Battery full-cycle health monitoring method based on multi-scale memory enhancement network

Abstract

The application provides a battery full-cycle health monitoring method based on a multi-scale memory enhancement network. Through multi-scale feature fusion and memory enhancement mechanism, high-precision voltage, current and temperature sensors are used for collecting battery operation data, a second-order Butterworth low-pass filter is used for removing high-frequency noise, a density-based spatial clustering algorithm is used for removing abnormal data points, and a standardization normalization method is used for normalizing the data. The preprocessed data is decomposed by discrete wavelet transform, the wavelet basis function suitable for the characteristics of the battery data is selected, and the feature components under different time scales are obtained. The memory module realizes the association between historical features and current input through the attention mechanism, optimizes the network training by using the adaptive learning rate adjustment strategy, and dynamically adjusts the learning rate according to the gradient change. Based on the battery type, specification and historical operation data, the threshold range of the health state index is determined by statistical analysis method to realize health state monitoring and early warning, and realize precise evaluation and dynamic early warning of the battery full cycle health.

Description

Technical Field

[0001] This invention relates to the field of battery health status monitoring technology, specifically to a battery full-cycle health monitoring system and method using a multi-scale memory enhancement network. By integrating multi-scale feature extraction and memory enhancement mechanisms, it achieves accurate health assessment and early warning throughout the battery's entire life cycle. Background Technology

[0002] With the rapid development of the new energy industry, electrochemical energy storage devices such as lithium-ion batteries and sodium-ion batteries have been widely used in new energy vehicles, large-scale energy storage power stations, and portable electronic devices. Their State of Health (SOH) directly affects the operational safety, energy efficiency, and lifespan of these devices. Industry data shows that a deviation of more than 5% in battery health status can lead to a decrease in energy storage system efficiency of more than 10%, and sudden health failures can even trigger safety accidents such as thermal runaway. Therefore, achieving accurate health monitoring throughout the entire battery lifecycle has significant engineering value and safety implications. Current battery health monitoring technologies are mainly divided into three categories: physical model-based methods, data-driven methods, and hybrid methods. However, all of these have significant technical limitations and cannot meet the high-precision monitoring requirements throughout the entire battery lifecycle.

[0003] Physical model-based methods, represented by equivalent circuit models (ECM) and electrochemical mechanism models (EMM), deduce the health status by establishing mathematical relationships between internal chemical reactions and external characteristics of the battery. The core drawbacks of these methods are: first, model parameters (such as ohmic resistance and diffusion coefficient) rely heavily on prior knowledge and exhibit nonlinear changes with battery aging, resulting in poor adaptability throughout the entire battery lifecycle; second, they cannot capture the multi-scale correlation between microscopic side reactions (such as lithium dendrite growth) and macroscopic performance degradation, often leading to prediction errors exceeding 8% for long-term capacity degradation; and third, they have high computational complexity, making them difficult to apply to real-time monitoring scenarios. For example, a certain EMM-based monitoring system takes hundreds of milliseconds for a single health assessment, far exceeding the engineering-allowed 10ms-level latency requirement.

[0004] Data-driven methods for assessing battery health status through mapping have become a research hotspot in recent years. However, existing data-driven methods have some problems. Monitoring methods based on single neural networks cannot fully utilize multi-scale feature information, while methods based on traditional machine learning lack long-term memory of battery degradation processes, resulting in large errors in health status assessment and making it difficult to meet the requirements of high-reliability scenarios. In addition, the hardware architecture of existing monitoring systems also has bottlenecks: feature extraction and model inference rely heavily on general-purpose processors (CPUs), which have low computational efficiency and cannot meet the parallel monitoring needs of large-scale battery packs (such as thousands of batteries in energy storage power stations); model parameters are fixed and cannot be dynamically updated with battery aging, leading to a continuous decline in monitoring accuracy in later stages.

[0005] Existing technologies suffer from limitations in multi-scale feature capture, historical information fusion, anti-interference capabilities, and engineering adaptability, making it difficult to simultaneously achieve accuracy, real-time performance, and robustness in battery lifecycle health monitoring. Therefore, developing a monitoring method capable of collaboratively extracting multi-scale features, dynamically fusing historical health information, and possessing efficient hardware implementation capabilities has become a critical issue urgently needing to be addressed in this field. Summary of the Invention

[0006] This invention aims to provide a battery full-cycle health monitoring system and method based on multi-scale memory enhancement networks. By using multi-scale feature fusion and memory enhancement mechanisms, it effectively solves the problems of insufficient multi-scale feature capture capability, lack of historical information utilization mechanism, poor anti-interference capability and engineering adaptability, effectively improves the accuracy and robustness of battery full-cycle health status monitoring, and realizes early fault warning and remaining life prediction.

[0007] A battery lifecycle health monitoring method based on multi-scale memory enhancement networks includes the following steps:

[0008] Step S1: Data Acquisition and Preprocessing. High-precision voltage, current, and temperature sensors are used to collect battery operating data. A second-order Butterworth low-pass filter (cutoff frequency 50Hz) is used to remove high-frequency noise. Density-based spatial clustering (DBSCAN) or isolated forest algorithms are used to identify and remove outlier data points. Interpolation algorithms are used to complete the data. A standardization normalization method is then used to normalize the data.

[0009] Step S2: Multi-scale feature extraction uses Discrete Wavelet Transform (DWT) to decompose the preprocessed data into multiple scales, selects wavelet basis functions suitable for the characteristics of battery data (such as db series, sym series), and obtains feature components at different time scales; combines GPU acceleration (such as NVIDIA CUDA framework) to realize parallel computation of wavelet transform and fast Fourier transform (FFT) to improve feature extraction efficiency.

[0010] Step S3: Construction and Training of the Memory-Enhanced Network. A deep learning network containing a memory module is constructed. The memory module uses an attention mechanism to associate historical features with the current input: The similarity (dot product or cosine similarity) between the current input features and historical features in the memory unit is calculated; the similarity is normalized using the Softmax function to generate attention weights; historical features are weighted and summed based on the attention weights to obtain memory information related to the current input; an adaptive learning rate adjustment strategy (such as Adam or RMSProp algorithms) is used to optimize network training, dynamically adjusting the learning rate according to gradient changes to improve convergence speed and stability; the network hardware implementation uses an FPGA acceleration platform to store the trained model parameters in on-chip memory for fast inference.

[0011] Step S4: Health status monitoring and early warning. Based on battery type, specifications and historical operating data, the threshold range of health status indicators (such as capacity decay rate and internal resistance growth rate) is determined through statistical analysis methods. When the real-time monitoring indicators exceed the threshold, a multi-level early warning mechanism is triggered, and data storage and early warning information are pushed through a server-side Java program.

[0012] Step S1 includes:

[0013] S11: High-precision voltage, current and temperature sensors are used to collect battery operation data in real time, with a sampling frequency of ≥1 kHz, to ensure that the data granularity meets the requirements of battery status analysis.

[0014] S12: Design of a second-order Butterworth low-pass filter. Based on the characteristic that high-frequency interference during battery operation is mainly concentrated in the frequency band above 50Hz, a second-order Butterworth low-pass filter is designed. The filter transfer function is:

[0015]

[0016] in The cutoff angular frequency, = 2π×50Hz = 314rad / s. Discretized coefficients are generated using MATLAB's butter function. The analog filter is converted into a digital filter using the bilinear transform method. When the sampling frequency is set to 1kHz, the normalized cutoff frequency of the digital filter is 0.1 (50Hz / 500Hz, i.e., half of the sampling frequency is the Nyquist frequency).

[0017] S13: When the data exhibits high-dimensional sparsity, the Isolation Forest algorithm is used: the number of trees is set to 100, the sample subset size of each tree is 256 data points, and the outlier threshold is set to 0.6 (determined through training with historical normal data). An outlier score is calculated for each sample in the dataset; when the score exceeds the threshold, it is considered outlier and removed.

[0018] For missing data after outlier removal, a piecewise linear interpolation algorithm is used to complete the missing data:

[0019]

[0020] in The two nearest valid data points are the two sides of the missing time period.

[0021] S14: Use the Z-score standardization method to unify the data scale and process it independently according to the feature dimension:

[0022] Splitting the dataset: Divide the complete dataset D into a training set. With test set ;

[0023] Calculate training set statistics: for each column j (voltage, current, temperature, etc.) in the training set;

[0024] Calculate the mean:

[0025] Calculate the standard deviation:

[0026] Transform the training set data: HK

[0027] Transform the test set and new data: using the data obtained from the training set. Perform the same conversion:

[0028] .

[0029] Step S2 includes:

[0030] S21: Wavelet basis function selection should be adapted to battery characteristics, prioritizing the Daubechies (db) series (e.g., db4) or Symlets (sym) series (e.g., sym5). Tight support and approximate symmetry should be chosen to accommodate the transient abrupt changes in battery data. Adaptive verification calculations of reconstruction errors for different wavelet bases should be performed, selecting the one with the smallest MSE. Discrete wavelet transform (DWT) should be used to perform multi-scale decomposition on the preprocessed data, selecting wavelet basis functions suitable for the battery data characteristics to obtain feature components at different time scales.

[0031] S22: Multi-scale decomposition of battery data based on DWT: Assume the preprocessed data is... (Normalized time series), the scaling function is The wavelet function is (Determined by the selected basis functions).

[0032] Iterative decomposition (j=1 to J)

[0033]

[0034] in This represents the low-frequency component (long-term feature) of the j-th layer. This represents the high-frequency components (short-term features) of the j-th layer.

[0035] S23: Task partitioning: DWT layers are decomposed and distributed to different CUDA Streams. FFT is used for frequency domain filtering (such as abnormal frequency band detection) and executed in parallel with wavelet transform. Core optimizations include DWT convolution kernels, shared memory to accelerate convolution operations, and FFT kernels that call the cufftExecR2C function of the cuFFT library. Key technologies for GPU-CUDA parallel acceleration include DWT convolution kernel optimization using texture memory to cache input data, reducing global memory access latency, unfolding convolution loops, and utilizing GPU warp-level parallelism. FFT and wavelet pipelines are used, with FFT applied to detail components to detect abnormal frequency bands in the 0.5-2Hz range (corresponding to micro-short circuits in batteries).

[0036] The multi-scale feature fusion mechanism employs attention weighting: (Weight allocation of importance across scales), weighted output: (f: Neural network function).

[0037] Step S3 includes:

[0038] S31: A memory enhancement module is added after the multi-scale extraction module. Its function is to efficiently record and store normal data within this module. Furthermore, the memory enhancement module can read and update the memory items during alternating training phases, thus enabling the module to adapt to new environments. The memory enhancement module is shown in the figure, and it can be divided into two parts: reading and updating memory items. First, the feature map of size H×W×C generated by the encoder is obtained. The memory item is denoted as... The query item is denoted as Both have a size of 1×1×C, with M and K representing the number of memory terms and query terms, respectively. Since the feature map consists of K query terms, K = H×W. Subsequently, based on the query terms... Memory items related to normal behavior The similarity is used for matching, and then the similarity is used as a weight to construct a new feature map. This feature map is not only more similar to the original image, but also fits video frames of normal behavior samples better.

[0039] S32: The memory item retrieval process generates entirely new feature maps by comparing the matching degree between query items and memory items. Specifically, this is achieved by calculating the matching degree between each query item and the memory items. With memory items The cosine values ​​between these features are used to generate an M×K two-dimensional cosine similarity matrix. Then, a softmax function is applied along the vertical direction, and the weights for each feature are calculated.

[0040]

[0041] This represents the k-th query item at step t. This represents the m-th memory item. express and The dot product is used to measure the similarity between them, and exp is an exponential function used to convert the similarity value to a non-negative value. Then, based on each query term... and their corresponding matching weights , with all memory items Calculate the weighted sum. Repeat the above operation for K features to obtain a new feature map, thereby achieving the reading of the memory item.

[0042]

[0043] By performing a memory item read operation on each query in the memory enhancement module, the transformed new features are obtained. The image will be shown later. With new features The images are stitched together, and the final stitched result is input into the encoder. When the feature map obtained by the decoder using the memory enhancement module is updated based on normal behavior patterns, it can reduce the generalization ability of the autoencoder. While enhancing the similarity to normal patterns, it also amplifies the difference between normal and abnormal patterns, thus achieving better results.

[0044] The memory update method involves adding variables to the existing memory base, using the matching weights mentioned above. As shown in the following formula:

[0045]

[0046] in It is the L2 norm. For memory module and query The m-th item in the corresponding index set, and through... The corresponding memory entries in the memory module are updated. Then, based on the above formula, the calculation direction of the softmax function is adjusted, changing the vertical direction to the horizontal direction, and calculations are performed to match the weights, resulting in the following formula:

[0047]

[0048] Again, query the index For reference, the matching weights calculated above are reconstructed using regularization, as shown in the following formula:

[0049]

[0050] The variable is obtained by multiplying and summing the corresponding query terms, and then adding this variable to the initial memory term and the change amount to obtain the updated memory term.

[0051] S33: The loss function is the "optimization objective" of model training, guiding the network to learn the core features of normal behavior through multi-dimensional constraints. This paper designs a combined loss function of feature consistency loss and feature separation loss to balance the correlation and diversity of memory modules.

[0052] Feature Consistency Loss Its function is to constrain the correlation between "query item" (current input feature) and "memory item" (historical normal feature) in the memory module, making the memory item closer to the core features of normal behavior.

[0053]

[0054] in express Norm, used to measure the Euclidean distance between two features.

[0055] Feature separation loss To avoid redundancy in memory items within the memory module (i.e., multiple memory items storing highly similar features), a feature separation loss is designed to increase the distance between the query item and the next most similar memory item, thus ensuring the diversity of the memory. Its mathematical expression is:

[0056]

[0057] in For secondary similarity memory index, =0.5 is the interval parameter.

[0058] Total loss function The total loss is a weighted sum of the feature consistency loss and the feature separation loss, calculated using a balance coefficient. Adjust the contributions of both.

[0059]

[0060] To achieve efficient model convergence and stable parameter updates, the Adam adaptive optimization algorithm and training process are adopted, and a dynamic learning rate adjustment strategy is designed. The optimizer selected is the Adam (Adaptive Moment Estimation) optimizer, which combines the advantages of momentum method and adaptive learning rate (RMSProp), and can effectively alleviate gradient oscillation and accelerate convergence.

[0061] S34: The network hardware implementation utilizes an FPGA acceleration platform, where trained model parameters are stored in on-chip memory for hardware acceleration and inference. The FPGA deployment scheme employs model embedding, with trained network parameters stored in the FPGA's on-chip memory (Block RAM). Parallel convolution and attention computations are performed using the FPGA pipeline architecture. Input video frames (256×256 resolution) are preprocessed, and low-latency anomaly detection is achieved through hardware-accelerated real-time inference.

[0062] Step S4 includes:

[0063] S41: Health Status Monitoring and Early Warning. Based on battery type, specifications, and historical operating data, statistical analysis methods are used to determine the threshold range of health status indicators (such as capacity decay rate and internal resistance growth rate). A comprehensive and systematic statistical analysis is conducted on health indicator data under historical normal operating conditions, based on equipment type (covering mainstream battery types such as lithium batteries and lead-acid batteries) and equipment specifications (including key parameters such as capacity and nominal voltage).

[0064] S42: The formula for calculating the capacity decay rate is as follows: ,in Represents the initial capacity of the equipment. This represents the current capacity of the device. Analysis of extensive experimental data and real-world operational cases shows that when the capacity decay rate... At this time, the equipment is usually within its normal operating range. The formula for calculating the internal resistance growth rate is: ,here Indicates the initial internal resistance of the device. This indicates the current internal resistance of the equipment. Similarly, based on extensive industry practice and research findings, the rate of increase in internal resistance is determined. At that time, the equipment was operating well.

[0065] S43: Status assessment provides a reliable data foundation. The status assessment process is based on the health score output from the network, combined with pre-determined indicator thresholds to comprehensively judge the health status of the equipment. When real-time monitored indicators exceed the thresholds, a multi-level early warning mechanism is triggered, and the real-time status assessment results are transmitted to the server at high speed via Ethernet. The server-side deploys a Java program developed based on the Spring Boot framework to realize data storage and early warning information push.

[0066] This invention provides a battery lifecycle health monitoring method based on a multi-scale memory enhancement network. Compared with existing technologies, the advantages of this invention are:

[0067] 1. Multi-scale decomposition of battery data is achieved through discrete wavelet transform, capturing degradation features at different time granularities and overcoming the limitations of traditional single-scale feature extraction.

[0068] 2. Construct a network structure containing historical feature memory units, and realize the temporal dependency modeling of battery degradation process through similarity calculation and attention weight allocation to improve the long-term monitoring accuracy.

[0069] 3. Based on statistical analysis of historical data throughout the battery's entire lifecycle, dynamically adjust the health status threshold to adapt to the monitoring needs of different usage stages. Attached Figure Description

[0070] Figure 1 Wavelet Transform (DWT) Multi-Scale Decomposition Battery Data Flowchart

[0071] Figure 2 System Flowchart of Health Status Monitoring and Early Warning System for Memory Enhancement Networks

[0072] Figure 3 GPU Parallel Accelerated Computation Flowchart Detailed Implementation

[0073] The following description, in conjunction with preferred embodiments of the present invention, should be understood as follows. These preferred embodiments are for illustration and explanation only and are not intended to limit the scope of the invention.

[0074] A battery lifecycle health monitoring system based on a multi-scale memory enhancement network includes a data acquisition module, a data preprocessing module, a multi-scale feature extraction module, a memory enhancement network module, and a health status monitoring and early warning module. The data acquisition module utilizes high-precision, high-reliability sensors such as voltage, current, and temperature sensors. It also features data caching capabilities, employing a high-speed cache chip capable of buffering at least 10 minutes of data to handle brief interruptions during data transmission. The data preprocessing module implements a hardware filtering circuit using a second-order Butterworth low-pass filter with a cutoff frequency set to 50Hz based on the noise characteristics of the battery data to remove high-frequency noise. The software algorithm is written in Python. The hardware acceleration unit of the multi-scale feature extraction module uses an NVIDIA GPU, leveraging the CUDA parallel computing framework to accelerate the computation of algorithms such as wavelet transform and fast Fourier transform. The software algorithm is implemented using TensorFlow or PyTorch deep learning frameworks, facilitating algorithm development and optimization. The hardware architecture of the memory augmentation network module utilizes an FPGA-based deep learning acceleration platform. The trained multi-scale memory augmentation network model parameters are stored in the FPGA's on-chip memory, enabling rapid inference computation. Furthermore, this module features model update capabilities; when new battery data or model optimization algorithms are available, the model parameters can be updated via the network or external storage devices. The health status monitoring and early warning module is a crucial component ensuring timely and accurate monitoring of equipment health status, primarily comprising two core steps: data acquisition and status assessment.

[0075] The data acquisition process employs advanced sensor technology to capture various parameters during equipment operation in real time. The sampling period is precisely set to 1 second to ensure the acquisition of real-time dynamic information on the equipment's operating status. The collected data is not directly used for subsequent analysis but undergoes a rigorous preprocessing procedure. The preprocessed data is then input into the FPGA inference module, providing a reliable data foundation for status assessment. The model training process aims to minimize a preset loss function through an optimization algorithm, enabling the network to stably learn characteristic patterns of normal behavior and effectively associate and update historical information through a memory module. This is mainly explained below:

[0076] 1. Dataset partitioning: The training data comes from time-series monitoring data under normal device operation (such as voltage-time curves and current-time curves for 500 complete charge-discharge cycles of a lithium-ion battery). The data sampling frequency is 1Hz, and the single sample length is 1000 time steps. The dataset is randomly divided into a training set and a validation set in an 8:2 ratio: the training set is used for iterative parameter updates, and the validation set is used to monitor the model overfitting trend and dynamically adjust the training strategy.

[0077] 2. The Adam optimizer is used with an initial learning rate (Lr=0.001). A dynamic adjustment strategy is employed: when the validation set loss decreases for three consecutive epochs, Lr is maintained; when the loss stagnates (rate of change < 1e-4), Lr is reduced to (Lr = Lr×0.5) (minimum Lr=1e-5). Iterative training is conducted with a batch size of 32, epochs=100, and an early stopping strategy (termination occurs if the validation set loss shows no improvement after 10 epochs). Hardware acceleration and inference implementation utilize an FPGA deployment scheme for model solidification, with trained network parameters stored in the FPGA's on-chip memory (Block RAM). Parallel convolution and attention operations are performed using an FPGA pipeline architecture. Input video frames (256×256 resolution) are preprocessed, and real-time inference achieves low-latency anomaly detection through hardware acceleration.

[0078] 3. Health Status Monitoring and Early Warning Implementation Process: When monitoring and issuing early warnings for equipment health status, the primary task is to accurately determine the health indicator thresholds. The status assessment stage is based on the health score S output by the network, combined with the pre-determined indicator thresholds, to comprehensively judge the health status of the equipment.

[0079] The specific judgment criteria are as follows:

[0080] When health score Meanwhile, capacity decay rate And the internal resistance growth rate At that time, it is determined that the equipment is in normal operating condition.

[0081] When the health score meets If either the capacity decay rate or the internal resistance growth rate exceeds the set threshold, the system will trigger an early warning mechanism.

[0082] When health score If either the capacity decay rate or the internal resistance growth rate exceeds the set threshold, the system will issue an alarm signal, indicating that the device may have a serious health problem.

[0083] 4. Construction and Application of a Multi-Level Early Warning Mechanism: To ensure timely and effective handling of abnormal equipment health status, a comprehensive multi-level early warning mechanism was constructed. For early warning push notifications, the FPGA inference module transmits real-time status assessment results to the server via high-speed Ethernet. A Java program developed based on the Spring Boot framework is deployed on the server side, which efficiently receives the status data transmitted from the FPGA. For data storage, a MySQL database is used as the data storage medium. The received equipment status data, including precise timestamps, specific values ​​of various health indicators, and corresponding status tags, is written to the database according to a standardized structure for subsequent querying, analysis, and traceability. For early warning information push notifications, modern communication and mobile application technologies are fully utilized. For early warning level information, the Alibaba Cloud SMS API is used to send the warning content as an SMS to the mobile terminals of relevant personnel; simultaneously, a dedicated APP can also push early warning information to maintenance terminals to ensure that relevant personnel receive warning alerts in a timely manner. For alarm-level information, in addition to the aforementioned SMS and APP push methods, an audible and visual alarm device is also triggered simultaneously. Through strong visual and auditory signals, the device immediately draws the attention of relevant personnel so that effective countermeasures can be taken quickly to ensure the safe and stable operation of the equipment.

[0084] This document describes the principles and implementation methods of the present invention. The description of the above-described implementation methods is intended to aid in understanding the methods and core ideas of the present invention. For those skilled in the art, variations in specific implementation methods and application scope will exist based on the concept of the present invention. Therefore, the content of this specification should not be considered as a limitation of the present invention. Those skilled in the art to which this invention pertains can make diverse modifications and additions to the described specific embodiments, or use similar methods to replace them, without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.

Claims

1. A method for full-cycle battery health monitoring based on multi-scale memory enhancement networks, characterized in that... include: S1: Data acquisition and preprocessing; S2: Multi-scale feature extraction; S3: Construction and training of memory enhancement networks; S4: Health status monitoring and early warning.

2. The battery life cycle health monitoring method based on multi-scale memory enhancement network according to claim 1, characterized in that... Step S1 includes: S11: Data acquisition and preprocessing uses high-precision voltage, current, and temperature sensors to collect battery operating data. It has a data caching function and uses a high-speed cache chip that can cache at least 10 minutes of data to cope with brief interruptions during data transmission. S12: The hardware filtering circuit uses a second-order Butterworth low-pass filter with a cutoff frequency set to 50Hz based on the noise characteristics of the battery data, in order to remove high-frequency noise. S13: For the collected abnormal data points, a density-based spatial clustering algorithm is used to identify and remove them, and then interpolation is performed to complete them; S14: During normalization, a standardized normalization method is used, and the formula is: Where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data. This is the normalized data.

3. The battery life cycle health monitoring method based on multi-scale memory enhancement network according to claim 1, characterized in that step S2 includes: S21: The preprocessed data is decomposed into multiple scales using Discrete Wavelet Transform (DWT), and wavelet basis functions suitable for the characteristics of battery data are selected to obtain the feature components at different time scales. S22: Multi-scale decomposition of battery data based on DWT: Assume the preprocessed data is... (Normalized time series), the scaling function is The wavelet function is (Determined by the selected basis functions). Iterative decomposition (j=1 to J) in This represents the low-frequency component (long-term characteristic) of the j-th layer. This represents the high-frequency component (short-term feature) of the j-th layer. S23: Combine GPU acceleration (such as NVIDIA CUDA framework) to achieve parallel computation of wavelet transform and fast Fourier transform (FFT) for multi-scale feature extraction.

4. The battery life cycle health monitoring method based on multi-scale memory enhancement network according to claim 1, characterized in that... Step S3 includes: S31: Construction and Training of Memory Enhancement Network. Construct a deep learning network containing a memory module. The memory module uses an attention mechanism to associate historical features with the current input. S32: Calculate the similarity between the current input feature and the historical features stored in the memory unit. Use methods such as dot product and cosine similarity to calculate the attention weight based on the similarity. Normalize the attention weight using the Softmax function to generate the attention weight. Based on the attention weight, sum the historical features in a weighted manner to obtain the memory information related to the current input. S33: When training a multi-scale memory reinforcement network, an adaptive learning rate adjustment strategy is adopted to automatically adjust the learning rate according to the changes in gradients during training, thereby improving the convergence speed and stability of training. S34: The network hardware implementation uses an FPGA acceleration platform to store the trained model parameters in on-chip memory, enabling fast inference.

5. The battery life cycle health monitoring method based on multi-scale memory enhancement network according to claim 1, characterized in that... Step S4 includes: S41: Health Status Monitoring and Early Warning. Based on battery type, specifications, and historical operating data, statistical analysis methods are used to determine the threshold range of health status indicators (such as capacity decay rate and internal resistance growth rate). A comprehensive and systematic statistical analysis is conducted on health indicator data under historical normal operating conditions, based on equipment type (covering mainstream battery types such as lithium batteries and lead-acid batteries) and equipment specifications (including key parameters such as capacity and nominal voltage). S42: The formula for calculating the capacity decay rate is as follows: ,in Represents the initial capacity of the equipment. This represents the current capacity of the device. Analysis of extensive experimental data and real-world operational cases shows that when the capacity decay rate... At this time, the equipment is usually within its normal operating range. The formula for calculating the internal resistance growth rate is: ,here Indicates the initial internal resistance of the device. This indicates the current internal resistance of the equipment. Similarly, based on extensive industry practice and research findings, the rate of increase in internal resistance is determined. At that time, the equipment was operating well. S43: Status assessment provides a reliable data foundation. The status assessment process is based on the health score output from the network, combined with pre-determined indicator thresholds to comprehensively judge the health status of the equipment. When real-time monitored indicators exceed the thresholds, a multi-level early warning mechanism is triggered, and the real-time status assessment results are transmitted to the server at high speed via Ethernet. The server-side deploys a Java program developed based on the Spring Boot framework to realize data storage and early warning information push.

Images