State data identification method and system applied to energy storage battery module

Abstract

The present invention discloses a state data identification method and system for energy storage battery modules, belonging to the field of data analysis technology. The present invention obtains comprehensive power storage working state monitoring data and safety situation decision information, uses a target state matching analysis algorithm to perform data processing and state matching analysis, and obtains accurate matching confidence weights, thereby achieving more comprehensive and accurate state evaluation and safety management of energy storage battery modules, and improving the safety, use efficiency and service life of energy storage batteries.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis, and in particular relates to a state data identification method and system applied to an energy storage battery module. Background Art

[0002] In today's energy storage field, the widespread application of energy storage batteries faces many challenges. With the large-scale use of energy storage batteries in many fields such as electric vehicles and renewable energy storage systems, higher requirements are placed on the safety and status management of energy storage battery modules.

[0003] First, the energy storage battery module will be affected by many factors during operation, and its working state is very complicated. Different types of energy storage batteries (such as lithium-ion batteries, lead-acid batteries, etc.) have different performance characteristics. Even for the same type of batteries, their performance is different due to differences in production batches, use environment and other factors. Traditional monitoring methods can often only obtain some basic physical quantity data, such as voltage, current and temperature, which is difficult to fully reflect the true state of the energy storage battery module, let alone accurately evaluate its safety.

[0004] Secondly, when faced with complex working conditions and potential safety hazards, existing technologies lack effective means to integrate various monitoring information and accurately judge the safety status of energy storage battery modules.

[0005] It can be seen that the existing technology has problems with comprehensiveness and low precision in status assessment and safety management, making it difficult to ensure the safety, efficiency and service life of energy storage batteries. Summary of the invention

[0006] The present invention provides a state data identification method and system applied to an energy storage battery module, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0007] The present invention provides a state data identification method applied to an energy storage battery module, which is applied to a state data identification system. The method comprises: acquiring power storage working state monitoring data including an energy storage battery module tag and a plurality of safety situation decision information, wherein the plurality of safety situation decision information comprises: first safety situation decision information indicating a stage safety and stability situation, and second safety situation decision information indicating each of the potential safety hazard situations in at least one potential safety hazard situation; Through the target state matching analysis algorithm, the unique hot knowledge embedding is performed on each of the security situation decision information to obtain the security situation decision knowledge vector of each of the security situation decision information, and the state monitoring vector mining is performed on the power storage working state monitoring data to obtain the IoT sensor state monitoring vector; Through the target state matching analysis algorithm, a state matching analysis is performed according to the IoT sensor state monitoring vector and each of the safety situation decision knowledge vectors to obtain a state matching analysis result of the power storage working state monitoring data; wherein the state matching analysis result includes the matching confidence weight of the energy storage battery module label corresponding to each safety situation, and the safety situation includes the stage safety and stability situation and the at least one potential safety hazard situation.

[0008] The present invention provides a state data identification system, comprising at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the above method.

[0009] The present invention provides a readable storage medium, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0010] The present invention obtains comprehensive power storage working status monitoring data and safety situation decision-making information, uses the target state matching analysis algorithm to perform data processing and state matching analysis, and obtains accurate matching confidence weights, thereby achieving more comprehensive and accurate state evaluation and safety management of energy storage battery modules, thereby improving the safety, efficiency and service life of energy storage batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The present invention provides a flow chart of a method for identifying state data of an energy storage battery module.

[0012] Figure 2 This is a structural schematic diagram of a status data identification system provided by the present invention. DETAILED DESCRIPTION

[0013] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0015] Figure 1 A state data identification method applied to an energy storage battery module is shown, and applied to a state data identification system. The method includes the following S102-S106.

[0016] S102, the status data identification system obtains the power storage working status monitoring data including the energy storage battery module label and a number of safety situation decision information.

[0017] The plurality of security situation decision information include: first security situation decision information indicating a stage security stability situation, and second security situation decision information indicating each of the potential security risk situations in at least one potential security risk situation.

[0018] In S102, the state data identification system can obtain the power storage working state monitoring data, which includes the energy storage battery module label. The power storage working state monitoring data is a comprehensive data set, which may cover information obtained from multiple sensors, such as voltage sensors, current sensors, temperature sensors, etc. Taking the lithium-ion energy storage battery module as an example, its monitoring data may include the voltage value of each single cell (such as the voltage of each single cell fluctuates between 2.5-3.7V, and the specific value depends on the charge and discharge state of the battery), the charge and discharge current of the battery module (may be in the range of 0-100A, depending on the power demand and working mode of the battery), the temperature of the battery module (the normal working temperature range is 0-45°C), etc. These data are associated with the energy storage battery module label. The energy storage battery module label can be a unique identifier used to distinguish different energy storage battery modules, and may include information such as battery type (such as lithium iron phosphate, ternary lithium, etc.), production batch, rated capacity (such as 100Ah, 200Ah, etc.).

[0019] In addition, the state data identification system can also obtain the first safety situation decision information (used to indicate or characterize the stage safety and stability situation) among several safety situation decision information. The safety situation decision information can be obtained based on the empirical data of the long-term stable operation of the energy storage battery module or the established safety standards. For example, when the voltage, current, temperature and other parameters of the energy storage battery module fluctuate within the normal range, and the various protection mechanisms inside the battery module (such as overcharge protection, over-discharge protection, overheating protection, etc.) are not triggered, it can be determined to be in a stage safety and stability situation. Specifically, if the voltage fluctuation amplitude of the battery module is within ±5% (based on the rated voltage), the current fluctuation is within ±10% (based on the rated current), and the temperature fluctuation is within ±3°C (based on the optimal operating temperature), and no abnormal situation occurs within a period of time (such as 10 consecutive hours), it can be considered to be in this safe and stable situation.

[0020] Accordingly, the second safety situation decision information indicates at least one potential safety hazard situation. These potential safety hazard situations can be caused by a variety of factors. For example, when the voltage difference of the single cells in the battery module exceeds a certain value (such as 50mV), there may be a battery balance problem, which is a potential safety hazard situation; or when the temperature of the battery module rises too fast (for example, the temperature rises by more than 10°C within 10 minutes), it may indicate that there are hidden dangers such as short circuits or heat dissipation system failures inside the battery. For different potential safety hazard situations, there are corresponding second safety situation decision information to identify them.

[0021] S104. The state data identification system performs unique hot knowledge embedding on each of the security situation decision information through the target state matching analysis algorithm to obtain the security situation decision knowledge vector of each of the security situation decision information, and performs state monitoring vector mining on the power storage working state monitoring data to obtain the IoT sensor state monitoring vector.

[0022] In S104, two core operations are involved: one-hot knowledge embedding and state monitoring vector mining.

[0023] 1. Unique knowledge embedding (for security situation decision-making information) The principle of one-hot knowledge embedding under the target state matching analysis algorithm: The state data recognition system performs one-hot knowledge embedding on each safety situation decision information through the target state matching analysis algorithm. One-hot knowledge embedding is a method of converting classified information into vector representation. Taking the safety situation decision information as an example, for the first safety situation decision information (stage safety and stability situation) and the second safety situation decision information (various potential safety hazard situations), for example, there are 3 potential safety hazard situations, then there are a total of 4 different safety situation categories. In the one-hot knowledge embedding, a 4-dimensional vector space is created. For the first safety situation decision information (stage safety and stability situation), its corresponding safety situation decision knowledge vector can be [1, 0, 0, 0], indicating that it is in the first safety situation category. For the second safety situation decision information indicating a potential safety hazard situation (such as unbalanced single cell voltage), its corresponding safety situation decision knowledge vector can be [0, 1, 0, 0]. The 1 in the present invention is located in the second position of the vector, indicating that it belongs to the second safety situation category (i.e., the potential safety hazard situation of unbalanced single cell voltage). This unique knowledge embedding method converts different security situation decision information into a vector form that is easy for computers to process, so that subsequent analysis and calculation can be based on a unified mathematical model.

[0024] (II) State monitoring vector mining (for power storage working state monitoring data) The principle of IoT sensor state monitoring vector mining: For the power storage working state monitoring data, the state data recognition system performs state monitoring vector mining. This process is to integrate the discrete data obtained by various sensors into a unified vector form. Taking the lithium-ion energy storage battery module mentioned above as an example, the monitoring data includes three parameters: voltage (V), current (I), and temperature (T). These data can be normalized and constructed into IoT sensor state monitoring vectors. For example, the voltage data is normalized to obtain (V_{norm}) (the value range is between 0-1, for example, the actual voltage is 3.2V, and after mapping to the 0-1 interval, (V_{norm}=0.6)), the current data is normalized to obtain (I_{norm}) (for example, the actual current is 50A and the rated current is 100A, then (I_{norm}=0.5)), and the temperature data is normalized to obtain (T_{norm}) (for example, the actual temperature is 25°C, the optimal operating temperature range is 0-45°C, and after calculation (T_{norm}=0.55)), then the IoT sensor state monitoring vector can be expressed as ([V_{norm}, I_{norm}, T_{norm}]=[0.6, 0.5, 0.55]). This vector contains the current comprehensive working state information of the battery module, providing basic data for subsequent state matching analysis.

[0025] S106. The state data identification system performs state matching analysis based on the IoT sensor state monitoring vector and each of the safety situation decision knowledge vectors through the target state matching analysis algorithm to obtain a state matching analysis result of the power storage working state monitoring data.

[0026] The state matching analysis result includes: the matching confidence weight of the energy storage battery module tag corresponding to each safety situation, and the safety situation includes: the stage safety and stability situation and the at least one potential safety hazard situation.

[0027] Finally, S106 involves state matching analysis and corresponding matching confidence weight determination.

[0028] 1. State matching analysis principle Operation based on the target state matching analysis algorithm: The state data recognition system performs state matching analysis based on the IoT sensor state monitoring vector and each security situation decision knowledge vector through the target state matching analysis algorithm. The present invention can adopt an algorithm model based on distance measurement, such as the Euclidean distance algorithm. For the IoT sensor state monitoring vector ([V_{norm}, I_{norm}, T_{norm}]) and each security situation decision knowledge vector, calculate the Euclidean distance between them. For example, there are (k) security situation decision knowledge vectors, namely (vec{v}_1, vec{v}_2, cdots, vec{v}_k), and the IoT sensor state monitoring vector is (vec{u}=[u_1, u_2, u_3]) (the present invention takes a 3-dimensional vector as an example). Then the Euclidean distance (d_i=sqrt{sum_{j=1}^{3}(u_j-v_{ij})^2}), where (i=1, 2, cdots, k).

[0029] By calculating these distances, we can get the degree of closeness between the IoT sensor state monitoring vector and each safety situation decision knowledge vector. The smaller the distance, the closer the current power storage working state monitoring data is to the safety situation.

[0030] 2. Determination of matching confidence weight Example of matching confidence weight calculation: Based on the above distance calculation results, the matching confidence weights of the energy storage battery module labels corresponding to each safety situation can be determined. For example, the Euclidean distance (d_1=0.2) between the IoT sensor state monitoring vector and the first safety situation decision knowledge vector (stage safety and stability situation) and the Euclidean distance (d_2=0.8) between the IoT sensor state monitoring vector and a second safety situation decision knowledge vector (such as the single cell battery voltage imbalance situation) are calculated. An exemplary weight calculation method can be used, such as: (w_i=frac{1 / d_i}{sum_{j=1}^{k}1 / d_j}).

[0031] Then the matching confidence weight corresponding to the first safety situation decision knowledge vector is: (w_1=frac{1 / 0.2}{1 / 0.2+1 / 0.8}=frac{5}{5+1.25}=frac{5}{6.25}=0.8), and the matching confidence weight corresponding to the second safety situation decision knowledge vector is: (w_2=frac{1 / 0.8}{1 / 0.2+1 / 0.8}=frac{1.25}{5+1.25}=frac{1.25}{6.25}=0.2). These matching confidence weights intuitively reflect the possibility of the current energy storage battery module being in different safety situations, thereby providing an important basis for the safety management and status assessment of the energy storage battery module. If the matching confidence weight indicates that the possibility of a potential safety hazard situation is high, corresponding measures can be taken in a timely manner, such as equalizing the battery module, checking the cooling system, etc., to ensure the safe and stable operation of the energy storage battery module.

[0032] In summary, the present invention can comprehensively grasp the working state and safety situation of the energy storage battery module by acquiring the power storage working state monitoring data including the energy storage battery module label and the safety situation decision information covering the stage safety and stability situation and the potential safety hazard situation. For example, in a large-scale energy storage system, different energy storage battery modules may have different characteristics (distinguished by module labels), and comprehensive monitoring data helps to accurately locate the status of each module.

[0033] The target state matching analysis algorithm is used to embed the security situation decision information into unique knowledge to obtain the security situation decision knowledge vector, and the state monitoring vector of the power storage working state monitoring data is mined to obtain the IoT sensor state monitoring vector. This process realizes the effective integration and transformation of data. Taking unique knowledge embedding as an example, converting different categories of security situation decision information into vector form can enable computer systems to process this information more efficiently, just like organizing various messy security situation information into a unified format that is easy for computers to recognize.

[0034] The matching confidence weight is obtained by performing state matching analysis based on the IoT sensor state monitoring vector and each safety situation decision knowledge vector. This result is of great significance. It can quantify the possibility of energy storage battery modules being in different safety situations. For example, in the energy storage battery management of electric vehicles, if it is accurately known that the matching confidence weight of the battery module in a potential safety hazard situation (such as battery overheating hazard) is high, timely measures can be taken, such as adjusting the charging strategy or strengthening heat dissipation, thereby improving the safety of the energy storage battery. At the same time, this quantitative result helps to conduct a more accurate state assessment of the energy storage battery module. In the entire life cycle management of the energy storage battery, whether in the initial performance evaluation or the later aging monitoring, it can provide more reliable data support, thereby optimizing the use efficiency of the energy storage battery and extending its service life.

[0035] In actual application, the debugging process of the above-mentioned target state matching analysis algorithm is the key to ensuring the technical stability of S102-S106. Based on this, the present invention also innovatively proposes debugging steps for the target state matching analysis algorithm, that is, the debugging of the original state matching analysis algorithm is realized through the following S202-S208, thereby obtaining the target state matching analysis algorithm applied to S102-S106.

[0036] S202, obtaining preceding working state monitoring data (such as sample working state monitoring data) for debugging the original state matching analysis algorithm, first monitoring judgment information of the preceding working state monitoring data, and second monitoring judgment information of the preceding working state monitoring data under the target monitoring task.

[0037] S204. Based on the first monitoring judgment information and the second monitoring judgment information, the state monitoring vector of the previous working state monitoring data is subjected to task attention processing to obtain a state monitoring attention vector, and based on the original task attention weight, the state monitoring vector is subjected to task attention processing to obtain a state attention prediction vector.

[0038] S206. Optimize the original task attention weight according to the difference between the state attention prediction vector and the state monitoring attention vector to obtain the target task attention weight.

[0039] S208. When debugging the original state matching analysis algorithm, the state monitoring vector is subjected to task attention processing according to the target task attention weight to obtain the target state monitoring attention vector of the previous working state monitoring data, and the original state matching analysis algorithm is debugged according to the target state monitoring attention vector to obtain the target state matching analysis algorithm.

[0040] First, S202 involves the acquisition of the preceding working state monitoring data: in the debugging process of the target state matching analysis algorithm, the preceding working state monitoring data used to debug the original state matching analysis algorithm must first be obtained. This data can be sample working state monitoring data, which is similar to the power storage working state monitoring data in the actual application scenario, but is specifically used for algorithm debugging purposes. The preceding working state monitoring data contains various parameter information related to the energy storage battery module. Taking lithium-ion energy storage batteries as an example, these data may include voltage data of multiple single cells, for example, the voltage range of each single cell is between 2.5-3.7V, and the collection of these voltage values ​​constitutes part of the preceding working state monitoring data. In addition, it may also include the charge and discharge current data of the battery module, for example, the current range is between 0-100A, and the temperature data of the battery module, and the normal operating temperature range may be 0-45°C. The accuracy and completeness of these data are crucial for subsequent algorithm debugging.

[0041] Based on the above content, the first monitoring judgment information of the previous working state monitoring data and the second monitoring judgment information under the target monitoring task can be obtained. The first monitoring judgment information may be obtained based on the known battery state judgment standard. For example, according to the technical specifications provided by the battery manufacturer, when the voltage of the single cell is lower than 2.8V, it is judged to be in a low-power state. This judgment information can be used as part of the first monitoring judgment information. The second monitoring judgment information is for the target monitoring task. For example, if the target monitoring task is to detect the state stability of the battery module during high-power charging and discharging, then the second monitoring judgment information may be based on a large amount of experimental data. For example, when the battery is working at a charge and discharge rate exceeding 0.5C (C is the current corresponding to the rated capacity of the battery), the battery temperature rise rate does not exceed 1°C / minute and is judged to be stable. This information is the second monitoring judgment information for the target monitoring task.

[0042] Furthermore, S204 may include the following two technical ideas.

[0043] (i) Task attention processing to obtain a state monitoring attention vector: According to the first monitoring discrimination information and the second monitoring discrimination information, the state monitoring vector of the previous working state monitoring data is subjected to task attention processing to obtain a state monitoring attention vector. The present invention uses an example numerical example to illustrate the task attention processing process. For example, the state monitoring vector of the previous working state monitoring data contains three elements, namely, the voltage value (V=3.2) (normalized to the 0-1 interval, for example, (0.6) after normalization), the current value (I=50) (for example, the rated current is 100A, and the normalized value is (0.5)), and the temperature value (T=25) (for example, the optimal operating temperature range is 0-45°C, and the normalized value is (0.55)), that is, the state monitoring vector (vec{x}=[0.6, 0.5, 0.55]). : The first monitoring discrimination information and the second monitoring discrimination information can be converted into a weight vector. For example, according to these discriminant information, the weights for voltage, current, and temperature are (w_1=0.3), (w_2=0.4), and (w_3=0.3) respectively (these weights are obtained based on the judgment of the importance of different parameters by the discriminant information). Then the state monitoring attention vector (vec{y}) is obtained by the following calculation: (vec{y}=[0.6*0.3, 0.5*0.4, 0.55*0.3]=[0.18, 0.2, 0.165]).

[0044] (ii) Obtaining the state attention prediction vector based on the original task attention weights: Based on the original task attention weights, the state monitoring vector is processed by task attention to obtain the state attention prediction vector. For example, the original task attention weights are (vec{w_0}=[0.25, 0.35, 0.4]) (these weights are originally set and have not been optimized). Similarly, for the state monitoring vector (vec{x}=[0.6, 0.5, 0.55]), the state attention prediction vector (vec{z}) is calculated as follows: (vec{z}=[0.6*0.25, 0.5*0.35, 0.55*0.4]=[0.15, 0.175, 0.22]).

[0045] Next, as for S206, the focus is on optimizing the original task attention weight to obtain the target task attention weight based on the difference between the state attention prediction vector and the state monitoring attention vector. There are many methods for calculating the difference between the two vectors, such as Euclidean distance or mean square error (MSE). The present invention takes the mean square error as an example, the mean square error (MSE=frac{1}{n}sum_{i=1}^{n}(y_i-z_i)^2), where (n=3) (vector dimension), (y_i) and (z_i) are elements of the state monitoring attention vector and the state attention prediction vector, respectively. For example, after calculation, (MSE=0.01) is obtained (this is just an example value). In order to optimize the original task attention weight, an example variant of the gradient descent algorithm is used. For example, the original task attention weight (vec{w_0}=[w_{01}, w_{02}, w_{03}]), the weight is updated according to the partial derivative of the mean square error with respect to the weight. For example, the update formula for (w_{01}) is: (w_{01}^{new}=w_{01}-alpha*frac{partialMSE}{partialw_{01}}); Where (alpha) is the learning rate, for example (alpha=0.01). After multiple iterations (for example, 10 iterations), the optimized target task attention weight (vec{w_1}) is finally obtained. For example, after optimization (vec{w_1}=[0.28, 0.32, 0.4]).

[0046] Finally, S208 involves the acquisition of the target state monitoring attention vector and the core steps of the target state matching analysis algorithm.

[0047] When debugging the original state matching analysis algorithm, according to the target task attention weight, the state monitoring vector is processed by task attention to obtain the target state monitoring attention vector of the previous working state monitoring data. Taking the previous state monitoring vector (vec{x}=[0.6, 0.5, 0.55]) and the target task attention weight (vec{w_1}=[0.28, 0.32, 0.4]) as an example, the target state monitoring attention vector (vec{y_1}) is calculated as: (vec{y_1}=[0.6*0.28, 0.5*0.32, 0.55*0.4]=[0.168, 0.16, 0.22]).

[0048] Further, according to the target state monitoring attention vector, the original state matching analysis algorithm is debugged to obtain the target state matching analysis algorithm. This process may involve adjusting some parameters in the algorithm or changing the internal structure of the algorithm. For example, the original state matching analysis algorithm may be based on a certain distance metric (such as Mahalanobis distance) for state matching analysis. During the debugging process, according to the information in the target state monitoring attention vector, the covariance matrix in the Mahalanobis distance calculation may be adjusted. For example, the original covariance matrix is ​​(Sigma_0), and the covariance matrix is ​​weighted and adjusted according to the elements in the target state monitoring attention vector. For example, (Sigma_1=beta*vec{y_1}circSigma_0), where (beta) is an adjustment coefficient (for example (beta=0.9)), and (circ) represents the corresponding multiplication of elements. Through such adjustments, the original state matching analysis algorithm is optimized to the target state matching analysis algorithm, making it more suitable for actual working state monitoring and analysis tasks.

[0049] Through the above debugging steps of S202-S208, the accuracy and stability of the target state matching analysis algorithm can be effectively improved. On the basis of obtaining the previous working state monitoring data and related discrimination information, the algorithm can process the data more accurately through task attention processing and optimization of the original task attention weights. For example, in the process of optimizing the original task attention weights, through the application of mean square error and gradient descent algorithm variants, the weights can be adjusted according to the actual data situation, so that the target task attention weights obtained are more in line with the data characteristics. The target state matching analysis algorithm finally obtained can improve the accuracy of the state judgment of the energy storage battery module when applied to the state matching analysis in S102-S106, so as to better ensure the safe operation of the energy storage battery, optimize its use efficiency and extend its service life. At the same time, this debugging process helps to improve the robustness of the entire technical solution, so that it can adapt to different energy storage battery types and working environments.

[0050] In the following design idea, the state monitoring vector of the preceding working state monitoring data is subjected to task attention processing based on the first monitoring judgment information and the second monitoring judgment information to obtain a state monitoring attention vector, as described in S204, including: performing one-hot knowledge embedding on the first monitoring judgment information to obtain a first safety situation decision knowledge vector, and performing one-hot knowledge embedding on the second monitoring judgment information to obtain a second safety situation decision knowledge vector; performing feature pooling processing on the state monitoring vector of the preceding working state monitoring data to obtain a state monitoring pooling vector; utilizing the second safety situation decision knowledge vector and the first safety situation decision knowledge vector to obtain a safety situation decision offset, and obtaining the state monitoring attention vector based on the safety situation decision offset and the state monitoring pooling vector.

[0051] Furthermore, the method described in S204 performs task attention processing on the state monitoring vector based on the original task attention weight to obtain a state attention prediction vector, including: obtaining a first weighted monitoring feature based on the original task attention weight and the state monitoring vector; and performing feature pooling processing on the first weighted monitoring feature to obtain the state attention prediction vector.

[0052] In the present invention, in the process of performing task attention processing on the state monitoring vector of the previous working state monitoring data to obtain the state monitoring attention vector, the first monitoring judgment information must first be subjected to one-hot knowledge embedding to obtain the first safety situation decision knowledge vector. One-hot knowledge embedding is an effective way to convert classified information into vector representation. For example, the first monitoring judgment information contains three different judgment states, such as low power state, normal power state, and high power state. When performing one-hot knowledge embedding, a three-dimensional vector space is created. If the current first monitoring judgment information indicates that the battery is in a normal power state, then the corresponding first safety situation decision knowledge vector may be [0, 1, 0], where 1 is located in the second position of the vector, indicating the category of normal power state.

[0053] Similarly, the second monitoring judgment information is embedded with one-hot knowledge to obtain the second safety situation decision knowledge vector. For example, the second monitoring judgment information is about the stability judgment of the battery under different charge and discharge rates, and there are three states: stable, relatively stable, and unstable. If the current second monitoring judgment information indicates that the battery is in a relatively stable state under the target monitoring task, after one-hot knowledge embedding, the corresponding second safety situation decision knowledge vector may be [0, 1, 0] (for example, the second position in the vector corresponding to the relatively stable state in the present invention).

[0054] In the feature pooling processing link, the state monitoring vector of the previous working state monitoring data can be subjected to feature pooling processing to obtain a state monitoring pooling vector. For example, the state monitoring vector contains multiple parameter information, such as the values ​​corresponding to the battery voltage ((V)), current ((I)) and temperature ((T)) and other parameters. Taking average pooling as an example, if the state monitoring vector is (vec{x}=[V, I, T]=[3.2, 50, 25]) (the numerical value of the present invention is consistent with the previous setting, (3.2) is the voltage value, (50) is the current value, and (25) is the temperature value), after average pooling processing, the state monitoring pooling vector (vec{x}_{pool}) is calculated as: (vec{x}_{pool}=frac{1}{3}(3.2+50+25)approx26.07) (the present invention is only an example of the calculation method of average pooling, and more complex calculations may be performed according to different pooling strategies in actual applications).

[0055] On the basis of the above content, calculating the security situation decision offset includes: using the second security situation decision knowledge vector and the first security situation decision knowledge vector to obtain the security situation decision offset. For example, the second security situation decision knowledge vector is (vec{v}_2=[0,1,0]), and the first security situation decision knowledge vector is (vec{v}_1=[0,1,0]), and the security situation decision offset can be calculated by vector subtraction (vec{Delta}=vec{v}_2-vec{v}_1=[0,0,0]) (the present invention is only an example of calculation method, and more complex calculations may be performed according to different logical relationships in actual applications).

[0056] In addition, calculating the state monitoring attention vector includes: obtaining the state monitoring attention vector according to the security situation decision offset and the state monitoring pooling vector. For example, the security situation decision offset is (vec{Delta}), and the state monitoring pooling vector is (vec{x}_{pool}), and the state monitoring attention vector can be obtained through a certain mapping relationship. For example, the mapping relationship is (vec{y}=vec{Delta}*vec{x}_{pool}) (the present invention is only an example relationship of an example, and in actual applications, it may be calculated based on more factors and complex algorithms). If (vec{Delta}=[0,0,0]) and (vec{x}_{pool}approx26.07), then (vec{y}=[0,0,0]) (the present invention is a special case calculated according to the example, and the actual situation will have different results according to different inputs).

[0057] Based on the above content, the detailed description of obtaining the state attention prediction vector in S204 is as follows.

[0058] (I) Obtaining the first weighted monitoring feature Calculation based on the original task attention weight and the state monitoring vector: In the process of performing task attention processing on the state monitoring vector according to the original task attention weight to obtain the state attention prediction vector, firstly, the first weighted monitoring feature is obtained based on the original task attention weight and the state monitoring vector. For example, the original task attention weight is (vec{w}_0=[0.25, 0.35, 0.4]) (consistent with the previous setting), and the state monitoring vector is (vec{x}=[3.2, 50, 25]) (the present invention also uses the previous example values). The calculation of the first weighted monitoring feature (vec{x}_w) is: (vec{x}_w=[3.2*0.25, 50*0.35, 25*0.4]=[0.8, 17.5, 10]).

[0059] (II) Feature pooling to obtain state attention prediction vector Perform feature pooling on the first weighted monitoring feature: Perform feature pooling on the first weighted monitoring feature to obtain a state attention prediction vector. For example, if the first weighted monitoring feature is (vec{x}_w=[0.8, 17.5, 10]), then the state attention prediction vector (vec{z}) is the maximum value in (vec{x}_w), that is, (vec{z}=17.5) (the present invention uses maximum pooling as an example for calculation, and different pooling strategies may be used in actual applications).

[0060] In this way, through the above-mentioned detailed operations in S204, including unique hot knowledge embedding, feature pooling processing and other steps, the previous working state monitoring data can be processed more accurately. Unique hot knowledge embedding converts the monitoring discrimination information into vector form, which is convenient for subsequent calculation and analysis. Feature pooling processing can effectively integrate and abstract the data, reduce the dimension of the data while retaining key information. In the process of calculating the state monitoring attention vector and the state attention prediction vector, through clear calculation steps and numerical calculations in examples, you can see how the data is converted and processed according to different weights, discrimination information and pooling strategies. This precise processing method helps to improve the accuracy and effectiveness of the entire algorithm, thereby optimizing the debugging process of the original state matching analysis algorithm, and ultimately improving the performance of the target state matching analysis algorithm in the state monitoring and analysis of the energy storage battery module, ensuring the stable operation of the energy storage battery system and improving its safety and reliability.

[0061] In another optional embodiment, the original task attention weight belongs to a self-attention network; based on this, the original task attention weight is optimized according to the difference between the state attention prediction vector and the state monitoring attention vector to obtain the target task attention weight, including: obtaining a training error representation of the self-attention network; determining an error variable of the training error representation according to the difference between the state attention prediction vector and the state monitoring attention vector; optimizing the original task attention weight of the self-attention network according to the error variable of the training error representation to obtain the target task attention weight.

[0062] In detail, the training error representation includes a first training error representation and a second training error representation; the determining the error variable of the training error representation based on the difference between the state attention prediction vector and the state monitoring attention vector includes: determining the error variable of the first training error representation based on the difference between the state attention prediction vector and the state monitoring attention vector; determining the error variable of the second training error representation based on the difference between the state attention prediction vector and the state monitoring vector; obtaining the error variable of the training error representation according to the error variable of the first training error representation and the error variable of the second training error representation.

[0063] In the above embodiment, an exemplary implementation method of optimizing the original task attention weight to obtain the target task attention weight is as follows.

[0064] 1. Obtaining the training error representation of the self-attention network 1) Relationship between the self-attention network and the original task attention weights: In this embodiment, the original task attention weights belong to the self-attention network. The self-attention network is a network structure that can automatically focus on the importance of different parts of the data when processing sequence data. In the process of optimizing the original task attention weights, the training error representation of the self-attention network must first be obtained. The training error representation is an indicator that measures the difference between the self-attention network and the expected output during the training process. It can contain multiple parts, such as the first training error representation and the second training error representation in this example.

[0065] For example, the input of the self-attention network is the processed previous working state monitoring data (such as data in vector form after some feature extraction), and the output is the result related to the state matching analysis (such as the predicted classification result of the battery state). The calculation of the training error representation may be based on the difference between the network output and the actual known correct result (such as the true state of the battery obtained through precise measurement and analysis).

[0066] 2. Determine the error variable for training error representation 1) Determine the error variable of the first training error representation: Determine the error variable of the first training error representation based on the difference between the state attention prediction vector and the state monitoring attention vector. For example, the state attention prediction vector is (vec{z}=[z_1, z_2, z_3]), and the state monitoring attention vector is (vec{y}=[y_1, y_2, y_3]). The mean square error (MSE) algorithm can be used to calculate the error variable of the first training error representation. The calculation formula of the mean square error is (MSE=frac{1}{n}sum_{i=1}^{n}(z_i-y_i)^2), the present invention (n=3) (for example, the vector dimension is 3). For example, if (vec{z}=[1, 2, 3]), (vec{y}=[1.5, 2.2, 2.8]), then (MSE=frac{1}{3}[(1-1.5)^2+(2-2.2)^2+(3-2.8)^2]).

[0067] Then the calculation can be obtained: (MSE=frac{1}{3}[0.25+0.04+0.04]=frac{1}{3}[0.33]approx0.11) The above (0.11) is an example value of the error variable represented by the first training error.

[0068] 2) Determine the error variable of the second training error representation: Determine the error variable of the second training error representation based on the difference between the state attention prediction vector and the state monitoring vector. For example, the state monitoring vector is (vec{x}=[x_1, x_2, x_3]), and the mean square error algorithm is still used as an example. The calculation method is (MSE'=frac{1}{n}sum_{i=1}^{n}(z_i-x_i)^2). For example, if (vec{x}=[0.8, 1.8, 2.5]), (vec{z}=[1, 2, 3]), then (MSE'=frac{1}{3}[(1-0.8)^2+(2-1.8)^2+(3-2.5)^2]).

[0069] Then the calculation can be obtained: (MSE'=frac{1}{3}[0.04+0.04+0.25]=frac{1}{3}[0.33]approx0.11). The above (0.11) is an example value of the error variable represented by the second training error.

[0070] 3) Obtain the error variable represented by the training error: Obtain the error variable represented by the training error according to the error variable represented by the first training error and the error variable represented by the second training error. The weighted summation method of the example can be used, for example, the weight of the error variable represented by the first training error is (w_1=0.5), and the weight of the error variable represented by the second training error is (w_2=0.5). Then the error variable represented by the training error (E=w_1*MSE+w_2*MSE') is substituted into the value calculated above (E=0.5*0.11+0.5*0.11=0.11).

[0071] (III) Optimizing the original task attention weight to obtain the target task attention weight 1) Optimization based on error variables: Based on the error variables represented by the training error, the original task attention weights of the self-attention network are optimized to obtain the target task attention weights. For example, if the original task attention weights are (vec{w}_0=[w_{01}, w_{02}, w_{03}]), a gradient descent-based optimization algorithm can be used. For example, the update formula for (w_{01}) is (w_{01}^{new}=w_{01}-alpha*frac{partialE}{partialw_{01}}), where (alpha) is the learning rate, for example (alpha=0.01).

[0072] In the present invention, to calculate (frac{partialE}{partialw_{01}}), it is necessary to determine it according to the structure of the self-attention network and the previous calculation relationship.

[0073] For example, after complex calculations (the present invention omits the specific derivation process related to the internal structure of the self-attention network), we get (frac{partialE}{partialw_{01}}=0.5), then (w_{01}^{new}=w_{01}-0.01*0.5). If (w_{01}=0.2), then (w_{01}^{new}=0.2-0.005=0.195).

[0074] The same method can be applied to the update of (w_{02}) and (w_{03}), and finally the optimized target task attention weight (vec{w}_1) is obtained.

[0075] It can be seen that by optimizing the original task attention weights under the framework of the self-attention network and using the training error representation and its error variables, the original task attention weights can be made more adaptable to the actual working state monitoring data. Through specific numerical examples, such as the error variables obtained by the mean square error calculation, the difference between the state attention prediction vector and the state monitoring attention vector and the state monitoring vector can be accurately quantified, providing a clear basis for weight optimization. Updating the original task attention weights based on the gradient descent algorithm can gradually adjust the weights to develop in a better direction. This optimization method helps to improve the accuracy of the entire algorithm when processing the working state monitoring data of the energy storage battery, thereby improving the performance of the target state matching analysis algorithm, enhancing the reliability of the energy storage battery state judgment, and ensuring the safe and stable operation of the energy storage system.

[0076] In an exemplary embodiment, the state monitoring vector is subjected to task attention processing based on the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data, including: determining whether the preceding working state monitoring data corresponding to the state monitoring vector belongs to a candidate preceding working state monitoring data set, the candidate preceding working state monitoring data set including at least one preceding working state monitoring data carrying an attention processing identifier, the at least one preceding working state monitoring data carrying an attention processing identifier being determined from the preceding working state monitoring data set in which the preceding working state monitoring data is located based on a target sampling strategy; when the preceding working state monitoring data corresponding to the state monitoring vector belongs to the candidate preceding working state monitoring data set, the state monitoring vector is subjected to task attention processing based on the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data.

[0077] Furthermore, a detailed description of obtaining the target state monitoring attention vector according to the target task attention weight is as follows.

[0078] 1. Determine whether the data corresponding to the condition monitoring vector belongs to the candidate data set 1) Composition of candidate previous working state monitoring data sets: In the process of performing task attention processing on the state monitoring vector according to the target task attention weight to obtain the target state monitoring attention vector of the previous working state monitoring data, it is first necessary to determine whether the previous working state monitoring data corresponding to the state monitoring vector belongs to the candidate previous working state monitoring data set. The candidate previous working state monitoring data set contains at least one previous working state monitoring data with an attention processing identifier. This data set is determined from the previous working state monitoring data set where the previous working state monitoring data is located according to the target sampling strategy.

[0079] For example, the previous working state monitoring data set may contain a large amount of working state data of energy storage battery modules, which covers different time points, different battery working modes and other situations. For example, this data set contains 1,000 data points (each data point corresponds to the previous working state monitoring data at a moment). The target sampling strategy can be selective sampling based on certain characteristics of the data. For example, stratified sampling is performed according to the charge and discharge state of the battery. If the charge and discharge state of the battery is divided into three levels: high, medium and low, according to the target sampling strategy, sampling may be performed from each level according to a certain proportion (such as 10% from each level), thus obtaining a candidate previous working state monitoring data set. For example, after such sampling, the candidate previous working state monitoring data set contains 30 previous working state monitoring data carrying attention processing identifiers.

[0080] 2) Determine the belonging relationship: For the previous working state monitoring data corresponding to the state monitoring vector, it is necessary to determine whether it belongs to this candidate previous working state monitoring data set. For example, the previous working state monitoring data corresponding to the state monitoring vector may be a data point composed of data such as the voltage, current, and temperature of the battery module at a specific moment. By comparing this data point with the data features in the candidate previous working state monitoring data set (such as by comparing whether the voltage value range, current value range, etc. in the data point overlap with the data in the candidate data set), determine whether it belongs to the candidate data set.

[0081] (ii) Obtaining the target state monitoring attention vector 1) Processing when it belongs to the candidate data set: When the previous working state monitoring data corresponding to the state monitoring vector belongs to the candidate previous working state monitoring data set, the state monitoring vector is processed with task attention according to the target task attention weight to obtain the target state monitoring attention vector of the previous working state monitoring data. For example, the state monitoring vector is (vec{x}=[x_1, x_2, x_3]), where (x_1) represents voltage data (for example, 3.2), (x_2) represents current data (for example, 50), and (x_3) represents temperature data (for example, 25), and the target task attention weight is (vec{w}_1=[w_{11}, w_{12}, w_{13}]) (for example, (w_{11}=0.3), (w_{12}=0.4), (w_{13}=0.3)).

[0082] The task attention processing is calculated as follows: the element (y_i=x_i*w_{1i}) of the target state monitoring attention vector (vec{y}).

[0083] but: (y_1=x_1*w_{11}=3.2*0.3=0.96); (y_2=x_2*w_{12}=50*0.4=20); (y_3=x_3*w_{13}=25*0.3=7.5); so the target state monitoring attention vector (vec{y}=[0.96, 20, 7.5]).

[0084] It can be seen that by determining whether the data corresponding to the state monitoring vector belongs to the candidate data set, task attention processing can be performed on some data in a targeted manner. This method selects data based on the target sampling strategy, avoiding the waste of computing resources and unnecessary computing complexity that may be caused by the unified processing of all previous working state monitoring data. In actual energy storage battery state monitoring, the amount of data is often large. Through this selective processing method, the efficiency of the algorithm can be improved. At the same time, when the data belongs to the candidate data set, task attention processing is performed according to the target task attention weight to obtain the target state monitoring attention vector, which can re-weight the data according to the importance of the data (reflected by the target task attention weight), so that subsequent operations such as algorithm debugging based on the target state monitoring attention vector are more accurate, and the effectiveness and reliability of the entire algorithm in energy storage battery state monitoring and analysis are improved.

[0085] In another exemplary embodiment, the state monitoring vector is subjected to task attention processing based on the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data, including: obtaining a second weighted monitoring feature using the target task attention weight and the state monitoring vector; and performing feature pooling processing on the second weighted monitoring feature to obtain the target state monitoring attention vector of the preceding working state monitoring data.

[0086] It can be understood that the core details of this exemplary implementation include the following two points.

[0087] (I) Obtaining the second weighted monitoring feature 1) Calculation based on target task attention weight and state monitoring vector: In the process of performing task attention processing on the state monitoring vector according to the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data, the target task attention weight and the state monitoring vector must first be used to obtain the second weighted monitoring feature.

[0088] For example, the state monitoring vector is (vec{x}=[x_1, x_2, x_3]), where (x_1) represents voltage data (e.g. (x_1=3.2)), (x_2) represents current data (e.g. (x_2=50)), and (x_3) represents temperature data (e.g. (x_3=25)). The target task attention weight is (vec{w}_1=[w_{11}, w_{12}, w_{13}]) (e.g. (w_{11}=0.3), (w_{12}=0.4), (w_{13}=0.3)).

[0089] The calculation formula of the second weighted monitoring feature (vec{x}_w) is: (vec{x}_w =[x_1*w_{11}, x_2*w_{12}, x_3*w_{13}] =[3.2*0.3, 50*0.4, 25*0.3]=[0.96, 20, 7.5]).

[0090] (II) Feature pooling to obtain target state monitoring attention vector 1) Type and calculation of feature pooling: The second weighted monitoring feature is subjected to feature pooling to obtain the target state monitoring attention vector of the previous working state monitoring data. Feature pooling can be performed in a variety of ways, such as average pooling, maximum pooling, etc.

[0091] If average pooling is used, the calculation method is: the value of the target state monitoring attention vector (vec{y}) is (vec{y}=frac{1}{n}sum_{i=1}^{n}x_{wi}), where (n=3) (because (vec{x}_w) has three elements). Substituting (vec{x}_w=[0.96, 20, 7.5]) into the calculation, we get (vec{y}=frac{1}{3}(0.96+20+7.5)=frac{1}{3}(28.46)approx9.49).

[0092] If max pooling is used, the value of the target state monitoring attention vector (vec{y}) is (vec{y}=max(x_{w1}, x_{w2}, x_{w3})). For (vec{x}_w=[0.96, 20, 7.5]), then (vec{y}=20).

[0093] In addition, other pooling methods can also be used, such as minimum pooling. The calculation method is: (vec{y}=min(x_{w1}, x_{w2}, x_{w3})).

[0094] For (vec{x}_w=[0.96, 20, 7.5]), then (vec{y}=0.96). Different pooling methods reflect different focuses on the second weighted monitoring feature. Average pooling considers the average value of all elements, maximum pooling focuses on the largest eigenvalue, and minimum pooling focuses on the smallest eigenvalue. These different target state monitoring attention vectors will have different effects in subsequent algorithm debugging and other operations.

[0095] In this way, by using the target task attention weight and the state monitoring vector to obtain the second weighted monitoring feature, the state monitoring vector is re-weighted according to the target task attention weight, so that the data can be adjusted according to the importance of different parameters. For example, in the above numerical example, the second weighted monitoring feature is obtained by weighting the voltage, current and temperature data according to the target task attention weight. The subsequent feature pooling processing, whether it is average pooling, maximum pooling or minimum pooling, further integrates the weighted features. This integration method can effectively reduce the dimension of the data and integrate multiple features into a target state monitoring attention vector that can represent the characteristics of the previous working state monitoring data. This helps to simplify the subsequent algorithm processing process, improve the computational efficiency and accuracy of the algorithm, and thus improve the effectiveness of the entire energy storage battery state monitoring system for battery state evaluation and analysis.

[0096] In some preferred embodiments, the method in S208 debugs the original state matching analysis algorithm based on the target state monitoring attention vector to obtain the target state matching analysis algorithm, including: performing state matching analysis based on the target state monitoring attention vector through the original state matching analysis algorithm to obtain the state matching analysis result of the preceding working state monitoring data; debugging the original state matching analysis algorithm based on the comparison viewpoint between the state matching analysis result and the prior training annotations of the preceding working state monitoring data to obtain the target state matching analysis algorithm.

[0097] In the above preferred embodiment, the original state matching analysis algorithm is debugged according to the target state monitoring attention vector to obtain a detailed description of the target state matching analysis algorithm through the following two aspects.

[0098] (I) The original algorithm uses the target state to monitor the attention vector and perform state matching analysis to obtain the result 1) Operation of the original state matching analysis algorithm: In the process of debugging the original state matching analysis algorithm according to the target state monitoring attention vector to obtain the target state matching analysis algorithm, the original state matching analysis algorithm must first be used to perform state matching analysis according to the target state monitoring attention vector to obtain the state matching analysis result of the previous working state monitoring data. The original state matching analysis algorithm may be based on some distance metric or classification algorithm.

[0099] For example, the original state matching analysis algorithm is an algorithm based on Euclidean distance. For example, the target state monitoring attention vector obtained by processing the previous working state monitoring data is (vec{y}=[9.49]) (the present invention uses the result obtained by the previous average pooling as an example), and there is a set of predefined standard state vectors (vec{s}_1=[8]), (vec{s}_2=

[10] ), (vec{s}_3=

[12] ), etc. (these standard state vectors represent different known battery working states).

[0100] Then calculate the Euclidean distance (d_i=sqrt{(vec{y}-vec{s}_i)^2} between the target state monitoring attention vector (vec{y}) and each standard state vector (vec{s}_i). For (vec{s}_1=[8]), (d_1=sqrt{(9.49-8)^2}=sqrt{1.49^2}=1.49); for (vec{s}_2=

[10] ), (d_2=sqrt{(9.49-10)^2}=sqrt{(-0.51)^2}=0.51); for (vec{s}_3=

[12] ), (d_3=sqrt{(9.49-12)^2}=sqrt{(-2.51)^2}=2.51), etc. Based on these distance calculation results, the closeness of the previous working state monitoring data to each predefined standard state can be determined, thereby obtaining the state matching analysis result. For example, if the standard state with the smallest distance is set as the matching state, then in this example, the state matching analysis result of the previous working state monitoring data may be the most matching with the state represented by (vec{s}_2=

[10] ).

[0101] (II) Debug the original algorithm based on the comparison point to obtain the target algorithm 1) Comparison between prior training annotations and state matching analysis results: Based on the comparison between the state matching analysis results and the prior training annotations of the previous working state monitoring data, the original state matching analysis algorithm is debugged to obtain the target state matching analysis algorithm. The prior training annotations are the pre-known identifiers of the correct state of the previous working state monitoring data.

[0102] For example, the prior training annotations indicate that the correct state corresponding to the previous working state monitoring data is consistent with the state represented by (vec{s}_1=[8]). However, according to the previous state matching analysis results, the best match is obtained with (vec{s}_2=

[10] ), which is a difference.

[0103] In order to adjust this difference, the original state matching analysis algorithm needs to be debugged. This can be achieved by adjusting the parameters in the algorithm. For example, in an algorithm based on Euclidean distance, if it is found that the distance calculation result does not match the prior training annotation, the weight in the distance calculation can be adjusted or the distance measurement method can be corrected. For example, in the original Euclidean distance calculation, the default weight for each element of the target state monitoring attention vector (vec{y}) is (w=[1]), which can now be adjusted according to the comparison point of view. If it is found that the difference between the value in (vec{y}) and the prior training annotation is mainly in a specific direction (for example, it may be due to improper processing of voltage-related data, and the contribution weight of voltage data in the target state monitoring attention vector (vec{y}) may need to be adjusted), the weight can be adjusted to (w=[0.8]) (the present invention is only an example of adjusting the weight).

[0104] After many such adjustments, repeated comparisons and debugging were performed based on different previous working state monitoring data and their corresponding prior training annotations, and finally the target state matching analysis algorithm was obtained. This target algorithm can more accurately perform state matching analysis based on the target state monitoring attention vector, so that its results are more consistent with the correct state represented by the prior training annotations.

[0105] By performing state matching analysis based on the target state monitoring attention vector through the original state matching analysis algorithm, the characteristic information of the previous working state monitoring data contained in the target state monitoring attention vector can be used to make state judgments. Then, the original algorithm is debugged based on the comparison viewpoint between the state matching analysis results and the prior training annotations, and the deviation of the original algorithm in processing data can be found and corrected. For example, by adjusting the weights in the Euclidean distance calculation, the algorithm can more accurately match and analyze the working state of the energy storage battery. This helps to improve the accuracy and reliability of the entire algorithm when processing energy storage battery state monitoring data, optimize the algorithm performance, and thus better meet the needs of energy storage battery state monitoring and management.

[0106] In an alternative technical solution, before the state matching analysis is performed according to the target state monitoring attention vector through the original state matching analysis algorithm to obtain the state matching analysis result of the previous working state monitoring data, the method also includes: obtaining a plurality of security situation decision information, the plurality of security situation decision information including: a first security situation decision information indicating a stage safety and stability situation, and a second security situation decision information indicating each of the potential safety hazard situations in at least one potential safety hazard situation. Based on this, the state matching analysis is performed according to the target state monitoring attention vector through the original state matching analysis algorithm to obtain the state matching analysis result of the preceding working state monitoring data, including: performing unique hot knowledge embedding on each of the safety situation decision information through the original state matching analysis algorithm to obtain the safety situation decision knowledge vector of each of the safety situation decision information; performing state matching analysis according to the target state monitoring attention vector and each of the safety situation decision knowledge vectors through the original state matching analysis algorithm to obtain the matching confidence weight of the preceding energy storage battery module label corresponding to each safety situation in the preceding working state monitoring data; wherein, the safety situation includes the stage safety and stability situation and the at least one potential safety hazard situation, and the state matching analysis result of the preceding working state monitoring data includes: the matching confidence weight of the preceding energy storage battery module label corresponding to each safety situation.

[0107] Furthermore, the state matching analysis is performed based on the target state monitoring attention vector and each of the safety situation decision knowledge vectors to obtain the matching confidence weights of the preceding energy storage battery module labels in the preceding working state monitoring data corresponding to each safety situation, including: performing full connection mapping on the target state monitoring attention vector to obtain a fully connected monitoring mapping attention vector, and performing state monitoring upsampling on the fully connected monitoring mapping attention vector to obtain a target state monitoring vector; obtaining a global monitoring vector through the target state monitoring attention vector and the target state monitoring vector; performing state matching analysis based on the global monitoring vector and each of the safety situation decision knowledge vectors to obtain the matching confidence weights of the preceding energy storage battery module labels in the preceding working state monitoring data corresponding to each safety situation.

[0108] Under a more detailed design concept, the target state monitoring attention vector is fully connected mapped to obtain a fully connected monitoring mapping attention vector, and the fully connected monitoring mapping attention vector is state-monitored upsampled to obtain a target state monitoring vector, including: through the fully connected mapping module of the original state matching analysis algorithm, the target state monitoring attention vector is fully connected mapped to obtain a fully connected monitoring mapping attention vector, and the fully connected monitoring mapping attention vector is state-monitored upsampled to obtain a target state monitoring vector. Based on this, the original state matching analysis algorithm is debugged according to the comparison point of view between the state matching analysis result and the prior training annotation of the preceding working state monitoring data to obtain the target state matching analysis algorithm, including: according to the comparison point of view between the state matching analysis result and the prior training annotation of the preceding working state monitoring data, the algorithm network configuration weight of the fully connected mapping module is optimized to debug the original state matching analysis algorithm to obtain the target state matching analysis algorithm.

[0109] The above-mentioned replaceable technical solution involves the following three technical points.

[0110] 1. Obtaining Security Situation Decision-Making Information 1) Composition of safety situation decision information: First, several safety situation decision information should be obtained. This information includes first safety situation decision information indicating the safety and stability situation of the stage and second safety situation decision information indicating each potential safety hazard situation in at least one potential safety hazard situation. This information is of great significance for comprehensively evaluating the status of the energy storage battery module.

[0111] For example, for an energy storage battery module, the first safety situation decision information indicates that the battery is in a stable state of normal operation, which may be based on the fact that the battery's voltage, current, temperature and other parameters are within the normal range, and no abnormal chemical or physical changes are detected inside the battery. The second safety situation decision information may target different potential safety hazards, such as excessive voltage differences between battery cells (which may lead to unbalanced charging and discharging of the battery pack, affecting battery life and performance), too fast a rise in battery temperature (which may indicate a short circuit inside the battery or a failure in the cooling system), etc. For example, for a lithium-ion energy storage battery module, the voltage range of the single cell is between 3.2-3.6V during normal operation. When the voltage difference of the single cell exceeds 0.2V, the corresponding second safety situation decision information will be triggered, indicating that there is a potential battery balancing safety hazard.

[0112] 2. Obtaining state matching analysis results through the original algorithm 1. Unique Hot Knowledge Embedding 1) Processing of security situation decision information: In the process of performing state matching analysis based on the target state monitoring attention vector through the original state matching analysis algorithm to obtain the state matching analysis results of the previous working state monitoring data, it is first necessary to perform unique hot knowledge embedding on each security situation decision information to obtain the security situation decision knowledge vector of each security situation decision information.

[0113] For example, there are three types of safety situations, namely, the stage safety and stability situation, the battery cell voltage imbalance hidden danger situation, and the battery temperature abnormal hidden danger situation. For the first safety situation decision information (stage safety and stability situation), if one-hot knowledge embedding is used, for example, it is mapped to a three-dimensional vector space, then the corresponding safety situation decision knowledge vector may be [1, 0, 0], indicating that it is in the first safety situation. For the second safety situation decision information indicating the battery cell voltage imbalance hidden danger situation, its corresponding safety situation decision knowledge vector may be [0, 1, 0], and for the second safety situation decision information indicating the battery temperature abnormal hidden danger situation, the corresponding safety situation decision knowledge vector may be [0, 0, 1]. This one-hot knowledge embedding method converts different types of safety situation decision information into a vector form that is easy for the algorithm to process.

[0114] (II) Perform state matching analysis based on vectors to obtain matching confidence weights 1) Fully connected mapping and state monitoring upsampling: Perform fully connected mapping on the target state monitoring attention vector to obtain a fully connected monitoring mapping attention vector, and perform state monitoring upsampling on the fully connected monitoring mapping attention vector to obtain a target state monitoring vector.

[0115] For example, the target state monitoring attention vector is (vec{y}=[9.49]) (the previous example numerical value is used). Fully connected mapping is performed through the fully connected mapping module of the original state matching analysis algorithm. Fully connected mapping can be regarded as a linear transformation. For example, the weight matrix of the fully connected mapping is (W=begin{bmatrix}2end{bmatrix}) (in order to simplify the example, the present invention is a one-dimensional weight matrix, and the actual situation may be a multi-dimensional matrix). Then the calculation of the fully connected monitoring mapping attention vector (vec{y'}) is (vec{y'}=vec{y}*W=[9.49*2]=[18.98]).

[0116] Then, the fully connected monitoring mapping attention vector (vec{y'}) is subjected to state monitoring upsampling. State monitoring upsampling can be performed using methods such as interpolation. For example, the linear interpolation of the example is used for upsampling to expand the dimension of the vector from 1 to 3 (the present invention is only an example of dimensional expansion). If the upsampling ratio is expanded at equal intervals, then the target state monitoring vector (vec{z}) may become ([18.98, 18.98, 18.98]) (the actual upsampling algorithm will obtain more reasonable results based on more factors and complex calculations).

[0117] 2) Obtain the global monitoring vector and perform state matching analysis: Obtain the global monitoring vector through the target state monitoring attention vector and the target state monitoring vector. For example, the addition operation is used to obtain the global monitoring vector (vec{g}), then (vec{g} =vec{y}+vec{z} =[9.49+18.98,9.49+18.98,9.49+18.98] =[28.47, 28.47, 28.47]).

[0118] According to the global monitoring vector and each safety situation decision knowledge vector, a state matching analysis is performed to obtain the matching confidence weights of the preceding energy storage battery module labels corresponding to each safety situation in the preceding working state monitoring data. For example, an algorithm based on cosine similarity is used to calculate the matching confidence weights. For example, the safety situation decision knowledge vector is (vec{v}_1=[1,0,0]) (stage safety and stability situation), (vec{v}_2=[0,1,0]) (battery cell voltage imbalance hidden danger situation), and (vec{v}_3=[0,0,1]) (battery temperature abnormal hidden danger situation) obtained previously.

[0119] Calculate the cosine similarity between the global monitoring vector (vec{g}) and (vec{v}_1) (s_1=frac{vec{g}cdotvec{v}_1}{vertvec{g}vertvertvec{v}_1vert}=frac{28.47*1+28.47*0+28.47*0}{sqrt{28.47^2+28.47^2+28.47^2}*1}=frac{28.47}{sqrt{3*28.47^2}}approx0.577).

[0120] Similarly, calculate the cosine similarity (s_2) between (vec{g}) and (vec{v}_2) and the cosine similarity (s_3) between (vec{v}_3). For example, (s_2=0.2), (s_3=0.3). Then normalize these similarities to get the matching confidence weight. The normalization method can be: (w_1 =frac{s_1}{s_1+s_2+s_3} =frac{0.577}{0.577+0.2+0.3}approx0.52), (w_2 =frac{s_2}{s_1+s_2+s_3}=frac{0.2}{1.077}approx0.19), (w_3 =frac{s_3}{s_1+s_2+s_3}=frac{0.3}{1.077}approx0.28).

[0121] These matching confidence weights indicate the likelihood that the preceding energy storage battery module is in different safety situations.

[0122] 3. Debug the original algorithm based on the comparison point to obtain the target algorithm 1) Optimizing the algorithmic network configuration weights of the fully connected mapping module: Based on the comparison between the state matching analysis results and the prior training annotations of the previous working state monitoring data, the algorithmic network configuration weights of the fully connected mapping module are optimized to debug the original state matching analysis algorithm and obtain the target state matching analysis algorithm.

[0123] For example, the prior training annotations indicate that the preceding energy storage battery module is actually in a stage of safe and stable status, that is, the corresponding correct matching confidence weight should be close to 1 for the first safety status (for example, ideally (w_{1}^{*}=0.9), (w_{2}^{*}=0.05), (w_{3}^{*}=0.05)), but there are differences according to the previously calculated (w_1=0.52), (w_2=0.19), (w_3=0.28).

[0124] In order to adjust this difference, it is necessary to optimize the algorithm network configuration weights of the fully connected mapping module. For example, the algorithm network configuration weights of the fully connected mapping module include the weight matrix of the fully connected mapping mentioned above (W=begin{bmatrix}2end{bmatrix}). The weights can be adjusted using an optimization algorithm based on gradient descent. First, calculate the loss function, for example, using the mean square error (MSE) as the loss function (L=frac{1}{n}sum_{i=1}^{n}(w_i-w_{i}^{*})^2), the present invention (n=3), substitute the numerical value to get (L=frac{1}{3}[(0.52-0.9)^2+(0.19-0.05)^2+(0.28-0.05)^2]) Then the calculation can be obtained: (L =frac{1}{3}[0.1444+0.0196+0.0529] =frac{1}{3}[0.2169]approx0.0723).

[0125] Then, according to the gradient descent algorithm, the gradient (nablaL) of the loss function to the weight matrix (W) is calculated (the calculation process of the present invention is omitted, and it is actually necessary to calculate according to the specific structure and calculation relationship of the fully connected mapping module), for example (nablaL=0.1), the formula for updating the weight matrix (W) is (W^{new}=W-alpha*nablaL), where (alpha) is the learning rate, for example (alpha=0.01), then (W^{new}=2-0.01*0.1=1.999). Through multiple such iterations, adjustments are made according to different previous working state monitoring data and their prior training annotations, and finally the algorithm network configuration weights of the optimized fully connected mapping module are obtained, thereby obtaining the target state matching analysis algorithm.

[0126] In this way, by obtaining the safety situation decision information, the different safety states of the energy storage battery module can be fully considered, including the stable state and the potential safety hazard state. Unique hot knowledge embedding converts the safety situation decision information into vector form, which is convenient for subsequent state matching analysis. The fully connected mapping and state monitoring upsampling operations process the target state monitoring attention vector to obtain the target state monitoring vector, and then obtain the global monitoring vector, which helps to integrate feature information at different levels. By calculating the matching confidence weight, the possibility of the preceding energy storage battery module being in different safety situations can be quantified. Finally, the algorithm network configuration weights of the fully connected mapping module are optimized according to the comparison point of view to debug the original algorithm, which improves the accuracy of the algorithm and makes the state matching analysis results more in line with the actual situation, thereby better ensuring the safe operation and effective management of the energy storage battery.

[0127] It is worth mentioning that in the present invention, S106 is the core invention point, which involves the state data recognition system performing state matching analysis based on the IoT sensor state monitoring vector and each safety situation decision knowledge vector through the target state matching analysis algorithm, thereby obtaining the state matching analysis result of the power storage working state monitoring data. This process is a key link in the entire energy storage battery module state monitoring and evaluation system, which will be described in detail below.

[0128] First of all, the IoT sensor state monitoring vector is the result of mining the state monitoring vector of the power storage working state monitoring data. It integrates various information about the energy storage battery module obtained from multiple sensors, such as voltage, current, temperature and other data, and forms a comprehensive vector after processing. Taking the lithium-ion energy storage battery module as an example, for example, the voltage data obtained from the voltage sensor is 3.2V (this value becomes an element in the IoT sensor state monitoring vector after normalization), the current obtained by the current sensor is 50A (also normalized), and the temperature obtained by the temperature sensor is 25°C (also normalized). These processed values ​​are combined to form the IoT sensor state monitoring vector. This vector fully reflects the working state information of the energy storage battery module at a certain moment.

[0129] The safety situation decision knowledge vector is obtained by performing unique hot knowledge embedding on each safety situation decision information. The safety situation decision information includes the first safety situation decision information indicating the stage safety and stability situation and the second safety situation decision information indicating at least one potential safety hazard situation. For example, for the stage safety and stability situation, if the safety situation is divided into three types (safe and stable, mild hidden danger, serious hidden danger), then the safety situation decision knowledge vector corresponding to the first safety situation decision information may be [1, 0, 0] after the unique hot knowledge embedding, indicating that it is in a safe and stable situation. For the second safety situation decision information of a potential safety hazard situation (such as the battery cell voltage imbalance hazard), the corresponding safety situation decision knowledge vector may be [0, 1, 0]. These vectors convert different safety situation decision information into mathematical forms that are convenient for algorithm processing.

[0130] Secondly, the target state matching analysis algorithm is an algorithm specially designed to process the IoT sensor state monitoring vector and the security situation decision knowledge vector to obtain accurate state matching analysis results. The algorithm may be constructed based on a variety of mathematical models and algorithm principles.

[0131] One possible approach is an algorithm based on distance measurement. For example, the Euclidean distance between the IoT sensor state monitoring vector and each safety situation decision knowledge vector is calculated. For example, the IoT sensor state monitoring vector is (vec{x}=[x_1, x_2, x_3]) (in the present invention, (x_1), (x_2), (x_3) are processed voltage, current, and temperature related values, respectively), and the safety situation decision knowledge vectors are (vec{v}_1=[v_{11}, v_{12}, v_{13}]) (corresponding to the stage safety and stability situation) and (vec{v}_2=[v_{21}, v_{22}, v_{23}]) (corresponding to a potential safety hazard situation), etc. Then the Euclidean distance between the IoT sensor state monitoring vector and (vec{v}_1) is (d_1=sqrt{sum_{i=1}^{3}(x_i-v_{1i})^2}), and the Euclidean distance between (vec{v}_2) is (d_2=sqrt{sum_{i=1}^{3}(x_i-v_{2i})^2}). By comparing the sizes of these distances, we can preliminarily judge the degree of closeness between the working state of the energy storage battery module and different safety situations. The smaller the distance, the closer the current working state is to the state corresponding to the safety situation.

[0132] Another possible method is to use an algorithm based on similarity calculation, such as the cosine similarity algorithm. The cosine similarity between the IoT sensor state monitoring vector and each security situation decision knowledge vector is calculated. The formula for calculating cosine similarity is: (s=frac{vec{x}cdotvec{v}}{vertvec{x}vertvertvec{v}vert}); where (vec{x}cdotvec{v}) is the dot product of the vectors, and (vertvec{x}vert) and (vertvec{v}vert) are the magnitudes of the vectors (vec{x}) and (vec{v}) respectively.

[0133] By calculating the cosine similarity between the IoT sensor state monitoring vector and the different safety situation decision knowledge vectors, a value between -1 and 1 can be obtained. The closer it is to 1, the more similar the two vectors are, that is, the closer the working state of the energy storage battery module is to the state corresponding to the safety situation.

[0134] Then, the state matching analysis result includes the matching confidence weight of the energy storage battery module tag corresponding to each safety situation. The safety situation of the present invention includes a stage safety stability situation and at least one potential safety hazard situation.

[0135] In the case of distance metrics (such as Euclidean distance), the matching confidence weight can be calculated based on the inverse of the distance. For example, the Euclidean distance between the IoT sensor state monitoring vector and (vec{v}_1) (stage safety and stability situation) is calculated to be (d_1), and the Euclidean distance between the IoT sensor state monitoring vector and (vec{v}_2) (a potential safety hazard situation) is (d_2). First, calculate the inverse of the distance (r_1=frac{1}{d_1}), (r_2=frac{1}{d_2}), and then calculate the matching confidence weight: (w_1=frac{r_1}{r_1+r_2}), (w_2=frac{r_2}{r_1+r_2}).

[0136] The matching confidence weight obtained in this way indicates the relative probability of the energy storage battery module being in different safety situations. For example, if (w_1=0.8) and (w_2=0.2), it means that the energy storage battery module is more likely to be in a stage of safe and stable situation, and the possibility of being in a potential safety hazard situation is relatively small.

[0137] In the case of similarity calculation (such as cosine similarity), the calculated similarity is directly used as the basis for matching confidence weights. For example, the calculated cosine similarity between the IoT sensor state monitoring vector and (vec{v}_1) is (s_1), and the cosine similarity with (vec{v}_2) is (s_2). Then normalization is performed to obtain matching confidence weights (w_1=frac{s_1}{s_1+s_2}), (w_2=frac{s_2}{s_1+s_2}). This matching confidence weight also reflects the possibility of the energy storage battery module being in different safety situations.

[0138] In addition, the target state matching analysis algorithm may also consider other factors to adjust the matching confidence weight. For example, the historical working state data of the energy storage battery module, the current working environment factors (such as humidity, air pressure, etc., which may be obtained through other sensors and included in the algorithm after processing), etc. If the energy storage battery module has been in a stable working state for a period of time in the past, then in the current state matching analysis, the matching confidence weight of the stage safety and stability situation may be appropriately improved.

[0139] Finally, by obtaining accurate state matching analysis results, especially the matching confidence weights of energy storage battery module labels corresponding to various safety situations, it can provide an important basis for the management and maintenance of energy storage batteries.

[0140] During the operation of the energy storage battery system, if it is found that a certain energy storage battery module is in a potential safety hazard situation with a high matching confidence weight (for example, exceeding a certain threshold, such as 0.5), corresponding measures can be taken in time. For example, if the matching confidence weight of the battery cell voltage imbalance hazard situation is high, the battery module can be charged evenly and the charging strategy in the battery management system can be adjusted to ensure the safe and stable operation of the battery module. For large-scale energy storage power stations, this precise state matching analysis result helps to improve the reliability and safety of the entire energy storage system and reduce the probability of safety accidents. It also helps to optimize the service life of the energy storage battery and reduce maintenance costs.

[0141] In summary, the state matching analysis technical solution in S106 effectively integrates the IoT sensor state monitoring vector and the safety situation decision-making knowledge vector through the target state matching analysis algorithm, and obtains the state matching analysis results of the power storage working state monitoring data with practical application value, which provides key technical support for the state monitoring, management and maintenance of energy storage battery modules.

[0142] In an extended technical solution based on the core invention point of the present invention, after the state matching analysis is performed according to the IoT sensor state monitoring vector and each of the safety situation decision knowledge vectors through the target state matching analysis algorithm to obtain the state matching analysis result of the power storage working state monitoring data, the method further includes: when the state matching analysis result includes a target potential safety hazard situation in the at least one potential safety hazard situation, based on the state marking data corresponding to the target potential safety hazard situation, performing abnormal risk detection processing on the energy storage system to which the energy storage battery module tag belongs.

[0143] In the extended technical solution of the core invention point of the present invention, first, the state matching analysis result includes the matching confidence weight of the energy storage battery module label corresponding to each safety situation (including the stage safety and stability situation and at least one potential safety hazard situation). When this state matching analysis result contains a target potential safety hazard situation in at least one potential safety hazard situation, it will trigger the subsequent abnormal risk detection process. The target potential safety hazard situation of the present invention refers to a specific hazard situation that needs to be focused on among many potential safety hazard situations, which is determined based on factors such as matching confidence weights. For example, if in the state matching analysis result, the matching confidence weight of the battery cell voltage imbalance hazard situation exceeds a preset threshold (for example, this threshold is 0.3), then the battery cell voltage imbalance hazard situation is determined as the target potential safety hazard situation.

[0144] Secondly, the status marking data is a specific data identifier associated with the target potential safety hazard situation. For different potential safety hazard situations, the status marking data has different definitions and acquisition methods. Taking the battery cell voltage imbalance hazard situation as an example, the status marking data may include the actual voltage value of each single cell in the battery module, the statistical information of the voltage difference, etc. For example, a storage battery module is composed of 10 single cells. When the battery cell voltage imbalance hazard situation is detected as the target potential safety hazard situation, the status marking data may be that the voltage values ​​of these 10 single cells are [3.2, 3.1, 3.3, 3.0, 3.4, 3.15, 3.25, 3.05, 3.35, 3.1] (in volts), and the calculated maximum voltage difference is 0.4 volts. These status marking data provide specific data basis for subsequent abnormal risk detection and processing.

[0145] Then, the process of abnormal risk detection and processing of the energy storage system to which the energy storage battery module tag belongs is carried out based on the status tag data corresponding to the target potential safety hazard situation.

[0146] One possible abnormal risk detection and processing method is an algorithm based on threshold comparison. Still taking the hidden danger situation of battery cell voltage imbalance as an example, in addition to the calculation of the maximum voltage difference mentioned above, different levels of voltage difference thresholds are set for risk assessment. For example, the voltage difference is divided into three levels: mild risk (0.2-0.3 volts), moderate risk (0.3-0.5 volts) and high risk (greater than 0.5 volts). According to the maximum voltage difference of 0.4 volts obtained above, it can be judged that the energy storage battery module is in a moderate risk state. In this case, the abnormal risk detection process may include sending an alarm signal to the management center of the energy storage system, indicating that the battery module needs to be further inspected or the charging strategy needs to be adjusted to balance the battery voltage.

[0147] Another possible approach is an algorithm based on data trend analysis. For some potential safety hazards, the changing trend of the status mark data is more important than the current value. For example, for abnormal battery temperature hazards, the status mark data records the temperature change trend over a period of time (such as the past 10 minutes) in addition to the current battery temperature value (for example, 40°C). For example, the temperature rises at a rate of 0.5°C per minute. According to the pre-set temperature change rate threshold (for example, the normal temperature change rate threshold is 0.2°C per minute), it can be judged that the battery temperature has an abnormal rising trend and is in a high-risk state. The abnormal risk detection process at this time may include initiating emergency heat dissipation measures, such as increasing the speed of the cooling fan or adjusting the cooling power of the liquid cooling system.

[0148] In addition, abnormal risk detection and processing may also involve the evaluation of the overall performance of the energy storage system. When a potential safety hazard situation of a target is detected, it is necessary not only to consider the impact of the hazard on a single battery module, but also to evaluate its impact on the entire energy storage system. For example, in an energy storage system composed of multiple battery modules, if a battery module has a hidden danger of battery cell voltage imbalance, it may affect the charging and discharging efficiency and stability of the entire energy storage system. By analyzing the relationship between the status marking data and the overall performance indicators of the energy storage system, the degree of impact of the abnormal risk on the entire system can be determined. For example, the normal range of the charging and discharging efficiency of the entire energy storage system is between 90% and 95%. Due to the voltage imbalance hazard of a certain battery module, the charging and discharging efficiency of the system drops to 85%. According to the pre-set system performance degradation threshold (for example, 5%), it can be judged that the abnormal risk has a greater impact on the entire energy storage system, and more comprehensive measures need to be taken, such as temporarily stopping the operation of the battery module or adjusting the control strategy of the entire energy storage system.

[0149] With this design, by performing abnormal risk detection based on the state marking data when the state matching analysis results contain the target potential safety hazard situation, the risk source in the energy storage system can be located more accurately. For example, the risk level can be accurately judged based on the state marking data such as the battery cell voltage difference or temperature change rate combined with the corresponding algorithm. This helps to take targeted measures in a timely manner to avoid the deterioration of potential safety hazards, such as sending alarms according to the risk level, adjusting heat dissipation or charging strategies, etc. At the same time, considering the impact on the overall performance of the energy storage system, the stability and charging and discharging efficiency of the system can be guaranteed, the safety and reliability of the energy storage system can be improved, and the service life of the energy storage battery can be extended.

[0150] In another extended embodiment, the abnormal risk detection processing is performed on the energy storage system to which the energy storage battery module tag belongs based on the status mark data corresponding to the target potential safety hazard situation, including: performing transient risk fault analysis and periodic risk fault analysis on multiple status data segments in the status mark data, respectively, to obtain a transient risk fault analysis viewpoint set and a periodic risk fault analysis viewpoint set; performing a first noise optimization operation on the transient risk fault analysis viewpoint set through a first noise optimization strategy, to obtain a first status data chain including a transient risk fault point; performing a second noise optimization operation on the periodic risk fault analysis viewpoint set through a second noise optimization strategy, to obtain a second status data chain including a periodic risk fault point; based on the intersection of the first status data chain and the second status data chain, a status mark data chain matching the target fault point in the status mark data is obtained; the target fault point includes at least one of a transient risk fault point and a periodic risk fault point; and generating an abnormal risk detection report for the energy storage system based on the status mark data chain.

[0151] In the above-mentioned extended embodiment, the process of performing abnormal risk detection and processing on the energy storage system based on the status mark data corresponding to the target potential safety hazard situation includes multiple steps.

[0152] First, transient risk fault analysis and periodic risk fault analysis are performed on multiple state data segments in the state marking data. For example, if the state marking data contains the voltage value and current value of the battery at different times, it is divided into multiple state data segments at a certain time interval. Transient risk fault analysis aims to find signs of risk faults that suddenly appear in a short period of time. Assume that the state marking data is the sampling value of the battery voltage every 1 minute within 10 minutes, such as [3.2, 3.15, 3.1, 3.05, 3.0, 2.95, 2.9, 2.85, 2.8, 2.75]. If a simple difference algorithm is used to calculate the difference between two adjacent values, it is found that the difference suddenly increases from 3.0 to 2.95, which may be a manifestation of transient risk faults, thus forming a transient risk fault analysis viewpoint set.

[0153] Periodic risk fault analysis focuses on the regular fault characteristics presented by the data over a longer period. For example, the number of battery charge and discharge times and the corresponding voltage fluctuation range are counted in an hourly cycle. If it is found that the voltage fluctuation exceeds the normal range once every 5 hours (assuming that the normal fluctuation range is ±0.1V, and the fluctuation reaches 0.2V at a specific moment), this is the characteristic of periodic risk fault, and thus a set of periodic risk fault analysis viewpoints is formed.

[0154] Then, the first noise optimization operation is performed on the transient risk fault analysis viewpoint set through the first noise optimization strategy to obtain the first state data chain. The first noise optimization strategy may be based on a filtering algorithm, such as a median filter. For some abnormal points in the transient risk fault analysis viewpoint set caused by measurement errors, etc. (for example, a sudden jump in a certain instantaneous voltage value may be caused by measurement noise), median filtering can effectively remove these noise points. Assuming that the original transient risk fault analysis viewpoint set contains some discrete abnormal points, after median filtering, a smooth data chain is obtained, namely the first state data chain, which contains the optimized transient risk fault points.

[0155] Similarly, the second noise optimization strategy is used to perform a second noise optimization operation on the periodic risk fault analysis viewpoint set to obtain a second state data chain. For example, the second noise optimization strategy uses mean filtering to smooth the fluctuations in the periodic risk fault analysis viewpoint set caused by environmental interference and other factors, thereby obtaining a second state data chain containing periodic risk fault points.

[0156] Next, the first state data chain and the second state data chain are crossed. By comparing the fault point location, fault characteristics and other information in the two data chains, the common parts or interrelated parts are found to obtain the state mark data chain that matches the target fault point (at least one of the transient risk fault point and the periodic risk fault point) in the state mark data. For example, if a transient fault point in the first state data chain is close in time to a periodic fault point in the second state data chain and the fault characteristics are related (such as abnormal voltage fluctuations at the same time), then these related data points are integrated into the state mark data chain.

[0157] Finally, an abnormal risk detection report for the energy storage system is generated based on the status tag data chain. This report records in detail the relevant information of the target fault point obtained from the status tag data, including the time of the fault, the fault type (transient or periodic), the possible impact of the fault on the energy storage system, etc., providing an accurate basis for subsequent maintenance and management.

[0158] With this design, risk information in the state tag data can be fully mined through transient and periodic risk fault analysis, and the characteristics of different types of faults can be accurately located. The noise optimization strategy removes interference factors and improves the accuracy of fault judgment. The data link cross-operation integrates related fault points to make the detection results more relevant. The abnormal risk detection report finally generated provides precise guidance for the maintenance and management of the energy storage system, which helps to repair faults in a timely manner, ensure the stable operation of the system, extend the life of the energy storage system and improve safety.

[0159] Further, Figure 2 FIG. 2 is a schematic diagram of a state data identification system 200 provided by the present invention. Figure 2 The status data identification system 200 shown includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method of the present invention.

[0160] Alternatively, if Figure 2 As shown, the state data identification system 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method of the present invention.

[0161] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0162] Alternatively, if Figure 2 As shown, the state data identification system 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices.

[0163] Optionally, the state data identification system 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with the storage engine in each method of the present invention, which will not be described in detail here for the sake of brevity.

[0164] It should be understood that the processor of the present invention may be an integrated circuit chip having signal processing capabilities.

[0165] It is understood that the memory in the present invention can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0166] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0167] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0168] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0169] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which are within the protection of the present invention.

Claims

1. A state data identification method applied to an energy storage battery module, characterized in that: The method is implemented by a state data identification system, and the method comprises: Acquire power storage working state monitoring data including energy storage battery module tags and a plurality of safety situation decision information, wherein the plurality of safety situation decision information includes: first safety situation decision information indicating a stage safety and stability situation, and second safety situation decision information indicating each of the potential safety hazard situations in at least one potential safety hazard situation; Through the target state matching analysis algorithm, the unique hot knowledge embedding is performed on each of the security situation decision information to obtain the security situation decision knowledge vector of each of the security situation decision information, and the state monitoring vector mining is performed on the power storage working state monitoring data to obtain the IoT sensor state monitoring vector; Through the target state matching analysis algorithm, a state matching analysis is performed according to the IoT sensor state monitoring vector and each of the safety situation decision knowledge vectors to obtain a state matching analysis result of the power storage working state monitoring data; wherein the state matching analysis result includes the matching confidence weight of the energy storage battery module label corresponding to each safety situation, and the safety situation includes the stage safety and stability situation and the at least one potential safety hazard situation.

2. The method according to claim 1, characterized in that The method further comprises: Acquire the preceding working state monitoring data for debugging the original state matching analysis algorithm, the first monitoring discrimination information of the preceding working state monitoring data, and the second monitoring discrimination information of the preceding working state monitoring data under the target monitoring task; According to the first monitoring discrimination information and the second monitoring discrimination information, a state monitoring vector of the preceding working state monitoring data is subjected to task attention processing to obtain a state monitoring attention vector, and according to an original task attention weight, a state monitoring vector is subjected to task attention processing to obtain a state attention prediction vector; According to the difference between the state attention prediction vector and the state monitoring attention vector, the original task attention weight is optimized to obtain the target task attention weight; When debugging the original state matching analysis algorithm, task attention processing is performed on the state monitoring vector according to the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data, and the original state matching analysis algorithm is debugged according to the target state monitoring attention vector to obtain the target state matching analysis algorithm.

3. The method according to claim 2, characterized in that The step of performing task attention processing on the state monitoring vector of the preceding working state monitoring data according to the first monitoring discrimination information and the second monitoring discrimination information to obtain the state monitoring attention vector includes: Performing one-hot knowledge embedding on the first monitoring and discrimination information to obtain a first security situation decision knowledge vector, and performing one-hot knowledge embedding on the second monitoring and discrimination information to obtain a second security situation decision knowledge vector; Performing feature pooling processing on the state monitoring vector of the preceding working state monitoring data to obtain a state monitoring pooling vector; The second security situation decision knowledge vector and the first security situation decision knowledge vector are used to obtain a security situation decision offset, and the state monitoring attention vector is obtained based on the security situation decision offset and the state monitoring pooling vector.

4. The method according to claim 2, characterized in that The step of performing task attention processing on the state monitoring vector according to the original task attention weight to obtain a state attention prediction vector includes: Obtaining a first weighted monitoring feature based on the original task attention weight and the state monitoring vector; Perform feature pooling processing on the first weighted monitoring feature to obtain the state attention prediction vector.

5. The method according to claim 2, characterized in that The original task attention weight belongs to the self-attention network; the original task attention weight is optimized according to the difference between the state attention prediction vector and the state monitoring attention vector to obtain the target task attention weight, including: Obtaining a training error representation of the self-attention network; Determining an error variable represented by the training error according to a difference between the state attention prediction vector and the state monitoring attention vector; According to the error variable represented by the training error, optimizing the original task attention weight of the self-attention network to obtain the target task attention weight; Among them, the training error representation includes a first training error representation and a second training error representation; the determining the error variable of the training error representation based on the difference between the state attention prediction vector and the state monitoring attention vector includes: determining the error variable of the first training error representation based on the difference between the state attention prediction vector and the state monitoring attention vector; determining the error variable of the second training error representation based on the difference between the state attention prediction vector and the state monitoring vector; obtaining the error variable of the training error representation according to the error variable of the first training error representation and the error variable of the second training error representation.

6. The method according to claim 2, characterized in that The step of performing task attention processing on the state monitoring vector according to the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data includes: Determine whether the preceding working state monitoring data corresponding to the state monitoring vector belongs to a candidate preceding working state monitoring data set, wherein the candidate preceding working state monitoring data set includes at least one preceding working state monitoring data carrying an attention processing identifier, and the at least one preceding working state monitoring data carrying an attention processing identifier is determined from the preceding working state monitoring data set in which the preceding working state monitoring data is located according to a target sampling strategy; When the preceding working state monitoring data corresponding to the state monitoring vector belongs to a candidate preceding working state monitoring data set, task attention processing is performed on the state monitoring vector according to the target task attention weight to obtain a target state monitoring attention vector of the preceding working state monitoring data.

7. The method according to claim 2, characterized in that The step of performing task attention processing on the state monitoring vector according to the target task attention weight to obtain the target state monitoring attention vector of the preceding working state monitoring data includes: Using the target task attention weight and the state monitoring vector, a second weighted monitoring feature is obtained; The second weighted monitoring feature is subjected to feature pooling processing to obtain the target state monitoring attention vector of the previous working state monitoring data.

8. The method according to claim 2, characterized in that The method of debugging the original state matching analysis algorithm based on the target state monitoring attention vector to obtain the target state matching analysis algorithm includes: By using the original state matching analysis algorithm, a state matching analysis is performed according to the target state monitoring attention vector to obtain a state matching analysis result of the preceding working state monitoring data; According to the comparison viewpoint between the state matching analysis result and the prior training annotation of the preceding working state monitoring data, the original state matching analysis algorithm is debugged to obtain the target state matching analysis algorithm.

9. The method according to claim 8, characterized in that Before performing state matching analysis according to the target state monitoring attention vector by the original state matching analysis algorithm to obtain the state matching analysis result of the preceding working state monitoring data, the method further includes: Acquire a plurality of security situation decision information, wherein the plurality of security situation decision information includes: First security situation decision information indicating a stage security and stability situation, and second security situation decision information indicating each of the potential security risk situations in at least one potential security risk situation; The state matching analysis is performed according to the target state monitoring attention vector by the original state matching analysis algorithm to obtain the state matching analysis result of the preceding working state monitoring data, including: By using the original state matching analysis algorithm, one-hot knowledge embedding is performed on each of the security situation decision information to obtain a security situation decision knowledge vector for each of the security situation decision information; By means of the original state matching analysis algorithm, a state matching analysis is performed according to the target state monitoring attention vector and each of the safety situation decision knowledge vectors, so as to obtain a matching confidence weight of each safety situation corresponding to the preceding energy storage battery module label in the preceding working state monitoring data; wherein the safety situation includes the stage safety and stability situation and the at least one potential safety hazard situation, and the state matching analysis result of the preceding working state monitoring data includes: a matching confidence weight of each safety situation corresponding to the preceding energy storage battery module label; The state matching analysis is performed based on the target state monitoring attention vector and each of the security situation decision knowledge vectors to obtain the matching confidence weight of the preceding energy storage battery module label in the preceding working state monitoring data corresponding to each security situation, including: performing full connection mapping on the target state monitoring attention vector to obtain a fully connected monitoring mapping attention vector, and performing state monitoring upsampling on the fully connected monitoring mapping attention vector to obtain a target state monitoring vector; obtaining a global monitoring vector through the target state monitoring attention vector and the target state monitoring vector; performing state matching analysis based on the global monitoring vector and each of the security situation decision knowledge vectors to obtain the matching confidence weight of the preceding energy storage battery module label in the preceding working state monitoring data corresponding to each security situation; The method of performing full-connection mapping on the target state monitoring attention vector to obtain a full-connection monitoring mapping attention vector, and performing state monitoring upsampling on the full-connection monitoring mapping attention vector to obtain a target state monitoring vector comprises: performing full-connection mapping on the target state monitoring attention vector through a full-connection mapping module of the original state matching analysis algorithm to obtain a full-connection monitoring mapping attention vector, and performing state monitoring upsampling on the full-connection monitoring mapping attention vector to obtain a target state monitoring vector; the method of debugging the original state matching analysis algorithm based on a comparison viewpoint between the state matching analysis result and a priori training annotation of the preceding working state monitoring data to obtain a target state matching analysis algorithm comprises: optimizing the algorithm network configuration weight of the full-connection mapping module based on a comparison viewpoint between the state matching analysis result and a priori training annotation of the preceding working state monitoring data to debug the original state matching analysis algorithm to obtain a target state matching analysis algorithm.

10. A status data identification system, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images