A Deep Learning-Based Method and Device for Predicting Remaining Life of UAV Lithium Batteries

Abstract

This application provides a method and apparatus for predicting the remaining lifespan of drone lithium batteries based on deep learning. It achieves accurate feature acquisition through cyclic identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable lifespan assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively addresses the shortcomings of traditional techniques in feature extraction, lifespan prediction, and uncertainty assessment, providing technical support for drone battery management.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for predicting the remaining lifespan of a drone's lithium battery based on deep learning. Background Technology

[0002] Existing methods for predicting lithium battery life have significant shortcomings. Traditional systems perform poorly in data acquisition and cycle identification, failing to effectively extract health factors and thus affecting prediction results.

[0003] Furthermore, existing technologies suffer from bottlenecks in temporal coding and state representation. Most systems lack robust attention mechanisms and physical constraint strategies, resulting in suboptimal lifetime prediction accuracy.

[0004] Existing systems have technical shortcomings in uncertainty assessment. The lack of in-depth analysis of prediction results makes it difficult to efficiently determine safety margins using confidence intervals, thus affecting prediction reliability. Solving these problems is crucial for improving battery management capabilities. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides a method and apparatus for predicting the remaining lifespan of UAV lithium batteries based on deep learning. This method can effectively solve the shortcomings of traditional technologies in feature extraction, lifespan prediction, and uncertainty assessment, and provide technical support for UAV battery management.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] In a first aspect, this application provides a deep learning-based method for predicting the remaining lifespan of a drone's lithium battery, comprising:

[0008] The raw operation data of charge and discharge cycles are obtained from the UAV battery management system and the timestamp alignment and resampling are performed to obtain standardized cycle data. The cycle data is then used to identify the cycle boundary based on the zero crossing point of the charge and discharge current to obtain a set of cycle data segments. Incremental capacity analysis and differential voltage analysis are performed on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number.

[0009] After normalizing the time series of health factors, the data is input into a long short-term memory network for time-series encoding to obtain a hidden state sequence. Attention weights are calculated and weighted aggregation is performed on the hidden state sequence to obtain an aging state representation vector. The aging state representation vector is input into a decoder network that incorporates physical constraints on capacity decay to obtain a predicted remaining lifespan. Multiple Monte Carlo random inactivation forward propagation is performed on the decoder network and the variance is statistically output to obtain a measurement of prediction uncertainty.

[0010] The remaining lifetime prediction value and the prediction uncertainty measure are used to calculate the remaining lifetime confidence interval based on a preset confidence level. The remaining lifetime prediction value and the remaining lifetime confidence interval are then sent to the UAV flight management system for battery safety margin determination.

[0011] Furthermore, it also includes: reading battery terminal voltage sampling data, charging and discharging current sampling data, and battery temperature sampling data from the data interface of the drone battery management system; performing outlier detection on the voltage sampling data, the current sampling data, and the temperature sampling data and marking sampling points that exceed the preset physical reasonable range as invalid; and performing linear interpolation to complete the marked invalid sampling points to obtain a complete set of sampling data.

[0012] The data of each type in the complete sampled data set is timestamped according to a unified time base. The resampling frequency is configured according to the data type differences. High-frequency resampling is performed on the voltage sampled data and the current sampled data, and low-frequency resampling is performed on the temperature sampled data. The resampled data of each type are merged according to the timestamp index to obtain standardized cyclic data.

[0013] Furthermore, it also includes: performing zero-crossing detection on the charging and discharging current in the standardized cyclic data, determining the charging start point based on the moment when the charging current changes from zero to positive, and determining the discharging end point based on the moment when the discharging current changes from negative to zero, and obtaining a set of cyclic data segments based on the data segments between adjacent charging start points;

[0014] For each cyclic data segment in the cyclic data segment set, the ratio of capacity change to voltage change is calculated to obtain the incremental capacity curve, and the peak position and peak amplitude are extracted as incremental capacity features. The ratio of voltage change to capacity change is calculated to obtain the differential voltage curve, and the feature point position and feature segment slope are extracted as differential voltage features. The incremental capacity features and the differential voltage features are assembled into a health factor vector and organized into a health factor time series according to the cyclic number.

[0015] Furthermore, it also includes: performing zero-mean unit variance standardization on each health factor vector in the health factor time series based on the training set statistics to obtain a normalized health factor sequence; inputting the normalized health factor sequence into a stacked long short-term memory network layer for sequential processing at each time step and outputting the hidden state at each time step to obtain a hidden state sequence.

[0016] A linear transformation of the query vector and key vector is performed on the hidden states at each time step in the hidden state sequence, and the dot product is calculated to obtain the attention score matrix. The attention score matrix is ​​then scaled and normalized to obtain the attention weights. Based on the attention weights, the hidden state sequence is weighted and aggregated to obtain the aging state representation vector.

[0017] Furthermore, it also includes: inputting the aging state characterization vector into the multi-layer fully connected layer of the decoder network for nonlinear mapping, wherein the decoder network performs forward propagation based on the network parameters trained according to the hybrid loss function of fusion capacity monotonically decreasing constraint and decay rate constraint to obtain the remaining lifetime prediction value.

[0018] During the prediction phase, the random deactivated layer activation state of the decoder network is maintained, and multiple forward propagations are performed on the same aging state representation vector to obtain a set of predicted values. The sample variance of the set of predicted values ​​is calculated to obtain a measure of prediction uncertainty.

[0019] Furthermore, it also includes: calculating the arithmetic mean of the sample set of predicted values ​​to obtain the predicted mean, and performing a square root operation on the predicted uncertainty measure to obtain the predicted standard deviation;

[0020] Based on the preset confidence level, the corresponding standard normal distribution quantile is queried, and the remaining life confidence interval is obtained by adding or subtracting the product of the quantile and the predicted standard deviation from the predicted mean.

[0021] Furthermore, it also includes: encapsulating the remaining lifetime prediction value and the remaining lifetime confidence interval into a prediction output message according to a preset message format, and sending the prediction output message to the UAV flight management system through a data communication interface;

[0022] The UAV flight management system parses the predicted output message to extract the remaining life prediction value and the lower bound of the confidence interval. It then compares the lower bound of the confidence interval with a preset safety margin threshold. When the lower bound of the confidence interval is lower than the safety margin threshold, a battery replacement warning signal is generated.

[0023] Secondly, this application provides a deep learning-based device for predicting the remaining lifespan of a drone's lithium battery, comprising:

[0024] The battery data acquisition module is used to obtain the raw operating data of the charge and discharge cycle from the UAV battery management system and perform timestamp alignment and resampling to obtain standardized cycle data. Based on the zero-crossing point of the charge and discharge current, the standardized cycle data is used to perform cycle boundary identification to obtain a set of cycle data segments. Incremental capacity analysis and differential voltage analysis are performed on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number.

[0025] The battery life prediction module is used to normalize the time series of health factors and input it into a long short-term memory network for time-series encoding to obtain a hidden state sequence. Attention weights are calculated and weighted aggregation is performed on the hidden state sequence to obtain an aging state representation vector. The aging state representation vector is input into a decoder network that integrates physical constraints of capacity decay to obtain a remaining life prediction value. Multiple Monte Carlo random inactivation forward propagation is performed on the decoder network and the variance is statistically output to obtain a prediction uncertainty measure.

[0026] The battery data feedback module is used to calculate the remaining life confidence interval based on the predicted remaining life value and the predicted uncertainty measure according to a preset confidence level, and send the predicted remaining life value and the remaining life confidence interval to the UAV flight management system for battery safety margin determination.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning-based method for predicting the remaining life of a drone lithium battery.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning-based method for predicting the remaining lifespan of a drone lithium battery.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the deep learning-based method for predicting the remaining lifespan of a drone lithium battery.

[0030] As described above, this application provides a method and apparatus for predicting the remaining lifespan of UAV lithium batteries based on deep learning. It achieves accurate feature acquisition through cyclic identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable lifespan assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively addresses the shortcomings of traditional techniques in feature extraction, lifespan prediction, and uncertainty assessment, providing technical support for UAV battery management. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the deep learning-based method for predicting the remaining lifespan of a drone lithium battery in an embodiment of this application.

[0033] Figure 2 This is a structural diagram of the deep learning-based drone lithium battery remaining life prediction device in the embodiments of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0036] To address the problems existing in current technologies, this application provides a method and apparatus for predicting the remaining lifespan of UAV lithium batteries based on deep learning. This method achieves accurate feature acquisition through cyclic identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable lifespan assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively solves the shortcomings of traditional technologies in feature extraction, lifespan prediction, and uncertainty assessment, providing technical support for UAV battery management.

[0037] To effectively address the shortcomings of traditional technologies in feature extraction, lifetime prediction, and uncertainty assessment, and to provide technical support for drone battery management, this application provides an embodiment of a deep learning-based method for predicting the remaining lifetime of drone lithium batteries. See [link to relevant documentation]. Figure 1 The deep learning-based method for predicting the remaining lifespan of drone lithium batteries specifically includes the following:

[0038] Step S101: Obtain the original operation data of charge and discharge cycles from the UAV battery management system and perform timestamp alignment and resampling to obtain standardized cycle data. Perform cycle boundary identification on the standardized cycle data based on the zero crossing point of the charge and discharge current to obtain a set of cycle data segments. Perform incremental capacity analysis and differential voltage analysis on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number.

[0039] In this embodiment, battery terminal voltage sampling data, charging / discharging current sampling data, and battery temperature sampling data are read from the data interface of the UAV battery management system. These three types of data originate from the independent sampling channels of the voltage sensor, current sensor, and temperature sensor, respectively. The data acquisition module performs outlier detection on the read sampling data, marking sampling points that exceed a preset physical reasonable range as invalid. For the marked invalid sampling points, a linear interpolation method using adjacent valid sampling points is used to complete the data, thereby obtaining a complete set of sampling data.

[0040] Based on the complete sampled data set, this embodiment performs timestamp alignment processing on each type of data according to a unified time base. Voltage and current sampled data are resampled at a higher frequency to capture the dynamic changes during charging and discharging, while temperature sampled data, due to its relatively gradual rate of change, is resampled at a lower frequency to reduce storage overhead. After resampling, this embodiment merges the data of each type according to their timestamp index to obtain standardized cyclic data.

[0041] Accordingly, this embodiment performs zero-crossing detection on the charging and discharging current in the standardized cyclic data to complete the cycle boundary identification. The charging start point is determined based on the moment when the charging current changes from zero to positive, and the discharging end point is determined based on the moment when the discharging current changes from negative to zero. This embodiment divides the complete charging and discharging process into independent cyclic data segments based on the data segments between adjacent charging start points, and all cyclic data segments are organized into a cyclic data segment set according to time sequence.

[0042] Based on the aforementioned set of cyclic data segments, this embodiment performs incremental capacity analysis on each cyclic data segment to extract features reflecting the battery's aging state. Incremental capacity is defined as the ratio of capacity change to voltage change. This embodiment calculates the incremental capacity curve for each cyclic data segment during the charging process, and extracts the peak position and peak amplitude from this curve as incremental capacity features. The peak position of the incremental capacity curve corresponds to the voltage plateau of a specific electrochemical reaction within the battery, and the peak amplitude reflects the activity level of that electrochemical reaction. Both features exhibit regular changes with battery aging.

[0043] Next, this embodiment performs differential voltage analysis on each cyclic data segment in the cyclic data segment set. Differential voltage is defined as the ratio of voltage change to capacity change. This embodiment calculates the differential voltage curve for each cyclic data segment and extracts the feature point positions and feature segment slopes from these curves as differential voltage features. The feature point positions reflect the voltage response of the phase transition process inside the battery, and the feature segment slopes reflect the changing trend of the battery's internal resistance.

[0044] Based on the outputs of the aforementioned incremental capacity analysis and differential voltage analysis, this embodiment assembles the incremental capacity features and the differential voltage features into a health factor vector. Each cyclic data segment corresponds to a health factor vector, and each component of the health factor vector records the peak position, peak amplitude, feature point position, and feature segment slope of that cycle. In this embodiment, each health factor vector is organized into a health factor time series according to the cycle number. The health factor time series is read by the normalization module and the long short-term memory network in subsequent step S201 as input data for time series encoding.

[0045] Step S102: After normalizing the health factor time series, input it into a long short-term memory network for time-series encoding to obtain a hidden state sequence. Perform attention weight calculation and weighted aggregation on the hidden state sequence to obtain an aging state representation vector. Input the aging state representation vector into a decoder network that integrates capacity decay physical constraints to obtain a remaining lifetime prediction value. Perform multiple Monte Carlo random inactivation forward propagation on the decoder network and statistically output the variance to obtain a prediction uncertainty measure.

[0046] In this embodiment, the health factor time series generated in step S101 is read, and each health factor vector in the sequence is standardized with zero mean and unit variance based on the statistics of the training set. The normalization module pre-calculates and stores the mean and standard deviation of each health factor component from the training dataset. In the online prediction stage, the input health factor vector is transformed based on the stored statistics, thereby obtaining the normalized health factor sequence.

[0047] Based on the normalized health factor sequence, this embodiment inputs it into a stacked long short-term memory (LSTM) network layer for temporal encoding. The stacked LSM network layer comprises multiple layers of sequentially connected LSM units. The first layer receives the normalized health factor sequence as input and processes the health factor vectors corresponding to each cycle sequentially according to time steps, outputting a hidden state vector at each time step. The second layer uses the hidden state sequence output by the first layer as input for higher-level temporal feature abstraction. In this embodiment, the outputs of each time step of the last LSM network layer are organized into a hidden state sequence.

[0048] Accordingly, this embodiment performs attention weight calculation on the hidden state sequence to identify historical cycles that significantly contribute to remaining lifetime prediction. The attention module performs linear transformations on the query vector and key vector for the hidden states at each time step in the hidden state sequence, and calculates the attention score matrix by the dot product of the query vector and key vector. This embodiment performs scaling on the attention score matrix to stabilize the gradient, and then performs normalization to obtain the attention weights.

[0049] Based on the attention weights, this embodiment performs weighted aggregation on the hidden state sequence according to these weights. Weighted aggregation sums the hidden state vectors at each time step according to their corresponding attention weights; time steps with larger weights contribute more significantly to the aggregation result. The output of weighted aggregation is an aging state representation vector, which encodes the cumulative aging information of the battery from the first cycle to the current cycle.

[0050] Next, in this embodiment, the aging state representation vector is input into a decoder network that incorporates physical constraints on capacity decay to generate a predicted remaining lifetime. The decoder network comprises multiple fully connected layers, with nonlinear activation functions introduced between layers to enable nonlinear mapping. Random deactivation layers are configured between hidden layers to enhance generalization performance. The parameters of the decoder network are obtained through training using a hybrid loss function that integrates monotonically decreasing capacity constraints and decay rate constraints. The monotonically decreasing capacity constraint requires that the predicted health state does not increase monotonically with the number of cycles, while the decay rate constraint requires that the predicted capacity decay rate be within a physically reasonable range.

[0051] Based on the aforementioned decoder network structure, this embodiment maintains the active state of the random deactivation layer during the prediction phase to perform Monte Carlo random deactivation inference. This embodiment performs multiple forward propagations on the same aging state representation vector. Each forward propagation produces a different prediction output due to the random shielding effect of the random deactivation layer, thus obtaining a sample set of predicted values. This embodiment calculates the sample variance of the sample set of predicted values ​​to obtain a prediction uncertainty measure. The prediction uncertainty measure reflects the model's confidence in the current prediction; a larger variance value indicates a lower model confidence. The remaining lifetime prediction value and the prediction uncertainty measure are read by subsequent step S103 for the calculation and construction of confidence intervals.

[0052] Step S103: Calculate the remaining lifetime confidence interval by combining the predicted remaining lifetime value and the prediction uncertainty measure according to a preset confidence level, and send the predicted remaining lifetime value and the remaining lifetime confidence interval to the UAV flight management system for battery safety margin determination.

[0053] In this embodiment, the predicted value sample set generated in step S102 is read, and the arithmetic mean of each predicted value in the set is calculated to obtain the predicted mean. The predicted mean is used as a point prediction output of the remaining lifespan, representing the expected number of remaining usable cycles of the battery in the current aging state.

[0054] Based on the predicted mean, this embodiment performs a square root operation on the prediction uncertainty metric generated in step S102 to obtain the prediction standard deviation. The prediction standard deviation quantifies the dispersion of the predicted value sampling set; a larger standard deviation indicates that the model has higher uncertainty in predicting the current input.

[0055] Accordingly, this embodiment constructs a confidence interval by querying the standard normal distribution quantiles corresponding to a preset confidence level. The preset confidence level is determined by the UAV operation and maintenance management strategy; the higher the confidence level, the larger the corresponding quantile, and the wider the generated confidence interval. In this embodiment, the upper and lower bounds of the remaining lifetime confidence interval are obtained by adding or subtracting the product of the quantile and the prediction standard deviation from the predicted mean. For example, when the quantile corresponding to the preset confidence level is a specific value, the lower bound of the confidence interval is equal to the predicted mean minus the product of the quantile and the prediction standard deviation, and the upper bound of the confidence interval is equal to the predicted mean plus the product of the quantile and the prediction standard deviation.

[0056] After the remaining lifetime confidence interval is calculated, this embodiment encapsulates the remaining lifetime prediction value and the remaining lifetime confidence interval into a prediction output message according to a preset message format. The message structure includes six fields: prediction timestamp, battery identifier, remaining lifetime prediction value, upper bound of the confidence interval, lower bound of the confidence interval, and prediction uncertainty measure. Each field is arranged in a fixed byte length and order for easy parsing by the receiving end.

[0057] Next, in this embodiment, the predicted output message is sent to the UAV flight management system via a data communication interface. The data communication interface uses the UAV onboard bus protocol to complete the message transmission, and waits for the flight management system to confirm receipt after transmission. The UAV flight management system parses the predicted output message and extracts two key data points—the predicted remaining lifetime value and the lower bound of the confidence interval—for safety margin determination.

[0058] Based on the extracted data, the UAV flight management system compares the lower bound of the confidence interval with a preset safety margin threshold. The safety margin threshold is configured according to the UAV mission type and flight environment risk level; a higher threshold is used for high-risk missions to increase safety redundancy. When the lower bound of the confidence interval falls below the safety margin threshold, the UAV flight management system generates a battery replacement warning signal and pushes it to the ground station maintenance terminal, prompting maintenance personnel to replace the battery pack before the next flight mission.

[0059] As described above, the deep learning-based method for predicting the remaining lifespan of UAV lithium batteries provided in this application can achieve accurate feature acquisition through cycle identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable lifespan assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively addresses the shortcomings of traditional technologies in feature extraction, lifespan prediction, and uncertainty assessment, providing technical support for UAV battery management.

[0060] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0061] Step S201: Read battery terminal voltage sampling data, charging and discharging current sampling data, and battery temperature sampling data from the data interface of the UAV battery management system. Perform outlier detection on the voltage sampling data, the current sampling data, and the temperature sampling data, and mark sampling points that exceed the preset physical reasonable range as invalid. Perform linear interpolation to complete the marked invalid sampling points to obtain a complete set of sampling data.

[0062] Step S202: Perform timestamp alignment on each type of data in the complete sampled data set according to a unified time base, perform high-frequency resampling on the voltage sampled data and the current sampled data and low-frequency resampling on the temperature sampled data according to the data type differences, and merge the resampled data of each type according to the timestamp index to obtain standardized cyclic data.

[0063] This embodiment establishes a communication connection with the onboard sensor network through the data interface of the UAV battery management system. It reads battery terminal voltage sampling data from the voltage sensor channel, charge / discharge current sampling data from the current sensor channel, and battery temperature sampling data from the temperature sensor channel. Each of the three types of sampling data carries independent timestamp information and sampling sequence number. The voltage sampling data records the terminal voltage change process of each cell in the battery pack, the current sampling data records the current amplitude and direction during charging and discharging, and the temperature sampling data records the temperature distribution on the battery surface and inside the battery.

[0064] After the voltage, current, and temperature sampling data are read, this embodiment performs outlier detection on each of the three types of data. The outlier detection module configures corresponding physical reasonable ranges for each type of data. The reasonable range for voltage sampling data is determined based on the cutoff voltage and full charge voltage of the battery cell; the reasonable range for current sampling data is determined based on the battery's maximum allowable charge / discharge rate; and the reasonable range for temperature sampling data is determined based on the battery's safe operating temperature range. In this embodiment, sampling points that exceed the corresponding physical reasonable range are marked as invalid, and these invalid marks serve as the basis for subsequent interpolation and completion processing.

[0065] Accordingly, this embodiment performs linear interpolation completion processing on the sampling points marked as invalid. The interpolation completion module locates the position of each invalid sampling point in the time series, finds the nearest valid sampling point before and after that position as the interpolation anchor point, calculates the linear interpolation coefficient based on the values ​​of the two anchor points and the time interval, and fills the invalid sampling point position with the interpolation result. When the duration of consecutive invalid sampling points exceeds a preset tolerance threshold, this embodiment marks the entire data segment as unusable to avoid interpolation distortion. After the interpolation completion processing is completed, this embodiment obtains a complete set of sampled data.

[0066] Based on the complete sampled data set, this embodiment performs timestamp alignment processing on each type of data according to a unified time reference. The timestamp alignment module selects the master clock signal of the battery management system as the unified time reference and converts the local timestamps of each sensor channel to this reference time coordinate system. The alignment process eliminates the time deviation between channels caused by clock drift and sampling delay, so that the voltage, current and temperature data at the same moment can be accurately matched.

[0067] Next, this embodiment configures the resampling frequency according to the differences in data types to balance data accuracy and storage overhead. Voltage and current sampling data undergo high-frequency resampling to capture transient changes during charging and discharging. High-frequency resampling uses interpolation to generate encrypted data points between the original sampling points. Temperature sampling data undergoes low-frequency resampling due to its large thermal inertia and relatively slow rate of change. Low-frequency resampling uses downsampling to extract data points from the original sampling points at fixed intervals.

[0068] Based on the output of the aforementioned resampling process, this embodiment merges the data of various types according to the timestamp index to obtain standardized cyclic data. The merging process uses the timestamp sequence generated by a unified time base as the index key, assembling the values ​​of each type of data at the corresponding timestamp position into a multi-dimensional data record. Each record of the standardized cyclic data includes four fields: timestamp, battery terminal voltage, charging / discharging current, and battery temperature. The standardized cyclic data is read by the cycle boundary identification module in subsequent step S301 as input data for charge / discharge cycle segmentation.

[0069] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0070] Step S301: Perform zero-crossing detection on the charging and discharging current in the standardized cyclic data, determine the charging start point based on the moment when the charging current changes from zero to positive, and determine the discharging end point based on the moment when the discharging current changes from negative to zero, and obtain a set of cyclic data segments based on the data segments between adjacent charging start points.

[0071] Step S302: Calculate the ratio of capacity change to voltage change for each cyclic data segment in the cyclic data segment set to obtain the incremental capacity curve, and extract the peak position and peak amplitude as incremental capacity features. Calculate the ratio of voltage change to capacity change to obtain the differential voltage curve, and extract the feature point position and feature segment slope as differential voltage features. Assemble the incremental capacity features and the differential voltage features into a health factor vector and organize them into a health factor time series according to the cyclic number.

[0072] In this embodiment, the standardized cyclic data generated in step S202 is read, and zero-crossing detection is performed on the charging and discharging current field to identify the switching moment of the charging and discharging state. The zero-crossing detection module traverses the current sampling sequence along the time axis and detects the change in the sign of the current value between adjacent sampling points. When the current at the previous sampling point is zero or negative and the current at the next sampling point is positive, the moment is determined to be the zero-crossing point where the charging current changes from zero to positive.

[0073] Based on the zero-crossing detection, this embodiment determines the charging start point according to the moment when the charging current changes from zero to positive. The charging start point marks the time boundary when the battery transitions from a resting or discharging state to a charging state, and this boundary serves as the starting segmentation position for the loop data segment. This embodiment also detects the moment when the discharging current changes from negative to zero and determines the discharging end point, which marks the time boundary when the battery transitions from a discharging state to a resting state.

[0074] Accordingly, this embodiment performs cyclic segmentation based on the data segments between adjacent charging start points. The cyclic segmentation module divides all data records between two adjacent charging start points on the time axis into a cyclic data segment. Each cyclic data segment contains one complete charging process, a possible resting process, and one complete discharging process. This embodiment assigns an incremental cyclic number to each cyclic data segment and organizes all cyclic data segments into a cyclic data segment set according to the cyclic number.

[0075] Based on the aforementioned cyclic data segment set, this embodiment performs incremental capacity analysis on each cyclic data segment to extract aging-sensitive features. The incremental capacity analysis module extracts the voltage and capacity sequences for the charging stage from the cyclic data segments. It calculates the difference between adjacent sampling points in the capacity sequence to obtain the capacity change, and calculates the difference between adjacent sampling points in the voltage sequence to obtain the voltage change. The ratio of the capacity change to the voltage change is arranged by voltage to obtain the incremental capacity curve. This embodiment extracts the peak position and peak amplitude from the incremental capacity curve as incremental capacity features. The peak position corresponds to the characteristic voltage of the phase transition reaction inside the battery, and the peak amplitude reflects the activity intensity of this phase transition reaction.

[0076] Next, this embodiment performs differential voltage analysis on each cyclic data segment in the cyclic data segment set. The differential voltage analysis module calculates the ratio of voltage change to capacity change to obtain a differential voltage curve. The shape of the differential voltage curve is sensitive to changes in battery internal resistance. This embodiment extracts the location of feature points and the slope of feature segments from the differential voltage curve as differential voltage features. The location of feature points marks the capacity coordinates corresponding to inflection points or extreme points on the curve, and the slope of feature segments quantifies the rate of change of a specific section of the curve.

[0077] Based on the outputs of the aforementioned incremental capacity analysis and differential voltage analysis, this embodiment assembles the incremental capacity features and the differential voltage features into a health factor vector. Each component of the health factor vector sequentially records four types of feature values: peak position, peak amplitude, feature point position, and feature segment slope. The order of the components is consistent with the output order of the feature extraction module. In this embodiment, the health factor vectors corresponding to each cyclic data segment are organized into a health factor time series according to the cyclic number. This health factor time series is read by the normalization module in subsequent step S401 and used as input data for the long short-term memory network time-series encoding.

[0078] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0079] Step S401: Perform zero-mean unit variance standardization on each health factor vector in the health factor time series according to the training set statistics to obtain a normalized health factor sequence. Input the normalized health factor sequence into the stacked long short-term memory network layer, process it step by step according to time steps, and output the hidden state of each time step to obtain the hidden state sequence.

[0080] Step S402: Perform a linear transformation of the query vector and the key vector on the hidden states at each time step in the hidden state sequence and calculate the dot product to obtain the attention score matrix. Perform scaling and normalization on the attention score matrix to obtain the attention weights. Perform weighted aggregation on the hidden state sequence based on the attention weights to obtain the aging state representation vector.

[0081] This embodiment reads the health factor time series generated in step S302 above, and performs zero-mean, unit-variance standardization on each health factor vector in the sequence to eliminate dimensional differences between different feature components. During the offline training phase, the normalization module traverses all health factor vectors in the training dataset, calculates the mean and standard deviation for each feature component, and stores them. During the online prediction phase, the normalization module performs a component-by-component transformation on the input health factor vector based on the stored training set statistics, subtracting the corresponding mean from each component value and dividing by the corresponding standard deviation, thereby obtaining the normalized health factor sequence.

[0082] Based on the normalized health factor sequence, this embodiment inputs it into a stacked long short-term memory (LSTM) network layer for temporal encoding. The stacked LSM network layer consists of multiple layers of sequentially connected LSM units. The first LSM unit receives the normalized health factor sequence as input and processes the health factor vectors corresponding to each cycle sequentially according to time steps. During the processing at each time step, the LSM unit updates the gating signal and cell state based on the current input and the hidden state of the previous time step, outputting the hidden state vector of the current time step. The second LSM unit uses the hidden state sequence output by the first layer as input for higher-level temporal pattern abstraction. In this embodiment, the hidden state vectors output by the last LSM unit at each time step are organized into a hidden state sequence in chronological order.

[0083] Accordingly, this embodiment performs attention weight calculation on the hidden state sequence to measure the contribution of each historical loop to the current prediction. The attention module configures query transformation matrix and key transformation matrix for the hidden state vector at each time step in the hidden state sequence, and maps the hidden state vector to query vector and key vector through matrix multiplication. In this embodiment, the dot product of the query vector and key vector is calculated to obtain the attention score matrix, and the element values ​​of the attention score matrix reflect the correlation strength between corresponding time steps.

[0084] Based on the attention score matrix, this embodiment performs a scaling process to stabilize the numerical range of subsequent normalization operations. The scaling process divides each element of the attention score matrix by the square root of the key vector dimension to prevent the dot product result from having excessively large values ​​due to high dimensionality, which could lead to gradient vanishing after normalization. This embodiment performs row-wise normalization on the scaled attention score matrix, ensuring that the sum of each row's elements is one, thereby obtaining the attention weights.

[0085] Based on the attention weights, this embodiment performs weighted aggregation on the hidden state sequence to generate an aging state representation vector. The weighted aggregation module multiplies the hidden state vector at each time step in the hidden state sequence by the corresponding attention weight, and then sums them along the time dimension to obtain the aggregation result. Time steps with larger attention weights contribute more significantly to the aggregation result. This mechanism enables the model to adaptively focus on historical cycles that have a significant impact on the remaining lifetime prediction. The aging state representation vector encodes the cumulative aging information of the battery from the first cycle to the current cycle, which is read by the decoder network in the subsequent step S501 to generate the remaining lifetime prediction value.

[0086] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0087] Step S501: Input the aging state characterization vector into the multi-layer fully connected layer of the decoder network for nonlinear mapping. The decoder network performs forward propagation based on the network parameters trained according to the hybrid loss function of fusion capacity monotonically decreasing constraint and decay rate constraint to obtain the remaining lifetime prediction value.

[0088] Step S502: During the prediction phase, maintain the random deactivation layer activation state of the decoder network, perform multiple forward propagations on the same aging state representation vector to obtain a set of predicted values, and calculate the sample variance of the set of predicted values ​​to obtain a prediction uncertainty measure.

[0089] In this embodiment, the aging state representation vector generated in step S402 is read and input into the multi-layer fully connected layers of the decoder network for nonlinear mapping. The first fully connected layer of the decoder network receives the aging state representation vector as input, applies a nonlinear activation function after a linear transformation of the weight matrix and bias vector, and outputs the first layer hidden representation. Subsequent fully connected layers sequentially perform the same linear transformation and nonlinear activation processing on the output of the previous layer, extracting higher-order features related to the remaining lifetime layer by layer. The output layer of the decoder network maps the last layer hidden representation to a single numerical value, which is the predicted remaining lifetime value.

[0090] In the structure of the decoder network, random deactivation layers are configured between each hidden layer to enhance the model's generalization ability. During training, these random deactivation layers randomly mask the outputs of some neurons with a preset probability, forcing the network to learn more robust feature representations. The parameters of the decoder network are obtained through training using a hybrid loss function that integrates a capacity monotonically decreasing constraint and a decay rate constraint. This hybrid loss function comprises two components: a data fitting term and a physical constraint term.

[0091] Accordingly, this embodiment describes the composition of the hybrid loss function. The data fitting term measures the deviation between the predicted value and the actual remaining lifetime label, calculated using the square of the prediction error. The physical constraint term includes a capacity monotonically decreasing constraint and a decay rate constraint. The capacity monotonically decreasing constraint penalizes cases where an upward trend appears in the predicted health state sequence, while the decay rate constraint penalizes cases where the predicted capacity decay rate exceeds a physically reasonable range. The hybrid loss function is a weighted sum of the data fitting term and the physical constraint term using preset weight coefficients. During training, the network parameters are updated by minimizing this loss function.

[0092] After the decoder network is trained, this embodiment maintains the activation state of the random deactivation layer during the prediction phase to perform Monte Carlo random deactivation inference. Unlike the conventional approach of disabling the random deactivation layer during the inference phase, Monte Carlo random deactivation inference utilizes the random masking characteristic of the random deactivation layer to sample and estimate model uncertainty. This embodiment performs multiple forward propagations on the same aging state representation vector, with each forward propagation producing a different prediction output due to the different locations of neurons masked by the random deactivation layer.

[0093] Based on the outputs of the aforementioned multiple forward propagations, this embodiment organizes the prediction results into a predicted value sampling set. The size of the predicted value sampling set is determined by the number of forward propagation executions; the more executions, the more fully the sampling set estimates the prediction distribution. This embodiment calculates the sample variance of the predicted value sampling set to obtain a prediction uncertainty measure. The sample variance is calculated by dividing the sum of squared deviations between the sampled values ​​and the sampled mean by the number of samples minus one. The prediction uncertainty measure quantifies the model's confidence in the current prediction; a larger variance value indicates a lower model confidence. The remaining lifetime prediction value and the prediction uncertainty measure are read by subsequent step S601 for the calculation and construction of confidence intervals.

[0094] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0095] Step S601: Calculate the arithmetic mean of the sample set of predicted values ​​to obtain the predicted mean, and perform a square root operation on the predicted uncertainty measure to obtain the predicted standard deviation;

[0096] Step S602: Based on the preset confidence level, query the corresponding standard normal distribution quantile, and add or subtract the product of the quantile and the predicted standard deviation to obtain the remaining life confidence interval.

[0097] In this embodiment, the predicted value sampling set generated in step S502 is read, and an arithmetic mean operation is performed on each predicted value in the set to obtain the predicted mean. The arithmetic mean operation sums all the sampled values ​​in the predicted value sampling set and divides by the number of samplings. The predicted mean is used as a point prediction output of the remaining lifespan, representing the expected number of usable cycles remaining in the current aging state of the battery.

[0098] Based on the predicted mean, this embodiment performs a square root operation on the prediction uncertainty metric generated in step S502 to obtain the prediction standard deviation. The prediction uncertainty metric is the sample variance of the predicted value sampling set. Performing a square root operation on it converts the variance into a standard deviation form with the same dimensions as the predicted values. The prediction standard deviation quantifies the degree of dispersion of the predicted value sampling set around the prediction mean. A larger standard deviation value indicates that the predicted outputs of each forward propagation are significantly different, and the model's prediction of the current input has high uncertainty.

[0099] Accordingly, this embodiment constructs a confidence interval by querying the corresponding standard normal distribution quantiles based on a preset confidence level. The preset confidence level is determined by the UAV operation and maintenance management strategy and represents the probability target of the confidence interval covering the actual remaining lifespan. There is a definite mapping relationship between the confidence level and the standard normal distribution quantiles; the higher the confidence level, the larger the absolute value of the corresponding quantile. This embodiment retrieves the quantile value corresponding to the preset confidence level from a pre-stored quantile lookup table.

[0100] After the quantile query is completed, this embodiment calculates the remaining lifetime confidence interval by adding or subtracting the product of the quantile and the predicted standard deviation from the predicted mean. The lower bound of the confidence interval is equal to the predicted mean minus the product of the quantile and the predicted standard deviation, and the upper bound is equal to the predicted mean plus the product of the quantile and the predicted standard deviation. The width of the remaining lifetime confidence interval is determined by both the predicted standard deviation and the quantile; when the prediction uncertainty is high, the confidence interval widens accordingly to reflect the range of confidence in the prediction results.

[0101] Based on the aforementioned confidence interval construction logic, this embodiment explains the physical meaning of the confidence interval. The remaining life confidence interval represents the probability that the battery's actual remaining life falls within this interval under a preset confidence level. The lower bound of the confidence interval is a conservative estimate, and the upper bound is an optimistic estimate. The predicted mean, the predicted standard deviation, and the remaining life confidence interval are read by subsequent step S701 for the encapsulation of the predicted output message and the determination of the battery safety margin.

[0102] In one embodiment of the deep learning-based method for predicting the remaining lifespan of drone lithium batteries in this application, the method may further include the following:

[0103] Step S701: Encapsulate the remaining lifetime prediction value and the remaining lifetime confidence interval into a prediction output message according to a preset message format, and send the prediction output message to the UAV flight management system through the data communication interface;

[0104] Step S702: The UAV flight management system parses the predicted output message to extract the remaining life prediction value and the lower bound of the confidence interval, compares the lower bound of the confidence interval with the preset safety margin threshold, and generates a battery replacement warning signal when the lower bound of the confidence interval is lower than the safety margin threshold.

[0105] This embodiment reads the predicted mean and remaining lifetime confidence interval generated in step S602, and encapsulates the remaining lifetime predicted value and remaining lifetime confidence interval into a prediction output message according to a preset message format. The message encapsulation module writes five fields sequentially: prediction timestamp, battery identifier, remaining lifetime predicted value, upper bound of the confidence interval, and lower bound of the confidence interval, according to the field structure and byte order rules defined in the preset message format. The prediction timestamp records the time of this prediction execution, the battery identifier distinguishes different battery packs carried by the drone, and each numerical field is encoded with a fixed byte length for location and parsing by the receiving end.

[0106] After the predicted output message is encapsulated, this embodiment sends it to the UAV flight management system via a data communication interface. The data communication interface establishes a communication link with the flight management system using the UAV onboard bus protocol, and the communication link supports bidirectional data transmission and a transmission acknowledgment mechanism. This embodiment writes the predicted output message into the transmission buffer and triggers a transmission request. The message transmission process is completed after the flight management system returns a receipt acknowledgment.

[0107] Accordingly, the UAV flight management system receives the predicted output message and performs parsing processing. The message parsing module extracts the values ​​of each field from the message data sequentially according to the field definitions of the preset message format, based on byte offsets. This embodiment focuses on extracting two key data points for subsequent safety margin determination: the predicted remaining lifetime value and the lower bound of the confidence interval. The predicted remaining lifetime value reflects the expected number of remaining usable cycles of the battery, and the lower bound of the confidence interval reflects a conservative estimate of the remaining lifetime at a preset confidence level.

[0108] After the remaining lifespan prediction value and the lower bound of the confidence interval are extracted, the UAV flight management system compares the lower bound of the confidence interval with a preset safety margin threshold. The safety margin threshold is configured differently according to the UAV mission type and the risk level of the flight environment. A higher threshold is used to increase safety redundancy when performing long-distance route missions or high-risk environment missions, while a relatively lower threshold is used to extend battery life when performing short-distance inspection missions or low-risk environment missions.

[0109] Based on the aforementioned comparison results, when the lower bound of the confidence interval falls below the safety margin threshold, the UAV flight management system generates a battery replacement warning signal. This warning signal carries three pieces of information: battery identifier, current remaining life prediction, and the lower bound of the confidence interval, and is pushed to the ground station maintenance terminal via the airborne data link. Upon receiving the warning signal, the ground station maintenance terminal displays a warning prompt on its interface, prompting maintenance personnel to replace or perform a deep inspection of the battery pack before the next flight mission to avoid flight mission interruptions or safety incidents due to insufficient battery performance.

[0110] To effectively address the shortcomings of traditional technologies in feature extraction, lifetime prediction, and uncertainty assessment, and to provide technical support for drone battery management, this application provides an embodiment of a deep learning-based drone lithium battery remaining lifetime prediction device for implementing all or part of the aforementioned deep learning-based drone lithium battery remaining lifetime prediction method. See [link to embodiment]. Figure 2 The deep learning-based drone lithium battery remaining life prediction device specifically includes the following components:

[0111] The battery data acquisition module 10 is used to acquire the original operating data of the charge and discharge cycle from the UAV battery management system and perform timestamp alignment and resampling to obtain standardized cycle data. Based on the zero-crossing point of the charge and discharge current, the standardized cycle data is used to perform cycle boundary identification to obtain a set of cycle data segments. Incremental capacity analysis and differential voltage analysis are performed on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number.

[0112] The battery life prediction module 20 is used to normalize the health factor time series and input it into a long short-term memory network for time-series encoding to obtain a hidden state sequence, perform attention weight calculation and weighted aggregation on the hidden state sequence to obtain an aging state representation vector, input the aging state representation vector into a decoder network that integrates capacity decay physical constraints to obtain a remaining life prediction value, and perform multiple Monte Carlo random inactivation forward propagation on the decoder network and statistically output the variance to obtain a prediction uncertainty measure.

[0113] The battery data feedback module 30 is used to calculate the remaining life confidence interval based on the predicted remaining life value and the prediction uncertainty measure according to a preset confidence level, and send the predicted remaining life value and the remaining life confidence interval to the UAV flight management system for battery safety margin determination.

[0114] As described above, the deep learning-based drone lithium battery remaining life prediction device provided in this application can achieve accurate feature acquisition through cycle identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable life assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively addresses the shortcomings of traditional technologies in feature extraction, life prediction, and uncertainty assessment, providing technical support for drone battery management.

[0115] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the deep learning-based method for predicting the remaining life of a drone lithium battery.

[0116] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned deep learning-based method for predicting the remaining lifespan of a drone's lithium battery.

[0117] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the aforementioned deep learning-based method for predicting the remaining lifespan of a drone's lithium battery.

[0118] In this embodiment of the invention, accurate feature acquisition is achieved through cyclic identification and health factor extraction. A prediction mechanism is constructed, combining temporal coding and physical constraints to establish a reliable lifetime assessment strategy. Uncertainty optimization is introduced, using random deactivation and confidence intervals to ensure continuous improvement in prediction. This method effectively addresses the shortcomings of traditional technologies in feature extraction, lifetime prediction, and uncertainty assessment, providing technical support for UAV battery management.

[0119] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining lifespan of a drone's lithium battery based on deep learning, characterized in that, The method includes: The raw operation data of charge and discharge cycles are obtained from the UAV battery management system and the timestamp alignment and resampling are performed to obtain standardized cycle data. The cycle data is then used to identify the cycle boundary based on the zero crossing point of the charge and discharge current to obtain a set of cycle data segments. Incremental capacity analysis and differential voltage analysis are performed on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number. After normalizing the time series of health factors, the data is input into a long short-term memory network for time-series encoding to obtain a hidden state sequence. Attention weights are calculated and weighted aggregation is performed on the hidden state sequence to obtain an aging state representation vector. The aging state representation vector is input into a decoder network that incorporates physical constraints on capacity decay to obtain a predicted remaining lifespan. Multiple Monte Carlo random inactivation forward propagation is performed on the decoder network and the variance is statistically output to obtain a measurement of prediction uncertainty. The remaining lifetime prediction value and the prediction uncertainty measure are used to calculate the remaining lifetime confidence interval based on a preset confidence level. The remaining lifetime prediction value and the remaining lifetime confidence interval are then sent to the UAV flight management system for battery safety margin determination.

2. The method for predicting the remaining lifespan of a drone lithium battery based on deep learning according to claim 1, characterized in that, The process of obtaining raw charge / discharge cycle data from the UAV battery management system and performing timestamp alignment and resampling to obtain standardized cycle data includes: The system reads battery terminal voltage sampling data, charging and discharging current sampling data, and battery temperature sampling data from the data interface of the drone battery management system. It performs outlier detection on the voltage sampling data, the current sampling data, and the temperature sampling data and marks sampling points that exceed the preset physical reasonable range as invalid. It then performs linear interpolation to complete the marked invalid sampling points to obtain a complete set of sampling data. The data of each type in the complete sampled data set is timestamped according to a unified time base. The resampling frequency is configured according to the data type differences. High-frequency resampling is performed on the voltage sampled data and the current sampled data, and low-frequency resampling is performed on the temperature sampled data. The resampled data of each type are merged according to the timestamp index to obtain standardized cyclic data.

3. The method for predicting the remaining lifespan of UAV lithium batteries based on deep learning according to claim 1, characterized in that, The standardized cyclic data is used to perform cyclic boundary identification based on the zero-crossing point of the charging and discharging current to obtain a set of cyclic data segments. Incremental capacity analysis and differential voltage analysis are then performed on the cyclic data segment set to extract health factor vectors, which are then organized into a health factor time series according to the cyclic number, including: Zero-crossing detection is performed on the charging and discharging currents in the standardized cyclic data. The charging start point is determined based on the moment when the charging current changes from zero to positive, and the discharging end point is determined based on the moment when the discharging current changes from negative to zero. A set of cyclic data segments is obtained by dividing the data segments between adjacent charging start points. For each cyclic data segment in the cyclic data segment set, the ratio of capacity change to voltage change is calculated to obtain the incremental capacity curve, and the peak position and peak amplitude are extracted as incremental capacity features. The ratio of voltage change to capacity change is calculated to obtain the differential voltage curve, and the feature point position and feature segment slope are extracted as differential voltage features. The incremental capacity features and the differential voltage features are assembled into a health factor vector and organized into a health factor time series according to the cyclic number.

4. The method for predicting the remaining lifespan of a drone lithium battery based on deep learning according to claim 1, characterized in that, The process involves normalizing the time-series sequence of health factors and then inputting it into a long short-term memory network for temporal encoding to obtain a hidden state sequence. Attention weights are then calculated and weighted aggregated on the hidden state sequence to obtain an aging state representation vector, including: The normalized health factor sequence is obtained by performing zero-mean unit variance standardization on each health factor vector in the health factor time series based on the training set statistics. The normalized health factor sequence is then input into a stacked long short-term memory network layer, processed step by step, and the hidden state of each time step is output to obtain the hidden state sequence. A linear transformation of the query vector and key vector is performed on the hidden states at each time step in the hidden state sequence, and the dot product is calculated to obtain the attention score matrix. The attention score matrix is ​​then scaled and normalized to obtain the attention weights. Based on the attention weights, the hidden state sequence is weighted and aggregated to obtain the aging state representation vector.

5. The method for predicting the remaining lifespan of a drone lithium battery based on deep learning according to claim 1, characterized in that, The process involves inputting the aging state characterization vector into a decoder network that incorporates physical constraints on capacity decay to obtain a predicted remaining lifetime. Multiple Monte Carlo random inactivation forward propagations are then performed on the decoder network, and the variance is statistically analyzed to obtain a measurement of prediction uncertainty. This includes: The aging state characterization vector is input into the multi-layer fully connected layer of the decoder network for nonlinear mapping. The decoder network performs forward propagation based on the network parameters trained according to the hybrid loss function of fusion capacity monotonically decreasing constraint and decay rate constraint to obtain the remaining lifetime prediction value. During the prediction phase, the random deactivated layer activation state of the decoder network is maintained, and multiple forward propagations are performed on the same aging state representation vector to obtain a set of predicted values. The sample variance of the set of predicted values ​​is calculated to obtain a measure of prediction uncertainty.

6. The method for predicting the remaining lifespan of a drone lithium battery based on deep learning according to claim 1, characterized in that, The step of calculating the remaining lifetime confidence interval by combining the predicted remaining lifetime value with the prediction uncertainty measure based on a preset confidence level includes: The arithmetic mean of the sampled set of predicted values ​​is calculated to obtain the predicted mean, and the square root operation is performed on the predicted uncertainty measure to obtain the predicted standard deviation. Based on the preset confidence level, the corresponding standard normal distribution quantile is queried, and the remaining life confidence interval is obtained by adding or subtracting the product of the quantile and the predicted standard deviation from the predicted mean.

7. The method for predicting the remaining lifespan of a drone lithium battery based on deep learning according to claim 1, characterized in that, The step of sending the predicted remaining lifetime value and the remaining lifetime confidence interval to the UAV flight management system for battery safety margin determination includes: The remaining lifetime prediction value and the remaining lifetime confidence interval are encapsulated into a prediction output message according to a preset message format, and the prediction output message is sent to the UAV flight management system through the data communication interface. The UAV flight management system parses the predicted output message to extract the remaining life prediction value and the lower bound of the confidence interval. It then compares the lower bound of the confidence interval with a preset safety margin threshold. When the lower bound of the confidence interval is lower than the safety margin threshold, a battery replacement warning signal is generated.

8. A deep learning-based device for predicting the remaining lifespan of a drone's lithium battery, characterized in that, The device includes: The battery data acquisition module is used to obtain the raw operating data of the charge and discharge cycle from the UAV battery management system and perform timestamp alignment and resampling to obtain standardized cycle data. Based on the zero-crossing point of the charge and discharge current, the standardized cycle data is used to perform cycle boundary identification to obtain a set of cycle data segments. Incremental capacity analysis and differential voltage analysis are performed on the set of cycle data segments to extract health factor vectors and organize them into a health factor time series according to the cycle number. The battery life prediction module is used to normalize the time series of health factors and input it into a long short-term memory network for time-series encoding to obtain a hidden state sequence. Attention weights are calculated and weighted aggregation is performed on the hidden state sequence to obtain an aging state representation vector. The aging state representation vector is input into a decoder network that integrates physical constraints of capacity decay to obtain a remaining life prediction value. Multiple Monte Carlo random inactivation forward propagation is performed on the decoder network and the variance is statistically output to obtain a prediction uncertainty measure. The battery data feedback module is used to calculate the remaining life confidence interval based on the predicted remaining life value and the predicted uncertainty measure according to a preset confidence level, and send the predicted remaining life value and the remaining life confidence interval to the UAV flight management system for battery safety margin determination.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the deep learning-based method for predicting the remaining lifespan of a drone lithium battery as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the deep learning-based method for predicting the remaining life of a drone lithium battery as described in any one of claims 1 to 7.

Images