Search

CN-122017598-A - Power battery residual service life prediction method based on big data analysis

CN122017598ACN 122017598 ACN122017598 ACN 122017598ACN-122017598-A

Abstract

The invention discloses a method for predicting the residual service life of a power battery based on big data analysis, and belongs to the technical field of battery health management. The method comprises the steps of synchronously collecting battery operation time sequence data, vehicle working condition data and environment data and preprocessing, extracting second-level transient characteristics and travel-level working condition characteristics, constructing calendar aging accumulation characteristics for quantifying high-temperature standing and high-charge state holding time, constructing a depth residual space-time network prediction model with enhanced physical information, guiding model learning to conform to a physical rule by introducing a physical consistency constraint item based on an experience degradation model into a loss function, and adopting a migration learning strategy combining cloud pre-training and edge side light weight fine adjustment to realize personalized adaptation of the model and dynamic recursion prediction of the residual service life. The invention improves the prediction precision and the model generalization capability, and realizes the accurate modeling of the coupling effect of the cyclic aging and the calendar aging.

Inventors

  • OUYANG YUFENG
  • WANG BO
  • WU SHIPING

Assignees

  • 湖南攻防新能源有限公司

Dates

Publication Date
20260512
Application Date
20260302

Claims (9)

  1. 1. A method for predicting remaining service life of a power battery based on big data analysis, the method comprising the steps of: Step 1, synchronously collecting running time sequence data, vehicle working condition time sequence data and historical environment data of a power battery, and cleaning, filling missing values and carrying out normalization pretreatment on multi-source time sequence data; step 2, dynamically extracting second-level or minute-level transient characteristics and travel-level working condition characteristics based on the preprocessed data, and constructing calendar aging accumulation characteristics to form multi-time-scale health characteristic factors; Step 3, constructing a depth residual space-time network prediction model with physical information enhancement, wherein the physical information enhancement is specifically that a physical consistency loss term is introduced into a total loss function trained by the model, the physical consistency loss term is calculated based on an experience degradation model and is used for restraining the health state change trend predicted by the model to accord with a physical rule; And 4, based on transfer learning, performing personalized fine adjustment on the pre-trained depth residual space-time network prediction model enhanced by physical information by utilizing recent historical data of the target vehicle, and recursively predicting the health state sequence of the target power battery by utilizing the fine-adjusted model until the health state is lower than a failure threshold value, so as to obtain the residual service life.
  2. 2. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 1, wherein in step1, the specific process of synchronous acquisition comprises: The method comprises the following steps of 1.1, acquiring operation time sequence data of power battery monomers in real time at a fixed sampling frequency of 1 Hz through a vehicle-mounted battery management system, wherein the operation time sequence data comprise terminal voltage, total current and temperature data measured by a temperature sensor on the surface of each single battery; Step 1.2, synchronously acquiring vehicle working condition time sequence data with the same sampling frequency as that of step 1.1 through a vehicle controller local area network bus, wherein the vehicle working condition time sequence data comprises vehicle speed, motor required power, accelerator pedal opening, brake pedal state and vehicle geographic position information; step 1.3, recording a time stamp accurate to a millisecond level for each piece of data acquired in the step 1.1 and the step 1.2, and uploading the data with the time stamp to a cloud server through a network communication module; Step 1.4, calling and matching historical environment data under the same time stamp from an environment data service platform according to the time stamp and the geographic position information of the vehicle at a cloud server, wherein the historical environment data at least comprises environment temperature data; the specific process of cleaning, missing value filling and normalization pretreatment comprises the following steps: step 1.5, adopting an outlier detection algorithm based on a sliding window and a 3 sigma criterion to detect outliers of all time sequence data obtained in the steps 1.1, 1.2 and 1.4, and eliminating data identified as outliers; Step 1.6, for missing data generated by data rejection or transmission loss, if the missing time is far smaller than the main data change period, filling by adopting a linear interpolation method, wherein the main data change period is determined according to the typical charge-discharge cycle time of the vehicle; And 1.7, performing minimum-maximum normalization processing on all the filled numerical field data, and linearly transforming all the data into the interval of [0,1 ].
  3. 3. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 2, wherein in step 2, the process for extracting transient characteristics of seconds or minutes comprises: Step 2.1.1, dividing the preprocessed time sequence data according to a fixed time length window of 60 seconds to obtain a continuous time window sequence; Step 2.1.2, calculating a group of transient response characteristics in each time window, wherein the transient response characteristics comprise standard deviation of current in the time window, average charge-discharge multiplying power in the time window and difference value between the highest temperature and the lowest temperature in the time window; 2.1.3, in each time window, based on a first-order RC equivalent circuit model, utilizing voltage and current data corresponding to pulse charge and discharge segments in the window, obtaining an ohmic internal resistance value of the power battery through online identification by a recursive least square method, and calculating the change rate of the ohmic internal resistance value relative to the ohmic internal resistance value in the last time window; the process for extracting the stroke-level working condition characteristics comprises the following steps: Step 2.2.1, dividing continuous time sequence data into a plurality of independent travel segments according to an upper electric signal and a lower electric signal of a vehicle; And 2.2.2, calculating a group of working condition aggregation characteristics in each stroke segment, wherein the working condition aggregation characteristics comprise the total mileage of the stroke, the average speed of the stroke, the total times of rapid acceleration events and rapid deceleration events in the stroke, the highest temperature reached by a power battery in the stroke and the square root of accumulated discharge capacity of the stroke, wherein the rapid acceleration events are defined as continuous events with acceleration of more than 2.5 meters per second, and the rapid deceleration events are defined as continuous events with deceleration of less than-2.5 meters per second.
  4. 4. A method for predicting remaining life of a power battery based on big data analysis according to claim 3, wherein in step 2, the specific process of constructing calendar aging accumulation features includes: calculating the high-temperature stationary accumulation duration of each day by taking 24 hours as a statistical period, wherein the calculation method of the high-temperature stationary accumulation duration comprises the steps of counting the accumulated hours of the power battery with the average temperature higher than 30 ℃ and the vehicle in a flameout stationary state within one day; Calculating the high state of charge holding accumulation duration of each day by taking 24 hours as a statistical period, wherein the calculation method of the high state of charge holding accumulation duration comprises the steps of counting the accumulated hours of which the state of charge value of the power battery is higher than 80 percent and estimated by an ampere-hour integration method and an open circuit voltage method within one day; And 2.3.3, respectively carrying out exponential weighted moving average treatment on the high-temperature stationary accumulated duration daily sequence obtained in the step 2.3.1 and the high-charge state maintaining accumulated duration daily sequence obtained in the step 2.3.2 to obtain two smoothed long-term accumulated sequences, wherein the attenuation factor of the exponential weighted moving average is set according to the time constant of the calendar aging process of the power battery.
  5. 5. The method for predicting remaining service life of a power battery based on big data analysis according to claim 4, wherein in step 3, the specific process of constructing the physical information enhanced depth residual space-time network prediction model comprises: Step 3.1, constructing a model input feature tensor, wherein the input feature tensor is formed by stacking daily feature vectors for continuous N days in time sequence, the daily feature vectors are composed of the transient feature statistic value extracted in the step 2, the travel level working condition feature and the calendar aging accumulation feature together, and the value of N is determined according to a history observation period required by prediction and is an integer not less than 30; Step 3.2, constructing a depth residual space-time network encoder, wherein the encoder is formed by sequentially connecting a one-dimensional convolutional neural network layer and a two-way long-short-term memory network layer, the one-dimensional convolutional neural network layer is used for extracting local correlation inside daily feature vectors, the two-way long-term memory network layer is used for extracting front-rear dependency of features in a time dimension, and jump connection is arranged in a network deep layer to form a residual structure; The method comprises the steps of 3.3, defining a total loss function enhanced by physical information, wherein the total loss function is obtained by weighted addition of a data loss term and a physical consistency loss term, the data loss term is a mean square error between a health state value predicted by a model and a real health state value, the physical consistency loss term is calculated based on a coupling aging equation and is used for punishing deviation between a health state change trend predicted by the model and a physical trend described by the coupling aging equation, and the coupling aging equation is an empirical model fused with a power law cyclic aging model and an Arrhenius temperature dependency relationship.
  6. 6. The method for predicting remaining service life of a power battery based on big data analysis according to claim 5, wherein the specific calculation and introduction process of the physical consistency loss term in step 3.3 comprises: Step 3.3.1, establishing a coupling aging equation, wherein the coupling aging equation represents the health state as a function of cyclic accumulated stress and temperature stress, and the mathematical form of the coupling aging equation comprises a power law term and an exponential term; step 3.3.2, in each forward propagation process of model training, inputting the historical circulation times and the temperature historical sequence corresponding to the input characteristic tensor of the current batch into the coupling aging equation, and calculating the gradient direction of the equation to the health state under the current condition; step 3.3.3, calculating the cosine distance or Euclidean distance between the future health state change trend predicted by the model for the same batch of data and the physical gradient direction obtained by calculation in the step 3.3.2; And 3.3.4, taking the distance value calculated in the step 3.3.3 as a physical consistency loss term, and carrying out weighted summation on the physical consistency loss term and a data loss term according to a preset weight coefficient to form a total loss function for back propagation optimization, wherein the weight coefficient is subjected to optimization selection in the range of 0.01 to 0.5 through cross verification.
  7. 7. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 5, wherein in step 4, the specific process of personalized fine tuning of the model based on transfer learning comprises: The method comprises the steps of (1) pre-training a cloud general model, collecting complete life cycle data of various vehicle power batteries covering different vehicle types, different regions and different driving habits at a cloud server, training according to the method from step 1 to step 3 to obtain a general physical information enhanced depth residual space-time network prediction model, and taking the model as a basic model; Step 4.2, light weight fine adjustment of a vehicle-mounted edge side model, loading the general model from a cloud to a vehicle-mounted computing unit or an edge server aiming at a specific target vehicle, and fixing characteristic encoder parameters consisting of a one-dimensional convolutional neural network layer and a two-way long-short-term memory network layer in the general model; step 4.3, extracting complete historical data of the target vehicle in M months recently, and processing according to the steps 1 and 2 to obtain a multi-time scale health characteristic factor sequence corresponding to the vehicle, wherein M is an integer not less than 3; And 4.4, taking the target vehicle characteristic data obtained in the step 4.3 as a new training data set, and performing supervised training with small learning rate and a moment rounds only on the fully connected regression layer parameters loaded at the tail end of the universal model in the step 4.2 to finish migration adaptation of the model from the universal domain to the target vehicle individual domain.
  8. 8. The method for predicting remaining life of a power battery based on big data analysis of claim 7, wherein in step 4, the specific process of recursively predicting the target power battery state of health sequence to obtain the remaining life comprises: step 4.5, constructing a daily feature vector of the target vehicle for N days recently and continuously as an input feature tensor at the current moment, and inputting the daily feature vector into the prediction model subjected to personalized fine adjustment in the step 4.4; Step 4.6, the model recursively predicts the health state values from the 1 st day to the K th day in the future in an autoregressive mode by taking the health state value predicted in the previous step as a part of the input in the next step to form a health state prediction sequence, wherein K is a preset maximum prediction horizon and is set according to the maintenance period of the vehicle; Step 4.7, searching a first health state value lower than a preset failure threshold value in the health state prediction sequence, wherein the failure threshold value is set to be 70% or 80% of rated capacity according to a power battery retirement standard; and 4.8, the prediction days corresponding to the health state value lower than the failure threshold value are the residual service life predicted by the model, and the residual service life result is output to the vehicle-mounted display terminal or the remote monitoring platform in real time.
  9. 9. The method for predicting the residual service life of the power battery based on big data analysis according to claim 8, wherein the method further comprises the steps of 5, rolling update of the prediction model and online correction of the residual service life; Step 5.1, setting a fixed model updating period, wherein the updating period is synchronous with a periodic maintenance period or a data accumulation period of the vehicle and is 1 month or 3 months; step 5.2, when the update period is reached, automatically collecting new running data of the target vehicle in the period, and executing the step 1 and the step 2 to extract multi-time scale health characteristic factors corresponding to the new running data of the period; Step 5.3, combining the newly added characteristic factors with the historical characteristic factors to form an extended training data set, triggering and executing the lightweight fine tuning flow in the step 4.4 again, and performing incremental update on parameters of the prediction model; and 5.4, re-executing the steps 4.5 to 4.8 by using the model with updated parameters, generating an updated residual service life predicted value, and realizing dynamic online correction of the service life prediction.

Description

Power battery residual service life prediction method based on big data analysis Technical Field The invention relates to the technical field of power battery health management, in particular to a power battery residual service life prediction method based on big data analysis. Background With the rapid development of the global new energy automobile industry, the power battery is used as a core component, and the health state of the power battery is directly related to the safety, reliability and economy of the whole automobile. The method has the advantages of accurately predicting the residual service life of the power battery, and has important significance for optimizing the battery use strategy, early warning potential faults, evaluating residual values and guiding echelon utilization. At present, the prediction methods of the residual service life of the power battery can be mainly divided into three types, namely a model driving method based on a physical model, a data driving method based on data mining and a hybrid method combining the model driving method and the data driving method. The model driving method is that the method constructs a state equation based on an electrochemical mechanism (such as a capacity attenuation equation and an equivalent circuit model), and carries out state estimation and life prediction through Kalman filtering and other algorithms. The method has the advantages that the physical significance is clear, the defects are obvious, the aging process in the battery is complex, the multi-physical field coupling is involved, the accurate mechanism model is extremely difficult to construct, model parameters (such as diffusion coefficient and reaction rate constant) generally depend on calibration under laboratory conditions, and the method is difficult to adapt to complex and changeable environments (such as severe temperature fluctuation and random change of charge and discharge multiplying power) and individual differences in real vehicle operation, so that the prediction precision and generalization capability of the model in practical application are limited. The method does not depend on an accurate physical model, but utilizes historical operation data to directly learn the implicit rule of battery aging from the data through machine learning or deep learning algorithms (such as a support vector machine and a long-short-term memory network LSTM). With the development of the Internet of vehicles and cloud big data technology, a data driving method based on massive real vehicle operation data becomes a research hotspot. However, the existing data driving methods still face a plurality of bottlenecks, namely firstly, uneven data quality, large amounts of noise, missing and abnormal values of real vehicle data, which can seriously affect prediction accuracy when being directly used for model training, secondly, most of the methods only use limited and single-source direct measurement data (such as voltage, current and temperature) or simple statistics thereof as characteristics, and fail to fully mine deep and multi-time scale degradation associated information hidden in multi-source heterogeneous data (such as working condition fragments, user driving habits and historical environment sequences), and furthermore, the existing models are mostly 'black boxes', the prediction results are lack of physical interpretability, and when battery types, material systems or usage scenarios change, the models usually need to be re-collected for training, and the migration capability and the adaptability are insufficient. Mixing method in order to combine the advantages of the two methods described above, a mixing method is proposed. However, the fusion depth is different, and part of methods are only simple model series connection or result weighting, so that the deep coupling of a physical mechanism and a data model in a characteristic layer and a loss function layer cannot be realized. In addition, the existing prediction method mostly separates cyclic aging from calendar aging, and actual battery aging is the result of coupling action of the cyclic aging and calendar aging under complex working conditions, and separation simplification reduces the reality of prediction. In summary, the prior art has core problems of insufficient large data utilization, weak prediction model generalization and self-adaptation capability, insufficient consideration of actual complex coupling aging mechanism and the like. Therefore, there is an urgent need for a remaining useful life prediction method that can deeply fuse multi-source big data, extract multi-time scale degradation features, and introduce physical constraints to enhance model generalization and interpretation. Disclosure of Invention Based on the above object, the present invention provides a method for predicting remaining service life of a power battery based on big data analysis, the method comprising the steps of: Step 1, synchrono