Battery abnormality early warning method and device based on ensemble learning

Abstract

The application provides a battery abnormality early warning method and device based on ensemble learning, and relates to the technical field of artificial intelligence, based on the sampling value instantaneous change of multiple source time series data preprocessing, which can avoid feature misjudgment caused by time dislocation, and can effectively retain key warning features, thereby effectively solving the problems of data asynchrony and feature loss. In addition, time-varying statistics are extracted for the preset state of charge interval, and are spliced with frequency domain adaptive entropy features to generate a target feature vector, which can effectively reflect the dynamic correlation between multiple parameters and significantly improve the representation ability of early composite abnormalities. Further, the battery abnormality prediction is based on the ensemble learning model of the base learning pool, the adaptive weight adjustment strategy and the result fusion unit, which can accurately match different structure samples, focus on key abnormal patterns, and the output battery abnormality probability has precision and stability, which can realize efficient and reliable early warning of battery early abnormalities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a battery early abnormality warning method and device based on ensemble learning. BACKGROUND

[0002] Under the global energy transformation, the safety and reliability of battery systems such as electric vehicles and energy storage power stations have become the core demand as the scale of their application increases. During long-term operation of the battery, there will be hidden electrochemical decay and material degradation, which will only manifest as weak abnormalities in monitoring parameters such as voltage and current in the early stage, and if not timely warned, it may lead to thermal runaway.

[0003] Among them, the monitoring data such as voltage and current collected by multiple sensors during the operation of the battery are naturally asynchronous and delayed, and the existing fixed window alignment or simple filtering method cannot effectively correct the asynchrony and will also smooth the weak mutation signal of the early abnormality, resulting in the loss of key features. In addition, there is a dynamic correlation between multi-dimensional monitoring parameters, and the conventional method simply concatenates the time domain features extracted independently, ignoring the parameter correlation and frequency energy change analysis, and it is difficult to represent the early composite abnormality. Moreover, the conventional algorithm does not consider the difference in abnormality representation difficulty of the battery under different states of charge, and processes all samples equally, resulting in insufficient learning of key abnormal patterns.

[0004] In summary, the existing method is difficult to accurately capture the early abnormality features of the battery, and cannot guarantee the safety of the system, so a new solution is needed to break through the technical bottleneck to meet the requirements of early warning accuracy and reliability for large-scale applications. SUMMARY

[0005] The present application relates to the technical field of artificial intelligence, and in particular to a battery early abnormality warning method and device based on ensemble learning.

[0006] In a first aspect, an embodiment of the present application provides a battery abnormality early warning method based on ensemble learning, which comprises: obtaining running data of a target battery; the running data comprising multi-source time series data of the target battery; based on instantaneous changes of sample values of each attribute data of the running data, performing path alignment and data cleaning on each attribute data to generate preprocessed data; calculating time-varying statistics of the preprocessed data with respect to a preset state of charge, performing feature splicing on the time-varying statistics and adaptive entropy features of the preprocessed data in a frequency domain to generate a target feature vector; inputting the target feature vector into a pre-constructed battery abnormality diagnosis model, determining a battery abnormality probability corresponding to the target feature vector through the battery abnormality diagnosis model, and performing abnormality warning on the target battery based on the battery abnormality probability; wherein the battery abnormality diagnosis model is constructed based on an ensemble learning architecture, the ensemble learning architecture comprising a base learner pool, an adaptive weight adjustment strategy of the base learner pool, and a result fusion unit; each base learner included in the base learner pool is used to process sample data of a preset structure.

[0007] In combination with the first aspect, an embodiment of the present application provides a first implementation manner of the first aspect, wherein the step of performing path alignment and data cleaning on each attribute data based on instantaneous changes of sample values of each attribute data of the running data comprises: determining instantaneous change amounts of sample values corresponding to each attribute data of the running data based on sampling time points of the running data; calculating optimal alignment paths corresponding to the instantaneous change amounts of sample values through a preset dynamic time warping algorithm; performing sampling time point alignment on each attribute data based on the optimal alignment paths to generate first preprocessed data; and performing data cleaning on the running data according to local outlier factors in the first preprocessed data to generate the preprocessed data.

[0008] In combination with the first aspect, an embodiment of the present application provides a second implementation manner of the first aspect, wherein the step of calculating time-varying statistics of the preprocessed data with respect to a preset state of charge comprises: dividing the preprocessed data into multiple intervals according to battery state of charge values, and calculating current distribution characteristics of each interval; performing weighted fusion on the current distribution characteristics based on center values of the state of charge corresponding to each interval to generate time-varying statistics corresponding to the preprocessed data; and the step of performing feature splicing on the time-varying statistics and adaptive entropy features of the preprocessed data in a frequency domain to generate a target feature vector comprises: performing splicing on the time-varying statistics, the adaptive entropy features, and time series coupling features corresponding to the preprocessed data to generate the target feature vector; wherein the adaptive entropy features are determined by performing adaptive wavelet packet decomposition on the preprocessed data.

[0009] With reference to the first aspect, in a third implementation form of the first aspect, the method further includes: calculating derivative features of the preprocessed data in multiple dimensions, reconstructing the preprocessed data into a spatio-temporal feature tensor based on the derivative features; the multiple dimensions include parameter type, time evolution and feature morphology dimensions; performing tensor decomposition on the spatio-temporal feature tensor to extract a core coupled feature tensor of the spatio-temporal feature tensor; and flattening the core coupled feature tensor into a one-dimensional vector to generate the time sequence coupled feature of the preprocessed data.

[0010] With reference to the first aspect, in a fourth implementation form of the first aspect, the method further includes: performing wavelet packet decomposition on the preprocessed data to determine node wavelet coefficients corresponding to multiple frequency bands respectively; determining an optimal wavelet packet decomposition tree corresponding to the node wavelet coefficients by using a minimum cost function; and extracting entropy values of all leaf nodes in the optimal wavelet packet decomposition tree to form the adaptive entropy feature corresponding to the preprocessed data.

[0011] With reference to the first aspect, in a fifth implementation form of the first aspect, the method further includes: inputting the target feature vector into a pre-constructed battery abnormality diagnosis model, determining an initial probability sequence of the target feature vector by using a base learner of the battery abnormality diagnosis model; wherein the initial probability sequence includes cluster prediction probabilities corresponding to multiple clusters of the target feature vector respectively; modeling time sequence dependence corresponding to the initial probability sequence by using a gated recurrent unit network of the battery abnormality diagnosis model to generate an initial output; and performing expert prior calibration on the initial output to determine the battery abnormality probability corresponding to the target feature vector.

[0012] With reference to the first aspect, in a sixth implementation form of the first aspect, the method further includes: obtaining a training sample set corresponding to the running data; performing feature clustering on the training sample set to generate multiple cluster features, and constructing a base learner for each cluster feature; determining a dynamic weight of the base learner by using a dynamic weight updating mechanism, and performing model training on the base learner; and constructing the battery abnormality diagnosis model based on the trained base learner.

[0013] With reference to the first aspect, in a seventh implementation form of the first aspect, the method further includes: using a cross-entropy loss function as a training optimization target of the base learner to perform model training on the base learner.

[0014] With reference to the first aspect, the eighth implementation of the first aspect is provided in the embodiments of the present application, wherein the step of determining the dynamic weight of the base learner by using the dynamic weight updating mechanism comprises: determining a local anomaly factor of the cluster feature based on the outlying degree corresponding to the cluster feature; determining an interval weight corresponding to the current cluster feature according to the importance of the state of charge interval to which the cluster feature belongs; and dynamically updating the weight coefficient of the base learner by using the interval weight and the local anomaly factor to generate the dynamic weight of the base learner.

[0015] In the second aspect, the embodiments of the present application provide a battery abnormality early warning device based on ensemble learning, which comprises: a data acquisition module configured to acquire running data of a target battery; the running data comprises multi-source time series data of the target battery; a preprocessing module configured to perform path alignment and data cleaning on each attribute data based on the instantaneous change of the sampling value of each attribute data of the running data to generate preprocessed data; a data processing module configured to calculate time-varying statistics of the preprocessed data for a preset state of charge, perform feature splicing on the time-varying statistics and adaptive entropy features in the frequency domain of the preprocessed data to generate a target feature vector; and an execution module configured to input the target feature vector into a pre-constructed battery abnormality diagnosis model, determine a battery abnormality probability corresponding to the target feature vector by the battery abnormality diagnosis model, and perform abnormality warning on the target battery based on the battery abnormality probability; wherein the battery abnormality diagnosis model is constructed based on an ensemble learning architecture, the ensemble learning architecture comprises a base learner pool, an adaptive weight adjustment strategy of the base learner pool, and a result fusion unit; each base learner included in the base learner pool is configured to process sample data of a preset structure.

[0016] The embodiments of the present application have the following beneficial effects: The embodiments of the present application provide a battery abnormality early warning method and device based on ensemble learning, which can avoid feature misjudgment caused by time misplacement by preprocessing the instantaneous change of the sampling value of multi-source time series data, and can effectively retain key warning features, thereby effectively solving the problems of data asynchrony and feature loss. In addition, the time-varying statistics for the preset state of charge (SOC) interval are extracted and spliced with the frequency domain adaptive entropy features to generate a target feature vector, which can effectively reflect the dynamic correlation between multiple parameters and significantly improve the representation ability for early composite abnormalities.

[0017] Further, the battery abnormality prediction is performed based on the ensemble learning model comprising the base learner pool, the adaptive weight adjustment strategy, and the result fusion unit, which can accurately match different structure samples and focus on key abnormal patterns, and the output battery abnormality probability has both accuracy and stability, thereby realizing efficient and reliable early warning of battery abnormalities.

[0018] Other features and advantages of the present application will be set forth in the descriptions that follow, and in part will be apparent from the description, or can be learned by practice of the application as hereinafter described, or can be learned by practice of the application. It is also to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application, as claimed.

[0019] In order to make the above objectives, features and advantages of the present application more apparent, the following will describe a preferred embodiment in detail, and the accompanying drawings will be described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings described in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative effort.

[0021] Figure 1 A flowchart of a battery abnormal early warning method based on ensemble learning provided by an embodiment of the present application;

[0022] Figure 2 A battery abnormal detection accuracy comparison diagram corresponding to an embodiment of the present application;

[0023] Figure 3 An abnormal sample quantity distribution diagram;

[0024] Figure 4 A comparison diagram of early abnormal detection capabilities corresponding to different methods;

[0025] Figure 5 A recognition performance diagram corresponding to an embodiment of the present application;

[0026] Figure 6 A structural diagram of a battery abnormal early warning device based on ensemble learning provided by an embodiment of the present application;

[0027] Figure 7 A structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more apparent, the following will describe the technical solutions in the embodiments of the present application clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0030] This invention provides a method and apparatus for early warning of battery anomalies based on ensemble learning, which can achieve efficient and reliable early warning of early battery anomalies.

[0031] To facilitate understanding, the early warning method for battery anomalies based on ensemble learning provided in this embodiment of the invention will be described first, referring to... Figure 1 The method includes the following steps:

[0032] Step S102: Obtain the operating data of the target battery.

[0033] The target battery is a specific battery cell or battery pack corresponding to the monitoring scenario, applicable to various battery systems such as electric vehicles and energy storage power stations. The operational data includes multi-source time-series data of the target battery. Operational data is a set of data reflecting the battery's operating status, collected in real-time by various sensors and data acquisition devices under actual operating conditions such as charge-discharge cycles and resting periods. Multi-source time-series data refers to sequential data continuously collected chronologically by multiple different types of data acquisition units (such as voltage sensors, current sensors, temperature sensors, and state-of-charge (SOC) estimation modules), reflecting the dynamic changes in battery status over time and used to capture early abnormal trends.

[0034] In one implementation, data acquisition can be performed in real time via a sensor network deployed on the battery module or system and a battery management system. The acquired parameters include the battery's total voltage, total current, temperature at each sampling point (located on the battery surface or at key hot spots), and the battery's state of charge (SOC) estimated using the ampere-hour integration method or a combination of the ampere-hour and open-circuit voltage methods. The data is continuously recorded at a fixed sampling frequency (e.g., 1Hz or 10Hz), covering the battery's complete operating cycle, including charging, discharging, and resting conditions, to comprehensively capture the battery's dynamic behavior under different states.

[0035] Step S104: Based on the instantaneous changes in the sampled values ​​of each attribute data in the running data, perform path alignment and data cleaning on each attribute data to generate preprocessed data.

[0036] The attribute data refers to a subdivided data type reflecting a specific running attribute of a battery in multi-source time-series data, such as voltage attribute data, current attribute data, temperature attribute data, state of charge attribute data, and the like. The sampling value instantaneous change refers to a numerical difference or a change rate of a certain attribute data at adjacent sampling time points, which can accurately reflect the dynamic fluctuation characteristics of the data and is used to distinguish normal fluctuations from early abnormal weak mutations.

[0037] The embodiment of the present application realizes path alignment based on the sampling value instantaneous change of the attribute data, can identify the time offset characteristics of each attribute data, realizes dynamic path alignment, effectively corrects the asynchronicity of multi-source data, avoids feature misjudgment caused by time misalignment. In addition, the sampling value instantaneous change can be used as a basis to distinguish noise and early abnormal weak mutations, so that only meaningless noise is removed and key mutation signals related to abnormalities are retained during data cleaning, completely solving the defect of key early warning feature loss in the existing method.

[0038] In step S106, the time-varying statistics of the preprocessed data for the preset state of charge are calculated, the time-varying statistics are feature spliced with the adaptive entropy features of the preprocessed data in the frequency domain, and a target feature vector is generated.

[0039] The preset state of charge (SOC) refers to the abnormal representation law according to different running stages of the battery; the time-varying statistics refer to feature indexes that dynamically change with time obtained by statistical analysis of the preprocessed data in the preset state of charge interval, such as mean, variance, peak value, kurtosis, etc. of the voltage in the interval, which can quantitatively reflect the time domain running characteristics of the battery in a specific SOC interval. Among them, the state of charge interval (such as 0-20%, 20%-80%, 80%-100%, etc.) can be pre-divided, and the electrochemical characteristics and abnormal performance of the battery in different intervals are significantly different. Further, entropy is an index for measuring signal uncertainty or complexity, and adaptive entropy refers to an entropy value feature that can automatically adjust the calculation parameters according to the dynamic change of the signal, which can accurately quantify the energy distribution complexity of the battery running signal in the frequency domain, and its change can effectively reflect the early electrochemical decay of the battery.

[0040] The embodiment of the present application extracts time-varying statistics based on the preset SOC interval, which can fully consider the differences of abnormal features under different states of charge, avoid the problem of poor feature generalization caused by ignoring the influence of SOC in the existing method. In addition, the introduction of the frequency domain adaptive entropy feature can make up for the deficiency of the existing technology which only focuses on time domain features, and capture the hidden early abnormal information in the time domain signal. The embodiment of the present application fuses multi-dimensional feature information through time domain-frequency domain feature splicing, can simultaneously reflect the dynamic correlation between multiple parameters, significantly improves the representation ability of early composite abnormalities, and solves the problem of low recognition accuracy of composite abnormalities in the existing method.

[0041] In step S108, the target feature vector is input into a pre-constructed battery abnormality diagnosis model, the battery abnormality probability corresponding to the target feature vector is determined through the battery abnormality diagnosis model, and the target battery is warned of abnormality based on the battery abnormality probability.

[0042] The battery abnormality diagnosis model is a model constructed based on a machine learning algorithm, capable of analyzing a target feature vector and outputting a battery abnormality probability, to realize battery abnormality warning. In the embodiment of the present application, the battery abnormality diagnosis model is constructed based on an ensemble learning architecture, which includes a base learner pool, an adaptive weight adjustment strategy of the base learner pool, and a result fusion unit; each base learner included in the base learner pool is used to process sample data of a preset structure.

[0043] The ensemble learning architecture refers to a machine learning framework that combines the output results of multiple base learners to obtain better performance than a single learner, by constructing multiple independent base learners and designing a reasonable weight adjustment and result fusion strategy. The base learner pool refers to a collection of base learners (such as decision trees, support vector machines, neural networks, etc.) of different types or different parameter settings, each of which has a specific sample processing advantage and can adapt to sample data of different structures (such as stationary working condition samples, dynamic working condition samples, and weak abnormality samples, etc.). The adaptive weight adjustment strategy refers to a dynamic adjustment mechanism that automatically adjusts the weight proportion of the output results of each base learner according to the diagnostic accuracy of each base learner in different SOC intervals and different sample types, so that the base learner with high diagnostic accuracy plays a greater role in the corresponding scene.

[0044] The embodiment of the present application enables the model to adapt to sample data of different structures based on the setting of the base learner pool, avoiding the diagnostic blind area of a single learner for specific samples. Combined with the adaptive weight adjustment strategy, the model can dynamically allocate weights according to the differences in SOC intervals and sample characteristics, focus on abnormal samples with high diagnostic difficulty, and solve the defect of insufficient learning of key abnormalities caused by the equal processing of all samples by existing algorithms. Further, the weighted output results of each base learner are analyzed by the result fusion unit, combined with time dimension smoothing processing and battery operation common sense (such as "the abnormal probability is usually low in the initial charging period"), and then a stable abnormal probability is output, which can effectively filter transient misjudgments caused by working condition fluctuations, significantly reduce the false alarm rate, and improve the warning stability. Finally, through the output of accurate and stable abnormal probability, reliable early warning of battery abnormalities is realized, providing protection for the safety of the battery system.

[0045] In summary, the embodiment of the present application can avoid feature misjudgment caused by time misalignment and effectively retain key early warning features based on the preprocessing of instantaneous changes of sampling values of multi-source time series data, thereby effectively solving the problems of data asynchrony and feature loss. In addition, time-varying statistics are extracted for a preset state of charge (SOC) interval, and are spliced with frequency domain adaptive entropy features to generate a target feature vector, which can effectively reflect the dynamic correlation between multiple parameters and significantly improve the representation ability of early composite anomalies.

[0046] Further, the integrated learning model based on the base learning model pool, the adaptive weight adjustment strategy and the result fusion unit is used for battery anomaly prediction, which can accurately match different structural samples and focus on key abnormal patterns. The output battery anomaly probability has accuracy and stability, and can realize efficient and reliable early warning of battery early anomalies.

[0047] Further, based on the above embodiment, the embodiment of the present application also provides another battery anomaly early warning method based on integrated learning. For the above step S104, the embodiment of the present application determines the instantaneous change amount of the sampling value corresponding to each attribute data of the running data based on the sampling time of the running data; calculates the optimal alignment path corresponding to the instantaneous change amount of the sampling value through a preset dynamic time warping algorithm; aligns the sampling time of each attribute data based on the optimal alignment path, and generates first preprocessed data. Further, the running data is cleaned according to the local outlier factor in the first preprocessed data, and preprocessed data is generated.

[0048] In the battery running data, the time series of multiple parameters such as current, voltage, temperature and state of charge often produce misalignment due to sensor sampling delay and working condition fluctuation. The conventional cleaning method based on a fixed window will smooth the local mutation in the data, thereby weakening the weak signal features related to the early anomalies of the battery. If the time series is directly aligned or filtered by a fixed window, the dynamic coupling relationship between the current, voltage and state of charge will be blurred, and the subtle features representing the early anomalies of the battery cannot be effectively retained.

[0049] The present application aligns multi-source time series through a dynamic time warping algorithm to eliminate the delay between sensors, and combines an adaptive cleaning window based on local gradient mutation detection to remove noise while retaining potential abnormal mutation points.

[0050] Corresponding to the above steps, the preprocessing process is as follows:

[0051] 1) Dynamic alignment of multi-source time series

[0052] Using the current sequence as a reference, the optimal alignment path between it and the voltage sequence is calculated using a dynamic time warping algorithm. Based on this path, the temperature sequence and the state of charge sequence are synchronously interpolated to correct the sampling delay between sensors. The time-aligned multidimensional data matrix is ​​then output, as follows:

[0053]

[0054] In the formula, This represents the optimal alignment path found by the dynamic time warping algorithm, which minimizes the cumulative alignment cost of all matching points on the path. This represents a candidate alignment path, which is a set of point pairs used to match the time indices of current and voltage sequences. This represents the weighting coefficient for temperature gradient differences, used to balance the synchronicity of current-voltage differences and temperature changes in path distance calculations to compensate for the hysteresis effect of the temperature sensor. The default value is [value missing]. ; This indicates the search for parameters that minimize the objective function. The operation.

[0055] The time index represents the reference current sequence; The time index represents the target voltage sequence; Indicates the optimal alignment path A matching point pair in a current sequence indicates the first matching point in the current sequence. The first time point and voltage sequence Each time point should be aligned. Represents current sequence In the Sampled values ​​at each time index; Represents voltage sequence In the The sampled value at each time index. Indicates the temperature sequence in the index The first-order forward difference at point is calculated as follows: , used to characterize the instantaneous rate of temperature change; Representing temperature sequence In the Sampled values ​​at each time index; Representing temperature sequence In the Sampled values ​​at each time index; Indicates the temperature sequence in the index The first-order forward difference at the location.

[0056] wherein, in the dynamic time warping algorithm, the path is a set of point pairs, each of which represents a time index of the reference current sequence and a time index of the target voltage sequence that should be aligned, the set defining a time mapping relationship between the two sequences, ensuring that the cumulative alignment cost is minimized. It should be noted that is the cost function calculation method adopted by the conventional dynamic time warping algorithm, which is used to align two sequences of different lengths. The core of the conventional dynamic time warping is to construct a cost matrix and find a path through dynamic programming, so that the cumulative distance of all point pairs on the path is minimized. This distance is usually the Euclidean distance, Manhattan distance or other point-to-point difference measure of the two sequences at the corresponding index points. The embodiments of the present application improve the cost function of the alignment path. Based on the above, the present application calculates the instantaneous rate of temperature change based on the first-order difference calculation method, and measures the difference between the temperature change trend of the i-th time point of the current sequence and the j-th time point of the voltage sequence by constructing , effectively solving the problems of "asynchronous delay between sensors" and "ignoring the dynamic interaction between electrical and thermal parameters". Temperature sensors usually have a large lag. If only the current and voltage are aligned, a mathematically minimal-cost path may be found, but this path may violate the physical timing logic of temperature change. By using the improved method of the present application, when finding the alignment path, it will tend to choose those time points that not only match the current and voltage values, but also synchronize the temperature change trend. The optimal alignment path and the synchronized data matrix obtained in this way can better reflect the real electrical-thermal coupling dynamic process inside the battery, and is no longer a simple signal alignment, but an intelligent synchronization based on the correlation of multiple physical fields.

[0057] Further, after time alignment, resample or interpolate all sequences according to the optimal alignment path to obtain a synchronized multi-dimensional data matrix (i.e., the first pre-processed data), wherein , , , respectively represent the aligned current, voltage, temperature and state of charge sequences.

[0058] 2) Adaptive cleaning based on gradient mutation

[0059] Calculate the first-order gradient of the aligned voltage sequence, and evaluate the mutation outlier degree of each gradient point based on the local outlier factor algorithm. Then, according to the local outlier factor value, dynamically set a threshold value to screen out data indexes that are judged as non-outlier points to form a cleaning window, and then output the cleaned time series data matrix, denoted as: ​

[0060]

[0061] In the formula, Indicates the first Voltage gradient values ​​at each time point The local outlier factor is used to quantify the degree of outlier of a data point relative to its local neighborhood. The larger the value, the higher the probability that the point is a mutation outlier. Indicates the first The first-order difference of the voltage sequence at each time point is calculated as follows: , used to characterize the instantaneous rate of voltage change; Represents voltage sequence In the Sampled values ​​at each time point; Represents voltage sequence In the Sampled values ​​at each time point; Indicates a point-in-time index; This represents the number of nearest neighbors considered when calculating the local outlier factor; the default value is [value to be filled in]. ; This represents the index of the nearest neighbor, used for traversal. The nearest neighbor point, ; Indicates the first The local reachability density of the nearest neighbor is used to measure the density of data around that neighbor. This represents the local reachability density of a data point, used to measure the density of data surrounding that point. Represents a data point in the feature space; This represents a data point in the feature space.

[0062] Furthermore, after the local outlier factor is calculated, the cleaning window is dynamically determined based on its value, resulting in the set of time point indices retained after cleaning, represented as:

[0063]

[0064] In the formula, This represents the set of time point indices retained after cleaning; points that meet the conditions are considered to be "normal" fluctuations. This represents the outlier threshold coefficient, used to adjust the strictness of the cleaning process; the default value is [value missing]. .

[0065] Based on this, from the synchronized multidimensional data matrix Extract all column indexes belonging to The data forms a cleaned time-series data matrix. (i.e., preprocessed data), which has a dimension of , is the number of elements in the middle of

[0066] Further, the early abnormal signal of the battery is weak in the time domain and is easily submerged by noise, but in the frequency domain, it can present a specific resonance frequency shift or energy distribution change. Meanwhile, the conventional statistical feature extraction method is easily contaminated by random noise and lacks dynamic relevance to the state of charge of the battery. In the conventional technology, the wavelet transform of the fixed basis function is difficult to adapt to the complex signal characteristics of the battery under different working conditions, and the isolated statistical features cannot effectively capture the time sequence correlation between the abnormality and the working condition, resulting in insufficient characterization of the early abnormality. For the above step S106, the present application extracts the adaptive entropy features of the signal in the frequency domain through adaptive wavelet packet decomposition, so as to capture the energy redistribution caused by the abnormality. At the same time, the time-varying statistics dynamically correlated with the state of charge of the battery are calculated to enhance the adaptability of the features to the operating conditions. Then, the two types of features are fused to form a high-dimensional derived feature set more sensitive to early abnormalities. For the time-varying statistics, the present application divides the preprocessed data into multiple intervals according to the state of charge value of the battery, calculates the current distribution characteristics of each interval to describe the statistical rules of the current data in a specific SOC interval, and quantifies the charge and discharge state stability of the battery in the interval. Further, based on the center value of the state of charge corresponding to each interval, the current distribution characteristics are weighted and fused to generate the time-varying statistics corresponding to the preprocessed data. The center value of the state of charge is used to indicate the midpoint value of the SOC interval. Here, as a representative indicator of the SOC level in the interval, it is used to reflect the importance of different intervals in the operation of the battery. The present application is based on the current distribution characteristics corresponding to the partition to correspond to the state of the battery under a specific SOC, and the feature weight is tilted to the regular operating interval of the battery through the weighted fusion of the SOC center value, which can effectively amplify the weak current change characteristics corresponding to the early abnormality in the interval (such as the slight increase of the current variance in the core interval under abnormal working conditions), while reducing the interference of the extreme SOC interval, and improving the sensitivity of the time-varying statistics to the early abnormality. In specific implementation, the cleaned current sequence (i.e., the above preprocessed data) can be divided into multiple intervals according to the state of charge value of the battery, and the mean, standard deviation and skewness of the current in each interval are calculated. Then, the weighted fusion is performed according to the distance between the center value of the state of charge of each interval and the reference value to generate a statistical feature vector dynamically related to the state of the battery, which is represented as:

[0067]

[0068] In the formula, represents the dynamic statistical feature vector, which has a dimension of , representing the distribution characteristics of the current under different states of charge; ​an index representing a state of charge interval, ; an index representing a total number of state of charge intervals, with a default value of ; an index representing the th state of charge interval, ; an index representing the th state of charge interval, ; an index representing the th state of charge interval, ; an index representing the th state of charge interval,

[0069] ; an index representing the th state of charge interval, ; an index representing the th state of charge interval, ; an index representing the th state of charge interval, ; ; an index representing the th state of charge interval, ;

[0070] Furthermore, regarding feature concatenation, this embodiment of the invention concatenates time-varying statistics, adaptive entropy features, and time-series coupling features corresponding to the preprocessed data to generate a target feature vector. This embodiment of the invention, by integrating features from the time domain, frequency domain, and parameter correlation domain, solves the information blind spot problem caused by existing methods relying solely on single-dimensional features, and can comprehensively cover the multi-dimensional manifestations of early battery anomalies (such as time-domain parameter fluctuations, frequency-domain energy mutations, and parameter correlation imbalances). Specifically, the adaptive entropy feature is determined by adaptive wavelet packet decomposition of the preprocessed data. For example, based on the frequency fluctuation characteristics of the input signal, the number of wavelet packet decomposition layers (e.g., 2-5 layers) and the basis function type are dynamically adjusted to adaptively decompose the original time-domain signal into multiple sub-signals of different frequency bands, ensuring that the decomposition results match the non-stationary characteristics of the signal. The entropy values ​​(e.g., sample entropy, approximate entropy) of each frequency band sub-signal are further calculated to quantify the complexity and uncertainty of the sub-signals, effectively mining the early anomaly frequency domain information hidden in the time-domain signal by capturing the energy distribution and complexity changes of the preprocessed data in the frequency domain. Correspondingly, the target feature vector is represented as: This formula represents the coupling eigenvector. Adaptive wavelet packet entropy feature vector and dynamic statistical eigenvectors The vectors are concatenated to obtain the derived feature vector (i.e., the target feature vector). Its dimension is... Based on multi-dimensional data integration, a more sensitive and comprehensive characterization of early battery anomalies has been formed.

[0071] Furthermore, regarding the adaptive entropy feature, this embodiment of the invention performs wavelet packet decomposition on the preprocessed data to determine the node wavelet coefficients corresponding to multiple frequency bands; and uses the minimization of the cost function to determine the optimal wavelet packet decomposition tree corresponding to the node wavelet coefficients. In one implementation, the data can be processed by a wavelet packet decomposition algorithm (such as adaptive decomposition to determine the number of decomposition layers and basis functions) to obtain a set of wavelet coefficients under each node, constituting the signal feature representation of that frequency band. The wavelet coefficients based on the node can reflect the energy intensity of the signal in the corresponding frequency band (node ​​p) and the corresponding feature component (index d). Furthermore, the frequency domain characteristics of the signal can be quantified by Shannon entropy. The node Shannon entropy can measure the degree of disorder of the signal in the frequency band corresponding to the p-th node in the wavelet packet decomposition tree. The level of entropy directly reflects the signal energy distribution characteristics: the higher the entropy value, the more uniform and random the distribution of signal energy in each wavelet coefficient in that frequency band; the lower the entropy value, the more concentrated the energy is in a few wavelet coefficients, and the signal has obvious regularity.

[0072]

[0073] In the formula, In the wavelet packet decomposition tree, the first... Shannon entropy of a node, used to measure the uniformity or complexity of the signal energy distribution in the frequency band corresponding to the node, the higher the entropy value, the more uniform and random the energy distribution is; represents the th node in the wavelet packet decomposition tree; represents the th wavelet coefficient in the th node; ; represents the logarithmic function, the default base is the natural constant. represents the th wavelet coefficient corresponding to the th node, obtained by wavelet packet decomposition; represents the th wavelet coefficient corresponding to the th node, obtained by wavelet packet decomposition; represents the index of the wavelet coefficient in the th node; represents the index of the wavelet coefficient different from in the th node. This Shannon entropy is used to measure the uncertainty of the signal (the degree of disorder of the signal energy distribution). When the battery is running normally, the monitoring signal (voltage, current) has a stable regularity in a certain frequency band, and the energy will be concentrated in a few wavelet coefficients, at this time, some of the values of p p p d are close to 1, and the rest are close to 0, and the entropy value is low; when the battery has early abnormalities, the regularity of the signal is destroyed, and the energy will spread to multiple wavelet coefficients, resulting in relatively uniform values of each p p p d, and the entropy value increases significantly. Therefore, the mutation of the entropy value can be used as an important feature signal of early abnormalities.

[0074] Further, the optimal wavelet packet decomposition tree is obtained by optimizing the following objective:

[0075]

[0076] wherein, represents the variance of all wavelet coefficients in the th node, representing the energy fluctuation degree of the signal in the frequency band corresponding to the node; represents the optimal wavelet packet decomposition tree obtained by optimizing the objective, i.e. the decomposition tree structure that minimizes the cost function; represents all leaf nodes in the decomposition tree ; represents the node index of the wavelet packet decomposition tree; represents the decomposition tree The operation of the product of the two matrices is performed. is a natural constant.

[0077] Further, the entropy values of all leaf nodes in the optimal wavelet packet decomposition tree are extracted to form the adaptive entropy feature corresponding to the preprocessed data. In specific implementation, the preprocessed voltage sequence is subjected to adaptive wavelet packet decomposition, the optimal wavelet packet decomposition tree is selected by minimizing a cost function that comprehensively considers the node information entropy and signal variance, and the entropy values of all leaf nodes in the optimal tree are calculated to form a feature vector representing the signal frequency domain complexity and energy distribution, i.e., the adaptive entropy feature, which is expressed as: ; wherein the adaptive wavelet packet entropy feature vector is formed by extracting the entropy values of all leaf nodes from represents the total number of leaf nodes in the optimal wavelet packet decomposition tree represents the entropy value of the first leaf node in the optimal wavelet packet decomposition tree represents the entropy value of the first leaf node in the optimal wavelet packet decomposition tree .

[0078] Further, for the time series coupling feature, the embodiment of the present application determines the multi-modal time series feature coupling enhancement of the preprocessed data. In specific implementation, the embodiment of the present application calculates the derived features of the preprocessed data in multiple dimensions, reconstructs the preprocessed data into a spatio-temporal feature tensor based on the derived features, performs tensor decomposition on the spatio-temporal feature tensor to extract the core coupling feature tensor of the spatio-temporal feature tensor, and flattens the core coupling feature tensor into a one-dimensional vector to generate the time series coupling feature of the preprocessed data. The multiple dimensions include parameter type, time evolution and feature morphology dimension.

[0079] There is a complex nonlinear coupling relationship between the parameters such as current, voltage, temperature and state of charge of the battery, for example, temperature rise causes internal resistance change, and then causes end voltage drop, and the single parameter feature extraction method cannot model the dynamic interaction across parameters, and the conventional technology independently extracts each parameter feature and then simply splices, ignoring the coupling effect between parameters, thereby reducing the sensitivity of the model to early composite abnormal signals.

[0080] The present application extracts low-dimensional fusion features that can represent the interaction of the internal state of the battery by constructing a high-order spatio-temporal feature tensor and performing constrained tensor decomposition, and jointly learning the coupling modes between parameters, time points and different feature types, and the specific steps are as follows:

[0081] 1) Constructing a spatio-temporal feature tensor ​​

[0082] The cleaned time series data matrix is combined with the preset derived features to reconstruct a three-order spatiotemporal feature tensor to simultaneously encapsulate information in three dimensions of parameter type, time evolution and feature morphology, and is expressed as:

[0083]

[0084] In the formula, represents the three-order spatiotemporal feature tensor, which can comprehensively represent the multi-modal coupling features of the battery operating state, and has a dimension of ; represents the element value of the spatiotemporal feature tensor at the parameter index , the time index and the feature index ; represents the three-order spatiotemporal feature tensor , the index of the first dimension, corresponding to the parameter type, respectively representing the current, voltage, temperature and state of charge; represents the time point index different from t; represents the three-order spatiotemporal feature tensor , the index of the third dimension, corresponding to the feature type, respectively representing the original observation value, the first-order difference and the local energy feature. represents the original observation value of the th parameter at the th time point in the cleaned time series data matrix , i.e., the pretreated data; represents the original observation value of the th parameter at the th time point in the cleaned time series data matrix .

[0085] represents the first-order difference of the th parameter at the th time point in the cleaned time series data matrix , i.e., ; represents the original observation value of the th parameter at the th time point in the cleaned time series data matrix . represents the energy sum of the data in the local window centered on the th parameter sequence at the time point , for representing the signal intensity or fluctuation intensity near the time point. Both of them represent the derived features of the pretreated data in the corresponding dimension. denotes a function that takes the larger of the two values; denotes a function that takes the smaller of the two values.

[0086] 2) Coupling tensor decomposition

[0087] In an embodiment, the spatiotemporal coupling tensor can be decomposed by Tucker decomposition (adapt to tensors of arbitrary order), and the core coupling tensor that condenses the coupling information in multiple dimensions is extracted by projecting and reducing dimensions in different dimensions through three factor matrices, and then it is flattened into a feature vector and represented as: .

[0088] In the formula, denotes the core coupling tensor, and the dimension is , which condenses the joint information among multiple parameters, long time span, and multiple feature types; denotes the product operation of the tensor and the matrix in the first mode, denotes the product operation of the tensor and the matrix in the second mode, denotes the product operation of the tensor and the matrix in the third mode; denotes the parameter coupling factor matrix, and the dimension is , which is used to learn and compress four original parameter types into a dimensional coupling parameter space; denotes the parameter coupling rank, and the default value is ; denotes the time-dependent factor matrix, and the dimension is , which is used to learn the evolution pattern in the time dimension and compress it to dimension; denotes the time rank, and the default value is ; denotes the feature interaction factor matrix, and the dimension is , which is used to learn the interaction relationship among the original value, the first-order difference, and the local energy of the three feature types and compress it to dimension; denotes the feature interaction rank, and the default value is . Further, the core coupling tensor is flattened into a one-dimensional vector to obtain the coupling feature vector , and the dimension is .

[0089] In combination with the above data processing steps, the embodiments of the present application also analyze the performance of each abnormality detection method in different state of charge intervals and the distribution of abnormal samples in different state of charge intervals. For details, refer to Figure 2, shows the battery anomaly detection accuracy comparison diagram of the embodiment of the application, and the experiment compares the three technologies of the method of the application, the convolutional neural network long short-term memory network method and the conventional statistical method. In the experiment, the battery state of charge is divided into five equal intervals, and the detection accuracy of each method is tested independently in each interval. Figure 2 The middle line chart shows that the method of the application maintains a high detection accuracy in all state of charge intervals, especially in the extreme state of charge interval, although there is a slight decline, but it is still significantly better than other methods.

[0090] Further, referring to Figure 3 The provided abnormal sample number distribution diagram shows the actual number distribution of abnormal samples in each state of charge interval, in units of pieces, and it can be seen that abnormal samples are more distributed in the extreme state of charge interval, and the comprehensive Figure 2 、 Figure 3 The two graphs show that the method of the application performs stably and has high performance in different state of charge intervals, which shows that the dynamic weight integration mechanism of state of charge perception and the statistical feature extraction dynamically associated with the state of charge effectively enhance the adaptability to abnormal patterns in different operating states. The performance of the conventional method decreases significantly in extreme working conditions, highlighting the robustness advantage of the method of the application under complex operating conditions.

[0091] Further, the early abnormal pattern of the battery is sparse, weak and strongly coupled with complex working conditions, and direct use of a single model is prone to overfitting on most "normal" samples, and cannot fully extract feature combinations with high discriminability for a small number of "early abnormal" samples. At the same time, the model lacks adaptability to battery aging cumulative effect and different operating stages, resulting in lag or false alarm in early warning. For the above step S108, the embodiment of the application inputs the target feature vector into the pre-constructed battery anomaly diagnosis model, determines the initial probability sequence of the target feature vector through the base learner of the battery anomaly diagnosis model; generates the initial output by modeling the time sequence dependence corresponding to the initial probability sequence through the gated recurrent unit network of the battery anomaly diagnosis model; calibrates the initial output by expert priori to determine the battery anomaly probability corresponding to the target feature vector. The initial probability sequence includes cluster prediction probabilities corresponding to multiple clusters of the target feature vector. The embodiment of the application uses multiple base learners to perform parallel prediction on different feature clusters of the target feature vector, and can obtain multiple perspective initial probability sequences. Further, the time dependence of the probability sequence can be captured through the gated recurrent unit (GRU) network to realize dynamic information fusion. Further, the fusion result is calibrated in combination with the experience rules of experts in the battery field to eliminate unreasonable prediction, and then the anomaly probability that fits the actual operating state of the battery is output. The anomaly probability not only ensures the comprehensiveness of feature interpretation, but also improves the stability and reliability of probability output.

[0092] Corresponding to the above steps, the application adopts a three-level ensemble learning architecture: first, a pool of base learners with structural differences is constructed based on derived features to capture diverse abnormal discrimination patterns, then a novel adaptive boosting strategy based on battery state of charge evolution and sample local anomaly factor weighting is adopted to dynamically adjust the base learner weights, strengthen the attention to difficult-to-classify, potentially early abnormal samples, and through a gated recurrent unit network, the time series prediction sequences of all base learners are fused, combined with expert prior based on operating phase division for decision calibration, and an abnormal probability sequence with time series consistency and working condition adaptability is output. In an embodiment, the target feature vector corresponding to the to-be-tested data is input into the trained ensemble learning model, the model internally sequentially performs feature-guided heterogeneous base learner prediction, state of charge-aware dynamic weight integration (weight reference for online incremental learning or historical data analysis), and time series gated fusion and expert prior decision calibration, and outputs an abnormal probability value between 0 and 1, which quantitatively represents the possibility of the battery being in an abnormal state at the current time or current analysis period. The early warning system continuously receives the abnormal probability sequence output by the model and applies a preset warning strategy, for example, a probability threshold (such as 0.7) can be set, when the abnormal probability of consecutive multiple time points exceeds the threshold, or the probability value sharply rises in a short period of time, the system triggers different levels of warning signals (such as prompt, warning, serious alarm). At the same time, the system can record and visualize the change curve of the abnormal probability, help the operation and maintenance personnel trace back the abnormal starting time and evolution trend, so as to realize early warning of battery abnormalities from weak signs to clear alarm, and provide key decision support for preventive maintenance and safety management of the battery system.

[0093] Wherein, the base learner generates multiple cluster features by clustering the training sample set corresponding to the operation data, and then constructs for each cluster feature. K-means clustering algorithm can be used to cluster the standardized feature matrix, and the number of clusters K is determined by the elbow rule (usually K=3-5, corresponding to time domain, frequency domain, and correlation domain feature types). After clustering, the features in each cluster are sorted by variance contribution, and the top 80% of the features with the highest variance contribution are retained to form multiple cluster features with clear structure and prominent commonality (such as cluster 1 for time domain feature cluster, cluster 2 for frequency domain feature cluster), reducing feature redundancy. Further, according to the characteristics of each cluster feature, a dedicated base learner is matched, such as using a heterogeneous learner combination strategy to improve model diversity: time domain feature cluster (numerical, regularity) matches random forest learner; frequency domain feature cluster (high-dimensional, nonlinear) matches lightweight neural network (such as 3-layer MLP); time series coupled feature cluster (correlation type feature) matches support vector machine (SVM). The initial parameters of each base learner can be preliminarily optimized by grid search (such as setting the number of decision trees of the random forest to 100 and the number of hidden layer neurons of the MLP to 64). Further, a dynamic weight updating mechanism is used to determine the dynamic weight of the base learner, and the base learner is trained; and a battery abnormality diagnosis model is constructed based on the trained base learner.

[0094] The dynamic weight updating mechanism is used to ensure the accuracy of the battery abnormality diagnosis model. The embodiments of the present application combine the essential characteristics of the feature clusters and the importance of the running scenarios, so that the model can focus on the diagnosis results corresponding to the high-value features. Among them, based on the outlying degree corresponding to the cluster feature, the local anomaly factor of the cluster feature is determined; according to the importance of the state of charge interval to which the cluster feature belongs, the interval weight corresponding to the current cluster feature is determined; and the interval weight and the local anomaly factor are used to dynamically update the weight coefficient of the base learner, to generate the dynamic weight of the base learner. The embodiments of the present application use two-dimensional indicators of cluster feature outlying degree and state of charge (SOC) interval importance to make the base learner weight not only fit the essential attributes of the features, but also match the actual running requirements, solving the limitations of fixed weight or single-dimensional weight.

[0095] In specific implementation, the model construction steps are as follows:

[0096] 1) Feature-guided heterogeneous base learner generation

[0097] A set of complementary base learners are automatically generated by using the nonlinear relationship between the dimensions of the derived feature vectors and the differences in statistical characteristics, avoiding the bias caused by manual selection of models and ensuring the diversity of the integrated system. This is achieved by feature clustering and model type mapping, which is represented as:

[0098]

[0099] In the formula, denotes the generated ensemble of base learners, which capture diverse anomaly discrimination patterns in the derived feature vector through different model architectures to improve the robustness and generalization capability of the ensemble model; denotes the set union operation; denotes the index of the base learner, ; denotes the total number of generated base learners, which is set to 15 by default; denotes the th base learner, which is a mapping function from the derived feature vector to its output, the anomaly probability estimate ; denotes the mapping relationship; denotes the anomaly probability prediction value of the th base learner for the current sample, ; denotes the distribution denotes the generation probability distribution of the base learner, denotes the hyperparameter set of the th base learner, such as the maximum depth of the decision tree, the number of hidden layer neurons of the neural network, etc. denotes the model type function, which assigns specific model architectures to base learners according to feature clustering results, such as gradient boosting trees, multi-layer perceptrons, linear discriminant analysis, etc. denotes the feature clustering on the derived feature vector , which divides the features into clusters, specifically by taking all dimension features of as input, dividing them into clusters according to feature similarity, and features within each cluster have high similarity, while features between clusters have large differences. Unsupervised clustering algorithms can use K-means clustering or hierarchical clustering. denotes the number of clusters for feature clustering, which is set to 3 by default.

[0100] It should be noted that the base learner is a basic classifier in ensemble learning, which can be a gradient boosting tree, a multi-layer perceptron, a linear discriminant analysis, a support vector machine, or a random forest, etc. It should also be noted that the derived feature vector usually contains three main features: coupled features, frequency domain entropy features, and dynamic statistical features. Setting the number of clusters for feature clustering to 3 clusters can correspond to these three types of features, thereby better matching the feature characteristics when assigning model types. Further, the base learner is trained on the training set by minimizing its specific loss function , where is the The true labels for each training sample are 0 for normal and 1 for abnormal. The number of training samples, For the first Derived feature vectors of each training sample For the index of the training samples, .

[0101] For the training dataset, the collected raw time-series data can be segmented and labeled. For example, the long-term series is first divided into sliding windows of fixed duration or fixed number of data points. Each window constitutes a sample segment to be analyzed. Then, each sample segment is labeled, with the labeling category determined based on the battery's actual operating state and known fault or anomaly records. It is mainly divided into two categories: "normal" and "abnormal." The "normal" category indicates that the battery performance parameters are within the safe threshold and no known faults have occurred during this time period. The "abnormal" category corresponds to periods when the battery exhibits early performance degradation or minor faults. The criteria for this can be combined with features such as abnormal voltage drops, abnormal temperature increases, accelerated capacity decay curves, and internal resistance growth exceeding thresholds, or aligned with the time points of early fault symptoms recorded in actual maintenance records. Furthermore, the constructed training dataset consists of a large number of labeled samples. Each sample contains a multi-dimensional time-series data segment (current, voltage, temperature, state of charge) and its corresponding category label (0 indicates normal, 1 indicates abnormal).

[0102] 2) Dynamic weighted integration of state-of-charge sensing

[0103] The conventional AdaBoost algorithm assigns the same weight increase to all misclassified samples, without considering the impact of the state of charge during battery operation on the difficulty of anomaly representation or the anomaly potential of the samples themselves.

[0104] This invention employs a dynamic weight update mechanism, which correlates the weight increase with the charge state interval to which the sample belongs and its local anomaly factor values. This makes the integration process more focused on potential anomalous samples under key operating states, as follows:

[0105]

[0106] In the formula, Indicates the first In the first round of integration iteration The weight of each training sample represents the importance of that sample in the current ensemble learning. The higher the weight, the more difficult or crucial the sample is to classify, and the more attention the subsequent base learner will pay to these samples. Indicates the first In the first round of integration iteration The weights of each training sample; This represents the index of the iteration rounds in ensemble learning. ; This represents the total number of iterations in the ensemble learning, preferably set to 50 to 100. The specific value depends on the number of training samples and the model complexity, and is usually selected through cross-validation to ensure that the ensemble model converges and does not overfit. Represents the natural exponential function; Indicates the first The current weak learner in the round of iteration The weighting coefficients are calculated as follows: ; for The weighted error rate is obtained by calculating the classification error rate of the current weak learner on the training samples, specifically the sum of the weights of all misclassified samples; It is an indicator function that returns 1 if the condition inside the parentheses is true, and 0 otherwise. Indicates the first The true label for each sample is either 0 or 1, where 0 indicates normal and 1 indicates abnormal. Indicates as of the date In round 1, the current ensemble model is for the 1st... The predicted category of each training sample is specifically determined by analyzing the previous... The threshold is determined by weighted voting results of all base learners. If the weighted probability is greater than 0.5, the prediction is 1; otherwise, it is 0. Indicates the first The charge state interval of each sample The corresponding weights are used to emphasize the importance of samples in extreme charge state intervals, and the calculation method is the same as that of the first... Fusion weights corresponding to each charge state interval same; Indicates based on sample The state of charge value determines the index of the state of charge interval to which it belongs; This represents the adjustment coefficient for local anomaly factors, used to control the intensity of the influence of local anomaly factors on weight updates. The default value is 0.5. Indicates the first The local anomaly factor value corresponding to the nth sample represents the degree of outlier of that sample point in its original time series context. The calculation method is the same as that of the nth sample. Voltage gradient values ​​at each time point Local outlier same; and respectively represent the minimum and maximum of all training sample local anomaly factor values. Through this rule, for misclassified samples, if they are in the extreme state of charge interval or the original signal shows high outlying, they will get a larger weight increase in the next round of iteration, driving the subsequent base learners to focus on learning the patterns of these "key and difficult" samples.

[0107] In specific implementation, when judging whether the sample is in the extreme state of charge interval, it is realized by comparing its state of charge value with the threshold value, for example, the state of charge close to 0 or 1 is considered as the extreme interval, and the original signal shows high outlying is judged by the local anomaly factor value, the larger the local anomaly factor value, the higher the outlying.

[0108] 3) Time series gating fusion and expert prior decision calibration

[0109] After round of iteration, a set of weighted base learners is obtained, and the weighted voting result is ; wherein is the total number of base learners, that is, the number of base learners in ensemble learning; is the index of the base learner, ; is the th base learner, which is a classification model; is the weight of the th base learner, which is calculated by the AdaBoost algorithm; is the abnormal probability after weighted voting, that is, the weighted average of all base learner prediction probabilities; is the abnormal probability prediction value of the th base learner for the sample.

[0110] Further, a time series gating fusion network is adopted, and a decision calibration is performed in combination with an expert prior in the running stage, and a smooth and consistent with working condition common sense abnormal probability sequence is output, represented as:

[0111]

[0112] In the formula, represents the calibrated abnormal probability output at time step , which is calibrated through time series fusion and expert prior, so that the probability sequence is smoother and more consistent with the battery running common sense; is the time step index, ; is the total length of the predicted time series; represents a Sigmoid activation function, which maps the output to the interval; represents the weight matrix of the output layer of the gated recurrent unit network, which are trainable parameters; This represents the bias vector of the output layer of the gated recurrent unit network, which is a trainable parameter; This represents a gated recurrent unit network used to model the temporal dependencies of the predicted probability sequences of the base learners; Indicates from time step 1 to time step 2. Weighted voting probability of all base learners The sequence formed; This indicates that the gated recurrent unit network is at time step The hidden state; This indicates the total number of preset battery operation phase types, for example... It can represent three typical stages: constant current charging, constant power discharging, and resting. Indexes representing runtime phase types, ; Indicates at time step The sample was determined to belong to the first The probability of each operational phase is given by a pre-defined lightweight phase classifier, such as a shallow network classifier based on current, voltage sign, and rate of change. Indicates at time step The battery's operating phases, such as charging, discharging, or resting; Indicates at time step The anomaly probability after weighted voting of the base learners, i.e. At time step The value; Indicates the first The "expert prior" calibration function corresponding to the class's runtime phase outputs a calibration offset. For example, during the charging phase, if the voltage continues to rise slowly but... A sudden increase, A negative offset may be applied to suppress false alarms caused by the rise in normal constant current charging voltage.

[0113] It should be noted that the specific structure of the gated recurrent unit network includes update gates and reset gates, which are used to capture long-term dependencies in the time series. The gated recurrent unit network receives the weighted voting probability sequence of the base learner as input and outputs the hidden state for final prediction.

[0114] Furthermore, refer to Figure 4The embodiment of the present application also shows the early abnormality detection ability of different methods in the actual battery operation process through time series analysis, especially the early warning and timeliness of the abnormality occurrence time point. The experiment adopts battery voltage time series data containing normal fluctuations and abnormal drops, and the time unit is minute. The abnormal probability curves output by the method of the present application, the convolutional neural network long short-term memory network method and the conventional statistical method are compared. The conventional statistical method adopts a threshold determination method based on sliding window statistical features and logistic regression. For the original time series data of current, voltage, temperature and state of charge, a fixed length sliding window is used for segmentation. Then in each window, the mean, standard deviation, maximum, minimum and range of each parameter are calculated independently, and these statistical quantities are simply spliced and input into a logistic regression classifier for normal and abnormal judgment.

[0115] The battery voltage time series graph shows the change of voltage with time, and the red shaded area identifies the period of actual abnormality. The early abnormality detection probability time series comparison graph of different methods shows the abnormal probability output corresponding to the three methods. In the experimental configuration, all methods receive the same voltage sequence as input and output an abnormal probability value between zero and one. From the probability curve, it can be seen that the method of the present application starts to rise in probability about 5 minutes before the actual abnormality occurs, maintains a high probability output during the abnormality duration, and gradually falls in probability after the abnormality ends. The curve is relatively smooth and stable throughout the process. The convolutional neural network long short-term memory network method reaches a high probability only after the abnormality occurs, which has obvious detection lag. The conventional statistical method frequently produces probability spikes and a large number of false positives during the entire time period, and the probability fluctuation during the actual abnormality period is also large. The experimental results show that the method of the present application captures early abnormality signs in the frequency domain through adaptive wavelet packet entropy extraction, smoothes the probability sequence through time series gating fusion network, and reduces false positives caused by working condition fluctuations by combining expert prior decision calibration, thereby realizing earlier, more accurate and more stable early warning. The conventional technology either lags behind in warning or has a high false positive rate, and cannot meet the dual demands of timeliness and reliability in actual engineering applications.

[0116] Further, the cross-entropy loss function can be used as the training optimization target of the base learner to train the base learner. In each training iteration, in order to quantify the gap between the prediction output of the ensemble learning model (including all base learners and the time-gated fusion network) and the true label, the loss function needs to be calculated. The present application uses the cross-entropy loss function as the optimization target of the model training. Specifically, for each sample in the training data set, the calibrated abnormal probability output after the time-gated fusion and expert prior calibration, and the known true class label of the sample are used to calculate the cross-entropy loss value of the sample. The size of the loss value directly reflects the uncertainty or error degree of the model prediction for the sample: when the model predicts a high abnormal probability for a normal sample or a low abnormal probability for an abnormal sample, a larger loss value will be generated; on the contrary, when the prediction probability is highly consistent with the true label, the loss value tends to zero. After each complete training iteration (i.e. all training samples are processed by the model once), the cross-entropy loss values of all training samples are summed or averaged to obtain the total training loss of the current model under this iteration. Then, using the back propagation algorithm, the gradient of all trainable parameters in the model with respect to the loss function is calculated according to the total loss value. The gradient represents the direction and amplitude of parameter adjustment to reduce the loss. Then, using the stochastic gradient descent optimization algorithm, all parameters in the model are updated at once according to the calculated gradient and the set learning rate. This parameter update process completes one model iteration. Through repeated execution of the cycle of "forward propagation to calculate prediction and loss, backward propagation to calculate gradient, and optimizer to update parameters", the model continuously adjusts its internal representation and decision boundary. The goal is to make the total cross-entropy loss function value continue to decrease, thereby improving the model's ability to distinguish between normal and early abnormal states of the battery.

[0117] The iteration process of model training is not infinite, and a clear stop condition needs to be set to prevent overfitting or resource waste. The stop iteration judgment of the present application is mainly based on the following two criteria: criterion one is based on the preset maximum iteration round, when the training iteration round reaches the preset upper limit value, the training is forced to stop regardless of the model performance, to ensure that the training process is completed within a controllable time; criterion two is based on the early stopping mechanism, which is realized by monitoring the performance of the model on a validation data set independent of the training set. After each training iteration is completed, in addition to calculating the training loss, the model with the current parameters is also evaluated on the validation set to calculate the cross-entropy loss or classification accuracy and other performance indicators on the validation set. The early stopping mechanism continuously tracks the trend of the performance indicator of the validation set, if the performance of the validation set no longer improves or even starts to decline in continuous multiple iterations (such as 10 rounds or 20 rounds), while the training loss is still continuously reduced, it indicates that the model may start to overfit the noise or specific patterns in the training data, and its generalization ability no longer improves, then the training process will be terminated in advance, and the model parameters corresponding to the iteration round with the best performance on the validation set will be rolled back, which will be used as the final trained model. By combining the maximum iteration round and the early stopping mechanism based on the performance of the validation set, the model can be fully learned to the effective mode while effectively avoiding overfitting, so as to obtain a battery abnormal early warning model with stronger generalization ability.

[0118] Further, the embodiments of the present application also compare the F1 values of different methods in each operating phase to evaluate the recognition performance of different abnormal detection methods in the three typical operating phases (charging, discharging and standing) of the battery. Referring to Figure 5 The comparison methods include: fixed window cleaning + logistic regression: using logistic regression classification after cleaning the data with a fixed time window; independent features + support vector machine: inputting the support vector machine after independently extracting each parameter feature; dynamic time warping + random forest: only using time alignment but not coupling feature enhancement; wavelet transform + gradient boosting: extracting frequency domain features based on fixed basis wavelet transform; the present application technology (integrated early warning): completely realizing dynamic alignment, feature coupling and integrated learning.

[0119] The vertical coordinate in the figure is the F1 value (comprehensive evaluation index), and the horizontal coordinate is the operating phase. The experimental results show that the F1 value of the present application technology is the highest in the charging phase, which is significantly better than other methods, proving the effectiveness of multi-stage adaptability and feature enhancement.

[0120] In summary, the embodiments of the present application have the following innovations compared with the prior art:

[0121] 1. In view of the problems of asynchronous and noise interference of battery sensor data, the dynamic time warping algorithm is innovatively used to correct the timing dislocation, and the adaptive cleaning strategy based on local outlier factor is combined to avoid smoothing out weak mutation signals related to early abnormalities while denoising.

[0122] 2. By constructing a high-order spatiotemporal feature tensor and performing constrained tensor decomposition, the coupling relationship among parameters, time points and feature types is jointly learned; at the same time, the adaptive wavelet packet entropy extraction and the dynamic statistical features associated with the state of charge are combined to form a comprehensive feature representation that is more sensitive to early abnormalities.

[0123] 3. The improved conventional AdaBoost algorithm is used to dynamically associate sample weight updating with state of charge interval importance and local anomaly factor value, so that the integration process focuses more on difficult-to-classify samples under extreme working conditions and potential outliers, and the learning ability for sparse abnormal patterns is improved.

[0124] 4. The gating recurrent unit network is used to fuse the timing prediction results of the base learner, and the decision calibration is performed in combination with the expert priori rule based on the running phase classification, so as to realize the smooth output of the abnormal probability sequence and the false alarm suppression, and enhance the timing consistency and working condition adaptability of the early warning result.

[0125] Based on the above system embodiment, the embodiment of the present application also provides a battery abnormal early warning device based on integrated learning, which is described with reference to Figure 6 The device comprises: a data acquisition module 10 for acquiring running data of a target battery; the running data comprises multi-source time series data of the target battery; a preprocessing module 20 for performing path alignment and data cleaning on each attribute data based on the instantaneous change of the sampling value of each attribute data of the running data, and generating preprocessed data; a data processing module 30 for calculating the time-varying statistics of the preprocessed data for a preset state of charge, and performing feature splicing on the time-varying statistics and the adaptive entropy features of the preprocessed data in the frequency domain to generate a target feature vector; an execution module 40 for inputting the target feature vector into a pre-constructed battery abnormal diagnosis model, determining the battery abnormal probability corresponding to the target feature vector through the battery abnormal diagnosis model, and performing abnormal early warning on the target battery based on the battery abnormal probability; wherein the battery abnormal diagnosis model is constructed based on an integrated learning architecture, and the integrated learning architecture comprises a base learner pool, and an adaptive weight adjustment strategy and a result fusion unit of the base learner pool; each base learner included in the base learner pool is used to process sample data of a preset structure.

[0126] The battery abnormal early warning device based on integrated learning provided by the embodiment of the present application has the same implementation principle and technical effects as the foregoing method embodiment. For brevity, the part of the device embodiment not mentioned can be referred to the corresponding content in the foregoing method embodiment.

[0127] The embodiment of the present application also provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method shown in the above Figure 1 The embodiment of the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the method shown in the above Figure 1 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in the above Figure 7 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in the above, wherein the electronic device comprises a processor 101 and a memory 100, the memory 100 stores computer executable instructions capable of being executed by the processor 101, and the processor 101 executes the computer executable instructions to implement the method shown in the above Figure 1 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in the above

[0128] In the embodiment shown in the above Figure 7 The electronic device also comprises a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103 and the memory 100 are connected through the bus 102. The memory 100 can comprise a high-speed random access memory (RAM) and can also comprise a non-volatile memory, for example, at least one disk memory. The communication connection between the system network element and at least one other network element is realized through the at least one communication interface 103 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus, wherein the AMBA defines three kinds of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus and an AXI (Advanced eXtensible Interface) bus. The bus 102 can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 7Only one bidirectional arrow is shown, but this does not mean there is only one bus or only one type of bus. The processor 101 can be an integrated circuit chip with signal processing capability. In the implementation process, the steps of the above method can be completed by the integrated logic circuit of the hardware in the processor 101 or the instructions in the form of software. The processor 101 described above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or other mature storage medium in the art. The storage medium is located in the memory, and the processor 101 reads the information in the memory, and combines its hardware to complete the method shown in the foregoing Figure 1

[0129] ​The computer program product of the battery abnormal early warning method and device based on ensemble learning provided by the embodiment of the application comprises a computer readable storage medium storing program codes, the program codes comprise instructions for executing the method described in the foregoing method embodiment, and specific implementation can be referred to the method embodiment, and details are not described herein again. Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiment, and details are not described herein again. In addition, in the description of the embodiment of the application, unless otherwise explicitly specified and limited, the terms "mounting", "connection" and "connection" should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integral connection; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances. The functions can be realized in the form of software function units and sold or used as independent products, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the application or the part of the technical solutions which essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of software products, the computer software product is stored in a storage medium, and comprises a plurality of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the application. The foregoing storage medium comprises a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk and various program code storage media.

[0130] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance. Finally, it should be noted that: the above examples are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, and are not a limitation on it, the protection scope of the present application is not limited to this, although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand: any person skilled in the art within the technical range disclosed by the present application, it can still modify or easily think of changes to the technical solutions recorded in the foregoing examples, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A battery abnormality early warning method based on ensemble learning, characterized in that, The method comprises: acquiring operation data of a target battery; the operation data comprises multi-source time series data of the target battery; based on the instantaneous change of the sampling value of each attribute data of the operation data, path alignment and data cleaning are performed on each attribute data to generate preprocessed data; the attribute data includes voltage attribute data, current attribute data, temperature attribute data and state of charge attribute data; dividing the preprocessed data into multiple intervals according to the battery state of charge value, and calculating the current distribution characteristics of each interval; based on the state of charge center value corresponding to each interval, the current distribution characteristics are weighted and fused to generate time-varying statistics corresponding to the preprocessed data; the time-varying statistics and the adaptive entropy features of the preprocessed data in the frequency domain, and the time series coupling features corresponding to the preprocessed data are spliced to generate a target feature vector; wherein the adaptive entropy features are determined by adaptive wavelet packet decomposition of the preprocessed data; inputting the target feature vector into a pre-constructed battery anomaly diagnosis model, determining the battery anomaly probability corresponding to the target feature vector through the battery anomaly diagnosis model, and performing anomaly warning on the target battery based on the battery anomaly probability; wherein the battery anomaly diagnosis model is constructed based on an ensemble learning architecture, the ensemble learning architecture includes a base learner pool, an adaptive weight adjustment strategy of the base learner pool and a result fusion unit; each base learner included in the base learner pool is used to process sample data of a preset structure; wherein the construction method of the battery anomaly diagnosis model comprises: acquiring a training sample set corresponding to the operation data; performing feature clustering on the training sample set to generate multiple cluster features, and constructing a base learner for each cluster feature; determining the dynamic weight of the base learner by using a dynamic weight updating mechanism, and performing model training on the base learner; constructing a battery anomaly diagnosis model based on the trained base learner; the step of determining the dynamic weight of the base learner by using a dynamic weight updating mechanism comprises: based on the outlying degree corresponding to the cluster feature, determining the local anomaly factor of the cluster feature; determine the interval weight corresponding to the current cluster feature according to the importance of the state of charge interval to which the cluster feature belongs; the interval weight and the local anomaly factor are used to dynamically update the weight coefficient of the base learner to generate the dynamic weight of the base learner; wherein for the misclassified samples, if they are in the extreme state of charge interval or the original signal shows high outlying, they will obtain a larger weight increment in the next iteration.

2. The method of claim 1, wherein, based on the instantaneous change of the sampling value of each attribute data of the operation data, path alignment and data cleaning are performed on each attribute data to generate preprocessed data; the attribute data includes voltage attribute data, current attribute data, temperature attribute data and state of charge attribute data; based on the sampling time of the operation data, determine the instantaneous change of the sampling value corresponding to each attribute data of the operation data; the optimal alignment path corresponding to the sampling value instantaneous change is calculated by using a preset dynamic time warping algorithm; Aligning, according to the optimal alignment path, a sampling time of each of the attribute data to generate first preprocessed data; According to the local outlier factor in the first preprocessed data, the running data is data cleaned to generate preprocessed data.

3. The method of claim 1, wherein, The step of calculating the time sequence coupling feature corresponding to the preprocessed data comprises: Calculating derivative features of the preprocessed data in multiple dimensions, and reconstructing the preprocessed data into a spatio-temporal feature tensor based on the derivative features; the multiple dimensions include parameter type, time evolution and feature morphology dimension; Tensor decomposition is performed on the spatio-temporal feature tensor to extract a core coupling feature tensor of the spatio-temporal feature tensor; The core coupling feature tensor is flattened into a one-dimensional vector to generate a time sequence coupling feature of the preprocessed data.

4. The method of claim 1, wherein, The step of determining the adaptive entropy feature of the preprocessed data in the frequency domain by adaptive wavelet packet decomposition of the preprocessed data comprises: Wavelet packet decomposition is performed on the preprocessed data to determine node wavelet coefficients corresponding to a plurality of frequency bands respectively; An optimal wavelet packet decomposition tree corresponding to the node wavelet coefficients is determined by minimizing a cost function; The entropy values of all leaf nodes in the optimal wavelet packet decomposition tree are extracted to form an adaptive entropy feature corresponding to the preprocessed data.

5. The method of claim 1, wherein, The step of determining the battery abnormality probability corresponding to the target feature vector through the battery abnormality diagnosis model comprises: The target feature vector is input into a pre-constructed battery abnormality diagnosis model, and an initial probability sequence of the target feature vector is determined through a base learner of the battery abnormality diagnosis model; wherein the initial probability sequence includes cluster prediction probabilities corresponding to a plurality of clusters of the target feature vector respectively; The initial probability sequence is modeled through a gating recurrent unit network of the battery abnormality diagnosis model to generate an initial output; The initial output is calibrated by an expert prior to determine the battery abnormality probability corresponding to the target feature vector.

6. The method of claim 1, wherein, The step of model training of the base learner comprises: A cross-entropy loss function is used as a training optimization target of the base learner to train the base learner.

7. An integrated learning-based early warning device for battery abnormalities, characterized in that, The device comprises: A data acquisition module for acquiring running data of a target battery; the running data includes multi-source time sequence data of the target battery; A preprocessing module for aligning and data cleaning of each attribute data based on instantaneous changes of sampling values of each attribute data of the running data to generate preprocessed data; the attribute data includes voltage attribute data, current attribute data, temperature attribute data and state of charge attribute data; The data processing module is configured to divide the preprocessed data into multiple intervals according to battery state of charge values, calculate current distribution characteristics of each interval, weight and fuse the current distribution characteristics based on a state of charge center value corresponding to each interval, generate a time-varying statistic corresponding to the preprocessed data, and splice the time-varying statistic with an adaptive entropy feature of the preprocessed data in a frequency domain and a time sequence coupling feature corresponding to the preprocessed data to generate a target feature vector. The adaptive entropy feature is determined by adaptively decomposing the preprocessed data using a wavelet packet. The execution module is configured to input the target feature vector into a pre-constructed battery anomaly diagnosis model, determine a battery anomaly probability corresponding to the target feature vector by using the battery anomaly diagnosis model, and perform an anomaly warning on the target battery based on the battery anomaly probability. The battery anomaly diagnosis model is constructed based on an ensemble learning architecture, which includes a base learner pool, an adaptive weight adjustment strategy of the base learner pool, and a result fusion unit. Each base learner in the base learner pool is configured to process sample data of a preset structure. The construction method of the battery anomaly diagnosis model includes the following steps: obtaining a training sample set corresponding to the operation data, performing feature clustering on the training sample set to generate multiple cluster features, constructing a base learner of each cluster feature, determining a dynamic weight of the base learner by using a dynamic weight updating mechanism, and performing model training on the base learner based on the dynamic weight. The step of determining the dynamic weight of the base learner by using the dynamic weight updating mechanism includes the following steps: determining a local anomaly factor of the cluster feature based on a degree of outlying corresponding to the cluster feature, determining an interval weight corresponding to the current cluster feature according to an importance of a state of charge interval to which the cluster feature belongs, and dynamically updating a weight coefficient of the base learner by using the interval weight and the local anomaly factor to generate the dynamic weight of the base learner. For a misclassified sample, if the sample is in an extreme state of charge interval or the original signal shows high outlying, the sample obtains a greater weight increment in the next iteration.

Images