Pure electric ship battery thermal runaway risk early warning method and device

Abstract

The invention relates to a battery thermal runaway risk early warning method and device for a pure electric ship. The method comprises the following steps: acquiring a sensor data set detected by various sensors on the pure electric ship; according to the fluctuation intensity of each sensor data in the sensor data set, the weight of each sensor is determined, and the weight is dynamically adjusted through the fluctuation intensity of each sensor, so that weak but key early thermal runaway signals can be identified, and feature fusion is performed on the data of various sensors in the sensor data set according to the weight. Obtaining a comprehensive feature vector after dynamic weight adjustment; the thermal runaway risk identification model can predict the temperature change trend of the battery in the future by analyzing the dynamic trajectory of the parameters in the battery charging and discharging process, so that the comprehensive feature vector is input into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, and risk early warning is performed on the pure electric ship according to the battery temperature sequence. And the early warning accuracy of the pure electric ship is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy ships, and particularly relates to a pure electric ship battery thermal runaway risk early warning method and device. BACKGROUND

[0002] The rapid development of global new energy ships is driven by the superposition of multiple strategic needs and technological changes. New energy and clean energy ships have become a research hotspot in the field of ships at present. However, the ship operating environment has characteristics such as high humidity, strong vibration and large temperature difference, and the battery system is in a high-power charging and discharging state for a long time, which is prone to cause thermal runaway accidents.

[0003] In the prior art, a pure electric ship is detected and warned by a high-precision battery thermal runaway prediction system, but the traditional sensor network has inconsistent collection frequencies, it is difficult to establish a continuous time sequence feature vector, it cannot analyze the time correlation of data, and it lacks a dynamic weight distribution mechanism, and it is not sensitive enough to the volatility of sensor data, it cannot prioritize capturing abnormal signals, resulting in weak but critical early thermal runaway features being drowned out by noise, and thus the detection and warning accuracy of the battery thermal runaway prediction system for the pure electric ship is low.

[0004] Therefore, it is necessary to propose a pure electric ship battery thermal runaway risk early warning method and device to solve the technical problem of low detection and warning accuracy of the battery thermal runaway prediction system for the pure electric ship due to the lack of sensitivity of the traditional sensor network to the volatility of sensor data and the inability to prioritize capturing abnormal signals. SUMMARY

[0005] Therefore, it is necessary to propose a pure electric ship battery thermal runaway risk early warning method and device to solve the technical problem of low detection and warning accuracy of the battery thermal runaway prediction system for the pure electric ship due to the lack of sensitivity of the traditional sensor network to the volatility of sensor data and the inability to prioritize capturing abnormal signals.

[0006] To solve the above problems, in a first aspect, the present application provides a pure electric ship battery thermal runaway risk early warning method, comprising: constructing a battery thermal runaway risk identification model to obtain a sensor data set detected by multiple sensors on a pure electric ship; determining the weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set; performing feature fusion on the data of the multiple sensors in the sensor data set according to the weight to obtain a comprehensive feature vector; The comprehensive characteristic vector is input into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, and a risk warning is issued for the pure electric ship based on the battery temperature sequence.

[0007] In one possible implementation, obtaining sensor data sets detected by multiple sensors on the pure electric ship includes: Acquire initial sensor data collected by the multiple sensors of each battery pack on the pure electric vessel through an initial sampling frequency; Determining the degree of deviation between the initial sensor data of each sensor and the overall data within the time window according to the distribution relationship of the initial sensor data within the window time; Adjusting the initial sampling frequency of the corresponding battery pack according to the degree of deviation to obtain a sampling frequency of the corresponding battery pack; The sensor data sets collected by the multiple sensors corresponding to the battery pack on the pure electric ship are obtained according to the collection frequency.

[0008] In a possible implementation, determining the weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set includes: Preprocessing the data in the sensor data set to obtain a preprocessed data set; Calculating each sensor data in the preprocessed data set respectively to obtain the corresponding variance; The weight of each sensor is determined according to the proportion of each variance in all variances; the proportion is in direct proportion to the weight.

[0009] In a possible implementation, preprocessing the data in the sensor data set to obtain a preprocessed data set includes: Determining the median of data within a preset length in the sensor data set according to the sliding window; and replacing the original data value of the center point of the sliding window with the median to obtain a first data set; Using a time series linear interpolation method to complete the missing data in the first data set to obtain a second data set; The data of the multiple sensors in the second data set are standardized to obtain a preprocessed data set.

[0010] In a possible implementation, the method of using a time series linear interpolation method to complete missing data in the first dataset to obtain a second dataset includes: Identifying the data in the first data set to obtain all missing points and corresponding first timestamp positions; Determine, based on each missing point and the timestamp position, a previous valid data point and a second timestamp and a next valid data point and a third timestamp; Obtaining a linear interpolation weight according to the first timestamp, the second timestamp, and the third timestamp; Obtaining a missing value corresponding to the missing point according to the previous valid data point, the next valid data point, and the linear interpolation weight; The missing data in the first data set are completed according to all missing values ​​to obtain a second data set.

[0011] In a possible implementation, the standardizing the data of the multiple sensors in the first data set to obtain a preprocessed data set includes: Building a sliding window within a preset time based on the first data set; Calculating the sensor data in the sliding window to obtain a mean and a standard deviation; The sensor data at each moment is calculated according to the mean and the standard deviation to obtain a preprocessed data set.

[0012] In one possible implementation, building a battery thermal runaway risk identification model includes: Acquire historical data sets of the multiple sensors of the pure electric ship based on a simulation platform; Performing data cleaning and feature fusion on the historical data set to obtain a historical feature vector; Dividing the historical feature vectors into a data set to obtain a training set, a validation set, and a test set; The initial battery thermal runaway risk identification model is trained on a CNN model, an RNN model, an LSTM model, and a Transformer model according to the training set, the validation set, and the test set to obtain corresponding training results; The training results of all models were compared by means of mean square error, mean absolute error and determination coefficient to obtain the optimal model, which was then determined as the battery thermal runaway risk identification model.

[0013] In a possible implementation, the acquiring of historical data sets of the multiple sensors of the pure electric ship based on a simulation platform includes: Determining multiple fault types based on the multiple sensors; the multiple sensors are respectively used to detect battery temperature, battery voltage, battery current, hydrogen concentration, and carbon dioxide concentration; Simulating at least one fault type according to the simulation platform to obtain all sensor data; annotating all the sensor data with risk levels to obtain a historical data set.

[0014] In a possible implementation, the risk warning for the pure electric ship according to the battery temperature sequence comprises: setting a plurality of risk warning levels and corresponding predicted temperature intervals; determining a target risk warning level according to the battery temperature sequence and the predicted temperature interval; warning the pure electric ship of risks according to the risk warning level.

[0015] In a second aspect, the present application also provides a pure electric ship battery thermal runaway risk warning device, comprising: a data acquisition module for constructing a battery thermal runaway risk identification model and acquiring a sensor data set detected by a plurality of sensors on a pure electric ship; a weight determination module for determining the weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set; a feature fusion module for fusing the data of the plurality of sensors in the sensor data set according to the weight to obtain a comprehensive feature vector; a model prediction module for inputting the comprehensive feature vector into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence and warning the pure electric ship of risks according to the battery temperature sequence.

[0016] The present application has the advantages that a battery thermal runaway risk identification model is constructed, a sensor data set detected by a plurality of sensors on a pure electric ship is acquired, the weight of each sensor is determined according to the fluctuation intensity of each sensor data in the sensor data set, the weight is dynamically adjusted through the fluctuation intensity of each sensor, weak but critical early thermal runaway signals can be identified, the data of the plurality of sensors in the sensor data set can be fused according to the weight to obtain a comprehensive feature vector after the weight is dynamically adjusted, further, the thermal runaway risk identification model can predict the temperature change trend of the future battery by analyzing the dynamic trajectory of the parameters in the battery charging and discharging process, the comprehensive feature vector can be input into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, the pure electric ship can be warned of risks according to the battery temperature sequence, and the warning accuracy of the pure electric ship is improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 An embodiment flowchart of the pure electric ship battery thermal runaway risk warning method provided by the present application is shown in the figure; Figure 2A structural diagram of an embodiment of the working principle diagram of the dynamic data acquisition module provided by the present invention; Figure 3 For the present invention Figure 1 A schematic flow chart of an embodiment of step S102; Figure 4 For the present invention Figure 3 A schematic flow chart of an embodiment of step S301; Figure 5 For the present invention Figure 4 A schematic flow chart of an embodiment of step S402; Figure 6 A schematic structural diagram of an embodiment of a thermal runaway risk identification model for a pure electric ship provided by the present invention; Figure 7 This is a schematic structural diagram of an embodiment of a pure electric ship battery thermal runaway risk warning device provided by the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0019] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for early warning of thermal runaway risk of batteries for pure electric ships, comprising: S101. Build a battery thermal runaway risk identification model and obtain sensor data sets detected by multiple sensors on pure electric ships.

[0020] The pure electric ship battery thermal runaway risk warning method provided in the embodiment of the present application can be applied to the pure electric ship battery thermal runaway risk warning system. The pure electric ship battery thermal runaway risk warning system can be based on a software system running on a terminal device. The terminal device can be a server, a tablet computer, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), a mobile phone, and other terminal devices. The embodiment of the present application does not impose any restrictions on the specific type of the terminal device.

[0021] The battery thermal runaway risk identification model can be determined through simulation experiments, and a Gaussian dynamic sensor network can be deployed on the pure electric ship. The network is composed of temperature sensors, current sensors, voltage sensors, hydrogen sensors, carbon dioxide sensors, and data collectors. The sensor is an element for measuring battery parameters, and the data collector converts the sensor signal into a digital signal at a certain frequency and uploads it to the battery thermal runaway risk identification model. The sensor data set refers to the original data set collected by various sensors such as temperature, voltage, and current in the ship battery system within a specified time period. Specifically, a distributed data acquisition unit can be used to collect real-time sensor signals and form a time-aligned multi-dimensional data matrix.

[0022] S102, according to the fluctuation intensity of each sensor data in the sensor data set, determining the weight of each sensor.

[0023] The fluctuation intensity refers to the dispersion degree of the sensor data in the time dimension, which can be quantified by calculating the variance or standard deviation of the data in the sliding window. The larger the variance value, the more intense the fluctuation of the sensor data. Weight determination refers to assigning an impact coefficient reflecting the data fluctuation of each sensor. Specifically, a normalization processing method can be used, and the proportion of the variance value of each sensor to the total variance is used as the weight value. The more volatile the sensor, the higher the weight.

[0024] S103, according to the weight, performing feature fusion on the data of various sensors in the sensor data set to obtain a comprehensive feature vector.

[0025] The feature fusion refers to converting multi-source heterogeneous sensor data into a unified dimensional feature vector. Specifically, a weighted average algorithm can be used to multiply each sensor data by its corresponding weight and then superimpose it to form a fusion feature representing the overall state of the system.

[0026] S104, inputting the comprehensive feature vector into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, and performing risk warning on the pure electric ship according to the battery temperature sequence.

[0027] Specifically, a Gaussian dynamic sensor network is first used to obtain sensor data sets for sensors such as temperature and voltage from the battery management system. The 24-hour sliding variance of the processed data for each sensor is also calculated. When the voltage sensor variance is 0.15 and the temperature sensor variance is 0.23, the temperature sensor receives a weight of 0.23 / (0.15+0.23)=60%. The feature fusion module multiplies the voltage data by 0.4 and the temperature data by 0.6, and then superimposes them to form a comprehensive feature vector containing multi-sensor information. This comprehensive feature vector is input into the battery thermal runaway risk identification model, which outputs a battery temperature sequence for the next 30 minutes. Risk warnings can then be issued based on the battery temperature sequence. For example, a level 3 warning can be triggered when the predicted value exceeds 85°C.

[0028] Compared with existing technologies, traditional methods use fixed-frequency data acquisition, resulting in a loss of temporal correlation. This solution constructs a continuous time series through dynamic interpolation and sliding window processing. While existing technologies treat sensor data with equal weights, this solution dynamically assigns weights based on fluctuation intensity, increasing the contribution of abnormal sensor features by 2.3 times. Conventional early warning systems directly train models using raw data, while this solution generates high signal-to-noise ratio inputs through feature fusion, improving the model's accuracy in identifying early thermal runaway signatures by 19 percentage points.

[0029] Compared with the prior art, the present embodiment provides a battery thermal runaway risk identification model for constructing a sensor data set detected by multiple sensors on a pure electric ship; the weight of each sensor is determined according to the fluctuation intensity of the data of each sensor in the sensor data set, and the weight is dynamically adjusted according to the fluctuation intensity of each sensor, so that weak but critical early thermal runaway signals can be identified, so that the data of multiple sensors in the sensor data set can be feature fused according to the weights to obtain a comprehensive feature vector after the weights are dynamically adjusted; further, the thermal runaway risk identification model can predict the future temperature change trend of the battery by analyzing the dynamic trajectory of the parameters during the battery charging and discharging process, so that the comprehensive feature vector can be input into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, and then a risk warning for the pure electric ship can be performed based on the battery temperature sequence, thereby improving the accuracy of the warning for the pure electric ship.

[0030] In some embodiments of the present invention, step S101 includes: Acquire initial sensor data from multiple sensors of each battery pack on the pure electric ship through an initial sampling frequency; According to the distribution relationship of the initial sensor data within the window time, the deviation degree between the initial sensor data of each sensor and the overall data within the time window is determined; Adjusting the initial sampling frequency of the corresponding battery group according to the degree of deviation to obtain the sampling frequency of the corresponding battery group; The sensor data sets collected by various sensors corresponding to the battery pack on the pure electric ship are obtained according to the collection frequency.

[0031] The window time refers to a preset period of time used to analyze the distribution of initial sensor data. This can be achieved using a fixed or dynamically adjustable time length, for example, from 30 seconds to 5 minutes. This window time is used to dynamically determine the range of data fluctuations, providing a basis for adjusting the acquisition frequency. The distribution relationship refers to the statistical distribution characteristics of the initial sensor data within the window time. This can be achieved using a probability density function or histogram comparison method. By analyzing the distribution differences between current and historical data, sudden changes or abnormal trends in the data can be identified, triggering an adjustment mechanism for the acquisition frequency. The acquisition frequency refers to the sampling interval of sensor data. This can be achieved using a dynamic adjustment strategy, for example, shortening the sampling interval from 1 second to 0.5 seconds when the data distribution deviates from a preset threshold. This frequency adjustment balances data acquisition efficiency with the ability to detect anomalies, avoiding redundant data processing caused by high-frequency sampling.

[0032] Specifically, if Figure 2 As shown, Figure 2 The working principle diagram of the dynamic data acquisition module is as follows. The data parameters of the sensor group of each battery pack can be collected. The initial sensor data is obtained, such as battery temperature, voltage and other parameters. The initial sampling frequency of the sensor network is 10HZ. Then, anomaly detection is performed on the initial sensor data collected by each sensor respectively. The deviation degree of the initial sensor data collected by each sensor from the overall data in its time window is analyzed. Then, the initial sampling frequency of the corresponding battery pack is adjusted according to the deviation degree to obtain the sampling frequency of the corresponding battery pack. For example, the mean and standard deviation of the sensor data in the battery pack are calculated based on the sliding window (the latest 30 data points). When there is data that does not exceed It is considered "normal" when a certain data exceeds It is considered as a "medium abnormality" when a certain data exceeds If the data distribution is stable, the collection frequency for that battery group is increased to a higher level. Conversely, if the data distribution is stable, the collection frequency for that battery group is reduced to reduce resource consumption. The sampling frequency of each battery group is controlled independently. For example, if the temperature in battery group 1 suddenly deviates from the normal distribution range, the collection frequency of the entire sensor group in battery group 1 is increased, while the collection frequency of the sensor network of other battery groups remains unchanged. Finally, data is collected from the corresponding sensors based on the adjusted collection frequency to form a sensor dataset. When the data is normal, the collection frequency can be 10 Hz, when it is a medium-level anomaly, the collection frequency can be 100 Hz, and when it is a high-level anomaly, the collection frequency can be 1 kHz. If the data value remains below the threshold for a period of time (five consecutive samples are normal), the frequency is switched back to 10 Hz to prevent frequent switching due to transient fluctuations.

[0033] Compared to existing technologies, traditional methods typically use a fixed sampling frequency and are unable to adjust the sampling interval based on dynamic data changes. For example, maintaining a high sampling frequency when the battery is stable results in wasted storage and computing resources; during unexpected events, insufficient sampling frequency may lead to the loss of critical data. This solution dynamically analyzes data distribution within the time window and adaptively adjusts the sampling frequency, ensuring data integrity while optimizing resource utilization.

[0034] Through the above technical solution, this application can effectively solve the problem of discontinuous time series features caused by fixed acquisition frequencies in sensor networks. By dynamically adjusting the acquisition strategy, it ensures data capture accuracy during abnormal fluctuations while avoiding the accumulation of redundant data during stable states, providing a high-quality data foundation for subsequent feature fusion and risk prediction.

[0035] In some embodiments of the present invention, Figure 3 , step S102 includes: S301 : Preprocess the data in the sensor data set to obtain a preprocessed data set.

[0036] Among them, the preprocessed data set refers to the sensor data set that has been processed through data cleaning and standardization. Specifically, it can be achieved by using sliding window denoising, missing value interpolation and standardization processing to eliminate noise interference and unify the data scale, providing reliable input for subsequent variance calculation.

[0037] S302 , respectively calculating each sensor data in the preprocessed data set to obtain the corresponding variance.

[0038] Variance refers to the degree of dispersion of sensor data in the time series. It can be achieved by calculating the variance value of each sensor data sequence. It is used to quantify the fluctuation intensity of sensor data. The larger the variance, the more obvious the abnormal fluctuation of the sensor data.

[0039] S303 , determining the weight of each sensor according to the proportion of each variance in all variances; the proportion and the weight are in direct proportion.

[0040] The weight refers to the contribution coefficient of each sensor in the feature fusion, which can be achieved by calculating the proportion of each sensor's variance to the overall variance. It is used to dynamically adjust the fusion weights of different sensors so that the sensor data with severe fluctuations occupies a higher proportion in the feature vector. The weight calculation is shown in formula (1): (1) Where: For sensors i Variance within the last 10 seconds; for t -9 sensor data after normalization; The average value of the normalized sensor data in the past 10 seconds.

[0041] The weight distribution is shown in formula (2): (2) Where: For sensors i The weight of the data, the sum of all sensor weights is 1.

[0042] Then the comprehensive feature vector can be ( 、 、 、 ).

[0043] Specifically, during the preprocessing phase, the raw sensor data first undergoes a sliding window denoising process, such as using a median filter with a window length of 5 seconds to eliminate transient noise. For missing data points, missing timestamps are identified and valid data points before and after are extracted for linear interpolation. After missing values ​​are filled, each sensor data series is normalized, for example, using a sliding window normalization method with a mean of 0 and a standard deviation of 1 to eliminate dimensional differences. The preprocessed dataset is input into the variance calculation module, which calculates the variance of each sensor's time series data. For example, the variance of the battery temperature sensor is σ1, and the variance of the voltage sensor is σ2. The final weight assignment is determined by the proportion of each sensor's variance to the total variance of all sensors. The proportion is directly proportional to the weight. For example, when σ1 accounts for 60% of the total variance, the weight of that sensor is set to 0.6, giving it a higher weight during feature fusion.

[0044] Compared with the prior art, the existing sensor network adopts a fixed weight distribution mechanism and cannot adjust the importance of the sensor according to real-time data fluctuations, resulting in a lag in abnormal signal identification. The scheme dynamically distributes weights according to the variance proportion, so that the sensor with strong fluctuation intensity obtains a higher weight in feature fusion. For example, when the temperature of a certain battery monomer abnormally rises, the variance of the corresponding temperature sensor will significantly increase, and the weight will automatically increase to a dominant position, so as to preferentially capture early thermal runaway features.

[0045] Through the above technical scheme, the application can dynamically adjust the weight distribution according to the real-time fluctuation characteristics of the sensor data, enhance the sensitivity to abnormal fluctuation signals, effectively reduce the interference of noise data on feature fusion, and improve the identification ability of the thermal runaway early warning system to early abnormal features.

[0046] In some embodiments of the application, as shown in Figure 4 Step S301 includes: S401, determining the median of the data in the preset length in the sensor data set according to the sliding window; and replacing the original data value of the center point of the sliding window with the median to obtain a first data set; The median denoising of the sliding window refers to replacing the original data value of the center point of the window with the median value of the data in the window by setting a data window of a fixed length. Specifically, a sliding median filter with a window length of 5 to 15 data points can be used to achieve this. This method can effectively eliminate pulse noise and outlier interference.

[0047] S402, the missing data in the first data set is completed by using a time series linear interpolation method to obtain a second data set.

[0048] The time series linear interpolation method refers to calculating the interpolation weight using the time stamp difference of the effective data points before and after the missing point. Specifically, the linear interpolation coefficient of the missing point can be determined by the time interval ratio of the previous effective data point and the next effective data point. This method can maintain the continuity characteristics of the time series.

[0049] S403, the data of multiple sensors in the second data set is standardized to obtain a preprocessed data set.

[0050] The standardization processing refers to normalizing the sensor data. Specifically, the mean and standard deviation of the data in the sliding window can be used for z-score standardization calculation. This method can eliminate the influence of the dimension difference of different sensors on subsequent feature fusion.

[0051] Specifically, during ship operation, the raw sensor data is first denoised using a sliding median filter to generate the first dataset. For example, when the window length is set to 9 data points, the sensor data at each moment is replaced with the median of the four preceding and four subsequent data points. This effectively suppresses transient data changes caused by ship vibration. For any missing data segments that remain after denoising, the valid data before and after the missing data are located based on the timestamp information. For example, if a temperature sensor has missing data between timestamps t=10.2 and 10.5, valid data points A at t=10.0 and B at t=10.6 are selected. Based on the time interval ratio, an interpolation weight of 0.5 is calculated for t=10.3 seconds. Finally, the missing values ​​are filled using the weighted average of data points A and B, resulting in the second dataset. After the missing values ​​are filled, a sliding window normalization method is used. For example, with a window length of 30 seconds, the mean and standard deviation of each sensor within the window are calculated in real time to convert the current data into a standard normal distribution, thus preprocessing the dataset.

[0052] In some embodiments of the present invention, Figure 5 As shown, step S402 includes: S501: Identify data in a first data set to obtain all missing points and corresponding first timestamp positions.

[0053] Missing points are data locations in the sensor data stream that are not effectively recorded. Specifically, they can be identified by traversing the nodes in the data sequence with empty values ​​to locate the data locations that need repair. For example, missing points in the data (such as NaN or blank values) can be identified and their timestamp locations recorded.

[0054] S502: Determine the previous valid data point and the second timestamp and the next valid data point and the third timestamp according to each missing point and the timestamp position.

[0055] The timestamp position refers to the time stamp of the data collection moment. Specifically, it can be implemented by recording millisecond timestamps to establish the temporal relationship between data points. For each missing value, find its previous and next valid data points and their corresponding timestamps.

[0056] S503: Obtain a linear interpolation weight according to the first timestamp, the second timestamp, and the third timestamp.

[0057] The effective data points refer to adjacent non-empty data points before and after the missing points, which can be determined by bidirectional search of the nearest non-empty values, and are used to provide the basis data for interpolation calculation. The linear interpolation weight refers to an interpolation coefficient calculated based on the time interval ratio, which can be realized by calculating the ratio of the difference between the missing time stamp and the front and rear effective time stamps, and is used to dynamically adjust the contribution degree of the front and rear data to the missing value. The linear interpolation weight calculation is shown in formula (3): (3) In the formula, is the time stamp of the missing value, is the time stamp of the previous effective data point, is the time stamp of the next effective data point.

[0058] S504, obtaining the missing value corresponding to the missing point according to the previous effective data point, the next effective data point and the linear interpolation weight.

[0059] The missing value refers to a filling value calculated by an interpolation algorithm, which can be realized by linear superposition of the front and rear effective data points according to the weight ratio, and is used to restore the integrity of the data sequence. The missing value calculation is shown in formula (4): (4) In the formula, is the missing value, is the value of the previous effective data point, is the value of the next effective data point.

[0060] S505, filling the missing data in the first data set according to all the missing values to obtain the second data set.

[0061] Specifically, after identifying the data missing points, the time stamp of each missing point is accurately recorded to determine the adjacent effective data points and their corresponding time stamps. By calculating the time interval ratio of the missing time stamp and the front and rear effective time stamps, an interpolation weight coefficient is dynamically generated. For example, if the time interval between a missing point and the previous effective data point accounts for 30% of the total interval, the previous data point participates in the calculation with a weight of 70%. Finally, the weighted front and rear data are superimposed to generate a filling value that conforms to the trend of the time series change, forming a continuous and complete second data set.

[0062] Compared with the prior art, the traditional method uses a fixed interpolation window or ignores the mean filling of the time stamp difference, which is easy to cause data mutation or trend distortion. The present scheme generates dynamic weights by accurately calculating the time stamp interval, so that the filling value is more consistent with the data change law under actual working conditions, avoiding the feature deviation problem caused by time axis misalignment.

[0063] Through the above technical solution, the present application can eliminate local omissions caused by equipment failure or signal interference during sensor data acquisition, restore the continuity of data in the time dimension, ensure the accuracy of time-series correlation features in the subsequent feature fusion process, and thus enhance the thermal runaway warning model's ability to capture early abnormal signals.

[0064] In some embodiments of the present invention, step S303 includes: Constructing a sliding window within a preset time according to the first data set; Calculate the sensor data in the sliding window to obtain the mean and standard deviation; The sensor data at each moment is calculated according to the mean and standard deviation to obtain the preprocessed data set.

[0065] Among them, the sliding window refers to the dynamic interception of fixed-length data blocks of adjacent time periods when processing time series data. Specifically, the window can be constructed using a preset time span of, for example, 5 minutes or 10 minutes. The sliding window can adapt to the dynamic changes in sensor data. The mean and standard deviation are statistics used to describe the distribution characteristics of the data within the sliding window. Specifically, the arithmetic mean and square root of the variance of the data within the window can be calculated point by point. These two parameters can eliminate the dimensional differences between different sensors. Normalization processing refers to converting the original data into standard normal distribution data with a mean of 0 and a standard deviation of 1. Specifically, the Z-score normalization formula can be used to perform a linear transformation on each data point. Through normalization processing, the data scale of different sensors can be unified.

[0066] Specifically, when constructing a sliding window, the preset time length can be set to, for example, 30 seconds based on the response speed of the battery system. Each window contains 30 consecutive sampling points. Let the current moment be t, and the sensor data within the last 30 seconds is taken to construct a sliding window. For the temperature, voltage, current and other sensor data in each window, the mean and standard deviation are calculated respectively. Subsequently, for the raw sensor data at each moment, the mean and standard deviation of the corresponding window are used for standardization calculation. For example, the current temperature value is subtracted from the window mean and then divided by the standard deviation, so that all sensor data are within the same dimensional range. The standard is shown in formula (5): (5) Where: After standardization t Momentary sensor data; for t Initial sensor data at the moment; is the mean value of sensor data in the 30s time window; is the standard deviation of the sensor data within the 30s time window.

[0067] The resulting preprocessed dataset eliminates the data distribution shift caused by differences in sensor types and enhances the stability of subsequent feature fusion and model prediction.

[0068] Compared with existing technologies, traditional methods typically use global normalization or fixed-time period statistics, which are unable to adapt to the dynamically changing operating conditions of battery systems. This solution, however, calculates local statistics using a sliding window, capturing data fluctuation trends in real time and avoiding the issue of standardized parameter failure caused by sudden environmental changes or operating condition changes. For example, when the battery enters a high-load discharge state, the sliding window dynamically adjusts the mean and standard deviation to accurately reflect the current data distribution characteristics, while traditional fixed-parameter methods can introduce errors due to statistical lag.

[0069] Through the above technical solution, this application effectively solves the difficulty of feature fusion caused by dimensional differences in multi-source sensor data, and improves the adaptability of data preprocessing to dynamic operating conditions. The standardized data can more realistically reflect the changing trends of battery status, reduce the impact of noise interference on subsequent model predictions, and thus provide a reliable data foundation for accurately identifying early thermal runaway risks.

[0070] In some embodiments of the present invention, step S101 includes: Obtain historical data sets of various sensors of pure electric ships based on the simulation platform.

[0071] Among them, the historical data set refers to a collection of sensor operating status under various working conditions. Specifically, it can be generated by injecting different fault types through the simulation platform, such as abnormal battery temperature, voltage drop and other scenarios, to cover potential risk modes in actual operation.

[0072] Perform data cleaning and feature fusion on historical data sets to obtain historical feature vectors.

[0073] Data cleaning is the process of removing noise and outliers, which can be achieved through sliding window filtering or outlier detection algorithms to ensure the validity of the input data. Feature fusion is the integration of multi-source sensor information into a unified representation, which can be achieved through weighted averaging or principal component analysis to extract relevant features from multidimensional data.

[0074] The historical feature vectors are divided into data sets to obtain training sets, validation sets, and test sets.

[0075] Data set partitioning refers to distributing training samples proportionally, which can be achieved by random sampling or time series segmentation to prevent overfitting during model training.

[0076] According to the training set, validation set and test set, the initial battery thermal runaway risk identification model is trained on the CNN model, RNN model, LSTM model and Transformer model respectively to obtain the corresponding training results.

[0077] Among them, the CNN model, RNN model, LSTM model and Transformer model refer to untrained machine learning frameworks, which can be implemented using long short-term memory neural network or convolutional neural network architecture.

[0078] The training results of all models were compared by means of mean square error, mean absolute error and determination coefficient to obtain the optimal model, which was then determined as the battery thermal runaway risk identification model.

[0079] Among them, the training results of each model can be calculated separately to obtain the mean square error, mean absolute error and determination coefficient corresponding to each model. Then, the mean square error, mean absolute error and determination coefficient of each model can be compared to obtain the comprehensive optimal model. The optimal model can be determined as the battery thermal runaway risk identification model.

[0080] In a specific embodiment of the present invention, Figure 6 As shown, the construction of a battery thermal runaway risk identification model includes data acquisition, data preprocessing, model construction, model training, and model evaluation. A pure electric ship simulation experiment can be conducted to collect battery parameter data from the ship during operation, i.e., a historical dataset. The data preprocessing phase involves data cleaning, such as removing noise and missing values. Data standardization is then performed, followed by feature engineering, which involves weight calculation and feature fusion to obtain a feature vector. The model construction phase can use a deep learning model such as CNN, RNN, LSTM, or Transformer. A loss function and optimizer are then designed. The model training phase then proceeds. The dataset is chronologically divided into a training set (70%), a validation set (15%), and a test set (15%). Time series feature vectors from 30-second sliding windows in the training, validation, and test sets are sampled as input data. These feature vectors contain the following five sensor values: ① battery temperature (°C), ② battery voltage (V), ③ battery current (A), ④ hydrogen concentration (ppm), and ⑤ carbon dioxide concentration (ppm). Output data: ① Prediction task: Based on the current 30-second window data, predict the battery temperature sequence within the next 10 seconds. ② Risk level classification: Synchronously output the risk level corresponding to the highest temperature in the next 10 seconds for early warning decision-making. Batch size: 32, training cycle: 200 epochs, early stopping (patience=10, terminate when the validation loss does not decrease), regularization: L2 regularization ( λ=1e-5), to prevent overfitting. Then in the model evaluation phase, the evaluation indicators are mainly the mean square error (MSE), mean absolute error (MAE) and To compare and evaluate the effects of different models, and select the model with the best comprehensive performance as the thermal runaway risk identification model for pure electric ships according to the user's specific situation and needs.

[0081] In some embodiments of the present invention, historical data sets of various sensors of a pure electric ship are obtained based on a simulation platform, including: Determine multiple fault types based on multiple sensors; the multiple sensors are used to detect battery temperature, battery voltage, battery current, hydrogen concentration, and carbon dioxide concentration; Simulate at least one fault type according to the simulation platform to obtain all sensor data; All sensor data are labeled with risk levels to obtain a historical data set.

[0082] In a specific embodiment of the present invention, the following monitoring data are selected based on the design of the system's sensor network: battery temperature, battery voltage, battery current, hydrogen concentration, and carbon dioxide concentration. By collecting and organizing case studies of thermal runaway risk accidents involving pure electric vessels, several typical accident scenarios (i.e., fault types) and triggering conditions are summarized in Table 1.

[0083] Table 1. Thermal runaway risk accident scenarios for pure electric ships

[0084] Scaled simulation cabin: Build a scaled simulation model of the key cabins of the ship (such as the battery compartment and cargo hold). The materials must be consistent with the actual ship (such as fireproof partitions and ventilation structures).

[0085] Multi-scenario fault injection: Dynamically adjust simulation parameters through programming interfaces (such as Python APIs), simulate five core fault scenarios, and generate multi-dimensional time series data including temperature, voltage, current, and gas concentration.

[0086] Sensor placement: A dynamic frequency regulation sensor network designed according to an embodiment of the present invention is deployed in each battery pack to obtain parameter data of the battery under various working conditions.

[0087] Normal operating condition data collection: Start the established simulation platform, run the equipment to a stable state, and then continuously collect all sensor data for at least 2 hours as a normal operating condition baseline.

[0088] Fault scenario simulation: ① Single fault trigger: for example, trigger S1 scenario: control the charging device to output over-limit current, then monitor the battery temperature, voltage, gas concentration change in real time until thermal runaway occurs. ② Multi-fault coupling: trigger two or more fault scenarios at the same time, and monitor the parameter changes of each battery pack in real time. ③ Data recording and labeling: record all sensor data to obtain a historical data set in CSV format, and label the risk level according to the rules in Table 2.

[0089] Table 2, risk level labeling rules

[0090] In some embodiments of the present application, the pure electric ship is warned according to the battery temperature sequence, comprising: setting multiple risk warning levels and corresponding predicted temperature intervals; determining the target risk warning level according to the battery temperature sequence and the predicted temperature interval; warning the pure electric ship according to the risk warning level.

[0091] In specific embodiments of the present application, a hierarchical warning mechanism is adopted, which is divided into three risk warning levels, namely low risk, medium risk and high risk, as shown in Table 2. Each level is set with a battery temperature interval, and the corresponding risk warning level is determined according to the temperature. Different levels of warning signals are sent according to the predicted risk warning level, so that the crew can take more effective safety measures according to different situations. The risk level warning prompt rules are shown in Table 3: Table 3, risk level warning prompt rules

[0092] In some specific embodiments, the predicted temperature interval can be dynamically adjusted, for example, the interval boundary value is adaptively corrected according to the environmental temperature data of the ship sailing area. When it is detected that the battery temperature sequence exceeds the upper limit value of the current warning level interval for three consecutive sampling periods, it can be automatically upgraded to a higher warning level.

[0093] Compared with the prior art, the traditional method only uses a single temperature threshold for warning, and cannot distinguish the difference in risk level. The present scheme can identify potential risks in advance according to the temperature rise gradient through a multi-level warning mechanism, for example, preventive cooling measures can be started when the temperature is in the primary warning interval, avoiding triggering emergency treatment after reaching the critical temperature.

[0094] Through the above technical solution, this application can achieve refined hierarchical management of risk warnings, solving the high false alarm rate and delayed response caused by the single threshold of traditional methods. By establishing a mapping relationship between multi-level temperature ranges and warning actions, targeted measures can be taken at different risk stages. For example, in the primary warning stage, only the battery charging and discharging strategy needs to be adjusted without interrupting the operation of the ship's power system, thereby maintaining the ship's normal operating capabilities while ensuring safety.

[0095] In order to better implement the pure electric ship battery thermal runaway risk warning method in the embodiment of the present invention, based on the pure electric ship battery thermal runaway risk warning method, the embodiment of the present invention also provides a pure electric ship battery thermal runaway risk warning device, such as Figure 7 As shown, the pure electric ship battery thermal runaway risk warning device 700 includes: The data acquisition module 701 is used to build a battery thermal runaway risk identification model and obtain sensor data sets detected by various sensors on the pure electric ship; A weight determination module 702 is configured to determine a weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set; A feature fusion module 703 is used to perform feature fusion on the data of multiple sensors in the sensor data set according to weights to obtain a comprehensive feature vector; The model prediction module 704 is used to input the comprehensive feature vector into the thermal runaway risk identification model for prediction, obtain the battery temperature sequence, and issue a risk warning for the pure electric ship based on the battery temperature sequence.

[0096] The pure electric ship battery thermal runaway risk warning device 700 provided in the above embodiment can implement the technical solution described in the above embodiment of the pure electric ship battery thermal runaway risk warning method. The specific implementation principles of the above modules or units can refer to the corresponding contents in the above embodiment of the pure electric ship battery thermal runaway risk warning method, which will not be repeated here.

[0097] The above is a detailed introduction to the pure electric ship battery thermal runaway risk warning method and device provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A method for early warning of thermal runaway risk of batteries for pure electric ships, characterized in that: include: Build a battery thermal runaway risk identification model and obtain sensor data sets detected by various sensors on pure electric ships; Determining a weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set; Performing feature fusion on the data of the multiple sensors in the sensor data set according to the weights to obtain a comprehensive feature vector; The comprehensive characteristic vector is input into the thermal runaway risk identification model for prediction to obtain a battery temperature sequence, and a risk warning is issued for the pure electric ship based on the battery temperature sequence.

2. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 1, characterized in that: The sensor data set detected by multiple sensors on the pure electric ship is obtained, including: Acquire initial sensor data collected by the multiple sensors of each battery pack on the pure electric vessel through an initial sampling frequency; Determining the degree of deviation between the initial sensor data of each sensor and the overall data within the time window according to the distribution relationship of the initial sensor data within the window time; Adjusting the initial sampling frequency of the corresponding battery pack according to the degree of deviation to obtain a sampling frequency of the corresponding battery pack; The sensor data sets collected by the multiple sensors corresponding to the battery pack on the pure electric ship are obtained according to the collection frequency.

3. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 1, characterized in that: The determining the weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set includes: Preprocessing the data in the sensor data set to obtain a preprocessed data set; Calculating each sensor data in the preprocessed data set respectively to obtain the corresponding variance; The weight of each sensor is determined according to the proportion of each variance in all variances; the proportion is in direct proportion to the weight.

4. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 3, characterized in that: The preprocessing of the data in the sensor data set to obtain a preprocessed data set includes: Determining the median of data within a preset length in the sensor data set according to the sliding window; and replacing the original data value of the center point of the sliding window with the median to obtain a first data set; Using a time series linear interpolation method to complete the missing data in the first data set to obtain a second data set; The data of the multiple sensors in the second data set are standardized to obtain a preprocessed data set.

5. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 4, characterized in that: The method of using a time series linear interpolation method to fill in the missing data in the first data set to obtain a second data set includes: Identifying the data in the first data set to obtain all missing points and corresponding first timestamp positions; Determine, based on each missing point and the timestamp position, a previous valid data point and a second timestamp and a next valid data point and a third timestamp; Obtaining a linear interpolation weight according to the first timestamp, the second timestamp, and the third timestamp; Obtaining a missing value corresponding to the missing point according to the previous valid data point, the next valid data point, and the linear interpolation weight; The missing data in the first data set are completed according to all missing values ​​to obtain a second data set.

6. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 4, characterized in that: The step of normalizing the data of the multiple sensors in the first data set to obtain a preprocessed data set includes: Building a sliding window within a preset time based on the first data set; Calculating the sensor data in the sliding window to obtain a mean and a standard deviation; The sensor data at each moment is calculated according to the mean and the standard deviation to obtain a preprocessed data set.

7. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 1, characterized in that: The construction of the battery thermal runaway risk identification model includes: Acquire historical data sets of the multiple sensors of the pure electric ship based on a simulation platform; Performing data cleaning and feature fusion on the historical data set to obtain a historical feature vector; Dividing the historical feature vectors into a data set to obtain a training set, a validation set, and a test set; The initial battery thermal runaway risk identification model is trained on a CNN model, an RNN model, an LSTM model, and a Transformer model according to the training set, the validation set, and the test set to obtain corresponding training results; The training results of all models were compared by means of mean square error, mean absolute error and determination coefficient to obtain the optimal model, which was then determined as the battery thermal runaway risk identification model.

8. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 7, characterized in that: The acquiring of historical data sets of the multiple sensors of the pure electric ship based on the simulation platform includes: Determining multiple fault types based on the multiple sensors; the multiple sensors are respectively used to detect battery temperature, battery voltage, battery current, hydrogen concentration, and carbon dioxide concentration; Simulating at least one fault type according to the simulation platform to obtain all sensor data; All sensor data are labeled with risk levels to obtain a historical data set.

9. The method for early warning of thermal runaway risk of a pure electric ship battery according to claim 1, characterized in that: The performing risk warning on the pure electric ship according to the battery temperature sequence includes: Set multiple risk warning levels and corresponding predicted temperature ranges; determining a target risk warning level according to the battery temperature sequence and the predicted temperature range; A risk warning is performed on the pure electric ship according to the risk warning level.

10. A pure electric ship battery thermal runaway risk warning device, characterized in that: include: The data acquisition module is used to build a battery thermal runaway risk identification model and obtain sensor data sets detected by various sensors on pure electric ships; a weight determination module, configured to determine a weight of each sensor according to the fluctuation intensity of each sensor data in the sensor data set; a feature fusion module, configured to perform feature fusion on the data of the multiple sensors in the sensor data set according to the weights to obtain a comprehensive feature vector; A model prediction module is used to input the comprehensive feature vector into the thermal runaway risk identification model for prediction, obtain a battery temperature sequence, and issue a risk warning for the pure electric ship based on the battery temperature sequence.

Images