Single-junction distributed sensing lead-acid battery full-dimension monitoring and evaluation method

Abstract

The present application relates to the technical field of substation lead-acid battery state evaluation, and particularly relates to a single-section distributed sensing lead-acid battery full-dimension monitoring and evaluation method. The method configures an independent intelligent sensor node for each single-section lead-acid battery in a battery pack, each node is connected in series through an RS485 bus to form a network and independently synchronously collects single-section voltage, single-section temperature and AC internal resistance under electrical isolation; after time alignment and preprocessing of multi-node data by an edge aggregation node, single-section voltage, temperature, internal resistance and capacity attenuation characteristics calculated based on historical data are extracted to construct a device-side multi-dimensional feature vector; a two-stage training health evaluation model is adopted to calculate the difference between single sections according to the evaluation results of each single section and the group average value, and to output a key maintenance prompt for single sections with a difference exceeding a difference threshold.

Description

Technical Field

[0001] This invention relates to the technical field of substation lead-acid battery condition assessment, and in particular to a method for comprehensive monitoring and assessment of a single distributed sensing lead-acid battery. Background Technology

[0002] Lead-acid battery banks are commonly used as backup power in DC systems of substations. These battery banks consist of multiple cells connected in series and operate under float charging high-voltage conditions. Traditional assessments of battery health often rely on single-parameter threshold judgments or simple statistical analyses, which fail to comprehensively reflect the actual aging and fault status of individual cells. Some solutions employ centralized data acquisition and cloud-based analysis, but these suffer from data transmission delays, strong communication dependencies, and difficulties in independently isolating and collecting individual cell parameters under series high-voltage conditions.

[0003] Existing single-cell data acquisition solutions are mostly geared towards automotive batteries or lithium battery packs, failing to fully consider the electrical isolation, synchronous sampling, and quantification requirements for inter-cell state differences under the high-voltage floating charge conditions of lead-acid batteries in substations. Some solutions introduce multi-sensor and artificial intelligence for health assessment, but often only provide functional descriptions of the sensors, AI algorithms, and communication solutions, without disclosing the A / D converter bit depth, specific internal resistance detection methods, model structure, and training details, resulting in insufficient disclosure. Furthermore, the generalized description of "multi-sensor + AI" easily overlaps with existing solutions, making the innovation unclear. In addition, how to establish personalized assessment models for target batteries with small sample sizes and how to provide interpretable prediction confidence intervals are also problems that existing solutions have not fully addressed.

[0004] Therefore, there is a need for a comprehensive monitoring and evaluation method for single-cell distributed sensing lead-acid batteries that can independently and synchronously collect and fuse multi-dimensional parameters, and perform personalized health assessments and differential quantification for each cell under series float charging conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a comprehensive monitoring and evaluation method for single-cell distributed sensing lead-acid batteries, in order to solve the problems mentioned in the background art, such as the difficulty in independently isolating and collecting individual cell parameters under series high-voltage float charging conditions, the difficulty in quantifying the state differences between individual cells, the insufficient accuracy of personalized evaluation under small sample conditions, and the lack of interpretable prediction intervals for health evaluation.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery, comprising: configuring an independent intelligent sensor node for each individual lead-acid battery in the battery pack; each intelligent sensor node integrating a voltage detection unit, a temperature detection unit, and an AC injection internal resistance detection unit, and connecting them in series via an RS485 bus; each intelligent sensor node independently and synchronously acquiring the voltage, temperature, and AC internal resistance of the corresponding cell under electrical isolation, obtaining a single-cell parameter sequence; the edge aggregation node performing timestamp alignment and preprocessing on each single-cell parameter sequence to obtain time-aligned multi-dimensional parameters of the single cell; extracting the voltage, temperature, internal resistance, and capacity decay characteristics calculated based on historical data for each cell, constructing a device-side multi-dimensional feature vector; inputting the device-side multi-dimensional feature vector into a health assessment model obtained through two-stage training, outputting the estimated health status value of the single-cell battery and the prediction range of the estimated value; calculating the inter-cell difference based on the health status estimation results of each cell and the group mean, and outputting a key maintenance prompt for the cell when the inter-cell difference exceeds a preset difference threshold.

[0007] Furthermore, independent synchronous acquisition includes: each smart sensor node uses a high-precision A / D converter to sample the voltage of individual cells; the internal resistance detection unit injects a sinusoidal AC signal of a set amplitude into the corresponding cell and collects the response voltage to calculate the AC internal resistance; each node starts acquisition according to the unified sampling trigger time issued by the edge aggregation node, so that the voltage, temperature and internal resistance sampling times of each cell in the same battery pack are aligned within the time window.

[0008] Furthermore, the data acquisition under electrical isolation conditions includes: the signal acquisition side of each smart sensor node is isolated from the bus communication side through an isolation power supply and signal isolation devices, so that the high voltage difference between the series-connected units does not form a common-mode path through the acquisition loop; the address of each node is allocated by the edge aggregation node through an automatic encoding process, which includes the node sending an address request frame carrying a physical identifier after power-on, the aggregation node responding with an unoccupied address, the node confirming and locking the address, and the aggregation node detecting address conflicts and reallocating them when conflicts occur.

[0009] Furthermore, the two-stage training health assessment model is obtained as follows: In the first stage, a general global model is trained using the local datasets of multiple distributed battery nodes as training samples and a collaborative training method with shared model parameters. During the collaborative training process, each node only exchanges model parameters and not the original collected data. In the second stage, a small amount of local labeled data of the target battery node is used as fine-tuning samples to fine-tune the general global model and obtain a personalized assessment model for the target battery node.

[0010] Furthermore, the health assessment model uses a lightweight convolutional neural network as its base network and introduces a channel attention mechanism into the network; the amount of locally labeled data is determined based on the historical capacity records of the target battery nodes, and each target node contains at least one labeled sample corresponding to a complete charge-discharge cycle; the model parameters shared during the collaborative training phase include at least the network weight parameters, but the original voltage, temperature, and internal resistance time series of each node are not shared.

[0011] Furthermore, the prediction interval of the health status estimate is obtained as follows: the health status point estimate and uncertainty characterization are output based on the output head of the enhanced Transformer architecture, and the upper and lower bounds of the prediction interval are calculated based on the uncertainty characterization; when the unit is in the stable aging stage, the width of the prediction interval is less than the set width threshold; when the unit is in the accelerated decay stage, the width of the prediction interval is not less than the set width threshold.

[0012] Furthermore, the inter-unit variability is calculated as follows: the mean of the estimated health status of all units within the group is taken as the group health mean, and the absolute value of the difference between the estimated health status of each unit and the group health mean is taken as the inter-unit variability of that unit; or the mean of the AC internal resistance of all units within the group is taken as the group internal resistance mean, and the ratio of the absolute value of the difference between the internal resistance of each unit and the group internal resistance mean to the group internal resistance standard deviation is taken as the internal resistance variability of that unit, and the health status variability and internal resistance variability are weighted and fused to obtain the comprehensive inter-unit variability.

[0013] Furthermore, the method also includes: based on the voltage, temperature, internal resistance, and health status estimates of each individual unit, real-time alarms are triggered for any of the following conditions: overvoltage, undervoltage, overtemperature, internal resistance exceeding limits, and health status below the health threshold; and the parameters, health status estimates, prediction ranges, and key maintenance prompts of each individual unit are organized into an operating status record according to the individual unit identifier, and reported to the monitoring center via an RS485 bus and an optical fiber link converted by the optical modem.

[0014] Furthermore, the measurement accuracy of AC internal resistance reaches ±1% under the set test conditions, and the voltage sampling accuracy reaches 0.2% under the set test conditions; the set test conditions include at least the ambient temperature range, the range of individual cell state of charge, float charging condition, and the amplitude and frequency of the injected sinusoidal AC signal.

[0015] Furthermore, the method also includes an evaluation report generation step: statistically analyzing the voltage distribution, temperature distribution, internal resistance distribution, and health status distribution of each individual unit at a preset period, generating a comprehensive evaluation report that includes statistical tables, trend curves, and key maintenance tips. The data sources, statistical definitions, calculation rules, and output fields of the comprehensive evaluation report are defined in the report template.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: by using edge aggregation node time alignment and multi-dimensional feature construction, it achieves full-dimensional fusion of voltage, temperature, internal resistance, and capacity decay features; through a two-stage trained health assessment model, it obtains a general model by using multi-node collaborative training while protecting the data privacy of each node, and obtains a personalized model by fine-tuning with a small amount of local data from the target node, thus balancing assessment accuracy and generalization ability under small sample conditions; by quantifying the uncertainty of the predicted interval output, it provides a confidence reference for operation and maintenance decisions; and by quantifying the differences between individual units, it identifies lagging units early and achieves on-demand maintenance. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery includes steps S1 to S6. Step S1: Configure an independent intelligent sensor node for each individual lead-acid battery in the battery pack. Each intelligent sensor node integrates a voltage detection unit, a temperature detection unit, and an AC injection internal resistance detection unit, and is networked in series via an RS485 bus. The signal acquisition side and bus communication side of each intelligent sensor node are isolated by an isolation power supply and signal isolation devices, ensuring that the high voltage difference between the series-connected cells does not form a common-mode path through the acquisition loop. The node address is allocated by the edge aggregation node through an automatic encoding process: after power-on, the node sends an address request frame carrying a physical identifier; the aggregation node responds that the address is not occupied; the node confirms and locks the address; the aggregation node detects address conflicts and reassigns addresses when conflicts occur, thus ensuring the uniqueness of node addresses within the same bus channel.

[0020] In one exemplary implementation, the microcontroller of the intelligent sensor node can be an STM32L4 series or GD32E230 series low-power MCU with a main frequency of not less than 48MHz, and built-in SRAM of not less than 32KB and Flash of not less than 128KB; the high-precision A / D converter of the voltage detection unit can be a 24-bit Σ-Δ ADC (such as ADS131M04 or CS1237), and the reference voltage source can be a low-temperature drift reference (such as REF3030, with a temperature drift of not more than 25ppm / ℃), with a single-unit voltage sampling resolution of... The efficiency is better than 0.5mV; the AC injection type internal resistance detection unit can inject a sinusoidal AC signal with an amplitude of 10mA to 100mA and a frequency of 1kHz into the corresponding cell. The response voltage is collected by a precision sampling resistor and an instrumentation amplifier (such as AD8421). The AC internal resistance is calculated as the ratio of the effective value of the response voltage to the effective value of the injected current. The temperature detection unit can use a DS18B20 one-wire digital temperature sensor or a PT100 with a Σ-ΔADC for four-wire acquisition. The temperature measurement range is -20℃ to +85℃ and the accuracy is ±0.5℃. The RS485 bus can use SP3485 or MAX3485 transceivers, with a maximum of 64 nodes connected to a single channel. Each end of the bus is equipped with a 120Ω terminating resistor. The signal acquisition side and the bus communication side are electrically isolated through an isolated DC-DC power supply module (such as B0505S) and a magnetic coupler isolator (such as ADuM1201) or optocoupler (such as 6N137). The communication protocol can be SNS protocol or ModbusRTU protocol, and the default baud rate can be configured to 9600, 19200 or 38400bps. The above MCU model, ADC model, isolation device and baud rate are all exemplary values. Those skilled in the art can adapt them according to the substation's rated voltage, capacity and manufacturer parameters. The specific model or parameters are alternative implementation methods and do not constitute a limitation on the protection scope.

[0021] Step S2: Each smart sensor node independently and synchronously acquires the voltage, temperature, and AC internal resistance of its corresponding cell under electrical isolation, obtaining a single-cell parameter sequence. The voltage detection unit uses a high-precision A / D converter to sample the cell voltage; the internal resistance detection unit calculates the AC internal resistance by injecting a sinusoidal AC signal of a set amplitude into the corresponding cell and acquiring the response voltage; the temperature detection unit uses a surface-mounted temperature sensor to detect the cell surface temperature. Each node starts acquisition according to the unified sampling trigger time issued by the edge convergence node, so that the sampling times of voltage, temperature, and internal resistance of each cell in the same battery pack are aligned within the time window. In this embodiment, the measurement accuracy of AC internal resistance reaches ±1% under the set test conditions, and the voltage sampling accuracy reaches 0.2% under the set test conditions; the set test conditions include the ambient temperature range, the cell state of charge range, float charging conditions, and the amplitude and frequency of the injected sinusoidal AC signal.

[0022] In one exemplary implementation, the unified sampling trigger time can be sent by the edge aggregation node via a broadcast sampling trigger frame. After receiving the trigger frame, each node starts sampling at the same clock edge, so that the sampling times of voltage, temperature, and internal resistance of each cell in the same battery pack fall within the same time window (e.g., the window width is no more than 100ms) to achieve time alignment. The sampling period can be configured from 1s to 60s, with a default of 5s. Digital filtering can use moving average filtering or first-order low-pass filtering, with a cutoff frequency no more than one-tenth of the sampling frequency. Outlier removal can use the 3σ criterion or a ±20% threshold based on the historical average. Missing value interpolation can use linear interpolation or forward padding. The test conditions can be set as follows: ambient temperature -10℃ to +55℃, cell state of charge 20% to 100%, float charge condition i.e., cell terminal voltage 2.20V to 2.30V, injected sinusoidal AC signal amplitude 50mA ±5%, frequency 1kHz ±1%. Under these test conditions, the AC internal resistance measurement accuracy reaches ±1%, and the voltage sampling accuracy reaches 0.2%.

[0023] Step S3: The edge aggregation node performs timestamp alignment and preprocessing on the parameter sequences of each segment to obtain time-aligned multidimensional parameters for each segment. Preprocessing includes digital filtering, outlier removal, missing value interpolation, and unit unification. Since each node starts acquisition based on a unified sampling trigger time, the edge aggregation node aligns the parameter sequences of each segment by timestamp using a unified clock reference, forming a multidimensional parameter table organized by individual unit identifiers.

[0024] In one exemplary implementation, the edge aggregation node can adopt an STM32H7 series MCU (Cortex-M7 core, main frequency not less than 400MHz) or an equivalent edge computing chip, with built-in SRAM not less than 512KB and Flash not less than 1MB, and equipped with an RTC real-time clock or IEEE1588 hardware timestamp unit to achieve microsecond-level time alignment; the preprocessed single-segment multidimensional parameters are organized into structured records according to the individual identifier (such as SNS address), the acquisition timestamp and the parameter values ​​of each dimension, which can be temporarily stored in the local cache of the edge aggregation node (such as SPIFlash or SD card, with a storage capacity of not less than 32MB, which can save no less than 30 days of historical data), and supports retrieval by time interval and individual identifier.

[0025] Step S4: Extract the voltage, temperature, internal resistance, and capacity decay characteristics of each cell based on historical data to construct a multi-dimensional feature vector on the device side. The capacity decay characteristics can be calculated based on the capacity decay trend with the number of cycles or operating time in the historical capacity record of the cell. The multi-dimensional feature vector on the device side is organized with the cell identifier, the data acquisition timestamp, and the feature values ​​of each dimension as fields.

[0026] In one exemplary implementation, the capacity decay feature can be calculated based on the capacity decay trend of the unit's historical capacity record with the number of cycles or operating time. For example, the capacity retention rate can be obtained by taking the initial rated capacity as a benchmark and the ratio of the actual available capacity to the rated capacity, or the decay rate can be characterized by the amount of capacity decay per unit time. The multidimensional feature vector on the equipment side can include the current unit voltage, the current unit temperature, the current AC internal resistance, the rate of change of the internal resistance relative to the initial value, the deviation of the temperature relative to the group mean, the capacity retention rate, and the decay rate. Each feature is normalized or standardized before being input into the model (e.g., z-score standardization based on the training set mean and standard deviation). Missing features can be filled with the mean of the corresponding features in the training set.

[0027] Step S5: Input the device-side multi-dimensional feature vector into the health assessment model obtained through two-stage training, and output the estimated health status of the single battery cell and the prediction range of the estimated value. The two-stage trained health assessment model is obtained as follows: In the first stage, a general global model is trained using local datasets from multiple distributed battery nodes as training samples and a collaborative training method with shared model parameters. During the collaborative training process, each node only exchanges model parameters and not the original collected data, thereby protecting data privacy. In the second stage, the general global model is fine-tuned using a small amount of locally labeled data from the target battery node as fine-tuning samples to obtain a personalized assessment model for the target battery node. The health assessment model uses a lightweight convolutional neural network as the base network and introduces a channel attention mechanism into the network. The amount of locally labeled data is determined based on the historical capacity records of the target battery node, and each target node contains at least one labeled sample corresponding to a complete charge-discharge cycle. The model parameters shared in the collaborative training stage include at least the network weight parameters, and the original voltage, temperature, and internal resistance time series of each node are not shared. During model inference, the output header based on the enhanced Transformer architecture outputs the health status point estimate and uncertainty representation, and calculates the upper and lower bounds of the prediction interval based on the uncertainty representation. When a single unit is in a stable aging stage, the width of the prediction interval is less than the set width threshold. When a single unit is in an accelerated decay stage, the width of the prediction interval is not less than the set width threshold to remind maintenance personnel to make careful decisions.

[0028] In one exemplary implementation, the lightweight convolutional neural network may contain 2 to 3 convolutional layers and 1 to 2 fully connected layers, with a kernel size of 3×1 or 5×1 and 16 to 64 channels. The channel attention mechanism may employ a Squeeze-and-Excitation (SE) module to weight each feature channel to highlight voltage and internal resistance channels sensitive to aging. The output header of the enhanced Transformer architecture may include a health state point estimation header and an uncertainty estimation header (such as uncertainty representation based on Monte Carlo Dropout or deep ensemble). The prediction interval may be calculated as point estimation ± k times the uncertainty representation, where k can be 1.96, corresponding to approximately 95% confidence level. The width threshold can be set according to the stable aging stage. The statistical mean of the predicted interval width is determined (e.g., the width threshold is 0.05); in the two-stage training, in the first stage, each node trains locally for 1 to 10 epochs and then uploads the model weight increment. The aggregation node aggregates the data using FedAvg or an equivalent aggregation strategy and then distributes the update, iterating until the global model converges; in the second stage, the last 1 to 2 fully connected layers and the output head are fine-tuned using a small amount of locally labeled data from the target node (e.g., 10 to 100 labeled samples corresponding to at least one complete charge-discharge cycle), and the fine-tuning epoch is 5 to 50; the above network layer number, channel number, k value, number of epochs and number of samples are all exemplary values, which can be adapted by those skilled in the art. The specific values ​​are alternative implementation methods and do not constitute a limitation on the scope of protection.

[0029] Step S6: Calculate the inter-unit variability based on the estimated health status of each individual unit and the group mean. When the inter-unit variability exceeds a preset threshold, output a key maintenance prompt for that unit. The inter-unit variability is calculated as follows: The mean of the estimated health status of all units within the group is used as the group health mean; the absolute value of the difference between the estimated health status of each unit and the group health mean is used as the inter-unit variability for that unit. Alternatively, the mean of the AC internal resistance of all units within the group is used as the group internal resistance mean; the ratio of the absolute value of the difference between the internal resistance of each unit and the group internal resistance mean to the group internal resistance standard deviation is used as the internal resistance variability for that unit. The health status variability and internal resistance variability are weighted and fused to obtain the comprehensive inter-unit variability. Units with a comprehensive inter-unit variability exceeding the threshold are identified as lagging units and a key maintenance prompt is output.

[0030] In one exemplary implementation, the preset difference threshold can be determined based on the statistical distribution of the estimated health status values ​​of individual units within the group. For example, the health status difference threshold can be 5% to 15% (e.g., 10%) of the group's health mean, and the internal resistance difference threshold can be 1.0 to 2.0 (e.g., 1.5). The overall difference between individual units can be a weighted fusion of the health status difference and the internal resistance difference (e.g., each weight is 0.5). The key maintenance prompt can include the individual unit identifier of the lagging unit, the difference value, the anomaly type (low health status or high internal resistance), and the suggested maintenance actions (e.g., equalization charging, internal resistance retesting, or replacement), and can be pushed to the operation and maintenance personnel through the monitoring center interface, SMS, or in-station message.

[0031] This embodiment also includes a real-time alarm and reporting step: based on the voltage, temperature, internal resistance, and health status estimates of each individual unit, real-time alarms are triggered for any of the following conditions: overvoltage, undervoltage, overtemperature, internal resistance exceeding limits, or health status below the health threshold. The parameters, health status estimates, prediction ranges, and key maintenance reminders for each individual unit are organized into an operational status record by unit identifier and reported to the monitoring center via an RS485 bus and, in the case of long-distance transmission, via a fiber optic link converted by an optical modem. The structure of the RS485 bus network, node address allocation, protocol frame format, terminating resistor, and isolation protection are configured during system deployment.

[0032] This embodiment also includes an evaluation report generation step: statistically analyzing the voltage distribution, temperature distribution, internal resistance distribution, and health status distribution of each individual unit at a preset period, generating a comprehensive evaluation report that includes statistical tables, trend curves, and key maintenance tips. The data sources, statistical calibers, calculation rules, and output fields of the comprehensive evaluation report are defined in the report template, thereby ensuring the reproducibility of the evaluation report.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery, characterized in that, Includes the following steps: Step S1: Configure an independent smart sensor node for each individual lead-acid battery in the battery pack. Each smart sensor node integrates a voltage detection unit, a temperature detection unit, and an AC injection internal resistance detection unit, and is connected in series via an RS485 bus. Step S2: Each smart sensor node independently and synchronously acquires the voltage, temperature and AC internal resistance of its corresponding unit under electrical isolation to obtain the single-unit parameter sequence; Step S3: The edge aggregation node performs timestamp alignment and preprocessing on the parameter sequences of each segment to obtain the time-aligned multidimensional parameters of each segment; Step S4: Extract the voltage, temperature, internal resistance, and capacity decay characteristics of each cell based on historical data to construct a multi-dimensional feature vector on the device side. Step S5: Input the multi-dimensional feature vector of the device side into the health assessment model obtained through two-stage training, and output the estimated health status of the single battery cell and the prediction range of the estimated value. Step S6: Calculate the inter-individual difference based on the estimated health status of each individual and the group mean. When the inter-individual difference exceeds the preset difference threshold, output a key maintenance prompt for that individual.

2. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The independent synchronous acquisition includes: each intelligent sensor node uses a high-precision A / D converter to sample the voltage of a single cell; the internal resistance detection unit injects a sinusoidal AC signal of a set amplitude into the corresponding single cell and collects the response voltage to calculate the AC internal resistance; each node starts acquisition according to the unified sampling trigger time issued by the edge aggregation node, so that the voltage, temperature and internal resistance sampling times of each single cell in the same battery pack are aligned within the time window.

3. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The data acquisition under the electrically isolated state includes: the signal acquisition side and the bus communication side of each smart sensor node are isolated by an isolation power supply and a signal isolation device, so that the high voltage difference between the series-connected units does not form a common-mode path through the acquisition loop; the address of each node is allocated by the edge aggregation node through an automatic encoding process, which includes: after power-on, the node sends an address request frame carrying a physical identifier, the aggregation node responds that the address is not occupied, the node confirms and locks the address, and the aggregation node detects address conflicts and reallocates the address when there is a conflict.

4. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The two-stage training health assessment model is obtained as follows: In the first stage, a general global model is trained using the local datasets of multiple distributed battery nodes as training samples and a collaborative training method with shared model parameters is adopted. During the collaborative training process, each node only exchanges model parameters and does not exchange the original collected data. In the second stage, a small amount of locally labeled data of the target battery node is used as a fine-tuning sample to fine-tune the general global model, thereby obtaining a personalized evaluation model for the target battery node.

5. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 4, characterized in that, The health assessment model is based on a lightweight convolutional neural network and introduces a channel attention mechanism into the network. The amount of locally labeled data is determined based on the historical capacity records of the target battery nodes, and each target node contains at least one labeled sample corresponding to a complete charge-discharge cycle. The model parameters shared during the collaborative training phase include at least the network weight parameters, but the original voltage, temperature, and internal resistance time series of each node are not shared.

6. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The prediction interval of the health status estimate is obtained as follows: the health status point estimate and uncertainty characterization are output based on the output head of the enhanced Transformer architecture, and the upper and lower bounds of the prediction interval are calculated based on the uncertainty characterization; when the single unit is in the stable aging stage, the width of the prediction interval is less than the set width threshold; when the single unit is in the accelerated decay stage, the width of the prediction interval is not less than the set width threshold.

7. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The inter-unit variability is calculated as follows: the mean of the estimated health status of all units within the group is taken as the group health mean, and the absolute value of the difference between the estimated health status of each unit and the group health mean is taken as the inter-unit variability of that unit; or, the mean of the AC internal resistance of all units within the group is taken as the group internal resistance mean, and the ratio of the absolute value of the difference between the internal resistance of each unit and the group internal resistance mean to the group internal resistance standard deviation is taken as the internal resistance variability of that unit, and the health status variability and internal resistance variability are weighted and fused to obtain the comprehensive inter-unit variability.

8. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The method further includes: based on the voltage, temperature, internal resistance, and health status estimates of each individual unit, real-time alarms are triggered for any of the following conditions: overvoltage, undervoltage, overtemperature, internal resistance exceeding limits, and health status below the health threshold of each individual unit; and the parameters, health status estimates, prediction ranges, and key maintenance prompts of each individual unit are organized into an operating status record according to the individual unit identifier and reported to the monitoring center via an RS485 bus and an optical fiber link converted by the optical modem.

9. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The measurement accuracy of the AC internal resistance reaches ±1% under the set test conditions, and the voltage sampling accuracy reaches 0.2% under the set test conditions; the set test conditions include at least: ambient temperature range, single-cell state of charge range, float charge condition, and the amplitude and frequency of the injected sinusoidal AC signal.

10. The method for comprehensive monitoring and evaluation of a single-cell distributed sensing lead-acid battery according to claim 1, characterized in that, The method also includes an evaluation report generation step: statistically analyzing the voltage distribution, temperature distribution, internal resistance distribution, and health status distribution of each individual unit at a preset period, and generating a comprehensive evaluation report containing statistical tables, trend curves, and key maintenance tips. The data sources, statistical calibers, calculation rules, and output fields of the comprehensive evaluation report are defined in the report template.

Images