Method and device for predicting remaining service life of battery, equipment, and storage medium

Abstract

The application relates to a battery residual service life prediction method and device, equipment and a storage medium. The battery residual service life prediction method comprises the following steps: obtaining battery data of a battery to be predicted; dividing the battery data into M first sequences based on a window length and a step length; performing step-by-step noise adding processing on the M first sequences by using a preset forward diffusion process included in a trained diffusion Transformer autoencoder (DTAE), to obtain M noise-added sequences; encoding the M noise-added sequences and / or the M first sequences by using an encoder included in the trained DTAE, to obtain M deterministic degradation features; obtaining N residual service life prediction sequences by using a predictor included in the trained DTAE based on the M deterministic degradation features; and performing uncertainty quantization processing based on the N residual service life prediction sequences, to obtain a residual service life prediction result comprising a first confidence interval.

Description

Technical Field

[0001] This application relates to the field of battery state prediction, specifically to a method, apparatus, device, and storage medium for predicting the remaining lifespan of a battery. Background Technology

[0002] Batteries are core components in modern electronic devices, electric vehicles, and renewable energy storage systems. Predicting battery performance degradation and remaining useful life (RUL) is crucial for the reliability, safety, and economic viability of these devices. Therefore, it is necessary to propose a method for predicting battery RUL. Summary of the Invention

[0003] This application provides a method, apparatus, device, and storage medium for predicting the remaining lifespan of a battery, which can improve the accuracy of predicting the remaining lifespan of a battery.

[0004] The technical solution of this application embodiment is implemented as follows:

[0005] This application provides a method for predicting the remaining lifespan of a battery, including:

[0006] Obtain battery data for the battery to be predicted; the battery data is at least one of voltage, current, temperature, and capacity collected at multiple sampling times; based on the window length and step size, divide the battery data into M first sequences; M is an integer greater than or equal to 2; using the preset forward diffusion process included in the trained Diffusion Transformer Autoencoder (DTAE), progressively add noise to the M first sequences to obtain M first noisy sequences; using the encoder included in the trained DTAE, encode the M first noisy sequences and / or the M first sequences to obtain M deterministic degradation features; based on the M deterministic degradation features, use the predictor included in the trained DTAE to obtain N remaining useful life prediction sequences; perform uncertainty quantization processing on the N remaining useful life prediction sequences to obtain the remaining useful life prediction result including the first confidence interval.

[0007] Based on the above technical means, a well-trained large language model is used to extract deterministic degradation features that can effectively capture battery performance degradation patterns. Then, time-series modeling is performed on the deterministic degradation features, which can accurately predict the remaining battery life. Furthermore, by introducing an uncertainty quantification mechanism, confidence interval estimation is added to the battery remaining life prediction sequence, which further improves the accuracy and reliability of the prediction and provides a scientific basis for battery management and maintenance.

[0008] Furthermore, based on M deterministic degradation features, using the predictor included in the trained DTAE, N remaining service life prediction sequences are obtained, including: parsing each deterministic degradation feature into physical parameters, including capacity decay rate, temperature fluctuation variance, and voltage relaxation time; substituting the physical parameters corresponding to each deterministic degradation feature into a preset battery feature semantic template to obtain a battery feature semantic description corresponding to each deterministic degradation feature; inputting the M battery feature semantic descriptions into the trained Large Language Model (LLM) for feature enhancement and semantic interpretation to obtain M temporal feature vectors; and inputting the M temporal feature vectors into the predictor to obtain N remaining service life prediction sequences.

[0009] Furthermore, based on M deterministic degradation features, and using the predictor included in the trained DTAE, N remaining lifespan prediction sequences are obtained, including: using the decoder included in the trained DTAE, performing N independent back-diffusion processes on the M deterministic degradation features to obtain N battery degradation trajectories at future time points; N is an integer greater than or equal to 2; inputting each battery degradation trajectory into the predictor to obtain N remaining lifespan prediction sequences; and performing uncertainty quantization processing on the N remaining lifespan prediction sequences to obtain the remaining lifespan prediction result.

[0010] Furthermore, uncertainty quantification is performed on the N remaining useful life prediction sequences to obtain the remaining useful life prediction results, including: decomposing each remaining useful life prediction sequence into low-frequency components and high-frequency components through wavelet transform; obtaining the low-frequency variance based on the low-frequency components corresponding to each of the N remaining useful life prediction sequences; obtaining the high-frequency variance based on the high-frequency components corresponding to each of the N remaining useful life prediction sequences; and obtaining the remaining useful life prediction results based on the low-frequency variance, high-frequency variance, and the first confidence interval.

[0011] Based on the above technical means, the remaining service life prediction sequence is decomposed into low-frequency components reflecting the degradation trend and high-frequency components representing the prediction uncertainty by wavelet transform. Then, variance processing is performed to further improve the accuracy of the battery remaining service life prediction sequence.

[0012] Furthermore, the method also includes: dividing the acquired battery data to be trained into P second sequences based on the window length and step size; where P is an integer greater than or equal to 2; using the preset forward diffusion process included in the DTAE, progressively adding noise to the P second sequences to obtain P second noisy sequences; using the encoder included in the DTAE, encoding the P second noisy sequences and / or the P second sequences to obtain P deterministic degradation feature representations; using the predictor included in the DTAE, performing noise prediction on the P deterministic degradation feature representations to obtain P remaining service life prediction values; determining a loss function based on the P true capacity decay rates and P remaining service life prediction values ​​corresponding to the acquired P second sequences, and training the DTAE based on the loss function to obtain a trained DTAE.

[0013] Furthermore, the method also includes: using the pre-defined forward diffusion process and encoder included in the trained DTAE, performing progressive noise addition and encoding on P second sequences to obtain P deterministic degradation features; based on each deterministic degradation feature, obtaining the battery feature semantic description corresponding to each deterministic degradation feature; inputting the P battery feature semantic descriptions into the LLM for feature enhancement and semantic interpretation to obtain P temporal feature vectors; and performing low-rank adaptive fine-tuning on the LLM based on the P temporal feature vectors to obtain the trained LLM.

[0014] Furthermore, the method also includes: obtaining the current health status of the battery to be predicted; and adjusting the first confidence interval based on the current health status of the battery to be predicted to obtain the adjusted second confidence interval.

[0015] Furthermore, based on the health status of the battery to be predicted, the first confidence interval is adjusted to obtain the adjusted second confidence interval, including: using the current health status of the battery to be predicted as the input of the adjustment function of the remaining service life to obtain the adjusted value of the remaining service life; and adjusting the first confidence interval based on the adjusted value of the remaining service life to obtain the adjusted second confidence interval.

[0016] Furthermore, based on the window length and step size, the battery data is divided into M first sequences, including: preprocessing the battery data to obtain processed battery data; and dividing the processed battery data into M first sequences based on the window length and step size.

[0017] This application provides a device for predicting the remaining lifespan of a battery, the device comprising:

[0018] The acquisition unit is used to acquire battery data of the battery to be predicted; the battery data is at least one of voltage, current, temperature and capacity collected at multiple sampling times;

[0019] The processing unit is used to divide the battery data into M first sequences based on the window length and step size; M is an integer greater than or equal to 2.

[0020] The noise unit is used to progressively add noise to M first sequences using the preset forward diffusion process included in the trained Diffusion Transformer Autoencoder (DTAE) to obtain M first noisy sequences.

[0021] The feature extraction unit is used to encode M first noisy sequences and / or M first sequences using the encoder included in the trained DTAE to obtain M deterministic degradation features;

[0022] The prediction unit is used to obtain N remaining lifetime prediction sequences based on M deterministic degradation features and using the predictors included in the trained DTAE.

[0023] The uncertainty quantification unit is used to perform uncertainty quantification processing based on N remaining useful life prediction sequences to obtain remaining useful life prediction results including the first confidence interval.

[0024] This application provides an electronic device, which includes: a memory for storing computer-executable instructions or computer programs; and a processor for executing the computer-executable instructions or computer programs stored in the memory to implement the battery remaining life prediction method provided in this application.

[0025] This application provides a computer-readable storage medium storing a computer program or computer-executable instructions, which, when executed by a processor, implements the battery remaining life prediction method provided in this application.

[0026] This application provides a computer program product, including a computer program or computer-executable instructions. When the computer program or computer-executable instructions are executed by a processor, they implement the battery remaining life prediction method provided in this application.

[0027] The embodiments of this application have the following beneficial effects:

[0028] (1) This application utilizes a well-trained large language model to extract deterministic degradation features that can effectively capture the battery performance decay mode, and then performs time series modeling on the deterministic degradation features, which can accurately predict the remaining battery life; furthermore, by introducing an uncertainty quantification mechanism, confidence interval estimation is added to the battery remaining life prediction sequence, which further improves the accuracy and credibility of the prediction and provides a scientific basis for battery management and maintenance.

[0029] (2) This application uses wavelet transform to decompose the remaining service life prediction sequence into a multi-scale decomposition, which decomposes it into a low-frequency component that reflects the degradation trend and a high-frequency component that represents the prediction uncertainty. Then, variance processing is performed to further improve the accuracy of the battery remaining service life prediction sequence. Attached Figure Description

[0030] Figure 1 A schematic diagram of the first process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0031] Figure 2 A schematic diagram of the second process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0032] Figure 3 A schematic diagram of the third process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0033] Figure 4 A schematic diagram of the fourth process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0034] Figure 5 A schematic diagram of the fifth process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0035] Figure 6 A schematic diagram of the sixth process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0036] Figure 7 A schematic diagram of the seventh process of the method for predicting the remaining battery life provided in the embodiments of this application;

[0037] Figure 8 A schematic diagram of the structure of the battery remaining life prediction device provided in the embodiments of this application;

[0038] Figure 9 This is a schematic diagram of the electronic device structure provided in the embodiments of this application.

[0039] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0041] This application provides a method for predicting the remaining lifespan of a battery. Figure 1 This is a first flowchart illustrating the method for predicting the remaining battery life provided in an embodiment of this application, as shown below. Figure 1 As shown, the method for predicting the remaining lifespan of the battery includes the following steps:

[0042] S101: Obtain battery data of the battery to be predicted; the battery data is at least one of voltage, current, temperature and capacity collected at multiple sampling times.

[0043] Based on a preset sampling frequency (e.g., 1Hz), data is collected within a preset time period, which includes several charge-discharge cycles, thereby collecting at least one of the voltage, current, temperature, and capacity of the battery to be predicted at multiple sampling times.

[0044] Additionally, it should be noted that battery data may also include battery electrochemical parameters, fiber optic sensing parameters, and is not limited to at least one of pressure, current, temperature, and capacity.

[0045] A complete charge-discharge cycle refers to the process of a battery going from a fully charged state to a fully discharged state, and then being recharged to a fully charged state.

[0046] S102: Based on the window length and step size, divide the battery data into M first sequences; M is an integer greater than or equal to 2.

[0047] Window length refers to the fixed size of the data segment covered during each sliding motion. For example, in time series analysis, if the window length is 5, then 5 consecutive data points will be selected as a window each time.

[0048] Step size refers to the number of data points moved each time the window slides. A step size of 1 means the window slides point by point; when the step size is the length of the window, the window does not overlap (i.e., it does not slide, but jumps directly).

[0049] The first sequence includes battery data at multiple sampling times corresponding to the window length.

[0050] In some embodiments of this application, dividing the battery data into M first sequences based on the window length and the step size includes: preprocessing the battery data to obtain processed battery data; and dividing the processed battery data into M first sequences based on the window length and the step size.

[0051] In this embodiment, the battery data can first be normalized, then outlier detection and repair can be performed to obtain processed battery data. Then, based on the window length and step size, the processed battery data is divided into M first sequences. This ensures data quality. The repair can be performed using interpolation between preceding and following values.

[0052] It should be noted that normalizing the battery data before performing outlier detection and repair is to transform the battery data into high-quality, model-processable time-series samples, eliminate the influence of units, and provide standard input for the trained large language model.

[0053] For example, if the window length is 100s and the step size is 50s, then the first sequence can be the voltage, current and temperature at time t1, up to the voltage, current and temperature at time t100; the second sequence can be the voltage, current and temperature at time t51, up to the voltage, current and temperature at time t150; and so on, without further explanation.

[0054] S103: Using the pre-defined forward diffusion process included in the trained Diffusion Transformer Autoencoder (DTAE), M first sequences are progressively denoised to obtain M first denoised sequences.

[0055] The Diffusion Transformer Autoencoder (DTAE) is a neural network architecture used for unsupervised learning. The encoder compresses the input data into a low-dimensional, dense latent feature representation, while the decoder reconstructs the original data from this representation. Diffusion involves two processes: forward propagation and backward propagation. The forward propagation adds noise to the data progressively, while the backward propagation removes noise from the data.

[0056] It should be noted that the trained large language model is based on the battery data to be trained. The DTAE is self-supervised pre-trained using the battery data to extract deterministic degradation features. Then, these features, along with semantic cues, are used to perform low-rank adaptation (LoRA) fine-tuning on the LLM. Finally, a predictor is jointly trained to predict the remaining lifetime. Specifically, the forward diffusion process included in the diffused Transformer autoencoder DTAE is used for noise reduction. The Transformer decoder is used for denoising. The predictor is used to predict the remaining lifetime.

[0057] Here, the first sequence is noise-added using a pre-defined forward diffusion process included in the Diffusion Transformer autoencoder (DTAE). This is not data corruption, but rather a feature enhancement and regularization technique. It forces the model to learn the ability to extract robust features even under noise interference, improving its generalization ability to real-world data noise. Furthermore, the pre-defined forward diffusion process is not stored separately in the encoder or decoder, but exists as a pre-defined flow independent of both, typically executed directly by the model framework (or algorithmic logic).

[0058] For details on the specific noise addition process, please refer to the specific steps in the following examples; they will not be repeated here.

[0059] S104: Using the encoder included in the trained DTAE, encode the M first noisy sequences and / or the M first sequences to obtain M deterministic degradation features.

[0060] M first noisy sequences and / or M first sequences, which can be M first noisy sequences or M first sequences and some noisy sequences.

[0061] It should be noted that the Transformer's self-attention mechanism can effectively capture long-term dependencies, and the encoder outputs a compressed representation of the degradation process.

[0062] Based on this, a Transformer encoder is used to encode each noisy sequence, mapping it to a low-dimensional latent space to obtain corresponding deterministic degradation features, which can effectively capture battery performance degradation patterns. Deterministic degradation features can capture the hidden structures and patterns in battery data, representing the battery's state and characteristics in a more concise and abstract way.

[0063] S105: Based on M deterministic degradation features, N remaining lifetime prediction sequences are obtained using the predictors included in the trained DTAE.

[0064] Accurate predictions can be made by leveraging the powerful sequence modeling and causal reasoning capabilities of a well-trained Large Language Model (LLM). Based on the semantic description of the input, the LLM can generate remaining lifespan prediction sequences directly or stepwise through its internal modeling and reasoning capabilities.

[0065] S106: Based on N remaining useful life prediction sequences, perform uncertainty quantification to obtain remaining useful life prediction results including the first confidence interval.

[0066] The first confidence interval of the remaining battery life (RUL) is used to quantify the uncertainty of the prediction, representing the range within which the true value of RUL may fall at a specific confidence level (e.g., 95%).

[0067] For example, at a 95% confidence level, the first confidence interval is [143, 157], which means that the true value of RUL may fall between 143 and 157 charge-discharge cycles.

[0068] In this embodiment, a well-trained large language model is used to extract deterministic degradation features that can effectively capture battery performance degradation patterns. Then, time-series modeling is performed on the deterministic degradation features, which can accurately predict the remaining battery life. Furthermore, by introducing an uncertainty quantification mechanism, confidence interval estimation is added to the battery remaining life prediction sequence, which further improves the accuracy and reliability of the prediction and provides a scientific basis for battery management and maintenance.

[0069] In some embodiments of this application, based on M deterministic degradation features, N remaining lifetime prediction sequences are obtained using the predictors included in the trained DTAE, including:

[0070] S201: Each deterministic degradation feature is resolved into physical parameters, including capacity decay rate, temperature fluctuation variance, and voltage relaxation time.

[0071] Deterministic degradation features refer to features extracted from battery data that can quantitatively characterize the degree of degradation, and usually have clear physical meaning or statistical regularity.

[0072] S202: Substitute the physical parameters corresponding to each deterministic degradation feature into the preset battery feature semantic template to obtain the battery feature semantic description corresponding to each deterministic degradation feature.

[0073] The preset battery feature semantic templates can include Template 1 (basic description), Template 2 (physical association), and Template 3 (prediction guidance). The basic description generates a natural language sentence containing cycle count, initial capacity, current capacity, and temperature fluctuations. The physical association generates sentences that associate physical parameters with battery aging mechanisms (such as SEI film growth and lithium-ion transport impedance). The prediction guidance generates interrogative or instructional sentences to guide LLM in RUL prediction and analysis.

[0074] For example, Template 1 (basic description) states, "Over the past {number of cycles} cycles, the battery capacity has decreased from {initial capacity} Ah to {current capacity} Ah, with an average weekly temperature fluctuation of {temperature variance} °C." Template 2 (physical correlation) states, "Based on feature analysis, the capacity decay rate ({α}% / cycle) is positively correlated with the SEI film growth rate, and the voltage relaxation time ({α}% / cycle) is positively correlated with the SEI film growth rate." The 'seconds' indicate an increase in lithium-ion transport impedance. Template 3 (prediction guide) asks, "Based on the above degradation trends, predict the remaining useful life (RUL) and explain the main influencing factors."

[0075] Battery feature semantic description refers to converting deterministic degradation features, i.e., low-level features, into natural language or standardized terms to describe the battery's degradation state, mechanism, or risk level. For example, degradation of the cathode material structure leads to increased lithium-ion diffusion resistance; SEI film thickening causes accelerated capacity decay (>3% / month); a sudden voltage surge at the end of charging poses a risk of lithium deposition (high danger level).

[0076] This application can also obtain abnormal information about the battery to be predicted during use, and adjust the preset battery feature semantic template based on the abnormal information. The abnormal information about the battery to be predicted during use is: "In the last 10 cycles, the battery experienced {number of voltage drops} times at {temperature} ℃, which may lead to a shortened RUL."

[0077] The anomaly information that the battery may exhibit during use is a core component of battery health management and fault prediction. This anomaly information is typically obtained through multi-source data fusion, real-time monitoring, and intelligent analysis technologies. Its goal is to identify unexpected battery behaviors (such as performance mutations, safety risks, or accelerated degradation) to provide a basis for subsequent Remaining Life (RUL) prediction and maintenance decisions.

[0078] Abnormal information that the battery is to be predicted to occur during use can be obtained through electrical signals (such as overcharge / over-discharge, voltage surge, abnormal current ripple), temperature signals, gas signals (such as electrolyte decomposition gas production, lithium precipitation side reaction) and historical operating data (frequent fast charging, deep discharging).

[0079] S203: Input the M battery feature semantic descriptions into the trained large language model LLM, perform feature enhancement and semantic interpretation, and obtain M temporal feature vectors.

[0080] LLM takes structured prompts as input, processes them through its multi-layer Transformer structure, and finally takes the average pooling result in the last hidden state or the vector corresponding to the special classification label as the temporal feature vector.

[0081] By using a trained LLM to perform time-series modeling of the semantic descriptions of M battery features, and mapping battery data to physical semantic descriptions through cue word engineering, the accuracy of battery remaining life prediction can be improved.

[0082] S204: Input M time-series feature vectors into the predictor to obtain N remaining useful life prediction sequences.

[0083] Accurate predictions can be made by leveraging the powerful sequence modeling and causal reasoning capabilities of a well-trained LLM.

[0084] In this embodiment, a time-series model of the semantic descriptions of M battery features is performed using a trained LLM, and the battery data is mapped to a physical semantic description through prompt word engineering, which can improve the accuracy of battery remaining life prediction.

[0085] In some embodiments of this application, based on M deterministic degradation features, N remaining lifetime prediction sequences are obtained using a pre-trained DTAE predictor, including:

[0086] S301: Using the decoder included in the trained DTAE, perform N independent reverse diffusion processes on M deterministic degradation features to obtain N battery degradation trajectories at future time points; N is an integer greater than or equal to 2.

[0087] The M deterministic degenerate features are transformed into a mode that the LLM (Limited Linear Modulation) is best suited for, and then used as the first input to the decoder to obtain the denoised features corresponding to each of the M deterministic degenerate features at the first time step. The M deterministic degenerate features are then transformed into a mode that the LLM is best suited for, and then used as the second input to the decoder to obtain the denoised features corresponding to each of the M deterministic degenerate features at the second time step. The M deterministic degenerate features are then transformed into a mode that the LLM is best suited for, and then used as the third input to the decoder to obtain the denoised features corresponding to each of the M deterministic degenerate features at the third time step. This process is repeated until the M deterministic degenerate features are transformed into a mode that the LLM is best suited for, and then used as the Nth input to the decoder to obtain the denoised features corresponding to each of the M deterministic degenerate features at the Nth time step.

[0088] Based on this, for each deterministic degradation feature, the corresponding N denoised features from the first time step to the Nth time step are connected to form a degradation trajectory.

[0089] S302: Input each battery degradation trajectory into the predictor to obtain N remaining lifespan prediction sequences.

[0090] The M denoised features corresponding to each battery degradation trajectory are input into the predictor to obtain the corresponding remaining lifespan prediction sequence.

[0091] For each time point, there are N predicted remaining useful lives corresponding to M deterministic degradation features, which are connected to form a degradation trajectory. Therefore, there are N degradation trajectories for N time points.

[0092] The degradation trajectory should show a gradual downward trend over time. Therefore, anomaly filtering is performed on the N degradation trajectories. Then, based on the remaining useful life prediction sequence on the filtered degradation trajectories, uncertainty quantification is performed to obtain the remaining useful life prediction result including the first confidence interval.

[0093] In this embodiment of the application, by introducing an uncertainty quantification mechanism, a confidence interval estimate is added to the battery remaining life prediction sequence, which further improves the accuracy and reliability of the prediction and provides a scientific basis for battery management and maintenance.

[0094] In some embodiments of this application, uncertainty quantification is performed on N remaining useful life prediction sequences to obtain remaining useful life prediction results, including:

[0095] S401: Decompose each remaining useful life prediction sequence into low-frequency and high-frequency components using wavelet transform.

[0096] Wavelet transform is a time-frequency analysis tool that decomposes a signal into multiple frequency bands using wavelet basis functions of different scales. Here, wavelet transform is used to decompose the remaining lifetime prediction sequence into low-frequency and high-frequency components.

[0097] Low-frequency components can reflect the long-term trend of RUL (such as slow capacity decay and overall degradation patterns). High-frequency components can capture short-term fluctuations or noise in RUL (such as measurement errors and sudden changes in operating conditions).

[0098] S402: Based on the low-frequency components corresponding to each of the N remaining useful life prediction sequences, the low-frequency variance is obtained.

[0099] Calculate the average value of the low-frequency components corresponding to each of the N remaining useful life prediction sequences to obtain the low-frequency mean. Then, substitute the low-frequency mean and the low-frequency components corresponding to each of the N remaining useful life prediction sequences into the variance calculation formula to obtain the low-frequency variance.

[0100] S403: Based on the high-frequency components corresponding to each of the N remaining useful life prediction sequences, the high-frequency variance is obtained.

[0101] Calculate the average of the high-frequency components corresponding to each of the N remaining useful life prediction sequences to obtain the high-frequency mean. Then, substitute the high-frequency mean and the high-frequency components corresponding to each of the N remaining useful life prediction sequences into the variance calculation formula to obtain the high-frequency variance.

[0102] S404: Based on the low-frequency variance, high-frequency variance, and the first confidence interval, the remaining useful life prediction results are obtained.

[0103] The variance of the low-frequency variance is calculated by summing the variance of the high-frequency variance to obtain the variance of the N remaining useful life prediction sequences corresponding to each cycle.

[0104] In this embodiment, wavelet transform is used to decompose each remaining useful life prediction sequence into a multi-scale component that reflects the degradation trend and a high-frequency component that characterizes the prediction uncertainty. The low-frequency variance and high-frequency variance are then determined, thereby improving the accuracy of the remaining useful life prediction results.

[0105] In some embodiments of this application, the method for predicting the remaining battery life further includes the following steps:

[0106] S501: Obtain the current health status of the battery to be predicted.

[0107] State of Health (SOH) measures the degree of performance degradation of a battery relative to its brand-new state, reflecting the aging of the battery.

[0108] S502: Based on the current health status of the battery to be predicted, the first confidence interval is adjusted to obtain the adjusted second confidence interval.

[0109] In this embodiment, the first confidence interval is adjusted or corrected based on the current health status of the battery to be predicted, resulting in an adjusted second confidence interval. This improves the accuracy of the second confidence interval, thereby enabling a correct assessment of the battery's remaining lifespan.

[0110] In some embodiments of this application, adjusting the first confidence interval based on the health state of the battery to be predicted to obtain an adjusted second confidence interval includes the following steps:

[0111] S601: The current health status of the battery to be predicted is used as the input to the adjustment function of the remaining service life to obtain the adjustment value of the remaining service life.

[0112] S602: Adjust the first confidence interval based on the adjustment value of the remaining useful life to obtain the adjusted second confidence interval.

[0113] The adjustment function for the remaining useful life can be expressed as Equation 31 below, which will be described in detail below.

[0114] Based on this, by substituting the current SOH of the battery to be predicted into the adjustment function for the remaining service life, the adjusted value for the remaining service life can be obtained. .

[0115] Furthermore, the minimum value of the first confidence interval minus the adjustment value of the remaining useful life. And the maximum value of the first confidence interval plus the adjustment value of the remaining useful life. This leads to the adjusted second confidence interval.

[0116] The training process of the trained diffusion Transformer autoencoder and the large language model will be explained next:

[0117] In some embodiments of this application, the method for predicting the remaining battery life further includes:

[0118] Based on the window length and step size, the acquired battery data to be trained is divided into P second sequences; P is an integer greater than or equal to 2.

[0119] The battery data to be trained is collected using an onboard BMS system or experimental platform. The battery data includes at least one of the following: voltage V(t), current I(t), temperature T(t), and capacity C(t). Here, t represents the time of the t-th sampling. The sampling frequency can be set to the level of 1Hz to 1min depending on the application scenario. After data acquisition, the data needs to be preprocessed to obtain the preprocessed dataset. The specific preprocessing procedure is as follows:

[0120] (1) Standardization of raw data

[0121] To eliminate the influence of different units of measurement and improve the model's convergence speed and generalization performance, the battery data to be trained is first subjected to Min-Max normalization, calculated as follows:

[0122] (1)

[0123] Where x represents the original acquired parameters (voltage, current, temperature, and capacity). and These are the minimum and maximum values ​​of the parameter in the sample set, respectively. This represents the parameters after normalization.

[0124] (2) Outlier detection and repair

[0125] The Z-score method is used to identify outliers in time series data. The calculation method is as follows:

[0126] (2)

[0127] Where μ is the mean, σ is the standard deviation, and z represents the standardized parameter. If |z| > 3 for a data point, it is identified as an outlier and repaired using interpolation between preceding and following values.

[0128] (3)

[0129] Where x represents the original acquisition parameters, and t represents the t-th sampling time. This represents the parameters after interpolation repair at the t-th sampling time.

[0130] (3) Construct a capacity decay model

[0131] In each complete charge-discharge cycle, the initial discharge capacity collected is used. and the discharge capacity of the nth cycle The discharge capacity is a cycle-level summary, not data collected per second. The capacity decay rate is calculated. ,pass It can construct a capacity degradation curve, which can be used for subsequent model training and target value calculation. Specifically, the capacity decay rate... The calculation method is as follows:

[0132] (4)

[0133] (4) Constructing time series samples using a sliding window

[0134] To be compatible with deep learning input structures, a sliding window method is used to process the raw data and transform it into multi-step sequence sample data.

[0135] Specifically, the data processing parameters are set as follows: window length is set to L, step size is s, and the window starts at the sampling point. The coverage of sampling points is The generated sample structure is as follows:

[0136] (5)

[0137] in, V represents voltage, I represents current, T represents temperature, and i is the sample index number.

[0138] (6)

[0139] Where i is the sample index number, i=1,2,…,n (n∈Z); It is the starting point of the first window sampling, that is, the initial sampling point of the entire sampling process.

[0140] This represents the sample data for the i-th time segment, containing data from L consecutive time points starting from i. T and d represent the number of feature dimensions.

[0141] This represents the label corresponding to the i-th time segment, i.e., the capacity decay rate.

[0142] The sample dataset is obtained by constructing a sliding window. The expression is:

[0143] (7)

[0144] It should be noted that the second sequence is constructed here. and corresponding This is for use in training large language models.

[0145] For example, when L=100 and s=50, if i=1, then , etc.

[0146] This is not limited to battery data V, I, and T; it can be replaced with any symbol, or in other words, it can be replaced with other collected battery data.

[0147] By utilizing the preset forward diffusion process contained in the Diffusion Transformer Autoencoder (DTAE), each second sequence is progressively noise-added to obtain a second noise-added sequence.

[0148] In the battery remaining life (RUL) prediction method based on the diffusion model, the forward diffusion process is the core encoding stage of self-supervised learning. This process forces the model to learn robust latent feature representations by progressively adding Gaussian noise to the battery time-series data, simulating the multi-scale noise distribution of data degradation. The specific process is as follows:

[0149] The mathematical process of forward diffusion. (This involves) processing the sample dataset. A noisy dataset is generated through a forward diffusion process, that is, by adding noise step by step in K steps. .

[0150] Specifically, the K-step noise addition process is as follows:

[0151] (8)

[0152] Where k represents the index of the current diffusion step, and K represents the total number of forward diffusion steps.

[0153] The following uses a sample dataset. The i-th sample Let's take an example to illustrate this.

[0154] After a forward diffusion process, i.e., by adding noise step by step in K steps, a noisy sequence is obtained. The expression is:

[0155] (9)

[0156] in, This represents the state of the i-th sample at step k. The relationship between them, satisfying the Markov chain, can be expressed by the following formula:

[0157] (10)

[0158] in, This represents the transition probability of forward diffusion. This represents the noise-adding path for the i-th sequence at step k. Indicates the given raw data Generate a noisy path The joint conditional probability. The noise addition at each step follows a Gaussian distribution:

[0159] (11)

[0160] (12)

[0161] in, , This is the cumulative noise attenuation factor. The noise scheduling parameter controls the noise intensity at step k. A cosine scheduling strategy is used here. . The ratio of signal to noise was controlled. As the step size s increases, the original signal component gradually weakens while the noise component strengthens. ε is standard Gaussian noise.

[0162] Specifically, the noise scheduling design process, taking into account the characteristics of lithium battery degradation, involves designing asymmetric noise scheduling, including: 1) Early stage (number of cycles t≤K / 2): using small step-size noise increase ( (Slow growth), maintaining a gradual capacity change trend. 2) Later stage (number of cycles t>K / 2): Accelerate noise addition ( (Rapid rise), simulating a capacity plunge.

[0163] Specifically, the diffusion process is performed independently at each time step t to obtain the noisy dataset. .

[0164] Using the encoder in the Initial Diffusion Transformer Autoencoder (DTAE), P second noisy sequences corresponding to P second sequences are encoded to obtain P initial deterministic degradation features.

[0165] Using the Transformer decoder included in the initial large language model, P second noisy sequences corresponding to P second sequences are mapped to P latent space vectors;

[0166] Mapping a noisy sequence to a latent space vector can be simply understood as transforming scrambled or "interference-added" data into a more concise and meaningful "abstract representation" through a process of "translation" or "encoding." This process is similar to compressing a garbled article into keywords or extracting the core features from a blurry image.

[0167] Using the Transformer decoder included in the initial large language model, noise prediction is performed based on P latent space vectors and P initial deterministic degradation features to obtain the predicted noise;

[0168] Based on the predicted noise and the real noise, as well as the P real capacity decay rates and P initial deterministic degradation features corresponding to the P second sequences, the initial large language model is trained to obtain the trained large language model; the initial large language model includes a target function layer, which is used to convert the capacity decay rate into the remaining lifetime.

[0169] In battery remaining life prediction methods based on diffusion models, the inverse denoising process is the core stage for generating high-quality degradation features. By combining the decoding capabilities of the Transformer with the iterative refinement mechanism of the diffusion model, this method achieves robust modeling of noise interference and accurate reconstruction of degradation trends. The specific process is as follows:

[0170] ① Mathematical definition of inverse denoising. The goal of inverse denoising is to denoise a noisy sequence... Gradually restore original battery data Its Markov chain representation is as follows:

[0171] (13)

[0172] in, This represents the joint probability distribution of the inverse denoising process. This represents the joint transition probability of back diffusion. This represents the noisy data at step k.

[0173] Each denoising operation is performed through a neural network. Predict the noise and restore the original sequence:

[0174] (14)

[0175] in z is Gaussian noise. This is the noise variance adjustment coefficient. For noise prediction neural networks, This is the cumulative noise attenuation factor. For noise scheduling parameters, This represents the noisy data at step k, where k represents the current time step.

[0176] ② Decoder Design. Traditional diffusion models use U-Net for denoising, but it has limitations in long-term modeling. This method uses a Causal Diffusion Transformer (CDT) as the decoder, and its core modules are as follows:

[0177] Input embedding and positional encoding. Noisy dataset. Each sequence Mapped to latent space vectors:

[0178] (15)

[0179] in, Represents a latent space vector. Indicates a fully connected layer. This indicates the position code.

[0180] The diffusion step K is injected into the model through sinusoidal position encoding to enhance the perception of noise intensity.

[0181] Degradation condition injection. This involves injecting deterministic degradation features extracted from the encoder. As a condition, cross-attention is incorporated into the decoding process:

[0182] (16)

[0183] in , , These are the learning parameters used to generate queries, keys, and values. , , Q is the query matrix, K is the key matrix, and V is the value matrix. As a dimension, This indicates a deterministic degradation characteristic.

[0184] Causal Multi-Head Diffusion Attention (CMDA). During decoding, temporal causality is enforced (the current time depends not only on historical information), and the attention mask is defined as:

[0185] (17)

[0186] in, Let i be the attention mask matrix, and j represent the dimensions.

[0187] Attention is calculated as follows:

[0188] (18)

[0189] The noise predicted by the decoder is , This represents the output of the attention module, and Linear represents the fully connected layer. Finally, in the diffusion model, the final denoised output is... Furthermore, a denoised dataset can be obtained. , This represents the battery data sequence generated after reverse noise reduction. This represents the entire denoised dataset.

[0190] (3) Denoising optimization of lithium battery time-series data. To enhance the ability of the denoising process to model the battery capacity decay trend, a constraint based on the first derivative is introduced:

[0191] (19)

[0192] in, This represents the trend-constrained loss, where T is the sequence length. This represents the noisy data at step k-1, where t and t1 represent time, compared to the sample dataset. of The values ​​of i in the two cases are equal.

[0193] This forces the local slope of the generated sequence to align with the true degradation trend. Coarse-grained stage (larger k): Employs step-attention, aggregating long-period features with a step size of s=4 to capture the overall trend of capacity decay. Fine-grained stage (smaller k): Uses full-resolution attention to repair local fluctuation details (such as voltage relaxation curves).

[0194] In battery remaining life prediction, Large Language Modeling (LLM) utilizes cue word engineering and temporal reasoning fine-tuning to deeply integrate battery degradation characteristics with physical mechanisms, significantly improving prediction accuracy and interpretability. The following are the specific implementation methods and technical details:

[0195] (1) Cue word engineering, from time-series data, i.e., P deterministic degradation features, to semantic description. Convert battery time-series data into natural language cues to guide the LLM in understanding degradation patterns. Specifically:

[0196] ① Feature extraction. Deterministic degradation features are extracted from the diffusion autoencoder (DTAE). The analysis is based on physical parameters: the capacity decay rate is... The variance of temperature fluctuation is The voltage relaxation time is .

[0197] ② Template design: Build multi-level prompt templates and gradually inject domain knowledge.

[0198] Level 1 (Basic Description):

[0199] "Over the past {number of cycles} cycles, the battery capacity has decreased from {initial capacity} Ah to {current capacity} Ah, with an average weekly temperature fluctuation of {temperature variance} °C."

[0200] Level 2 (Physical Connection):

[0201] According to feature analysis, the capacity decay rate ({α}% / cycle) is positively correlated with the SEI film growth rate, and the voltage relaxation time ({α}% / cycle) is also positively correlated with the SEI film growth rate. (} seconds) suggests an increase in lithium-ion transport impedance.

[0202] Level 3 (Predictive Guidance):

[0203] Based on the aforementioned degradation trends, please predict the remaining useful life (RUL) and explain the main influencing factors.

[0204] ③ Dynamic context injection. Introduce real-time sensor data update prompts, for example:

[0205] "In the last 10 cycles, the battery experienced {number of voltage drops} at {temperature} °C, which may shorten the RUL."

[0206] ③ Dynamic context injection. Introduce real-time sensor data update prompts, for example:

[0207] "In the last 10 cycles, the battery experienced {number of voltage drops} at {temperature} °C, which may shorten the RUL."

[0208] (2) Temporal reasoning fine-tuning, domain adaptation, and knowledge fusion. Using the pre-trained DTAE including the preset forward diffusion process and the encoder, gradually add noise to and encode the P second sequences to obtain P deterministic degradation features; based on each deterministic degradation feature, obtain the battery feature semantic description corresponding to each deterministic degradation feature; input the P battery feature semantic descriptions into the LLM for feature enhancement and semantic interpretation to obtain P temporal feature vectors; perform low-rank adaptation fine-tuning on the LLM based on the P temporal feature vectors to obtain the trained LLM. In addition, the pre-trained LLM can be adjusted by combining the P true capacity attenuation rates corresponding to the P second sequences to make it suitable for the battery degradation temporal reasoning task. For example, the LLM can be GLM-4. Specifically:

[0209] ① Low-Rank Adaptation (LoRA). Parameter-efficient fine-tuning: Only train the low-rank matrix , keep the original weight W0 frozen, and inject a trainable adapter module into the Attention layer of the model. This module consists of two small matrices A and B. At this time, the forward propagation formula becomes: <--snip--><000> (20)

[0211] Among them, is the dimensionality reduction matrix, initially a random Gaussian distribution; is the dimensionality increase matrix, initially all zeros (to ensure that the weight change amount ΔW = 0 at the beginning of training and does not affect the original model); the rank r of the matrix << d (in this paper, r = 8), and r controls the number of adapted parameters and capabilities. The advantage of this method is to avoid catastrophic forgetting and reduce the number of training parameters (only 0.1% of full-scale fine-tuning). h is the result of the low-rank adaptation forward calculation, and w0 represents the original weight

[0212] ② Temporal-aware loss function. This loss function is divided into two parts: one is the early training stage, which emphasizes feature alignment (MSE loss).

[0213] One is the later training stage, which focuses on the RUL prediction accuracy (Huber loss). The loss function formula is as follows:

[0214] ---snip-- (21)

[0215] (22)

[0216] Among them, is the feature alignment loss, is the deterministic degradation feature vector, is the hidden state after mapping, is the hidden state of the LLM, For Huber loss, y represents the actual remaining battery life. This is the predicted remaining battery life value. This is the threshold parameter.

[0217] ③ Knowledge Graph Enhancement. Electrochemical knowledge is introduced, specifically the battery aging mechanism (such as SEI growth and lithium deposition) is constructed into a knowledge graph, and interacts with the hidden states of the LLM through graph attention mechanism (GAT):

[0218] (twenty three)

[0219] A knowledge graph representing chemical entities (such as "anode" and "electrolyte") and their relationships. This indicates the hidden state of the LLM. The enhanced state.

[0220] Based on the above embodiments, this application specifically illustrates a method for predicting the remaining lifespan of a battery, such as... Figure 7 As shown, it includes the following steps:

[0221] S701, battery data acquisition and preprocessing.

[0222] Referring to the data acquisition and preprocessing procedures described above, they will not be repeated here.

[0223] S702, the training process of a large language model, includes forward diffusion, inverse denoising, reconstructing the error between predicted noise and real noise, latent space vector features (i.e., deterministic degradation features), LLM cue word engineering, LLM inference fine-tuning, and RUL regression to obtain prediction results.

[0224] The training process for the large language model can be referred to the above embodiments, and will not be repeated here.

[0225] It should be noted that when performing the RUL prediction process, the data preprocessing process of the training process is referenced to preprocess the battery data of the battery to be predicted and perform sequence partitioning to obtain M first sequences.

[0226] Furthermore, using the pre-defined forward diffusion process included in the Diffusion Transformer Autoencoder (DTAE), M first sequences are progressively denoised to obtain M denoised sequences. Then, using the encoder in the trained Diffusion Transformer Autoencoder (DTAE), the M denoised sequences (or M first sequences and some denoised sequences) are encoded to obtain M deterministic degradation features. Based on the M deterministic degradation features, the predictor included in the trained large language model is used to perform N independent sampling processes (generating N degradation trajectories) to obtain the remaining lifetime prediction sequence corresponding to each cycle; N is an integer greater than or equal to 2.

[0227] Next, execute S703.

[0228] S703, uncertainty quantification, multi-trajectory diffusion sampling and confidence interval calculation.

[0229] In predicting the remaining battery life (RUL), uncertainty quantification generates a probability distribution of battery degradation through multi-track diffusion sampling and calculates confidence intervals based on statistical methods. This method combines the generative capabilities of the diffusion model with the temporal inference of the Large Language Model (LLM), significantly improving the reliability and interpretability of the prediction results. The specific process is as follows:

[0230] (1) Multi-trajectory diffusion sampling. Multiple possible degradation trajectories are generated to characterize the uncertainty of battery aging. The steps are as follows:

[0231] ① Reverse denoising sampling. For the same initial noise Perform N independent reverse denoising processes (each time injecting different random noise). Generate N degenerate trajectories:

[0232] (twenty four)

[0233] in, Given a set of N degenerate trajectories, This represents the noisy data at step k.

[0234] Conditional injection: Each sampling is guided by semantic cues generated by LLM (such as "high risk of lithium dendrite growth") to generate a trajectory that conforms to physical laws.

[0235] ② Trajectory post-processing. First, trend alignment is performed by aligning the time axes of each trajectory using dynamic time adjustment (DTW) to eliminate timing jitter. Then, anomaly filtering is performed to remove trajectories that severely conflict with known electrochemical mechanisms (such as capacity surges).

[0236] (2) Calculation of confidence intervals. Suppose that N RUL prediction sequences of degenerate trajectories are obtained through N independent samplings. ={RUL1 ,RUL 2 ,…,RUL N The confidence interval for output RUL is: , It is the adaptive confidence interval adjustment coefficient. The mean of the RUL predicted sequence, Let be the standard deviation of the RUL predicted sequence. The steps for calculating the confidence interval are as follows:

[0237] a. Calculation of data uncertainty: For the M predicted values ​​obtained in the i-th iteration, calculate their first mean. and first variance (This variance quantifies the uncertainty of the data in this forecast.)

[0238] b. Multiple iterations and model uncertainty calculation: The above prediction process is repeated N times (i from 1 to N) to obtain N means. and N variances .

[0239] (25)

[0240] (26)

[0241] Overall mean (i.e., second mean):

[0242] Total variance (i.e., the sum of the second and third variances): Total uncertainty consists of the sum of model uncertainty and average data uncertainty.

[0243] (27)

[0244] Model uncertainty (i.e., third difference): Calculate the mean of N values. The sample variance is used as an estimate of the model uncertainty:

[0245] (28)

[0246] Average data uncertainty (i.e., second variance): Calculate the variance of N data points. average :

[0247] (29)

[0248] (3) Dynamic uncertainty modeling

[0249] ① Multi-scale uncertainty decomposition. The N remaining lifetime prediction sequences are decomposed into low-frequency (trend) and high-frequency (noise) components using wavelet transform (a signal processing technique), and the uncertainties are calculated for each component.

[0250] (30)

[0251] Low-frequency uncertainty ( : Determined by the variance of the trend trajectory generated by the diffusion model, reflecting the long-term uncertainty of capacity decay. High-frequency uncertainty ( : Determined by the sensitivity of LLM to local fluctuations (such as temperature abrupt changes).

[0252] ② Adaptive confidence interval coefficient adjustment dynamically adjusts the interval width based on the battery's current state of health (SOH):

[0253] (31)

[0254] in, It is the adaptive confidence interval adjustment coefficient. The base confidence coefficient (usually taken as the quantile of the standard normal distribution, for example, 1.96 for a 95% confidence level); γ is the decay rate, controlling the decrease when SOH decreases. The rate at which the value increases is a hyperparameter greater than 0, and its optimal value can be obtained through validation set tuning; SOH is the battery health state, which is generally a scalar value between 0 and 1, and can be obtained from the battery management system.

[0255] The input is SOH, and the base confidence interval is obtained from Formula 27. The dynamic adjustment pattern describes that a higher SOH indicates a healthier battery, and model predictions are generally more certain. Conversely, a lower SOH indicates severe battery aging, more complex evolution behavior, and increased uncertainty. As the value approaches 1, the confidence interval widens.

[0256] The goal is to make state-specific predictions that better reflect actual working conditions.

[0257] Implementation Case 1: Real-time RUL Prediction of Electric Vehicle Power Batteries

[0258] Application scenario: New energy vehicle manufacturers need to predict the remaining life of the power battery in real time in the vehicle battery management system (BMS) in order to optimize charging and discharging strategies and provide early warning of potential faults.

[0259] (1) Data preparation.

[0260] Data source: 50 sets of lithium battery data collected by vehicle-mounted sensors (cycle count: 1000 times / set), including voltage, current, temperature, and capacity sequences.

[0261] Data augmentation: 200 sets of composite data are generated through a diffusion model to simulate extreme operating conditions such as high and low temperatures (-20℃~60℃) and fast charging (2C).

[0262] Preprocessing: Sliding window segmentation (window length = 50 loops, step size = 10 loops).

[0263] Normalization: Perform Min-Max normalization on temperature data and Z-Score normalization on voltage / current.

[0264] (2) Model configuration.

[0265] Diffusion Transformer Autoencoder (DTAE): Diffusion steps K=100, latent dimension h=64, causal attention head number=4.

[0266] Noise scheduling: Cosine rule (initial β1=1e) 4. Final value βK = 0.02.

[0267] Large Language Model (LLM): The base model uses GLM-4, with a LoRA adaptation layer (rank r=8). Cue word templates: Dynamically injected with real-time context such as vehicle speed and ambient temperature.

[0268] (3) Deployment plan.

[0269] Edge device: NVIDIA Jetson AGX Orin (32GB memory, 200 TOPS computing power).

[0270] Lightweight strategy: Knowledge distillation compresses the DTAE teacher model (12 layers) into a student model (4 layers), with an accuracy loss of <3%. Quantization converts the FP16 accuracy, reducing the model size by 50%.

[0271] Real-time prediction stream: Data acquisition and feature extraction are triggered every 10 loops (latency = 2ms). The diffusion model generates 5 degradation trajectories (time = 15ms). The LLM outputs RUL values ​​and a text report (time = 8ms).

[0272] (4) Verification results.

[0273] Accuracy metrics: MAE = 1.3% (compared to LSTM's 2.9%).

[0274] Interval coverage (PICP) = 96.2%, average width (MPIW) = 18 cycles.

[0275] Real-time performance: Total delay for a single prediction = 25ms, which meets the ≤30ms requirement of the vehicle system.

[0276] Interpretability: Example of LLM generated report: "Current battery capacity degradation rate is 0.42% / cycle, which is higher than the normal value (0.3%~0.35%). It is recommended to reduce the fast charging frequency to extend the RUL to 152 cycles (95% confidence interval: [138,166])."

[0277] (5) Technological advantages

[0278] Noise robustness: MAE fluctuation rate <5% under vehicle vibration-induced sensor noise (SNR=15dB).

[0279] Early warning: Capacity plunge was detected 50 cycles in advance (KneePoint), with an accuracy of 92%.

[0280] Resource efficiency: The model uses 1.5GB of memory and is compatible with automotive embedded hardware.

[0281] Implementation Case 2: Health Management of Battery Packs in Energy Storage Power Stations

[0282] Application scenario: A 100MWh energy storage power station needs to monitor the health status of thousands of lithium battery packs, predict the RUL (Recovery Duration and Upgrade) and optimize the replacement plan.

[0283] (1) Data preparation.

[0284] Data source: 300 sets of battery data from the power station monitoring system (cycle count: 2000 times / set), including the following parameters: electrical parameters (voltage, current, internal resistance), thermal parameters (surface temperature, coolant flow rate), and environmental parameters (humidity, altitude).

[0285] Preprocessing: 1) Multi-cell alignment: Align the capacity decay curves of different cells using the DTW algorithm. 2) Anomaly detection: Remove 5% of outlier data using the Isolation Forest algorithm.

[0286] (2) Model configuration.

[0287] Diffusion Transformer Autoencoder (DTAE): 1) Multi-scale architecture: shallow layer captures single-cell fluctuations, deep layer models battery pack co-degradation. 2) Physical constraints: coupled equivalent circuit model (ECM) constrains the ohmic polarization characteristics of the generated trajectory.

[0288] Large Language Model (LLM): 1) Domain Adaptation: Fine-tuning LLaMA-3 on battery maintenance logs to learn fault description semantics. 2) Knowledge Enhancement: Integrating an electrochemical knowledge graph (containing 300+ entities, such as "SEI film thickness" and "lithium deposition rate").

[0289] (3) Deployment plan.

[0290] Cloud-Edge Collaboration: 1) Edge: Each battery cabinet deploys a lightweight DTAE (4-layer Transformer) to perform real-time feature extraction. 2) Cloud: An LLM cluster processes all site data and generates maintenance strategies.

[0291] Dynamic scheduling: Prioritize battery packs with SOH < 90% during high load periods. Regularly (weekly) reassess the entire site's RUL and update the confidence interval.

[0292] (4) Verification results

[0293] Accuracy metrics: 1) Battery pack level MAE = 1.8% (single cell level MAE = 1.5%). 2) Interval coverage (PICP) = 94.7%, with a 40% reduction in false alarm rate.

[0294] Economic benefits: 1) 22% reduction in replacement costs: Premature replacements are avoided through accurate prediction (saving $120,000 annually). 2) 35% reduction in downtime due to failures: Early detection of thermal runaway risk (89% accuracy).

[0295] Explainable output: Battery pack #B23-07 has a RUL of 243 cycles. Due to a sudden increase in internal resistance (+15%), priority maintenance is recommended.

[0296] Cause of association: Insufficient coolant flow leading to localized overheating (LLM related knowledge graph entity)

[0297] (5) Technological advantages

[0298] Scalability: Supports parallel prediction for thousands of battery packs, with cloud processing latency of <2s.

[0299] Multimodal fusion: Joint analysis of electrical-thermal-environmental data reduces MAE by 18% compared to single-modal models.

[0300] Dynamic reliability: Automatically expands the confidence interval width based on humidity changes (e.g., MPIW+20% when humidity > 80%).

[0301] Based on the above embodiments, this application also provides a device for predicting the remaining battery life. Figure 8 This is a schematic diagram of the structure of the battery remaining life prediction device provided in the embodiments of this application, as shown below. Figure 8 As shown, the battery remaining life prediction device 80 includes:

[0302] The acquisition unit 801 is used to acquire battery data of the battery to be predicted; the battery data is at least one of voltage, current, temperature and capacity collected at multiple sampling times;

[0303] Processing unit 802 is used to divide the battery data into M first sequences based on window length and step size; M is an integer greater than or equal to 2;

[0304] The noise unit 803 is used to perform stepwise noise addition processing on the M first sequences using the preset forward diffusion process included in the trained diffusion Transformer autoencoder DTAE, to obtain M first noise-added sequences.

[0305] The feature extraction unit 804 is used to encode the M first noisy sequences and / or the M first sequences using the encoder included in the trained DTAE to obtain M deterministic degradation features;

[0306] The prediction unit 805 is used to obtain N remaining lifetime prediction sequences based on the M deterministic degradation features and using the predictors included in the trained DTAE.

[0307] Uncertainty quantification unit 806 is used to perform uncertainty quantification processing based on the N remaining useful life prediction sequences to obtain remaining useful life prediction results including a first confidence interval.

[0308] In this embodiment, a well-trained large language model is used to extract deterministic degradation features that can effectively capture battery performance degradation patterns. Then, time-series modeling is performed on the deterministic degradation features, which can accurately predict the remaining battery life. Furthermore, by introducing an uncertainty quantification mechanism, confidence interval estimation is added to the battery remaining life prediction sequence, which further improves the accuracy and reliability of the prediction and provides a scientific basis for battery management and maintenance.

[0309] In some embodiments of this application, the prediction unit 805 is further configured to parse each deterministic degradation feature into physical parameters, the physical parameters including capacity decay rate, temperature fluctuation variance, and voltage relaxation time; substitute the physical parameters corresponding to each deterministic degradation feature into a preset battery feature semantic template to obtain a battery feature semantic description corresponding to each deterministic degradation feature; input the M battery feature semantic descriptions into a trained large language model (LLM) for feature enhancement and semantic interpretation to obtain M time-series feature vectors; and input the M time-series feature vectors into the predictor to obtain the N remaining service life prediction sequences.

[0310] In some embodiments of this application, the uncertainty quantization unit 806 is further configured to use the decoder included in the trained DTAE to perform N independent reverse diffusion processes on the M deterministic degradation features to obtain N battery degradation trajectories at future times; N is an integer greater than or equal to 2; and input each of the battery degradation trajectories into the predictor to obtain N remaining lifespan prediction sequences.

[0311] In some embodiments of this application, the uncertainty quantization unit 806 is specifically used to decompose each remaining useful life prediction sequence into low-frequency components and high-frequency components through wavelet transform; obtain low-frequency variance based on the low-frequency components corresponding to each of the N remaining useful life prediction sequences; obtain high-frequency variance based on the high-frequency components corresponding to each of the N remaining useful life prediction sequences; and obtain the remaining useful life prediction result based on the low-frequency variance, the high-frequency variance, and the first confidence interval.

[0312] In some embodiments of this application, the processing unit 802 is further configured to: divide the acquired battery data to be trained into P second sequences based on the window length and the step size; where P is an integer greater than or equal to 2; perform progressive noise addition on the P second sequences using a preset forward diffusion process included in the DTAE to obtain P second noisy sequences; encode the P second noisy sequences and / or the P second sequences using an encoder included in the DTAE to obtain P deterministic degradation feature representations; perform noise prediction on the P deterministic degradation feature representations using a predictor included in the DTAE to obtain P remaining service life prediction values; determine a loss function based on the P real capacity decay rates corresponding to the acquired P second sequences and the P remaining service life prediction values, and train the DTAE based on the loss function to obtain the trained DTAE.

[0313] In some embodiments of this application, the noise unit 803 is further configured to use the preset forward diffusion process and encoder included in the trained DTAE to perform progressive noise addition and encoding on the P second sequences to obtain P deterministic degradation features; based on each deterministic degradation feature, obtain the battery feature semantic description corresponding to each deterministic degradation feature; input the P battery feature semantic descriptions into the LLM for feature enhancement and semantic interpretation to obtain P temporal feature vectors; and perform low-rank adaptive fine-tuning on the LLM based on the P temporal feature vectors to obtain the trained LLM.

[0314] In some embodiments of this application, the acquisition unit 801 is further configured to acquire the current health status of the battery to be predicted; and based on the current health status of the battery to be predicted, adjust the first confidence interval to obtain an adjusted second confidence interval.

[0315] In some embodiments of this application, the processing unit 802 is further configured to use the current health state of the battery to be predicted as input to the adjustment function of the remaining service life to obtain an adjusted value of the remaining service life; and based on the adjusted value of the remaining service life, adjust the first confidence interval to obtain an adjusted second confidence interval.

[0316] In some embodiments of this application, the processing unit 802 is further configured to preprocess the battery data to obtain processed battery data; and to divide the processed battery data into M first sequences based on the window length and the step size.

[0317] This application also provides another electronic device. Figure 9 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 9 As shown, the electronic device 90 includes: a processor 901 and a memory 902 configured to store computer programs capable of running on the processor;

[0318] The processor 901 is configured to execute the method steps in the foregoing embodiments when running a computer program.

[0319] Of course, in practical applications, such as Figure 9 As shown, the various components in the electronic device 90 are coupled together via a bus system 903. It is understood that the bus system 903 is used to achieve communication between these components. In addition to a data bus, the bus system 903 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 9 The general designated all buses as Bus System 903.

[0320] In an exemplary embodiment, this application also provides a computer-readable storage medium for storing a computer program.

[0321] Optionally, the computer-readable storage medium can be applied to any of the methods in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the processor in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0322] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for predicting the remaining lifespan of a battery, characterized in that, The method includes: Obtain battery data of the battery to be predicted; the battery data is at least one of voltage, current, temperature and capacity collected at multiple sampling times; Based on the window length and step size, the battery data is divided into M first sequences; M is an integer greater than or equal to 2; Using the pre-defined forward diffusion process included in the trained Diffusion Transformer Autoencoder (DTAE), the M first sequences are progressively denoised to obtain M first denoised sequences. Using the encoder included in the trained DTAE, the M first noisy sequences and / or the M first sequences are encoded to obtain M deterministic degradation features; Based on the M deterministic degradation features, and using the predictors included in the trained DTAE, N remaining lifetime prediction sequences are obtained. Uncertainty quantification is performed on the N remaining useful life prediction sequences to obtain remaining useful life prediction results including the first confidence interval; Based on the M deterministic degradation features, and using the predictors included in the trained DTAE, N remaining lifetime prediction sequences are obtained, including: The M deterministic degradation features are enhanced and semantically interpreted to obtain M temporal feature vectors; Using the decoder included in the trained DTAE, N independent reverse diffusion processes are performed on the M deterministic degradation features to obtain N battery degradation trajectories at future time points; N is an integer greater than or equal to 2. The M time-series feature vectors or each of the battery degradation trajector are input into the predictor to obtain N remaining lifespan prediction sequences.

2. The method for predicting the remaining service life of a battery according to claim 1, characterized in that, The process of enhancing and semantically interpreting the M deterministic degradation features yields M temporal feature vectors, including: Each of the deterministic degradation features is parsed as a physical parameter, which includes capacity decay rate, temperature fluctuation variance, and voltage relaxation time. Substitute the physical parameters corresponding to each deterministic degradation feature into the preset battery feature semantic template to obtain the battery feature semantic description corresponding to each deterministic degradation feature. The M battery feature semantic descriptions are input into the trained large language model LLM for feature enhancement and semantic interpretation, resulting in M ​​temporal feature vectors.

3. The method for predicting the remaining service life of a battery according to claim 1, characterized in that, The uncertainty quantification process performed on the N remaining useful life prediction sequences to obtain the remaining useful life prediction results includes: Each remaining useful life prediction sequence is decomposed into low-frequency and high-frequency components using wavelet transform. Based on the low-frequency components corresponding to each of the N remaining useful life prediction sequences, the low-frequency variance is obtained; Based on the high-frequency components corresponding to each of the N remaining useful life prediction sequences, the high-frequency variance is obtained; The remaining useful life prediction result is obtained based on the low-frequency variance, the high-frequency variance, and the first confidence interval.

4. The method for predicting the remaining service life of a battery according to claim 1, characterized in that, The method further includes: Based on the window length and the step size, the acquired battery data to be trained is divided into P second sequences; P is an integer greater than or equal to 2. Using the preset forward diffusion process included in DTAE, P second sequences are progressively noise-added to obtain P second noise-added sequences; Using the encoder included in DTAE, the P second noisy sequences and / or the P second sequences are encoded to obtain P deterministic degradation feature representations; Using the predictor included in DTAE, noise prediction is performed on the P deterministic degradation feature representations to obtain P remaining useful life prediction values; Based on the P actual capacity decay rates corresponding to the P second sequences and the P predicted remaining lifetime values, a loss function is determined, and the DTAE is trained based on the loss function to obtain the trained DTAE.

5. The method for predicting the remaining service life of a battery according to claim 4, characterized in that, The method further includes: Using the pre-defined forward diffusion process and encoder included in the trained DTAE, the P second sequences are progressively noise-added and encoded to obtain P deterministic degradation features; Based on each deterministic degradation feature, a semantic description of the battery features corresponding to each deterministic degradation feature is obtained; P battery feature semantic descriptions are input into the LLM for feature enhancement and semantic interpretation, resulting in P time-series feature vectors. The LLM is then subjected to low-rank adaptive fine-tuning based on the P temporal feature vectors to obtain the trained LLM.

6. The method for predicting the remaining service life of a battery according to any one of claims 1 to 5, characterized in that, The method for predicting the remaining battery life also includes: Obtain the current health status of the battery to be predicted; Based on the current health status of the battery to be predicted, the first confidence interval is adjusted to obtain the adjusted second confidence interval.

7. The method for predicting the remaining service life of a battery according to claim 6, characterized in that, The step of adjusting the first confidence interval based on the health state of the battery to be predicted to obtain an adjusted second confidence interval includes: The current health status of the battery to be predicted is used as the input to the adjustment function of the remaining service life to obtain the adjustment value of the remaining service life. Based on the adjustment value of the remaining useful life, the first confidence interval is adjusted to obtain the adjusted second confidence interval.

8. The method for predicting the remaining service life of a battery according to any one of claims 1 to 5, characterized in that, The battery data is divided into M first sequences based on window length and step size, including: The battery data is preprocessed to obtain processed battery data; Based on the window length and the step size, the processed battery data is divided into M first sequences.

9. A device for predicting the remaining lifespan of a battery, characterized in that, The device includes: An acquisition unit is used to acquire battery data of the battery to be predicted; the battery data is at least one of voltage, current, temperature and capacity collected at multiple sampling times; The processing unit is used to divide the battery data into M first sequences based on the window length and step size; M is an integer greater than or equal to 2. The noise unit is used to progressively add noise to the M first sequences using a preset forward diffusion process included in the trained Diffusion Transformer Autoencoder (DTAE) to obtain M first noisy sequences. The feature extraction unit is used to encode the M first noisy sequences and / or the M first sequences using the encoder included in the trained DTAE to obtain M deterministic degradation features; The prediction unit is used to obtain N remaining lifetime prediction sequences based on the M deterministic degradation features and using the predictors included in the trained DTAE. An uncertainty quantification unit is used to perform uncertainty quantification processing based on the N remaining useful life prediction sequences to obtain remaining useful life prediction results including a first confidence interval. The prediction unit is also used to perform feature enhancement and semantic interpretation on the M deterministic degradation features to obtain M temporal feature vectors; Using the decoder included in the trained DTAE, N independent reverse diffusion processes are performed on the M deterministic degradation features to obtain N battery degradation trajectories at future time points; N is an integer greater than or equal to 2. The M time-series feature vectors or each of the battery degradation trajector are input into the predictor to obtain N remaining lifespan prediction sequences.

10. An electronic device, characterized in that, The electronic device includes: a processor and a memory configured to store computer programs capable of running on the processor. Wherein, when the processor is configured to run the computer program, it performs the steps of the method for predicting the remaining battery life as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting the remaining battery life as described in any one of claims 1 to 8.

Images