A battery safety risk prediction method based on full life cycle data

Abstract

The application discloses a battery safety risk prediction method based on full life cycle data, and relates to the technical field of data analysis. By dividing multi-source heterogeneous data into a hierarchical structure of different time scales according to variable evolution characteristics, and constructing an evolution chain based on cross-scale trend coupling relationship, not only can the risk precursor events occurring at different stages and different time scales be accurately identified, but also the linkage risk index can be calculated in time when the chain evolution is activated, and then the early warning of potential safety risks of the battery can be realized. Compared with the existing method which is mostly based on unified time window modeling, the application can effectively capture the nonlinear interaction relationship between long-term slowly changing factors and short-term sudden factors, solve the risk path breaking problem caused by scale mismatch in the prior art, significantly improve the risk identification ability and warning accuracy in complex scenarios, and help prevent serious failures such as thermal runaway caused by the accumulation of multi-stage weak anomalies.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and specifically to a method for predicting battery safety risks based on full lifecycle data. Background Technology

[0002] Currently, the widespread application of battery systems in new energy vehicles, energy storage power stations, and other scenarios has made battery safety a core focus of industry attention. To improve the safety and reliability of battery operation, an increasing number of systems are introducing battery safety risk prediction methods based on full lifecycle data. This method collects and integrates multi-source data from various stages of battery lifecycle, including production, transportation, installation, use, and maintenance, such as voltage, current, temperature, internal resistance, SOC / SOH evolution, and historical alarm logs, to construct a multi-dimensional feature map. Combined with time series modeling, trend analysis, or deep learning methods, it assesses potential battery safety risks and provides early warnings of abnormal states. This method not only expands the temporal scope of risk identification but also enhances the model's ability to perceive latent faults in complex scenarios, gradually becoming an important technical component of intelligent battery management systems (BMS).

[0003] However, in reality, key risk factors during battery operation often exhibit a long-term evolution coupled with short-term abrupt triggering mechanisms. For example, a seemingly harmless physical impact, followed by months of thermal gradient accumulation, and then a sudden high-rate discharge event, can ultimately lead to thermal runaway. Since these three events occur at different time scales and stages, conventional models often fail to identify them as a single risk chain. Current risk prediction models are mostly based on uniform time windows or data of equal granularity, making it difficult to effectively capture the nonlinear interactions between variables across cycles and scales. This leads to the omission or misjudgment of some key warning signs, thus affecting the overall accuracy of early warnings and the timeliness of response. Summary of the Invention

[0004] The purpose of this invention is to address the problem mentioned in the background art that most current risk prediction models are based on a unified time window or equal-granularity data modeling, which makes it difficult to effectively capture the nonlinear interaction between variables across cycles and scales, resulting in the omission or misjudgment of some key early warning information. Therefore, this invention proposes a battery safety risk prediction method based on full life cycle data.

[0005] In terms of implementation, this invention provides a battery safety risk prediction method based on full lifecycle data, the method comprising: S1: Acquire multi-source data throughout the entire battery lifecycle and process it uniformly according to a preset structural standard to form a data set indexed by the battery identifier; S2: Divide the variables in the dataset into three levels according to their rate of change: slow-changing variables, medium-changing variables, and fast-changing variables, and extract the evolution trend information of variables at each level. S3: Based on the trend information of variables at each level, identify trend disturbance point pairs that have the same direction but occur at different times across time scales, forming a set of trend disturbance coupling points; S4: Based on the set of trend perturbation coupling points, construct a set of cross-scale evolution chains in historical samples or current data, which are connected in sequence by slow-changing variables, medium-changing variables and fast-changing variables. The chains reflect potential risk evolution paths. S5: Determine whether each chain in the evolution chain set is activated at the current moment. If the trend state of all nodes in the chain meets the preset activation conditions, mark it as an activated chain and record the activation intensity. S6: Calculate the current linkage risk index of the battery based on the activation intensity of all activated chains, and output the risk index as the final prediction result of battery safety risk.

[0006] Optionally, S1: The steps of acquiring multi-source data throughout the entire battery lifecycle and processing it uniformly according to a preset structural standard to form a data set indexed by the battery identifier are as follows: Collect raw data of the battery during the manufacturing, transportation, installation, operation and maintenance stages. The raw data includes voltage, current, temperature, SOC, SOH, internal resistance, consistency parameters, environmental information and event logs. The raw data is grouped according to the life cycle stage, and field standardization mapping is performed on each group of data to convert the raw fields of each stage into a preset data structure field format, resulting in standardized data records with a unified structure. Using the battery's unique identifier as the primary index key, all standardized data records are integrated to construct a multi-variable, multi-time-point dataset, where each dataset includes the values ​​of multiple variables at different time points; To address the issue of inconsistent sampling frequencies among variables, a time synchronization method is employed for data alignment. Specifically, this includes: using forward imputation to fill in missing values ​​for slowly changing variables and using moving average interpolation to fill in missing values ​​for rapidly changing variables. The synchronized data set is established as a data structure mapping relationship with the battery's unique identifier as the key, serving as a unified input data source for subsequent variable trend extraction and risk prediction models.

[0007] Optionally, S2: The step of dividing the variables in the dataset into three levels according to their rate of change—slow-changing variables, medium-changing variables, and fast-changing variables—is as follows: From the data structure indexed by battery identifiers obtained in step S1, extract the time series of each variable on a unified time axis, and then normalize the time series to obtain a set of variable sequences, where the first... The first battery The variable in the first... The values ​​at each time point are denoted as ; For each variable sequence, its average rate of change within the sampling period is calculated as an indicator of the variable's evolution rate. The specific calculation formula is as follows: ; Indicates the first The average rate of change of each variable (evolutionary rate index); Indicates the total number of sampling time points; The variables are divided into three evolution levels based on their evolution rate index: if the evolution rate index of a variable is less than the first threshold, the corresponding evolution level of the variable is recorded as a slow-changing variable. If the evolution rate index of a variable is not less than the first threshold but less than the second threshold, then the corresponding variable evolution level is recorded as a medium-level variable. If the evolution rate index of a variable is not less than the second threshold, then the corresponding variable evolution level is recorded as a fast-changing variable; and the first threshold is less than the second threshold.

[0008] Optionally, the steps for extracting the evolution trend information of variables at each level are as follows: For each level of variable, a trend direction symbol sequence S(j) is constructed on a unified time axis to express the short-term change direction of the variable. The trend is calculated as follows: In the formula, The sign of the trend direction of the j-th variable at time point k is used to describe the micro-trend of the variable. +1 indicates an upward trend, -1 indicates a downward trend, and 0 indicates that it is stable and unchanged. The preset minimum significant change threshold is used to indicate the trend direction; the trend direction sign sequence of each variable is used as evolutionary trend information.

[0009] Optionally, S3: Based on the trend information of variables at each level, the steps to identify trend disturbance point pairs that have the same direction but occur at different times across time scales and to form a set of trend disturbance coupling points are as follows: Between each pair of slowly changing and rapidly changing variables, a trend direction matching and identification mechanism is established. For any slowly changing variable and any rapidly changing variable, based on their trend sign sequences, trend disturbance point pairs that satisfy the condition of consistent direction are found. ; That is, it exists: ,and ; In the formula, Indicates the first A slowly changing variable at time point Trend symbols; Indicates the first A fast variable at time point Trend symbols; and This serves as the time index for the occurrence of the corresponding trend abrupt change; A preset minimum time misalignment window is used to exclude synchronous changes at the same point in time, retaining only trend coupling point pairs with sequential logic; when the condition is met... and And time difference Within the preset maximum misalignment window When the range is within the specified range, the point pair is determined to be a valid trend perturbation coupling point, and is represented in the form of a quintuple ( , , , ,s) Add to the set of trend perturbation coupling points, where, Indicates the number of a slowly changing variable; Indicates the number of the rapidly changing variable; This indicates the point in time when a sudden change in the trend of a slowly changing variable occurs; , indicates the time point when the direction of the rapidly changing variable changes abruptly; s is the trend direction sign (+1 or -1), used to characterize whether the trend pair is rising or falling synchronously; All trend perturbation point pairs that meet the conditions are output as a set of trend perturbation coupling points for use in the construction of cross-scale evolution chains in subsequent steps, and the hierarchical source and temporal misalignment interval of each point pair are marked.

[0010] Optionally, S4: Based on the set of trend perturbation coupling points, the steps for constructing a set of cross-scale evolutionary chains in historical samples or current data, consisting of slowly changing variables, moderately changing variables, and rapidly changing variables connected in sequence, with the chains reflecting potential risk evolution paths, are as follows: Based on the set of trend disturbance coupling points output in step S3, extract all point pairs that satisfy the condition that there is a consistent trend direction and a time misalignment between the fast-changing variable and the slow-changing variable. Let each point pair be a quintuple ( , , , ,s), where, Indicates the number of the slowly changing variable. Indicates the number of the rapidly changing variable. This indicates the time point at which a sudden change in the trend of a slowly changing variable occurs. This indicates the time point at which abrupt changes in the trend of a rapidly changing variable occur, where 's' represents the same trend direction sign for both; in the current battery data or historical sample data, for each group ( , , , ,s), in the set of intermediate variables, search for variables that satisfy the condition of consistent trend direction and time point located at and Let the intermediate variable be numbered c1, and its trend sequence be S(c1, ...). If a time point exists 3. The following conditions must be met: Then it is assumed that variable c1 in time... The trend changes at point 3 can be considered intermediate nodes in this evolutionary path; under the above conditions, the variables... c1 They are respectively used as variables at three levels: slow-change, medium-change, and fast-change, in chronological order. Construct a cross-scale evolutionary chain, defined as follows: Among them, P x Let x represent the xth evolutionary chain, which consists of three variable trend mutation events at different scales, and these three events have the same trend direction and are connected in chronological order; All evolutionary chains that meet the above structure and conditions are collected into a cross-scale evolutionary chain set. Each chain in the set retains complete information such as variable number, mutation time point, and trend direction.

[0011] Optionally, S5: The step of determining whether each chain in the evolutionary chain set is activated at the current moment, and if the trend state of all nodes in the chain meets the preset activation condition, is to mark it as an activated chain and record the activation intensity. This step is based on the cross-scale evolutionary chain set constructed in step S4, for each chain P x It contains a trend-change triplet consisting of a slowly changing variable, a moderately changing variable, and a rapidly changing variable. Let the chain structure be: ,in, Indicates the number of the slowly changing variable. The point in time when its trend suddenly changed; Indicates the variable number in the intermediate variable. The point in time when its trend suddenly changed; Indicates the number of the rapidly changing variable. Its trend abrupt change point; at the current moment. To determine whether the chain meets the activation condition, specifically: obtain the trend direction symbol at the current moment from the variable trend direction symbol sequence. If all three variables in the chain satisfy the following condition at the current moment: If so, the chain is determined to be in an activated state, where The symbol represents the consistent direction of the original trend change in the chain (+1 indicates an upward trend, -1 indicates a downward trend).

[0012] Optionally, the steps for recording activation intensity are as follows: For each activated chain, calculate the trend drag synchronization index and the directional imbalance offset index for each activated chain, and add the trend drag synchronization index and the directional imbalance offset index to obtain the strength index.

[0013] Optionally, the calculation steps for the trend lag synchronization index are as follows: At the current detection moment, extract the indices of the slow-changing, medium-changing, and fast-changing variables contained in the activated chain, denoted as vs, vm, and vf, respectively. Then, extract the trend direction sequence of each variable for the L time points preceding the current detection moment, with the trend direction being +1, -1, or 0, used to represent the rising, falling, or no-trend change state of the variable at the corresponding time point. The sequence construction is as follows: , In the formula, , , These are respectively the numbers for slow-changing variables, medium-changing variables, and fast-changing variables; For the current detection time, Indicates the length of the trend observation window. Indicates the time of slow variable A. The trend direction, the intermediate variable B at time The trend direction, the rapidly changing variable C at time trend direction The trend direction of variable v at time t, taking values ​​of +1 / -1 / 0; Based on the trend direction sequence, a trend lag difference sequence is constructed. Specifically, for slow-changing and medium-changing variables, the difference between adjacent trend values ​​is calculated by subtracting the current trend value of the medium-changing variable from the previous trend value of the slow-changing variable. The same method is used for medium-changing and fast-changing variables, subtracting the current trend value of the fast-changing variable from the previous trend value of the medium-changing variable. The corresponding formula is: ; In the formula, The difference in trend between slow-changing and medium-changing variables under lag structure is given by the formula. The trend lag difference between medium- and fast-changing variables under a lag structure is used. Based on the trend lag difference, a trend consistency judgment result is constructed. Specifically, if the trend difference of any item is zero, the trend consistency result of the item is 1; if the absolute value of the difference is 1, the result of the item is 0.5; if the absolute value of the difference is 2, the result of the item is 0. The judgment result is calculated separately for the slow- and medium-changing variables and the medium- and fast-changing variables, respectively, to obtain two consistency scores corresponding to each time step. Multiply the two consistency scores at each time step to obtain the joint consistency score sequence at that time step; then, calculate the sum of the absolute values ​​of the differences between adjacent terms in the joint consistency score sequence to obtain the total number of bends in the trend consistency sequence, and normalize it according to the maximum number of bends allowed under the window to obtain the stability score of the trend drag structure. The closer the stability score is to 1, the more stable the trend drag structure is. The joint consistency score at each time step is incremented by 1 and then multiplied together. The natural logarithm of the product is then taken and divided by the product of the sequence length and the natural logarithm 2 to obtain the trend dragging continuous kernel value. The larger the kernel value, the more concentrated the trend consistency segment is on the time axis. Multiplying the stability score by the continuous kernel value yields the final trend drag synchronization index, which is a dimensionless quantity between 0 and 1. This index is used to characterize whether the activated chain possesses continuous and structurally stable misaligned trend transmission characteristics at the current moment.

[0014] Optionally, the calculation steps of the directional imbalance offset index are as follows: at the current detection time, extract the slow-changing variables, medium-changing variables and fast-changing variables contained in the activation chain, and construct the trend direction sequence of the three types of variables respectively. Each trend direction takes three states: rising, falling or no trend. The trend direction sequence is the time series of the variable within a preset window length range from the current detection time. Based on the trend direction sequence of each variable, identify the sub-intervals that maintain the same trend direction continuously. Represent each continuous trend segment as the interval in which the variable continuously rises or falls within a certain time period, and record the start and end time of each interval and the corresponding trend direction to form a set of continuous trend intervals for the variable. For any combination of two different variables, determine whether there exists a continuous trend interval of one variable that, under the condition that the trend direction is the same, completely covers the continuous trend interval of the other variable. If such a covering relationship exists, it is considered that the former exerts trend suppression on the latter. For each pair of variables, the trend suppression index is calculated. The trend suppression index is defined as the ratio of the number of trend intervals that are completely covered in the suppressed variable to the total number of trend intervals. It is used to measure whether the independent trend of the variable is dominated by the other variable. Among the three variables—slow-changing, medium-changing, and fast-changing—trend suppression mechanisms are constructed sequentially for three sets of variable pairs: slow-changing suppression mechanism against medium-changing, medium-changing suppression mechanism against fast-changing, and slow-changing suppression mechanism against fast-changing. The one with the largest suppression mechanism is taken as the dominant trend suppression strength of the current chain. Based on the trend-restraint relationship, determine whether there is a trend-dominant structure chain in the chain, that is, a certain variable not only exerts trend suppression on another variable, but the suppressed variable also continues to exert trend suppression on a third variable. If this structure is satisfied, the length of the trend-dominant structure chain is 2; if there is only a one-way suppression relationship, the chain length is 1; if there is no directional coverage relationship among the three variables, the chain length is 0. Multiply the length of the dominant structural chain by the strength of the dominant trend suppression, and divide by 2 to obtain the final directional imbalance offset index. This index is a dimensionless quantity between 0 and 1, used to measure whether there is a significant dominant phenomenon of a certain variable in the chain on the overall trend path. The larger the value, the more serious the directional offset of the chain and the more unbalanced the trend evolution structure.

[0015] Optionally, S6: Calculate the current linkage risk index of the battery based on the activation intensity of all activated chains, and output the risk index as the final prediction result of battery safety risk.

[0016] Divide the strength index of the activated chain by a constant 2 to obtain the normalized standard strength index, such that the value of the normalized standard strength index ranges from 0 to 1. The normalized standard strength index is used to represent the standard contribution strength of the chain to the overall risk. The mean of all normalized standard strength indices is used as the linked risk index. This linked risk index is compared to a preset threshold. If the linked risk index is not less than the preset threshold, it indicates a risk to battery safety, and this is output as the final prediction result. If the linked risk index is less than the preset threshold, it indicates no risk to battery safety, and this is output as the final prediction result. The beneficial effects of this invention are: This invention proposes a battery safety risk prediction method based on full lifecycle data. By dividing multi-source heterogeneous data into hierarchical structures at different time scales according to the evolutionary characteristics of variables, and constructing an evolutionary chain based on cross-scale trend coupling relationships, it can not only accurately identify risk precursor events occurring at different stages and time scales, but also calculate the linkage risk index in a timely manner when the chain evolution is activated, thereby achieving early warning of potential battery safety risks. Compared with most existing methods based on unified time window modeling, this invention can effectively capture the nonlinear interaction between long-term slowly changing factors and short-term abrupt factors, solving the problem of risk path breakage caused by scale mismatch in existing technologies. It significantly improves the risk identification capability and early warning accuracy in complex scenarios, and helps prevent serious failures such as thermal runaway caused by the accumulation of weak anomalies at multiple stages. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a battery safety risk prediction method based on full lifecycle data, provided in an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides a method for predicting battery safety risks based on full lifecycle data. See also... Figure 1 , Figure 1 A flowchart illustrating a battery safety risk prediction method based on full lifecycle data, provided in this embodiment of the invention. The method includes the following steps: S1: Acquire multi-source data throughout the entire battery lifecycle and process it uniformly according to a preset structural standard to form a data set indexed by the battery identifier; S2: Divide the variables in the dataset into three levels according to their rate of change: slow-changing variables, medium-changing variables, and fast-changing variables, and extract the evolution trend information of variables at each level. S3: Based on the trend information of variables at each level, identify trend disturbance point pairs that have the same direction but occur at different times across time scales, forming a set of trend disturbance coupling points; S4: Based on the set of trend perturbation coupling points, construct a set of cross-scale evolution chains in historical samples or current data, which are connected in sequence by slow-changing variables, medium-changing variables and fast-changing variables. The chains reflect potential risk evolution paths. S5: Determine whether each chain in the evolution chain set is activated at the current moment. If the trend state of all nodes in the chain meets the preset activation conditions, mark it as an activated chain and record the activation intensity. S6: Calculate the current linkage risk index of the battery based on the activation intensity of all activated chains, and output the risk index as the final prediction result of battery safety risk.

[0020] This invention provides a battery safety risk prediction method based on full lifecycle data. By dividing multi-source heterogeneous data into hierarchical structures at different time scales according to variable evolution characteristics, and constructing an evolutionary chain based on cross-scale trend coupling relationships, this method can not only accurately identify risk precursor events occurring at different stages and time scales, but also calculate the linkage risk index in a timely manner when the chain evolution is activated, thereby achieving early warning of potential battery safety risks. Compared with most existing methods based on unified time window modeling, this invention can effectively capture the nonlinear interaction between long-term slowly changing factors and short-term abrupt factors, solving the problem of risk path breakage caused by scale mismatch in existing technologies. It significantly improves the risk identification capability and early warning accuracy in complex scenarios, and helps prevent serious failures such as thermal runaway caused by the accumulation of weak anomalies at multiple stages.

[0021] In one embodiment, in step S1, it should be noted that during the process of acquiring and uniformly processing multi-source data from all stages of the battery's life cycle, raw battery data is first collected from five stages: manufacturing, transportation, installation, operation, and maintenance. This raw data includes, but is not limited to, initial voltage, cell internal resistance, and consistency level during manufacturing; ambient temperature, vibration intensity, and impact records during transportation; initial SOC and battery module number during installation; real-time temperature, voltage, current, SOC, and SOH curves during operation; and maintenance logs, alarm records, and component replacement information during maintenance. This ensures comprehensive coverage of the battery's state evolution information from factory to retirement. After collection, this data is grouped according to life cycle stages, and a field standardization mapping operation is performed. Differences in field naming, units, and formats from different data sources are uniformly converted into a preset data structure field format. For example, the cell_IR_avg field in the manufacturing stage is mapped to Resistance_Avg, the shock_value field in the transportation record is mapped to Transport_Shock_Intensity, and the temp_max field in the operation stage is mapped to Max_Temperature_Celsi. This ensures that all data can be analyzed and processed under a unified structure. After field standardization, the unique identifier of each battery is used as the primary key for indexing. Standardized data from each stage are vertically integrated to construct a lifecycle dataset containing multiple variables and spanning multiple time points. Each variable has corresponding value records at different time points. For example, the dataset for battery pack numbered P001 includes initial internal resistance of 3.1mΩ, transportation impact strength of 7g, installation SOC of 65%, daily SOC, voltage, and temperature sequence values ​​during the operation phase, and records the cooling maintenance event on day 50 during the maintenance phase. Due to the sampling of different variables... Due to frequency differences, such as SOH being updated daily while current is recorded every minute, time synchronization processing is required for all variables. For slowly changing variables such as SOH and internal resistance, forward padding is used to complete their values ​​at unsampled times. For rapidly changing variables such as current and temperature, moving average interpolation is used to smooth their abrupt missing values, ultimately forming a complete and usable dataset on a unified time axis. All processed data will be organized into a data structure mapping relationship with the battery number as the key. For example, P001 is mapped to its standardized and time-aligned complete data volume, serving as a unified input source for subsequent variable trend extraction, cross-scale chain identification, and risk prediction models.

[0022] In one embodiment, it should be noted that in step S2, based on the data structure indexed by the battery identifier constructed in step S1 as input, the original time series of each variable on a unified time axis is first extracted, and the time series is normalized to eliminate differences in numerical units and magnitudes among different variables. For example, the temperature variable may range from 20 to 80 degrees Celsius, the current variable may be between -100 and +100 amps, and the internal resistance variable may be only a few milliohms. To avoid these different units interfering with subsequent comparisons, a normalization method (such as minimum-maximum scaling) is used to normalize all variables. The range of variable values ​​is uniformly mapped to 0 to 1. Assuming the battery number is P001 and the variable is SOC, its original time series over 60 days is [62.1, 63.3, 65.0, 92.4], which, after normalization, becomes [0.01, 0.03, 0.08, 1.00]. After normalization, the average rate of change for each variable sequence within that period is calculated to determine the degree of activity of the variable in actual operation. This average rate of change is called the variable's evolution rate index, and it is calculated by averaging the absolute values ​​of the differences between the variable's values ​​at each adjacent time point. The specific calculation formula is V(j) = (1 / (T-1)) × ∑|D(i,j,k+1) - D(i,j,k)|, where V(j) is the average rate of change of the j-th variable, T is the number of time points for that variable, and D(i,j,k) represents the normalized value of the j-th variable at the k-th time point in battery i. For example, in battery P001, the SOC variable has values ​​of 0.42 and 0.45 at k=3 and k=4, respectively, and the difference between them is 0.03, which is included in the cumulative term of the rate of change. After the calculation is completed, the V(j) results for each variable are classified according to... Two preset rate-of-change thresholds divide all variables into three categories: if V(j) is less than the first threshold (e.g., 0.01), the variable is a slow-changing variable, indicating that its value evolves slowly over time; if V(j) is between the first and second thresholds (e.g., 0.01~0.05), it is a medium-changing variable, indicating moderate change; if V(j) is greater than the second threshold (e.g., exceeding 0.05), it is a fast-changing variable, indicating frequent fluctuations or drastic changes in value. Taking P001 as an example, if its SOH variable V(j) is 0.002, its temperature variable V(j) is 0.037, and its charging current V(j) is 0.12 variables are then classified into slow-changing, medium-changing, and fast-changing variables. After classification, the trend direction information on a unified time axis is further extracted for each level of variable. The trend extraction method is to determine whether the variable is rising, falling, or stationary between adjacent time points, that is, to construct a trend direction symbol sequence S(j,k), which is defined as follows: if D(i,j,k+1)-D(i,j,k)>ε, then S(j,k)=+1 indicates an upward trend; if it is less than -ε, then S(j,k)=-1 indicates a downward trend; if the absolute value of the difference is less than or equal to ε, then S(j,k)=0 indicates stationary with no significant change. Here, ε is a preset threshold for the minimum significant change, used to exclude misjudgments caused by minor fluctuations. For example, if the SOC variable changes from 0.42 to 0.421 between two days, and the threshold is set to 0.005, the change is judged as 0, indicating a stable trend. Ultimately, each variable has a complete sequence of trend direction signs. For instance, the trend direction of a temperature variable over 10 days is [+1,+1,0,-1,-1,0,+1,+1,0], representing periods of warming, cooling, and stabilization. This trend information will serve as the basic data input for subsequently constructing trend disturbance pairs and cross-scale linkage paths.

[0023] In one embodiment, in step S3, it should be noted that the process of identifying trend disturbance point pairs across time scales based on the trend symbol sequences of variables at each level obtained in step S2 is implemented as follows: First, select any slow-changing variable and any fast-changing variable. Assume the slow-changing variable is numbered j1 and the fast-changing variable is numbered j2. Extract each position point where a trend change occurs from their trend direction symbol sequences, that is, find all k1 in S(j1,k) such that S(j1,k1)≠0, and at the same time find all k2 in S(j2,k) such that S(j2,k2)≠0. Then, compare the trend directions corresponding to k1 and k2 one by one to see if they are the same, that is, determine whether the trend direction is satisfied. The condition S(j1,k1) = S(j2,k2) is satisfied, and further, it is determined whether k2 occurs after k1 and whether the time interval between the two exceeds the preset minimum time misalignment window Δmin. This is to exclude synchronous changes that occur at the same moment, ensuring that we only retain point pairs with sequential evolution logic. Then, it is determined whether the time difference between k2 and k1 is within the preset maximum time misalignment window Δmax. That is, we only retain trend linkages with moderate time misalignment and exclude accidental changes with excessive time intervals. If all three conditions are met at the same time, namely, consistent trend direction, sequential occurrence, and time difference within the allowable range, then this is considered a valid trend disturbance coupling point, and the point pair is set. The data is recorded as a quintuple (j1,j2,k1,k2,s), where j1 represents the number of the slowly changing variable, j2 represents the number of the rapidly changing variable, k1 represents the time point of the trend abrupt change of the slowly changing variable, k2 represents the time point of the trend abrupt change of the rapidly changing variable, and s represents the trend direction sign (i.e., +1 or -1), used to distinguish whether the trend is rising or falling synchronously. For example, in a set of battery data, the slowly changing variable is SOH (number j1=3), which shows a downward trend at time point k1=20, i.e., S(3,20)=-1, while the rapidly changing variable is the charging current (number j2=8), which also shows a downward trend at time point k2=24, i.e., S(8,24)=-1, and the two trends are consistent, with k2 being larger than s. Since k1 and k2−k1=4 fall between the preset Δmin=2 and Δmax=10, this pair of points is identified as a valid trend perturbation coupling point, indicating that the decrease in battery SOH is accompanied by a decrease in charging current 4 days later, and there may be a causal relationship. This quintuple (3,8,20,24,-1) will be added to the trend perturbation coupling point set S. In this way, all combinations of slow and fast variables are traversed and matched, and finally all point pairs that meet the conditions are combined into a trend perturbation coupling point set for use in subsequent steps to construct cross-scale evolution chains. The set retains variable hierarchical source markers and time misalignment interval information to ensure that causal chain logic analysis and risk inference can be performed during subsequent path construction.

[0024] In one embodiment, in S4, it should be noted that, based on the set of trend disturbance coupling points obtained in step S3, in order to further construct a complete cross-scale risk evolution path, first, all point pairs that satisfy the same trend direction and have a chronological order need to be extracted from the set of trend disturbance coupling points. Let each point pair be represented as (j1, j2, k1, k2, s), where j1 is the slow-varying variable number, j2 is the fast-varying variable number, k1 is the time point when the slow-varying variable undergoes a trend mutation, k2 is the time point when the fast-varying variable undergoes a trend mutation, and s is the symbol of the same trend direction for both, +1 represents an increase, -1 represents a decrease, and 0 is not considered; then, taking each point pair (j1, j2, k1, k2, s) as input, search in the current battery data or historical sample data to find whether there is a medium-varying variable, whose number is denoted as c1. It is required that there is a time point k3 in the trend sequence S(c1, k) of this medium-varying variable, satisfying S(c1, k3) = s, and k1 < k3 < k2, that is, the medium-varying variable also undergoes a trend mutation in the same direction between the two trend disturbance points; if this condition holds, it indicates that the change of this medium-varying variable may be the intermediate link before the change of the slow-varying variable is transmitted to the fast-varying variable, and there is a possibility of a causal chain in physical or data evolution. Therefore, j1, c1, and j2 are respectively used as the slow-varying, medium-varying, and fast-varying variable numbers of this chain, k1, k3, and k2 are respectively used as the trend mutation time points of the corresponding variables, and s is used as the unified trend direction of the entire chain to construct a three-segment evolution chain P, in the form of: P = (j1, k1) → (c1, k3) → (j2, k2), which means that after a slow-varying variable of the battery undergoes an increasing or decreasing trend at k1, the medium-varying variable also undergoes a同向mutation at k3, and finally the fast-varying variable undergoes a drastic trend change in a responsive manner at k2. Such chains reflect a possible cross-time-scale risk evolution path inside the battery; for example, in the data of a certain battery, the slow-varying variable is SOH (j1 = 2) and shows a decreasing trend S(2, 15) = -1 on the 15th day k1, the fast-varying variable is the charging current (j2 = 8) and shows a decreasing trend S(8, 21) = -1 on the 21st day k2, and the medium-varying variable, the temperature change rate (c1 = 5), also shows a decreasing trend S(5, 18) = -1 on the 18th day k3. Then it is considered that the three of them can form an effective cross-scale evolution chain P = (2, 15) → (5, 18) → (8, 21), indicating that the decrease of SOH causes the decrease of temperature rise, and finally leads to a sharp decrease in the charging current, which may reflect the complete risk path that the degradation of the battery cell causes temperature response changes and finally affects the BMS regulation behavior; add all chains that meet the above conditions to the chain set, and each chain retains the variable number, trend mutation time point, and trend direction symbol for subsequent chain activation judgment and risk index calculation, ensuring that subsequent analysis can perform risk evolution modeling and linkage deduction based on the real time series data chain.

[0025] In one embodiment, it should be noted that, based on the cross-scale evolutionary chain set constructed in step S4, each chain P x It consists of variables at three different evolutionary rate levels: a slow-changing variable, a medium-changing variable, and a fast-changing variable. The chain structure is represented by P. x =(vs,ts)→(vm,tm)→(vf,tf), where vs represents the number of the slow-changing variable in the chain, ts is the time point when the slow-changing variable shows a significant trend change in its historical trend, vm represents the number of the medium-changing variable, tm is the time point when the medium-changing variable changes abruptly, vf represents the number of the fast-changing variable, and tf is the time point when its trend changes abruptly. The entire chain reflects the consistent trend direction change process between variables of different scales in past data. To determine whether the chain is activated at the current moment, it is necessary to combine the real-time trend direction information of the three variables under the current time tnow. The specific method is: first, extract the trend direction symbols of the three types of variables at the current time tnow from their trend symbol sequences, and denote them as S(vs,tnow), S(vf,tf), etc. The trend direction symbols (m,tnow) and S(vf,tnow) are obtained from the trend sequence S(j,k) generated in step S2. They represent the trend direction of a variable at a certain moment, with a value of +1 indicating an upward trend, -1 indicating a downward trend, and 0 indicating a stable trend. Then, these three extracted trend direction symbols are compared to determine if they are consistent with the original trend direction sx of the chain. sx represents the trend direction exhibited by the three variables at the three historical abrupt change points ts, tm, and tf when the chain was constructed. If at the current moment, all three variables satisfy S(vs,tnow) = S(vm,tnow) = S(vf,tnow) = sx, meaning the current trend direction is completely consistent with the original trend direction and none are 0, then the chain P can be considered complete. x The fact that it has been activated at the current moment indicates that the current state exhibits a trend combination highly similar to the historical evolution path, possessing the reproducibility of potential risk evolution. For example, if a chain consists of a slowly changing variable SOH (variable number vs=2), a moderately changing variable temperature gradient (vm=5), and a rapidly changing variable discharge current (vf=9), and in historical samples, these three variables experience abrupt changes in the downward trend (sx=-1) at days 20, 23, and 25 respectively—that is, SOH decreases on day 20, temperature decreases on day 23, and current drops sharply on day 25—then this chain is recorded as P. x=(2,20)→(5,23)→(9,25), sx=-1. If on the 70th day of the current operation (tnow=70), the trend signs of SOH S(2,70)=-1, temperature gradient S(5,70)=-1, and discharge current S(9,70)=-1 are all detected, and all three are consistent with the historical trend direction sx, then it indicates that this chain has been triggered again, the link is activated, and it shows that the potential structure that caused the current drop in the past is recurring. This chain will be marked as activated. It should be noted that, at the current detection time, in order to determine whether an activated chain has significant misalignment trend linkage characteristics, three types of variables are first extracted from the chain: slow-changing variables, medium-changing variables, and fast-changing variables, which are numbered as vs, vm, and vf, respectively. Then, the trend direction sequence of these three variables is extracted for L consecutive time points forward from the current detection time τ. The trend direction value is +1 to indicate that the variable is showing an upward trend at that time, -1 to indicate a downward trend, and 0 to indicate no obvious trend change. For example, the SOH variable may have decreased continuously over the past 3 days, and its trend direction sequence is [-1, -1, -1], while the current variable may be [0, -1, +1], indicating unstable fluctuations. The trend directions of these three variables are denoted as Qa(p), Qb(p), and Qc(p), respectively, where p is the time step index, ranging from 0 to L-1. Qa(p) represents the trend direction of variable vs at time τ−p, and Qb(p) and Qc(p) represent the trend directions of variables vm and vf at time τ−p, respectively. Next, a trend lag difference sequence is constructed to characterize the time misalignment relationship between variables. First, the trend difference between the slow-changing variable and the medium-changing variable is calculated, that is, for each p, Qa(p+1) minus Qb(p) is calculated to obtain D1(p). Then, the trend difference between the medium-changing variable and the fast-changing variable is calculated, that is, Qb(p+1) minus Qc(p) is calculated to obtain D2(p), where p ranges from 0 to L−2. For example, if the trend sequence of the slow-changing variable is [-1,-1,0] and the medium-changing variable is [0,-1,-1], then at p=0, D1(0)=Qa(1)-Qb(0)=-1-0=-1, D2(0)=Qb(1)-Qc(0)=-1-Qc(0)=+1=-2.Next, a trend consistency score is constructed, and a value is assigned by judging whether the above difference indicates that the trend direction is consistent: if the trend difference is 0, it means that the trend is completely consistent, and the score for this item is 1; if the difference is ±1, it means that there is a slight difference or intermediate state, and the score for this item is 0.5; if the difference is ±2, it means that the trend is completely opposite, and the score is 0; in this way, two trend consistency scores are obtained at each time p, representing the direction consistency of slow-medium and medium-fast respectively. For example, D1(0)=-1 ⇒ score is 0.5, D2(0)=-2 ⇒ score is 0. Then, the two scores are multiplied to obtain the joint consistency score at that time. That is, if both segments are completely consistent, it is 1×1=1, and if one segment is consistent and the other segment is opposite, it is 1×0=0, forming a joint consistency sequence of length L-2. Next, bending detection is performed on the joint sequence, that is, the absolute value of the change in consistency score between adjacent time points is calculated item by item, and these differences are summed to form the total bending amount. For example, if the joint consistency sequence is [1, [0.5,1], then the change is |0.5-1|=0.5, |1-0.5|=0.5, the total bend is 1.0; then divide the bend value by the theoretical maximum bend number L-2 to obtain the normalized stability score. If the sequence is stable and consistent, the value approaches 1, and if it fluctuates violently, it approaches 0; further, in order to evaluate whether the trend consistency is concentrated in the whole time window, add 1 to all the joint consistency scores and multiply them together. For example, [1,0.5,1] becomes [2,1.5,2], the multiplication is 6, then take the natural logarithm and normalize it into a continuous kernel value between 0 and 1, representing the degree of trend consistency concentration. For example, if the window length is 3, divide by ln(2)×2. Finally, the trend one The stability score of the structure is multiplied by the trend continuity kernel value to obtain the final trend lag synchronization index. The index value ranges from 0 to 1 and is used to reflect whether the activation chain has a clear lag structure, strong trend consistency, and high linkage concentration at the current moment. The closer the value is to 1, the stronger the actual linkage of the chain and the more stable and reliable the trend transmission. For example, if the variables in a chain have continuously shown progressive trend linkage in the past 5 days, the trend direction is stable, and the time misalignment between the three variables is fixed, with little fluctuation and short intervals between trend changes, the final index value may reach above 0.85, reflecting that the chain is in a highly reliable activation transmission state at the current moment, which is suitable as the main criterion for inclusion in the linkage risk index.

[0026] The Trend Lag Synchronization Index is an indicator used to measure whether there is a structural misalignment and linkage between slow-changing, medium-changing, and fast-changing variables in an activation chain. Specifically, it characterizes whether these three variables undergo asynchronous but orderly trend abrupt changes over a period of time; that is, the slow-changing variable changes first, the medium-changing variable responds subsequently, and finally the fast-changing variable jumps. This transmission relationship is not only consistent in trend direction but also exhibits a clear sequential lag in time structure. Its essential function is to identify risk chains from the trend evolution sequence of multiple variables where trends do not occur simultaneously but are structurally highly coupled. These chains often represent the true manifestation of failure, anomalies, or deterioration processes. The reason why a higher Trend Lag Synchronization Index is more meaningful is that it comprehensively measures the consistency of the three types of variables in trend direction (i.e., whether they all develop in the same direction, such as all decreasing or all increasing), the orderly misalignment in time (e.g., SOH decreases first, then temperature gradually increases, and finally current suddenly drops), and the overall trend... Whether the trend process is consistently stable (rather than just coincidentally occurring on a particular day); for example, in a certain chain, SOH began to slowly decrease 5 days ago, the temperature continued to rise after the 3rd day, and the discharge current has been fluctuating and increasing in the last 2 days. Although the trend directions do not occur synchronously, the three variables constitute a consistent response sequence of slow change → medium change → fast change. If this structure persists near the current moment, is consistent in direction, and fluctuates stably, the index value will be close to 1, indicating that this chain is not a random resonance, but the result of the real activation of the trend transmission chain. This indicates that the current state is repeating the risk path in history and is a strongly activated chain worthy of close attention. Conversely, if the three types of variables are currently consistent in trend direction, but the trend changes occur at the same time point, or the fluctuations are violent and the order is disordered, or only one variable has a strong trend while the other variables follow weakly, the trend lag synchronization index will be significantly reduced. This indicates that although this chain appears to be activated, it lacks a transmission mechanism behind it and belongs to a loose structure or occasional resonance, with low credibility of the warning.

[0027] It should be noted that at the current detection moment, three core variables are extracted from an activated chain: a slowly changing variable, a moderately changing variable, and a rapidly changing variable, denoted as X, Y, and Z. Then, a trend direction sequence is constructed for each of these three variables. This sequence represents the trend direction of the variable each day for a preset window period (e.g., 10 days) from the current moment. A trend direction of +1 indicates an upward trend, -1 indicates a downward trend, and 0 indicates no significant trend change. These trend directions are extracted using the temporal difference sign of the variable values. For example, if the SOH variable gradually decreases over several consecutive days, its trend direction sequence might be [-1, -1, -1, -1, 0, ...]. [0, -1, -1, -1, -1]; After obtaining the trend direction sequence, the next step is to identify continuous trend segments in the trend sequence of each variable, that is, to extract the time intervals in which the same direction (+1 or -1) is maintained continuously. For example, in the trend sequence of SOH above, there is a time period of [-1, -1, -1, -1], which corresponds to day 1 to day 4, indicating that SOH is continuously in a downward state during this period. Such trend segments are represented as a set of continuous trend intervals for each variable, each including the trend direction, start time, and end time. After completing the extraction of trend segments for the three variables, the next step is to perform pairwise analysis on the trend segments between any two variables to determine whether there is a continuous trend interval for a certain variable. Under the premise that the trend direction is the same, the trend interval of one variable completely covers the continuous trend interval of the other variable. For example, if variable X has a +1 trend interval from day 3 to day 9, and variable Y has a +1 trend interval from day 4 to day 6, then the trend interval of variable X completely covers the trend interval of variable Y, and the trend direction of both is the same. Therefore, variable X is considered to have exerted downward pressure on variable Y in the trend direction. Each such covering relationship is recorded as a valid trend pressure record. Subsequently, the pressure is calculated for each pair of variables (i.e., X to Y, Y to Z, X to Z), defined as: the number of completely covered trend intervals of the suppressed variable divided by its total number of trend intervals. For example, if variable Y has a total number of... There are 5 trend segments, 3 of which are completely covered by the trend segment of variable X and are in the same direction. Therefore, the trend suppression force of X on Y is 3 / 5 = 0.6. After calculating the suppression force in the three directions, the direction pair with the largest suppression force value is selected as the dominant trend suppression path in this activation chain. For example, if the suppression force of X on Y is 0.6, Y on Z is 0.8, and X on Z is 0.4, then the Y on Z path is selected, with a dominant suppression force of 0.8. Then, it is further determined whether there is a structural transmission suppression path chain, that is, a variable not only exerts trend suppression on a second variable, but the second variable also exerts suppression on a third variable, forming a chain transmission relationship. For example, if the trend suppression force of X on Y is 0.6 and Y on Z is 0.If 8 is the value, then a suppression chain path X→Y→Z is considered to exist, and the length of the suppression chain is set to 2. If only one suppression system exists (e.g., X to Z), the chain length is 1. If there is no consistent suppression coverage relationship in any direction among the three, the chain length is 0. Finally, the dominant trend suppression system is multiplied by the structural chain length and divided by 2 to obtain the final directional imbalance offset index. This index is a dimensionless quantity ranging from 0 to 1. The larger the value, the more obvious the dominant trend variable is in the trend structure. That is, the trend direction of a certain variable is not only strong and continuous, but also structurally occupies the evolution space of other variables, thus forming a single trend-dominated imbalance pattern. For example, if SOH continues to decrease throughout the window, and the temperature change and current trend are exactly within the decreasing segment of SOH, it is judged that the SOH direction completely suppresses the trend expression of the other two variables. The directional imbalance offset index will be close to 1, indicating that the risk transmission in this chain is almost unidirectionally dominated by one variable, lacking a balanced structure, and belongs to a high-risk chain with a relatively serious trend structure imbalance.

[0028] The Directional Imbalance Shift Index is an index used to measure whether there is a trend-dominant structural imbalance among three variables in a chain. Specifically, when the trend of one variable (such as continuous upward or downward) covers and suppresses the trend space of other variables over time, that variable constitutes trend dominance and exerts structural control over the overall chain evolution. The Directional Imbalance Shift Index is a quantitative expression of this trend dominance phenomenon. It analyzes the trend direction sequence of the three variables near the current moment, extracts continuous trend segments, and determines whether a certain variable's trend segment completely envelops the trend segments of other variables over time while maintaining the same direction. Based on this, it calculates the trend suppression force of that variable on the other variables. If one variable simultaneously suppresses the other two variables, or even forms a chain-like trend control structure (such as X→Y→Z), it indicates that the current evolutionary path of the activated chain is severely controlled by the direction of a certain variable, lacking the necessary structure for multi-variable collaborative evolution, thus resulting in a higher Directional Imbalance Shift Index value. The larger the value, the stronger the control of a certain variable over the chain trend, the more unbalanced the overall trend structure, and the more easily it is misled by misjudgments of a single variable. For example, if the SOH in a chain continuously decreases for 10 days, and the trends of temperature and current both fall within this 10-day decrease range and are in the same direction, then the SOH exerts a suppressive effect on other variables, and the directional imbalance offset index will be close to 1. This indicates that this is a unidirectional dominant activation chain, and the warning result needs careful evaluation to determine whether it is a true linkage. Conversely, if the three variables each form trends independently in different time periods without a clear suppressive structure, then the index value will be close to 0, indicating that the chain structure is relatively balanced and the risk of directional imbalance is low.

[0029] In S6, it should be noted that at the current detection moment, the system first extracts all evolutionary chains that have been determined to be "active" and normalizes the strength index of each chain. The strength index is the sum of the trend drag synchronization index and the directional imbalance offset index of the chain, with a value range between 0 and 2. To make the strength values ​​of all chains comparable under the same standard, the system divides the strength index of each chain by a constant 2 to obtain a normalized standard strength index, whose value range is compressed to between 0 and 1. This standard strength index represents the degree of standardized impact of the chain on the overall battery risk. For example, if the trend drag synchronization of a chain is 0.8 and the directional imbalance offset index is 0.6, its total strength index is 1.4. After normalization, 1.4 divided by 2 equals 0.7, indicating that the chain's impact on the current risk structure is 0.7. Next, the system summarizes all the normalized standard strength indices and calculates their average value to obtain the linkage risk index of the battery at the current moment. For example, if there are 5 activated chains... The normalized intensity values ​​of the bars are 0.2, 0.6, 0.5, 0.3, and 0.4, respectively. The linkage risk index is the average of these five values, i.e., (0.2+0.6+0.5+0.3+0.4) / 5=0.4. Subsequently, the system compares this linkage risk index with a preset risk judgment threshold. If the linkage risk index is not less than the preset threshold, for example, the preset threshold is 0.35, it indicates that the current battery has a linkage safety risk under the synergistic effect of multiple activation chains, and the system outputs this judgment result as "Battery safety is at risk". If the linkage risk index is less than the preset threshold, for example, only 0.22, the system determines that although there is an activation chain, the overall strength is weak and the linkage structure is insufficient, and the output result is "Battery safety is not at risk". In this way, the system can aggregate the intensity information of multiple risk evolution chains into a unified risk index in a simple, controllable, and effective manner, and combine it with the threshold to make structural risk decision outputs, thereby improving the perception and response efficiency of potential battery safety hazards.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A battery safety risk prediction method based on full life cycle data, characterized in that, Includes the following steps: Acquire multi-source data throughout the entire battery lifecycle and process it uniformly according to a preset structural standard to form a data set indexed by battery identifier; The variables in the dataset are divided into three levels according to their rate of change: slow-changing variables, medium-changing variables, and fast-changing variables, and the evolution trend information of the variables at each level is extracted. Based on the trend information of variables at each level, we identify trend disturbance point pairs that have the same direction but occur at different times across time scales, forming a set of trend disturbance coupling points; Based on the set of trend perturbation coupling points, a set of cross-scale evolution chains is constructed, which are connected sequentially by slow-changing variables, medium-changing variables and fast-changing variables. The chains reflect potential risk evolution paths. Determine whether each chain in the evolution chain set is activated at the current moment. If the trend state of all nodes in the chain meets the preset activation conditions, mark it as an activated chain and record the activation intensity. Based on the activation intensity of all activated chains, the current linkage risk index of the battery is calculated, and the risk index is output as the final prediction result of battery safety risk.

2. The battery safety risk prediction method based on full life cycle data according to claim 1, characterized in that, The steps to classify variables in a dataset into three levels—slow-changing variables, medium-changing variables, and fast-changing variables—according to their rate of change are as follows: Extract the time series of each variable on a unified time axis, and then normalize the time series to obtain a set of variable sequences; For each variable sequence, calculate its average rate of change within the sampling period as an indicator of the variable's evolution rate; The variables are divided into three evolution levels based on their evolution rate index: if the evolution rate index of a variable is less than the first threshold, the corresponding evolution level of the variable is recorded as a slow-changing variable. If the evolution rate index of a variable is not less than the first threshold but less than the second threshold, then the corresponding variable evolution level is recorded as a medium-level variable. If the evolution rate index of a variable is not less than the second threshold, then the corresponding variable evolution level is recorded as a fast-changing variable; and the first threshold is less than the second threshold.

3. The battery safety risk prediction method based on full life cycle data according to claim 2, characterized in that, The steps for extracting the evolution trend information of variables at each level are as follows: For each level of variable, a trend direction symbol sequence S(j) is constructed on a unified time axis to express the short-term change direction of the variable. The trend is calculated as follows: ; In the formula, This indicates the sign of the trend direction of the j-th variable at time point k. Indicates the first The first battery The variable in the first... The values ​​at each time point; The preset minimum significant change threshold for determining the trend direction; The trend direction sign sequence of each variable is used as evolutionary trend information.

4. The battery safety risk prediction method based on full life cycle data according to claim 1, characterized in that, Based on the trend information of variables at each level, the steps to identify trend disturbance point pairs that have the same direction but occur at different times across time scales and to construct a set of trend disturbance coupling points are as follows: For each pair of slow-changing and fast-changing variables, a trend direction matching and identification mechanism is established. For any slow-changing variable and any fast-changing variable, based on their trend sign sequences, trend disturbance point pairs that satisfy the condition of consistent direction are searched. That is, it exists: ,and ; In the formula, Indicates the first A slowly changing variable at time point Trend symbols; Indicates the first A fast variable at time point Trend symbols; and This serves as the time index for the occurrence of the corresponding trend abrupt change; The preset minimum time misalignment window; when the condition is met... and And time difference Within the preset maximum misalignment window When within the range, the corresponding point pair is determined to be a valid trend disturbance coupling point, and is represented in the form of a quintuple ( , , , ,s) Add to the set of trend perturbation coupling points, where, Indicates the number of a slowly changing variable; Indicates the number of the rapidly changing variable; This indicates the point in time when a sudden change in the trend of a slowly changing variable occurs; , represents the time point when the direction of the rapidly changing variable changes uniformly; s is the trend direction symbol; all trend disturbance point pairs that meet the conditions are taken as the set of trend disturbance coupling points.

5. The battery safety risk prediction method based on full life cycle data according to claim 1, characterized in that, The steps for constructing a set of cross-scale evolutionary chains consisting of slowly changing, moderately changing, and rapidly changing variables are as follows: Based on the set of trend perturbation coupling points, extract all point pairs that satisfy the condition that there is a consistent trend direction and a time misalignment between the rapidly changing and slowly changing variables. Let each point pair be a quintuple (...). , , , ,s), where, Indicates the number of the slowly changing variable. Indicates the number of the rapidly changing variable. This indicates the time point at which a sudden change in the trend of a slowly changing variable occurs. This indicates the time point at which abrupt changes in the trend of a rapidly changing variable occur, where 's' represents the same trend direction sign for both; in the current battery data or historical sample data, for each group ( , , , ,s), in the set of intermediate variables, search for variables that satisfy the condition of consistent trend direction and time point located at and Let the intermediate variable be numbered c1, and its trend sequence be S(c1, ...). If a time point exists 3. The following conditions must be met: Then it is assumed that variable c1 in time... The trend changes at point 3 can be considered intermediate nodes in the group's evolutionary path; under the above conditions, the variables... c1 They are respectively used as variables at three levels: slow-change, medium-change, and fast-change, in chronological order. Construct a cross-scale evolutionary chain, defined as: collecting all evolutionary chains that satisfy the above structure and conditions into a cross-scale evolutionary chain set.

6. The battery safety risk prediction method based on full life cycle data according to claim 1, characterized in that, To determine whether each chain in the evolutionary chain set is activated at the current moment, if the trend states of all nodes in the chain meet the preset activation conditions, then mark it as an activated chain. The steps are as follows: Based on the cross-scale evolutionary chain set, for each chain, it contains a trend mutation triplet of a slowly changing variable, a moderately changing variable, and a rapidly changing variable. Let the chain structure be: ,in, Indicates the number of the slowly changing variable. Its trend change time point Indicates the variable number in the intermediate variable. The point in time when its trend suddenly changed; Indicates the number of the rapidly changing variable. Its trend abrupt change point; at the current moment. To determine whether the chain meets the activation condition, specifically: obtain the trend direction symbol at the current moment from the variable trend direction symbol sequence. If all three variables in the chain satisfy the following condition at the current moment: If so, the chain is determined to be in an activated state.

7. The battery safety risk prediction method based on full life cycle data according to claim 6, characterized in that, The steps for recording activation intensity are as follows: For each activated chain, calculate the trend drag synchronization index and the directional imbalance offset index for each activated chain, and add the trend drag synchronization index and the directional imbalance offset index to obtain the strength index.

8. The battery safety risk prediction method based on full life cycle data according to claim 7, characterized in that, The calculation steps for the trend lag synchronization index are as follows: at the current detection time, extract the numbers of the slow-changing variables, medium-changing variables and fast-changing variables contained in the activated chain, and extract the trend direction sequence of each variable from the current detection time forward L time points. Based on the trend direction sequence, a trend lag difference sequence is constructed. Specifically, for slow-changing and medium-changing variables, the difference between adjacent trend values ​​is calculated by subtracting the trend value of the medium-changing variable at the current moment from the trend value of the slow-changing variable at the previous moment. The trend lag difference is obtained by processing the medium-changing and fast-changing variables in the same way, that is, subtracting the trend value of the fast-changing variable at the current moment from the trend value of the medium-changing variable at the previous moment. Based on the trend drag difference, a trend consistency judgment result is constructed, specifically: if any trend difference is zero, then the trend consistency result is 1. If the absolute value of the difference is 1, the result of the item is 0.5; if the absolute value of the difference is 2, the result of the item is 0. The discrimination results are calculated separately for the slow-changing and medium-changing variables and for the medium-changing and fast-changing variables, to obtain two consistency scores for each time step. Multiply the two consistency scores at each time step to obtain the joint consistency score sequence at that time step; then, calculate the sum of the absolute values ​​of the differences between adjacent terms in the joint consistency score sequence to obtain the total bending of the trend consistency sequence, and normalize it according to the maximum number of bending steps allowed under the window to obtain the stability score of the trend drag structure. The joint consistency score at each time step is incremented by 1 and then multiplied together. The natural logarithm of the product is then taken and divided by the product of the sequence length and the natural logarithm 2 to obtain the trend dragging continuous kernel value. Multiplying the stability score by the continuous kernel value yields the final trend drag synchronization index.

9. A battery safety risk prediction method based on full life cycle data according to claim 7, characterized in that, The calculation steps for the directional imbalance offset index are as follows: At the current detection moment, extract the slow-changing, medium-changing, and fast-changing variables contained in the activation chain, and construct the trend direction sequence of the three types of variables respectively. Each trend direction takes three states: rising, falling, or no trend. Based on the trend direction sequence of each variable, identify the sub-intervals that maintain the same trend direction continuously. Represent each continuous trend segment as the interval in which the variable continuously rises or falls within a certain time period, and record the start and end time of each interval and the corresponding trend direction to form a set of continuous trend intervals of the variable. For any combination of two different variables, determine whether there exists a continuous trend interval of one variable that, under the condition that the trend direction is the same, completely covers the continuous trend interval of the other variable. If such a covering relationship exists, it is considered that the former exerts trend suppression on the latter. For each pair of variables, calculate its trend suppression level, which is defined as the ratio of the number of trend intervals that are completely covered in the suppressed variable to the total number of trend intervals. Among the three variables—slow-changing, medium-changing, and fast-changing—trend suppression mechanisms are constructed sequentially for three sets of variable pairs: slow-changing suppression mechanism against medium-changing, medium-changing suppression mechanism against fast-changing, and slow-changing suppression mechanism against fast-changing. The one with the largest suppression mechanism is taken as the dominant trend suppression strength of the current chain. Based on the trend-restraint relationship, determine whether there is a trend-dominant structure chain in the chain, that is, a certain variable not only exerts trend suppression on another variable, but the suppressed variable also continues to exert trend suppression on a third variable. If this structure is satisfied, the length of the trend-dominant structure chain is 2; if there is only a one-way suppression relationship, the chain length is 1. If there is no directional overlap relationship among the three variables, the chain length is 0; Multiply the length of the dominant structural chain by the strength of the dominant trend suppression, and divide by 2 to obtain the directional imbalance offset index.

10. The battery safety risk prediction method based on full life cycle data according to claim 1, characterized in that, The steps to calculate the current linkage risk index of the battery based on the activation intensity of all activated chains and output the risk index as the final prediction result of battery safety risk are as follows: divide the intensity index of the activated chains by a constant 2 to obtain the normalized standard intensity index. The mean of all normalized standard strength indices is used as the linkage risk index. The linkage risk index is compared with a preset threshold. If the linkage risk index is not less than the preset threshold, it indicates that there is a risk to battery safety, and the existence of battery safety risk is output as the final prediction result. If the linkage risk index is less than the preset threshold, it means that there is no risk to battery safety, and the absence of battery safety risk will be output as the final prediction result.

Images