actual operating conditions and to collect data synchronously during the charging and discharging process. This invention does not limit the device to be tested.

[0057] The data acquisition unit connected to the lithium battery is used to acquire voltage, current and surface temperature data during the charging and discharging process of the lithium battery, and transmit the voltage, current and surface temperature data to the lithium battery remaining life prediction device via wired or wireless means to predict the remaining life of the lithium battery.

[0058] The data acquisition unit may include a voltage sensor, a current sensor, and a temperature sensor.

[0059] The voltage sensor is used to collect the battery voltage, obtain the battery voltage data, and transmit it to the lithium battery remaining life prediction device.

[0060] The current sensor is used to collect the battery current, obtain the battery current data, and transmit it to the lithium battery remaining life prediction device.

[0061] The temperature sensor is used to collect the surface temperature of the battery, obtain the surface temperature data of the battery, and transmit it to the lithium battery remaining life prediction device.

[0062] Please also refer to Figure 1 and Figure 2 , Figure 1 This is a simplified structural diagram of the lithium battery remaining life prediction device described in this invention. Figure 2 This is a simplified flowchart illustrating the lithium battery remaining lifespan prediction method described in this invention.

[0063] The lithium battery remaining life prediction device includes a charge / discharge data feature extension unit 1, a cycle position encoding preprocessing unit 2, a backbone network based on dual-channel attention fusion 3, and a battery life prediction unit 4.

[0064] The charge / discharge data feature extension unit 1 is used to perform step S1: extend the features of the battery's voltage, current and surface temperature to obtain the battery's electrothermal coupling data.

[0065] Specifically, the feature is extended to: integrating the battery current to obtain the instantaneous charge / discharge capacity; and using the finite difference method to perform first-order and second-order finite difference calculations on the battery voltage and corresponding capacity respectively, obtaining the first and second derivatives of the voltage-capacity relationship. The battery current, voltage, temperature, instantaneous charge / discharge capacity, and the first and second derivatives of the voltage-capacity relationship are then combined to obtain the battery's electrothermal coupling data, which is specifically represented as follows:

[0066]

[0067] In the formula, Indicates the first The electrothermal coupling data of the battery at the sampling time, its Indicates the first The battery current value at the sampling time is used to calculate the battery current at the first sampling time. The charge and discharge intensity at the sampling time; Indicates the first The voltage value at each sampling time is used to characterize the battery's potential state; It represents the surface temperature of the battery and is used to characterize the heat release and conduction effects during the charging and discharging process;

[0068] Indicates the first The instantaneous charge and discharge capacity at the sampling moment is used to characterize the coulomb quantity transferred by the cell within an extremely short time window. Its specific calculation is expressed as follows:

[0069]

[0070] In the formula, This represents the time interval between adjacent sampling points, used for integral approximation of capacity increments, and can be determined based on the sensor's sampling accuracy. This invention does not... The specific values ​​are strictly limited, as long as the time frame for sampling current, voltage and temperature is consistent.

[0071] The first derivative, representing the voltage-capacity relationship, characterizes the rate of change of potential with capacity during battery charging and discharging. It can reveal the drift characteristics of the voltage plateau during degradation. Its specific calculation is expressed as follows:

[0072]

[0073] In the formula, and They represent in and The voltage value at the sampling time, The span of the difference window is used for finite difference calculations; and They represent in and The capacity at each sampling time;

[0074] The second derivative, representing the voltage-capacity relationship, is used to characterize the voltage microstructure of the battery at the phase transition-potential plateau and the irreversible lithiumization reaction. It can effectively reflect the irreversible reaction and degradation mechanism. Its specific calculation is expressed as follows:

[0075]

[0076] It should be noted that the current of the aforementioned battery... ,Voltage With surface temperature The data acquisition is based on the battery's charge-discharge cycle and is recorded at fixed sampling intervals within a unified time frame. Each sampling index... This corresponds to a sampling moment in a charge-discharge cycle; and the data collected during the charge-discharge cycle of this battery can comprehensively reflect the electrochemical and thermal changes of the battery under actual operating conditions, including both the macroscopic characteristics of capacity transfer and the microscopic features of voltage microstructure and temperature response.

[0077] Therefore, through the electrothermal coupling data of the battery A data vector, or data vector, can simultaneously describe the potential changes, capacity transfer, and heat accumulation of a battery during charging and discharging, thus forming a comprehensive feature characterizing the electro-thermal coupling effect.

[0078] It should be further noted that the present invention does not specifically limit the sampling method and data source. The above-mentioned feature extension can be achieved by using voltage sensors, current sensors and temperature sensors to obtain synchronous data such as current, voltage and temperature of the battery during charging and discharging.

[0079] In some embodiments, the electrothermal coupling data of the battery can also be used to calculate the local temperature rise characteristics through the surface temperature of the battery, thereby obtaining local temperature rise characteristics, and combining them with the electrothermal coupling data of the battery to obtain extended electrothermal coupling data of the battery. The specific calculation representation of its local temperature rise characteristics is as follows:

[0080]

[0081] In the formula, Indicates the first The temperature difference between the sampling time and the previous sampling interval is used to capture the temperature accumulation rate, thereby further enhancing the model's ability to characterize the electro-thermal coupling failure mechanism.

[0082] The cyclic position encoding preprocessing unit 2 is used to perform step S2: perform cyclic position encoding preprocessing on the electrothermal coupling data of the battery to obtain the electrothermal coupling embedding vector of the battery.

[0083] Specifically, the cyclic position encoding preprocessing includes the following sub-steps:

[0084] First, the electrothermal coupling data of the battery is subjected to linear transformation and layer normalization operations in sequence to obtain the normalized embedding representation, which is as follows:

[0085]

[0086] In the formula, Indicates the first The normalized embedding representation at the sampling time is used to ensure that all features have the same dimensions and differentiable learning properties; and Let represent the trainable linear weight matrix and bias matrix, respectively; The representation layer normalization operation is used to suppress the distribution bias between different feature channels.

[0087] Simultaneously, cyclic position encoding is performed on each sampling moment of the battery's electrothermal coupling data to obtain the cyclic position encoding vector corresponding to each sampling moment, which is specifically represented as follows:

[0088]

[0089] In the formula, Indicates the first Sampling time and charge / discharge cycle number are The cyclic position encoding vector is used to ensure that even if similar charge and discharge rates occur within different cycles, they can be distinguished in the encoding space. Its essence is: This represents the fundamental frequency, used to determine the basic period of the sine and cosine functions, ensuring the continuity of the timing coding. It indicates the sequence number of the charge / discharge cycle, used to characterize the cycle progress of the battery during its lifespan. Indicates the electrical cycle number The corresponding smoothing function is used to map different loops into a separable coding space.

[0090] Finally, the normalized embedding representation is fused with the cyclic position encoding vector to obtain the battery's electrothermal coupling embedding vector, the specific representation of which is as follows:

[0091]

[0092] In the formula, Indicates the first The charge / discharge cycle number at the sampling time is The battery's electrothermal coupling embedding vector is used as the subsequent input token; Indicates the first The charge / discharge cycle number at the sampling time is The normalized embedding representation is used to provide a standardized representation of the electro-thermal coupling characteristics; This represents a trainable positional encoding fusion projection matrix, used to map positional information and feature signals to the same latent space, so that deep networks can utilize both temporal and physical features simultaneously.

[0093] It should be noted that the current, voltage, and surface temperature of the battery mentioned above are all sampled based on the charge-discharge cycle process, and each cycle of data corresponds to a charge-discharge cycle number. This allows the battery's lifespan stages to be characterized by these serial numbers. Furthermore, without cyclic position encoding, segments with similar charge / discharge rates in different cycles will overlap in the encoding space, making it difficult for the model to distinguish the degradation characteristics of the battery at different lifespan stages, thus affecting the accuracy of lifespan prediction. Therefore, this invention employs cyclic position encoding to preserve the temporal order within a cycle, enabling separation across cycle dimensions and allowing the model to better capture lifespan evolution patterns.

[0094] Furthermore, by introducing a trainable linear weight matrix and bias vector during the embedding mapping process, and a learnable projection matrix during the fusion process, the original physical features and positional features can be unified into a continuous and differentiable embedding vector. This enables the battery's electrothermal coupling embedding vector to have a unified dimension and reflect the electrothermal coupling features and cyclic temporal features, thereby providing high-quality input for subsequent deep feature extraction based on the attention mechanism.

[0095] The backbone network 3 based on dual-channel attention fusion is used to perform step S3: extract dual-channel attention features several times from the electrothermal coupling embedding vector of the battery to obtain the dual-channel fusion feature tensor of the battery.

[0096] Specifically, the backbone network based on dual-channel attention fusion includes several layers of stacked lightweight dual-channel attention modules. Each layer of lightweight dual-channel attention modules includes a cross-channel attention submodule and a local convolutional feedforward submodule.

[0097] The cross-channel attention submodule is used to process the dual-channel fusion feature tensor of the current battery. Layer normalization is performed to stabilize the distribution of different channels, thereby avoiding interference from numerical imbalance in attention calculation, resulting in a normalized dual-channel fusion feature tensor. ;

[0098] Wherein, if the current battery's dual-channel fused feature tensor is the first... The layer is the electrothermal coupling embedding vector of the battery, i.e. ,That This represents the sequence of electrothermal coupling embedding vectors at all sampling times;

[0099] Next, the embedding representations belonging to the voltage channel in the normalized dual-channel fused feature tensor are used for query mapping to obtain the query matrix, the specific calculation of which is as follows:

[0100]

[0101] In the formula, The current number The query matrix of the layer is used to calculate the correlation between the voltage channel and other channels; Indicates the current number The normalized dual-channel fused feature tensor of the layer; This represents a trainable linear projection matrix used to extract query features from voltage channels;

[0102] Simultaneously, the embedding representations belonging to the current channel and surface temperature channel in the normalized dual-channel fused feature tensor are concatenated, and the concatenated embedding representations are mapped to the key matrix and value matrix respectively. The specific calculation representation is as follows:

[0103]

[0104] In the formula, This represents the key matrix, which is used to calculate similarity with the query matrix; Indicates the first Embedded representation of current channels in the normalized dual-channel fused feature tensor of the layer Embedded representation of surface temperature channels The embedded representation after concatenation; This represents a value matrix used for weighted combination based on attention weights; and These are trainable linear mapping matrices for the key matrix and the value matrix, respectively.

[0105] Then, based on the similarity weights between the query matrix and the key matrix, the value matrix is ​​weighted and summed to obtain the cross-channel fused attention representation. The cross-channel fused attention representations corresponding to all attention heads are then concatenated to obtain the attention weight vector. For the ... The attention weight vector for the layer can be found in the following expression:

[0106]

[0107] In the formula, For the first The attention weight vector of the layer; This indicates a concatenation operation along the channel dimension; This represents a trainable projection matrix used to map the concatenated multi-head attention results back to a unified dimension; Indicates the first The first layer To the The cross-channel fused attention representation of the attention head output, for the . The attention representation of the cross-channel fusion of the attention head outputs can be found in the following expression:

[0108]

[0109] In the formula, Indicates the first The first layer A cross-channel fused attention representation of the attention head output is used to capture the interaction between the voltage channel and the current-temperature channel; This represents the dimension of the key vector, used to scale the similarity to maintain numerical stability; This represents the normalization operation, used to map similarity to a weight distribution, i.e., the normalization exponential function;

[0110] Finally, the current attention weight vector is fused with the current dual-channel fusion feature tensor of the battery to obtain the current cross-channel fusion representation, which can be referred to in the following expression:

[0111]

[0112] In the formula, Indicates the first The fusion representation of layers across channels.

[0113] The local convolutional feedforward submodule is used for cross-channel fusion representation. Layer normalization is performed to stabilize the distribution of different channels, resulting in a normalized cross-channel fusion representation. ;

[0114] Next, the normalized cross-channel fusion representation is subjected to depthwise convolution and linear transformation to obtain local convolutional features;

[0115] Simultaneously, point convolution is performed on the cross-channel fusion representation to obtain point convolution features;

[0116] Then, the local convolutional features are multiplied element-wise with the point convolutional features to obtain the local feature vector of charging and discharging, which for the th The compressed feature vector of the layer's charging and discharging can be found in the following expression:

[0117]

[0118] In the formula, Indicates the first The compressed eigenvectors of the layer's charging and discharging; This indicates a depthwise convolution operation with a kernel length of 3, used to capture local degradation information of adjacent sampling points during battery charging and discharging. Used to represent the SiLU activation function, which is used as a gating factor to control the flow of convolutional features; This represents point convolutional features, used to compress channel dimensions and enhance non-linear representations; This represents an element-wise multiplication operation, used to combine local features with a global mapping;

[0119] Finally, the current cross-channel fusion representation is fused with the current charge / discharge local feature vectors to obtain the dual-channel fusion feature tensor of the battery in the next layer. For details, please refer to the following expression:

[0120]

[0121] It should be noted that this invention, by using the voltage channel as the query matrix and the current and surface temperature channels as keys and values ​​for weighted calculation, embodies the "potential-first" mechanism assumption. This is because most degradation signs tend to appear first in the voltage curve, while current and temperature provide mechanistic information leading to these signs. Therefore, this invention achieves efficient fusion between the voltage and current-temperature channels through a cross-attention mechanism, thereby capturing degradation signals while maintaining strong interpretability; simultaneously, it integrates the attention output with the original input. Adding the results to form residual updates can enhance feature interactions while avoiding information loss and improving the training stability of deep networks.

[0122] In addition, in the local convolutional feedforward submodule, this invention extracts local degradation features (such as micro-discharge platforms and micro-temperature rises) between adjacent sampling points through depthwise convolution, while using point convolution to obtain compressed global features, and achieving complementary fusion through element-wise multiplication. This can highlight local morphological features that are sensitive to degradation and reduce the number of redundant parameters, which is equivalent to compressing the number of parameters to 1 / 4 of that of the traditional FFN (two-layer fully connected).

[0123] It should be further clarified that although this invention uses the voltage channel embedding representation as the query vector and the current and surface temperature channel embedding representations as keys and values ​​for weighted fusion during the calculation of the cross-attention submodule, this does not mean that other auxiliary features (such as instantaneous charge / discharge capacity, and the first and second derivatives of the voltage-capacity relationship) are not utilized. In fact, these auxiliary features have been uniformly encoded into the battery's electrothermal coupling embedding vector during the feature expansion in step S1 and the embedding mapping in step S2, and are input into the attention network along with the main channel features. Therefore, the cross-attention mechanism only logically uses the voltage channel as the "dominant query" and does not discard any auxiliary physical features. Its overall model still models the complete electrothermal coupling representation containing all extended features, thereby ensuring that the model can comprehensively capture the multi-scale degradation information of the battery in the electro-thermal-chemical dimensions.

[0124] The battery life prediction unit 4 is used to perform step S4: predict the battery life of the dual-channel fused feature tensor of the battery to obtain the remaining life prediction value of the current battery.

[0125] Specifically, the battery life prediction includes the following sub-steps:

[0126] First, the dual-channel fused feature tensor of the battery is averaged and pooled according to the time dimension to obtain a recurrent global representation, which is as follows:

[0127]

[0128] In the formula, This indicates that the charge / discharge cycle number is A global representation of the cycle level at that time, used to reflect the overall degradation characteristics of the battery under that cycle; Indicates the first The layer output in charge / discharge cycle number is At that time, the corresponding dual-channel fused feature tensor of the battery, its The last layer is a lightweight dual-channel attention module, in which, Arranged in chronological order It consists of hidden vectors, This indicates the number of sampling points within a single loop; Indicates the above along the time dimension The operation of summing each hidden vector element by element is used to obtain global statistics across a time window.

[0129] Next, the global representation at the cyclic level is subjected to health state classification and probability distribution mapping to obtain the current battery health state and remaining life interval probability, which are specifically represented as follows:

[0130]

[0131] In the formula, Indicates that the charge / discharge sequence number is The battery's health status at that time is used to reflect the battery's capacity retention level and is a continuous value within the range [0,1]. and These are the trainable classification weight matrix and bias matrix, respectively; The Sigmoid function is used to map health status to probability values ​​between [0,1] to achieve health status classification or regression. Indicates that the charge / discharge sequence number is The remaining lifetime interval probability at a given time is used to characterize the likelihood that the battery's remaining lifetime falls into each interval, in order to avoid the instability that may be caused by direct regression values, while providing a measure of uncertainty for different lifetime intervals. and These are the trainable regression weight matrix and bias matrix, respectively; This indicates a normalization operation.

[0132] Finally, based on a preset interval center weight, a weighted regression mapping is performed on the remaining life interval probability to obtain the remaining life prediction value, which is specifically calculated as follows:

[0133]

[0134] In the formula, Indicates that the charge / discharge sequence number is At that time, the predicted remaining lifespan of the battery; Indicates the first The center value of a preset lifespan interval is used as the first... The interval represents the lifespan, its This indicates the number of preset intervals, used to divide the entire lifespan into multiple intervals; Indicates that the charge / discharge sequence number is Battery life falls to the first The remaining lifetime probability value for each interval.

[0135] It should be noted that the preset interval center weights are manually preset and are used to discretize the battery's entire lifespan into several intervals, with the center value of each interval being... As a representative measure of battery lifespan, this invention does not specifically limit the number of intervals or the method of division; it can be equal-width division, equal-frequency division, or adaptive division based on the stage of battery lifespan degradation. Furthermore, the dual-output structure of the battery's health status and the predicted remaining battery lifespan improves the interpretability and reliability of the prediction results.

[0136] It should be further noted that the backbone network and battery life prediction unit based on dual-channel attention fusion of the present invention can be trained end-to-end. The present invention does not limit the specific training algorithm. For example, according to the actual data scale and task requirements, cross-entropy loss, mean squared error loss or a combination thereof can be selected for loss calculation, and the trainable parameters in the backbone network and battery life prediction unit can be updated by gradient in combination with the optimizer (Adam) to realize model training. At the same time, during the training process, overfitting can be suppressed by applying Dropout (p=0.1) or other regularization methods after all weight matrices or projection matrices, while such redundant calculations are omitted in the inference stage to reduce inference latency.

[0137] Compared to existing technologies, this invention introduces physically relevant features such as capacity increment, first derivative and second derivative of voltage-capacity relationship, on the basis of original current, voltage and temperature sampling data. This adds a feature dimension reflecting electrochemical reaction kinetics and thermal changes to the single time-series signal, ensuring that the model can accurately capture the microscopic degradation information of battery internal material decay and thermal effect response.

[0138] In addition, this invention adds cyclic index weights to the standard sine and cosine time encoding, enabling the model to distinguish temporal fluctuations within the same cycle and to connect lifetime stage information of different cycles, thus solving the problems of spatiotemporal information dilution and gradient decay in long sequences.

[0139] Meanwhile, in the dual-channel attention fusion, the present invention uses voltage sequence as query and current and temperature concatenation as key and value to perform cross-channel attention to mine global dependencies. Then, it extracts local degradation features through lightweight deep separable convolution and SiLU gated feedforward module. The two work together to achieve efficient coupling and accurate fusion of electro-thermal-chemical data.

[0140] To verify the effectiveness of this invention, four sets of battery data (B101–B104) were selected, and comparative experiments were conducted with Support Vector Regression (SVR), Long Short-Term Memory (LSTM) networks, and Transformer models. Please refer to [link to relevant documentation]. Figure 3 , Figure 3This diagram illustrates the comparison of experimental results for predicting remaining battery life using four sets of battery data. As shown in the figure, across all tested batteries, this invention achieved superior results in terms of mean absolute error (MAE), root mean square error (RMSE), and relative error (RE). Specifically, taking the B101 battery as an example, the MAE of this invention is 0.0103, which is approximately 33% and 6% lower than that of LSTM (0.0153) and Transformer (0.0110), respectively. For the B102 battery, the MAE decreases to 0.0310, representing reductions of approximately 14% and 9% compared to LSTM and Transformer, respectively. For the B103 and B104 batteries, this invention also achieved error reductions of approximately 10%–30%.

[0141] Furthermore, both RMSE and RE indices showed a significant decrease, indicating that the method of the present invention can stably capture degradation characteristics on different battery samples and improve the accuracy and reliability of remaining lifetime prediction.

[0142] Based on the same inventive concept, this application also provides an electronic device, which can be a server, desktop computing device, or mobile computing device (e.g., laptop computing device, handheld computing device, tablet computer, netbook, etc.). The device includes one or more processors and a memory, wherein the processor is used to execute a program to implement the lithium battery remaining life prediction method of the embodiments of the present invention; the memory is used to store a computer program executable by the processor.

[0143] Based on the same inventive concept, this application also provides a computer-readable storage medium corresponding to the aforementioned embodiment of a method for predicting the remaining lifespan of a lithium battery. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the method for predicting the remaining lifespan of a lithium battery as described in any of the above embodiments.

[0144] This application may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0145] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A method for predicting the remaining lifespan of a lithium battery, characterized in that, Includes the following steps: S1: Feature extension of the battery's voltage, current, and surface temperature to obtain the battery's electrothermal coupling data; The feature is extended to obtain the instantaneous charge and discharge capacity by integrating the battery current, and to perform first-order and second-order finite difference calculations on the battery voltage and corresponding capacity respectively using the finite difference method to obtain the first-order and second-order derivatives of the voltage-capacity relationship. The battery current, voltage, temperature, instantaneous charge and discharge capacity, and the first-order and second-order derivatives of the voltage-capacity relationship are combined to obtain the battery's electrothermal coupling data. S2: Perform cyclic position encoding preprocessing on the battery's electrothermal coupling data to obtain the battery's electrothermal coupling embedding vector; The cyclic position encoding preprocessing includes the following sub-steps: First, the electrothermal coupling data of the battery is subjected to linear transformation and layer normalization operations in sequence to obtain the normalized embedded representation; Simultaneously, cyclic position encoding is performed on each sampling moment of the battery's electrothermal coupling data to obtain the cyclic position encoding vector corresponding to each sampling moment; Finally, the normalized embedding representation is fused with the cyclic position encoding vector to obtain the battery's electrothermal coupling embedding vector; S3: Perform several dual-channel attention feature extractions on the battery's electrothermal coupling embedding vector to obtain the battery's dual-channel fusion feature tensor. S4: Perform battery lifetime prediction on the dual-channel fused feature tensor of the battery to obtain the predicted remaining lifetime value of the current battery. The dual-channel attention feature extraction is performed by a lightweight dual-channel attention module. Each layer of the lightweight dual-channel attention module includes a cross-channel attention submodule and a local convolutional feedforward submodule. The cross-channel attention submodule is used to perform layer normalization on the current dual-channel fusion feature tensor of the battery to obtain the normalized dual-channel fusion feature tensor. If the current dual-channel fusion feature tensor of the battery is the first layer, then it is the battery's electrothermal coupling embedding vector. Next, the embedding representations belonging to the voltage channel in the normalized dual-channel fused feature tensor are queried and mapped to obtain the query matrix; Meanwhile, the embedding representations belonging to the current channel and the surface temperature channel in the normalized dual-channel fusion feature tensor are concatenated, and the concatenated embedding representations are mapped to the key matrix and the value matrix, respectively. Then, based on the similarity weights between the query matrix and the key matrix, the value matrix is ​​weighted and summed to obtain the cross-channel fused attention representation. The cross-channel fused attention representations corresponding to all attention heads are then concatenated to obtain the attention weight vector. Finally, the current attention weight vector is fused with the current dual-channel fusion feature tensor of the battery to obtain the current cross-channel fusion representation.

2. The method for predicting the remaining service life of a lithium battery according to claim 1, characterized in that, The local convolutional feedforward submodule is used to perform layer normalization on the cross-channel fusion representation to stabilize the distribution of different channels and obtain the normalized cross-channel fusion representation. Next, the normalized cross-channel fusion representation is subjected to depthwise convolution and linear transformation to obtain local convolutional features; Simultaneously, point convolution is performed on the cross-channel fusion representation to obtain point convolution features; Then, the local convolutional features are multiplied element-wise with the point convolutional features to obtain the local feature vector of charging and discharging. Finally, the current cross-channel fusion representation is fused with the current charge / discharge local feature vectors to obtain the dual-channel fusion feature tensor of the battery in the next layer.

3. A device for predicting the remaining lifespan of a lithium battery, characterized in that, It includes a charge / discharge data feature extension unit, a cyclic position encoding preprocessing unit, a backbone network based on dual-channel attention fusion, and a battery life prediction unit; The charge / discharge data feature extension unit is used to extend the voltage, current and surface temperature of the battery to obtain the battery's electrothermal coupling data. The feature is extended to obtain the instantaneous charge and discharge capacity by integrating the battery current, and to perform first-order and second-order finite difference calculations on the battery voltage and corresponding capacity respectively using the finite difference method to obtain the first-order and second-order derivatives of the voltage-capacity relationship. The battery current, voltage, temperature, instantaneous charge and discharge capacity, and the first-order and second-order derivatives of the voltage-capacity relationship are combined to obtain the battery's electrothermal coupling data. The cyclic position coding preprocessing unit is used to perform cyclic position coding preprocessing on the electrothermal coupling data of the battery to obtain the electrothermal coupling embedding vector of the battery; wherein, the cyclic position coding preprocessing includes the following sub-steps: First, the electrothermal coupling data of the battery is subjected to linear transformation and layer normalization operations in sequence to obtain the normalized embedded representation; Simultaneously, cyclic position encoding is performed on each sampling moment of the battery's electrothermal coupling data to obtain the cyclic position encoding vector corresponding to each sampling moment; Finally, the normalized embedding representation is fused with the cyclic position encoding vector to obtain the battery's electrothermal coupling embedding vector; The backbone network based on dual-channel attention fusion is used to perform several dual-channel attention feature extractions on the battery's electrothermal coupling embedding vector to obtain the battery's dual-channel fusion feature tensor. The battery life prediction unit is used to predict the battery life of the battery by performing battery life prediction on the dual-channel fused feature tensor, and to obtain the predicted value of the remaining life of the current battery. The backbone network based on dual-channel attention fusion also includes several layers of stacked lightweight dual-channel attention modules. Each layer of lightweight dual-channel attention module includes a cross-channel attention submodule and a local convolutional feedforward submodule. The cross-channel attention submodule is used to perform layer normalization on the current dual-channel fusion feature tensor of the battery to obtain the normalized dual-channel fusion feature tensor. If the current dual-channel fusion feature tensor of the battery is the first layer, then it is the battery's electrothermal coupling embedding vector. Next, the embedding representations belonging to the voltage channel in the normalized dual-channel fused feature tensor are queried and mapped to obtain the query matrix; Meanwhile, the embedding representations belonging to the current channel and the surface temperature channel in the normalized dual-channel fusion feature tensor are concatenated, and the concatenated embedding representations are mapped to the key matrix and the value matrix, respectively. Then, based on the similarity weights between the query matrix and the key matrix, the value matrix is ​​weighted and summed to obtain the cross-channel fused attention representation. The cross-channel fused attention representations corresponding to all attention heads are then concatenated to obtain the attention weight vector. Finally, the current attention weight vector is fused with the current dual-channel fusion feature tensor of the battery to obtain the current cross-channel fusion representation.

4. The lithium battery remaining lifespan prediction device according to claim 3, characterized in that, The local convolutional feedforward submodule is used to perform layer normalization on the cross-channel fusion representation to stabilize the distribution of different channels and obtain the normalized cross-channel fusion representation. Next, the normalized cross-channel fusion representation is subjected to depthwise convolution and linear transformation to obtain local convolutional features; Simultaneously, point convolution is performed on the cross-channel fusion representation to obtain point convolution features; Then, the local convolutional features are multiplied element-wise with the point convolutional features to obtain the local feature vector of charging and discharging. Finally, the current cross-channel fusion representation is fused with the current charge / discharge local feature vectors to obtain the dual-channel fusion feature tensor of the battery in the next layer.

5. The lithium battery remaining life prediction device according to any one of claims 3-4, characterized in that, It also includes a data acquisition module, which is equipped with a voltage sensor, a current sensor and a temperature sensor; The voltage sensor is used to collect the battery voltage, obtain the battery voltage data, and transmit it to the charge / discharge data feature extension unit. The current sensor is used to collect the battery current, obtain the battery current data, and transmit it to the charge / discharge data feature extension unit. The temperature sensor is used to collect the surface temperature of the battery, obtain the surface temperature data of the battery, and transmit it to the charge / discharge data feature extension unit.

6. A lithium battery remaining lifespan prediction system, characterized in that, It includes the device to be tested, and the lithium battery remaining life prediction device according to any one of claims 3-4; The device under test is equipped with a lithium battery and a data acquisition unit connected to the lithium battery; The data acquisition unit connected to the lithium battery is used to acquire voltage, current and surface temperature data during the charging and discharging process of the lithium battery, and transmit the voltage, current and surface temperature data to the lithium battery remaining life prediction device via wired or wireless means to predict the remaining life of the lithium battery.

7. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements a method for predicting the remaining lifespan of a lithium battery as described in any one of claims 1-2.

8. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed by the processor, they implement a method for predicting the remaining lifespan of a lithium battery as described in any one of claims 1-2.

Images