Battery life prediction method and system of battery swap cabinet supporting big data analysis

Abstract

The application provides a battery life prediction method and system of a battery swap cabinet supporting big data analysis, relates to the technical field of big data analysis, and comprises the following steps: real-time monitoring of the battery swap cabinet battery through an Internet of Things sensing array and sensing interference correction; EMD decomposition of the battery monitoring sequence; life prediction of the battery swap cabinet battery according to the monitoring high-frequency component and the monitoring low-frequency component based on an equalization countermeasure learning channel; global fusion under certainty analysis according to the first residual life, the second residual life and the third residual life to obtain the fourth residual life. The application solves the technical problem that, in the prior art, it is difficult to extract comprehensive degradation characteristics due to abnormal fluctuation interference under complex working conditions, which further affects the battery life prediction accuracy of the battery swap cabinet. The application separates and accurately captures degradation characteristics of different scales from complex working condition data, thereby improving the battery life prediction accuracy of the battery swap cabinet.

Description

Technical Field

[0001] This application relates to the field of big data analytics technology, specifically to a method and system for predicting battery life in battery swapping cabinets that supports big data analytics. Background Technology

[0002] Currently, existing methods for predicting the remaining lifespan of batteries in battery swapping cabinets typically rely on simple statistical models or basic machine learning methods. When faced with abnormal fluctuations under complex operating conditions, these methods struggle to effectively distinguish between normal battery degradation processes and abnormal fluctuations caused by external disturbances. Battery degradation is a complex process spanning multiple time scales. Its early weak signals, mid-term cyclic decay, and long-term aging trends are easily obscured by sensor noise and complex usage conditions. Existing models designed to capture long-term trends may be insensitive to early high-frequency anomalies, while models focusing on short-term fluctuations struggle to grasp the overall degradation trajectory. This mismatch between model capabilities and degradation mechanisms leads to a lag in early warning of battery degradation. Furthermore, these methods lack sufficient generalization ability when faced with diverse factors such as different user habits and environmental temperature changes, making it difficult for traditional lifespan prediction methods to extract comprehensive and accurate degradation characteristics, thus affecting the accuracy of lifespan predictions.

[0003] In summary, existing technologies suffer from the technical problem that, under complex operating conditions, abnormal fluctuations can easily cause interference, making it difficult to extract comprehensive degradation characteristics and further affecting the accuracy of battery life prediction for battery swapping cabinets. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for predicting the battery life of battery swapping cabinets that supports big data analysis, in order to solve the technical problem in the prior art that the battery life prediction accuracy is affected by the difficulty in extracting comprehensive degradation characteristics due to the susceptibility to abnormal fluctuations under complex operating conditions.

[0005] To achieve the above objectives, this application provides a method and system for predicting the battery life of a battery swapping cabinet that supports big data analysis.

[0006] Firstly, this application provides a method for predicting the battery life of a battery swapping cabinet that supports big data analysis. This method is implemented through a battery life prediction system that supports big data analysis. The method includes: real-time monitoring of the battery in the swapping cabinet using an IoT sensor array to obtain battery monitoring data; performing sensor interference correction based on the battery monitoring data to obtain a battery monitoring sequence; performing EMD decomposition on the battery monitoring sequence to obtain a high-frequency monitoring component, a low-frequency monitoring component, and a monitoring trend component; predicting the battery life of the swapping cabinet based on the high-frequency monitoring component using an equalization adversarial learning channel to obtain a first remaining lifespan; predicting the battery life of the swapping cabinet based on the low-frequency monitoring component using the equalization adversarial learning channel to obtain a second remaining lifespan; predicting the battery life of the swapping cabinet based on the monitoring trend component using the equalization adversarial learning channel to obtain a third remaining lifespan; and performing global fusion under confidence analysis based on the first remaining lifespan, the second remaining lifespan, and the third remaining lifespan to obtain a fourth remaining lifespan.

[0007] Optionally, scene information is collected from the IoT sensor array based on each monitoring parameter in the battery monitoring data to obtain scene information for each sensor; anomaly detection is performed based on the scene information for each sensor to obtain anomaly detection results for each sensor; interference analysis is performed on the battery monitoring data based on the anomaly detection results for each sensor to obtain interference analysis results for each anomaly; adaptive correction is performed on the battery monitoring data based on the interference analysis results for each anomaly to generate the battery monitoring sequence.

[0008] Optionally, the IoT sensing array includes multiple IoT sensors.

[0009] Optionally, the battery monitoring sequence is traversed to extract a first monitoring signal; empirical mode decomposition is performed on the first monitoring signal to obtain a first intrinsic mode function component set and a first residual term; the first intrinsic mode function component set is classified according to frequency distribution characteristics to obtain a first high-frequency component set and a first low-frequency component set; amplitude normalization and time synchronization processing are performed on the first high-frequency component set to generate a first signal high-frequency component, and the first signal high-frequency component is incorporated into the monitoring high-frequency component; amplitude normalization and time synchronization processing are performed on the first low-frequency component set to generate a first signal low-frequency component, and the first signal low-frequency component is incorporated into the monitoring low-frequency component; amplitude normalization and time synchronization processing are performed on the first residual term to generate a first signal trend component, and the first signal trend component is incorporated into the monitoring trend component.

[0010] Optionally, a big data search is performed based on the battery model of the battery in the battery swapping cabinet to obtain multiple high-frequency component lifetime records. Each high-frequency component lifetime record includes historical high-frequency components and historical remaining lifetime. The balanced adversarial learning channel is activated, which includes a balanced checker, an adversarial example generator, and a learner. The learner is trained under supervision based on the multiple high-frequency component lifetime records to obtain a first high-frequency component lifetime prediction model. The first high-frequency component lifetime prediction model is optimized using balanced adversarial methods based on the balanced checker and the adversarial example generator to obtain a high-frequency component lifetime prediction channel. The monitored high-frequency components are input into the high-frequency component lifetime prediction channel to obtain the first remaining lifetime.

[0011] Optionally, the lifetime stages are classified according to the multiple high-frequency component lifetime records to obtain the sample distribution of each lifetime stage; the sample quantity balance is calculated according to the sample distribution of each lifetime stage to obtain the sample balance coefficient; the sample balance coefficient is input into the balance checker to obtain the balance deviation feature, the balance checker including the predetermined balance coefficient; based on the balance deviation feature, the multiple high-frequency component lifetime records are balanced and optimized according to the adversarial example generator to obtain the balanced optimized lifetime record set; the learner is supervised and trained according to the balanced optimized lifetime record set to obtain the second high-frequency component lifetime prediction model; the output fusion training is performed according to the first high-frequency component lifetime prediction model and the second high-frequency component lifetime prediction model to generate the high-frequency component lifetime prediction channel.

[0012] Optionally, the confidence level of the first remaining lifetime is evaluated based on the monitored high-frequency components to obtain a first lifetime confidence level; the confidence level of the second remaining lifetime is evaluated based on the monitored low-frequency components to obtain a second lifetime confidence level; the confidence level of the third remaining lifetime is evaluated based on the monitored trend components to obtain a third lifetime confidence level; a global percentage calculation is performed based on the first lifetime confidence level, the second lifetime confidence level, and the third lifetime confidence level to obtain each lifetime confidence weight; and the first remaining lifetime, the second remaining lifetime, and the third remaining lifetime are weighted and fused based on each lifetime confidence weight to generate the fourth remaining lifetime.

[0013] Optionally, a lifespan big data retrieval of batteries of the same model is performed based on the monitored high-frequency components to obtain a first retrieval remaining lifespan distribution; a confidence level evaluation is performed on each retrieval remaining lifespan within the first retrieval remaining lifespan distribution to obtain each lifespan confidence level; based on each lifespan confidence level, the first retrieval remaining lifespan distribution is cleaned according to a predetermined confidence level to obtain a first retrieval confidence lifespan distribution; a central tendency calculation is performed on the first retrieval confidence lifespan distribution to obtain the high-frequency component fitted remaining lifespan; and a cosine similarity calculation is performed on the first remaining lifespan and the high-frequency component fitted remaining lifespan to generate the first lifespan certainty level.

[0014] Optionally, the battery in the battery swapping cabinet is maintained and managed according to the fourth remaining lifespan.

[0015] Secondly, this application also provides a battery life prediction system for battery swapping cabinets supporting big data analysis, used to execute the battery life prediction method for battery swapping cabinets supporting big data analysis as described in the first aspect, wherein the battery life prediction system for battery swapping cabinets supporting big data analysis includes: a real-time monitoring module, used to monitor the battery of the battery swapping cabinet in real time through an Internet of Things sensor array, obtain battery monitoring data, and perform sensor interference correction based on the battery monitoring data to obtain a battery monitoring sequence; an EMD decomposition module, used to perform EMD decomposition on the battery monitoring sequence to obtain monitoring high-frequency components, monitoring low-frequency components, and monitoring trend components; and a first prediction module, used to... Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitored high-frequency components to obtain a first remaining lifespan; a second prediction module is used to predict the battery life of the battery swapping cabinet according to the monitored low-frequency components based on the balanced adversarial learning channel to obtain a second remaining lifespan; a third prediction module is used to predict the battery life of the battery swapping cabinet according to the monitored trend components based on the balanced adversarial learning channel to obtain a third remaining lifespan; a global fusion module is used to perform global fusion based on the first remaining lifespan, the second remaining lifespan, and the third remaining lifespan under confidence analysis to obtain a fourth remaining lifespan.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] The battery in the battery swapping cabinet is monitored in real time using an IoT sensor array to obtain battery monitoring data. Sensor interference correction is performed based on this data to obtain a battery monitoring sequence. EMD decomposition is then performed on the battery monitoring sequence to obtain high-frequency, low-frequency, and trend components. Based on an equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the high-frequency components to obtain a first remaining lifetime. Based on the equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the low-frequency components to obtain a second remaining lifetime. Based on the equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the trend components to obtain a third remaining lifetime. Finally, a global fusion based on confidence analysis is performed on the first, second, and third remaining lifetimes to obtain a fourth remaining lifetime. In other words, the battery monitoring data is cleaned and corrected through a sensor interference correction step, and the monitoring sequence is decomposed into high-frequency, low-frequency and trend components using EMD decomposition. The lifetime prediction is performed according to different components through an equalized adversarial learning channel. The confidence analysis of the prediction results of different components is performed, and the weights are calculated according to the confidence of each component. Finally, a fused fourth remaining lifetime is obtained, which improves the accuracy and robustness of battery lifetime prediction for battery swapping cabinets.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the battery life prediction method for battery swapping cabinets that supports big data analysis, as described in this application.

[0021] Figure 2 This is a schematic diagram of the battery life prediction system for a battery swapping cabinet that supports big data analysis, as described in this application.

[0022] Explanation of reference numerals in the attached figures: Real-time monitoring module 11, EMD decomposition module 12, first prediction module 13, second prediction module 14, third prediction module 15, global fusion module 16. Detailed Implementation

[0023] This application addresses the technical problem in existing technologies where battery life prediction for swapping cabinets is hampered by abnormal fluctuations under complex operating conditions, making it difficult to extract comprehensive degradation characteristics and further affecting the accuracy of battery life prediction. The method and system provide support for big data analysis to predict battery life in swapping cabinets. The battery monitoring data is cleaned and corrected through a sensor interference correction step. EMD decomposition is used to decompose the monitoring sequence into high-frequency, low-frequency, and trend components. Life prediction is performed separately for each component using an equalized adversarial learning channel. Confidence analysis is conducted on the prediction results for each component, and a weighted average is calculated based on the confidence level of each component to obtain a fused fourth remaining lifespan, thus improving the accuracy and robustness of battery life prediction for swapping cabinets.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a method for predicting the battery life of a battery swapping cabinet that supports big data analysis. The method is applied to a battery life prediction system for a battery swapping cabinet that supports big data analysis. The method specifically includes the following steps:

[0026] The battery in the battery swapping cabinet is monitored in real time using an IoT sensor array to obtain battery monitoring data. Based on the battery monitoring data, sensor interference correction is performed to obtain a battery monitoring sequence.

[0027] Furthermore, this application also includes the following steps: collecting scene information from the IoT sensor array based on each monitoring parameter in the battery monitoring data to obtain scene information for each sensor; performing anomaly detection based on the scene information for each sensor to obtain anomaly detection results for each sensor; performing interference analysis on the battery monitoring data based on the anomaly detection results for each sensor to obtain interference analysis results for each anomaly; and adaptively correcting the battery monitoring data based on the interference analysis results for each anomaly to generate the battery monitoring sequence.

[0028] Furthermore, this application also includes the following step: the IoT sensing array includes multiple IoT sensors.

[0029] Specifically, an IoT sensor array—comprising multiple IoT sensors—monitors the batteries within the battery swapping cabinet in real time. The collected battery monitoring data includes parameters such as battery voltage, current, temperature, and state of charge (SOC). The IoT sensor array is a collection of sensors designed with a specific layout, integrated within the battery swapping cabinet, and communicating via IoT technology. It is not a single sensor, but a network of multiple IoT sensors connected through wireless communication protocols. This network enables real-time acquisition of environmental or battery monitoring data and transmits this data to a central processing center.

[0030] Multiple IoT sensors deployed at key battery nodes in the battery swapping cabinet will start synchronously and perform measurements at a preset sampling frequency (e.g., 1Hz). Each IoT sensor performs its specific function. For example, the total voltage sensor continuously measures the voltage of the positive and negative terminals of the battery pack; the current sensor is connected in series in the main circuit to accurately measure the current flowing in (positive for charging) and flowing out (negative for discharging); and multiple temperature sensors are respectively attached to key hot spots such as the surface of the battery cells, the positive and negative electrode connecting pieces, and the battery pack casing. For example, for a battery A, the IoT sensor array starts working. At a specific discharge moment, the raw battery monitoring data reported by each sensor is as follows: the voltage sensor reports a voltage value of 49.15V, the current sensor reports a current value of -20.1A (the negative sign indicates that it is in a discharge state), the sensor located on the surface of battery cell 1 reports 36.8℃, the sensor located on the surface of battery cell 2 reports 35.2℃, the sensor located on the positive electrode connector reports 34.5℃, and the sensor located in the center of the battery pack casing reports 33.1℃. These discrete data points with timestamps, when continuously sampled, constitute the battery monitoring data used for subsequent analysis.

[0031] In addition to battery monitoring data, it is also necessary to collect scene information related to the sensors, including sensor status, sensor environment, and external interference factors around the battery. Scene information collection refers to collecting data on the sensor's own operating status and environmental conditions while collecting battery data, including the sensor's own status and external factors such as temperature, humidity, and electromagnetic interference.

[0032] Anomaly detection is performed based on information from various sensing scenarios to determine whether the sensor itself or the data it collects is in an abnormal state, identifying outliers in the monitored data. In other words, by using a predefined rule base, statistical model, or machine learning algorithm, information from various sensing scenarios is analyzed to determine whether the sensor is in an abnormal operating state or whether there is a drastic disturbance in its measurement environment. The anomaly detection results for each sensor are qualitative or quantitative descriptions of the abnormal state of each sensor, typically a structured judgment. For example, continuously monitoring the power supply voltage of each sensor and immediately triggering a power abnormality alarm if its value exceeds the specified range of [4.98V, 5.02V]; monitoring the internal chip temperature and triggering a sensor overheat alarm if it exceeds the safety threshold of 85℃ or if a temperature rise of more than 20℃ is detected within a short period; analyzing environmental electromagnetic noise indicators and marking current and voltage sensor data as having a high probability of transient interference risk if their intensity exceeds a critical value. Through these parallel rule sets built upon physical knowledge and historical experience, the operating status of each sensor is monitored in real time, and a structured report of sensor anomaly detection results is output, indicating which sensor is malfunctioning and what problem it is. For example, the current sensor's power supply voltage is 4.90V, which is abnormal and below the lower limit; the internal chip temperature is 48℃, which is normal; and the ambient electromagnetic noise intensity is 55dB, which is abnormal and too high. The anomaly detection results include sensor ID C1, anomaly type of power instability and electromagnetic interference, anomaly level of moderate, and confidence level of 90%. Before a sensor completely fails or provides seriously erroneous data, potential faults are detected in advance through its own status parameters, transforming passive handling into proactive early warning.

[0033] Interference analysis is performed on the battery monitoring data based on the anomaly detection results of each sensor to identify anomalies caused by external interference or sensor malfunctions, such as sensor noise, internal battery faults, or changes in the external environment. Based on the anomaly analysis results, the battery monitoring data is adaptively corrected by adjusting the parameters of the data analysis model to eliminate anomalies and ensure more accurate and reliable final battery monitoring data. For example, for power instability and electromagnetic interference anomalies in current sensor C1, the specific value of its supply voltage (4.90V) and the historical spectrum of electromagnetic noise are retrieved. Using a pre-set error model library, the anomaly analysis results show that the current reading under this condition has a -1% gain error and ±2A of random spike noise. For the current data, gain compensation is first performed (reading / 0.99), and then a median filter with a window size of 5 sampling points is activated to eliminate random spikes. The original current data reported as -20.5A, accompanied by a momentary spike of -23.0A. By dividing -20.5A by 0.99, we get -20.71A. Within the window containing the spike [-20.7, -20.8, -23.0, -20.6, -20.7], the median value is -20.7A. This value is used to replace the -23.0A spike, resulting in a final output current of -20.7A.

[0034] By collecting scene information and detecting anomalies, the system effectively identifies and eliminates data anomalies caused by external environmental interference or sensor malfunctions, ensuring the accuracy of the final battery monitoring data. An adaptive correction mechanism enables the monitoring data to eliminate interference from abnormal fluctuations through real-time dynamic adjustments, avoiding the limitations of traditional methods in dealing with such interference. Through comprehensive collection and correction of scene information, the system adapts to changes in different environments, usage conditions, and equipment states, enhancing the robustness and adaptability of the entire battery monitoring and prediction system.

[0035] The battery monitoring sequence is decomposed using EMD to obtain the high-frequency monitoring component, the low-frequency monitoring component, and the monitoring trend component.

[0036] Furthermore, this application also includes the following steps: traversing the battery monitoring sequence to extract a first monitoring signal; performing empirical mode decomposition on the first monitoring signal to obtain a first intrinsic mode function component set and a first residual term; classifying the first intrinsic mode function component set according to frequency distribution characteristics to obtain a first high-frequency component set and a first low-frequency component set; performing amplitude normalization and time synchronization processing on the first high-frequency component set to generate a first signal high-frequency component, and incorporating the first signal high-frequency component into the monitoring high-frequency component; performing amplitude normalization and time synchronization processing on the first low-frequency component set to generate a first signal low-frequency component, and incorporating the first signal low-frequency component into the monitoring low-frequency component; performing amplitude normalization and time synchronization processing on the first residual term to generate a first signal trend component, and incorporating the first signal trend component into the monitoring trend component.

[0037] Specifically, a monitoring signal is randomly extracted from the battery monitoring sequence as the first monitoring signal. Empirical Mode Decomposition (EMD) is performed on the first monitoring signal, an iterative sieving process that continuously extracts the highest frequency oscillations in the signal until the remaining signal becomes monotonic, thus outputting a set of first intrinsic mode function (IMF) components arranged from high to low frequency and a first residual term. Empirical Mode Decomposition (EMD) is an adaptive signal processing method suitable for analyzing non-stationary and nonlinear data. It decomposes complex signals into a series of IMFs with different oscillation frequencies from high to low frequency and a residual term representing the overall trend. The first IMF component set is a collection of multiple IMF components generated by EMD decomposition. Each IMF component must satisfy two conditions: the number of extrema is equal to or differs from the number of zero-crossings by at most one; and at any point, the mean of the envelope defined by local maxima and minima is zero. They represent oscillation modes at different time scales in the original signal. The first residual term is the last component obtained from the decomposition, usually representing the remaining trend or low-frequency components in the signal, representing the long-term, slowly changing trend components in the signal.

[0038] Based on the frequency distribution characteristics of the signal, the obtained first intrinsic mode function component set is classified into a first high-frequency component set and a first low-frequency component set. The intrinsic mode function components with shorter average periods are assigned to the first high-frequency component set, representing the rapid dynamics during the charging and discharging process; the intrinsic mode function components with longer average periods are assigned to the first low-frequency component set, representing the capacity decay fluctuations with a period of dozens of cycles.

[0039] To eliminate dimensions, amplitude normalization and time synchronization are performed on the first high-frequency component set, the first low-frequency component set, and the first residual term, respectively. By adjusting the amplitude range of the signal, the amplitude differences between different components are unified. Simultaneously, time synchronization is performed to ensure time alignment of each component, resulting in the first signal high-frequency component, the first signal low-frequency component, and the first signal trend component. These are then incorporated into the global monitoring high-frequency component, monitoring low-frequency component, and monitoring trend component sets, respectively. High-frequency components typically represent rapid changes or noise, low-frequency components represent long-term trends, and trend components represent long-term stable changes.

[0040] For example, taking the cycle test of battery M as an example, its initial capacity is 20.0 Ah. The data is traversed, and the capacity of each complete discharge cycle is extracted as the first monitoring signal, forming a capacity decay sequence: 20.00 Ah, 19.98 Ah, 19.97 Ah, 19.95 Ah, etc., recording the capacity decay data of the first 100 cycles. EMD decomposition is performed to obtain the first intrinsic mode function component set. IMF1 is a high-frequency oscillation with an amplitude of approximately ±0.02 Ah and a period of 2-5 cycles, reflecting random fluctuations in charging and discharging; IMF2 is an oscillation with an amplitude of approximately ±0.06 Ah and a period of 15-25 cycles, reflecting periodic capacity fluctuations. The first residual term shows a monotonically decreasing trend, with the capacity decaying from 19.96 Ah to 19.47 Ah, demonstrating an overall linear decay trend from 20.0 Ah to 19.47 Ah. IMF1 is assigned to the first high-frequency component set; IMF2 is assigned to the first low-frequency component set. The high-frequency component set (IMF1) is normalized to obtain values ​​of 0.82, 0.45, 0.91, ..., 0.38; the low-frequency component set (IMF2) is normalized and scaled to the [0,1] interval to obtain values ​​of 0.62, 0.65, 0.68, ..., 0.12; the residuals (trends) are normalized and scaled to the [0,1] interval to obtain values ​​of 1.00, 0.99, 0.98, ..., 0.02; all components retain the same timestamps as the original capacity sequence. Using the Min-Max normalization formula, the amplitude of the high-frequency component is ±0.02Ah, and the rapid fluctuation characteristics are preserved after normalization; the amplitude of the low-frequency component is ±0.03Ah to +0.06Ah, and the periodic changes are highlighted after normalization; the trend component decays from 19.96Ah to 19.47Ah, and the normalization shows a smooth downward curve.

[0041] EMD decomposition separates the high-frequency, low-frequency, and trend components in the battery signal. Through amplitude normalization and time synchronization, it ensures the consistency and comparability of each component at different time points, eliminating interference caused by amplitude differences or time inconsistencies. The EMD method is entirely driven by the data itself, unlike Fourier transform which requires pre-setting sinusoidal basis functions. Therefore, it can better handle non-stationary and nonlinear signals such as battery degradation, and the extracted features are more representative and robust.

[0042] Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitored high-frequency components to obtain the first remaining life.

[0043] Furthermore, this application also includes the following steps: performing big data retrieval based on the battery model of the battery in the battery swapping cabinet to obtain multiple high-frequency component lifetime records, each high-frequency component lifetime record including historical high-frequency components and historical remaining lifetime; activating the balanced adversarial learning channel, the balanced adversarial learning channel including a balanced checker, an adversarial example generator, and a learner; performing supervised training on the learner based on the multiple high-frequency component lifetime records to obtain a first high-frequency component lifetime prediction model; performing balanced adversarial optimization on the first high-frequency component lifetime prediction model based on the balanced checker and the adversarial example generator to obtain a high-frequency component lifetime prediction channel; inputting the monitored high-frequency components into the high-frequency component lifetime prediction channel to obtain the first remaining lifetime.

[0044] Furthermore, this application also includes the following steps: classifying lifetime stages according to the multiple high-frequency component lifetime records to obtain sample distributions for each lifetime stage; calculating sample balance based on the sample distributions for each lifetime stage to obtain sample balance coefficients; inputting the sample balance coefficients into the balance checker to obtain balance deviation features, wherein the balance checker includes predetermined balance coefficients; based on the balance deviation features, performing balance optimization on the multiple high-frequency component lifetime records according to the adversarial example generator to obtain a balanced optimized lifetime record set; performing supervised training on the learner based on the balanced optimized lifetime record set to obtain a second high-frequency component lifetime prediction model; and performing output fusion training based on the first high-frequency component lifetime prediction model and the second high-frequency component lifetime prediction model to generate the high-frequency component lifetime prediction channel.

[0045] Specifically, based on battery model, a big data search is performed to extract tens of thousands of high-frequency component lifespan records for the same battery model from the database, forming multiple high-frequency component lifespan records. The battery model in the battery swapping cabinet is the battery's specification identifier; batteries of the same model have the same chemical system, structure, and manufacturing process, and their aging patterns are similar. By analyzing a large amount of historical data, especially historical high-frequency component and remaining lifespan records related to specific battery models, characteristics and patterns related to battery lifespan prediction are extracted. High-frequency component lifespan records are historical data records related to the battery's high-frequency components. During battery degradation, high-frequency components typically represent rapidly fluctuating parts of the battery; these fluctuations may be related to short-term performance degradation or external environmental influences. Each record includes historical high-frequency components and their corresponding remaining lifespan.

[0046] The balanced adversarial learning pathway comprises three core components: a balance checker, an adversarial example generator, and a learner. The balance checker evaluates the performance of the trained model, ensuring it can adapt to different input data and avoiding overfitting. The adversarial example generator produces adversarial examples based on existing training data to enhance the model's generalization ability. The learner, the core prediction model (e.g., LSTM or GRU), is responsible for learning the mapping relationship from high-frequency components to remaining lifetime. The learner is trained under supervision using multiple high-frequency component lifetime records. Through known high-frequency components and their corresponding remaining lifetime data, the learner learns the patterns of battery degradation and forms a preliminary lifetime prediction model, thus obtaining the first high-frequency component lifetime prediction model.

[0047] Multiple high-frequency component lifetime records are input into the framework, and the learner within the framework begins supervised training. It reads each historical high-frequency component sequence, calculates a predicted remaining lifetime through the network, and then compares this predicted value with the actual historical remaining lifetime labels. Based on the calculated loss, i.e., the error, it continuously adjusts its weights and bias parameters through the backpropagation algorithm. After multiple rounds of iterative training, the learner gradually masters the ability to infer the battery degradation state from high-frequency signal fluctuation patterns, thus obtaining a preliminary first high-frequency component lifetime prediction model.

[0048] Based on multiple high-frequency component lifetime records, lifetime stages are classified into several lifetime stages, including the initial lifetime stage, the stable lifetime stage, and the decay lifetime stage. This yields the sample distribution for each lifetime stage, which is the distribution of high-frequency component lifetime records collected under different lifetime stages. The number of data samples in each lifetime stage may vary.

[0049] The sample distribution for each lifespan stage is statistically analyzed, and a sample balance calculation is performed to obtain a sample balance coefficient, which determines the degree of imbalance in sample size across different lifespan stages. For example, this could be the ratio of the number of samples in each stage to the maximum number of samples. This sample balance coefficient is input into a balance checker and compared with a predetermined balance coefficient to obtain a balance deviation characteristic, i.e., determining the difference between the current data distribution and the ideal distribution. If a stage has too many or too few samples, the balance checker will output the balance deviation characteristic, indicating the imbalance of the current dataset. The balance checker includes a predetermined balance coefficient, which is an idealized target, typically expected to be 1 for all lifespan stages, indicating perfect sample balance.

[0050] Based on the equilibrium deviation characteristics, an adversarial example generator is used to generate optimized samples, balancing the number of samples across different lifetime stages and preventing model bias due to data imbalance. The generation of adversarial examples ensures that the number of data samples across all stages tends towards balance, thereby improving the model's predictive accuracy. For stages with sparse samples, the adversarial example generator learns the characteristics of real data from that stage to generate a large number of realistic high-frequency component sequences, thus obtaining a balanced optimized lifetime record set. For example, assuming there are 40,000 samples in the initial stage, 55,000 in the stable stage, and 5,000 in the decline stage, the sample equilibrium coefficients are calculated to be 1.0, 1.375, and 0.125, respectively. The equilibrium checker finds that the equilibrium coefficient of 0.125 in the decline stage deviates significantly from the predetermined value of 1.0. The adversarial example generator is then instructed to generate data for the decline stage, producing 40,000 high-frequency component sequences that match the characteristics of the decline stage, filling the data gaps and forming a balanced optimized lifetime record set. The number of decline stage samples increases from 5,000 to 45,000, and the distribution tends towards equilibrium.

[0051] The learner is trained under supervision using a balanced optimized lifespan record set to obtain a second high-frequency component lifespan prediction model, which enhances its predictive ability during periods of scarcity. The training process for the second high-frequency component lifespan prediction model is similar to that of the first model, but the difference lies in learning the patterns of battery lifespan based on the optimized dataset for accurate prediction. The outputs of the first and second high-frequency component lifespan prediction models are fused and trained to combine their advantages. By fusing the outputs of different models, a comprehensive and robust high-frequency component lifespan prediction channel is generated, fully considering the degradation patterns of batteries at different usage stages and outputting more accurate lifespan prediction results. For example, assuming monitoring of a battery in a battery swapping cabinet with battery model X200, the separated high-frequency component data includes 0.75, 0.72, 0.78, 0.71, and 0.74, a normalized high-frequency sequence of the last five cycles. The volatility is slightly increased compared to when the battery is new, and the first remaining lifespan prediction is 325 cycles. The abnormal fluctuations in the high-frequency signal detected early signs of recession faster than the macro trend, thus predicting a relatively short remaining lifespan and providing a conservative early warning value that focuses on near-term risks.

[0052] When it is necessary to predict the battery life of the battery swapping cabinet, the decomposed high-frequency monitoring components are input into the finally determined high-frequency component life prediction channel to obtain the first remaining life. The first remaining life is the predicted value of the battery's remaining life based on the high-frequency component life prediction channel. By introducing an equalization adversarial learning channel, the robustness and accuracy of the battery life prediction model are effectively improved. Through supervised training of the high-frequency component life records, a preliminary prediction model is generated, and the problems of data imbalance and noise interference are solved through equalization and adversarial optimization, thereby improving the model's prediction accuracy.

[0053] Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitored low-frequency components to obtain a second remaining life.

[0054] Specifically, similarly to the process of acquiring high-frequency component data, a big data search is performed based on the battery model of the battery in the battery swapping cabinet to obtain multiple low-frequency component lifespan records. Each low-frequency component lifespan record includes historical low-frequency components and historical remaining lifespan. These multiple low-frequency component lifespan records are derived from the long-term usage of the battery, such as changes in long-term trends in battery temperature, aging degree, and charge / discharge cycles.

[0055] The optimized balanced adversarial learning pathway is activated, comprising a balance checker, an adversarial example generator, and a learner. The learner is trained based on multiple low-frequency component lifetime records, a process similar to that of high-frequency component data, but focusing on the long-term degradation trend of the battery. The balance checker examines the balance of the low-frequency component data, ensuring a reasonable distribution of data across lifetime stages and avoiding the impact of data imbalance. The adversarial example generator generates balanced and challenging adversarial examples based on the current distribution of low-frequency component data across lifetime stages, thereby enhancing the lifetime prediction model's adaptability to different degradation patterns. The learner undergoes supervised training based on the low-frequency component data, establishing a battery lifetime prediction model that monitors the low-frequency components. Multiple low-frequency component lifetime records extracted from historical data are input into the learner for supervised training. The goal of supervised training is to enable the model to predict the remaining battery lifetime using long-term varying low-frequency component data. Based on supervised training, the balance checker and adversarial example generator are used to optimize the model. The balance checker examines the balance of the training data and optimizes based on balance deviation characteristics. Adversarial example generators enhance a model's ability to handle complex and disruptive data by generating adversarial examples.

[0056] Real-time acquired low-frequency component data is input into an optimized equalization adversarial learning channel. An optimized second remaining lifetime prediction model then predicts the battery's second remaining lifetime. For example, assuming monitoring of a battery in a battery swapping cabinet (model X200), the separated low-frequency component data includes normalized low-frequency sequences of 0.65, 0.64, 0.63, 0.65, and 0.62 over five cycles, showing a stable, slow decreasing trend. The predicted second remaining lifetime is 310 cycles. Predicting based on a relatively stable, medium-term decay rate filters out short-term noise, resulting in a smoother prediction compared to the high-frequency channel, reflecting the battery's health status over the medium term. The low-frequency component lifetime prediction channel focuses on learning the battery's decay rate and stability from periodic fluctuations. Its prediction results reflect the battery's healthy evolution trend over the medium term, complementing the short-term fluctuations of the high-frequency channel.

[0057] Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitoring trend component to obtain the third remaining life.

[0058] Specifically, similarly, the trend component data of a battery obtained from a battery monitoring system typically comes from a long-term battery usage history and represents the battery's long-term health status. For example, suppose the battery's trend component includes long-term temperature changes, such as the long-term average temperature change during battery use, or the number of battery charge-discharge cycles, such as the degradation trend exhibited by the battery after multiple charge-discharge cycles.

[0059] The equilibrium adversarial learning channel is activated by inputting the monitored trend component data. Similar to the processing of high-frequency and low-frequency component data, the equilibrium checker examines whether the input trend component data is balanced, outputting equilibrium deviation features to indicate whether data imbalance exists. The adversarial example generator generates adversarial examples based on the current data to improve the model's robustness in complex data environments. The learner uses the trend component data for supervised learning to construct a battery life prediction model corresponding to the trend component data.

[0060] The extracted historical trend component lifetime records are input into the learner for supervised training. Using the known trend component data and corresponding remaining lifetime data, the model is trained to predict the remaining battery life. The trained model then undergoes an balance checker to examine the data balance and outputs balance deviation features. If data imbalance is found, the balance checker proposes optimization solutions. An adversarial example generator produces adversarial examples to challenge the model's predictive ability under different battery usage scenarios, thereby improving the model's robustness. Through these optimizations, the model is tuned into a more generalized battery life prediction model, better able to handle various possible battery usage conditions.

[0061] Real-time trend component data is input into an optimized equalization adversarial learning channel. The optimized model then predicts the remaining battery life, yielding the third remaining lifespan. For example, assuming monitoring of a battery in a battery swapping cabinet (model X200), the separated monitoring trend components are 0.95, 0.94, 0.93, 0.925, and 0.92. The long-term normalized degradation trend from cycle 0 to the present is currently in the linear degradation region, and the predicted third remaining lifespan is 295 cycles. Extrapolation is performed based on the overall degradation slope of the battery from its new state to the present. Because it does not consider medium-term fluctuations and short-term noise, its prediction results are usually the most objective, indicating the end of lifespan under ideal conditions. The third remaining lifespan is the remaining battery lifespan predicted through an equalization adversarial learning channel based on the monitored battery trend components, i.e., the battery's long-term degradation trend data. Unlike the first and second remaining lifespans, the third remaining lifespan focuses more on the long-term degradation process of the battery. These three predicted values ​​(325, 310, and 295) constitute a multi-scale prediction view. The first remaining lifetime (325) is greater than the third remaining lifetime (295), which is a typical and reasonable phenomenon. This indicates that the early decay detected by the high-frequency signal has not yet been fully reflected in the macro trend, and plays an early warning role.

[0062] A fourth remaining lifetime is obtained by global fusion of the first remaining lifetime, the second remaining lifetime, and the third remaining lifetime under confidence analysis.

[0063] Furthermore, this application also includes the following steps: evaluating the confidence level of the first remaining lifetime based on the monitored high-frequency components to obtain a first lifetime confidence level; evaluating the confidence level of the second remaining lifetime based on the monitored low-frequency components to obtain a second lifetime confidence level; evaluating the confidence level of the third remaining lifetime based on the monitored trend components to obtain a third lifetime confidence level; calculating the global proportion of the first lifetime confidence level, the second lifetime confidence level, and the third lifetime confidence level to obtain each lifetime confidence weight; and weighting and fusing the first remaining lifetime, the second remaining lifetime, and the third remaining lifetime based on each lifetime confidence weight to generate the fourth remaining lifetime.

[0064] Furthermore, this application also includes the following steps: performing a lifespan big data retrieval of batteries of the same model based on the monitored high-frequency components to obtain a first retrieval remaining lifespan distribution; evaluating the confidence level of each retrieval remaining lifespan within the first retrieval remaining lifespan distribution to obtain each lifespan confidence level; cleaning the first retrieval remaining lifespan distribution based on each lifespan confidence level according to a predetermined confidence level to obtain a first retrieval confidence lifespan distribution; performing central tendency calculation based on the first retrieval confidence lifespan distribution to obtain the high-frequency component fitted remaining lifespan; and performing cosine similarity calculation on the first remaining lifespan and the high-frequency component fitted remaining lifespan to generate the first lifespan certainty level.

[0065] Specifically, based on the monitored high-frequency components, a large-scale lifespan retrieval of batteries of the same model is performed. Battery records with similar high-frequency fluctuation patterns are searched in the historical database to obtain a first-level remaining lifespan distribution containing the remaining lifespan values ​​of multiple similar records. This first-level remaining lifespan distribution, obtained through the large-scale data retrieval, shows how the remaining lifespan of the batteries is distributed under different high-frequency component conditions, representing the remaining lifespan data of all batteries of the same model related to the high-frequency components.

[0066] The confidence level of each remaining search lifetime within the first remaining search lifetime distribution is evaluated to obtain the confidence level of each lifetime. Each lifetime confidence level represents the frequency of each remaining search lifetime occurring within the first remaining search lifetime distribution. For example, the first remaining search lifetime distribution might be 310 cycles, 315 cycles, 320 cycles, 320 cycles, 325 cycles, 325 cycles, 330 cycles, 335 cycles, 340 cycles, 295 cycles, and 350 cycles. The lifetime confidence levels for the rings are as follows: 1 / 12 for 295 cycles, 1 / 12 for 310 cycles, 1 / 12 for 315 cycles, 1 / 6 for 320 cycles, 1 / 4 for 325 cycles, 1 / 12 for 330 cycles, 1 / 12 for 335 cycles, 1 / 12 for 340 cycles, and 1 / 12 for 350 cycles.

[0067] Based on each lifetime confidence level, the first retrieval remaining lifetime distribution is cleaned according to a predetermined confidence level, i.e., outliers below the predetermined confidence level are removed to obtain the first retrieval confidence lifetime distribution. The predetermined confidence level is a preset threshold, such as 0.1, used to filter out outlier lifetime values ​​that occur too infrequently. Remaining lifetime data below the predetermined confidence level will be considered unreliable or invalid and thus discarded.

[0068] The data in the first retrieval confidence lifetime distribution are subjected to ensemble value calculation to obtain a representative remaining lifetime value. The ensemble value calculation is usually the average value. For example, after removing all lifetime values ​​with a frequency below 0.1, the first retrieval confidence lifetime distribution is obtained as 320, 320, 325, 325, 325. The ensemble value calculation yields a remaining lifetime of 323 cycles for the high-frequency components.

[0069] The cosine similarity is calculated for the first remaining lifetime and the high-frequency component fitted remaining lifetime. This treats the two lifetime values ​​as vectors and calculates their directional consistency, generating a first lifetime confidence score between 0 and 1. A value closer to 1 indicates a stronger consistency between the model prediction and historical statistical patterns. For example, if the first remaining lifetime is 325 and the high-frequency component fitted remaining lifetime is 323, the calculated cosine similarity is 0.9999. Cosine similarity measures the directional similarity between two scalar values; in practice, it can be considered as calculating their consistency with an ideal reference direction.

[0070] Similarly, the aforementioned steps are repeated for monitoring low-frequency components and trend components respectively to obtain the second lifetime confidence and the third lifetime confidence, quantifying the reliability of the current prediction results for each channel. A global percentage calculation is performed based on the first lifetime confidence, second lifetime confidence, and third lifetime confidence to obtain the lifetime confidence weight. The confidence of each channel is divided by the sum of the three confidences to obtain the lifetime confidence weight. First lifetime confidence weight = First lifetime confidence / (First lifetime confidence + Second lifetime confidence + Third lifetime confidence), Second lifetime confidence weight = Second lifetime confidence / (First lifetime confidence + Second lifetime confidence + Third lifetime confidence), Third lifetime confidence weight = Third lifetime confidence / (First lifetime confidence + Second lifetime confidence + Third lifetime confidence).

[0071] The first remaining lifetime, second remaining lifetime, and third remaining lifetime are weighted and fused according to their respective lifetime confidence weights to obtain the fourth remaining lifetime. That is, the fourth remaining lifetime = first remaining lifetime * first lifetime confidence weight + second remaining lifetime * second lifetime confidence weight + third remaining lifetime * third lifetime confidence weight. For example, assuming the total confidence score is 0.85 + 0.95 + 0.90 = 2.70, then the first lifetime confidence weight is 0.315, the second lifetime confidence weight is 0.352, and the third lifetime confidence weight is 0.333. Therefore, the fourth remaining lifetime is calculated as 325 * 0.315 + 310 * 0.352 + 295 * 0.333 = 310 cycles.

[0072] By combining the prediction results of different components with their corresponding confidence levels and employing a weighted fusion method, more accurate and reliable battery remaining life prediction results can be obtained. Through confidence level evaluation and global proportion calculation, the reasonable weight of each prediction result in the comprehensive prediction is ensured, resulting in higher accuracy and credibility of the final fourth remaining life prediction.

[0073] Furthermore, this application also includes the following steps: maintaining and managing the battery in the battery swapping cabinet based on the fourth remaining lifespan.

[0074] Specifically, the battery management system assesses the current health status of the battery based on its fourth remaining lifespan to determine whether maintenance, replacement, or optimized usage strategies are needed. If the fourth remaining lifespan is low, maintenance or replacement should be initiated in advance. A periodic inspection plan is established based on the predicted remaining lifespan. If the predicted fourth remaining lifespan is low, an early warning is issued, requiring staff to inspect the battery and perform necessary maintenance. When the battery's fourth remaining lifespan approaches a predetermined lifespan threshold, the battery management system can automatically trigger battery replacement or charging. Depending on battery usage and predicted lifespan, the battery can be transferred to a charging cabinet or directly scheduled for replacement. Maintenance strategies are dynamically adjusted based on the fourth remaining lifespan of different batteries and actual workload. For example, for batteries with longer remaining lifespans, the inspection cycle can be extended; while for batteries with shorter remaining lifespans, the inspection frequency can be increased or replacement can be initiated earlier. Based on accurate lifespan predictions, it is determined when batteries need maintenance or replacement, avoiding equipment downtime or safety hazards caused by battery failure or excessive aging.

[0075] In summary, the battery life prediction method for battery swapping cabinets that supports big data analysis provided in this application has the following technical advantages:

[0076] The battery in the battery swapping cabinet is monitored in real time using an IoT sensor array to obtain battery monitoring data. Sensor interference correction is performed based on this data to obtain a battery monitoring sequence. EMD decomposition is then performed on the battery monitoring sequence to obtain high-frequency, low-frequency, and trend components. Based on an equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the high-frequency components to obtain a first remaining lifetime. Based on the equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the low-frequency components to obtain a second remaining lifetime. Based on the equalization adversarial learning channel, the battery life of the swapping cabinet is predicted according to the trend components to obtain a third remaining lifetime. Finally, a global fusion based on confidence analysis is performed on the first, second, and third remaining lifetimes to obtain a fourth remaining lifetime. In other words, the battery monitoring data is cleaned and corrected through a sensor interference correction step, and the monitoring sequence is decomposed into high-frequency, low-frequency and trend components using EMD decomposition. The lifetime prediction is performed according to different components through an equalized adversarial learning channel. The confidence analysis of the prediction results of different components is performed, and the weights are calculated according to the confidence of each component. Finally, a fused fourth remaining lifetime is obtained, which improves the accuracy and robustness of battery lifetime prediction for battery swapping cabinets.

[0077] Example 2: Based on the same inventive concept as the battery life prediction method for battery swapping cabinets supporting big data analysis in Example 1, this application also provides a battery life prediction system for battery swapping cabinets supporting big data analysis. Please refer to the appendix. Figure 2The battery life prediction system for battery swapping cabinets that supports big data analysis includes:

[0078] The real-time monitoring module 11 is used to monitor the battery in the battery swapping cabinet in real time through an IoT sensor array, obtain battery monitoring data, and perform sensor interference correction based on the battery monitoring data to obtain a battery monitoring sequence; the EMD decomposition module 12 is used to perform EMD decomposition on the battery monitoring sequence to obtain monitoring high-frequency components, monitoring low-frequency components, and monitoring trend components; the first prediction module 13 is used to predict the battery life of the battery swapping cabinet based on the monitoring high-frequency components using an equalization adversarial learning channel to obtain a first remaining lifespan; the second prediction module 14 is used to predict the battery life of the battery swapping cabinet based on the monitoring low-frequency components using the equalization adversarial learning channel to obtain a second remaining lifespan; the third prediction module 15 is used to predict the battery life of the battery swapping cabinet based on the monitoring trend components using the equalization adversarial learning channel to obtain a third remaining lifespan; and the global fusion module 16 is used to perform global fusion based on the first remaining lifespan, the second remaining lifespan, and the third remaining lifespan under confidence analysis to obtain a fourth remaining lifespan.

[0079] Furthermore, the real-time monitoring module 11 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used to: collect scene information from the IoT sensor array according to each monitoring parameter in the battery monitoring data to obtain scene information for each sensor; perform anomaly detection based on the scene information for each sensor to obtain anomaly detection results for each sensor; perform interference analysis on the battery monitoring data based on the anomaly detection results for each sensor to obtain interference analysis results for each anomaly; and adaptively correct the battery monitoring data based on the interference analysis results for each anomaly to generate the battery monitoring sequence.

[0080] Furthermore, the real-time monitoring module 11 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used for: the IoT sensor array includes multiple IoT sensors.

[0081] Furthermore, the EMD decomposition module 12 in the battery life prediction system for the battery swapping cabinet supporting big data analysis is also used for: traversing the battery monitoring sequence to extract a first monitoring signal; performing empirical mode decomposition on the first monitoring signal to obtain a first intrinsic mode function component set and a first residual term; classifying the first intrinsic mode function component set according to frequency distribution characteristics to obtain a first high-frequency component set and a first low-frequency component set; performing amplitude normalization and time synchronization processing on the first high-frequency component set to generate a first signal high-frequency component, and incorporating the first signal high-frequency component into the monitoring high-frequency component; performing amplitude normalization and time synchronization processing on the first low-frequency component set to generate a first signal low-frequency component, and incorporating the first signal low-frequency component into the monitoring low-frequency component; performing amplitude normalization and time synchronization processing on the first residual term to generate a first signal trend component, and incorporating the first signal trend component into the monitoring trend component.

[0082] Furthermore, the first prediction module 13 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used for: performing big data retrieval based on the battery model of the battery swapping cabinet to obtain multiple high-frequency component life records, each high-frequency component life record including historical high-frequency components and historical remaining life; activating the balanced adversarial learning channel, the balanced adversarial learning channel including a balanced checker, an adversarial sample generator, and a learner; performing supervised training on the learner based on the multiple high-frequency component life records to obtain a first high-frequency component life prediction model; performing balanced adversarial optimization on the first high-frequency component life prediction model based on the balanced checker and the adversarial sample generator to obtain a high-frequency component life prediction channel; and inputting the monitored high-frequency components into the high-frequency component life prediction channel to obtain the first remaining life.

[0083] Furthermore, the first prediction module 13 in the battery life prediction system for the battery swapping cabinet supporting big data analysis is also used for: classifying life stages according to the multiple high-frequency component life records to obtain sample distributions for each life stage; performing sample quantity balancing calculations according to the sample distributions for each life stage to obtain sample balancing coefficients; inputting the sample balancing coefficients into the balancing checker to obtain balancing deviation features, wherein the balancing checker includes predetermined balancing coefficients; based on the balancing deviation features, performing balancing optimization on the multiple high-frequency component life records according to the adversarial example generator to obtain a balancing optimized life record set; performing supervised training on the learner according to the balancing optimized life record set to obtain a second high-frequency component life prediction model; and performing output fusion training based on the first high-frequency component life prediction model and the second high-frequency component life prediction model to generate the high-frequency component life prediction channel.

[0084] Furthermore, the global fusion module 16 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used to: evaluate the confidence level of the first remaining lifespan based on the monitored high-frequency components to obtain a first lifespan confidence level; evaluate the confidence level of the second remaining lifespan based on the monitored low-frequency components to obtain a second lifespan confidence level; evaluate the confidence level of the third remaining lifespan based on the monitored trend components to obtain a third lifespan confidence level; calculate the global proportion of the first lifespan confidence level, the second lifespan confidence level, and the third lifespan confidence level to obtain each lifespan confidence weight; and perform weighted fusion of the first remaining lifespan, the second remaining lifespan, and the third remaining lifespan based on each lifespan confidence weight to generate the fourth remaining lifespan.

[0085] Furthermore, the global fusion module 16 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used for: performing a life big data retrieval of batteries of the same model based on the monitored high-frequency components to obtain a first retrieval remaining life distribution; evaluating the confidence level of each retrieval remaining life within the first retrieval remaining life distribution to obtain each life confidence level; cleaning the first retrieval remaining life distribution based on each life confidence level according to a predetermined confidence level to obtain a first retrieval confidence life distribution; performing central tendency calculation based on the first retrieval confidence life distribution to obtain the high-frequency component fitted remaining life; and performing cosine similarity calculation on the first remaining life and the high-frequency component fitted remaining life to generate the first life certainty level.

[0086] Furthermore, the global fusion module 16 in the battery life prediction system for the battery swapping cabinet that supports big data analysis is also used to: perform maintenance management on the battery swapping cabinet battery based on the fourth remaining life.

[0087] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1 The battery life prediction method and specific examples for battery swapping cabinets that support big data analysis in Example 1 are also applicable to the battery life prediction system for battery swapping cabinets that support big data analysis in this example. Through the foregoing detailed description of the battery life prediction method for battery swapping cabinets that support big data analysis, those skilled in the art can clearly understand the battery life prediction system for battery swapping cabinets that support big data analysis in this example. Therefore, for the sake of brevity, it will not be described in detail here.

[0088] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0089] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting battery life in battery swapping cabinets that supports big data analysis, characterized in that: include: The battery in the battery swapping cabinet is monitored in real time by an IoT sensor array to obtain battery monitoring data. Based on the battery monitoring data, sensor interference correction is performed to obtain a battery monitoring sequence. The battery monitoring sequence was decomposed using EMD to obtain the high-frequency monitoring component, the low-frequency monitoring component, and the monitoring trend component. Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitored high-frequency components to obtain the first remaining life, including: Based on the battery model of the battery in the battery swapping cabinet, a big data search is performed to obtain multiple high-frequency component lifetime records. Each high-frequency component lifetime record includes historical high-frequency components and historical remaining lifetime. Activate the balanced adversarial learning channel, which includes a balanced checker, an adversarial sample generator, and a learner; The learner is trained under supervision based on the multiple high-frequency component lifetime records to obtain a first high-frequency component lifetime prediction model. The first high-frequency component lifetime prediction model is optimized by equalization and adversarial optimization based on the equalization checker and the adversarial example generator to obtain the high-frequency component lifetime prediction channel, including: Based on the multiple high-frequency component lifetime records, lifetime stages are classified to obtain the sample distribution of each lifetime stage. Based on the sample distribution of each lifespan stage, a sample balance calculation is performed to obtain the sample balance coefficient. The sample balance coefficient is input into the balance tester to obtain the balance deviation characteristics. The balance tester includes a predetermined balance coefficient. Based on the aforementioned balance deviation characteristics, the multiple high-frequency component lifetime records are balanced and optimized according to the adversarial example generator to obtain a balanced and optimized lifetime record set. The learner is trained under supervision based on the balanced optimized lifetime record set to obtain the second high-frequency component lifetime prediction model. The high-frequency component lifetime prediction channel is generated by performing output fusion training based on the first high-frequency component lifetime prediction model and the second high-frequency component lifetime prediction model. The monitored high-frequency component is input into the high-frequency component lifetime prediction channel to obtain the first remaining lifetime; Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitored low-frequency components to obtain the second remaining life. Based on the balanced adversarial learning channel, the battery life of the battery swapping cabinet is predicted according to the monitoring trend component to obtain the third remaining life. A fourth remaining lifetime is obtained by global fusion of the first remaining lifetime, the second remaining lifetime, and the third remaining lifetime under confidence analysis.

2. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 1, characterized in that, Based on the battery monitoring data, sensor interference correction is performed to obtain a battery monitoring sequence, including: Based on each monitoring parameter in the battery monitoring data, scene information is collected from the IoT sensor array to obtain scene information for each sensor. Anomaly detection is performed based on the information of each sensing scenario to obtain the anomaly detection results for each sensing. Based on the abnormal detection results of each sensor, the battery monitoring data is subjected to interference analysis to obtain the interference analysis results of each abnormality. The battery monitoring data is adaptively corrected based on the analysis results of each abnormal interference to generate the battery monitoring sequence.

3. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 1, characterized in that, The battery monitoring sequence is decomposed using EMD to obtain high-frequency monitoring components, low-frequency monitoring components, and monitoring trend components, including: Extract the first monitoring signal by traversing the battery monitoring sequence; Perform empirical mode decomposition on the first monitoring signal to obtain the first intrinsic mode function component set and the first residual term; The first intrinsic mode function component set is classified according to its frequency distribution characteristics to obtain the first high-frequency component set and the first low-frequency component set; Amplitude normalization and time synchronization processing are performed on the first high-frequency component set to generate a first signal high-frequency component, and the first signal high-frequency component is incorporated into the monitoring high-frequency component. Amplitude normalization and time synchronization processing are performed on the first low-frequency component set to generate a first signal low-frequency component, and the first signal low-frequency component is incorporated into the monitoring low-frequency component. Amplitude normalization and time synchronization are performed based on the first residual term to generate a first signal trend component, and the first signal trend component is incorporated into the monitoring trend component.

4. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 1, characterized in that, A fourth remaining lifetime is obtained by global fusion of the first, second, and third remaining lifetimes under confidence analysis, including: The confidence level of the first remaining lifetime is evaluated based on the monitored high-frequency components to obtain the first lifetime confidence level. The confidence level of the second remaining lifetime is evaluated based on the monitored low-frequency components to obtain the confidence level of the second lifetime. The confidence level of the third remaining lifetime is evaluated based on the monitored trend components to obtain the confidence level of the third lifetime. The global proportion is calculated based on the first lifetime confidence level, the second lifetime confidence level, and the third lifetime confidence level to obtain the lifetime confidence weight; The first remaining lifetime, the second remaining lifetime, and the third remaining lifetime are weighted and fused according to the lifetime certainty weights to generate the fourth remaining lifetime.

5. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 4, characterized in that, The confidence level of the first remaining lifetime is evaluated based on the monitored high-frequency components to obtain the first lifetime confidence level, including: Based on the monitored high-frequency components, perform a large-scale lifespan retrieval of the same type of battery to obtain the first retrieval of remaining lifespan distribution; Confidence is evaluated for each remaining lifetime within the first remaining lifetime distribution to obtain the confidence level for each lifetime. Based on the lifetime confidence levels, the first retrieval remaining lifetime distribution is cleaned according to a predetermined confidence level to obtain the first retrieval confidence lifetime distribution; Based on the first search confidence lifetime distribution, the central value is calculated to obtain the high-frequency component fitting remaining lifetime; The cosine similarity between the first remaining lifetime and the high-frequency component fitted remaining lifetime is calculated to generate the first lifetime confidence score.

6. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 1, characterized in that, The IoT sensor array includes multiple IoT sensors.

7. The battery life prediction method for battery swapping cabinets supporting big data analysis as described in claim 1, characterized in that, The battery in the battery swapping cabinet is maintained and managed according to the fourth remaining lifespan.

8. A battery life prediction system for battery swapping cabinets supporting big data analysis, characterized in that: The steps for implementing the battery life prediction method for a battery swapping cabinet supporting big data analysis as described in any one of claims 1 to 7, wherein the battery life prediction system for a battery swapping cabinet supporting big data analysis comprises: The real-time monitoring module is used to monitor the batteries in the battery swapping cabinet in real time through an IoT sensor array, obtain battery monitoring data, and perform sensor interference correction based on the battery monitoring data to obtain a battery monitoring sequence. The EMD decomposition module is used to perform EMD decomposition on the battery monitoring sequence to obtain the monitoring high-frequency component, monitoring low-frequency component and monitoring trend component. The first prediction module is used to predict the lifespan of the battery in the battery swapping cabinet based on the monitored high-frequency components using the balanced adversarial learning channel, and to obtain the first remaining lifespan. The second prediction module is used to predict the lifespan of the battery swapping cabinet battery based on the balanced adversarial learning channel and the monitored low-frequency component to obtain the second remaining lifespan. The third prediction module is used to predict the lifespan of the battery swapping cabinet battery based on the balanced adversarial learning channel and the monitoring trend component to obtain the third remaining lifespan. The global fusion module is used to perform global fusion based on the first remaining lifetime, the second remaining lifetime, and the third remaining lifetime under a confidence analysis to obtain a fourth remaining lifetime.

Images