Lithium battery bicycle overcharge thermal runaway intelligent early warning method and system based on multi-modal sensor data fusion

Abstract

This invention relates to the field of lithium battery safety monitoring and intelligent early warning technology, and in particular to an intelligent early warning method and system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion. The method synchronously collects hydrogen, carbon monoxide, and hydrogen fluoride concentrations, as well as battery voltage and temperature, using a multimodal sensor array to construct a multidimensional feature vector containing physical quantities, statistical quantities, and interactive features. The feature vector is mapped to a high-dimensional space and positional encoding is added to generate an embedded feature sequence characterizing the temporal evolution of thermal runaway. This sequence is then input in parallel to a multi-head self-attention and multi-scale convolution module to extract the global dependency and local temporal features of the multi-physics coupling evolution of thermal runaway. After fusion, the thermal runaway risk probability is output through a fully connected network. An early warning is triggered when the probability exceeds a preset threshold n times consecutively, where n≥2. This invention achieves accurate early warning of thermal runaway, reduces false alarm and missed alarm rates, and exhibits good real-time performance and integration.

Description

Technical Field

[0001] This invention relates to the field of lithium battery safety monitoring and intelligent early warning technology, and in particular to an intelligent early warning method and system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion. Background Technology

[0002] Lithium-ion electric bicycles have become an important mode of transportation for short-distance urban travel due to their high energy density, long cycle life, and lightweight design. However, under overcharge conditions, lithium batteries are highly susceptible to thermal runaway due to the cascading effects of multiple physicochemical reactions, including lithium plating at the negative electrode, electrolyte decomposition, and separator failure, leading to fires and even explosions. According to statistics from the Ministry of Emergency Management, between 2020 and 2025, the annual growth rate of fires caused by lithium-ion electric bicycle battery failures in my country exceeded 30%, seriously threatening personal safety and property.

[0003] Existing overcharge thermal runaway warning technologies for lithium-ion electric bicycles generally suffer from the following significant drawbacks: Single signal source: Most existing solutions rely solely on battery surface temperature or cell voltage as early warning criteria, failing to fully utilize the various characteristic gases (such as H2, CO, HF), higher-order rate of change of temperature / voltage, and multi-physics coupling information released during thermal runaway, resulting in a severe lack of ability to perceive the early stages of thermal runaway.

[0004] Fixed threshold setting: Traditional methods often use fixed thresholds (such as temperature ≥80℃ or voltage ≥4.25V) as alarm criteria. However, the overcharge response characteristics of batteries vary significantly under different aging levels, ambient temperatures, and charging rates. Fixed threshold methods are difficult to adapt to the complexity of actual operating conditions and are prone to missed or false alarms.

[0005] Lack of temporal modeling: Thermal runaway is a typical nonlinear dynamic evolution process with time dependence and state lag. Existing methods fail to effectively model the time-series characteristics of sensor data, relying only on instantaneous values ​​at the current moment or within a very short window for judgment. The response lag is usually more than 2-3 minutes, missing the optimal intervention window.

[0006] Insufficient fusion of multi-source information: There is a lack of effective fusion mechanism between small-scale laboratory experimental data, numerical simulation results and field operation data. The evaluation model is difficult to reflect the evolution law of thermal runaway in real scenarios, the model has poor generalization ability, and the prediction error generally exceeds 30%.

[0007] Therefore, there is an urgent need in this field for a method and system that can integrate multimodal sensing information, capture the dynamic evolution of thermal runaway, and achieve early and accurate warning. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, this invention provides an intelligent early warning method and system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion. By combining multimodal sensor data with the physical process of thermal runaway in lithium-ion batteries, interpretable modeling and accurate early warning of the early stage of thermal runaway are achieved, thereby solving the problems existing in the background art.

[0009] This invention provides the following technical solution: an intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion, comprising the following steps: S1: Synchronously collect hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature of lithium-ion bicycle batteries through a multimodal sensor array to form multimodal time-series data; S2: Construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the multimodal time series data; S3: The multidimensional feature vector is mapped to a high-dimensional hidden representation space through a linear projection layer, and positional encoding is added to generate an embedded feature sequence that characterizes the temporal information and dynamic change characteristics of the thermal runaway evolution process; S4: Input the embedded feature sequence in parallel into the multi-head self-attention mechanism module and the multi-scale convolution module to extract the global temporal dependency features and multi-scale local temporal features of the thermal runaway multi-physics field coupled evolution, respectively. S5: The global temporal dependency features and multi-scale local temporal features are concatenated and fused, and the probability of thermal runaway risk is output through a fully connected classification network; S6: When the probability of thermal runaway exceeds a preset threshold more than n times, where n≥2, an early warning signal is triggered.

[0010] Furthermore, the physical quantity characteristics mentioned in step S2 include: hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, battery surface temperature, temperature change rate, voltage change rate, acceleration of temperature change, and acceleration of voltage change, for a total of 9 dimensions.

[0011] Furthermore, the statistical features mentioned in step S2 include: temperature moving average, voltage moving average, temperature standard deviation, temperature 75th percentile, and temperature 25th percentile, totaling 10 dimensions, all within a preset sliding window.

[0012] Furthermore, the interactive features mentioned in step S2 include: the product of temperature and voltage, and the total gas concentrations of hydrogen, carbon monoxide, and hydrogen fluoride, totaling two dimensions.

[0013] Furthermore, the multi-head self-attention mechanism module described in step S4 has 4 attention heads and a hidden dimension of 128.

[0014] Furthermore, the multi-scale convolution module in step S4 contains three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, respectively. Each convolutional layer has 64 output channels and is connected sequentially to a batch normalization layer and a GELU activation function.

[0015] Furthermore, the multi-scale convolution module in step S4 contains three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, respectively. Each convolutional layer has 64 output channels and is connected sequentially to a batch normalization layer and a GELU activation function.

[0016] An intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion includes: A multimodal sensor array is used to collect data on hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature. The data acquisition and preprocessing module is used to synchronously acquire multimodal sensor data and perform filtering and normalization processing; The feature construction module is used to construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the collected multimodal data; The embedding and encoding module is used to map multidimensional feature vectors to a high-dimensional representation space and add positional encoding; The temporal modeling module includes a multi-head self-attention unit and a multi-scale convolutional unit, which are used to extract global temporal dependency features and multi-scale local temporal features in parallel. The feature fusion and classification module is used to concatenate and fuse the extracted global and local features and output the probability of thermal runaway risk. The early warning output module is used to trigger an early warning based on the comparison result between the probability of thermal runaway risk and a preset threshold. Edge computing units are used to deploy and run the computing tasks of the above modules.

[0017] Furthermore, the multimodal sensor array includes an electrochemical hydrogen sensor, an electrochemical carbon monoxide sensor, an electrochemical hydrogen fluoride sensor, a voltage sensor, and a temperature sensor, with a sampling frequency of 10 Hz.

[0018] Furthermore, it also includes a cloud platform, which is used to receive early warning information, store historical data, provide a remote monitoring interface, and enable online updates and iterative optimization of models.

[0019] The technical effects and advantages of this invention are as follows: Compared with existing technologies, this invention constructs a 21-dimensional multimodal feature vector encompassing gas concentration, voltage, temperature and their higher-order derivatives, sliding window statistics, and multi-physical quantity interaction characteristics. Combined with a dual-branch temporal modeling architecture using multi-head self-attention and multi-scale convolution parallelism, it achieves holographic perception and dynamic capture of the overcharge thermal runaway evolution process of lithium-ion electric bicycles. It replaces the traditional fixed single-threshold threshold with a continuous multiple-threshold exceedance criterion, further reducing the probability of false triggering. Compared to traditional early warning methods relying on a single physical signal and a fixed threshold, this invention provides at least 30% earlier warning time under the same experimental conditions, effectively reducing false alarm and false negative rates. Furthermore, it exhibits good cross-type generalization ability across different lithium-ion battery material systems. Simultaneously, the system adopts an edge computing and cloud-based collaborative architecture, with inference latency meeting real-time application requirements, possessing excellent engineering integration and maintainability, providing reliable technical assurance for the charging safety of lithium-ion electric bicycles. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion, as described in this invention. Figure 2 This is a system block diagram of the intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion, as described in this invention. Detailed Implementation

[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, the intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion includes the following steps: S1: Synchronously collect hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature of lithium-ion bicycle batteries through a multimodal sensor array to form multimodal time-series data; S2: Construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the multimodal time series data; S3: The multidimensional feature vector is mapped to a high-dimensional hidden representation space through a linear projection layer, and positional encoding is added to generate an embedded feature sequence that characterizes the temporal information and dynamic change characteristics of the thermal runaway evolution process; S4: Input the embedded feature sequence in parallel into the multi-head self-attention mechanism module and the multi-scale convolution module to extract the global temporal dependency features and multi-scale local temporal features of the thermal runaway multi-physics field coupled evolution, respectively. S5: The global temporal dependency features and multi-scale local temporal features are concatenated and fused, and the probability of thermal runaway risk is output through a fully connected classification network; S6: When the probability of thermal runaway exceeds a preset threshold more than n times, where n≥2, an early warning signal is triggered.

[0023] In one specific embodiment, system deployment and data acquisition are performed first. Three DS18B20 digital temperature sensors are attached to the center of the lithium-ion bicycle battery pack. An array of electrochemical gas sensors, including an ME3-H2 hydrogen sensor, an ME3-CO carbon monoxide sensor, and an ME3-HF hydrogen fluoride sensor, is installed in the near-field region inside the battery pack. An ADuM4190 isolated voltage sampling module is connected in parallel to the positive and negative terminals of the battery pack. All sensors are connected to a data acquisition board via shielded cables. The data acquisition board uses an STM32H743 microcontroller, which has a built-in multi-channel ADC and multiple digital interfaces. A 10Hz sampling trigger signal is generated by the microcontroller's timer to synchronously start data acquisition from all sensors. In each sampling cycle, the microcontroller reads the analog voltage value of the gas sensor and converts it to PPM units, reads the temperature values ​​of the three temperature sensors via a single-bus protocol and averages them, reads the output value of the voltage sampling module via the ADC and converts it to the actual voltage, and packages all the data and adds a unified timestamp to form a multimodal time-series data stream.

[0024] Feature engineering is then performed. The microcontroller maintains a circular buffer of length 100 in memory, storing the raw data of the most recent 100 sampling points. For each new sampling point, the microcontroller extracts the current time and historical data from the buffer and calculates a 21-dimensional feature vector. Specifically, this includes directly reading the current hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature as 5-dimensional raw physical quantities; and using the central difference method to calculate the rate of temperature change and the rate of voltage change, as shown in the formula. ,in These are the values ​​for the next and previous time steps, respectively. For boundary points, forward or backward differencing is used. Second-order central differencing is used to calculate the acceleration due to temperature change and voltage change, using the following formula: This constitutes a 4-dimensional higher-order derivative feature. Sliding window sizes of 3, 5, and 10 sampling points are set, and the moving average of temperature, the moving average of voltage, and the standard deviation of temperature within the window are calculated. The moving average is the arithmetic mean of the data within the window, and the standard deviation is calculated using the sample standard deviation formula. The following features are constructed: 9-dimensional statistical features; 2-dimensional quantile features are constructed by calculating the 75th and 25th quantiles of temperature values ​​within the window, obtained by sorting the temperature values ​​within the window and taking the values ​​at the 75th and 25th percentile positions; 3-dimensional temperature-voltage product features are obtained by multiplying the current temperature value by the voltage value; and 2-dimensional interactive features are obtained by summing the hydrogen concentration, carbon monoxide concentration, and hydrogen fluoride concentration. In total, there are 21 feature vectors.

[0025] After feature construction, feature embedding and positional encoding are performed. The microcontroller sends the 21-dimensional feature vector to the edge computing unit via the SPI interface. The edge computing unit uses an NVIDIA Jetson Orin Nano module, runs on the Ubuntu 20.04 operating system, and deploys the PyTorch 2.0 deep learning framework and the TensorRT inference engine. In the edge computing unit, the input features are first transformed in dimension through a linear projection layer. The linear projection layer is a fully connected network with an input dimension of 21 and an output dimension of 128, and the weight matrix... The shape is 21×128, and the bias vector is... The length is 128, and the calculation formula is: ,in This represents the 21-dimensional feature vector of the current sampling point. For an input sequence of length L, the L feature vectors are sequentially passed through a linear projection layer to obtain the embedded feature sequence. Then, sine and cosine position encoding is added. The position encoding matrix PE has a shape of L×128, where the encoding value is calculated for position pos and dimension index i using the following formula: The positional encoding matrix is ​​added to the embedded feature sequence to obtain the final embedded feature sequence. This sequence characterizes the temporal information and dynamic changes of the thermal runaway evolution process.

[0026] Next, parallel temporal modeling is performed, embedding feature sequences. Simultaneously, the data is processed through two branches. The first branch is a multi-head self-attention mechanism module, used to extract the global temporal dependency features of the multiphysics coupling evolution of thermal runaway. This module sets the number of attention heads h=4, and the dimension dk=32 for each head. First, the query matrix Q, key matrix K, and value matrix V are generated through three linear transformations, calculated using the following formula: ,in All are learnable weight matrices. Q, K, and V are split into 4 groups based on the number of heads, each group having a shape of L×32. Scaled dot product attention is calculated for each group using the following formula: The output shape is L×32. The outputs of the four heads are concatenated along the feature dimensions to obtain a tensor of shape L×128. Then, a linear transformation is performed through the output projection matrix WO∈R128×128 to obtain the multi-head attention output. This output is added to the input through a residual connection, i.e., Hattn=H0+MultiHead(Q,K,V), followed by layer normalization. After layer normalization, the input is fed forward into a feedforward network, which contains two fully connected layers. The first layer maps 128 dimensions to 512 dimensions. After GELU activation, the second layer maps 512 dimensions back to 128 dimensions. Then, it is processed through residual connections and layer normalization to complete the computation of one Transformer layer. In this embodiment, two such Transformer layers are stacked to output a global feature sequence. .

[0027] The second branch is a multi-scale convolutional module used to extract multi-scale local temporal features of the coupled evolution of thermal runaway multiphysics fields. The embedded feature sequence H0H0 is simultaneously input into three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, a stride of 1, and padding of "same" to ensure the output sequence length remains LL. Each convolutional layer has 128 input channels and 64 output channels. Each convolutional layer is followed by a batch normalization layer and a GELU activation function. The batch normalization layer normalizes each channel using the following formula: , where μ and ε are the mean and variance of the channel, respectively, and ϵ is taken as 1×10. −5 Then, a linear transformation is performed using learnable scaling and offset factors. The formula for calculating the GELU activation function is as follows: The three convolutional layers each output a feature map of shape L×64, denoted as... .

[0028] After completing temporal modeling, feature fusion and classification are performed. For global feature sequences... Average pooling is performed along the sequence dimension, and the calculation formula is as follows: This yields a global feature vector of shape 128. For the three multi-scale convolutional output feature maps, adaptive average pooling is performed, compressing each feature map to 1 in the sequence dimension, i.e., averaging the L values ​​of each channel, resulting in three feature vectors of shape 64, denoted as follows: The global feature vector is concatenated with the three local feature vectors along the feature dimension to obtain a fused feature vector with a shape of 128+64+64+64=320. .Will Input a three-layer fully connected classification network. The first fully connected layer has an input dimension of 320 and an output dimension of 256, followed by a GELU activation function and a Dropout layer with a Dropout probability of 0.3. The second fully connected layer has an input dimension of 256 and an output dimension of 128, followed by a GELU activation function and a Dropout layer with a Dropout probability of 0.3. The third fully connected layer has an input dimension of 128 and an output dimension of 1, followed by a Sigmoid activation function to output the probability of thermal runaway. The above calculation process is executed once every 0.1 seconds on the edge computing unit, processing an input sequence with a sliding window length of L=64 (corresponding to 6.4 seconds of historical data).

[0029] Finally, a warning trigger judgment is performed. The warning threshold θ = 0.85 is set, and the consecutive judgment criterion n = 3 (n ≥ 2, 3 in this embodiment). The system maintains a circular queue of length 3 to store the risk probability values ​​output by the model for the most recent 3 times. Each time a new risk probability value is obtained, the system checks whether all values ​​in the circular queue are greater than 0.85; if so, a warning signal is triggered. This judgment method corresponds to the condition that the risk probability exceeds the preset threshold n times consecutively. The warning signal outputs a high level through the GPIO interface of the edge computing unit, driving the buzzer to emit a continuous alarm sound for 5 seconds. Simultaneously, the relay is controlled through the I / O port to disconnect the charger's input power, physically cutting off the charging circuit. The warning signal is also sent to the vehicle's instrument panel via the CAN bus, and the instrument panel screen displays the text "Battery overheating risk, please stop charging immediately." All alarm events and feature snapshots of the alarm time are uploaded to the cloud platform via the 4G module using the MQTT protocol. The cloud platform records the alarm time, device ID, risk probability value, and feature data, and pushes the message to the user's mobile phone via WeChat official account.

[0030] Furthermore, the physical quantity characteristics mentioned in step S2 include: hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, battery surface temperature, temperature change rate, voltage change rate, acceleration of temperature change, and acceleration of voltage change, for a total of 9 dimensions.

[0031] In one specific embodiment, the details of constructing the physical quantity characteristics are as follows: The hydrogen concentration characteristic is directly read from the analog voltage value output by the ME3-H2 sensor, which outputs a voltage of 0-2.5V in the range of 0-40000ppm, and is determined using a linear calibration formula. Converted to PPM value. The carbon monoxide concentration characteristic is directly read from the analog voltage value output by the ME3-CO sensor, which outputs a voltage of 0-3.0V within the range of 0-10000ppm, using a linear calibration formula. Converted to PPM value. The hydrogen fluoride concentration characteristic is directly read from the analog voltage value output by the ME3-HF sensor, which outputs a voltage of 0-2.5V within the range of 0-100ppm, using a linear calibration formula. Converted to PPM value. Battery voltage characteristics are obtained through an ADuM4190 isolated voltage sampling module, which linearly converts a 0-60V input voltage to a 0-3.3V output voltage. The microcontroller's ADC reads this voltage and then uses the formula... This is converted to an actual voltage value, where 18.18 is the voltage division ratio. The battery surface temperature characteristics are read by three DS18B20 digital temperature sensors. The microcontroller reads the temperature values ​​of the three sensors sequentially via a single-bus protocol, and takes the arithmetic mean as the current battery surface temperature, in degrees Celsius, rounded to one decimal place.

[0032] The temperature change rate characteristic is calculated using the central difference method. The system maintains the temperature values ​​of the three most recent sampling points in memory. The sampling interval Δt = 0.1 seconds, and the formula for calculating the rate of temperature change is: For the starting point of the sequence, the forward difference formula is used; For the end point of the sequence, the backward difference formula is used. The voltage change rate is calculated using the same method, and the formula is: Forward or backward difference is also used at the boundary.

[0033] The acceleration characteristics of temperature change are calculated using the second-order central difference formula, which is as follows: This characteristic reflects the accelerating trend of the rate of temperature change. The acceleration characteristic of voltage change is calculated using the same method, and the formula is: The above nine physical quantity characteristics are recalculated once at each sampling time (every 0.1 seconds), forming a set of basic physical quantities that describe the dynamic process of battery overcharge thermal runaway.

[0034] Furthermore, the statistical features mentioned in step S2 include: temperature moving average, voltage moving average, temperature standard deviation, temperature 75th percentile, and temperature 25th percentile, totaling 10 dimensions, all within a preset sliding window.

[0035] In one specific embodiment, the details of constructing statistical features are as follows: The system maintains circular buffers of length 10 in memory for both temperature and voltage values, storing data from the 10 most recent sampling points. At each sampling time, the system calculates statistics for three different sliding window sizes based on the current buffer contents.

[0036] For a statistic with a window size of 3, the temperature values ​​of the three most recent sampling points (corresponding to 0.3 seconds) in the buffer are taken, and their arithmetic mean is calculated as the temperature moving average feature. The calculation formula is as follows: For window sizes of 5 and 10, the temperature values ​​of the most recent 5 sampling points (0.5 seconds) and 10 sampling points (1.0 seconds) are taken respectively, and the same arithmetic mean method is used to calculate the moving average of temperature. The voltage moving average is calculated using the exact same method, with the voltage arithmetic mean calculated for window sizes of 3, 5, and 10 respectively.

[0037] The temperature standard deviation characteristic is calculated separately for each window size. For a window size of 3, the arithmetic mean of the temperature values ​​within the window is calculated first. Then calculate the standard deviation. , where n is the window size. When the window size is 3, the sample size is small, but the sample standard deviation formula is still used to obtain an unbiased estimate. For window sizes of 5 and 10, the same standard deviation calculation formula is used. It should be noted that when there is insufficient data within the window, the maximum amount of data currently available is used for calculation.

[0038] The 75th and 25th percentile temperature features are calculated using only a buffer with a window size of 10. The system copies the 10 most recent temperature values ​​from the buffer into a temporary array and uses a Quickselect algorithm to find the values ​​corresponding to the 25th and 75th percentiles in O(n) time complexity. Specifically, for 10 data points, the 25th percentile is the 3rd value after sorting (index 2), and the 75th percentile is the 8th value after sorting (index 7). If there are fewer than 10 data points, the corresponding quantiles are calculated using all current data points. This calculation is performed once at each sampling time, forming a 10-dimensional statistical feature.

[0039] Furthermore, the interactive features mentioned in step S2 include: the product of temperature and voltage, and the total gas concentrations of hydrogen, carbon monoxide, and hydrogen fluoride, totaling two dimensions.

[0040] In one specific embodiment, the details of constructing the interaction features are as follows: The temperature-voltage product feature is obtained at each sampling time by multiplying the current battery surface temperature value (in °C) by the current battery voltage value (in V), and the calculation formula is as follows: This feature is used to characterize the coupling effect of temperature and voltage. When a battery is overcharged, temperature and voltage often rise simultaneously, and their product can amplify abnormal signals.

[0041] The total gas concentration characteristic is obtained at each sampling time by summing the current hydrogen concentration, carbon monoxide concentration, and hydrogen fluoride concentration. The calculation formula is as follows: The unit is PPM. This feature comprehensively reflects the intensity of gas release from electrolyte decomposition. When side reactions occur inside the battery, multiple characteristic gases are released simultaneously, and their total concentration can provide a more stable indication of anomalies than a single gas.

[0042] The aforementioned 2D interactive features, together with the aforementioned 9D physical quantity features and 10D statistical quantity features, constitute a complete 21D feature vector. In actual implementation, after completing the data acquisition of each sampling point, the microcontroller sequentially calculates the aforementioned 21D features and sends them to the edge computing unit for subsequent processing via the SPI interface.

[0043] Furthermore, the multi-head self-attention mechanism module described in step S4 has 4 attention heads and a hidden dimension of 128.

[0044] In one specific embodiment, the multi-head self-attention mechanism module is implemented as follows. In the edge computing unit, this module is implemented using the PyTorch framework, and the module is defined as a class inheriting from `torch.nn.Module`. The initialization parameters of this module include: input dimension... The number of attention heads h=4, and the dimension of each head. The module defines three linear transformation layers. Each linear layer has an input dimension of 128, an output dimension of 128, and an output projection layer. Input dimension 128, output dimension 128.

[0045] During forward propagation, the input tensor x has a shape of (L, 128), where L is the sequence length. First, Q, K, and V are computed through three linear layers, each with a shape of (L, 128). The shapes of Q, K, and V are then transformed... That is, (L, 4, 32), and by swapping the first and second dimensions, we obtain the shape as The tensor. When calculating the scaled dot product attention, first calculate the... and The dot product of these values ​​yields an attention score matrix of shape (L,L), which is then divided by... Scale the matrix, then apply the softmax function along the last dimension (i.e., the key dimension) to obtain the attention weight matrix, and then combine it with... Multiply by each head to obtain the output of each head, which has the following shape: The outputs of hh heads are concatenated along the feature dimension to obtain a tensor of shape (L, 128), which is then passed through the output projection layer. Perform a linear transformation to obtain the output of multi-head attention.

[0046] In the implementation, to improve computational efficiency, a batch processing approach is used to process multiple heads simultaneously, employing `torch.matmul` for batch matrix multiplication. This module is followed by residual connections and layer normalization, implemented using `torch.nn.LayerNorm` with a normalization dimension of 128. The feedforward network consists of two linear layers with the GELU activation function in between. The first layer maps the 128-dimensional network to 512-dimensional, and the second layer maps the 512-dimensional network back to 128-dimensional, also through residual connections and layer normalization. In this embodiment, two complete Transformer layers are stacked, ultimately outputting a global feature sequence.

[0047] Furthermore, the multi-scale convolution module in step S4 contains three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, respectively. Each convolutional layer has 64 output channels and is connected sequentially to a batch normalization layer and a GELU activation function.

[0048] In one specific embodiment, the multi-scale convolution module is implemented as follows. This module is implemented in the PyTorch framework and is defined as three parallel convolutional branches, each containing a 1D convolutional layer, a batch normalization layer, and a GELU activation function.

[0049] The branch with a kernel size of 3 is implemented as follows: The convolutional layer uses torch.nn.Conv1d, with 128 input channels and 64 output channels. The kernel size is 3, the stride is 1, and the padding is 1 (using "same" padding, meaning the sequence length remains unchanged before and after convolution). The convolutional layer is followed by torch.nn.BatchNorm1d(64), a batch normalization layer that independently normalizes each channel. The learnable parameters are initialized to a scaling factor of 1 and an offset factor of 0. After batch normalization, torch.nn.GELU() is used as the activation function.

[0050] The branch with a kernel size of 5 is implemented as follows: The convolutional layer uses torch.nn.Conv1d, with 128 input channels and 64 output channels. The kernel size is 5, the stride is 1, and the padding is 2 to ensure that the length of the output sequence is consistent with the input. This is followed by BatchNorm1d(64) and the GELU activation function.

[0051] The branch with a kernel size of 7 is implemented as follows: The convolutional layer uses torch.nn.Conv1d, with 128 input channels and 64 output channels. The kernel size is 7, the stride is 1, and the padding is 3 to ensure that the length of the output sequence is consistent with the input. This is followed by BatchNorm1d(64) and the GELU activation function.

[0052] During forward propagation, the input embedded feature sequence H0H0 has a shape of (L, 128). First, a permute operation transforms the shape to (1, 128, L), where the first dimension is the batch size (batch size = 1), the second is the channel size, and the third is the sequence length. Three branches process this tensor in parallel, with each branch outputting a shape of (1, 64, L). The output of each branch is then permuteed back to (L, 64), resulting in three feature maps. This module uses three convolutional kernels of different sizes to capture local temporal patterns within time windows of 0.3 seconds (kernel size 3 corresponds to 3 sampling points), 0.5 seconds (kernel size 5), and 0.7 seconds (kernel size 7), respectively.

[0053] Furthermore, the multi-scale convolution module in step S4 contains three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, respectively. Each convolutional layer has 64 output channels and is connected sequentially to a batch normalization layer and a GELU activation function.

[0054] In a specific embodiment, the fully connected classification network is implemented as follows. The network is defined as a three-layer fully connected structure, implemented using the PyTorch framework, with the parameters of each layer as follows: The first layer: The input dimension is the total dimension of the fused feature vector. Depending on the specific configuration, the fused features are composed of global features (128 dimensions) and three multi-scale convolutional features (64 dimensions each), totaling 128 + 64 + 64 + 64 = 320 dimensions. The first fully connected layer uses `torch.nn.Linear(320, 256)`, initialized using Kaiming uniform initialization. The activation function is GELU, implemented using `torch.nn.GELU()`. A Dropout layer with a probability of 0.3 is added after this layer to prevent overfitting.

[0055] The second layer is a fully connected layer, torch.nn.Linear(256, 128), also initialized using Kaiming uniformly. The activation function is GELU, followed by a Dropout layer with a Dropout probability of 0.3.

[0056] The third layer is a fully connected layer, torch.nn.Linear(128,1), initialized using Xavier uniform initialization. The activation function is Sigmoid, implemented using torch.nn.Sigmoid(), with an output range of [0,1], representing the probability of thermal runaway.

[0057] During model training, the classification network is jointly trained with the aforementioned feature embedding layer, Transformer layer, and multi-scale convolutional layer, using a binary cross-entropy loss function. The optimizer is AdamW, with an initial learning rate of 1e-4, weight decay of 1e-5, batch size of 64, and 200 training epochs. The early stopping strategy stops training when the validation set loss does not decrease for 10 consecutive epochs. During inference, the network receives fused feature vectors and outputs a single probability value, with an inference latency of less than 5 milliseconds.

[0058] Please see Figure 2 As shown, the intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion includes: A multimodal sensor array is used to collect data on hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature. The data acquisition and preprocessing module is used to synchronously acquire multimodal sensor data and perform filtering and normalization processing; The feature construction module is used to construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the collected multimodal data; The embedding and encoding module is used to map multidimensional feature vectors to a high-dimensional representation space and add positional encoding; The temporal modeling module includes a multi-head self-attention unit and a multi-scale convolutional unit, which are used to extract global temporal dependency features and multi-scale local temporal features in parallel. The feature fusion and classification module is used to concatenate and fuse the extracted global and local features and output the probability of thermal runaway risk. The early warning output module is used to trigger an early warning based on the comparison result between the probability of thermal runaway risk and a preset threshold. Edge computing units are used to deploy and run the computing tasks of the above modules.

[0059] In one specific embodiment, in terms of system implementation, the entire early warning system consists of three parts: a hardware layer, an embedded software layer, and a cloud software layer.

[0060] The hardware layer includes a multimodal sensor array and a data acquisition and preprocessing module. The specific configuration of the multimodal sensor array is as follows: ME3-H2 electrochemical hydrogen sensor, operating voltage 5V, output 0-2.5V analog voltage corresponding to 0-40000ppm; ME3-CO electrochemical carbon monoxide sensor, operating voltage 5V, output 0-3.0V analog voltage corresponding to 0-10000ppm; ME3-HF electrochemical hydrogen fluoride sensor, operating voltage 5V, output 0-2.5V analog voltage corresponding to 0-100ppm; ADuM4190 isolated voltage sampling module, input 0-60V, output 0-3.3V isolated analog signal; three DS18B20 digital temperature sensors, encapsulated in stainless steel and attached to the battery surface with thermal grease. The data acquisition and preprocessing module uses an STM32H743 microcontroller as its core. This chip has a main frequency of 480MHz, three built-in 12-bit ADCs, a maximum sampling rate of 3.6Msps, and supports 16-channel simultaneous sampling. The microcontroller reads analog signals from the gas and voltage sensors via the ADC channel and temperature values ​​from the DS18B20 sensor via a single-bus interface. Sampling from all sensors is uniformly triggered by a 10Hz interrupt generated by timer TIM2 to ensure data synchronization. The microcontroller performs data reading, Kalman filtering, and z-score normalization within the interrupt service routine. The state transition matrix for the Kalman filter is set to A=1, the observation matrix to H=1, and the process noise covariance is set to... Observation noise covariance The initial state estimate is set to the first observation. z-score normalization uses pre-calculated mean and standard deviation, performed independently for each sensor channel, as shown in the formula: ,in and They are stored in the microcontroller's internal Flash memory.

[0061] The feature construction module also runs on the STM32H743 microcontroller and is implemented in C. The module maintains multiple circular buffers, each with a length of 100, to store historical sensor data. Upon triggering at each sampling point, the feature construction module sequentially calculates a 21-dimensional feature vector. Floating-point arithmetic is used during the calculation, with hardware acceleration provided by the STM32H743's FPU unit. The calculated feature vector is then sent to the edge computing unit via the SPI interface using DMA. The SPI communication rate is set to 10MHz, transmitting 84 bytes (21-dimensional × 4 bytes) every 0.1 seconds, with communication overhead less than 1%.

[0062] The edge computing unit utilizes an NVIDIA Jetson Orin Nano module, equipped with a 6-core ARM Cortex-A78AE CPU and a 1024-core NVIDIA Ampere architecture GPU, running Ubuntu 20.04 operating system, and installing PyTorch 2.0, TensorRT 8.5, and cuDNN 8.9. The embedding and encoding module, temporal modeling module, and feature fusion and classification module are implemented using Python scripts. Model parameters are stored in eMMC memory in PyTorch's state_dict format. Upon system startup, the edge computing unit loads the model and performs inference optimization using TensorRT, converting the model to the TensorRT engine to achieve FP16 precision inference, reducing inference latency from approximately 15 milliseconds in the original PyTorch to less than 5 milliseconds. The edge computing unit receives feature vectors via the SPI interface. Upon receiving each feature vector, it appends it to a feature sequence buffer in memory, with a buffer length of 64. When the buffer is full, the 64 feature vectors are concatenated into a tensor of shape (64, 21) and input into the model for inference.

[0063] The warning output module is implemented through the GPIO interface of the edge computing unit. The GPIO pin is configured for output mode, with an initial low level. When the warning condition is met (risk probability greater than 0.85 for three consecutive times), the edge computing unit pulls the GPIO pin high. This high-level signal is connected to the buzzer driver circuit and the relay driver circuit. The buzzer driver circuit uses an S8050 transistor switching circuit. When the GPIO is high, the transistor conducts, the buzzer emits an alarm sound, and after 5 seconds, the edge computing unit pulls the GPIO low to turn it off. The relay driver circuit uses optocoupler isolation and MOSFET switching. When the GPIO is high, the optocoupler conducts, the MOSFET gate voltage rises, the relay coil is energized, the normally closed contact opens, and the AC input power to the charger is cut off. A normally closed relay with a DC 24V coil voltage is selected to ensure that the charging circuit remains disconnected when the system is powered off.

[0064] Furthermore, the multimodal sensor array includes an electrochemical hydrogen sensor, an electrochemical carbon monoxide sensor, an electrochemical hydrogen fluoride sensor, a voltage sensor, and a temperature sensor, with a sampling frequency of 10 Hz.

[0065] In one specific embodiment, the specific selection, parameter configuration, and installation method of the multimodal sensor array are as follows: The hydrogen sensor selected is the ME3-H2 electrochemical sensor, with the following technical parameters: measurement range 0-40000ppm, resolution 10ppm, response time T90≤30 seconds, operating voltage 5V DC, and output signal 0-2.5V linear analog voltage. This sensor is installed inside the battery pack near the battery cells, fixed in the gap between the battery modules by a bracket, with the sensor head facing the battery cells for rapid detection of leaked gas.

[0066] The carbon monoxide sensor is an ME3-CO type electrochemical sensor with the following technical parameters: measurement range 0-10000ppm, resolution 5ppm, response time T90≤30 seconds, operating voltage 5V DC, and output signal 0-3.0V linear analog voltage. It is installed adjacent to the hydrogen sensor, also inside the battery pack.

[0067] The hydrogen fluoride sensor selected is the ME3-HF electrochemical sensor, with the following technical parameters: measurement range 0-100ppm, resolution 0.5ppm, response time T90≤60 seconds, operating voltage 5V DC, and output signal 0-2.5V linear analog voltage. Due to the corrosive nature of HF gas, the sensor probe is protected with a Teflon coating and is installed near the battery pack exhaust valve.

[0068] The voltage sensor uses the ADuM4190 isolated voltage sampling module, which is based on Analog Devices' iCoupler technology and provides a 5kV isolation voltage. The module has an input range of 0-60V and an output of a 0-3.3V isolated analog signal with an accuracy of 0.1% and a nonlinearity error of less than 0.05%. The module's input is connected in parallel to the positive and negative terminals of the battery pack, and its output is connected to the microcontroller's ADC input pin.

[0069] The temperature sensor uses three DS18B20 digital temperature sensors in a TO-92 package. These sensors have a measurement range of -55℃ to +125℃, an accuracy of ±0.5℃, and a programmable resolution to 0.0625℃. They employ a single-bus communication protocol, and each sensor has a unique 64-bit ROM serial number, allowing them to be mounted on the same bus. The three sensors are respectively attached to the center, edge, and area near the vent valve of the battery pack. Thermally conductive silicone grease is used to fill the contact surfaces, and high-temperature resistant tape is used for fixation. The sensor leads are made of silicone-shielded wire and are 30cm in length.

[0070] All sensors are uniformly configured with a sampling frequency of 10Hz, and synchronous sampling is achieved through the microcontroller's timer interrupt. For analog output gas and voltage sensors, the microcontroller's ADC is configured in continuous scan mode, initiating ADC conversion upon each sampling trigger, and transferring data via DMA after conversion. For digital temperature sensors, the temperature values ​​of the three sensors are sequentially read via a single-bus protocol upon each sampling trigger, with each read taking approximately 10 milliseconds, completed within the interrupt service routine. The total execution time of the interrupt service routine is controlled within 15 milliseconds to meet the 10Hz sampling frequency requirement.

[0071] Furthermore, it also includes a cloud platform, which is used to receive early warning information, store historical data, provide a remote monitoring interface, and enable online updates and iterative optimization of models.

[0072] In one specific embodiment, the cloud platform is deployed on an Alibaba Cloud ECS server with an instance specification of ecs.c7.xlarge (4 cores, 8GB RAM), an operating system of Ubuntu 20.04 LTS, a 100GB SSD system disk, and a 500GB high-efficiency cloud disk for data storage. The server has a public network bandwidth of 10Mbps and is assigned a fixed public IP address.

[0073] The cloud platform software architecture adopts a microservice design, and its main components include: MQTT message broker, data access service, time-series database, object storage, web backend service, and web frontend interface.

[0074] The MQTT message broker uses EMQ X Enterprise 4.4, configured in cluster mode, listening on port 1883 (MQTT) and port 8883 (MQTT over TLS). The edge computing unit acts as an MQTT client, establishing an MQTT connection with the cloud via a 4G module. It subscribes to the topic "device / {device_id} / command" to receive downlink commands and publishes the topic "device / {device_id} / telemetry" to upload telemetry data. MQTT connections use username and password authentication, with each device assigned a unique Client ID and password. The data access service is developed based on the Python Flask framework. It subscribes to telemetry topics in EMQX, parses the JSON-formatted data payload when a new message is received, extracts information such as device ID, timestamp, risk probability value, 21-dimensional feature snapshot, and alarm status, and then writes it to a time-series database.

[0075] The time-series database uses InfluxDB 2.0 and the Flux query language. A bucket named "battery_telemetry" is created in the database, with a data retention policy of 365 days. Each data record contains the following tags: device_id (unique device identifier), location (device location), and alert_status (alarm status). Fields include: risk_probability (risk probability), temperature, voltage, gas_h2 (hydrogen concentration), gas_co (carbon monoxide concentration), and gas_hf (hydrogen fluoride concentration). Timestamps are in nanosecond precision and are uploaded after synchronization with the device's local time.

[0076] The object storage service uses Alibaba Cloud OSS, creating a bucket named "model-update" to store deep learning model files. Model files are saved in PyTorch's `state_dict` format, named "model_v{version}.pt", with version numbers incrementing from 1. When a new version of the model is deployed in the cloud, the system uploads the model file to OSS and sends an update command to all online devices via MQTT. Upon receiving the command, the edge computing unit downloads the new model file via HTTPS, verifies file integrity (MD5 checksum), and then reloads the model using TensorRT, achieving hot updates. The entire process is seamless for the user.

[0077] The web backend service is developed based on the Flask framework and provides RESTful API interfaces, including: device list query, device details query, historical alarm query, characteristic curve data acquisition, and model version management interfaces. The API interfaces use JWT for authentication, and access logs are recorded to Alibaba Cloud Log Service.

[0078] The web front-end interface is developed based on the Vue.js 3.0 framework and uses the Element Plus UI component library. Interface functions include: a device status dashboard (displaying statistics such as the number of online devices and the number of alarms today), a map visualization (displaying device locations and alarm points based on the Amap API), a real-time monitoring page (displaying the real-time risk probability curve and 21-dimensional feature curve of the selected device), a historical alarm page (supporting filtering alarm records by time range and device, and supporting export to CSV), and a model management page (displaying the current model version and supporting uploading new model files).

[0079] The cloud platform also integrates WeChat Official Account push functionality, sending alarm information as template messages to users' WeChat accounts via the WeChat Official Accounts Platform API. Users can bind their WeChat OpenID on the web frontend, and the cloud platform will automatically send a push message when an alarm is triggered, including the device name, alarm time, risk probability value, and handling suggestions.

[0080] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent early warning of overcharge thermal runaway in lithium-ion electric bicycles based on multimodal sensor data fusion, characterized in that, Includes the following steps: S1: Synchronously collect hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature of lithium-ion bicycle batteries through a multimodal sensor array to form multimodal time-series data; S2: Construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the multimodal time series data; S3: The multidimensional feature vector is mapped to a high-dimensional hidden representation space through a linear projection layer, and positional encoding is added to generate an embedded feature sequence that characterizes the temporal information and dynamic change characteristics of the thermal runaway evolution process; S4: Input the embedded feature sequence in parallel into the multi-head self-attention mechanism module and the multi-scale convolution module to extract the global temporal dependency features and multi-scale local temporal features of the thermal runaway multi-physics field coupled evolution, respectively. S5: The global temporal dependency features and multi-scale local temporal features are concatenated and fused, and the probability of thermal runaway risk is output through a fully connected classification network; S6: When the probability of thermal runaway exceeds a preset threshold more than n times, where n≥2, an early warning signal is triggered.

2. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The physical quantity characteristics mentioned in step S2 include: hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, battery surface temperature, temperature change rate, voltage change rate, acceleration of temperature change, and acceleration of voltage change, totaling 9 dimensions.

3. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The statistical features mentioned in step S2 include: temperature moving average, voltage moving average, temperature standard deviation, temperature 75th percentile, and temperature 25th percentile, totaling 10 dimensions, all within a preset sliding window.

4. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The interactive features mentioned in step S2 include: the product of temperature and voltage, and the total gas concentrations of hydrogen, carbon monoxide and hydrogen fluoride, totaling two dimensions.

5. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The multi-head self-attention mechanism module described in step S4 has 4 attention heads and 128 hidden dimensions.

6. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The multi-scale convolution module in step S4 contains three parallel 1D convolutional layers with kernel sizes of 3, 5, and 7, respectively. Each convolutional layer has 64 output channels and is connected to a batch normalization layer and a GELU activation function in sequence.

7. The intelligent early warning method for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 1, characterized in that, The fully connected classification network described in step S5 is a three-layer fully connected structure with hidden layer dimensions of 256 and 128, and the output layer uses the Sigmoid activation function.

8. An intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion, characterized in that: include: A multimodal sensor array is used to collect data on hydrogen concentration, carbon monoxide concentration, hydrogen fluoride concentration, battery voltage, and battery surface temperature. The data acquisition and preprocessing module is used to synchronously acquire multimodal sensor data and perform filtering and normalization processing; The feature construction module is used to construct a multidimensional feature vector containing physical quantity features, statistical features, and interaction features based on the collected multimodal data; The embedding and encoding module is used to map multidimensional feature vectors to a high-dimensional representation space and add positional encoding; The temporal modeling module includes a multi-head self-attention unit and a multi-scale convolutional unit, which are used to extract global temporal dependency features and multi-scale local temporal features in parallel. The feature fusion and classification module is used to concatenate and fuse the extracted global and local features and output the probability of thermal runaway risk. The early warning output module is used to trigger an early warning based on the comparison result between the probability of thermal runaway risk and a preset threshold. Edge computing units are used to deploy and run the computing tasks of the above modules.

9. The intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 8, characterized in that, The multimodal sensor array includes an electrochemical hydrogen sensor, an electrochemical carbon monoxide sensor, an electrochemical hydrogen fluoride sensor, a voltage sensor, and a temperature sensor, with a sampling frequency of 10 Hz.

10. The intelligent early warning system for overcharge thermal runaway of lithium-ion electric bicycles based on multimodal sensor data fusion according to claim 8, characterized in that, It also includes a cloud platform, which is used to receive early warning information, store historical data, provide a remote monitoring interface, and enable online updates and iterative optimization of models.

Images