Battery fault identification method and system

Abstract

The invention relates to a battery fault identification method and system, and relates to the field of battery faults, and the method comprises the steps: receiving a manually detected fault type after a preset test period, collecting battery feature information and time information, and forming multi-dimensional tensor data; initializing a dynamic weight matrix and an offset value based on a random number generation function; inputting the multi-dimensional tensor data into a deep learning model to calculate and obtain an output array; obtaining a fault probability array based on the output array; determining a predicted fault type based on the fault probability array; when the predicted fault type is consistent with the artificially detected fault type, continuing to execute the model identification method until the next test period; and when the predicted fault type is inconsistent with the artificially detected fault type, searching the fault type weight, and adjusting the dynamic weight matrix and the offset value until the predicted fault type is consistent with the artificially detected fault type. The method has the effect of continuously adjusting and optimizing fault identification according to the actual operation state of the battery.

Description

Technical Field

[0001] This invention relates to the field of battery faults, and in particular to a battery fault identification method and system. Background Art

[0002] With the development of new energy vehicles, the health status of their batteries plays a crucial role in vehicle performance and safety. In the existing technological system, battery fault detection mainly relies on two traditional methods: One method is the manual parameter testing method, where technicians use equipment such as multimeters and internal resistance testers to measure the basic parameters of each battery cell offline, such as voltage, internal resistance, and temperature, and then judge whether the parameters exceed the threshold according to industry standards.

[0003] The second method is machine-automated threshold analysis, which involves setting fixed parameter thresholds in the battery management system, collecting battery operation data in real time and comparing it with the thresholds, and triggering a fault alarm when the parameters exceed the thresholds.

[0004] In recent years, with the penetration of deep learning and big data technologies into the new energy field, some research has begun to attempt to overcome the aforementioned limitations through data-driven intelligent recognition methods. The core logic of such solutions is to collect multi-dimensional parameters throughout the battery's entire life cycle to construct multi-dimensional tensor data, and then assign dynamic weights to different parameters through neural networks.

[0005] Regarding the aforementioned technologies, although deep learning considers the impact of battery aging and environmental parameters and dynamically adjusts the weight values ​​of parameters to accurately identify fault types, the rapid pace of battery technology updates makes it impossible for neural network models to quickly adapt to the feature parameter distribution of new batteries, resulting in inaccurate subsequent fault type identification. Summary of the Invention

[0006] To achieve accurate identification of battery fault types, this invention provides a battery fault identification method and system.

[0007] In a first aspect, the present invention provides a battery fault identification method, which adopts the following technical solution: A battery fault identification method, comprising: Step 1: After a preset test cycle, receive manually detected fault types and collect battery characteristic information and time information to form multidimensional tensor data; Step 2: Initialize the dynamic weight matrix and bias values ​​based on the preset random number generation function; Step 3: Input multidimensional tensor data into a preset deep learning model to calculate the output array; Step 4: Obtain the fault probability array based on the output array and the preset activation function; Step 5: Determine the expected fault type based on the fault probability array and preset sample labels; Step 6: If the expected fault type and the manually detected fault type are consistent, continue to execute the preset model identification method until the next test cycle; Step 7: When the expected fault type and the manually detected fault type are inconsistent, find the fault category weight corresponding to the sample label, and adjust the dynamic weight matrix and bias value until the expected fault type and the manually detected fault type are consistent.

[0008] By employing the above technical solution, battery faults can be intelligently identified based on a deep learning model, and the accuracy of fault identification can be optimized by dynamically adjusting the weight matrix and bias values. After each test cycle, if the expected fault type is inconsistent with the manually detected fault type, this invention will automatically find the fault category weight corresponding to the sample label and adjust the model parameters accordingly to improve the accuracy of subsequent identification.

[0009] Optionally, methods for adjusting the dynamic weight matrix and bias values ​​include: Step 8: When no fault category weights exist, preset corresponding fault category weights for different fault types using a random number generation function; Step 9: Obtain the loss function value corresponding to the fault category based on the preset loss function, fault category weights, sample labels, and fault probability array; Step 10: When the loss function value corresponding to the fault category exceeds the preset standard threshold of the loss function, calculate the gradient of the dynamic weight matrix based on the output array, the fault probability array, and the dynamic weight matrix, and calculate the gradient of the bias value based on the output array and the fault probability array. Step 11: Calculate the corrected dynamic weight matrix based on the gradient of the dynamic weight matrix, the loss function, and the preset learning rate, and output it as the dynamic weight matrix. Calculate the corrected bias value based on the gradient of the bias value, the loss function, and the learning rate, and output it as the bias value. Step 12: Repeat steps 2 to 11 until the loss function values ​​corresponding to all fault categories fall into the standard threshold of the loss function, or the expected fault type is consistent with the manually detected fault type.

[0010] By adopting the above technical solution, the parameters of the deep learning model can be further optimized, improving the accuracy and stability of fault identification. In the absence of fault category weights, fault category weights are preset using a random number generation function. When the loss function value corresponding to a fault category exceeds the preset standard threshold, the gradient of the dynamic weight matrix and the gradient of the bias value are calculated. The model parameters are then adjusted based on these gradients and the preset learning rate until the loss function values ​​for all fault categories meet the requirements, or the predicted fault type matches the manually detected fault type. This achieves continuous optimization and improvement of the deep learning model.

[0011] Optionally, it also includes a preset information compensation method when multidimensional tensor data cannot be collected, the method including: Step 13: Determine if there is unidentifiable data in the multidimensional tensor data; Step 14: Determine the data loss time based on the time information in the lost multidimensional tensor data; Step 15: Determine the adjacent time range based on the data loss time; Step 16: Find multidimensional tensor data in adjacent time ranges to form multidimensional tensor data intervals; Step 17: Determine adjacent data based on multidimensional tensor data intervals; Step 18: Assign a preset weight decay coefficient based on the time information in the neighboring data, and calculate the reference neighboring data; Step 19: Calculate the reference multidimensional tensor data based on the reference neighbor data, and output it as multidimensional tensor data.

[0012] By adopting the above technical solution, when unidentifiable data occurs during the acquisition of multidimensional tensor data, the adjacent valid data intervals are located by utilizing temporal correlation, and the weight coefficients are dynamically allocated to generate compensation data in combination with the time decay characteristics. This effectively avoids the model input gap problem caused by data loss, ensures that the deep learning model can still maintain stable fault identification capability in scenarios with incomplete data, and improves the robustness and reliability of the method in practical engineering applications.

[0013] Optionally, a fault warning method may also be included, which includes: Step 20: Determine the warning threshold range based on the output array and the fault probability array; Step 21: Determine the risky multidimensional tensor data when the multidimensional tensor data exceeds the warning threshold range; Step 22: Determine the risk multidimensional tensor data interval when all multidimensional tensor data are risk multidimensional tensor data within a preset continuous time period; Step 23: Determine the minimum and maximum reference data based on the risk multidimensional tensor data and the risk multidimensional tensor data interval; Step 24: Obtain the battery time-varying data by subtracting the minimum and maximum reference data; Step 25: Determine the battery data mutation rate when the battery time-varying data exceeds the preset battery time-varying data threshold; Step 26: Determine the expected fault based on the battery data mutation rate and the corresponding fault data features in the preset fault data feature library, and output the corresponding expected fault risk signal.

[0014] By adopting the above technical solution, and constructing a dynamic early warning threshold range based on the output array and the fault probability array, combined with a continuous anomaly monitoring mechanism using multidimensional tensor data, early detection of battery fault risks is achieved. By calculating battery time-varying data and mutation rates, fault warning is upgraded from simple threshold judgment to in-depth analysis of data trend changes. Combined with a pre-set fault data feature library, accurate fault location prediction and early warning are achieved.

[0015] Optionally, it also includes a method for modifying the fault data feature library, the method comprising: Step 27: Determine the standard deviation of battery parameters based on the dynamic weight matrix and multidimensional tensor data; Step 28: Calculate the parameter correlation degree based on the standard deviation of battery parameters and the dynamic weight matrix; Step 29: Sort and filter based on parameter correlation to obtain key feature parameters; Step 30: Determine the corrected fault data features based on key characteristic parameters and expected fault types; Step 31: Re-determine the fault data feature library based on the corrected fault data features.

[0016] By employing the aforementioned technical solution, and through the coupled calculation of dynamic weight matrices and multidimensional tensor data, the dispersion of battery parameters is accurately quantified to obtain the standard deviation of battery parameters. This is then combined with a parameter correlation evaluation model constructed using the dynamic weight matrix, enabling in-depth mining of the intrinsic correlations between multidimensional feature parameters. By ranking the parameter correlations, the key feature parameters that contribute the most to fault identification are selected, ensuring the targetedness and effectiveness of subsequent fault feature corrections. Based on the key feature parameters and the actual predicted fault types, feature matching and optimization are performed to generate dynamically adaptive corrected fault data features, thereby completing the iterative update of the fault data feature library. This effectively improves the fault detection accuracy of the fault identification model during the evolution of battery technology.

[0017] Optional, also includes: Step 32: In response to a manual maintenance signal, identify the faulty battery based on the expected fault type; Step 33: When a faulty battery exists, a feature 3D projection is formed based on the faulty battery. The feature 3D projection is a 3D projection of the battery pack, and the projection light of the faulty battery flickers. Step 34: Determine the projected light color based on the expected fault type; Step 35: Determine the location of the faulty battery based on the color of the projected light, and perform preset maintenance operations based on the location of the faulty battery.

[0018] By adopting the above technical solution, the faulty battery location process is triggered in response to manual maintenance signals, and the faulty battery is accurately selected based on the expected fault type. A visual model of the battery pack is constructed using 3D projection technology, utilizing the flickering characteristics of light to highlight the spatial location of the faulty battery. Differentiated light color codes are matched according to different fault types, forming an intuitive fault location marker. Maintenance personnel can quickly locate the faulty battery based on the color and flickering status of the projected light, significantly reducing the time cost of manual troubleshooting. At the same time, standardized maintenance operation guidelines ensure the standardization and accuracy of fault handling, effectively improving the efficiency and safety of battery pack maintenance.

[0019] Optionally, it also includes a method for adjusting the flicker frequency of the light, the method comprising: Step 36: Establish risk levels based on fault data features in the fault feature library; Step 37: Calculate the comprehensive risk index based on the failure probability array and battery mutation rate; Step 38: Match the comprehensive risk index with the risk level to determine the current risk level; Step 39: Determine the light flicker frequency based on the current risk level; Step 40: Control the lighting of the faulty battery in the 3D projection to flicker according to the lighting flicker frequency.

[0020] By employing the above technical solution, a comprehensive risk index is calculated through a weighted fusion of the fault probability array and the battery mutation rate, enabling a quantitative assessment of battery fault risk. Combined with a preset risk level classification standard, the comprehensive risk index is mapped to a specific risk level, and then the corresponding light flicker frequency is dynamically matched. When the comprehensive risk index is higher, the light flicker frequency increases accordingly, so that the intensity of the visual signal intuitively reflects the urgency of the fault, helping maintenance personnel quickly determine the fault priority and improve fault response efficiency.

[0021] Optionally, it also includes a method for lightweight model learning after battery failure repair, the method comprising: Step 41: When a faulty battery exists, extract the multidimensional tensor data within the preset reliable time range before fault repair, as well as the fault type, fault probability array, and comprehensive risk index corresponding to the multidimensional tensor data, and form historical fault data; Step 42: Classify and store historical fault data according to fault type, occurrence time, and battery number to establish a historical feature database; Step 43: In response to the battery repair completion signal, the forgetting time is obtained by subtracting the time information in the historical fault data from the current time, and a forgetting curve function is constructed based on the forgetting time and the historical fault data; Step 44: Calculate the weight decay coefficient corresponding to the historical fault data based on the forgetting curve function; Step 45: Multiply the fault category weights corresponding to the multidimensional tensor data in the historical fault data with the weight decay coefficient to obtain the current fault category weights; Step 46: Execute steps 9 to 12 based on the current fault category weight; Step 47: In response to the battery repair completion signal and when the expected fault type and the manually detected fault type are inconsistent, the corresponding fault category weight in the historical feature library is searched based on the manually detected fault type and used as the expected fault category weight; Step 48: Perform steps 9 through 12 based on the expected fault category weights.

[0022] By adopting the above technical solution, and through the effective management and utilization of historical fault data after battery fault repair, lightweight learning and updating of model knowledge are achieved. Fault category weights are quickly determined, and the model is executed based on these weights to accelerate iterative updates.

[0023] Optional, also includes: Step 49: When a faulty battery exists, search the historical fault data corresponding to the faulty battery in the historical feature database; Step 50: Determine the fault interval time based on the occurrence time of historical fault data and the current time; Step 51: A battery is defined as frequently failing when the fault interval time falls within a preset range of frequent fault times; Step 52: Send a preset alarm signal to a preset terminal device based on the frequent battery failures.

[0024] By employing the above technical solution and conducting in-depth analysis of historical fault data, the system can identify batteries that frequently fail. When a battery's failure interval is short and falls within a preset frequent failure time range, the system will mark it as a frequently failing battery. To ensure the safe operation of the battery pack, the system will immediately send an alarm signal to preset terminal devices, reminding relevant personnel to focus on monitoring and handling the frequently failing battery. This function helps to promptly identify and address potential safety hazards, improving the overall reliability and safety of the battery pack.

[0025] Secondly, the present invention provides a battery fault identification system, which adopts the following technical solution: A battery fault identification system, comprising: The acquisition module is used to acquire battery characteristic information and time information; A memory used to store a program that implements a battery fault identification method; The processor loads and executes programs from memory.

[0026] By adopting the above technical solution, the system uses an acquisition module to collect real-time characteristic and temporal information of the battery, which is then used as input data for a deep learning model. The memory stores the control program for the battery fault identification method. By loading and executing this program, the processor can automatically complete a series of operations, including constructing multidimensional tensor data, intelligently predicting fault types, dynamically adjusting fault category weights, and issuing fault warnings. The entire system has a compact structure and a high degree of automation, enabling efficient and accurate identification of battery faults and providing strong technical support for battery safety management.

[0027] In summary, the present invention has at least one of the following beneficial technical effects: Through model training and manual verification, the parameter configuration of the deep learning model is continuously optimized, enabling the model to exhibit higher recognition accuracy and stronger generalization ability when faced with different types of battery faults. After the battery is repaired, the model extracts reliable historical fault data, constructs a forgetting curve to decay weights, and focuses on updating weights based on effective historical information. This eliminates the need for retraining on all data and significantly reduces the consumption of computing resources. By visually marking the location of faulty batteries using 3D projection, matching the expected fault type with a specific lighting color, and dynamically adjusting the flashing frequency based on the comprehensive risk index, maintenance personnel can quickly locate high-risk faulty batteries, avoiding blind disassembly and troubleshooting, and shortening the maintenance cycle. Attached Figure Description

[0028] Figure 1 This is a flowchart of a battery fault identification method according to an embodiment of this application; Figure 2 This is a flowchart of the method for adjusting the dynamic weight matrix and bias value in the embodiments of this application; Figure 3 This is a flowchart of a method for modifying a fault data feature library, which is also included in the embodiments of this application; Figure 4 This is a flowchart of a method for determining the location of a faulty battery based on feature 3D projection in an embodiment of this application; Figure 5 This application also includes a flowchart of a method for lightweight model learning after repairing battery faults in the embodiments of this application. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0030] This invention discloses a battery fault identification method and system. (Refer to...) Figure 1 A battery fault identification method includes: Step 1: After a preset test cycle, receive manually detected fault types and collect battery characteristic information and time information to form multidimensional tensor data.

[0031] The testing cycle refers to the fixed time interval for periodic manual inspection of the battery pack. Manually detected fault types refer to battery fault types determined through manual inspection. These manually detected fault types are battery fault types obtained from analysis by experienced experts and serve as a benchmark for training and validating deep learning models.

[0032] Battery characteristic information refers to key parameters including but not limited to battery voltage, current, temperature, and internal resistance, which are acquired through real-time monitoring using sensors. Time information records the specific moment the battery characteristic information is acquired; the system records the time when the battery characteristic information is acquired. Multidimensional tensor data is composed of battery characteristic information and time information, forming a dataset containing multiple dimensions. Here, we use a two-dimensional tensor data formed by battery internal resistance information and time information as an example. Let the battery internal resistance information be r, and the time information be t; the two-dimensional tensor data is [r, t].

[0033] Step 2: Initialize the dynamic weight matrix and bias values ​​based on the preset random number generation function.

[0034] A random number generation function is used to generate random numerical values. This function can set specific intervals to limit the maximum and minimum values ​​of the random numbers. The dynamic weight matrix and bias values ​​are training parameters in a deep learning model established for preset fault types. Fault types refer to common battery fault types, which are preset manually.

[0035] Step 3: Input multidimensional tensor data into a preset deep learning model to calculate the output array.

[0036] Deep learning models are models built on multi-layer neural network structures and are specifically designed to handle the mapping relationship between multi-dimensional tensor data and fault types.

[0037] For ease of understanding, the two-dimensional tensor data [r, t] mentioned above is used as a reference. The dynamic weight matrix is ​​represented by W. Since the input data is two-dimensional tensor data, the number of rows in the dynamic weight array is 2, and the number of columns is set by the preset sample labels. The sample labels refer to the number of fault types that the deep learning model needs to predict. The deep learning model is specifically divided into an input layer, a hidden layer, and an output layer. The input layer receives the two-dimensional tensor data [r, t]. The hidden layer performs a weighted integration of the input data through a linear calculation of [r, t] × W + b. The output layer outputs the linear calculation result of [r, t] × W + b. The number of results obtained through linear calculation is consistent with the number of fault types in the sample labels. Here, the linear calculation results are y1, y2, ..., yn. The linear calculation results are converted into an array form [y] = [y1, y2, ..., yn] for output. Each linear calculation result corresponds to a different fault type.

[0038] Step 4: Obtain the fault probability array based on the output array and the preset activation function.

[0039] The output array refers to the array of linear calculation results output by the deep learning model, represented as [y] = [y1, y2, ..., yn]. The activation function is a function used to introduce nonlinear factors, transforming the linear calculation results into probability values. The fault probability array is obtained by performing a nonlinear transformation on the output array using the activation function, where each element represents the probability of occurrence of the corresponding fault type. Specifically, the activation function here refers to the sigmoid function, which maps the linear calculation results [y1, y2, ..., yn] in the output array [y] to probability values ​​between 0 and 1, forming the fault probability array [ŷ] = [ŷ1, ŷ2, ..., ŷn].

[0040] Step 5: Determine the expected fault type based on the fault probability array and preset sample labels.

[0041] The sample labels have been introduced above and will not be repeated here. The predicted fault type refers to the predicted fault obtained through deep learning model analysis; specifically, it refers to the fault type corresponding to the maximum probability value in the fault probability array.

[0042] Step 6: If the expected fault type and the manually detected fault type are consistent, continue to execute the preset model recognition method until the next test cycle.

[0043] Model-based identification methods refer to methods that use deep learning models to determine the current battery fault type. When the predicted fault type matches the manually detected fault type, it indicates that the deep learning model's prediction results have high accuracy. In this case, there is no need to make large-scale adjustments to the model, and the current model-based identification method can continue to be used for subsequent battery fault identification tasks.

[0044] Step 7: When the expected fault type and the manually detected fault type are inconsistent, find the fault category weight corresponding to the sample label, and adjust the dynamic weight matrix and bias value until the expected fault type and the manually detected fault type are consistent.

[0045] Fault category weights refer to the weight values ​​assigned to different fault types, reflecting the importance of each fault type in the model's predictions. Fault category weights are independent weights, distinct from the dynamic weight matrix. By adjusting these weights, the model can focus more on fault types prone to bias in subsequent predictions, thereby improving overall prediction accuracy.

[0046] When the predicted fault type and the manually detected fault type are inconsistent, it indicates that there is a deviation in the prediction results of the deep learning model. In this case, it is necessary to find the fault category weights corresponding to the sample labels in order to make targeted adjustments to the model until the predicted fault type is consistent with the manually detected fault type.

[0047] Reference Figure 2 Methods for adjusting the dynamic weight matrix and bias values ​​include: Step 8: When there are no fault category weights, preset the corresponding fault category weights for different fault types using a random number generation function.

[0048] If no fault category weight exists, it means that the fault category weight has not been initialized. Therefore, a value should be assigned to it using a random number generation function.

[0049] Step 9: Obtain the loss function value corresponding to the fault category based on the preset loss function, fault category weights, sample labels, and fault probability array.

[0050] The loss function is a function used to evaluate the difference between the prediction results and the actual results of a deep learning model. The loss function value corresponding to the fault category is the loss function value calculated based on a specific fault category. The specific calculation method is Lᵢ=-[wᵢ×labelᵢ×log(ŷᵢ)+(1-wᵢ)×(1-labelᵢ)×log(1-ŷᵢ)], where label refers to the sample label, Lᵢ refers to the loss function value corresponding to the i-th fault category, labelᵢ refers to the probability of the i-th fault category obtained by the sample label through the fault probability array, ŷᵢ is the predicted probability of the i-th fault category, and wᵢ is the fault category weight corresponding to the i-th fault category.

[0051] Step 10: When the loss function value corresponding to the fault category exceeds the preset standard threshold of the loss function, calculate the gradient of the loss function with respect to the dynamic weight matrix based on the output array, the fault probability array, and the dynamic weight matrix, and define it as the gradient of the dynamic weight matrix. Calculate the gradient of the loss function with respect to the bias value based on the output array and the fault probability array, and define it as the gradient of the bias value.

[0052] The standard threshold of the loss function refers to the critical value used to evaluate whether the loss function value corresponding to the fault category is abnormal. The gradient of the dynamic weight matrix refers to the rate of change of the loss function relative to the dynamic weight matrix, reflecting the degree of influence of the dynamic weight matrix on the loss function value. The dynamic weight matrix is ​​specifically calculated as ∂Lᵢ / ∂Wᵢ=[∂Lᵢ / ∂ŷᵢ×∂ŷᵢ / ∂yᵢ]×[r,t]ᵀ, where ∂Lᵢ / ∂Wᵢ refers to the gradient value of the dynamic weight matrix corresponding to the i-th fault category, and [r,t]ᵀ is the transpose of [r,t]. The two-dimensional tensor data [r,t] is a 1-row, 2-column matrix, and [r,t]ᵀ is a 2-row, 1-column matrix. The bias gradient refers to the rate of change of the loss function relative to the bias value, and also reflects the degree of influence of the bias value on the loss function value. The bias gradient is specifically calculated as: ∂Lᵢ / ∂bᵢ=∂Lᵢ / ∂ŷᵢ×∂ŷᵢ / ∂yᵢ×1 (since yᵢ=[t,r]×Wᵢ+bᵢ, the derivative of bᵢ is 1). Here, ∂Lᵢ / ∂bᵢ refers to the bias gradient value corresponding to the i-th fault category.

[0053] When the loss function value corresponding to the fault category exceeds the standard threshold, it indicates that there is a significant difference between the prediction result of the deep learning model and the actual result. At this time, the model needs to be trained, specifically by modifying variables such as the dynamic weight matrix, fault category weights, and bias values.

[0054] Step 11: Calculate the corrected dynamic weight matrix based on the gradient of the dynamic weight matrix, the loss function, and the preset learning rate, and output it as the dynamic weight matrix. Calculate the corrected bias value based on the gradient of the bias value, the loss function, and the learning rate, and output it as the bias value.

[0055] The learning rate is a parameter in a deep learning model used to control model parameters such as the dynamic weight matrix and bias value. It adjusts these parameters in each iteration, specifically when the loss function value exceeds a threshold. The corrected dynamic weight matrix is ​​the dynamic weight matrix after adjustment using the learning rate, calculated as Wᵢ = Wᵢ - η × ∂Lᵢ / ∂Wᵢ. The corrected bias value is the bias value after adjustment using the learning rate, calculated as bᵢ = bᵢ - η × ∂Lᵢ / ∂bᵢ, where η represents the learning rate.

[0056] Step 12: Repeat steps 2 to 11 until the loss function values ​​corresponding to all fault categories fall into the standard threshold of the loss function, or the expected fault type is consistent with the manually detected fault type.

[0057] When the corrected dynamic weights and corrected bias values ​​are obtained, it indicates that the deep learning model has undergone one round of optimization iteration. To ensure that the model can continuously improve its prediction accuracy, the system will continuously execute the operations from step 2 to step 11 until the loss function values ​​corresponding to all fault categories are lower than the preset standard threshold of the loss function or the predicted fault type is consistent with the manually detected fault type.

[0058] This also includes a pre-defined information compensation method for situations where multidimensional tensor data cannot be collected. This method includes: Step 13: Determine if there is unidentifiable data in the multidimensional tensor data.

[0059] Loss of multidimensional tensor data refers to the situation where, during the acquisition of battery characteristic information and time information, some data may not be properly identified or acquired due to factors such as sensor failure, data transmission errors, or environmental interference.

[0060] Step 14: Determine the time of data loss based on the time information in the lost multidimensional tensor data.

[0061] Data loss time refers to the point in time when multidimensional tensor data is lost. It is determined by the time information recorded by the system. If the time information in the lost multidimensional tensor data cannot be parsed, the estimated loss time is obtained by analyzing adjacent data of the lost multidimensional tensor data. The specific format of multidimensional tensor data is array-based. Therefore, the corresponding parsing format is set according to the data collected. When the format of the time information is garbled or when an entire multidimensional tensor data set is garbled, the above method can be used to determine the data loss time. The array format allows calculation of the length of unrecognizable data, thus inferring the number of lost multidimensional tensor data. Here, the array uses the double data type, with each data point corresponding to 8 bytes. The length of the multidimensional tensor data can be obtained based on its dimensions and the 8-byte size.

[0062] Step 15: Determine the adjacent time range based on the data loss time.

[0063] The adjacent time range refers to a period of time adjacent to the time of data loss, and this range is preset by humans.

[0064] Step 16: Find multidimensional tensor data in adjacent time ranges to form multidimensional tensor data intervals.

[0065] A multidimensional tensor data interval refers to an interval formed by multiple multidimensional tensor data. Specifically, it means that within an adjacent time range, other multidimensional tensor data that are temporally adjacent to the lost multidimensional tensor data can be found, and these data form a data interval.

[0066] Step 17: Determine adjacent data based on the multidimensional tensor data interval.

[0067] Adjacent data refers to the data that is closest in time to the lost multidimensional tensor data within the multidimensional tensor data interval.

[0068] Step 18: Assign a preset weight decay coefficient based on the time information in the neighboring data, and calculate the reference neighboring data.

[0069] The weight decay coefficient is a weighting factor specifically used to adjust the impact of adjacent data on the compensation of missing multidimensional tensor data based on temporal information. Reference adjacent data refers to the adjacent data after adjustment using the weight decay coefficient. Since there is a certain time interval between adjacent data and the missing multidimensional tensor data, their data values ​​may deviate. To prevent sudden data changes and more accurately compensate for missing multidimensional tensor data, different weight decay coefficients need to be assigned to adjacent data based on the length of the time interval. The weight decay coefficient is inversely proportional to the length of the time interval; that is, the larger the time interval, the smaller the weight decay coefficient, indicating a smaller impact of adjacent data on the compensation of missing multidimensional tensor data.

[0070] Step 19: Calculate the reference multidimensional tensor data based on the reference neighbor data, and output it as multidimensional tensor data.

[0071] Reference multidimensional tensor data refers to estimated data used to replace lost multidimensional tensor data, calculated by averaging neighboring reference data. When lost multidimensional tensor data occurs, the system does not directly ignore this data, but instead uses the aforementioned information compensation method to estimate the lost data using valid data within adjacent time ranges. This also includes a fault early warning method, which includes: Step 20: Determine the warning threshold range based on the output array and the fault probability array.

[0072] The warning threshold range refers to the numerical range used to determine whether the probability of battery failure has reached the warning standard.

[0073] Step 21: Determine the risky multidimensional tensor data when the multidimensional tensor data exceeds the warning threshold range.

[0074] Risk multidimensional tensor data refers to multidimensional tensor data with a high probability of failure, specifically multidimensional tensor data that has exceeded the warning threshold range. When multidimensional tensor data exceeds the warning threshold range, it indicates a high probability of battery failure, hence it is classified as risk multidimensional tensor data.

[0075] Step 22: Determine the risk multidimensional tensor data interval when all multidimensional tensor data are risk multidimensional tensor data within a preset continuous time period.

[0076] Continuous time refers to a continuously defined period of time. The risk multidimensional tensor data interval refers to the interval within this continuous time period where all multidimensional tensor data are risk multidimensional tensor data. This avoids misjudgments caused by short-term jumps in data.

[0077] Step 23: Determine the minimum and maximum reference data based on the risk multidimensional tensor data and the risk multidimensional tensor data interval.

[0078] The minimum reference data refers to the data with the lowest failure probability among all data within the risk multidimensional tensor data range. The maximum reference data, on the other hand, refers to the data with the highest failure probability among all data within the risk multidimensional tensor data range.

[0079] Step 24: Obtain the battery time-varying data by subtracting the minimum reference data and the maximum reference data.

[0080] Battery time-varying data refers to the difference between the minimum reference data and the maximum reference data, which reflects the magnitude of change in battery failure probability within the risk multidimensional tensor data range.

[0081] Step 25: Determine the battery data mutation rate when the battery time-varying data exceeds the preset battery time-varying data threshold.

[0082] The battery time-varying data threshold is a critical value used to determine whether the probability of battery failure will change abruptly. The battery data mutation rate is the rate at which the probability of battery failure changes abruptly, calculated by dividing the battery time-varying data by the time difference between the minimum and maximum reference data. When the battery time-varying data exceeds this threshold, it indicates that the battery data has changed significantly in a short period, which usually suggests a potential battery failure or performance problem; therefore, the battery data mutation rate is calculated at this point.

[0083] Step 26: Determine the expected fault based on the battery data mutation rate and the corresponding fault data features in the preset fault data feature library, and output the corresponding expected fault risk signal.

[0084] The fault data feature library is a database storing fault data features corresponding to various battery faults. These features are pre-initialized and then improved based on historical fault data. Historical fault data refers to battery data from batteries that have already experienced faults. Predicted faults refer to possible fault types obtained by matching the battery data mutation rate with the fault data feature library. A predicted risk signal is a signal used to indicate an impending risk to the battery. When the battery data mutation rate matches a fault feature in the fault data feature library, it indicates that the battery is likely to have experienced that fault, and the system will output the corresponding predicted fault risk signal.

[0085] Reference Figure 3 It also includes a method for correcting the fault data feature library, the method comprising: Step 27: Determine the standard deviation of battery parameters based on the dynamic weight matrix and multidimensional tensor data.

[0086] Battery parameter standard deviation is a statistic used to quantify the dispersion of each feature parameter in multidimensional tensor data. Specifically, it refers to the average deviation of the i-th parameter in the multidimensional tensor data over all values ​​taken during training. The larger the standard deviation of the battery parameter, the wider the fluctuation range of the parameter; the smaller the standard deviation of the battery parameter, the more concentrated the value of the parameter and the higher its stability.

[0087] Step 28: Calculate the parameter correlation degree based on the standard deviation of battery parameters and the dynamic weight matrix.

[0088] Parameter correlation degree refers to an indicator used to quantify the degree of correlation between feature parameters in multidimensional tensor data. Because there is often a certain correlation between feature parameters during battery fault identification, calculating the parameter correlation degree can reveal the intrinsic relationship between these parameters, thus providing a more accurate basis for fault identification. The calculation of parameter correlation degree considers the standard deviation of battery parameters and the dynamic weight matrix, reflecting the mutual influence between parameters and their importance in fault identification.

[0089] Step 29: Sort and filter based on parameter correlation to obtain key feature parameters.

[0090] Key feature parameters refer to the feature parameters that play a leading role in the battery fault identification process, and are obtained by sorting the parameters in descending order of their correlation.

[0091] Step 30: Determine the corrected fault data features based on key feature parameters and expected fault types.

[0092] Corrected fault data features refer to the fault data features updated in the fault data feature library. When there is a clear correspondence between key feature parameters and expected fault types, it indicates that these feature parameters have high sensitivity in identifying the fault type. Therefore, they can be used as corrected fault data features to update the fault data feature library.

[0093] Step 31: Re-determine the fault data feature library based on the corrected fault data features.

[0094] Reference Figure 4 It also includes: Step 32: In response to a manual maintenance signal, identify the faulty battery by predicting the fault type.

[0095] A manual maintenance signal is a signal used to confirm the current operating status of the system. A faulty battery is a battery that has malfunctioned. When the system receives a manual maintenance signal, it indicates that a faulty battery is present. By predicting the type of fault, the type of battery fault can be determined so that personnel can quickly locate and take repair measures.

[0096] Step 33: When a faulty battery exists, a feature 3D projection is formed based on the faulty battery. The feature 3D projection is a 3D projection of the battery pack, and the projection light of the faulty battery flickers.

[0097] Feature-based 3D projection refers to the projection of a digital 3D model corresponding to the physical spatial location of the battery pack, constructed using 3D modeling technology. This model is based on the actual physical structure of the battery pack, such as its spatial arrangement, the number and location of cells, and the layout of connecting harnesses. Projection illumination flicker refers to the flickering of light in a specific projection area. Specifically, in feature-based 3D projection, the projection area of ​​the faulty battery will flicker at a specific frequency and brightness. Feature-based 3D projection not only visually presents the three-dimensional structure of the entire battery pack but also highlights identified faulty batteries through visually indicated flickering light, allowing staff to quickly locate the spatial position of the faulty cells without disassembling the battery pack.

[0098] The coordinate system, cell arrangement, and outer casing contour of the feature 3D projection are all strictly matched to the actual battery pack where the faulty battery is located. For example, a new energy vehicle power battery pack contains 200 cells arranged in 5 layers and 40 columns. The cell distribution of the 5 layers and 40 columns will also be displayed synchronously in the projection, with each projected cell corresponding one-to-one with the actual cell.

[0099] Step 34: Determine the projected light color based on the expected fault type.

[0100] Projected lighting color refers to the lighting color preset by humans to identify different types of faults. By assigning specific colors to different types of expected faults, staff can quickly identify the fault type by color.

[0101] Step 35: Determine the location of the faulty battery based on the color of the projected light, and perform preset maintenance operations based on the location of the faulty battery.

[0102] The location of a faulty battery refers to its precise position in the feature's 3D projection. Maintenance operations refer to the actions performed to repair or replace the faulty battery.

[0103] This also includes a method for adjusting the flicker frequency of the light, the method comprising: Step 36: Establish risk levels based on fault data features in the fault feature library.

[0104] Risk level refers to a grade used to quantify the severity of a battery failure. Specifically, risk level refers to the severity of the failure. The failure data feature database stores data features of various failures, including failure severity level labels. These labels reflect the degree of severity of the failure, and the risk level is determined through these labels.

[0105] Step 37: Calculate the comprehensive risk index based on the failure probability array and battery mutation rate.

[0106] The comprehensive risk index is an index derived by comprehensively considering both the probability of battery failure and the urgency of its development. The failure probability array itself quantifies the likelihood of a battery belonging to a certain type of failure; this is the probability of battery failure occurring in the calculation of the comprehensive risk index. The battery mutation rate reflects the magnitude and speed of abnormal changes in key battery parameters within a short period of time; this is the urgency of the failure's development in the calculation of the comprehensive risk index. The specific calculation method uses the average value method, which will not be elaborated further here.

[0107] Step 38: Match the comprehensive risk index with the risk level to determine the current risk level.

[0108] The current risk level refers to the current risk status of the battery, which is obtained by matching the comprehensive risk index with the preset risk level.

[0109] Step 39: Determine the light flicker frequency based on the current risk level.

[0110] The illumination flicker frequency refers to the frequency of illumination flickering in the projected area of ​​the faulty battery in a characteristic 3D projection. The illumination flicker frequency is directly proportional to the risk level; that is, the higher the risk level, the faster the illumination flicker frequency.

[0111] Step 40: Control the lighting of the faulty battery in the 3D projection to flicker according to the lighting flicker frequency.

[0112] Reference Figure 5 It also includes a method for lightweight model learning after battery failure repair, which includes: Step 41: When a faulty battery exists, extract the multidimensional tensor data within the preset reliable time range before fault repair, as well as the fault type, fault probability array, and comprehensive risk index corresponding to the multidimensional tensor data, and form historical fault data.

[0113] The reliable time range refers to a pre-defined time range, specifically the time range of data collected during normal battery operation for subsequent model training. Historical fault data, as explained above, refers to battery data that has already experienced faults. Here, historical fault data refers to the comprehensive set of data directly related to the current fault, collected within the reliable time range from the occurrence of the current fault to its repair completion. It does not refer to data from all batteries that have experienced faults, but rather focuses on valid data within the current fault cycle, extracting features such as fault type, fault probability array, and comprehensive risk index.

[0114] Step 42: Classify and store historical fault data according to fault type, occurrence time, and battery number to establish a historical feature database.

[0115] The historical feature database refers to a database that stores historical fault data. This database is classified and stored according to key information such as fault type, occurrence time, and battery number.

[0116] Step 43: In response to the battery repair completion signal, the forgetting time is obtained by subtracting the time information in the historical fault data from the current time, and a forgetting curve function is constructed based on the forgetting time and the historical fault data.

[0117] The battery repair completion signal is a signal used to confirm the battery repair status and complete data updates. Current time refers to the current system time. Forgotten time refers to the difference between the time information in historical fault data and the current time. The forgetting curve function is a mathematical function used to describe the degree of forgetting of historical fault data over time; specifically, it quantifies the importance of historical fault data at different time points by mapping the forgetting time (independent variable) to data weights (dependent variable).

[0118] Step 44: Calculate the weight decay coefficient corresponding to the historical fault data based on the forgetting curve function.

[0119] The weight decay coefficient is a coefficient used to adjust the impact of historical fault data on model training according to the forgetting curve function. As time progresses, the importance of historical fault data gradually decreases; therefore, a weight decay coefficient is needed to reduce its impact on model training, ensuring the model can more accurately reflect the current battery state. The weight decay coefficient is directly proportional to the forgetting time; that is, the longer the forgetting time, the smaller the weight decay coefficient. Since the data after battery repair differs significantly from that before repair, affecting the learning speed of subsequent dynamic weight matrices, bias values, and fault category weights, it is necessary to select data from before battery repair to accelerate model learning efficiency. Furthermore, during fault type identification, the fault category weight corresponding to the current battery fault type will be higher than the fault category weight for other fault types. After battery fault repair, a considerable amount of time will be spent modifying the fault category weights.

[0120] Step 45: Multiply the fault category weights corresponding to the multidimensional tensor data in the historical fault data with the weight decay coefficient to obtain the current fault category weights.

[0121] The current fault category weight refers to the fault category weight after weight decay adjustment. By multiplying the fault category weight by the weight decay coefficient, the weight of fault categories that are older in historical data and have a smaller impact on the current battery state can be reduced, thus focusing more on recent fault data and the actual state of the current battery.

[0122] Step 46: Execute steps 9 to 12 based on the current fault category weight.

[0123] Once the battery fault is repaired, it indicates that the dynamic weight matrix and bias value need to be updated, so steps 9 to 12 are executed.

[0124] Step 47: In response to the battery repair completion signal and when the expected fault type and the manually detected fault type are inconsistent, the corresponding fault category in the historical feature library is searched based on the manually detected fault type to optimize the weight and use it as the expected fault category weight.

[0125] Fault category optimization weights refer to the fault category weights relearned by the model through the loss function after repairing a faulty battery following a human-detected fault type during model training. Each time a human-detected fault type is encountered and a faulty battery is repaired, the subsequent model training process is recorded, obtaining the corresponding fault category weights after the repair. In other words, experience is accumulated through historical faults, so that when the same fault type occurs again, training can begin based on the fault category optimization weights without needing to perform random number generation and steps 43 to 45, thus increasing model learning efficiency.

[0126] The predicted fault category weights refer to the fault category weights predicted based on the fault types detected by humans. Specifically, after the battery repair is completed, when the fault type predicted by the model is inconsistent with the fault type detected by humans, the fault category weights corresponding to the actual fault type confirmed by humans are matched from the historical feature library and optimized. These optimized weights are then directly used as the initial fault category weights for the model to be retrained after the battery fault repair.

[0127] Step 48: Perform steps 9 through 12 based on the expected fault category weights.

[0128] The dynamic weight matrix and bias values ​​are trained by predicting the fault category weights.

[0129] This also includes: Step 49: When a faulty battery exists, search the historical fault data corresponding to the faulty battery in the historical feature database.

[0130] When a faulty battery is present, the system automatically retrieves and extracts all historical fault data related to that battery from the historical feature database based on the battery's unique identifier, such as the battery number.

[0131] Step 50: Determine the fault interval time based on the occurrence time of historical fault data and the current time.

[0132] The fault interval time refers to the difference between the time of the current fault of the battery and the time of the most recent fault of the battery recorded in the historical feature database.

[0133] Step 51: A battery is defined as a frequent failure battery when the failure interval falls within the preset frequent failure time range.

[0134] The frequent failure time range refers to a pre-set time threshold interval for judging whether a battery falls into the category of frequent failures, based on the battery's normal lifespan, maintenance standards, and failure characteristics. A frequently failing battery is one that experiences failures frequently. When the interval between failures falls within the frequent failure time range, it indicates that the battery may need to be replaced, and its potential safety hazards should be manually inspected.

[0135] Step 52: Send a preset alarm signal to a preset terminal device based on the frequent battery failures.

[0136] An alarm signal is a signal used to indicate frequent battery malfunctions. Terminal devices refer to the equipment used by personnel responsible for battery maintenance and management, such as monitoring center displays, maintenance personnel's mobile phones, tablets, etc.

[0137] Based on the same inventive concept, embodiments of the present invention provide a battery fault identification system.

[0138] A battery fault identification system, comprising: The acquisition module is used to acquire battery characteristic information and time information; A memory used to store a program that implements a battery fault identification method; The processor loads and executes programs from memory.

[0139] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A battery fault identification method, characterized in that, include: Step 1: After a preset test cycle, receive manually detected fault types and collect battery characteristic information and time information to form multidimensional tensor data; Step 2: Initialize the dynamic weight matrix and bias values ​​based on the preset random number generation function; Step 3: Input multidimensional tensor data into a preset deep learning model to calculate the output array; Step 4: Obtain the fault probability array based on the output array and the preset activation function; Step 5: Determine the expected fault type based on the fault probability array and preset sample labels; Step 6: If the expected fault type and the manually detected fault type are consistent, continue to execute the preset model identification method until the next test cycle; Step 7: When the expected fault type and the manually detected fault type are inconsistent, find the fault category weight corresponding to the sample label, and adjust the dynamic weight matrix and bias value until the expected fault type and the manually detected fault type are consistent.

2. The battery fault identification method according to claim 1, characterized in that, Methods for adjusting the dynamic weight matrix and bias values ​​include: Step 8: When no fault category weights exist, preset corresponding fault category weights for different fault types using a random number generation function; Step 9: Obtain the loss function value corresponding to the fault category based on the preset loss function, fault category weights, sample labels, and fault probability array; Step 10: When the loss function value corresponding to the fault category exceeds the preset standard threshold of the loss function, calculate the gradient of the dynamic weight matrix based on the output array, the fault probability array, and the dynamic weight matrix, and calculate the gradient of the bias value based on the output array and the fault probability array. Step 11: Calculate the corrected dynamic weight matrix based on the gradient of the dynamic weight matrix, the loss function, and the preset learning rate, and output it as the dynamic weight matrix. Calculate the corrected bias value based on the gradient of the bias value, the loss function, and the learning rate, and output it as the bias value. Step 12: Repeat steps 2 to 11 until the loss function values ​​corresponding to all fault categories fall into the standard threshold of the loss function, or the expected fault type is consistent with the manually detected fault type.

3. The battery fault identification method according to claim 2, characterized in that, It also includes a preset information compensation method for situations where multidimensional tensor data cannot be collected, the method comprising: Step 13: Determine if there is unidentifiable data in the multidimensional tensor data; Step 14: Determine the data loss time based on the time information in the lost multidimensional tensor data; Step 15: Determine the adjacent time range based on the data loss time; Step 16: Find multidimensional tensor data in adjacent time ranges to form multidimensional tensor data intervals; Step 17: Determine adjacent data based on multidimensional tensor data intervals; Step 18: Assign a preset weight decay coefficient based on the time information in the neighboring data, and calculate the reference neighboring data; Step 19: Calculate the reference multidimensional tensor data based on the reference neighbor data, and output it as multidimensional tensor data.

4. The battery fault identification method according to claim 3, characterized in that, It also includes a method for fault early warning, which includes: Step 20: Determine the warning threshold range based on the output array and the fault probability array; Step 21: Determine the risky multidimensional tensor data when the multidimensional tensor data exceeds the warning threshold range; Step 22: Determine the risk multidimensional tensor data interval when all multidimensional tensor data are risk multidimensional tensor data within a preset continuous time period; Step 23: Determine the minimum and maximum reference data based on the risk multidimensional tensor data and the risk multidimensional tensor data interval; Step 24: Obtain the battery time-varying data by subtracting the minimum and maximum reference data; Step 25: Determine the battery data mutation rate when the battery time-varying data exceeds the preset battery time-varying data threshold; Step 26: Determine the expected fault based on the battery data mutation rate and the corresponding fault data features in the preset fault data feature library, and output the corresponding expected fault risk signal.

5. The battery fault identification method according to claim 4, characterized in that, It also includes a method for correcting the fault data feature library, which includes: Step 27: Determine the standard deviation of battery parameters based on the dynamic weight matrix and multidimensional tensor data; Step 28: Calculate the parameter correlation degree based on the standard deviation of battery parameters and the dynamic weight matrix; Step 29: Sort and filter based on parameter correlation to obtain key feature parameters; Step 30: Determine the corrected fault data features based on key characteristic parameters and expected fault types; Step 31: Re-determine the fault data feature library based on the corrected fault data features.

6. The battery fault identification method according to claim 5, characterized in that, Also includes: Step 32: In response to a manual maintenance signal, identify the faulty battery based on the expected fault type; Step 33: When a faulty battery exists, a feature 3D projection is formed based on the faulty battery. The feature 3D projection is a 3D projection of the battery pack, and the projection light of the faulty battery flickers. Step 34: Determine the projected light color based on the expected fault type; Step 35: Determine the location of the faulty battery based on the color of the projected light, and perform preset maintenance operations based on the location of the faulty battery.

7. The battery fault identification method according to claim 6, characterized in that, It also includes a method for adjusting the flicker frequency of the light, the method comprising: Step 36: Establish risk levels based on fault data characteristics in the fault feature library; Step 37: Calculate the comprehensive risk index based on the failure probability array and battery mutation rate; Step 38: Match the comprehensive risk index with the risk level to determine the current risk level; Step 39: Determine the light flicker frequency based on the current risk level; Step 40: Control the lighting of the faulty battery in the 3D projection to flicker according to the lighting flicker frequency.

8. A battery fault identification method according to claim 7, characterized in that, It also includes a method for lightweight model learning after battery failure repair, which includes: Step 41: When a faulty battery exists, extract the multidimensional tensor data within the preset reliable time range before fault repair, as well as the fault type, fault probability array, and comprehensive risk index corresponding to the multidimensional tensor data, and form historical fault data; Step 42: Classify and store historical fault data according to fault type, occurrence time, and battery number to establish a historical feature database; Step 43: In response to the battery repair completion signal, the forgetting time is obtained by subtracting the time information in the historical fault data from the current time, and a forgetting curve function is constructed based on the forgetting time and the historical fault data; Step 44: Calculate the weight decay coefficient corresponding to the historical fault data based on the forgetting curve function; Step 45: Multiply the fault category weights corresponding to the multidimensional tensor data in the historical fault data with the weight decay coefficient to obtain the current fault category weights; Step 46: Execute steps 9 to 12 based on the current fault category weight; Step 47: In response to the battery repair completion signal and when the expected fault type and the manually detected fault type are inconsistent, the corresponding fault category weight in the historical feature library is searched based on the manually detected fault type and used as the expected fault category weight; Step 48: Perform steps 9 through 12 based on the expected fault category weights.

9. A battery fault identification method according to claim 8, characterized in that, Also includes: Step 49: When a faulty battery exists, search the historical fault data corresponding to the faulty battery in the historical feature database; Step 50: Determine the fault interval time based on the occurrence time of historical fault data and the current time; Step 51: A battery is defined as frequently failing when the fault interval time falls within a preset range of frequent fault times; Step 52: Send a preset alarm signal to a preset terminal device based on the frequent battery failures.

10. A battery fault identification system, characterized in that, include: The acquisition module is used to acquire battery characteristic information and time information; A memory for storing a program of a battery fault identification method as described in any one of claims 1 to 9; The processor loads and executes programs from memory.

Images