Battery health state evaluation method and system for power supply device of electric tool

Abstract

The invention relates to the technical field of battery management, and discloses a battery health state evaluation method and system for an electric tool energy supply device, and the method comprises the steps: obtaining multi-source sensor data in the operation process of an electric tool, and carrying out the preprocessing of the data; identifying and classifying multi-stage pulse discharge events, extracting multi-scale voltage response features and performing temperature compensation; equivalent circuit model parameter identification is carried out, abnormal identification results are checked and removed, and parameter aging characteristics are extracted; constructing a time sequence prediction model for prediction; evaluating the overall health state and the monomer consistency of the battery pack; defining a judgment standard of effective cycles, predicting the number of remaining effective cycles, calculating a confidence interval of a prediction result, and generating a battery management suggestion; formatting into a standardized evaluation report, carrying out data synchronization and model optimization through edge cloud collaboration, and recording a system operation log; according to the invention, accurate health state evaluation and residual life prediction are realized.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and more specifically, to a method and system for assessing the battery health status of an electric tool power supply device. Background Technology

[0002] With the widespread application of power tools in industrial manufacturing, construction, and equipment maintenance, the importance of lithium-ion batteries as their core power supply devices is becoming increasingly prominent. The operating characteristics of power tools mean that their battery packs face multiple challenges, including high-frequency pulse discharge, high-current instantaneous loads, intermittent operating modes, and complex and variable environmental temperature conditions. These factors combined result in the aging mechanism of power tool batteries exhibiting characteristics different from those in other application scenarios such as electric vehicles and energy storage systems. Accurately assessing the battery health status is of great significance for ensuring the safe operation of equipment and optimizing battery asset management.

[0003] Most existing battery health assessment methods are developed based on standardized charge-discharge test conditions, assuming the battery operates in a constant current and constant voltage mode, which makes it difficult to adapt to the non-standard pulse operating conditions of power tools in actual use. Furthermore, existing methods typically use general health indicators, failing to fully explore the aging characteristics under the specific operating conditions of power tools, thus limiting the accuracy of the assessment. In addition, power tool battery packs consist of multiple individual cells connected in series and parallel; the inconsistency between individual cells is a key factor affecting pack-level performance and lifespan, and existing methods lack an effective mapping mechanism from individual cell health status to pack-level health status.

[0004] Therefore, there is a need for a battery health status assessment method that can adapt to the special working conditions of power tools, make full use of pulse discharge characteristic information, and achieve collaborative assessment from individual cells to the group level, so as to solve the shortcomings of existing technologies in terms of working condition adaptability, feature extraction targeting, and individual cell-group level correlation. Summary of the Invention

[0005] This invention provides a method and system for assessing the battery health status of power tools, addressing the technical problems of insufficient adaptability to operating conditions, targeted feature extraction, and individual cell-level correlation in related technologies.

[0006] This invention provides a method for assessing the battery health status of a power tool power supply device, comprising: Acquire multi-source sensor data during the operation of power tools and preprocess it to obtain a normalized time-series dataset; Based on a normalized time-series dataset, multi-level pulse discharge events are identified and classified, multi-scale voltage response features are extracted and temperature compensation is performed to obtain a temperature-compensated health feature vector. Based on the normalized time-series dataset and the temperature-compensated health feature vector, the equivalent circuit model parameters are identified, the abnormal identification results are verified and eliminated, and the parameter aging features are extracted. The health feature vector after temperature compensation and the parameter aging feature are fused to construct a time-series prediction model for prediction, and the health status estimation results of each individual cell are obtained. Based on the health status estimation results of each individual cell, the overall health status and individual cell consistency of the battery pack are evaluated, and the overall health status evaluation result of the battery pack is output. Based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, the criteria for determining effective cycles are defined, the remaining number of effective cycles is predicted, the confidence interval of the prediction results is calculated, and battery management recommendations are generated. The remaining effective cycle count prediction results and battery management recommendations are formatted into a standardized evaluation report. Data synchronization and time series prediction model optimization are performed through edge-cloud collaboration, and system operation logs are recorded.

[0007] In a preferred embodiment, acquiring and preprocessing multi-source sensor data during the operation of the power tool includes: The terminal voltage, charging and discharging current, and surface temperature of the power tool battery pack are collected in real time, and a unified clock source is used to drive the analog-to-digital converters of each channel. The median filtering algorithm is used to denoise the original sensor signal; Missing value detection and imputation are performed. When a data point is null, exceeds the physical reasonable range, or the rate of change between adjacent data points exceeds the preset maximum rate of change threshold, it is determined to be a missing value. Based on the load status identification signal, continuous operating data is divided into independent work cycle segments.

[0008] In a preferred embodiment, obtaining the temperature-compensated health feature vector specifically includes: Based on the current time series data within the working cycle segment, multi-level pulse discharge events are identified and classified. A multi-scale and multi-condition joint judgment method is adopted to calculate the first-order and second-order difference sequences of the current signal and search for positive and negative jump points in the difference sequences. Based on the pulse current amplitude, the pulses are divided into three levels: light load pulse, medium load pulse, and heavy load pulse. Multi-scale voltage response features were extracted for pulses of different levels, including transient voltage drop amplitude, steady-state voltage drop amplitude, voltage drop rate, multi-stage voltage recovery time constant, graded dynamic internal resistance, and pulse energy efficiency. Based on battery temperature data and a temperature compensation model, the equivalent values ​​of the eigenvalues ​​at the reference temperature are calculated.

[0009] In a preferred embodiment, the method for extracting the multi-stage voltage recovery time constant is as follows: During the voltage recovery phase after the pulse ends, a double exponential function model is used to fit the recovery curve, and nonlinear least squares method is used for parameter identification. The iterative process adopts the Levenberg-Marquardt algorithm, and the convergence criterion is that the parameter change between two adjacent iterations is less than the preset convergence threshold or the number of iterations exceeds the preset maximum number of iterations.

[0010] In a preferred embodiment, the extracted parameter aging features specifically include: A second-order RC equivalent circuit model adaptive to pulse operating conditions is established, and current dependence and frequency dependence corrections are introduced. The model structure includes an ideal voltage source, a current-dependent ohmic internal resistance, a first RC parallel link, and a second RC parallel link. Based on the discretized state-space model and measured voltage and current data, the recursive least squares method is used to identify the model parameters online, and a forgetting factor mechanism is introduced. The parameter estimates are validated to eliminate outlier results. The validation uses multiple criteria, including physical range validation, residual validation and rate of change validation. Based on the validated model parameter sequence, parameter aging features are extracted.

[0011] In a preferred embodiment, obtaining the health status estimation results of each individual battery cell specifically includes: Feature fusion is performed to construct the model input. Feature fusion adopts a vector concatenation method, and each feature is standardized before concatenation. A sliding window method is used to extract a fixed-length subsequence from the feature vector sequence as the model input. The structure of the Long Short-Term Memory (LSTM) network is designed. The hidden layer of the network is composed of stacked LSM units. Each LSM unit contains four core components: a forget gate, an input gate, an output gate, and a cell state. An ensemble prediction method is used to quantify uncertainty by training multiple long short-term memory network models with different initialization parameters and using the distribution of prediction results from multiple models to estimate prediction uncertainty.

[0012] In a preferred embodiment, the overall health status assessment result of the output battery pack specifically includes: Collect voltage and temperature data of each individual cell in the battery pack, and calculate the dispersion index between individual cells. The dispersion of the health status of individual cells is analyzed to identify abnormal cells. The abnormality determination adopts a statistical threshold method. When the health status of a cell deviates from the mean by more than a preset multiple threshold, it is determined to be an abnormal cell. A graph neural network model is constructed to capture the mutual influence relationships between individual cells. The battery pack is modeled as a graph, with each individual battery cell as a node in the graph. Edges connect the cells to represent the relationships between them. The forward propagation process of the graph neural network adopts a message passing mechanism. Based on the updated feature vectors of each individual entity output by the graph neural network, a group-level health status aggregation function is designed to aggregate the health status of each individual entity.

[0013] In a preferred embodiment, the generated battery management recommendations specifically include: Define the criteria for effective cycle. The core indicator of the criteria is the battery's available energy. When the available energy is not less than the product of the rated energy and the effective energy threshold coefficient, it is judged as an effective cycle. Extrapolation is used to predict the time point when the health status drops to the failure threshold. The latter part of the health status prediction curve is selected as the fitting sample, and the extrapolation model is selected to fit the fitting sample. Convert the failure time point into the remaining effective cycle count, count the historical usage frequency of the battery, and convert the failure time point into the remaining cycle count based on the usage frequency; The confidence intervals for the predicted remaining effective number of cycles are calculated comprehensively, and the uncertainty is comprehensively quantified using the Monte Carlo simulation method; Based on the predicted remaining effective cycle count and confidence interval, battery management recommendations are generated.

[0014] In a preferred embodiment, the step of formatting the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, performing data synchronization and time-series prediction model optimization through edge-cloud collaboration, and recording system operation logs specifically includes: The data is formatted to generate a standardized evaluation report, which uses a structured data format. Data is uploaded to the cloud platform via a wireless communication module. After receiving the uploaded data, the cloud platform aggregates, stores, and analyzes the data. The model update is sent to the local device through the downlink channel. The model update includes the update of the weight parameters of the deep learning model, the adjustment of the parameters of the feature extraction algorithm, and the optimization of the decision threshold. Based on the status information of each stage of the evaluation process, a system operation log is recorded. The log records include data acquisition status, feature extraction results, model inference status, communication status, and system resource status.

[0015] This invention provides a battery health status assessment system for power tool power supply devices, used to perform the aforementioned battery health status assessment method for power tool power supply devices, comprising: The data acquisition module is used to acquire multi-source sensor data during the operation of power tools and preprocess it to obtain a normalized time-series dataset; The feature extraction module, based on a normalized time-series dataset, identifies and classifies multi-level pulse discharge events, extracts multi-scale voltage response features and performs temperature compensation to obtain a temperature-compensated health feature vector. The parameter identification module identifies the parameters of the equivalent circuit model based on the normalized time-series dataset and the temperature-compensated health feature vector, verifies and removes the abnormal identification results, and obtains the parameter aging characteristics. The health prediction module is used to fuse the temperature-compensated health feature vector and parameter aging features to construct a time-series prediction model for prediction, and obtain the health status estimation results of each individual cell. The consistency analysis module evaluates the overall health status and individual cell consistency of the battery pack based on the health status estimation results of each individual cell, and outputs the overall health status assessment results of the battery pack. The lifespan prediction module defines the criteria for determining effective cycles based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, predicts the remaining effective cycle count, calculates the confidence interval of the prediction results, and generates battery management recommendations. The results output module is used to format the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, perform data synchronization and time series prediction model optimization through edge-cloud collaboration, and record system operation logs.

[0016] The beneficial effects of this invention are as follows: A dedicated health feature extraction method was designed for the pulse discharge characteristics of power tools. By analyzing the voltage response curve during the pulse and the rest relaxation characteristics during the working interval, characteristic parameters that can reflect the aging state of the battery, such as voltage drop amplitude, recovery time constant, and dynamic internal resistance, are extracted. Furthermore, the interference of environmental factors is eliminated through a temperature compensation model. Compared with traditional methods that use general health indicators, this method can more accurately characterize the health status of power tool batteries under actual working conditions, thereby improving the relevance and reliability of the evaluation results.

[0017] A single-cell correlation model based on graph neural network and a group-level health status aggregation function considering the barrel effect were established. This model can effectively capture the mutual influence between single cells in the battery pack, accurately quantify the impact of single-cell inconsistency on group-level performance, and achieve a precise mapping from single-cell health status to group-level health status. This solves the problem of the disconnect between single-cell and group-level assessment in the health assessment of multi-cell battery packs in existing methods, and provides a reliable basis for the overall health management and balanced maintenance decision of the battery pack. Attached Figure Description

[0018] Figure 1This is a main flowchart of a battery health status assessment method for an electric tool power supply device according to the present invention; Figure 2 This is a detailed flowchart of a battery health status assessment method for an electric tool power supply device according to the present invention; Figure 3 This is a block diagram of a battery health status assessment system for an electric tool power supply device according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for assessing the battery health status of a power tool power supply device, such as... Figures 1 to 2 As shown, it includes: Step 1: Acquire multi-source sensor data during the operation of the power tool and preprocess it to obtain a normalized time-series dataset; Based on the hardware acquisition capabilities of the power tool battery management system, multi-channel signals such as voltage, current, and temperature are synchronously acquired during battery operation. A median filtering algorithm with adaptive window width is used to eliminate sensor noise and electromagnetic interference. Missing value interpolation is used to ensure data integrity. Finally, the continuous data is segmented into independent work cycle segments according to the working status of the power tool, providing a high-quality data foundation for subsequent feature extraction.

[0021] Step 1.1: Synchronous acquisition of multi-channel sensor signals; The terminal voltage, charging / discharging current, and surface temperature of the power tool battery pack are acquired in real time. A unified clock source drives the analog-to-digital converters of each channel to ensure time synchronization of data from multiple channels.

[0022] The sampling frequency of the voltage acquisition channel is determined based on the time constant of the power tool load response, and needs to be able to capture rapid voltage changes during pulse discharge. The sampling frequency is determined as follows: perform a load step response test on the target power tool, record the time required for the voltage to drop from steady state to its lowest point, and multiply the reciprocal of this time by an oversampling factor as the preferred value for the sampling frequency. The typical value range for the oversampling factor is 10 to 20 times. The current acquisition channel uses the same sampling frequency as the voltage channel to maintain data alignment. Since temperature changes relatively slowly, the temperature acquisition channel uses a lower sampling frequency to reduce data storage and processing burden; the preferred value for the temperature sampling frequency is set to one-hundredth of the voltage sampling frequency. Simultaneously, a load status indicator signal is obtained from the power tool's main controller. This signal indicates whether the power tool is currently in working, standby, or off state, and is used for subsequent work cycle segmentation.

[0023] Step 1.2, median filtering for noise reduction; Based on the acquired raw sensor signals, a median filtering algorithm is used for denoising. Median filtering effectively suppresses impulse noise and electromagnetic interference while preserving signal edge features relatively well. The filtering window width is adaptively determined according to the sampling frequency and noise characteristics, with an optimal window width of 5 to 11 sampling points. For voltage and current signals, the same window width is used for filtering; for temperature signals, due to their slower change, a larger window width can be used to enhance the filtering effect. The filtering process is performed point-by-point in chronological order. For the current sampling point, the median of all sampled values ​​within the window range before and after it is taken as the filtered output.

[0024] Step 1.3, Missing value detection and imputation; Based on the filtered time-series data, missing value detection and imputation are performed. The causes of missing values ​​include sensor failure, communication interruption, data storage overflow, etc., which manifest as null values ​​or abnormal jumps in the data sequence.

[0025] Missing value detection employs a dual determination mechanism. A data point is considered missing if any of the following conditions are met: the data point is null or exceeds the physically reasonable range; the reasonable range for voltage is determined based on the nominal voltage of the battery pack and the allowable overcharge and over-discharge margin; the reasonable range for current is determined based on the maximum load capacity of the power tool; and the reasonable range for temperature is determined based on the battery's safe operating temperature range. The rate of change between adjacent data points exceeds the physically possible maximum rate of change, which is determined based on the dynamic response characteristics of the battery and the load.

[0026] For detected missing values, an imputation method based on local weighted regression is used. The specific implementation process is as follows: Centering on the missing point, several valid data points before and after it are selected as reference points, with an optimal number of reference points being five before and after. The temporal distance between each reference point and the missing point is calculated, and a weight is calculated based on this temporal distance. The weighting function uses a Gaussian kernel function, with reference points having greater weights the closer they are in time. A weighted least squares method is used to fit a polynomial to the reference points, with the optimal polynomial order being second. The value of the fitted polynomial at the time of the missing point is used as the imputation value. When the number of consecutive missing points exceeds a preset maximum consecutive missing threshold, this data segment is marked as unusable and not included in subsequent feature extraction.

[0027] Step 1.4, Work cycle segmentation; Based on the preprocessed time-series data and load status identification signals, continuous operating data is segmented into independent work cycle segments. A work cycle is defined as a complete usage process of a power tool from start to stop, including the start-up phase, the working phase, and the stop phase.

[0028] The segmentation algorithm is based on the state transition of the load state identification signal: when the load state changes from standby or off to working, it is marked as the start point of the working cycle; when the load state changes from working to standby or off, it is marked as the end point of the working cycle; the data between the start point and the end point constitutes a complete working cycle segment.

[0029] For situations where the load status indicator signal is unavailable or unreliable, an alternative segmentation method based on the current signal is adopted: when the current rises from below the no-load current threshold to above the starting current threshold and continues to exceed the minimum working time threshold, the working cycle is determined to begin; when the current falls below the no-load current threshold and continues to exceed the minimum stop time threshold, the working cycle is determined to end. The no-load current threshold, starting current threshold, minimum working time threshold, and minimum stop time threshold are set according to the specific model and working characteristics of the power tool.

[0030] After the segmentation is completed, a unique cycle identifier is assigned to each work cycle segment, and its start and end timestamps, duration, number of data points, and other metadata are recorded to form a work cycle index table.

[0031] Step 1 outputs a normalized time-series dataset, containing: voltage, current, and temperature time-series data after denoising and missing value imputation; a duty cycle index table recording the boundaries and metadata of each duty cycle; and a load state identifier signal sequence. This dataset possesses time synchronization, data integrity, and structural normalization, providing a reliable data foundation for subsequent feature extraction and health status assessment.

[0032] Step 2: Based on the normalized time series dataset, identify and classify multi-level pulse discharge events, extract multi-scale voltage response features and perform temperature compensation to obtain the temperature-compensated health feature vector. Specifically, the following steps are included: Step 2.1, Identification and classification of multi-level pulse discharge events; Based on current timing data within a work cycle segment, multi-stage pulse discharge events are identified and classified. Power tools exhibit typical multi-stage pulse load characteristics during operation; for example, an electric wrench generates start-up pulses, loading pulses, and surge pulses with noticeable differences when tightening bolts.

[0033] The identification of pulse discharge events employs a multi-scale, multi-condition joint judgment method: First-order and second-order difference sequences of the current signal are calculated. The first-order difference captures pulse edge features, while the second-order difference identifies the rate of change features within the pulse. Positive and negative transition points are searched in the difference sequences. A positive transition point is defined as a sampling point where the first-order difference value exceeds a positive transition threshold and the absolute value of the second-order difference is lower than a stationarity threshold. A negative transition point is defined as a sampling point where the first-order difference value is lower than a negative transition threshold. The time interval between positive and negative transition points is the pulse duration. When the pulse duration and average current meet preset conditions, it is determined to be a valid pulse discharge event.

[0034] The method for determining the positive jump threshold is as follows: Perform a typical load switching test on the power tool, record the change in current from no-load to working load, and take 80% of the minimum change as the preferred value of the positive jump threshold. For an 18-volt power wrench, the typical positive jump threshold is 3 to 5 amps, and the stability threshold is set to 20% to 30% of the positive jump threshold.

[0035] Identified pulse events are classified into three levels based on their pulse current amplitude: light-load pulses, medium-load pulses, and heavy-load pulses. Light-load pulses have a current amplitude less than 30% of the rated current, corresponding to no-load operation or light load conditions. Medium-load pulses have a current amplitude between 30% and 70% of the rated current, corresponding to normal operating loads. Heavy-load pulses have a current amplitude exceeding 70% of the rated current, corresponding to peak loads or stall conditions. For an 18-volt, 20-amp electric wrench, the light-load pulse threshold is 6 amps, and the heavy-load pulse threshold is 14 amps. Different pulse levels cause different stress modes in the battery. Light-load pulses mainly cause a voltage drop across the ohmic internal resistance, medium-load pulses activate both ohmic and electrochemical polarization processes, and heavy-load pulses induce concentration polarization and temperature rise effects.

[0036] Simultaneously, the temporal pattern characteristics of the pulse sequence are analyzed, including the pulse interval time distribution, pulse repetition frequency, and pulse amplitude variation trend. These sequence characteristics reflect the dynamic response and recovery capabilities of the battery under continuous pulse load.

[0037] Step 2.2, Multi-scale pulse voltage response feature extraction; Based on the identified graded pulse discharge events, multi-scale voltage response features are extracted for pulses of different grades. For each pulse event, a voltage segment is extracted from the corresponding voltage time series data, from several sampling points before the pulse starts to several sampling points after the pulse ends. The extent of the segmentation range is adaptively determined according to the pulse grade and the time constant of the voltage response. The extent of the segmentation is smaller for light-load pulses and larger for heavy-load pulses, ensuring that the voltage drop and recovery process under pulses of different intensities can be completely captured.

[0038] Based on the extracted voltage segments, the following multi-scale feature parameters are extracted: Transient voltage drop amplitude: The average voltage of several sampling points before the pulse starts is taken as the initial voltage, and the minimum voltage in the first sampling period after the pulse starts is taken as the transient voltage drop. The difference between the initial voltage and the transient voltage drop is the transient voltage drop amplitude. This feature mainly reflects the ohmic internal resistance of the battery and has low temperature sensitivity.

[0039] Steady-state voltage drop amplitude: The average value of the voltage after stabilization during the pulse period is taken as the steady-state voltage drop. The difference between the initial voltage and the steady-state voltage drop is the steady-state voltage drop amplitude. This characteristic comprehensively reflects the combined effect of ohmic internal resistance and polarization internal resistance. The difference between the transient voltage drop amplitude and the steady-state voltage drop amplitude is the polarization voltage drop, which is more sensitive to changes in battery aging.

[0040] Voltage sag rate: The transient sag rate is the time it takes for the voltage to drop from the initial voltage to the transient sag voltage. The voltage sag amplitude divided by this time is the transient sag rate. The polarization voltage drop rate is the time it takes for the voltage to drop from the transient sag voltage to the steady-state sag voltage. The polarization establishment rate is the polarization settling rate, which reflects the high-frequency response characteristics of the battery's internal resistance.

[0041] Multi-stage voltage recovery time constant: In the voltage recovery stage after the pulse ends, a double exponential function model is used to fit the recovery curve to obtain the fast time constant and the slow time constant. The form of the fitting function is that the voltage equals the steady-state voltage minus the exponential decay term of the fast recovery amplitude multiplied by the fast time constant, and then minus the exponential decay term of the slow recovery amplitude multiplied by the slow time constant. The steady-state voltage, fast recovery amplitude, fast time constant, slow recovery amplitude, and slow time constant are the parameters to be fitted. Nonlinear least squares method is used for parameter identification. The initial parameter values ​​are set as follows: steady-state voltage is the endpoint voltage value of the recovery curve; fast recovery amplitude is 70% of the total change in the recovery curve; the initial value of the fast time constant is set to 10 milliseconds; the slow recovery amplitude is 30% of the total change in the recovery curve; and the initial value of the slow time constant is set to 100 milliseconds. The iterative process uses the Levenberg-Marquardt algorithm, with the convergence criterion being that the parameter change between two adjacent iterations is less than one-thousandth or the number of iterations exceeds 50. The fast time constant obtained from the fitting reflects the relaxation characteristics of electrochemical polarization, while the slow time constant reflects the relaxation characteristics of concentration polarization. The two time constants exhibit different evolution patterns as the battery ages.

[0042] Graded Dynamic Internal Resistance: Dynamic internal resistance is calculated separately for different pulse levels. The dynamic internal resistance of a light-load pulse mainly reflects the ohmic internal resistance; the dynamic internal resistance of a medium-load pulse reflects the superposition of the ohmic internal resistance and the electrochemical polarization internal resistance; and the dynamic internal resistance of a heavy-load pulse reflects the superposition of all internal resistance components. By comparing the differences in dynamic internal resistance among different pulse levels, the contribution of each internal resistance component is separated, enabling differentiated assessment of different aging mechanisms within the battery. Specifically, the dynamic internal resistance of a light-load pulse is approximately equal to the ohmic internal resistance; the estimated value of the electrochemical polarization internal resistance is obtained by subtracting the dynamic internal resistance of a light-load pulse from the dynamic internal resistance of a medium-load pulse; and the estimated value of the concentration polarization internal resistance is obtained by subtracting the dynamic internal resistance of a medium-load pulse from the dynamic internal resistance of a heavy-load pulse.

[0043] Pulse energy efficiency: The ratio of the electrical energy released by the battery during a pulse to the electrical energy that should be released under ideal conditions with no internal resistance. Ideally, the released electrical energy equals the initial voltage multiplied by the integral of the pulse current; the actual released electrical energy equals the actual voltage multiplied by the integral of the pulse current. The ratio of these two is the pulse energy efficiency. Pulse energy efficiency comprehensively reflects the degree of energy loss of the battery under pulsed operating conditions.

[0044] Step 2.3, analysis of static relaxation characteristics; Voltage relaxation characteristics were analyzed based on rest period data between work cycles. When a power tool stops working and enters a resting state, the battery's terminal voltage undergoes a relaxation process, gradually recovering from the transient voltage at the end of the load to the open-circuit voltage. This relaxation process reflects the dissipation of the concentration gradient and the reconstruction of electrochemical equilibrium within the battery, and its characteristic parameters are related to the battery's health status. Extraction of resting relaxation characteristics requires meeting a minimum resting time requirement. The minimum resting time is determined based on the battery's relaxation time constant and needs to be long enough to observe a significant relaxation process; its preferred value can be set as a multiple of the voltage recovery time constant. For a resting period that meets the minimum resting time requirement, the following characteristic parameters were extracted: The method for extracting the relaxation voltage change is as follows: take the voltage at the beginning of the resting period as the initial relaxation voltage, and take the voltage at the end of the resting period or after relaxation has stabilized as the termination relaxation voltage. The difference between the termination relaxation voltage and the initial relaxation voltage is the relaxation voltage change. The relaxation voltage change reflects the battery's voltage recovery capability after the load is removed.

[0045] The relaxation time constant is extracted by fitting an exponential function to the voltage relaxation curve during the resting period. The fitting method is the same as that used to extract the voltage recovery time constant, and the resulting time constant is the relaxation time constant. The relaxation time constant reflects the rate of diffusion processes inside the battery and is related to the degree of aging of the electrode material.

[0046] Step 2.4, Periodic-level feature statistics; Based on the features of multiple pulse events extracted within a single work cycle, a periodic feature vector is formed through statistical summarization. Since a single work cycle may contain multiple pulse discharge events, and the load intensity and duration of each pulse event may differ, directly using the feature value of a single pulse would introduce significant random fluctuations. Periodic feature statistics extract more stable and representative health indicators by statistically processing multiple pulse features.

[0047] Specific statistics include: the arithmetic mean of each pulse feature, reflecting the average level of battery response characteristics within that period; the standard deviation of each pulse feature, reflecting the stability and consistency of battery response characteristics; the skewness of each pulse feature, reflecting the degree of asymmetry in the feature distribution; and the kurtosis of each pulse feature, reflecting the sharpness of the feature distribution. These statistics are concatenated with the resting relaxation features to form the original feature vector at the periodic level.

[0048] Step 2.5, temperature compensation processing; Temperature compensation is performed based on the original eigenvectors at the periodic level and the corresponding battery temperature data. The electrochemical characteristics of the battery are sensitive to temperature, and the purpose of temperature compensation is to eliminate the influence of temperature and normalize the eigenvalues ​​collected at different temperatures to the equivalent values ​​at the reference temperature.

[0049] The method for establishing the temperature compensation model is as follows: collect characteristic data of the same battery under different temperature conditions, and for each characteristic parameter, perform regression analysis with temperature as the independent variable and the characteristic value as the dependent variable. The regression model can be a linear model or a multinomial model; the model parameters obtained from the regression are the temperature compensation coefficients. The implementation method of temperature compensation is as follows: based on the battery temperature at the time of data collection and the temperature compensation model, calculate the equivalent value of the characteristic value at a reference temperature. The preferred reference temperature can be set to room temperature.

[0050] Step 2 outputs a temperature-compensated health feature vector, which includes: temperature-compensated pulse voltage response characteristics; temperature-compensated rest relaxation characteristics; and periodic statistical characteristics. This feature vector eliminates the influence of temperature and can accurately reflect the battery's health status under different operating conditions, providing reliable feature input for subsequent parameter identification and health status prediction.

[0051] Furthermore, since linear temperature compensation models may suffer from insufficient compensation accuracy over a wide temperature range, piecewise linear compensation models can be used as an alternative. Piecewise linear compensation models divide the temperature range into several sub-intervals, establishing a linear compensation model within each sub-interval. The division of sub-intervals is determined based on the degree of nonlinearity in the relationship between eigenvalues ​​and temperature. Specifically, linear fitting is performed on the data across the entire temperature range, and the fitting residuals are calculated. Segmentation points are set in temperature regions with large residuals, dividing the temperature range into multiple sub-intervals. Linear fitting is then performed again within each sub-interval to obtain the piecewise compensation coefficients. Piecewise linear compensation models can better adapt to the nonlinear relationship between eigenvalues ​​and temperature, improving the accuracy of temperature compensation.

[0052] Step 3: Based on the normalized time series dataset and the temperature-compensated health feature vector, perform equivalent circuit model parameter identification, verify and remove the anomaly identification results, and extract parameter aging features. Specifically, the following steps are included: Step 3.1, Construction of the adaptive equivalent circuit model for pulse operating conditions; Based on the electrochemical characteristics of lithium-ion batteries under pulsed discharge conditions, an adaptive second-order RC equivalent circuit model is established. This model is specially designed for the high-frequency pulsed load characteristics of power tools, and introduces current-dependent and frequency-dependent corrections on the basis of the traditional equivalent circuit model, which can more accurately describe the dynamic response behavior of the battery under pulses of different intensities and frequencies.

[0053] The model structure comprises the following components: an ideal voltage source representing the battery's open-circuit voltage, which is related to the battery's state of charge (SOC) and is determined using a method based on ampere-hour integration and an open-circuit voltage-SOC lookup table; and a current-dependent ohmic internal resistance representing the battery's purely resistive characteristics. This internal resistance is no longer a constant but a function of the current, and a piecewise linear model describes the relationship between the internal resistance and the current. In the low-current region, the resistance is small and changes slowly, while in the high-current region, the resistance increases and changes rapidly. This current dependence stems from the nonlinear conductivity of the electrode material and the electrolyte concentration polarization effect. The segment points and the slopes of each segment are determined by fitting pulse response data at different current levels. The first parallel resistor-capacitor link represents the electrochemical polarization process, where the resistance represents the charge transfer resistance and the capacitance represents the double-layer capacitance. The time constant of this link reflects the kinetic characteristics of the charge transfer process, and the typical value of the time constant is on the order of milliseconds to tens of milliseconds. The second parallel resistor-capacitor link represents the concentration polarization process, where the resistance represents the diffusion resistance and the capacitance represents the diffusion capacitance. The time constant of this link reflects the kinetic characteristics of ion diffusion in the electrode material, and the typical value of the time constant is on the order of tens of milliseconds to hundreds of milliseconds.

[0054] To adapt to the high-frequency pulse conditions of power tools, a frequency-dependent correction is introduced into the first RC circuit. The values ​​of charge transfer resistance and double-layer capacitance vary with the pulse frequency. Under high-frequency pulses, the double-layer capacitance does not have enough time to fully build up, exhibiting a smaller equivalent capacitance value. The charge transfer resistance is affected by the dynamic occupancy of active sites on the electrode surface, exhibiting a larger equivalent resistance value under high-frequency pulses. Frequency dependence is achieved by introducing a frequency correction factor, which is a function of the pulse frequency and described using a first-order inertial circuit. The correction factor is equal to 1 divided by the square of the ratio of the pulse frequency to the characteristic frequency. The characteristic frequency reflects the response speed of the electrochemical process and is determined by fitting response data from pulse sequences of different frequencies.

[0055] The output voltage of the model is calculated as follows: the output voltage equals the open-circuit voltage minus the voltage drop across the current-dependent ohmic internal resistance, then minus the voltage drop across the first RC link, and finally minus the voltage drop across the second RC link. The voltage drop of each RC link is obtained by solving differential equations. For the first RC link, the rate of change of the capacitor voltage equals the current minus the capacitor voltage divided by the resistance and then divided by the capacitance. A similar equation is used for the second RC link. Under pulsed operating conditions, due to the rapid change of current, the voltage response of the RC link exhibits obvious dynamic characteristics. Accurately describing these dynamic characteristics is crucial to the model's accuracy.

[0056] Step 3.2, Model Discretization and State-Space Representation; Based on the equivalent circuit model in the continuous-time domain, discretization is performed to meet the requirements of digital signal processing. A zero-order hold method is used to discretize the model, with the discretization step size equal to the sampling period of the voltage and current data.

[0057] The discretized model can be represented in state-space form: the state vector contains the voltages of two RC circuits, the state transition equation describes the evolution of the state vector from the current time step to the next time step, and the elements of the state transition matrix are determined by the time constants of the RC circuits and the discretization step size; the output equation describes the relationship between the terminal voltage, the state vector, and the input current, and the elements of the output matrix are determined by the model parameters. This state-space representation provides a unified mathematical framework for subsequent parameter identification algorithms.

[0058] Step 3.3, recursive least squares parameter identification; Based on a discretized state-space model and measured voltage and current data, recursive least squares method is used for online identification of model parameters. Recursive least squares is a parameter estimation method suitable for online applications, capable of incrementally updating parameter estimates as new data arrives, without needing to store and process all historical data. The core idea of ​​the algorithm is to minimize the sum of squared errors between the model-predicted voltage and the measured voltage.

[0059] During the algorithm initialization phase, initial values ​​for the parameter estimates and the covariance matrix are set. The initial parameter values ​​can be set based on the battery's nominal parameters or previous identification results. The initial covariance matrix is ​​set to a large diagonal matrix to represent the higher uncertainty of the initial estimates.

[0060] During the algorithm iteration phase, for each newly arrived sampling point, the following calculations are performed: calculate the model-predicted voltage based on the current parameter estimates and input current; calculate the prediction error between the predicted voltage and the measured voltage; calculate the gain vector, which involves the covariance matrix and regression vector, and the regression vector is determined by the model structure and input / output data; update the parameter estimates using the gain vector and prediction error, with the update amount equal to the product of the gain vector and the prediction error; and update the covariance matrix, which reflects the uncertainty of the parameter estimates.

[0061] To enhance the algorithm's ability to track time-varying parameter characteristics, a forgetting factor mechanism is introduced. The forgetting factor is a positive number less than or equal to one, which applies exponential decay weighting to historical information during covariance matrix updates, ensuring that recent data contributes more to parameter estimation than older data. The value of the forgetting factor needs to be balanced between tracking speed and estimation stability. For cases where parameters change slowly during battery aging, a recommended forgetting factor value is 0.99 to 0.999, with a typical value of 0.995.

[0062] Step 3.4, parameter validity verification; The parameter estimates output by the recursive least squares method are validated to remove outliers. The parameter identification process may produce unreasonable parameter estimates due to data anomalies, numerical calculation problems, or model mismatch, which can interfere with subsequent health status assessments. Validity validation employs multiple criteria: Physical Range Verification: Check whether the estimated values ​​of each parameter are within the physically reasonable range. The ohmic internal resistance should be positive and not exceed the maximum internal resistance at the time of battery failure. The maximum internal resistance can be determined based on battery specifications or historical data. The polarization resistance and diffusion resistance should also be positive and within a reasonable range. The time constant should be positive and conform to the dynamic response characteristics of the battery. Parameter estimates that exceed the physical range are marked as invalid.

[0063] Residual verification: This checks whether the fitting residuals of the parameter identification are below the residual threshold. The fitting residual is defined as the root mean square value of the difference between the model-predicted voltage and the measured voltage during the identification period. The residual threshold is determined based on the sensor accuracy and model accuracy. Excessively large residuals indicate a mismatch between the model and the actual data, which may be due to data anomalies or an inapplicable model structure. In such cases, the parameter estimates are marked as invalid.

[0064] Rate of Change Verification: This checks whether the rate of change of the parameter estimate relative to the previous valid estimate exceeds the abnormal change threshold. Battery parameter aging changes are gradual and should not exhibit drastic changes in a short period; an excessively high rate of change usually indicates interference with the identification process. The abnormal change threshold is determined based on the battery's normal aging rate and identification cycle. Parameter estimates with a rate of change exceeding the threshold are marked as suspicious and require further confirmation with subsequent data.

[0065] Parameter estimates that pass the validity check are included in the parameter sequence, while estimates that fail the check are removed or marked and do not participate in subsequent aging characteristic analysis.

[0066] Step 3.5, Extraction of Parameter Aging Features; Based on the validated model parameter sequence, parameter aging features are extracted. The model parameters exhibit a regular trend of change with battery cycle aging, and these trends reflect the degree and rate of battery aging. The method for extracting parameter aging features is as follows: For the ohmic internal resistance parameter, calculate its growth rate relative to the initial value, which is the average parameter value at the beginning of battery use. The ohmic internal resistance growth rate is an important indicator of the battery's aging degree, and its growth is mainly related to the growth of the solid electrolyte interface film and the decrease in the conductivity of the electrode materials. Simultaneously calculate the ohmic internal resistance growth rate, i.e., the increase per unit number of cycles; the growth rate reflects the battery's aging speed.

[0067] The growth rate and growth rate of polarization resistance and diffusion resistance parameters were calculated using a similar method. The increase in polarization resistance is related to the reduction of active sites on the electrode surface, while the increase in diffusion resistance is related to the growth of ion diffusion paths caused by the degradation of the electrode material structure.

[0068] For the time constant parameter, analyze its changing trend. The change in the time constant reflects the evolution of the internal dynamic characteristics of the battery, and it usually shows an increasing trend with aging, indicating that the battery's response speed slows down.

[0069] Step 3 outputs the model parameter sequence characterizing the internal state of the battery and its aging characteristics, including: the equivalent circuit model parameter sequence after validity verification, including parameters such as ohmic internal resistance, polarization resistance, diffusion resistance and time constant; parameter aging characteristics, including the growth rate and growth rate of each parameter relative to the initial value.

[0070] Step 4: The temperature-compensated health feature vector and parameter aging features are fused to construct a time-series prediction model for prediction, and the health status estimation results of each individual cell are obtained. Specifically, the following steps are included: Step 4.1, Feature fusion and input construction; Based on the temperature-compensated health feature vector output from step 2 and the parameter aging features output from step 3, feature fusion is performed to construct the model input. Feature fusion employs a vector concatenation method, concatenating the impulse response features, rest relaxation features, and equivalent circuit model parameter features in a fixed order into a high-dimensional feature vector. Before concatenation, each feature is standardized by subtracting the feature mean and then dividing by the feature standard deviation. The mean and standard deviation are obtained statistically from historical data. Standardization eliminates differences in dimensions and numerical ranges between different features, which is beneficial for the subsequent training and convergence of the neural network.

[0071] When constructing time-series input samples, a sliding window method is used to extract a fixed-length subsequence from the feature vector sequence as model input. The sliding window length determines the range of historical information the model can utilize. For industrial power tool batteries, changes in health status typically require ten to hundreds of charge-discharge cycles. The recommended sliding window length is 20 to 50 work cycles, with 30 work cycles being the preferred value. The sliding step size is set to one work cycle, meaning one input sample is generated per work cycle.

[0072] Step 4.2, Long Short-Term Memory Network Structure Design; Based on the requirements of time series prediction tasks, a Long Short-Term Memory (LSTM) network structure was designed. LSTM is a special type of recurrent neural network that solves the gradient vanishing problem in standard recurrent neural networks when processing long sequences through a gating mechanism, effectively capturing long-term dependencies in time series data.

[0073] The network's input layer receives a standardized sequence of feature vectors. The input dimension is equal to the dimension of the fused feature vectors, and the sequence length is equal to the sliding window length.

[0074] The hidden layers of the network consist of multiple stacked Long Short-Term Memory (LSTM) units. Each LSM unit contains four core components: a forget gate, an input gate, an output gate, and a cell state. The forget gate determines which information to discard from the cell state; its output is calculated by concatenating the previous hidden state and the current input, followed by a linear transformation and activation function. The input gate determines which new information to add to the cell state, including the input gate activation value and candidate cell states, obtained through similar calculations. The cell state is updated by element-wise multiplying the forget gate output by the previous cell state, plus the element-wise multiplication of the input gate activation value and the candidate cell state. The output gate determines which information to output from the cell state as the current hidden state; its calculation process is similar to that of the forget gate.

[0075] The number of hidden layers and the number of hidden units per layer are key hyperparameters of a network. Too few layers may lead to insufficient model expressive power, while too many layers may result in overfitting and increased computational cost. The choice of the number of hidden units also requires a trade-off between model capacity and generalization ability. The optimal values ​​of these hyperparameters can be determined by performing a grid search on the validation set, with the search range set according to the task complexity and available computational resources.

[0076] The network's output layer is a fully connected layer that maps the hidden state of the last time step to a predicted health state. The number of neurons in the output layer is determined by the prediction task. For a single-step prediction task, the output layer contains one neuron that outputs the current health state estimate; for a multi-step prediction task, the output layer contains multiple neurons that output the predicted health state values ​​for each future time step.

[0077] Step 4.3, quantify the uncertainty of prediction; Based on a pre-trained Long Short-Term Memory (LSTM) network model, an ensemble prediction method is employed to quantify uncertainty. Deep learning models typically produce point estimates, lacking the quantification of prediction uncertainty. However, in battery health assessment applications, understanding the reliability of prediction results is crucial for decision-making.

[0078] The ensemble prediction method estimates prediction uncertainty by training multiple LSTM models with different initialization parameters and utilizing the distribution of prediction results from these models. The specific implementation process is as follows: train three to five LSTM models, each with different random initialization parameters and training data sampling; for a given input sample, use all models to make predictions, and collect the prediction results to form a sample set; calculate the mean of the sample set as a point estimate of the health state; calculate the standard deviation of the sample set as a measure of prediction uncertainty; and calculate confidence intervals based on the sample distribution, for example, using the 5th percentile and 95th percentile of the sample as the lower and upper bounds of the 90% confidence interval. This method is computationally efficient and suitable for implementation in embedded systems.

[0079] Step 4.4, Incremental learning mechanism; Based on newly acquired online data, an incremental learning mechanism is designed to support continuous model updates. Battery aging characteristics may change with usage conditions and environmental factors, and pre-trained models may experience performance degradation over long-term use. The incremental learning mechanism enables the model to be updated online using newly acquired data, adapting to changes in battery characteristics.

[0080] Incremental learning is implemented using a cloud-edge collaborative strategy. The local embedded system continuously collects runtime data and evaluation results, and uploads them to the cloud platform once a certain amount of new data has accumulated. The cloud platform uses data from a large number of devices to retrain or fine-tune the model, obtaining updated model parameters. The updated model parameters are then distributed to the local devices, which update the model parameters during idle periods. This cloud-edge collaborative approach ensures continuous model optimization while avoiding the computational burden on local devices.

[0081] Step 4 outputs the health status estimation results for each individual cell, including: the estimated health status of each individual cell at the current moment; the predicted health status curves for several future operating cycles; and the predicted mean, standard deviation, and confidence interval obtained based on the ensemble prediction method. These outputs provide crucial health status information for subsequent cell consistency analysis and remaining lifetime prediction.

[0082] Step 5: Based on the health status estimation results of each individual battery cell, evaluate the overall health status of the battery pack and the consistency of individual cells, and output the overall health status assessment results of the battery pack. Specifically, the following steps are included: Step 5.1, Individual data acquisition and dispersion calculation; The system collects voltage and temperature data from each individual cell within the battery pack. Power tool battery packs typically consist of multiple individual cells connected in series or parallel; the number of cells is determined by the pack's voltage and capacity specifications. Individual cell voltage data is acquired through the multi-channel voltage measurement function of the battery management chip, with the acquisition accuracy and sampling rate meeting the requirements for detecting individual cell differences. Individual cell temperature data can be acquired by placing temperature sensors at representative locations. For cells without directly placed sensors, their temperature can be estimated using a thermal model.

[0083] Based on the collected individual unit data, the dispersion index between units is calculated. The voltage dispersion is calculated as follows: the arithmetic mean of all unit voltages is calculated as the group average voltage; the difference between each unit voltage and the group average voltage is calculated; and the standard deviation of the difference is calculated as the voltage dispersion. Voltage dispersion reflects the voltage consistency between units; a larger dispersion indicates more significant differences between units. Temperature dispersion is calculated using a similar method, reflecting the uniformity of temperature distribution between units.

[0084] In addition to static dispersion, dynamic dispersion is also calculated. Dynamic voltage dispersion focuses on the consistency of individual cell voltage changes during charging and discharging. The calculation method involves calculating the voltage dispersion at multiple time points during the charging and discharging process and then calculating the magnitude of these dispersion values. Dynamic dispersion reflects the differences in load response characteristics among individual cells and is an important indicator of inconsistencies in cell aging.

[0085] Step 5.2, Individual health status dispersion analysis; Based on the estimated health status of each individual cell output from step 4, the dispersion of the individual cell health status is analyzed. The calculation method for the dispersion of health status is similar to that for voltage dispersion: the mean of all estimated health status values ​​of each individual cell is calculated as the average health status within the group; the difference between the health status of each individual cell and the mean is calculated; and the standard deviation of the difference is calculated as the dispersion of health status.

[0086] Based on the dispersion of health status, anomalous individuals are identified. An anomalous individual is defined as one whose health status deviates significantly from the group average, which may be a weak individual aging too rapidly or a faulty individual with abnormal data. Anomaly determination uses a statistical threshold method: assuming that the health status of normal individuals follows a distribution centered on the group mean, when the health status of an individual deviates from the mean by more than a certain number of standard deviations, it is determined to be an anomalous individual. The threshold for anomaly determination is determined based on a trade-off between false positive and false negative rates, with an optimal value set at two to three standard deviations.

[0087] For identified abnormal units, their abnormality type is further analyzed. By comparing the health status of the abnormal unit with the group mean, it is determined whether it is an abnormality of rapid aging or slow aging; by analyzing the historical health status trajectory of the abnormal unit, it is determined whether it is a gradual abnormality or a sudden abnormality. The determination of the abnormality type provides a basis for subsequent fault diagnosis and maintenance decisions.

[0088] Step 5.3, graph neural network for single-unit association modeling; Based on the physical structure of the battery pack and the health status data of individual cells, a graph neural network model is constructed to capture the interrelationships between cells. The individual cells within a battery pack are not independent but influence each other through electrical connections and thermal conduction. For example, cells connected in series carry the same current; a change in the internal resistance of one cell will affect the current distribution of the entire pack. Thermal conduction exists between adjacent cells; a temperature increase in one cell will affect the temperature of adjacent cells. Graph neural networks can effectively model these structured relationships.

[0089] The method for constructing the graph structure is as follows: the battery pack is modeled as a graph, with each individual battery cell serving as a node in the graph; the node feature vector contains information such as the estimated health status, voltage, and temperature of the individual cell; edge connections represent the relationships between individual cells, which can be defined based on physical adjacency, i.e., establishing edge connections between physically adjacent cells, or based on electrical connectivity, i.e., establishing edge connections between cells connected in series or parallel; edge features can include physical parameters such as the distance between individual cells and thermal resistance.

[0090] The forward propagation process of a graph neural network employs a message-passing mechanism. In each graph convolutional operation, each node collects information from its neighboring nodes and updates its own feature representation. The message-passing mechanism is implemented as follows: it aggregates the feature vectors of all neighboring nodes of each node, using methods such as summation, mean, or maximum value; it concatenates or adds the aggregated result to the node's own feature vector; and it obtains the updated node features through linear transformations and non-linear activation functions. After multiple layers of graph convolutional operations, each node's feature vector incorporates information from its multi-hop neighbors, reflecting the node's state and influence within the entire group.

[0091] Furthermore, since standard graph convolutional networks assign the same aggregation weights to all neighboring nodes, they may not accurately reflect the differences in the influence of different neighbors. Graph attention networks can be used as an alternative. Graph attention networks introduce an attention mechanism during message passing, adaptively learning the importance weights of different neighboring nodes. Specifically, for each pair of adjacent nodes, an attention coefficient is calculated by concatenating the feature vectors of the two nodes and then applying a linear transformation and activation function. The attention coefficients of all neighbors of each node are normalized. The normalized attention coefficients are then used to perform weighted aggregation of neighbor features. Graph attention networks can automatically identify the neighbors with the greatest influence on the current node, improving the accuracy of single-node association modeling.

[0092] Step 5.4, Group-level health status aggregation; Based on the updated feature vectors of each individual cell output by a graph neural network, a group-level health state aggregation function is designed. The overall performance of the battery pack is limited by the weakest cell, a characteristic known as the "weakest link" effect. Therefore, the group-level health state cannot simply be the arithmetic mean of the health states of each individual cell; instead, an aggregation function that reflects the weakest link effect needs to be designed.

[0093] The group-level health status aggregation function is designed using a weighted aggregation method, assigning higher weights to individuals with poorer health statuses. The weight allocation function is designed as follows: calculate the distance between each individual's health status and the lowest health status within the group; the smaller the distance, the higher the weight of the individual, and vice versa; the weight function can adopt an exponential decay form, i.e., the weight equals the negative distance multiplied by the decay coefficient and then the exponent; normalize the weights of all individuals so that their sum is one; the group-level health status is equal to the weighted sum of each individual's health status and its corresponding normalized weight.

[0094] The attenuation coefficient is a key parameter of the aggregation function, determining the strength of the "weakest link" effect. A larger attenuation coefficient results in a more concentrated weight distribution on the weakest cell, and the group-wide health state is closer to the health state of the weakest cell; a smaller attenuation coefficient results in a more uniform weight distribution, and the group-wide health state is closer to the average cell health state. The optimal value of the attenuation coefficient can be determined based on the series-parallel structure of the battery pack and application requirements. For battery packs with predominantly series connection, the "weakest link" effect is more pronounced, and a larger attenuation coefficient should be chosen.

[0095] Step 5.5, prediction of monomer inconsistency evolution; Based on historical data on the dispersion of individual cell health status, the evolution trend of cell inconsistency is predicted. Cell inconsistency typically intensifies gradually with cycle aging. When inconsistency reaches a critical level, the overall performance of the battery pack will decline, and may even lead to safety issues. Predicting the evolution trend of inconsistency helps in planning maintenance or replacement in advance.

[0096] Inconsistency evolution prediction employs a time-series extrapolation method. Historical health status dispersion data at various time points are collected to form a dispersion time series. Trend fitting is performed on the dispersion time series; the fitting model can be a linear model, an exponential model, or a multinomial model, with the model selection determined based on the actual evolution pattern of the dispersion. The fitted model is then used to extrapolate and predict future dispersion values, determining the time point when the dispersion reaches a critical threshold. The critical threshold is determined based on the battery pack's performance requirements and safety standards. When the dispersion exceeds the critical threshold, the battery pack should be taken out of service for maintenance or replacement.

[0097] The prediction results are output as a single-cell consistency score. The single-cell consistency score comprehensively reflects the current level of dispersion and the predicted evolution trend. The score is calculated as follows: the ratio of the current dispersion to the critical threshold is used as the current state component; the ratio of the predicted remaining safe usage time to the reference time is used as the trend component; the current state component and the trend component are weighted and summed to obtain the consistency score, with the weights determined according to application requirements. A higher consistency score indicates better single-cell consistency and a more reliable overall health status of the battery pack.

[0098] Step 5 outputs the overall health status assessment results of the battery pack, including: a group-level health status assessment value obtained by weighted aggregation based on the "weakest link" effect, reflecting the overall performance level of the battery pack; a cell consistency score, comprehensively reflecting the current dispersion level between cells and the future evolution trend of inconsistency; abnormal cell identification results, identifying cells whose health status deviates significantly from the group average level and their abnormality types; and the predicted time point when cell inconsistency reaches the critical threshold. These outputs provide crucial group-level health status information for subsequent remaining lifetime prediction and battery management decisions.

[0099] Step 6: Based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, define the criteria for determining effective cycles, predict the remaining number of effective cycles, calculate the confidence interval of the prediction results, and generate battery management recommendations. Specifically, the following steps are included: Step 6.1, Define the criteria for valid cycle determination; Based on the operational requirements of power tools, a criterion for determining an effective cycle is defined. An effective cycle refers to a cycle in which the battery can meet the minimum operational requirements of the power tool during a single charge-discharge cycle. The core indicator of the criterion is the battery's usable energy, defined as the total energy released by the battery during discharge from a fully charged state to the cutoff voltage. When the usable energy is not less than the product of the rated energy and the effective energy threshold coefficient, the cycle is determined to be an effective cycle; when the usable energy is lower than this threshold, the battery cannot meet the operational requirements of the power tool and is determined to be an ineffective cycle.

[0100] The effective energy threshold coefficient is determined by analyzing the minimum energy required for a power tool to complete a typical work task, and then dividing this minimum energy by the battery's rated energy to obtain the threshold coefficient. The threshold coefficient reflects the application scenario's requirements for battery performance; higher requirements result in a larger threshold coefficient, and the battery reaches its end-of-life threshold earlier. The optimal value for the threshold coefficient can be set according to the type of power tool and typical usage scenarios. For high-power, intermittently operating power tools, the threshold coefficient is typically set at a higher level to ensure operational reliability.

[0101] Specifically, for high-power, intermittently operating power tools, such as electric wrenches and hammer drills, the recommended effective energy threshold coefficient is 70% to 80%, with a typical value of 75%. This means that the battery is considered dead when its usable energy drops to 75% of its rated energy. For continuously operating power tools, such as angle grinders, the threshold coefficient can be appropriately reduced to 60% to 70%, with a typical value of 65%, because continuous operation requires relatively lower instantaneous power. The power judgment threshold is set to 1.1 to 1.2 times the peak power demand of the power tool, with a typical value of 1.15 times, to ensure sufficient power margin. For an electric wrench with a rated power of 1000 watts and a peak power of 1500 watts, the power judgment threshold is set at 1725 watts. The determination of an effective cycle uses a dual criterion of energy and power: when the usable energy is below the energy threshold or the maximum discharge power is below the power threshold, the cycle is considered invalid.

[0102] In addition to energy indicators, power indicators can be introduced as an auxiliary judgment criterion. Power indicators focus on the battery's instantaneous power supply capability under pulsed loads. When the battery's maximum discharge power is lower than the power tool's peak power requirement, it cannot function properly even with sufficient energy. The power judgment threshold is determined based on the power tool's peak power requirement and the battery's power characteristics.

[0103] Step 6.2, extrapolation and prediction of health status; Based on the health status prediction curve output in step 4, an extrapolation method is used to predict the time point when the health status drops to the failure threshold. The health status prediction curve provides predicted health status values ​​for several future working cycles, but the prediction range is usually limited and may not cover the time point when the health status drops to the failure threshold. The extrapolation method estimates health status values ​​further in the future by trend fitting and extending the prediction curve.

[0104] The extrapolation method is implemented as follows: The latter part of the health status prediction curve is selected as the fitting sample, as this data better reflects the current aging trend. A suitable extrapolation model is selected to fit the fitting sample. The selection of the extrapolation model needs to consider the nonlinear characteristics of health status decline. Optional models include linear models, exponential decay models, and double exponential models. Linear models assume that health status declines at a constant rate, suitable for the stable decline stage in the middle of aging. Exponential decay models assume that the rate of health status decline accelerates as health status decreases, suitable for the accelerated decline stage in the later stages of aging. Double exponential models combine rapid and slow decay components, capable of describing more complex aging behaviors. Model selection can be determined by comparing the goodness of fit of different models on validation data, or specified based on prior knowledge of the battery's aging mechanism. The extrapolation model obtained from the fitting is used to calculate the time point at which the health status drops to the failure threshold. The failure threshold is determined according to the effective cycle judgment criteria, usually set as a certain percentage level of the health status.

[0105] Furthermore, since a single extrapolation model may not accurately describe the changes in health status at different aging stages, a piecewise extrapolation method can be used as an alternative. The piecewise extrapolation method selects different extrapolation models based on the current level of health status: a linear model is used when health status is high; an exponential decay model is switched when health status drops to a moderate level; and a steeper decay model is used when health status approaches the failure threshold. Specifically, several health status segmentation points are set, and model fitting and extrapolation are performed separately within each segment interval. The segmentation points are determined based on the typical shape of the battery aging curve. The piecewise extrapolation method can more accurately capture the decay characteristics of health status at different stages, improving the accuracy of remaining life prediction.

[0106] Step 6.3, conversion of remaining valid loop counts; Based on the failure time point obtained from extrapolation prediction, the time is converted into the remaining effective cycle count. The failure time point is expressed in terms of work cycles or calendar time, and needs to be converted into cycle count according to the usage frequency of the power tool.

[0107] The conversion method is implemented as follows: statistically analyze the historical usage frequency of the battery. The usage frequency is defined as the number of charge-discharge cycles or working cycles per unit calendar time. The usage frequency can be calculated by analyzing historical usage data. The calculation method is to divide the total number of cycles by the total usage time. Based on the usage frequency, the failure time point is converted into the remaining number of cycles. The remaining number of cycles is equal to the remaining time multiplied by the usage frequency.

[0108] For cases where usage frequency fluctuates significantly, a weighted average method is used to calculate the representative usage frequency. The weighting method assigns higher weight to recent usage frequencies to reflect current usage patterns. The weighting function can employ an exponential decay form, with more recent data having a greater weight than older data.

[0109] In addition, a usage frequency prediction function can be introduced to predict future usage frequency based on the periodic patterns of historical usage patterns. For example, construction projects may experience changes in usage intensity due to milestones in the project schedule; identifying such periodic patterns can improve the accuracy of cycle number conversion.

[0110] Step 6.4, predict uncertainty by comprehensive quantification; Based on the uncertainties in health status prediction and usage frequency, the confidence interval for the prediction of the remaining effective number of cycles is calculated comprehensively. The uncertainty in the prediction of the remaining number of cycles stems from multiple sources: the uncertainty of the health status prediction model itself, the uncertainty of the extrapolation model, and the uncertainty of the usage frequency estimation. Quantifying these uncertainties comprehensively is crucial for assessing the reliability of the prediction results.

[0111] Uncertainty quantification employs the Monte Carlo simulation method. The simulation process is as follows: A health status trajectory is randomly sampled from the distribution of health status predictions, using the predicted mean and variance output in step 4, assuming the predicted values ​​follow a normal distribution; the sampled health status trajectory is extrapolated to obtain a failure time point sample; a usage frequency value is randomly sampled from the distribution of usage frequency, the distribution of usage frequency being determined based on the statistical characteristics of historical data; the failure time point sample and the usage frequency sample are multiplied to obtain a sample of remaining cycle counts; the above process is repeated several times to obtain a sample set of remaining cycle counts; the predicted mean, standard deviation, and confidence interval are calculated based on the sample set.

[0112] The confidence interval is calculated as follows: The sample set is sorted, and certain percentiles are used as the lower and upper bounds of the confidence interval. For example, a 90% confidence interval uses the 5th and 95th percentiles. The width of the confidence interval reflects the degree of uncertainty in the prediction; a wider interval indicates greater uncertainty in the prediction.

[0113] Step 6.5: Battery management recommendations are generated; Based on the predicted remaining effective cycle count and confidence intervals, battery management recommendations are generated. These recommendations are automatically generated according to the prediction results and preset decision rules, providing users with actionable guidance.

[0114] The suggested decision-making rules include: when the predicted mean of the remaining effective cycles is lower than the emergency replacement threshold, an immediate replacement suggestion is generated, prompting the user that the battery is nearing the end of its life and should be replaced as soon as possible to avoid equipment failure; when the predicted mean of the remaining effective cycles is between the planned replacement threshold and the emergency replacement threshold, a planned replacement suggestion is generated, prompting the user to schedule battery replacement within a future time window; when the predicted mean of the remaining effective cycles is higher than the planned replacement threshold but the lower bound of the confidence interval is lower than the emergency replacement threshold, an increased monitoring suggestion is generated, prompting the user that the prediction uncertainty is high and the monitoring frequency should be increased to obtain more data support; when the individual cell consistency score is lower than the consistency threshold, a balanced maintenance suggestion is generated, prompting the user that the individual cells in the battery pack have significant inconsistencies and may require equalization or individual cell replacement.

[0115] The threshold settings are determined based on the risk tolerance and economic considerations of the application scenario. The emergency replacement threshold is set relatively low to ensure a safety margin, while the planned replacement threshold is set relatively high to allow sufficient preparation time. The consistency threshold is determined based on the battery pack's performance requirements. Recommended decision threshold settings are as follows: the emergency replacement threshold is 50 to 100 remaining effective cycles, with a typical value of 75 cycles; an immediate replacement recommendation is generated when the predicted mean is below this threshold. The planned replacement threshold is 150 to 300 remaining effective cycles, with a typical value of 200 cycles; a planned replacement recommendation is generated when the predicted mean is between the emergency replacement threshold and the planned replacement threshold. The consistency threshold is defined based on the health status dispersion; when the dispersion (standard deviation) exceeds 15% to 20%, with a typical value of 18%, it is considered that the individual cell inconsistency is too large, and a balanced maintenance recommendation is generated. For applications with high reliability requirements, each threshold should be appropriately increased to increase the safety margin; the emergency replacement threshold can be increased to 100 to 150 cycles, the planned replacement threshold can be increased to 300 to 500 cycles, and the consistency threshold can be reduced to 10% to 15%. For cost-sensitive applications, the thresholds can be appropriately lowered, but the emergency replacement threshold should not be lower than 50 times to ensure basic safety.

[0116] Step 6 outputs the predicted remaining effective battery cycles and battery management recommendations, including: the mean remaining effective battery cycles predicted using an extrapolation method; the prediction confidence interval calculated using Monte Carlo simulation, reflecting the degree of uncertainty in the prediction results; battery management recommendations automatically generated according to decision rules, including recommendation types such as immediate replacement, planned replacement, increased monitoring, or balanced maintenance; and the current settings for various decision thresholds. These outputs provide users with actionable battery management guidance, supporting equipment maintenance decisions and asset management optimization.

[0117] Step 7: Format the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, perform data synchronization and time series prediction model optimization through edge-cloud collaboration, and record system operation logs; Specifically, the following steps are included: Step 7.1, Format the evaluation results; Based on the evaluation results output from each step, a standardized evaluation report is generated through formatting. The evaluation report uses a structured data format and includes the following fields: The battery identification field records the unique identifier of the battery pack being evaluated, used to distinguish different battery packs and trace historical records; the evaluation timestamp field records the time when this evaluation was completed, using a standard time format to support time-series analysis; the health status percentage field records the current health status estimate, expressed as a percentage, with 100% corresponding to full health status; the remaining effective cycles field records the predicted average number of remaining effective cycles; the confidence interval field records the upper and lower bounds of the confidence interval for the predicted number of remaining effective cycles, as well as the corresponding confidence levels; the individual cell consistency score field records the individual cell consistency score of the battery pack, reflecting the degree of health status balance among individual cells; the management recommendation field records the management recommendation code and recommendation text generated based on the evaluation results; and the key feature field records the key health feature values ​​extracted in this evaluation, including core indicators such as dynamic internal resistance and voltage recovery time constant.

[0118] The formatted evaluation report can be output in multiple formats, including structured data formats suitable for machine parsing and text formats suitable for human reading.

[0119] Step 7.2, Local display and user interaction; Based on the formatted evaluation report, the evaluation results are displayed to the user through a local display interface. The display interface can be the battery management system's built-in display, the power tool's status indicator, or a mobile terminal application connected via Bluetooth.

[0120] The display modes are divided into a concise mode and a detailed mode. The concise mode is suitable for quick viewing scenarios, displaying only the percentage of health status and summary information on management recommendations. Health status levels are intuitively represented by color coding: green indicates good health, yellow indicates attention is needed, and red indicates immediate action is required. The detailed mode is suitable for in-depth analysis scenarios, displaying complete assessment indicators, historical trend charts, and predictive curves. It allows users to view the health status distribution of individual units and inconsistency analysis results.

[0121] The user interaction features support users in querying historical evaluation records, setting alarm thresholds, and triggering manual evaluations. The historical record query function allows users to view evaluation results within a specified time range, supporting trend analysis and anomaly tracing. The alarm threshold setting function allows users to adjust various threshold parameters according to actual needs, enabling personalized alarm strategies. The manual evaluation trigger function allows users to proactively initiate evaluations when needed, rather than waiting for the automatic evaluation cycle.

[0122] Step 7.3, cloud data synchronization; Based on local assessment results and key feature data, the data is uploaded to the cloud platform via a wireless communication module. The cloud synchronization function enables data aggregation and analysis across devices, supporting rental service providers in centralized monitoring and management of all devices.

[0123] The uploaded data includes: formatted assessment reports, time-series data of key health characteristics, and abnormal event records. The uploaded data has undergone anonymization to remove sensitive information that may involve user privacy, retaining only technical data related to battery health. Lossless compression algorithms are used to reduce transmission bandwidth consumption. Standard encryption protocols are employed to ensure the security of the transmission process.

[0124] Data upload timing employs an intelligent scheduling strategy. Batch uploads are performed when network connectivity is good and devices are idle to avoid impacting normal device operation. For urgent alarm information, an immediate upload strategy ensures timely notification. When network connectivity is unavailable, data is temporarily stored in a local cache and automatically uploaded once network access is restored.

[0125] After receiving the uploaded data, the cloud platform aggregates, stores, and analyzes the data. Cross-device analysis compares the health status distribution of different devices, identifying abnormal devices and common problems. The group aging pattern mining function analyzes aging data from a large number of devices, extracting aging patterns and influencing factors to provide data support for battery design optimization and usage strategy improvement.

[0126] Step 7.4, online model update; Based on the analysis results and optimized model parameters from the cloud platform, model updates are sent to local devices via the downlink channel. This online model update function enables continuous improvement of the local evaluation model, adapting to new battery characteristics and usage patterns.

[0127] The model update includes: updating the weight parameters of the deep learning model, adjusting the parameters of the feature extraction algorithm, and optimizing the decision threshold. The update package uses an incremental update method, transmitting only the changed parameters to reduce the amount of data transmitted. The update package includes a version number and a checksum to ensure the completeness and correctness of the update.

[0128] Model updates employ a secure update mechanism. The update process is performed during device idle periods to avoid impacting normal evaluation functions. Current model parameters are backed up before the update, and an automatic rollback to the backup version is initiated in case of update failure. Functional verification is conducted after the update, and the new model can only be switched to after confirming its proper operation.

[0129] The update policy supports two modes: forced updates and optional updates. Forced updates are used to fix known defects or security vulnerabilities; the device must perform the update after receiving a forced update notification. Optional updates are used for performance optimization or feature enhancement; users can choose whether to accept the update.

[0130] Step 7.5, System operation log recording; Based on the status information of each stage of the evaluation process, a system operation log is recorded. The operation log is used for system status monitoring, performance analysis, and fault diagnosis, and is an important support for ensuring the reliable operation of the system.

[0131] The log records include: data acquisition status, recording the acquisition success rate and data quality indicators of each sensor channel; feature extraction results, recording the extraction success rate of various features and the statistical distribution of feature values; model inference status, recording the model inference time, memory usage, and abnormal situations; communication status, recording the connection status with the cloud platform and data transmission; and system resource status, recording processor utilization, memory usage, and storage space, etc.

[0132] Log levels are divided into debug, information, warning, and error levels. Debug logs record detailed process information for development debugging; information logs record normal operating status for daily monitoring; warning logs record abnormal situations that do not affect functionality, indicating situations that require attention; and error logs record serious problems that cause functional abnormalities, requiring timely handling.

[0133] Log storage employs a cyclic overwrite strategy; when storage space reaches its limit, the oldest log record is automatically deleted. Critical logs can be marked as protected and are not subject to automatic deletion. Logs support local viewing and remote export, facilitating problem analysis by technical support personnel.

[0134] Step 7 outputs a user-readable assessment report and system operation log, including: a formatted, standardized assessment report with complete information such as battery identification, health status, remaining cycle count, confidence interval, consistency score, and management recommendations; assessment results displayed through a local interface, supporting user interaction in both concise and detailed modes; key data uploaded to the cloud platform for cross-device analysis and group aging pattern mining; model update packages received from the cloud to support continuous optimization of the local assessment model; and system operation logs recording information such as data acquisition, feature extraction, model inference, communication, and resource status. These outputs provide users with comprehensive battery health management services, supporting device maintenance decisions, asset management optimization, and system reliability assurance.

[0135] A battery health status assessment system for an electric tool power supply device is used to perform the aforementioned battery health status assessment method for an electric tool power supply device, such as... Figure 3 As shown, it includes: The data acquisition module is used to acquire multi-source sensor data during the operation of power tools and preprocess it to obtain a normalized time-series dataset; The feature extraction module, based on a normalized time-series dataset, identifies and classifies multi-level pulse discharge events, extracts multi-scale voltage response features and performs temperature compensation to obtain a temperature-compensated health feature vector. The parameter identification module identifies the parameters of the equivalent circuit model based on the normalized time-series dataset and the temperature-compensated health feature vector, verifies and removes the abnormal identification results, and obtains the parameter aging characteristics. The health prediction module is used to fuse the temperature-compensated health feature vector and parameter aging features to construct a time-series prediction model for prediction, and obtain the health status estimation results of each individual cell. The consistency analysis module evaluates the overall health status and individual cell consistency of the battery pack based on the health status estimation results of each individual cell, and outputs the overall health status assessment results of the battery pack. The lifespan prediction module defines the criteria for determining effective cycles based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, predicts the remaining effective cycle count, calculates the confidence interval of the prediction results, and generates battery management recommendations. The results output module is used to format the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, perform data synchronization and time series prediction model optimization through edge-cloud collaboration, and record system operation logs.

[0136] In one embodiment of the present invention, a specific example is provided: A power tool rental service provider deployed several electric wrenches equipped with the evaluation system of this invention at a large construction site. These electric wrenches are used for tightening bolts on steel structures. Each wrench is equipped with two replaceable lithium-ion battery packs, with a nominal voltage of 18 volts and a nominal capacity of 5 amp-hours. The ambient temperature at the construction site varies with the seasons, reaching over 40 degrees Celsius in summer and dropping to minus 5 degrees Celsius in winter. The equipment is used by different operators in rotation, with the intensity of use varying according to the progress of the construction.

[0137] To verify the effectiveness of the method of this invention, a three-month field test was conducted at the construction site. During the test, ten electric wrenches equipped with the evaluation system were deployed, and twenty sets of batteries were used in rotation. The test covered different environmental conditions from high summer temperatures to normal autumn temperatures. Battery operation data was continuously collected during the test, accumulating complete data for over five thousand working cycles.

[0138] The raw sensor data collected by a battery pack during a typical pulse discharge cycle are shown in Table 1: Table 1: Raw sensor data collected by a battery pack during a typical pulse discharge cycle; The working cycle includes two pulse discharge events, each lasting about one second, with a pulse current of 45 to 50 amperes and a voltage drop of about 1.3 to 1.6 volts. During the resting relaxation phase, the voltage gradually recovers.

[0139] Table 2 shows the voltage recovery characteristics of a battery pack during the resting relaxation phase after pulse discharge: Table 2: Voltage recovery characteristics of a battery pack during the resting relaxation phase after pulse discharge; The relaxation process shows that the voltage recovers rapidly after the pulse ends, with a high initial recovery rate followed by a gradual slowdown, consistent with exponential decay characteristics. The standard deviation of the individual cell voltage decreases with increasing relaxation time, reflecting the consistency level among individual cells.

[0140] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for assessing the battery health status of a power tool power supply device, characterized in that, Includes the following steps: Acquire multi-source sensor data during the operation of power tools and preprocess it to obtain a normalized time-series dataset; Based on a normalized time-series dataset, multi-level pulse discharge events are identified and classified, multi-scale voltage response features are extracted and temperature compensation is performed to obtain a temperature-compensated health feature vector. Based on the normalized time-series dataset and the temperature-compensated health feature vector, the equivalent circuit model parameters are identified, the abnormal identification results are verified and eliminated, and the parameter aging features are extracted. The health feature vector after temperature compensation and the parameter aging feature are fused to construct a time-series prediction model for prediction, and the health status estimation results of each individual cell are obtained. Based on the health status estimation results of each individual cell, the overall health status and individual cell consistency of the battery pack are evaluated, and the overall health status evaluation result of the battery pack is output. Based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, the criteria for determining effective cycles are defined, the remaining number of effective cycles is predicted, the confidence interval of the prediction results is calculated, and battery management recommendations are generated. The remaining effective cycle count prediction results and battery management recommendations are formatted into a standardized evaluation report. Data synchronization and time series prediction model optimization are performed through edge-cloud collaboration, and system operation logs are recorded.

2. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The process of acquiring and preprocessing multi-source sensor data during the operation of power tools includes: The terminal voltage, charging and discharging current, and surface temperature of the power tool battery pack are collected in real time, and a unified clock source is used to drive the analog-to-digital converters of each channel. The median filtering algorithm is used to denoise the original sensor signal; Missing value detection and imputation are performed. When a data point is null, exceeds the physical reasonable range, or the rate of change between adjacent data points exceeds the preset maximum rate of change threshold, it is determined to be a missing value. For the detected missing values, an imputation method based on local weighted regression is used to fill them; Based on the load status identification signal, continuous operating data is divided into independent work cycle segments.

3. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The obtained temperature-compensated health feature vector specifically includes: Based on the current time series data within the working cycle segment, multi-level pulse discharge events are identified and classified. A multi-scale and multi-condition joint judgment method is adopted to calculate the first-order and second-order difference sequences of the current signal and search for positive and negative jump points in the difference sequences. Based on the pulse current amplitude, the pulses are divided into three levels: light load pulse, medium load pulse, and heavy load pulse. Multi-scale voltage response features were extracted for pulses of different levels, including transient voltage drop amplitude, steady-state voltage drop amplitude, voltage drop rate, multi-stage voltage recovery time constant, graded dynamic internal resistance, and pulse energy efficiency. Based on battery temperature data and a temperature compensation model, the equivalent values ​​of the eigenvalues ​​at the reference temperature are calculated.

4. The method for assessing the battery health status of a power tool power supply device according to claim 3, characterized in that, The method for extracting the multi-stage voltage recovery time constant is as follows: During the voltage recovery phase after the pulse ends, a double exponential function model is used to fit the recovery curve, and nonlinear least squares method is used for parameter identification. The iterative process adopts the Levenberg-Marquardt algorithm, and the convergence criterion is that the parameter change between two adjacent iterations is less than the preset convergence threshold or the number of iterations exceeds the preset maximum number of iterations.

5. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The extracted aging features specifically include: A second-order RC equivalent circuit model adaptive to pulse operating conditions is established, and current dependence and frequency dependence corrections are introduced. The model structure includes an ideal voltage source, a current-dependent ohmic internal resistance, a first RC parallel link, and a second RC parallel link. Based on the discretized state-space model and measured voltage and current data, the recursive least squares method is used to identify the model parameters online, and a forgetting factor mechanism is introduced. The parameter estimates are validated to eliminate outlier results. The validation uses multiple criteria, including physical range validation, residual validation and rate of change validation. Based on the validated model parameter sequence, parameter aging features are extracted.

6. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The specific results of obtaining the health status estimation of each individual battery cell include: Feature fusion is performed to construct the model input. Feature fusion adopts a vector concatenation method, and each feature is standardized before concatenation. A sliding window method is used to extract a fixed-length subsequence from the feature vector sequence as the model input. The structure of the Long Short-Term Memory (LSTM) network is designed. The hidden layer of the network is composed of stacked LSM units. Each LSM unit contains four core components: a forget gate, an input gate, an output gate, and a cell state. An ensemble prediction method is used to quantify uncertainty by training multiple long short-term memory network models with different initialization parameters and using the distribution of prediction results from multiple models to estimate prediction uncertainty.

7. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The overall health status assessment results of the output battery pack specifically include: Collect voltage and temperature data of each individual cell in the battery pack, and calculate the dispersion index between individual cells. The dispersion of the health status of individual cells is analyzed to identify abnormal cells. The abnormality determination adopts a statistical threshold method. When the health status of a cell deviates from the mean by more than a preset multiple threshold, it is determined to be an abnormal cell. A graph neural network model is constructed to capture the mutual influence relationships between individual cells. The battery pack is modeled as a graph, with each individual battery cell as a node in the graph. Edges connect the cells to represent the relationships between them. The forward propagation process of the graph neural network adopts a message passing mechanism. Based on the updated feature vectors of each individual entity output by the graph neural network, a group-level health status aggregation function is designed to aggregate the health status of each individual entity.

8. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The battery management recommendations specifically include: Define the criteria for effective cycle. The core indicator of the criteria is the battery's available energy. When the available energy is not less than the product of the rated energy and the effective energy threshold coefficient, it is judged as an effective cycle. Extrapolation is used to predict the time point when the health status drops to the failure threshold. The latter part of the health status prediction curve is selected as the fitting sample, and the extrapolation model is selected to fit the fitting sample. Convert the failure time point into the remaining effective cycle count, count the historical usage frequency of the battery, and convert the failure time point into the remaining cycle count based on the usage frequency; The confidence intervals for the predicted remaining effective number of cycles are calculated comprehensively, and the uncertainty is comprehensively quantified using the Monte Carlo simulation method; Based on the predicted remaining effective cycle count and confidence interval, battery management recommendations are generated.

9. The method for assessing the battery health status of a power tool power supply device according to claim 1, characterized in that, The process of formatting the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, performing data synchronization and time-series prediction model optimization through edge-cloud collaboration, and recording system operation logs specifically includes: The data is formatted to generate a standardized evaluation report, which uses a structured data format. Data is uploaded to the cloud platform via a wireless communication module. After receiving the uploaded data, the cloud platform aggregates, stores, and analyzes the data. The model update is sent to the local device through the downlink channel. The model update includes the update of the weight parameters of the deep learning model, the adjustment of the parameters of the feature extraction algorithm, and the optimization of the decision threshold. Based on the status information of each stage of the evaluation process, the system operation log is recorded. The log records include data acquisition status, feature extraction results, model inference status, communication status, and system resource status.

10. A battery health status assessment system for an electric tool power supply device, characterized in that, A method for assessing the battery health status of a power tool power supply device according to any one of claims 1-9 includes: The data acquisition module is used to acquire multi-source sensor data during the operation of power tools and preprocess it to obtain a normalized time-series dataset; The feature extraction module, based on a normalized time-series dataset, identifies and classifies multi-level pulse discharge events, extracts multi-scale voltage response features and performs temperature compensation to obtain a temperature-compensated health feature vector. The parameter identification module identifies the parameters of the equivalent circuit model based on the normalized time-series dataset and the temperature-compensated health feature vector, verifies and removes the abnormal identification results, and obtains the parameter aging characteristics. The health prediction module is used to fuse the temperature-compensated health feature vector and parameter aging features to construct a time-series prediction model for prediction, and obtain the health status estimation results of each individual cell. The consistency analysis module evaluates the overall health status and individual cell consistency of the battery pack based on the health status estimation results of each individual cell, and outputs the overall health status assessment results of the battery pack. The lifespan prediction module defines the criteria for determining effective cycles based on the health status estimation results of each individual battery cell and the overall health status assessment results of the battery pack, predicts the remaining effective cycle count, calculates the confidence interval of the prediction results, and generates battery management recommendations. The results output module is used to format the remaining effective cycle count prediction results and battery management recommendations into a standardized evaluation report, perform data synchronization and time series prediction model optimization through edge-cloud collaboration, and record system operation logs.

Images