Unmanned aerial vehicle battery life prediction method based on monotonic feature and adaptive dtw
By constructing a multi-channel feature matrix and using an adaptive DTW algorithm, the matching distortion problem caused by non-monotonic features and fixed penalty boundaries in UAV battery life prediction was solved, achieving high-precision remaining life prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU INTERNATIONAL INNOVATION INSTITUTE OF BEIHANG UNIVERSITY
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-07
Smart Images
Figure CN122063440B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery life prediction technology, specifically a method for predicting the battery life of drones based on monotonic features and adaptive DTW. Background Technology
[0002] The safe operation of a drone's power system highly depends on the stability of its battery's electrical performance. During battery charging and discharging, continuous measurement and signal acquisition of electrical variables such as voltage and discharge capacity can effectively assess the battery's internal health. Remaining life prediction is based on these measured electrical observation data and historical degradation patterns to infer the number of charge-discharge cycles remaining before the battery's electrical performance degrades to the failure threshold. However, the raw electrical characteristics collected by sensors are often accompanied by significant environmental noise and local fluctuations, masking the irreversible, monotonically decreasing nature of battery physical damage. Therefore, extracting strictly monotonically decreasing degradation characteristics and combining them with adaptive dynamic time warping (adaptive DTW) technology, which can dynamically sense local degradation rates, has become a key direction for achieving high-precision trajectory matching and remaining life prediction. Accurately acquiring the electrical characteristics of drone batteries and predicting their remaining life is fundamental to preventing equipment power depletion and developing scientific maintenance strategies.
[0003] Currently, trajectory similarity-based matching algorithms are commonly used to infer battery life. However, due to the highly nonlinear characteristics of real battery degradation, conventional distance metrics cannot handle timeline misalignments between different devices. Dynamic Time Warping (DTW) technology has thus been introduced to allow for local timeline scaling. To further address alignment distortion at the end of degradation, existing techniques typically attempt to introduce improved DTW methods with nonlinear penalty mechanisms to optimize the model's similarity assessment accuracy.
[0004] However, existing lifetime prediction methods have significant shortcomings: directly using non-monotonic features containing local bounce and noise for matching violates the physical principle that internal battery damage is irreversible, and is highly susceptible to causing serious interference with distance calculations; secondly, existing improved DTW methods heavily rely on globally fixed penalty boundary mechanisms. When the feature decay trajectory of the device is relatively smooth, the small fixed penalty is difficult to activate, leading to algorithm degradation and over-regulation. This computational mechanism easily misaligns the early health state of the device with the final degradation stage, resulting in severely overestimated lifetime predictions.
[0005] Therefore, this invention proposes a method for predicting the battery life of drones based on monotonic features and adaptive DTW to address the shortcomings of existing technologies. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting the battery life of drones based on monotonic features and adaptive DTW. This method solves the problems caused by existing methods directly using non-monotonic features that contain local fluctuations, and by the over-regulation of existing DTW algorithms due to their reliance on fixed penalty boundaries. These problems result in battery degradation feature alignment distortion and severely overestimated remaining life prediction results.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the battery life of a drone based on monotonic features and adaptive DTW, comprising the following steps:
[0008] The voltage time series, discharge capacity time series, and temperature time series of the drone battery are obtained and preprocessed to obtain a multi-channel feature matrix;
[0009] The multi-channel feature matrix is input into a one-dimensional convolutional autoencoder, which outputs a latent feature sequence.
[0010] Calculate the local absolute gradient of the latent feature sequence and accumulate it to construct an exponential damage accumulation. Perform analytical boundary normalization and flip truncation on the exponential damage accumulation to obtain a monotonically decreasing health factor sequence.
[0011] A local acceleration factor is calculated based on the monotonically decreasing health factor sequence, and a nonlinear penalty upper limit is dynamically generated based on the local acceleration factor.
[0012] The state transition cost between the test equipment and the historical trajectory is calculated using the nonlinear penalty upper limit and spherical distance method, and the shortest matching distance is found through the cumulative cost matrix.
[0013] Extract the remaining lifetime of the historical trajectory corresponding to the shortest matching distance, and output the predicted remaining lifetime based on the remaining lifetime of the historical trajectory.
[0014] Preferably, the step of obtaining and preprocessing the voltage time series, discharge capacity time series, and temperature time series of the UAV battery to obtain a multi-channel feature matrix includes:
[0015] The voltage time series, discharge capacity time series, and temperature time series of the UAV battery are subjected to linear interpolation and resampling processing to be unified to a fixed length;
[0016] Based on the physical operating limits, the resampled voltage time series and the resampled temperature time series are truncated and mapped to obtain the truncated and mapped voltage time series and the truncated and mapped temperature time series.
[0017] The discharge capacity time series after resampling is normalized by using the factory rated capacity to obtain the normalized discharge capacity time series.
[0018] The voltage time series after truncation mapping, the temperature time series after truncation mapping, and the discharge capacity time series after per-unit processing are spliced and aligned according to the channel dimension to obtain the multi-channel feature matrix.
[0019] Preferably, the step of inputting the multi-channel feature matrix into a one-dimensional convolutional autoencoder and outputting a latent feature sequence includes:
[0020] The multi-channel feature matrix is input into the encoder of the one-dimensional convolutional autoencoder, and after one-dimensional convolution operation, flattening layer and fully connected layer, scalar latent features are output.
[0021] The multi-channel feature matrices generated by the drone battery under multiple consecutive charge-discharge cycles are sequentially input into the encoder, and the corresponding scalar latent features are output sequentially.
[0022] The scalar latent features are spliced and combined in sequence to obtain the latent feature sequence.
[0023] Preferably, the step of calculating the local absolute gradient of the latent feature sequence and accumulating it to construct the exponential damage accumulation includes:
[0024] Calculate the absolute value of the change in the potential feature sequence at adjacent charge-discharge cycle times to obtain the local absolute gradient;
[0025] The global maximum cumulative gradient limit, constrained by a safety margin, is calculated using a pre-established training set.
[0026] The exponential damage accumulation is constructed using the local absolute gradient and the global maximum cumulative gradient limit.
[0027] Preferably, the step of performing analytical boundary normalization and flipping truncation on the accumulated exponential damage to obtain a monotonically decreasing health factor sequence includes:
[0028] The accumulated exponential damage is normalized using a purely mathematical analytical boundary to obtain the normalized feature quantity.
[0029] The normalized feature values are subjected to a flip mapping and interval truncation to obtain the health factor;
[0030] The health factors at different times are combined in chronological order to obtain the monotonically decreasing health factor sequence.
[0031] Preferably, the step of calculating a local acceleration factor based on the monotonically decreasing health factor sequence and dynamically generating a nonlinear penalty upper limit based on the local acceleration factor includes:
[0032] Set a final time window for observation, and extract the average rate of change of the monotonically decreasing health factor sequence within the final time window as the recent local degradation rate;
[0033] The rate of change of the values at the beginning and end of the monotonically decreasing health factor sequence is calculated as the global average degradation rate.
[0034] The local acceleration factor is obtained by dividing the recent local degradation rate by the global average degradation rate and adding a very small normal number.
[0035] The local acceleration factor is linearly mapped and truncated with a threshold to obtain the upper limit of the nonlinear penalty.
[0036] Preferably, the step of calculating the state transition cost of the test equipment and historical trajectory using the nonlinear penalty upper limit and spherical distance method includes:
[0037] The monotonically decreasing health factor sequence extracted from the testing device is set as the test sequence, and the historical trajectory is extracted from the pre-established training library as the reference sequence.
[0038] Initialize the cumulative cost matrix and search constraint window;
[0039] A linear interpolation array is established using the aforementioned nonlinear penalty upper limit, and the nonlinear penalty weight is calculated using the aforementioned spherical distance method.
[0040] Within the search constraint window, the weighted local cost between the test sequence and the reference sequence is calculated based on the nonlinear penalty weight as the state transition cost.
[0041] Preferably, the step of finding the shortest matching distance through the cumulative cost matrix includes:
[0042] Using the weighted local cost, dynamic programming state transition equations are updated in the cumulative cost matrix to complete the numerical filling of the cumulative cost matrix;
[0043] The shortest matching distance is obtained by extracting the last element of the cumulative cost matrix and performing a square root operation.
[0044] Preferably, the step of extracting the remaining lifetime of the historical trajectory corresponding to the shortest matching distance includes:
[0045] The sliding window mechanism is used to traverse and scan all the historical trajectories in the pre-established training library;
[0046] The reference subsequence is extracted using the sliding window mechanism, and the matching distance between the monotonically decreasing health factor sequence and the reference subsequence is calculated.
[0047] The shortest matching distance is selected as the distance with the smallest value among all the matching distances, and the end position of the sliding window corresponding to the shortest matching distance is defined as the matching endpoint.
[0048] The remaining lifespan of the historical trajectory is obtained by calculating the step difference between the matching endpoint and the actual lifespan termination point.
[0049] Preferably, the step of predicting the remaining lifetime based on the remaining lifetime output of the historical trajectory includes:
[0050] Construct a global candidate set that includes multiple shortest matching distances and the corresponding remaining lifetime of the historical trajectories;
[0051] The global candidate set is sorted in ascending order according to the numerical value of the shortest matching distance, and the two shortest matching distances with the smallest values and the corresponding remaining lifetime of the historical trajectory are extracted.
[0052] The reciprocal of the distance between the two shortest matching distances with the smallest values is used as a weight, and the remaining lifetime of the corresponding historical trajectories is weighted and summed to output the predicted remaining lifetime.
[0053] This invention provides a method for predicting the battery life of unmanned aerial vehicles (UAVs) based on monotonic features and adaptive DTW (Dynamic Time-Depth Wave). It has the following beneficial effects:
[0054] 1. This invention addresses the degradation status assessment of drone batteries. To achieve highly reliable remaining lifespan prediction, a data processing framework based on monotonic feature extraction is constructed. This invention extracts a strictly monotonically decreasing sequence of health factors through a one-dimensional convolutional autoencoder and absolute gradient accumulation. This process effectively filters out random noise from multivariate sensors, aligns with the irreversible physical nature of equipment damage, and eliminates the computational interference of local feature fluctuations on subsequent similarity matching algorithms.
[0055] 2. This invention addresses the timeline misalignment problem in UAV battery operation by utilizing an adaptive DTW (Time-Driven Weathering) dynamic sensing mechanism to improve the accuracy of remaining battery life prediction. This invention constructs a local acceleration factor by comparing the recent local degradation rate with the global average degradation rate, thereby dynamically generating a nonlinear penalty upper limit. This mechanism overcomes the limitation of traditional algorithms that heavily rely on fixed penalty boundaries, enabling adaptive warping of the device's accelerated decay period and preventing matching distortion during feature alignment.
[0056] 3. To avoid the overestimation of UAV battery life in asymmetric risk assessment, this invention deeply integrates the anti-disturbance advantage of monotonic features with the optimization mechanism of adaptive DTW (Dynamic Time-to-Wave) to output robust remaining lifetime prediction results. When calculating the cumulative cost matrix of the test equipment and historical trajectories, the pure and stable monotonic features ensure accurate calculation of state transition costs. Based on this, this invention combines nonlinear penalty weights to extract historical experience corresponding to the shortest matching distance for weighted fusion. This method avoids the over-regularization problem of smooth trajectories in traditional similarity algorithms, effectively improving the engineering reliability of modern equipment health management. Attached Figure Description
[0057] Figure 1 The flowchart of the UAV battery life prediction method based on monotonic features and adaptive DTW of the present invention is shown below.
[0058] Figure 2 This is a graph showing the degradation of the original capacity of the battery according to the present invention.
[0059] Figure 3 This is a schematic diagram of the health factor trajectory extracted based on the one-dimensional convolutional autoencoder of the present invention;
[0060] Figure 4 The image shows the predicted remaining lifetime fitting trajectory of the test battery Cell3 of this invention at different prediction starting points.
[0061] Figure 5 The following is a comparison chart of the multi-stage prediction performance of the present invention: (a) shows the MAE error trend, and (b) shows the NASA score trend. Detailed Implementation
[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0063] See attached document Figure 1 This invention provides a method for predicting the battery life of drones based on monotonic features and adaptive DTW, comprising the following steps:
[0064] S10: Obtain the voltage time series, discharge capacity time series, and temperature time series of the drone battery and preprocess them to obtain a multi-channel feature matrix;
[0065] S20 inputs the multi-channel feature matrix into a one-dimensional convolutional autoencoder and outputs a latent feature sequence.
[0066] S30, calculate the local absolute gradient of the latent feature sequence and accumulate it to construct the exponential damage accumulation. Perform analytical boundary normalization and flip truncation on the exponential damage accumulation to obtain the monotonically decreasing health factor sequence.
[0067] S40, calculate the local acceleration factor based on the health factor sequence, and dynamically generate the nonlinear penalty upper limit based on the local acceleration factor;
[0068] S50 calculates the state transition cost between the test equipment and the historical trajectory using the nonlinear penalty upper limit and spherical distance method, and finds the shortest matching distance through the cumulative cost matrix;
[0069] S60 extracts the remaining lifetime of the historical trajectory corresponding to the shortest matching distance, and outputs the predicted remaining lifetime based on the remaining lifetime of the historical trajectory.
[0070] See attached document Figure 1 In this embodiment, step S10 is specifically executed through the following sub-steps:
[0071] S101, Acquire multivariate sensor data during the operation of the drone battery. Data-driven methods that integrate prior physical knowledge have become a mainstream trend in prediction and health management. According to the physical degradation law of lithium batteries, their usable capacity typically exhibits a nonlinear exponential decay characteristic in the later stages of their lifespan. For the multivariate time-series sensor data during the charging and discharging process of the drone battery, i.e., the continuous electrical and thermal sampling signals characterizing the physical state of the battery acquired in real time by sensors, voltage, discharge capacity, and temperature are selected as the original degradation characteristics. During charging and discharging, voltage changes, discharge capacity decay, and temperature distribution directly reflect the loss of active materials and the increase in polarization resistance inside the battery. For the process of acquiring device operating status data using sensors and acquisition circuits, in one embodiment, the sensors can be voltage sensors, Hall current sensors, and patch thermocouple temperature sensors. Those skilled in the art can use conventional analog-to-digital converters and data loggers for continuous sampling and measurement. The hardware acquisition methods and basic data extraction processes are well-known technologies in the field and will not be elaborated here.
[0072] S102, after obtaining the above-mentioned original degradation characteristics, linear interpolation resampling processing is performed on the voltage time series, discharge capacity time series, and temperature time series. The charge / discharge rate of the UAV varies dynamically in the actual working environment, resulting in differences in the number of sampling data points for a single charge / discharge cycle. To ensure the consistency of the input dimension of the subsequent model structure, a one-dimensional linear interpolation algorithm is used to realign the non-uniform length original sequences along the time axis, unifying them to a fixed length. In this embodiment, as a preferred approach, the fixed length is set to... .
[0073] S103, to eliminate dimensional differences and preserve the absolute decay trend within the lifetime, a channel-independent scaling strategy based on physical common sense is adopted. Specifically, for the resampled voltage and temperature time series, truncation mapping is performed based on the physical operating limits of the device. Conventional minimum-maximum normalization methods heavily rely on the extreme values within the current input sequence, which can cause data from different degradation stages to lose their original absolute physical amplitude differences after normalization. Dimensionless mapping is performed using a pre-determined absolute operating limit interval; the relevant truncation mapping calculation logic is as follows:
[0074] ;
[0075] in, For the resampled voltage time series or temperature time series at the 1st Observed values at each sampling point; and These respectively characterize the lower and upper limits of the physical operating limits of the corresponding channel given in the factory design for this type of battery; This represents a truncation function used to forcibly constrain abnormal data exceeding the design limits within the physical operating boundary. The specific operational logic of this truncation function is defined as follows: when the input value satisfies... When, output When the input value satisfies When, output When the input value satisfies When, output . These are the normalized feature values obtained after truncation mapping. Specific physical limit parameters are set based on the inherent properties of the test object. For example, for a soft-pack lithium-ion battery with a rated capacity of 0.74Ah, the set absolute voltage limit is 4.2V at the upper limit and 2.7V at the lower limit; the set absolute temperature limit is 60℃ at the upper limit and 20℃ at the lower limit.
[0076] Unlike the voltage and temperature channels, S104 directly normalizes the resampled discharge capacity time series using the factory rated capacity. Capacity characteristics exhibit monotonically decaying properties; preserving the relative true decay between cycles is crucial for extracting effective degradation features. Therefore, the capacity channel uses static physical constants for proportional numerical scaling, and its normalization calculation is as follows:
[0077] ;
[0078] in, The resampled discharge capacity time series is at the 1st Observed values at each sampling point; These are the standard factory rated capacity parameters for drone batteries. This represents the dimensionless capacity characteristic value obtained after per-unit processing. The feature extracted through this division operation maintains the gradient ratio state of the battery's evolution from the early to the late stages of its lifespan.
[0079] S105, after independent scaling of each channel, the truncated and mapped voltage time series, the truncated and mapped temperature time series, and the normalized discharge capacity time series are concatenated and aligned according to the channel dimension to obtain a multi-channel feature matrix. This process filters out the network training gradient update imbalance caused by different unit scales, while avoiding statistical feature transformations that are detached from physical reality. The specific dimensions of the obtained multi-channel feature matrix are as follows: These three channels correspond to the preprocessed voltage, capacity, and temperature data streams, respectively. This multi-channel feature matrix will serve as the standard input format for the next step of the one-dimensional convolutional autoencoder to extract latent features.
[0080] See attached document Figure 1 In this embodiment, step S20 is specifically executed through the following sub-steps:
[0081] S201, Construct the encoder network part of the one-dimensional convolutional autoencoder and set the relevant network structure parameters. The multivariate time series data of drone batteries exhibits strongly coupled physical characteristics within local time intervals, requiring feature extraction tools to capture local degradation features. The preprocessed multi-channel feature matrix is input into the one-dimensional convolutional autoencoder. In this embodiment, this input is the dimension of the output from the aforementioned preprocessing step. The encoder contains two one-dimensional convolutional layers for local temporal feature extraction. As a preferred approach, the kernel size of both one-dimensional convolutional layers is set to 5, and the stride is set to 2. After one-dimensional convolution, the number of channels in the input feature map is mapped sequentially from 3 to 16 to 32, with ReLU activation function used for both. Subsequently, a flattening layer and two fully connected layers are applied, ultimately outputting scalar latent features. The scalar's latent characteristics In terms of business meaning, it represents the compressed value of high-dimensional degradation information corresponding to the current single charge-discharge cycle.
[0082] S202, construct the decoder network part of the one-dimensional convolutional autoencoder. This is to verify the scalar latent features extracted by the encoder. Whether the original degraded information is truly preserved in its entirety needs to be verified through a reconstruction mechanism. The decoder employs a symmetrical structure, reshaping the input through a fully connected layer before entering two one-dimensional transposed convolutional layers to reconstruct the original input. During this reconstruction process, the input's scalar latent features... The fully connected layer remaps the feature map to the same dimension as the output of the flattened layer. The kernel size and stride parameters of the two subsequent one-dimensional transposed convolutional layers are consistent with the encoder. Through upsampling, the number of feature channels is reverse-mapped sequentially from 32→16→3, resulting in a final decoded output dimension that is also [missing information]. The reconstructed feature matrix is used to restore the shape and value of the multi-channel feature matrix.
[0083] S203 specifies the training samples, supervision labels, and key network parameters of the one-dimensional convolutional autoencoder to perform network parameter optimization in an unsupervised state. The one-dimensional convolutional autoencoder adopts the self-supervised training paradigm in unsupervised learning. Specifically, a dataset containing multiple known complete lifecycles of similar drone batteries needs to be pre-established as the training set. The historical battery degradation data stored within this training set consists of multivariate sensor features such as voltage, capacity, and temperature collected from these historical drone batteries during their continuous operation from brand new to the end of their lifespan. The training samples are defined as the multi-channel feature matrix extracted from the aforementioned historical battery degradation data through the multi-channel data preprocessing step; the supervision labels are directly set to the multi-channel feature matrix itself, meaning that the model output should reproduce the input as closely as possible. For the underlying code implementation of the neural network model and the gradient backpropagation mechanism, those skilled in the art can use conventional deep learning frameworks (such as TensorFlow or PyTorch) to build it. The basic programming and computation processes are well-known technologies in this field and will not be elaborated here. Key network parameters include the loss function used to evaluate model reconstruction error, the learning rate controlling the weight update step size, the type of optimization algorithm, and the total number of network iterations. The model training process constructs this reconstruction loss function using mean squared error, and the specific calculation logic is as follows:
[0084] ;
[0085] in, This represents the reconstruction loss value for the current iteration batch; The total number of samples in a single training batch; 3 represents the number of feature channels in the input data; Represents a uniform and fixed sequence length; The multi-channel feature matrix of the real input is in the first... The sample, the first The first channel, the first The specific actual value of each time step; This represents the predicted value at the same position in the reconstructed feature matrix output by the decoder. In this embodiment, the model training uses the Adam optimizer with a learning rate of 0.001 and a total number of iterations of 50 epochs. By continuously minimizing the reconstruction loss function through the backpropagation algorithm, the encoder can effectively filter random environmental noise in multi-channel sensor signals and learn its inherent, minimal degradation rules.
[0086] S204 utilizes a trained one-dimensional convolutional autoencoder to perform forward inference, constructing and outputting a latent feature sequence. During the actual online testing and feature extraction phases, only the trained and solidified encoder network structure and its corresponding weight parameters are retained; the decoder part no longer participates in the computation. Multiple multi-channel feature matrices generated by the drone battery during consecutive charge-discharge cycles are sequentially input into the encoder. The encoder performs independent forward propagation operations on the data from each cycle, outputting the corresponding scalar latent features sequentially. Based on the chronological order of equipment operation, the scalar latent characteristics of each cycle output are... By splicing and combining the time sequences, a one-dimensional time series that increases with the number of charge-discharge cycles is obtained, which is the output latent feature sequence. This latent feature sequence physically represents the fluctuating downward trajectory of battery health in a high-dimensional feature space, and also constitutes the direct data source for subsequent monotonic mapping and strict physical boundary constraints.
[0087] See attached document Figure 1 In this embodiment, step S30 is specifically executed through the following sub-steps:
[0088] S301, Extract the local absolute gradient of the latent feature sequence. The latent feature sequence output by unsupervised learning in the previous step often exhibits non-monotonic local oscillations. These fluctuations cannot directly and accurately characterize the irreversible physical degradation process inside the battery device. To eliminate the interference caused by the aforementioned non-monotonic oscillations, calculate the absolute value of the change in latent features at adjacent charge-discharge cycles. Let... For the potential feature sequence at the current charge / discharge cycle time The observed scalar value (since the latent feature sequence is composed of features extracted sequentially from each charge-discharge cycle in step S20, the latent feature sequence is at a specific time) The observed scalar value is physically derived from the scalar latent feature of the corresponding loop output in step S20. Given the observed scalar value at the previous adjacent time step, the specific formula for calculating the local absolute gradient is as follows:
[0089] ;
[0090] in, That is, at time The extracted local absolute gradient represents the absolute physical quantity of the characteristic fluctuations between two adjacent cycles.
[0091] S302 calculates the global maximum cumulative gradient limit based on the training set, constrained by a safety margin. To impose scale constraints on subsequent accumulated features, a reasonable baseline upper limit needs to be determined. Considering the potential differences in actual degradation rates among different test batteries, a safety margin coefficient is introduced to scale this limit value to prevent the degradation rate of the test batteries from exceeding the prior distribution of the historical training set, thus causing the features to prematurely fail. The calculation logic for the global maximum cumulative gradient limit is as follows:
[0092] ;
[0093] in, This represents the set of all known complete lifecycle battery samples of the same type within a pre-established training set. This represents a specific battery sample within the set of battery samples of the same type. Indicates battery sample The sum of local absolute gradients accumulated throughout its entire lifecycle; 1.05 is the set safety margin coefficient used to extend the empirical boundary; The minimum positive constant is set to prevent mathematical anomalies such as a denominator of zero in subsequent division operations. In this embodiment, as a preferred approach, the minimum positive constant can be a value of ; This is the global maximum cumulative gradient limit obtained by solving the problem.
[0094] S303 constructs an exponentially increasing cumulative damage amount by accumulating local absolute gradients. Lithium-ion batteries often exhibit accelerated degradation in the later stages of their lifespan. By introducing an exponential mapping mechanism, this nonlinear degradation characteristic can be effectively amplified, making it more consistent with the actual internal wear and tear. The exponential cumulative damage amount is constructed using the extracted local absolute gradient and the global maximum cumulative gradient limit. The specific construction formula is as follows:
[0095] ;
[0096] in, For at any time The cumulative amount of exponential damage constructed; Represents the time from the initial moment to the current moment. The accumulated value of the local absolute gradient; This is the exponential amplification factor, used to adjust the curvature boundary of exponential growth. In this embodiment, as a preferred approach, the exponential amplification factor is set to [value missing]. .
[0097] S304 employs a purely mathematical boundary normalization method to analyze the cumulative exponential damage. Conventional dynamic normalization methods typically rely on the global extremum of the current test sequence, which introduces statistics dependent on future unknown test data in real-time online prediction scenarios, leading to serious data leakage problems. Given the cumulative exponential damage... The theoretical analytical range is strictly fixed at . Within a closed interval, normalization is performed using fixed purely mathematical analytic boundaries:
[0098] ;
[0099] in, This refers to the dimensionless characteristic quantity obtained after normalizing the analytical boundary; natural constant of The power is a factor that is set according to the exponent. The calculated constant has a theoretical upper limit. This calculation process does not require prediction of the future degradation extreme value of the test battery, fundamentally eliminating the risk of data leakage during online testing.
[0100] S305, perform a flip mapping and interval truncation to obtain the final monotonically decreasing health factor sequence. The health factor, in its physical sense, represents the remaining health of the equipment and should exhibit a strictly monotonically decreasing evolution trend with increasing usage time. Perform a reciprocal flip mapping on the normalized feature values and forcibly truncate them to eliminate extreme anomalies caused by extreme burst noise. The calculation formula is:
[0101] ;
[0102] in, The feature direction was flipped; This is a range truncation function used to force the flipped value to be bound to a closed interval between 0 and 1; that is, output 0 when the value is less than 0 and output 1 when the value is greater than 1. That is, time. The extracted monotonically decreasing health factors can be combined sequentially at different times to obtain a monotonically decreasing health factor sequence.
[0103] The aforementioned continuous mapping chain, from absolute gradient extraction, exponential accumulation, analytical boundary normalization to flip truncation, collectively ensures that the final output monotonically decreasing health factor sequence maintains a strictly monotonically decreasing trend throughout the entire lifespan of the drone battery. Simultaneously, this health factor sequence exhibits a significant nonlinear drop pattern at the end of its lifespan, consistent with the physical laws of irreversible equipment damage. This provides high-quality foundational data support for subsequent high-precision trajectory matching using the improved dynamic time warping algorithm.
[0104] See attached document Figure 1 In this embodiment, step S40 is specifically executed through the following sub-steps:
[0105] S401, set the end-of-observation time window and extract the recent local degradation rate of the health factor sequence. While the conventional baseline spherical distance can be calculated using normalized features, it suffers from a penalty failure issue when processing smoothed degradation data, failing to handle abrupt changes in local degradation rates. For the accelerated capacity drop commonly observed in lithium batteries at the end of their lifespan, accurate quantification of the recent degradation rate is essential. Let the length of the currently acquired health factor sequence from the test equipment be... (This length, in a physical sense, represents the total number of charge-discharge cycles that the test equipment has currently undergone.) The health factor elements within the sequence are represented sequentially according to time. Define the length of the end window used to capture recent trends as... The specific calculation method is as follows:
[0106] ;
[0107] in, This represents a function that takes the minimum value. This indicates a floor operation. The adaptive length design of the terminal window ensures that the window will not go out of bounds when the test sequence is short, while fixing it within a reasonable local range (e.g., 5 loops) when the sequence is sufficiently long. It is particularly important to note that, to ensure the completeness of the algorithm logic and avoid the sequence length being too short in the initial stage of the test device's operation (i.e., when...), the window is designed to be rounded down. This results in a calculated window length of zero, leading to a division-by-zero dead zone. In actual calculations, when the round-down result equals 0, the end window length is forcibly set. Based on the determined end window length, the recent local degradation rate is calculated:
[0108] ;
[0109] in, This indicates the recent local degradation rate (i.e., the average rate of change of the health factor sequence within the end window). This represents the latest observed values of health factors (i.e., the last element of the health factor sequence). The observed values of health factors at the start of the final window; This indicates the absolute value operation.
[0110] S402, calculate the global average degradation rate of the health factor sequence over the entire known observation period. To assess whether recent local degradation has accelerated abnormally, a global reference baseline needs to be established. Extract the values at the beginning and end of the health factor sequence and calculate the overall average rate of change:
[0111] ;
[0112] in, Indicates the global average degradation rate; This represents the baseline values for the health factor sequence at the initial time. This indicates an absolute value operation; the global average degradation rate reflects the rate at which the overall health deterioration of the test equipment evolves from a brand-new state to the present moment.
[0113] S403. Based on the relative proportion between the recent local degradation rate and the global average degradation rate, a local acceleration factor is constructed. The extracted recent degradation features are compared with the global benchmark by division, which dynamically quantifies the severity of the degradation rate at the current moment. The specific logic for constructing the local acceleration factor is as follows:
[0114] ;
[0115] in, This is the calculated local acceleration factor. When recent degradation is significantly faster than the historical average, the value of the local acceleration factor will be significantly amplified. This is a set minimum normal number to prevent mathematical operations where the divisor is zero from occurring during the initial stages of device operation or when the overall degradation is extremely small. In this embodiment, as a preferred approach, this minimum normal number can be set to a value of... .
[0116] S404 uses a local acceleration factor for linear mapping and threshold truncation to dynamically generate a nonlinear penalty upper limit. In the improved sequence matching algorithm, a higher alignment penalty weight needs to be assigned to the degradation acceleration stage to improve the model's matching sensitivity to nonlinear plunges at the end of the lifespan. The mapping expression for calculating the corresponding penalty upper limit using the local acceleration factor is as follows:
[0117] ;
[0118] in, The nonlinear penalty upper limit is dynamically generated; the constant 0.5 represents the starting benchmark of the basic penalty upper limit, and 0.15 is the linear mapping slope coefficient that controls the growth of the penalty upper limit with the local acceleration factor. The operation aims to forcibly truncate and constrain the mapping results to within 0.9. This truncation constraint mechanism mathematically prevents a mathematical explosion in the penalty weights for calculating the final spherical distance when the local acceleration factor is abnormally large due to extreme noise, thus ensuring the stability of the prediction algorithm under complex conditions. The upper limit of the nonlinear penalty in the calculation output is also considered. This will be used directly as the core input parameter for the next step of constructing the feature-aligned cumulative cost matrix.
[0119] See attached document Figure 1 In this embodiment, step S50 is specifically executed through the following sub-steps:
[0120] S501, extract the test sequence and reference sequence, and initialize the cumulative cost matrix and search constraint window. Let the monotonically decreasing health factor sequence extracted by the test equipment at the current time be the test sequence. Its sequence length is (here) This represents the total number of charge-discharge cycles that the test equipment has currently completed, and its value ranges from [value missing]. (Integers). Extract a historical battery degradation trajectory with a known complete lifespan from the pre-established training library (i.e., the dataset containing historical UAV batteries with multiple known complete lifespans in step S20) as a reference sequence. Its sequence length is set to ( The total number of cycles in the complete trajectory, with a value range of [value missing]. (Integers). To perform dynamic programming optimization, construct a dimension of... Cumulative cost matrix The cumulative cost matrix All internal elements are initially assigned the value of positive infinity. ), and set the starting point. To improve the algorithm's search efficiency and avoid ill-conditioned over-alignment during the time warping process, the size of the Sakoe-Chiba search constraint window is set. The specific calculation formula is as follows:
[0121] ;
[0122] in, This is a function to find the maximum value. This is a floor function. The constant 0.15 is the search constraint scaling factor. In specific numerical terms, this refers to the size of the pre-defined search constraint window (e.g., when the test sequence length...). At that time, the size of the search constraint window can be calculated. This constraint window limits the search range for sequence matching to near the diagonal, ensuring the rationality of the matching path on the physical time axis.
[0123] S502 constructs a linear interpolation array based on the upper bound of the nonlinear penalty and calculates the nonlinear penalty weights using the spherical distance method. The upper bound of the nonlinear penalty is utilized. Construct a linear interpolation array to represent the relative time progress of the sequence. The array starts from a fixed point. Starting from the upper limit of non-linear penalty End, the total length generated by internal uniform interpolation is The sequence. The total length is The specific linear interpolation formula for generating the sequence is: , where array index The range of values is In the test sequence of the first... time steps ( When performing traversal operations, extract the corresponding interpolation elements of the array. The spherical distance method is introduced to calculate the nonlinear penalty weight for the current time step. The calculation logic is as follows:
[0124] ;
[0125] in, Indicates scalar and The normalized spherical distance function calculated after mapping to a unit sphere. Specifically, the mathematical analytical expression of this normalized spherical distance function is defined as follows: For the specific projection coordinate calculation of this normalized spherical distance, those skilled in the art can use conventional polar coordinate transformation and arc length calculation formulas, which are well-known techniques in the field and will not be elaborated here. A constant of 1.0 is the basic invariant weight, and 5.0 is the set penalty amplification factor. This nonlinear penalty weight... Able to move with time As the lifespan progresses, the alignment penalty gradually increases, thus imposing a more severe alignment penalty on the characteristic differences at the end of the lifespan.
[0126] S503 calculates the weighted local cost between the test equipment and historical trajectories within the search constraint window. This is based on the predefined search constraint window size. Determine the current time step In the reference sequence The corresponding alignment search boundary. Define the search starting point of the current time step as... The search endpoint is Within a defined search range (i.e., for an index) satisfy Within the range), calculate the weighted local cost of each state transition:
[0127] ;
[0128] in, This represents the weighted local cost of the test equipment and historical trajectories at the current alignment node. and These represent the feature observation values of the test sequence and the reference sequence at corresponding array index positions, respectively. This is achieved by first calculating the square of the feature difference and then multiplying it by a non-linear penalty weight. It achieves adaptive amplification and quantification of local degradation mutations and morphological deviations in the middle and late stages of life.
[0129] S504 updates the cumulative cost matrix based on the state transition equation, and finally extracts and outputs the shortest matching distance. Using the weighted local cost obtained from the above calculation, the cumulative cost matrix... Dynamic programming is used to perform state transitions. The specific state transition equation is:
[0130] ;
[0131] in, Indicates accumulation to node The overall state transition cost; The mechanism is used to select the minimum effective path from the accumulated costs of previous histories from the horizontal, vertical, and diagonal directions. It iterates through all time steps of the test sequence using nested loops. and the corresponding search range Gradually complete the full cumulative cost matrix Numerical padding. After all time steps have been calculated, the last element in the bottom right corner of the matrix is directly extracted and its square root is taken, thus minimizing the shortest matching distance. In this embodiment, the shortest matching distance In terms of business implications, it accurately quantifies the overall health evolution similarity between the current test equipment and the selected specific historical trajectory after comprehensively considering time offset, local degradation acceleration, and nonlinear degradation morphology characteristics.
[0132] See attached document Figure 1 In this embodiment, step S60 is specifically executed through the following sub-steps:
[0133] S601 uses a sliding window mechanism to traverse and scan all historical degradation trajectories in a pre-established training library to find the optimal local alignment state within each trajectory. This is taken into account the current total number of running cycles on the test equipment. The total loop length of historical trajectories, which is typically much shorter than the known complete lifespan, can lead to severe distortion of the physical time scale if direct end-to-end global alignment is performed. Therefore, a sliding window is introduced for local feature extraction. To ensure the completeness of the algorithm logic and avoid sliding window dead zones caused by excessively long test sequences, a validity filter is performed on the historical trajectories to be matched before traversal, retaining only those with a total loop length greater than the current test sequence length. The historical trajectory. Let's assume the current iteration reaches the th element that has passed the validity filter. The total number of loop steps corresponding to the actual end point of the historical trajectory is: (satisfy Set the length of the sliding window and the length of the test sequence. To maintain consistency, let the sliding window be in the first... The system slides along the historical trajectory in single-step loops. The initial capture index range for the window is 1 to... The length obtained for each sliding segment is... For each reference subsequence, the nonlinear penalty and spherical distance calculation logic in step S50 above is invoked to solve for the shortest matching distance between the test sequence and the reference subsequence. After traversing the first... After tracing all feasible windows on a historical trajectory, the smallest distance value is selected as the representative distance for that historical trajectory. The index position of the sliding window corresponding to this smallest distance on the original historical trajectory is recorded, and this index is defined as the matching endpoint. .
[0134] S602, based on the determined matching endpoint, extract the remaining lifetime of the historical trajectory corresponding to the shortest matching distance. For the... This historical trajectory represents a known and complete physical degradation pattern. The matching endpoint is determined based on the optimization steps from the previous step. This allows for the accurate determination of the remaining usable time of the historical trajectory in the current most similar degradation state. The remaining lifetime of the historical trajectory at the matching point is also determined. The calculation method is the difference in the number of steps between the actual end of the lifetime and the matching end point:
[0135] ;
[0136] in, For the extracted first The remaining lifetime of a historical trajectory is a physical representation of the actual number of charge-discharge cycles that a particular historical device experiences from the end of the matching process until complete failure, thus providing a reference benchmark for the remaining lifetime of the current test device based on the same degradation state.
[0137] S603, perform global candidate sorting and extract the globally optimal matching distance and its corresponding historical trajectory remaining lifetime. Repeat the above sliding optimization and remaining lifetime extraction operations for all historical trajectories filtered for validity within the pre-established training library to construct a global candidate set containing multiple shortest matching distances and their corresponding historical trajectory remaining lifetimes. Sort all candidate results in this set in strict ascending order according to the numerical value of the shortest matching distance. Simply selecting a single optimal matching trajectory is easily affected by noise from individual random features in historical data; using a multi-trajectory fusion strategy can effectively improve the robustness of the prediction results. Extract the top-ranked optimal solution from the sorted results. In this embodiment, as a preferred method, extract the two globally smallest distance values (i.e., the Top-2 minimum distances), and denote them as follows: and (satisfy ), and record their corresponding historical trajectory remaining lifespans as , respectively and .
[0138] S604 calculates the reciprocal of the distance as a weight, performs a weighted summation of the remaining lifetime of historical trajectories, and outputs the final predicted remaining lifetime. A reciprocal weight mapping is constructed based on the extracted Top-2 minimum distances, ensuring that historical device degradation experience with smaller matching distances carries a higher weight in the final prediction result. To ensure the completeness of the algorithm logic and prevent division by zero anomalies when a perfect match occurs (i.e., distance is zero), a very small positive constant is introduced into the denominator. The specific fusion calculation expression is as follows:
[0139] ;
[0140] ;
[0141] ;
[0142] in, and They are respectively assigned to the minimum distances corresponding to the Top-2 above. and The fusion weight of historical trajectories; To prevent division by zero anomalies, a very small positive number is set (similar to the previous steps, this is a preferred value). ); This is a predicted remaining lifetime output based on a weighted sum of remaining lifetimes from historical trajectories. This predicted remaining lifetime integrates multiple historical equipment degradation experiences with the highest similarity in health evolution, eliminating the influence of a single extreme value. In terms of specific operational meaning, it provides a high-precision quantification of the remaining number of cycles that the currently tested equipment can continue to operate safely, thus providing a reliable decision-making basis for early warning and condition-based maintenance of the equipment.
[0143] To further verify the effectiveness and superiority of the UAV battery life prediction method based on monotonic features and adaptive DTW provided by this invention, the beneficial technical effects of this invention will be described in detail below with reference to specific experimental data, figures, and comparative test results.
[0144] I. Experimental Data Preparation and Basic Parameter Setting
[0145] In this verification embodiment, the publicly available Oxford University battery degradation dataset is used for algorithm verification. This dataset contains multi-channel continuous monitoring data of a 0.74Ah pouch lithium-ion battery under actual operating conditions. The test objects consist of 8 battery samples (referred to as Cell1 to Cell8).
[0146] To ensure the homogeneity of the verification and exclude heterogeneous extreme anomalies in the test equipment at the physical level, this embodiment pre-performed anomaly removal. Specifically, this embodiment removed heterogeneous outliers (Cell2, Cell4, Cell5, and Cell6) with substandard initial capacity or severely abnormal feature gradients. This embodiment used four core battery samples (Cell1, Cell3, Cell7, and Cell8) with representative remaining degradation trajectories for leave-one-out cross-validation. That is, in each test, the data from three battery samples were added to the training library, and the remaining one was used as the current test equipment for online evaluation of predicted remaining lifetime.
[0147] In the preprocessing of the multi-channel feature matrix and the setting of network parameters, the aforementioned steps were strictly followed: a uniform and fixed sequence length was set as follows. =200; The set cutoff mapping physical operating limits are: upper voltage limit 4.2V and lower limit 2.7V, upper temperature limit 60℃ and lower limit 20℃; the rated capacity is set to the rated capacity. =0.74Ah. The kernel size of the one-dimensional convolutional layer in the one-dimensional convolutional autoencoder is set to 5, the stride to 2, and the number of feature channels is mapped sequentially from 3 to 16 to 32. The model uses the Adam optimizer with a learning rate of 0.001 and a total number of iterations of 50 epochs. During the construction of the health factor sequence, the exponential amplification factor is set to... The smallest normal number is set as In the global maximum cumulative gradient limit calculated based on the training library, the safety margin coefficient is set to 1.05 to preserve physical margin for test cells whose degradation rate exceeds the prior distribution of the training library.
[0148] II. Validation of Degradation Feature Extraction and Physical Constraints
[0149] Combined with appendix Figure 2 With appendix Figure 3 The actual effectiveness of the feature extraction in this invention will be verified and explained.
[0150] like Figure 2 As shown, normal battery samples Cell1, Cell3, Cell7, and Cell8 exhibit highly smooth quasi-linear degradation characteristics. However, some heterogeneous batteries, Cell4, Cell5, and Cell6, experienced sudden capacity drop failures in the middle of their lifespan. This smooth degradation trajectory, lacking local peak and trough markers, is the core challenge that makes traditional DTW algorithms prone to over-alignment (severe time axis distortion) when searching for similarity.
[0151] like Figure 3 As shown, after inputting the raw data into a one-dimensional convolutional autoencoder, and performing local absolute gradient calculation, accumulation to construct exponential damage accumulation, and analytical boundary normalization and flipping truncation, the health factor sequences extracted from the core test batteries Cell1, Cell3, Cell7, and Cell8 were successfully constrained into a strictly monotonically decreasing pattern. This not only completely eliminates the interference of local environmental noise on similarity matching, but also ensures that the feature trajectory fully conforms to the physical nature of irreversible device degradation, providing high-quality data input for subsequent shortest matching distance calculation.
[0152] III. Real-time tracking, evaluation, and verification of predicted remaining lifetime
[0153] Combined with appendix Figure 4 The stability of real-time prediction in this invention is verified and explained.
[0154] like Figure 4 As shown in the figure, the solid black line represents the true remaining lifetime (RUL), and the dashed red line represents the predicted value output by this invention. As the loop progresses, the predicted trajectory output by this invention consistently closely matches the true remaining lifetime path. Due to the introduction of a local acceleration factor and the Sakoe-Chiba search constraint window mechanism, the model did not exhibit any significant overestimation of lifetime or sudden drops in instability throughout its entire lifecycle. Its average tracking error is only 172.3 loops, and the relative error is strictly controlled at an extremely low level of 4.4%.
[0155] IV. Validation of Early Prediction and Asymmetric Risk Control in Multiple Degradation Stages
[0156] Combined with appendix Figure 5 This paper verifies and explains the robustness and risk control capability of the present invention in the complex nonlinear degradation stage. Euclidean distance method, standard DTW method, and traditional spherical DTW method are introduced as comparative methods, and the method of the present invention (see appendix) is compared with those of the present invention. Figure 5 A comparative test was conducted with the one marked A-SDTW.
[0157] The stage differences between stiffness measurement and elastic regularity (30% and 50% stages): such as Figure 5 As shown in (a), in the early to mid-stages when the device operates at 30% and 50%, the Euclidean distance achieves a relatively low prediction error due to its rigid point-to-point matching characteristic, as the battery degradation trajectory is in a highly flat quasi-linear range. The method of this invention, through reasonable basic weight settings, also performs excellently in this stage.
[0158] Robustness performance during the core degradation phase (70% stage): When the prediction node advances to the 70% EOL (End of Life) stage, the battery enters a nonlinear accelerated degradation phase, and different batteries exhibit severe time axis phase shifts due to individual differences. At this point, the rigidity of the Euclidean distance method is fully exposed, with its MAE (Mean Absolute Error) soaring to 333.86 cycles; although standard DTW and traditional spherical DTW allow for time axis scaling, they still produce errors exceeding 285 cycles. In contrast, this invention, through dynamic sensing of local acceleration factors and the resulting dynamically generated nonlinear penalty upper limit, successfully achieves accurate alignment of the accelerated degradation phase. In this critical stage, this invention achieves the lowest MAE of 280.99 cycles, effectively overcoming the late-stage inaccuracy problem of the Euclidean distance method.
[0159] Global Optimality and Asymmetric Statistical Significance: Based on a comprehensive lifecycle evaluation, this invention achieved a globally optimal average MAE of 262.43 cycles. In the Wilcoxon signed-rank test for all prediction starting points, the statistical significance probability value (i.e., p-value; when this probability value is less than 0.01, it indicates that the difference in accuracy between the two algorithms is highly statistically significant, rather than a random error) for the difference between the two sets of data was calculated as p = 0.0024 < 0.01, confirming that this invention has strong statistical significance in improving accuracy. Regarding risk control, such as... Figure 5As shown in the NASA scoring trend in (b), under the constraints of monotonic features and nonlinear penalties, the asymmetric score of the present invention remains stable and excellent, effectively avoiding the serious underestimation of the lifetime of Euclidean distance and the extremely dangerous "lifetime overestimation" phenomenon of standard DTW, which proves that it has extremely high practical engineering application potential in modern equipment PHM (Prognostics and Health Management) systems.
[0160] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting UAV battery life based on monotonic features and adaptive DTW, characterized in that, Includes the following steps: The voltage time series, discharge capacity time series, and temperature time series of the drone battery are obtained and preprocessed to obtain a multi-channel feature matrix; The multi-channel feature matrix is input into a one-dimensional convolutional autoencoder, which outputs a latent feature sequence. Calculate the local absolute gradient of the latent feature sequence and accumulate it to construct an exponential damage accumulation. Perform analytical boundary normalization and flip truncation on the exponential damage accumulation to obtain a monotonically decreasing health factor sequence. A local acceleration factor is calculated based on the monotonically decreasing health factor sequence, and a nonlinear penalty upper limit is dynamically generated based on the local acceleration factor. The state transition cost between the test equipment and the historical trajectory is calculated using the nonlinear penalty upper limit and spherical distance method, and the shortest matching distance is found through the cumulative cost matrix. Extract the remaining lifetime of the historical trajectory corresponding to the shortest matching distance, and output the predicted remaining lifetime based on the remaining lifetime of the historical trajectory. The step of calculating a local acceleration factor based on the monotonically decreasing health factor sequence and dynamically generating a nonlinear penalty upper limit based on the local acceleration factor includes: Set a final time window for observation, and extract the average rate of change of the monotonically decreasing health factor sequence within the final time window as the recent local degradation rate; The rate of change of the values at the beginning and end of the monotonically decreasing health factor sequence is calculated as the global average degradation rate. The local acceleration factor is obtained by dividing the recent local degradation rate by the global average degradation rate and adding a very small normal number. The nonlinear penalty upper limit is obtained by linearly mapping and thresholding the local acceleration factor.
2. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 1, characterized in that, The steps of acquiring and preprocessing the voltage time series, discharge capacity time series, and temperature time series of the UAV battery to obtain a multi-channel feature matrix include: The voltage time series, discharge capacity time series, and temperature time series of the UAV battery are subjected to linear interpolation and resampling processing to be unified to a fixed length; Based on the physical operating limits, the resampled voltage time series and the resampled temperature time series are truncated and mapped to obtain the truncated and mapped voltage time series and the truncated and mapped temperature time series. The discharge capacity time series after resampling is normalized by using the factory rated capacity to obtain the normalized discharge capacity time series. The voltage time series after truncation mapping, the temperature time series after truncation mapping, and the discharge capacity time series after per-unit processing are spliced and aligned according to the channel dimension to obtain the multi-channel feature matrix.
3. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 1, characterized in that, The step of inputting the multi-channel feature matrix into a one-dimensional convolutional autoencoder and outputting a latent feature sequence includes: The multi-channel feature matrix is input into the encoder of the one-dimensional convolutional autoencoder, and after one-dimensional convolution operation, flattening layer and fully connected layer, scalar latent features are output. The multi-channel feature matrices generated by the drone battery under multiple consecutive charge-discharge cycles are sequentially input into the encoder, and the corresponding scalar latent features are output sequentially. The scalar latent features are spliced and combined in sequence to obtain the latent feature sequence.
4. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 1, characterized in that, The step of calculating the local absolute gradient of the latent feature sequence and accumulating it to construct the exponential damage accumulation includes: Calculate the absolute value of the change in the potential feature sequence at adjacent charge-discharge cycle times to obtain the local absolute gradient; The global maximum cumulative gradient limit, constrained by a safety margin, is calculated using a pre-established training set. The exponential damage accumulation is constructed using the local absolute gradient and the global maximum cumulative gradient limit.
5. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 4, characterized in that, The step of performing analytical boundary normalization and flipping truncation on the accumulated exponential damage to obtain a monotonically decreasing health factor sequence includes: The accumulated exponential damage is normalized using a purely mathematical analytical boundary to obtain the normalized feature quantity. The normalized feature values are subjected to a flip mapping and interval truncation to obtain the health factor; The health factors at different times are combined in chronological order to obtain the monotonically decreasing health factor sequence.
6. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 1, characterized in that, The steps for calculating the state transition cost of the test equipment and historical trajectory using the nonlinear penalty upper limit and spherical distance method include: The monotonically decreasing health factor sequence extracted from the testing device is set as the test sequence, and the historical trajectory is extracted from the pre-established training library as the reference sequence. Initialize the cumulative cost matrix and search constraint window; A linear interpolation array is established using the aforementioned nonlinear penalty upper limit, and the nonlinear penalty weight is calculated using the aforementioned spherical distance method. Within the search constraint window, the weighted local cost between the test sequence and the reference sequence is calculated based on the nonlinear penalty weight as the state transition cost.
7. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 6, characterized in that, The steps for finding the shortest matching distance using the cumulative cost matrix include: Using the weighted local cost, dynamic programming state transition equations are updated in the cumulative cost matrix to complete the numerical filling of the cumulative cost matrix; The shortest matching distance is obtained by extracting the last element of the cumulative cost matrix and performing a square root operation.
8. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 1, characterized in that, The step of extracting the remaining lifetime of the historical trajectory corresponding to the shortest matching distance includes: The sliding window mechanism is used to traverse and scan all the historical trajectories in the pre-established training library; The reference subsequence is extracted using the sliding window mechanism, and the matching distance between the monotonically decreasing health factor sequence and the reference subsequence is calculated. The shortest matching distance is selected as the distance with the smallest value among all the matching distances, and the end position of the sliding window corresponding to the shortest matching distance is defined as the matching endpoint. The remaining lifespan of the historical trajectory is obtained by calculating the step difference between the matching endpoint and the actual lifespan termination point.
9. The UAV battery life prediction method based on monotonic features and adaptive DTW according to claim 8, characterized in that, The step of predicting the remaining lifetime based on the historical trajectory includes: Construct a global candidate set that includes multiple shortest matching distances and the corresponding remaining lifetime of the historical trajectories; The global candidate set is sorted in ascending order according to the numerical value of the shortest matching distance, and the two shortest matching distances with the smallest values and the corresponding remaining lifetime of the historical trajectory are extracted. The reciprocal of the distance between the two shortest matching distances with the smallest values is used as a weight, and the remaining lifetime of the corresponding historical trajectories is weighted and summed to output the predicted remaining lifetime.