A method and apparatus for handling thermal runaway events in autonomous vehicles.
By deploying a multimodal sensor network and intelligent prediction model on autonomous vehicles, the problem of low accuracy in thermal runaway event identification in existing technologies has been solved, achieving higher prediction accuracy and flexibility, and making it applicable to vehicles with different computing power levels.
Patent Information
- Application Number
- CN202511926123.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-12-19
AI Technical Summary
Existing solutions for handling thermal runaway events in autonomous vehicles rely on single-point sensors, which have low recognition accuracy and are prone to missed or false alarms, lacking flexible and intelligent recognition methods.
A multimodal sensor network is employed, including temperature sensors inside and around the battery pack, pressure sensors in the powertrain compartment, smoke sensors in the passenger compartment, and gas sensors in the trunk. Data prediction is performed using SVM and LSTM models, and data accuracy is improved through Kalman filtering and time alignment. Three intelligent prediction modes are provided to adapt to vehicles with different computing power levels.
It improves the flexibility and accuracy of predicting thermal runaway events, reduces the probability of false alarms and missed alarms, adapts to autonomous vehicles with different computing power levels, and ensures safety and reliability.
Smart Images

Figure CN121492664B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for handling thermal runaway events in autonomous vehicles. Background Technology
[0002] With the maturity and development of autonomous driving technology, low-speed driverless vehicles for logistics delivery and driverless taxis for passenger transport have already been put into operation in some cities. Both types of vehicles mostly use batteries as their power source. If a thermal runaway event occurs in the vehicle's battery pack, it can not only cause damage to goods / vehicles but also potentially cause personal injury and even fire. Therefore, these two types of vehicles without dedicated drivers must be equipped with thermal runaway event handling devices or systems.
[0003] However, the survey revealed some problems with the thermal runaway event handling solutions for these two types of operating vehicles: 1) Traditional solutions mostly rely on single-point temperature sensors or smoke sensors. The event recognition accuracy of this single-sensor solution is not high, and it is easy to miss or false alarms; 2) Traditional solutions mostly rely on fixed thresholds for fixed logic judgments, lacking flexible intelligent recognition methods, which also easily leads to missed or false alarms. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, apparatus, electronic device, and computer-readable storage medium for handling thermal runaway events in autonomous vehicles. This invention employs a multimodal sensor network to construct a monitoring sensor network: first and second temperature sensors are deployed inside and around the battery pack; a pressure sensor is deployed in the powertrain compartment; a smoke sensor is deployed in the passenger compartment; and a gas sensor is deployed in the trunk. Sensor data from the monitoring sensor network is continuously collected; and time alignment and Kalman filtering are continuously performed on the most recent sensor data to obtain the corresponding sensor data array. Three intelligent prediction modes are provided to predict the probability of thermal runaway events for each obtained sensor data array: 1) First mode: prediction is performed by a first event prediction model based on an SVM model; 2) Second mode: prediction is performed by a first event prediction model based on an LSTM model; 3) Third mode: prediction is performed by a two-level linkage of the first and second event prediction models. When the predicted event probability exceeds a predetermined threshold, an alarm emergency is triggered locally on the vehicle side, alarm data is uploaded to a remote platform, and alarm information is sent to vehicle management personnel. This invention uses an intelligent model to make predictions based on multimodal perception data, which can improve prediction flexibility, prediction accuracy, and reduce the probability of false positives and false negatives. The three prediction modes of this invention actually correspond to three model configuration schemes. The first mode can be configured on low-speed unmanned vehicles with lower computing power, while the second and third modes can be configured on unmanned taxis with higher computing power. Through these three model configuration schemes, the vehicle's computing power and prediction accuracy can be taken into account, and the adaptability of model deployment can be improved.
[0005] To achieve the above objectives, a first aspect of the present invention provides a method for handling thermal runaway events in autonomous vehicles, the method comprising:
[0006] A monitoring sensor network for thermal runaway events is formed by a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and the sensor data of the monitoring sensor network is continuously collected and stored in the sensor dataset.
[0007] According to a preset data analysis frequency, the sensor data in the sensor dataset that is within the most recent first time period is periodically extracted to form a corresponding data acquisition array;
[0008] First, the time alignment of various sensor data in the acquired data array is performed by interpolation, and then the time-aligned sensor data is filtered by Kalman filtering to obtain the corresponding sensor data array.
[0009] The system identifies preset prediction modes. If the prediction mode is a first mode, a preset first event prediction model predicts the probability of an event corresponding to the sensor data array. If the prediction mode is a second mode, a preset second event prediction model predicts the probability of an event corresponding to the sensor data array. If the prediction mode is a third mode, the first and second event prediction models work together to predict the probability of an event corresponding to the sensor data array. The prediction modes include first, second, and third modes. The first event prediction model is a thermal runaway event probability prediction model based on an SVM model. The second event prediction model is a thermal runaway event probability prediction model based on an LSTM model. Both the first and second event prediction models are used to predict the probability of a thermal runaway event based on the sensor data array input to the model and output the corresponding event probability.
[0010] When the probability of the event exceeds a preset event probability threshold, an alarm emergency is triggered locally on the vehicle side, alarm data is uploaded to a remote platform, and alarm information is sent to vehicle management personnel.
[0011] Preferably, the first temperature sensor is used to sense the internal temperature of the battery pack; the second temperature sensor is used to sense the external temperature of the battery pack; the pressure sensor is used to sense the atmospheric pressure inside the power compartment; the smoke sensor is used to sense the concentration of smoke particles inside the passenger compartment; and the gas sensor is used to sense the concentration of carbon dioxide (CO2) and hydrogen (H2) gas inside the trunk.
[0012] The sensor dataset includes a first temperature sequence, a second temperature sequence, a first pressure sequence, a first smoke concentration sequence, and a first gas concentration sequence; the first temperature sequence corresponds to the first temperature sensor, the second temperature sequence corresponds to the second temperature sensor, the first pressure sequence corresponds to the pressure sensor, the first smoke concentration sequence corresponds to the smoke sensor, and the first gas concentration sequence corresponds to the gas sensor.
[0013] The first temperature sequence consists of multiple internal temperature data, including a timestamp and the battery's internal temperature; the second temperature sequence consists of multiple external temperature data, including a timestamp and the battery's external temperature; the first pressure sequence consists of multiple atmospheric pressure data, including a timestamp and an atmospheric pressure value; the first smoke concentration sequence consists of multiple smoke concentration data, including a timestamp and a smoke concentration value; the first gas concentration sequence consists of multiple gas concentration data, including a timestamp, a CO2 concentration value, and an H2 concentration value.
[0014] The data acquisition array consists of an internal temperature acquisition sequence, an external temperature acquisition sequence, an atmospheric pressure acquisition sequence, a smoke concentration acquisition sequence, a CO2 concentration acquisition sequence, and an H2 concentration acquisition sequence; all time periods of the acquisition data array are the most recent time periods corresponding to the first duration.
[0015] The sensor data array consists of temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F; temperature sequence A consists of multiple temperature a t The temperature sequence B is formed by sequentially sorting multiple temperatures b. t The pressure sequence C is formed by sequentially sorting multiple atmospheres c. t The smoke sequence D is formed by sequentially sorting multiple smoke concentrations d. t The gas sequence E is formed by sequentially sorting multiple CO2 concentrations e. t The gas sequence F is formed by sequentially sorting multiple H2 concentrations f. t The data is ordered sequentially; the six types of sequences in the sensor data array have the same sequence length and the sequence data are time-aligned; the time scale t is an integer, 1≤t≤N, N is the total number of moments in the most recent time period, N=first duration / τ, and the preset time interval τ is divisible by the first duration.
[0016] Preferably, the first event prediction model has been trained before use;
[0017] The first event prediction model is composed of a feature extraction layer and the SVM model connected sequentially;
[0018] The feature extraction layer is used to identify the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of each sequence in the sensor data array, and to form corresponding sequence feature vectors from the identification results; and to form corresponding sensor feature vectors by sequentially concatenating all the obtained sequence feature vectors; and to normalize the features of each dimension of the sensor feature vectors and send the processed sensor feature vectors to the SVM model;
[0019] The SVM model is used to predict the probability of thermal runaway based on the sensor feature vectors, obtain the corresponding probability of the event, and output it.
[0020] Specifically, the process of identifying the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of each sequence in the sensor data array and forming a corresponding sequence feature vector from the identification results involves: taking the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array as the corresponding current sequences; calculating the mean and standard deviation of the current sequences; calculating the slope of the current sequences using a linear regression method; identifying the maximum and minimum values of the current sequences; taking the absolute difference between the maximum and minimum values as the corresponding maximum difference value; identifying the 75th and 95th percentiles of the current sequences; and forming the corresponding sequence feature vector from the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of the current sequences.
[0021] Preferably, the second event prediction model has been trained before use;
[0022] The second event prediction model is composed of an embedding coding layer, the LSTM model, an attention layer, a fully connected network, and a prediction output layer connected in sequence.
[0023] The embedding coding layer is used to perform sequence data normalization processing on the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array to obtain corresponding normalized vectors A', B', C', D', E', and F'; and six normalized vectors are combined to form an N×6 embedding coding vector X, which is then sent to the LSTM model; the normalized vector A' is composed of N normalized codes a' t Composition; the normalized vector B' consists of N normalized codes b' t Composition; the normalized vector C' consists of N normalized codes c' t Composition; the normalized vector D' consists of N normalized codes d' t Composition; the normalized vector E' consists of N normalized codes e' t Composition; the normalized vector F' consists of N normalized codes f' t Composition; the embedding encoding vector X consists of N embedding codes x t Composition, the embedded encoding x t Consists of six corresponding normalized codes a' t b' t c' t ,d' t et '、f t 'composition;
[0024] The LSTM model is used to perform feature encoding processing on the embedded encoding vector X to obtain the corresponding encoded feature vector H, which is then sent to the attention layer. The shape of the encoded feature vector H is N×M, where M is a preset encoding feature dimension. The encoded feature vector H consists of N feature vectors h. t Composition, the feature vector h t It consists of M eigenvalues;
[0025] The attention layer is used to process each of the feature vectors h of the encoded feature vector H. t attention weight wa t Perform calculations; and based on the calculation results, apply the calculations to the N feature vectors h. t The corresponding feature vector Y is obtained by performing a weighted summation calculation and then sent to the fully connected network; the feature vector Y has a shape of N×M.
[0026] The attention weight wa t The calculation method is as follows:
[0027] q t =W q h t k t =W k h t v t =W v h t ,
[0028]
[0029] Among them, W q W k W v These are the model's pre-defined query weight parameter vector, key weight parameter vector, and value weight parameter vector; q t k t v t These are the corresponding query vector, key vector, and value vector; W1 and W2 are the model's pre-set first and second weight parameter vectors, respectively; tanh() is the tanh activation function; u t Score for attention;
[0030] The weighted summation calculation method corresponding to the feature vector Y is as follows:
[0031] The fully connected network is composed of one or more fully connected activation layers connected sequentially; each fully connected activation layer consists of one fully connected layer and one ReLU activation layer connected sequentially; the fully connected network is used to perform a one-dimensional feature vector mapping on the feature vector Y to obtain the corresponding feature vector Z, which is then sent to the prediction output layer; wherein, the vector shape of the feature vector Z is N. Z ×1, the preset vector length N Z N is a positive integer, 1 < N Z <N;
[0032] The prediction output layer performs regression calculation on the feature vector Z based on the built-in fully connected layer to obtain the corresponding feature value R; and then substitutes the feature value R into the Sigmoid function to calculate the corresponding probability of the event occurrence and outputs it.
[0033] The characteristic value R is calculated as follows:
[0034] R = W FC Z+b FC ;
[0035] Among them, W FC b FC The weight parameter vector and offset vector of the fully connected layer built into the prediction output layer are respectively.
[0036] The probability of the event occurring is calculated as follows:
[0037]
[0038] Preferably, the step of continuously collecting sensor data from the monitoring sensor network and storing the collected data in a sensor dataset specifically includes:
[0039] The internal temperature of the battery pack is periodically collected by the first temperature sensor at a preset first collection frequency, and the collected data is stored in the first temperature sequence of the sensor dataset.
[0040] The battery pack outer edge temperature is periodically collected by the second temperature sensor at a preset second collection frequency, and the collected data is stored in the second temperature sequence of the sensor dataset.
[0041] The atmospheric pressure inside the engine compartment is periodically collected by the pressure sensor at a preset third collection frequency, and the collected data is stored in the first pressure sequence of the sensor dataset.
[0042] The smoke sensor periodically collects data on the concentration of smoke particles inside the passenger compartment at a preset fourth collection frequency and stores the collected data in the first smoke concentration sequence of the sensor dataset.
[0043] At a preset fifth sampling frequency, the gas sensor periodically collects data on the CO2 and H2 gas concentrations inside the trunk and stores the collected data in the first gas concentration sequence of the sensor dataset.
[0044] Preferably, the step of periodically extracting sensor data from the sensor dataset within the most recent first time period according to a preset data analysis frequency to form a corresponding acquisition data array specifically includes:
[0045] According to the data analysis frequency, the current time is periodically used as the current end time, the time obtained by subtracting the first duration from the current end time is used as the current start time, and the time period composed of the current start time and the current end time is used as the corresponding recent time period.
[0046] The internal temperature acquisition sequence is composed of all the internal temperature data in the most recent time period in the first temperature sequence;
[0047] The outer edge temperature acquisition sequence is composed of all the outer edge temperature data in the most recent time period in the second temperature sequence;
[0048] The atmospheric pressure acquisition sequence is composed of all atmospheric pressure data in the most recent time period in the first pressure sequence;
[0049] The smoke concentration collection sequence is composed of all the smoke concentration data in the most recent time period in the first smoke concentration sequence;
[0050] All gas concentration data in the most recent time period in the first gas concentration sequence are recorded as the current sampling data; a corresponding CO2 concentration data is formed by the timestamp and CO2 concentration value of each current sampling data; a corresponding H2 concentration data is formed by the timestamp and H2 concentration value of each current sampling data; a corresponding CO2 concentration acquisition sequence is formed by all the CO2 concentration data obtained this time; and a corresponding H2 concentration acquisition sequence is formed by all the H2 concentration data obtained this time.
[0051] The internal temperature acquisition sequence, the outer edge temperature acquisition sequence, the atmospheric pressure acquisition sequence, the smoke concentration acquisition sequence, the CO2 concentration acquisition sequence, and the H2 concentration acquisition sequence obtained in this study constitute the corresponding acquisition data array.
[0052] Preferably, the step of first performing time alignment on various types of sensor data in the acquired data array using interpolation, and then filtering the time-aligned sensor data using Kalman filtering to obtain the corresponding sensor data array, specifically includes:
[0053] On the time axis of the most recent time period corresponding to the data acquisition array, N time scales t are set based on the time interval τ.
[0054] Using linear interpolation, the internal temperature, external temperature, atmospheric pressure, atmospheric pressure, smoke concentration, CO2 concentration, and H2 concentration values at each time scale t are interpolated and estimated based on the internal temperature acquisition sequence, the external temperature acquisition sequence, the atmospheric pressure acquisition sequence, the smoke concentration acquisition sequence, the CO2 concentration acquisition sequence, and the H2 concentration acquisition sequence. The estimation results are then used to form corresponding internal temperature estimation sequences, external temperature estimation sequences, atmospheric pressure estimation sequences, smoke concentration estimation sequences, CO2 concentration estimation sequences, and H2 concentration estimation sequences.
[0055] The internal temperature estimation sequence, the outer temperature estimation sequence, the atmospheric pressure estimation sequence, the smoke concentration estimation sequence, the CO2 concentration estimation sequence, and the H2 concentration estimation sequence are each subjected to Kalman filtering; and the six data sequences obtained after filtering are taken as the corresponding temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F;
[0056] The temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F obtained in this study form the corresponding sensor data array.
[0057] Preferably, the step of predicting the probability of occurrence of the event corresponding to the sensor data array by linking the first and second event prediction models at two levels specifically includes:
[0058] The sensor data array is input into the first event prediction model for prediction processing, and the event occurrence probability obtained from this prediction processing is used as the corresponding current probability; it is then identified whether the current probability is less than a preset first probability threshold or greater than a preset second probability threshold; if so, the current probability is output as the final event occurrence probability; if not, the sensor data array is input into the second event prediction model for prediction processing, and the event occurrence probability obtained from this prediction processing is output as the final event occurrence probability; wherein, 0 < first probability threshold < second probability threshold < 1.
[0059] Preferably, the step of performing local alarm emergency handling on the vehicle side, uploading alarm data to a remote platform, and sending alarm information to vehicle management personnel specifically includes:
[0060] The vehicle will simultaneously trigger audible and visual alarms and deceleration and parking operations on the local side; and upon completion of parking, the vehicle doors will be unlocked, passengers will be reminded to disembark, and cooling spray operations will be performed simultaneously.
[0061] The system obtains the current vehicle location information and battery status information through the vehicle's autonomous driving system or vehicle CAN bus; uses the preset unique vehicle identifier as the corresponding current vehicle identifier; and uploads the corresponding remote alarm data, composed of the current time, the current vehicle identifier, the vehicle location information, the battery status information, and the most recent data collection array, to the remote platform.
[0062] The alarm notification SMS, consisting of the current time, the current vehicle identifier, the vehicle location information, and the battery status information, is sent to the preset vehicle management personnel SMS notification interface.
[0063] A second aspect of the present invention provides an apparatus for implementing the method for handling thermal runaway events of autonomous vehicles as described in the first aspect above. The apparatus includes: a multimodal data acquisition module, an acquired data extraction module, a data alignment and filtering module, an event prediction module, and an event early warning module.
[0064] The multimodal data acquisition module is used to monitor a thermal runaway event by a monitoring sensor network consisting of a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and to continuously collect sensor data from the monitoring sensor network and store the collected data in a sensor dataset.
[0065] The data acquisition extraction module is used to periodically extract the sensor data in the sensor dataset that is in the most recent first time period according to a preset data analysis frequency to form a corresponding data acquisition array.
[0066] The data alignment and filtering module is used to first perform time alignment on various types of sensor data in the acquired data array using interpolation, and then filter the time-aligned sensor data using Kalman filtering to obtain the corresponding sensor data array.
[0067] The event prediction module is used to identify preset prediction modes. If the prediction mode is a first mode, a preset first event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is a second mode, a preset second event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is a third mode, the first and second event prediction models work together to predict the probability of the event corresponding to the sensor data array. The prediction modes include first, second, and third modes. The first event prediction model is a thermal runaway event probability prediction model based on an SVM model. The second event prediction model is a thermal runaway event probability prediction model based on an LSTM model. Both the first and second event prediction models are used to predict the probability of thermal runaway events based on the sensor data array input to the model and output the corresponding event probability.
[0068] The event warning module is used to perform local alarm emergency processing on the vehicle side, upload alarm data to the remote platform, and send alarm information to vehicle operation management personnel when the probability of the event exceeds a preset event probability threshold.
[0069] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;
[0070] The processor is used to couple with the memory, read and execute instructions in the memory to implement the steps of the method described in the first aspect above;
[0071] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.
[0072] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a computer, cause the computer to perform the instructions described in the first aspect.
[0073] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for handling thermal runaway events in autonomous vehicles. As described above, this invention employs a multimodal sensor network to construct a monitoring sensor network: first and second temperature sensors are deployed inside and around the battery pack; a pressure sensor is deployed in the powertrain compartment; a smoke sensor is deployed in the passenger compartment; and a gas sensor is deployed in the trunk. Sensor data from the monitoring sensor network is continuously collected; and time alignment and Kalman filtering are continuously performed on the most recent sensor data to obtain the corresponding sensor data array. Three intelligent prediction modes are provided to predict the probability of thermal runaway events for each obtained sensor data array: 1) First mode: prediction is performed by a first event prediction model based on an SVM model; 2) Second mode: prediction is performed by a first event prediction model based on an LSTM model; 3) Third mode: prediction is performed by a two-level linkage of the first and second event prediction models. When the predicted event probability exceeds a predetermined threshold, an alarm emergency is triggered locally on the vehicle side, alarm data is uploaded to a remote platform, and alarm information is sent to vehicle management personnel. The embodiments of this invention use intelligent models to make predictions based on multimodal perception data, which improves prediction flexibility, improves prediction accuracy, and reduces the probability of false positives and false negatives. The three prediction modes of the embodiments of this invention correspond to three model configuration schemes. The first mode can be configured on low-speed unmanned vehicles with low computing power, while the second and third modes can be configured on unmanned taxis with high computing power. Based on these three model configuration schemes, both the vehicle's computing power and prediction accuracy are taken into account, while the flexibility and adaptability of model deployment are improved. Attached Figure Description
[0074] Figure 1 This is a schematic diagram of a method for handling thermal runaway events in an autonomous vehicle according to Embodiment 1 of the present invention;
[0075] Figure 2 This is a schematic diagram of the structure of the first event prediction model provided in Embodiment 1 of the present invention;
[0076] Figure 3 This is a schematic diagram of the structure of the second event prediction model provided in Embodiment 1 of the present invention;
[0077] Figure 4 This is a module structure diagram of a device for handling thermal runaway events in an autonomous vehicle, provided in Embodiment 2 of the present invention.
[0078] Figure 5 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation
[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0080] Embodiment 1 of the present invention provides a method for handling thermal runaway events in autonomous vehicles, such as... Figure 1 The schematic diagram of a method for handling thermal runaway events in an autonomous vehicle according to Embodiment 1 of the present invention is shown, including the following main steps:
[0081] Step 1: A monitoring sensor network for thermal runaway events is formed by a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and the sensor data of the monitoring sensor network is continuously collected and stored in the sensor dataset.
[0082] Specifically, it includes: Step 11, which consists of a monitoring sensor network for thermal runaway events, consisting of a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk.
[0083] Here, in this embodiment of the invention, the first temperature sensor is used to sense the internal temperature of the battery pack; the second temperature sensor is used to sense the external temperature of the battery pack; the pressure sensor is used to sense the atmospheric pressure inside the power compartment; the smoke sensor is used to sense the concentration of smoke particles inside the passenger compartment; and the gas sensor is used to sense the concentration of carbon dioxide (CO2) and hydrogen (H2) gas inside the trunk.
[0084] Step 12, and continuously collect sensor data from the monitoring sensor network and store the collected data in the sensor dataset;
[0085] Here, the sensor dataset of this embodiment of the invention includes a first temperature sequence, a second temperature sequence, a first pressure sequence, a first smoke concentration sequence, and a first gas concentration sequence; the first temperature sequence corresponds to a first temperature sensor, the second temperature sequence corresponds to a second temperature sensor, the first pressure sequence corresponds to a pressure sensor, the first smoke concentration sequence corresponds to a smoke sensor, and the first gas concentration sequence corresponds to a gas sensor.
[0086] The first temperature sequence consists of multiple internal temperature data points, including timestamps and battery internal temperatures; the second temperature sequence consists of multiple external temperature data points, including timestamps and battery external temperatures; the first pressure sequence consists of multiple atmospheric pressure data points, including timestamps and atmospheric pressure values; the first smoke concentration sequence consists of multiple smoke concentration data points, including timestamps and smoke concentration values; the first gas concentration sequence consists of multiple gas concentration data points, including timestamps, CO2 concentration values, and H2 concentration values.
[0087] The current step 12 specifically includes:
[0088] Step 121: Collect data on the internal temperature of the battery pack periodically using the first temperature sensor at a preset first collection frequency, and store the collected data in the first temperature sequence of the sensor dataset.
[0089] Step 122: At a preset second acquisition frequency, periodically collect data on the temperature of the outer edge of the battery pack using the second temperature sensor and store the collected data in the second temperature sequence of the sensor dataset.
[0090] Step 123: At a preset third acquisition frequency, the atmospheric pressure inside the power compartment is periodically collected by the pressure sensor and the collected data is stored in the first pressure sequence of the sensor dataset.
[0091] Step 124: At a preset fourth acquisition frequency, the smoke particle concentration inside the crew cabin is periodically collected by the smoke sensor and the collected data is stored in the first smoke concentration sequence of the sensor dataset.
[0092] Step 125: At the preset fifth acquisition frequency, the gas sensor periodically collects data on the CO2 gas concentration and H2 gas concentration inside the trunk and stores the collected data into the first gas concentration sequence of the sensor dataset.
[0093] Here, the first, second, third, fourth, and fifth acquisition frequencies in this embodiment of the invention are five preset time frequency parameters; typically, these five time frequency parameters are partially or completely different, which means that the sequence data of the five sequences in the sensor dataset will not be fully time-aligned.
[0094] Step 2: Periodically extract the sensor data that is in the most recent first time period from the sensor dataset according to the preset data analysis frequency to form the corresponding acquisition data array.
[0095] Here, the data analysis frequency in this embodiment of the invention is a pre-set time frequency parameter. This parameter can be dynamically set based on the application scenario or application requirements. The higher the data analysis frequency, the stronger the real-time performance.
[0096] The data acquisition array in this embodiment of the invention consists of an internal temperature acquisition sequence, an external temperature acquisition sequence, an atmospheric pressure acquisition sequence, a smoke concentration acquisition sequence, a CO2 concentration acquisition sequence, and an H2 concentration acquisition sequence. All time periods in the data acquisition array are the most recent time periods corresponding to the most recent first duration. The first duration is a pre-set time length parameter, such as 30 seconds, 1 minute, 5 minutes, etc.
[0097] The current step 2 specifically includes:
[0098] Step 21: Periodically use the current time as the current end time according to the data analysis frequency, subtract the first duration from the current end time as the current start time, and use the time period composed of the current start time and the current end time as the corresponding most recent time period;
[0099] Step 22, and the corresponding internal temperature acquisition sequence is composed of all the internal temperature data in the most recent time period in the first temperature sequence;
[0100] Step 23, and the corresponding outer edge temperature acquisition sequence is composed of all the outer edge temperature data in the most recent time period in the second temperature sequence;
[0101] Step 24, and the corresponding atmospheric pressure acquisition sequence is composed of all atmospheric pressure data in the most recent time period in the first pressure sequence;
[0102] Step 25, and the corresponding smoke concentration acquisition sequence is composed of all the smoke concentration data in the most recent time period in the first smoke concentration sequence;
[0103] Step 26: Record all gas concentration data in the most recent time period in the first gas concentration sequence as the current sampling data; combine the timestamp and CO2 concentration value of each current sampling data to form a corresponding CO2 concentration data; combine the timestamp and H2 concentration value of each current sampling data to form a corresponding H2 concentration data; combine all the CO2 concentration data obtained this time to form a corresponding CO2 concentration acquisition sequence; and combine all the H2 concentration data obtained this time to form a corresponding H2 concentration acquisition sequence.
[0104] Step 27, and the corresponding acquisition data array is composed of the internal temperature acquisition sequence, the outer edge temperature acquisition sequence, the atmospheric pressure acquisition sequence, the smoke concentration acquisition sequence, the CO2 concentration acquisition sequence, and the H2 concentration acquisition sequence obtained in this step.
[0105] Step 3: First, time alignment of various sensor data in the data acquisition array is performed using interpolation, and then Kalman filtering is used to filter the time-aligned sensor data to obtain the corresponding sensor data array.
[0106] Here, the sensor data array in this embodiment of the invention consists of temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F. Temperature sequence A comprises multiple temperature a... t The temperature sequence B is formed by sequentially sorting the data; the temperature sequence B consists of multiple temperature b. t The pressure sequence C is composed of multiple atmospheres c. t The smoke sequence D is formed by sequentially sorting the smoke; the smoke sequence D consists of multiple smoke concentrations d. t The gas sequence E is formed by sequentially sorting the gases; the gas sequence E consists of multiple CO2 concentrations e. t The gas sequence F is formed by sequentially sorting the gases; the gas sequence F consists of multiple H2 concentrations f. t The data is ordered sequentially. The six types of sequences in the sensor data array have the same sequence length and the sequence data are time-aligned; the time scale t is an integer, 1≤t≤N, where N is the total number of moments in the most recent time period, N = first duration / τ, and the preset time interval τ is divisible by the first duration.
[0107] The current step 3 specifically includes:
[0108] Step 31: On the time axis of the most recent period corresponding to the data collection array, set N corresponding time scales t based on the time interval τ;
[0109] Step 32: Using linear interpolation, estimate the internal temperature, external temperature, atmospheric pressure, smoke concentration, CO2 concentration, and H2 concentration values at each time scale t based on the internal temperature acquisition sequence, external temperature acquisition sequence, atmospheric pressure acquisition sequence, smoke concentration acquisition sequence, CO2 concentration acquisition sequence, and H2 concentration acquisition sequence. Then, use the estimation results to form the corresponding internal temperature estimation sequence, external temperature estimation sequence, atmospheric pressure estimation sequence, smoke concentration estimation sequence, CO2 concentration estimation sequence, and H2 concentration estimation sequence.
[0110] Step 33: Perform Kalman filtering on the internal temperature estimation sequence, external temperature estimation sequence, atmospheric pressure estimation sequence, smoke concentration estimation sequence, CO2 concentration estimation sequence, and H2 concentration estimation sequence respectively; and use the six data sequences obtained after filtering as the corresponding temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F.
[0111] Step 34, and the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E and gas sequence F obtained this time are used to form the corresponding sensor data array.
[0112] Step 4: Identify the preset prediction mode; if the prediction mode is the first mode, the preset first event prediction model predicts the probability of the event corresponding to the sensor data array; if the prediction mode is the second mode, the preset second event prediction model predicts the probability of the event corresponding to the sensor data array; if the prediction mode is the third mode, the first and second event prediction models work together to predict the probability of the event corresponding to the sensor data array.
[0113] Specifically, this includes: Step 41, identifying the preset prediction pattern;
[0114] The prediction models include the first, second, and third models;
[0115] It should be noted that the first event prediction model in this embodiment of the invention is a thermal runaway event probability prediction model implemented based on the SVM model; the second event prediction model in this embodiment of the invention is a thermal runaway event probability prediction model implemented based on the LSTM model; both the first and second event prediction models are used to predict the probability of thermal runaway events based on the sensor data array input to the model and output the corresponding event probability.
[0116] It should also be noted that both the first and second event prediction models have been trained before use. During training, these two types of models can be trained using the same dataset or different datasets can be constructed for each model and trained separately. Regardless of whether the same dataset is used for training, the positive and negative sample collection methods for the datasets used by the first and second event prediction models are the same. Positive samples refer to data samples where thermal runaway events have occurred, and negative samples refer to data samples where thermal runaway events have not occurred.
[0117] Positive samples are primarily collected through experimental methods, supplemented by methods based on real-world incidents.
[0118] When acquiring positive samples through experimental data collection, a denser multimodal sensor network than that used in mass-produced vehicles is first deployed in a test environment. Then, multiple vehicle-mounted individual batteries or battery packs are subjected to overcharging, over-discharging, heating, short-circuiting, needle penetration, and compression operations to induce thermal runaway events. The multimodal sensor data collected before and after each event, within specified timeframes (e.g., 30 seconds, 60 seconds, 3 minutes, or 5 minutes), is converted into sensor data arrays, and each of these arrays is assigned a corresponding event probability label of 1. Here, each sensor data array obtained from the test environment and its corresponding event probability label constitutes a positive sample. Positive samples obtained through experimental methods can also be expanded by orders of magnitude through data augmentation.
[0119] When acquiring positive samples through real-world accident data collection, the multimodal sensor data synchronized in the cloud for low-speed autonomous vehicles or driverless taxis that experienced thermal runaway events is examined. If the time point of a thermal runaway event for a certain vehicle is determined, and the vehicle synchronized a segment of sensor data to the cloud before the accident time point, and this sensor data contains some or all of the following data types: battery internal temperature, battery outer edge temperature, vehicle interior air pressure, vehicle interior smoke concentration, vehicle interior CO2 concentration, and vehicle interior H2 concentration, then the multimodal sensor data within a specified time period (e.g., 30 seconds, 60 seconds, 3 minutes, or 5 minutes) before the event can be converted into a sensor data array, and the corresponding event occurrence probability label for this type of sensor data array is set to 1. Here, each sensor data array obtained from the accident data and its corresponding event occurrence probability label can form a corresponding positive sample.
[0120] Negative sample collection is accomplished by collecting normal operating data from the vehicles. Simply put, it involves deploying a monitoring sensor network on a regularly operating low-speed unmanned vehicle or driverless taxi, according to the deployment method described in this invention embodiment. The sensor data from this monitoring sensor network is then continuously collected and stored. The stored multimodal sensor data is converted into a large number of sensor data arrays at specified intervals (e.g., 30 seconds, 60 seconds, 3 minutes, or 5 minutes), and the corresponding event probability label for these sensor data arrays is set to 0. Here, each sensor data array obtained from the normal operating data and its corresponding event probability label constitute a negative sample.
[0121] When training the model based on a dataset consisting of positive and negative samples, the dataset is divided into a training set and an evaluation set. The model is trained using the training set, and the training effect is evaluated using the evaluation set. When training the first event prediction model based on the SVM model, conventional SVM training algorithms (such as the SMO algorithm) are used to optimize the model parameters. When training the second event prediction model based on the LSTM model, L1 loss function, L2 loss function, or cross-entropy loss are used to optimize the model parameters. When evaluating the first and second event prediction models, conventional F1 and / or AUC metrics can be used for evaluation. It should be noted that when evaluating the model, a binary classification judgment of whether a thermal runaway event has occurred or not needs to be made based on the model's predicted probability. In this case, the event probability threshold used in step 5 below is used as the binary classification watershed threshold. Exceeding this threshold is considered as a thermal runaway event occurring, and not exceeding this threshold is considered as a thermal runaway event not occurring.
[0122] Step 42: If the prediction mode is the first mode, then the probability of the event corresponding to the sensor data array is predicted by the preset first event prediction model.
[0123] like Figure 2 As shown in the schematic diagram of the structure of the first event prediction model provided in Embodiment 1 of the present invention, the first event prediction model is composed of a feature extraction layer and an SVM model connected sequentially;
[0124] The feature extraction layer in this embodiment of the invention is used to identify the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of each sequence in the sensor data array, and to form the corresponding sequence feature vector from the identification results; and to form the corresponding sensor feature vector by sequentially concatenating all the obtained sequence feature vectors; and to normalize the features of each dimension of the sensor feature vector and send the processed sensor feature vector to the SVM model;
[0125] Specifically, the feature extraction layer is used to identify the mean, standard deviation, slope, maximum value, minimum value, maximum difference, 75th percentile, and 95th percentile of each sequence in the sensor data array and to form the corresponding sequence feature vector from the identification results. The specific processing steps include: taking the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array as the corresponding current sequences; calculating the mean and standard deviation of the current sequences; calculating the slope of the current sequences using linear regression; identifying the maximum and minimum values of the current sequences; taking the absolute difference between the maximum and minimum values as the corresponding maximum difference; identifying the 75th and 95th percentiles of the current sequences; and forming the corresponding sequence feature vector from the mean, standard deviation, slope, maximum value, minimum value, maximum difference, 75th percentile, and 95th percentile of the current sequences.
[0126] The SVM model in this embodiment of the invention is used to predict the probability of thermal runaway based on sensor feature vectors, obtain the corresponding event occurrence probability, and output it.
[0127] Step 43: If the prediction mode is the second mode, the probability of the event corresponding to the sensor data array is predicted by the preset second event prediction model.
[0128] like Figure 3 As shown in the structural diagram of the second event prediction model provided in Embodiment 1 of the present invention, the second event prediction model is composed of an embedding coding layer, an LSTM model, an attention layer, a fully connected network, and a prediction output layer connected in sequence.
[0129] The embedding coding layer in this embodiment of the invention is used to perform sequence data normalization processing on the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array to obtain corresponding normalized vectors A', B', C', D', E', and F'; and six normalized vectors are combined to form an N×6 embedding coding vector X, which is then sent to the LSTM model; the normalized vector A' is composed of N normalized codes a' t composition;
[0130] Here, the normalized vector B' in this embodiment of the invention consists of N normalized codes b' t Composition; the normalized vector C' consists of N normalized codes c' t Composition; the normalized vector D' consists of N normalized codes d' t Composition; the normalized vector E' consists of N normalized codes e' t Composition; the normalized vector F' consists of N normalized codes f' t Composition; the embedding encoding vector X consists of N embedding codes x t Composition, embedding encoding x t Consists of six corresponding normalized codes a' t b t '、c t '、d' t e t '、f t 'composition;
[0131] The LSTM model in this embodiment of the invention is used to perform feature encoding processing on the embedded encoding vector X to obtain the corresponding encoded feature vector H and send it to the attention layer;
[0132] Here, in this embodiment of the invention, the encoded feature vector H has a shape of N×M, where M is a preset encoded feature dimension; the encoded feature vector H consists of N feature vectors h t Composition, eigenvector h t It consists of M eigenvalues;
[0133] The attention layer in this embodiment of the invention is used to process each feature vector h of the encoded feature vector H. t attention weight wa t Perform calculations; and based on the calculation results, apply the calculations to the N feature vectors h. t The corresponding feature vector Y is obtained by performing a weighted summation calculation and then sent to the fully connected network.
[0134] Here, in the embodiment of the present invention, the attention weight wa t The calculation method is as follows:
[0135] qt =W q h t k t =W k h t v t =W v h t ,
[0136]
[0137] Among them, W q W k W v These are the model's pre-defined query weight parameter vector, key weight parameter vector, and value weight parameter vector; q t k t v t These are the corresponding query vector, key vector, and value vector; W1 and W2 are the model's pre-set first and second weight parameter vectors, respectively; tanh() is the tanh activation function; u t Score for attention;
[0138] The weighted summation calculation method corresponding to the feature vector Y in this embodiment of the invention is as follows:
[0139]
[0140] In this embodiment of the invention, the shape of the feature vector Y is N×M;
[0141] The fully connected network in this embodiment of the invention is composed of one or more fully connected activation layers connected sequentially; each fully connected activation layer is composed of a fully connected layer and a ReLU activation layer connected sequentially; the fully connected network is used to perform one-dimensional feature vector mapping on feature vector Y to obtain the corresponding feature vector Z and send it to the prediction output layer;
[0142] Here, in this embodiment of the invention, the feature vector Z has a vector shape of N. Z ×1, the preset vector length N Z N is a positive integer, 1 < N Z <N;
[0143] In this embodiment of the invention, the prediction output layer performs regression calculation on the feature vector Z based on the built-in fully connected layer to obtain the corresponding feature value R; and then substitutes the feature value R into the Sigmoid function to calculate the corresponding event occurrence probability and outputs it.
[0144] Here, the characteristic value R in this embodiment of the invention is calculated as follows:
[0145] R = W FC Z+b FC ;
[0146] Among them, W FC b FC Predict the weight parameter vector and offset vector of the fully connected layer built into the output layer, respectively;
[0147] The method for calculating the probability of event occurrence in the predicted output layer in this embodiment of the invention is as follows:
[0148]
[0149] Step 44: If the prediction mode is the third mode, the probability of the event corresponding to the sensor data array is predicted by the two-level linkage of the first and second event prediction models.
[0150] Specifically, this includes: inputting the sensor data array into the first event prediction model for prediction processing and using the event occurrence probability obtained from this prediction processing as the corresponding current probability; identifying whether the current probability is less than a preset first probability threshold or greater than a preset second probability threshold; if so, outputting the current probability as the final event occurrence probability; if not, inputting the sensor data array into the second event prediction model for prediction processing and using the event occurrence probability obtained from this prediction processing as the final event occurrence probability output.
[0151] Here, the first and second probability thresholds in this embodiment of the invention are two pre-set threshold parameters, and 0 < first probability threshold < second probability threshold < 1.
[0152] Step 5: When the probability of an event occurring exceeds the preset event probability threshold, perform local alarm emergency handling on the vehicle side, upload alarm data to the remote platform, and send alarm information to the vehicle operation management personnel.
[0153] The event probability threshold is a pre-set threshold parameter.
[0154] Specifically, this includes: Step 51, synchronously performing audible and visual alarms and deceleration and parking operations on the vehicle side; and simultaneously performing door unlocking, passenger disembarkation reminder, and cooling spray operations when parking is completed.
[0155] It should be noted that the specific spraying strategy of the cooling spraying operation in the embodiments of the present invention can be customized based on the application scenario and / or application requirements; for example, a parallel spraying strategy that sprays multiple locations simultaneously can be adopted, or a directional spraying strategy for high-temperature locations can be adopted; it should also be noted that the cooling spraying liquid selected in the cooling spraying operation of the embodiments of the present invention must be suitable for high-temperature and thermal runaway battery scenarios, and cooling spraying materials that will produce a conductive effect and directly or indirectly cause the battery to heat up cannot be used.
[0156] Step 52: Obtain the current vehicle location information and battery status information through the vehicle's autonomous driving system or vehicle CAN bus; use the preset vehicle unique identifier as the corresponding current vehicle identifier; and upload the corresponding remote alarm data, composed of the current time, current vehicle identifier, vehicle location information, battery status information, and the most recently collected data array, to the remote platform.
[0157] Step 53: The alarm notification SMS, consisting of the current time, current vehicle identifier, vehicle location information, and battery status information, is sent to the preset vehicle management personnel SMS notification interface.
[0158] Figure 4 This is a module structure diagram of a device for handling thermal runaway events in an autonomous vehicle according to Embodiment 2 of the present invention. This device can be a terminal device or server implementing the aforementioned method embodiments, or it can be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiments. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 4 As shown, the device for handling thermal runaway events of autonomous vehicles provided in Embodiment 2 of the present invention includes: a multimodal data acquisition module 201, a data extraction module 202, a data alignment and filtering module 203, an event prediction module 204, and an event early warning module 205.
[0159] The multimodal data acquisition module 201 is used to monitor a thermal runaway event by a monitoring sensor network consisting of a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and to continuously collect sensor data from the monitoring sensor network and store the collected data in a sensor dataset.
[0160] The data acquisition and extraction module 202 is used to periodically extract the sensor data that is concentrated in the most recent first time period according to the preset data analysis frequency and form a corresponding data acquisition array.
[0161] The data alignment and filtering module 203 is used to first perform time alignment on various types of sensor data in the acquired data array using interpolation, and then filter the time-aligned sensor data using Kalman filtering to obtain the corresponding sensor data array.
[0162] The event prediction module 204 is used to identify preset prediction modes. If the prediction mode is the first mode, the preset first event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is the second mode, the preset second event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is the third mode, the first and second event prediction models work together to predict the probability of the event corresponding to the sensor data array. The prediction modes include the first, second, and third modes. The first event prediction model is a thermal runaway event probability prediction model based on the SVM model. The second event prediction model is a thermal runaway event probability prediction model based on the LSTM model. Both the first and second event prediction models are used to predict the probability of thermal runaway events based on the sensor data array input to the model and output the corresponding event probability.
[0163] The event warning module 205 is used to perform local alarm emergency processing on the vehicle side, upload alarm data to the remote platform, and send alarm information to vehicle operation management personnel when the probability of an event exceeds the preset event probability threshold.
[0164] The present invention provides a device for handling thermal runaway events of autonomous vehicles, which can execute the method steps in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.
[0165] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, the data acquisition and extraction module can be a separate processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.
[0166] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-on-a-Chip (SOC).
[0167] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the foregoing method embodiments are generated. The computer described above can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The aforementioned computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the aforementioned computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, Bluetooth, microwave, etc.) means. The aforementioned computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).
[0168] Figure 5 This is a schematic diagram of an electronic device provided in Embodiment 3 of the present invention. This electronic device can be a terminal device or server implementing the methods of the aforementioned embodiments, or it can be a terminal device or server connected to the aforementioned terminal device or server implementing the methods of the aforementioned embodiments. Figure 5As shown, the electronic device may include: a processor 301 (e.g., CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transmission and reception operations of the transceiver 303. The memory 302 may store various instructions for performing various processing functions and implementing the processing steps described in the foregoing embodiments. Preferably, the electronic device involved in the embodiments of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The communication port 306 is used for communication between the electronic device and other peripherals.
[0169] exist Figure 5 The system bus 305 mentioned can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write libraries, and read-only libraries). Memory may include Random Access Memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.
[0170] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0171] It should be noted that the embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when run on a computer, cause the computer to perform the methods and processes provided in the above embodiments.
[0172] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for handling thermal runaway events in autonomous vehicles. As described above, this invention employs a multimodal sensor network to construct a monitoring sensor network: first and second temperature sensors are deployed inside and around the battery pack; a pressure sensor is deployed in the powertrain compartment; a smoke sensor is deployed in the passenger compartment; and a gas sensor is deployed in the trunk. Sensor data from the monitoring sensor network is continuously collected; and time alignment and Kalman filtering are continuously performed on the most recent sensor data to obtain the corresponding sensor data array. Three intelligent prediction modes are provided to predict the probability of thermal runaway events for each obtained sensor data array: 1) First mode: prediction is performed by a first event prediction model based on an SVM model; 2) Second mode: prediction is performed by a first event prediction model based on an LSTM model; 3) Third mode: prediction is performed by a two-level linkage of the first and second event prediction models. When the predicted event probability exceeds a predetermined threshold, an alarm emergency is triggered locally on the vehicle side, alarm data is uploaded to a remote platform, and alarm information is sent to vehicle management personnel. The embodiments of this invention use intelligent models to make predictions based on multimodal perception data, which improves prediction flexibility, improves prediction accuracy, and reduces the probability of false positives and false negatives. The three prediction modes of the embodiments of this invention correspond to three model configuration schemes. The first mode can be configured on low-speed unmanned vehicles with low computing power, while the second and third modes can be configured on unmanned taxis with high computing power. Based on these three model configuration schemes, both the vehicle's computing power and prediction accuracy are taken into account, while the flexibility and adaptability of model deployment are improved.
[0173] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0174] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for handling thermal runaway events in autonomous vehicles, characterized in that, The method includes: A monitoring sensor network for thermal runaway events is formed by a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and the sensor data of the monitoring sensor network is continuously collected and stored in the sensor dataset. According to a preset data analysis frequency, the sensor data in the sensor dataset that is within the most recent first time period is periodically extracted to form a corresponding data acquisition array; First, the time alignment of various sensor data in the acquired data array is performed by interpolation, and then the time-aligned sensor data is filtered by Kalman filtering to obtain the corresponding sensor data array. The system identifies preset prediction modes. If the prediction mode is a first mode, a preset first event prediction model predicts the probability of an event corresponding to the sensor data array. If the prediction mode is a second mode, a preset second event prediction model predicts the probability of an event corresponding to the sensor data array. If the prediction mode is a third mode, the first and second event prediction models work together to predict the probability of an event corresponding to the sensor data array. The prediction modes include first, second, and third modes. The first event prediction model is a thermal runaway event probability prediction model based on an SVM model. The second event prediction model is a thermal runaway event probability prediction model based on an LSTM model. Both the first and second event prediction models are used to predict the probability of a thermal runaway event based on the sensor data array input to the model and output the corresponding event probability. When the probability of the event exceeds a preset event probability threshold, an alarm emergency is triggered locally on the vehicle side, alarm data is uploaded to a remote platform, and alarm information is sent to vehicle management personnel. Specifically, the prediction of the probability of occurrence of the event corresponding to the sensor data array by the two-level linkage of the first and second event prediction models includes: The sensor data array is input into the first event prediction model for prediction processing, and the event occurrence probability obtained from this prediction processing is used as the corresponding current probability; it is then identified whether the current probability is less than a preset first probability threshold or greater than a preset second probability threshold; if so, the current probability is output as the final event occurrence probability; if not, the sensor data array is input into the second event prediction model for prediction processing, and the event occurrence probability obtained from this prediction processing is output as the final event occurrence probability; wherein, 0 < first probability threshold < second probability threshold < 1.
2. The method for handling thermal runaway events in autonomous vehicles according to claim 1, characterized in that, The first temperature sensor is used to sense the internal temperature of the battery pack; the second temperature sensor is used to sense the external temperature of the battery pack; the pressure sensor is used to sense the atmospheric pressure inside the power compartment; the smoke sensor is used to sense the concentration of smoke particles inside the passenger compartment; and the gas sensor is used to sense the concentration of carbon dioxide (CO2) and hydrogen (H2) gas inside the trunk. The sensor dataset includes a first temperature sequence, a second temperature sequence, a first pressure sequence, a first smoke concentration sequence, and a first gas concentration sequence; the first temperature sequence corresponds to the first temperature sensor, the second temperature sequence corresponds to the second temperature sensor, the first pressure sequence corresponds to the pressure sensor, the first smoke concentration sequence corresponds to the smoke sensor, and the first gas concentration sequence corresponds to the gas sensor. The first temperature sequence consists of multiple internal temperature data, which include a timestamp and the internal temperature of the battery; the second temperature sequence consists of multiple external temperature data, which include a timestamp and the external temperature of the battery; the first pressure sequence consists of multiple atmospheric pressure data, which include a timestamp and the atmospheric pressure value. The first smoke concentration sequence consists of multiple smoke concentration data, which include timestamps and smoke concentration values. The first gas concentration sequence consists of multiple gas concentration data, including timestamps, CO2 concentration values, and H2 concentration values. The data acquisition array consists of an internal temperature acquisition sequence, an external temperature acquisition sequence, an atmospheric pressure acquisition sequence, a smoke concentration acquisition sequence, a CO2 concentration acquisition sequence, and an H2 concentration acquisition sequence; all time periods of the acquisition data array are the most recent time periods corresponding to the first duration. The sensor data array consists of temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F; temperature sequence A consists of multiple temperature a t The temperature sequence B is formed by sequentially sorting multiple temperatures b. t The pressure sequence C is formed by sequentially sorting multiple atmospheres c. t The smoke sequence D is formed by sequentially sorting multiple smoke concentrations d. t The gas sequence E is formed by sequentially sorting multiple CO2 concentrations e. t The gas sequence F is formed by sequentially sorting multiple H2 concentrations f. t The data is ordered sequentially; the six types of sequences in the sensor data array have the same sequence length and the sequence data are time-aligned; the time scale t is an integer, 1≤t≤N, N is the total number of moments in the most recent time period, N=first duration / τ, and the preset time interval τ is divisible by the first duration.
3. The method for handling thermal runaway events in autonomous vehicles according to claim 2, characterized in that, The first event prediction model has been trained before it is used. The first event prediction model is composed of a feature extraction layer and the SVM model connected sequentially; The feature extraction layer is used to identify the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of each sequence in the sensor data array, and to form corresponding sequence feature vectors from the identification results; and to form corresponding sensor feature vectors by sequentially concatenating all the obtained sequence feature vectors; and to normalize the features of each dimension of the sensor feature vectors and send the processed sensor feature vectors to the SVM model; The SVM model is used to predict the probability of thermal runaway based on the sensor feature vectors, obtain the corresponding probability of the event, and output it. Specifically, the process of identifying the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of each sequence in the sensor data array and forming a corresponding sequence feature vector from the identification results involves: taking the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array as the corresponding current sequences; calculating the mean and standard deviation of the current sequences; calculating the slope of the current sequences using a linear regression method; identifying the maximum and minimum values of the current sequences; taking the absolute difference between the maximum and minimum values as the corresponding maximum difference value; identifying the 75th and 95th percentiles of the current sequences; and forming the corresponding sequence feature vector from the mean, standard deviation, slope, maximum value, minimum value, maximum difference value, 75th percentile, and 95th percentile of the current sequences.
4. The method for handling thermal runaway events in autonomous vehicles according to claim 2, characterized in that, The second event prediction model has been trained before use; The second event prediction model is composed of an embedding coding layer, the LSTM model, an attention layer, a fully connected network, and a prediction output layer connected in sequence. The embedded coding layer is used to perform sequence data normalization processing on the temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F of the sensor data array to obtain the corresponding normalized vector A. ’ Normalized vector B ’ Normalized vector C ’ Normalized vector D ’ Normalized vector E ’ and normalized vector F ’ The embedded encoding vector X, consisting of six normalized vectors, is composed of an N×6 shape and sent to the LSTM model; the normalized vector A... ’ Consists of N normalized codes Composition; the normalized vector B ’ Consists of N normalized codes Composition; the normalized vector C ’ Consists of N normalized codes Composition; the normalized vector D ’ Consists of N normalized codes Composition; the normalized vector E ’ Consists of N normalized codes Composition; the normalized vector F ’ Consists of N normalized codes Composition; the embedding encoding vector X consists of N embedding codes x t Composition, the embedded encoding x t Consisting of six corresponding normalized codes , , , , , composition; The LSTM model is used to perform feature encoding processing on the embedded encoding vector X to obtain the corresponding encoded feature vector H, which is then sent to the attention layer. The vector shape of the encoded feature vector H is N×M, where M is a preset encoding feature dimension. The encoded feature vector H consists of N feature vectors h. t Composition, the feature vector h t It consists of M eigenvalues; The attention layer is used to process each of the feature vectors h of the encoded feature vector H. t Attention weights wa t Perform calculations; and based on the calculation results, apply the calculations to the N feature vectors h. t The corresponding feature vector Y is obtained by performing a weighted summation calculation and then sent to the fully connected network; the feature vector Y has a shape of N×M. The attention weight wa t The calculation method is as follows: , , , , ; Among them, W q W k W v These are the model's pre-defined query weight parameter vector, key weight parameter vector, and value weight parameter vector; q t k t v t These are the corresponding query vector, key vector, and value vector; W1 and W2 are the model's pre-set first and second weight parameter vectors, respectively; tanh() is the tanh activation function; u t Score for attention; The weighted summation calculation method corresponding to the feature vector Y is as follows: ; The fully connected network is composed of one or more fully connected activation layers connected sequentially; each fully connected activation layer consists of one fully connected layer and one ReLU activation layer connected sequentially; the fully connected network is used to perform a one-dimensional feature vector mapping on the feature vector Y to obtain the corresponding feature vector Z, which is then sent to the prediction output layer; wherein, the vector shape of the feature vector Z is N. Z ×1, the preset vector length N Z N is a positive integer, 1 < N Z <N; The prediction output layer performs regression calculation on the feature vector Z based on the built-in fully connected layer to obtain the corresponding feature value R; and substitutes the feature value R into the Sigmoid function to calculate the corresponding probability of the event occurrence and outputs it. The characteristic value R is calculated as follows: ; Among them, W FC b FC These are the weight parameter vector and offset vector of the fully connected layer built into the prediction output layer, respectively; The probability of the event occurring is calculated as follows: 。 5. The method for handling thermal runaway events in autonomous vehicles according to claim 2, characterized in that, The process of first time-aligning the various sensor data in the acquired data array using interpolation, and then filtering the time-aligned sensor data using Kalman filtering to obtain the corresponding sensor data array, specifically includes: On the time axis of the most recent time period corresponding to the data acquisition array, N time scales t are set based on the time interval τ. Using linear interpolation, the internal temperature, external temperature, atmospheric pressure, atmospheric pressure, smoke concentration, CO2 concentration, and H2 concentration values at each time scale t are interpolated and estimated based on the internal temperature acquisition sequence, the external temperature acquisition sequence, the atmospheric pressure acquisition sequence, the smoke concentration acquisition sequence, the CO2 concentration acquisition sequence, and the H2 concentration acquisition sequence. The estimation results are then used to form corresponding internal temperature estimation sequences, external temperature estimation sequences, atmospheric pressure estimation sequences, smoke concentration estimation sequences, CO2 concentration estimation sequences, and H2 concentration estimation sequences. The internal temperature estimation sequence, the outer temperature estimation sequence, the atmospheric pressure estimation sequence, the smoke concentration estimation sequence, the CO2 concentration estimation sequence, and the H2 concentration estimation sequence are each subjected to Kalman filtering; and the six data sequences obtained after filtering are taken as the corresponding temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F; The temperature sequence A, temperature sequence B, pressure sequence C, smoke sequence D, gas sequence E, and gas sequence F obtained in this study form the corresponding sensor data array.
6. The method for handling thermal runaway events in autonomous vehicles according to claim 2, characterized in that, The process of handling alarms and emergencies locally on the vehicle side, uploading alarm data to a remote platform, and sending alarm information to vehicle management personnel specifically includes: Simultaneously, the vehicle side performs audible and visual alarms and deceleration and stopping operations locally; and when the vehicle stops, the doors are unlocked, passengers are reminded to disembark, and cooling spray operations are performed simultaneously. The system obtains the current vehicle location information and battery status information through the vehicle's autonomous driving system or vehicle CAN bus; uses the preset unique vehicle identifier as the corresponding current vehicle identifier; and uploads the corresponding remote alarm data, composed of the current time, the current vehicle identifier, the vehicle location information, the battery status information, and the most recently collected data array, to the remote platform. The alarm notification SMS, consisting of the current time, the current vehicle identifier, the vehicle location information, and the battery status information, is sent to the preset vehicle management personnel SMS notification interface.
7. An apparatus for performing the method for handling thermal runaway events in an autonomous vehicle according to any one of claims 1-6, characterized in that, The device includes: a multimodal data acquisition module, an acquired data extraction module, a data alignment and filtering module, an event prediction module, and an event early warning module; The multimodal data acquisition module is used to monitor a thermal runaway event by a monitoring sensor network consisting of a first temperature sensor and a second temperature sensor deployed inside and outside the vehicle battery pack, a pressure sensor deployed in the power compartment, a smoke sensor deployed in the passenger compartment, and a gas sensor deployed in the trunk; and to continuously collect sensor data from the monitoring sensor network and store the collected data in a sensor dataset. The data acquisition extraction module is used to periodically extract the sensor data in the sensor dataset that is in the most recent first time period according to a preset data analysis frequency to form a corresponding data acquisition array. The data alignment and filtering module is used to first perform time alignment on various types of sensor data in the acquired data array using interpolation, and then filter the time-aligned sensor data using Kalman filtering to obtain the corresponding sensor data array. The event prediction module is used to identify preset prediction modes. If the prediction mode is a first mode, a preset first event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is a second mode, a preset second event prediction model predicts the probability of the event corresponding to the sensor data array. If the prediction mode is a third mode, the first and second event prediction models work together to predict the probability of the event corresponding to the sensor data array. The prediction modes include first, second, and third modes. The first event prediction model is a thermal runaway event probability prediction model based on an SVM model. The second event prediction model is a thermal runaway event probability prediction model based on an LSTM model. Both the first and second event prediction models are used to predict the probability of thermal runaway events based on the sensor data array input to the model and output the corresponding event probability. The event warning module is used to perform local alarm emergency processing on the vehicle side, upload alarm data to the remote platform, and send alarm information to vehicle operation management personnel when the probability of the event exceeds a preset event probability threshold.
8. An electronic device, characterized in that, include: Memory, processor, and transceiver; The processor is configured to be coupled to the memory, read and execute instructions in the memory to implement the method according to any one of claims 1-6; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1-6.
Citation Information
Patent Citations
New energy automobile battery detection method and system based on integrated sensor
CN118244126A
Industrial and commercial energy storage thermal runaway early warning method based on multi-source cooperation
CN119007397A
