Battery thermal runaway early warning system and method

Abstract

The invention discloses a battery thermal runaway early warning system and method.In the battery thermal runaway early warning system, at least one battery intelligent end cover device comprises a state information sampling module, and a gas sensor signal acquisition unit is arranged near a pressure release valve of a battery end cover so as to measure battery fault release gas to obtain a characteristic gas content signal; the battery surface temperature acquisition unit is attached to the surface of a battery end cover to acquire surface temperature information of a single cell, the battery pressure acquisition unit is attached to the surface of the battery end cover to acquire stress information of the single cell, and the battery voltage acquisition unit measures battery voltage data; fusing data of various sensors based on Bayesian deep learning of potential Dirichlet allocation of a time sequence, and extracting features to establish a multi-parameter lithium battery thermal runaway early warning model; and the thermal runaway early warning unit is connected with the time sequence potential Dirichlet allocation-based Bayesian deep learning module so as to early warn thermal runaway based on a model result.

Description

Technical Field

[0001] The present invention relates to the technical field of battery status monitoring, and in particular to a battery thermal runaway early warning system and method. Background Art

[0002] In recent years, the rapid development of industries such as electrochemical energy storage power stations and electric vehicles has led to the continuous expansion of the application scale of lithium batteries. However, lithium batteries inevitably have internal defects during large-scale production. Moreover, the working environment of lithium batteries is complex, and it is difficult to avoid harsh working conditions such as overcharging, over-discharging, overheating, and external stress stimulation. Under the influence of internal and external factors, lithium batteries will have the risk of thermal runaway failures, which will cause fires or even explosions, resulting in casualties and property losses. Therefore, it is of great significance to monitor the status of lithium batteries and provide timely warnings for thermal runaway failures.

[0003] Research shows that thermal runaway of lithium batteries can be divided into four stages: in the first stage, the battery temperature continues to rise, a series of chemical reactions occur inside, a large amount of heat and gas is generated, and the internal pressure of the battery increases; in the second stage, the internal pressure of the battery reaches the threshold, the battery safety valve opens, and the gas and electrolyte vapor are released; the third stage is the stage of severe thermal runaway of the battery, the surface temperature of the battery rises exponentially, and a large amount of gas is released, and the pressure increases sharply; the fourth stage is the battery cooling stage. In the second stage, the battery safety valve opens, and after the internal gas and electrolyte vapor of the battery are released, these gases can be detected. Since the above-mentioned gases originate from the internal chemical reactions of the battery and the volatilization of the electrolyte, when the battery is operating normally, these gases are not contained outside. Therefore, the gas composition and content of the normal operating state of the battery and the thermal runaway failure will be significantly different, and the battery state can be judged based on this. Moreover, there is a certain development time from the opening of the battery safety valve to the severe thermal runaway. Therefore, if the gas released in the early stage of thermal runaway can be detected in time and an early warning is issued, a certain amount of time can be gained for the handling of the thermal runaway failure.

[0004] Common lithium battery state monitoring systems often only collect information such as battery voltage and surface temperature, lacking information such as gas and pressure. The monitoring dimension is single, and there is a lack of unified data processing and diagnostic analysis, making it difficult to accurately reflect the battery operating state and difficult to achieve early warning of battery thermal runaway faults. In addition, different physical quantities often require different devices for detection, which have the disadvantages of complex structure, low integration, large volume, and high cost. Thermal runaway faults often originate from a fault in a single cell and its propagation among surrounding cells. Currently, most detection devices are placed inside the battery pack to detect changes in the internal environment of the battery pack, making it difficult to detect the state of individual cells and give early warnings of faults in a timely manner. Existing battery end caps mainly consist of a cover plate, positive and negative electrode posts, a pressure relief valve, a liquid injection hole, etc., and cannot collect battery state information, with a single function. Integrating the battery state detection module with the battery end cap can achieve detection and fault warning at the individual cell level. The existing technology only uses gas sensors and also has problems such as insufficient information and a high risk of false alarms and missed alarms.

[0005] The information disclosed in the background art section is only used to enhance the understanding of the background of the present invention and may therefore include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] In view of the deficiencies or defects of the existing technology, a battery thermal runaway warning system and method are provided, which collect various information including gas signals, realize the state detection of individual batteries and timely warning of faults, and have high sensitivity, fast detection speed, and can realize the thermal runaway fault detection of large-scale lithium battery energy storage systems.

[0007] The object of the present invention is achieved through the following technical solutions.

[0008] A battery thermal runaway warning system includes,

[0009] At least one battery intelligent end cap device, which includes a state information sampling module, and the state information sampling module includes,

[0010] A gas sensor signal acquisition unit, which is arranged near the pressure relief valve of the battery end cap to measure the characteristic gas content signal obtained from the gas released by the battery fault. Among them, the characteristic gas is the gas released by the battery fault and its decomposition products, and is different from the gas components contained in the normal air environment.

[0011] A battery surface temperature acquisition unit, which is attached to the surface of the battery end cap to collect the surface temperature information of the individual cell.

[0012] A battery pressure acquisition unit, which is attached to the surface of the battery end cap to collect the stress information of the individual cell.

[0013] A battery voltage acquisition unit, which measures the battery voltage data;

[0014] A host computer, which interacts with and processes data of the battery intelligent end cover device via a wireless communication system. The host computer includes

[0015] A Bayesian deep learning module based on temporal Latent Dirichlet Allocation, which accepts the input of temporal data within a time window, predicts the latent topics of the time window, and outputs the results to the Bayesian deep learning module, which gives the final prediction result and uncertainty. A thermal runaway warning unit, which is connected to the Bayesian deep learning based on temporal Latent Dirichlet Allocation to warn of thermal runaway based on the model results.

[0016] In the battery thermal runaway warning system described above, the battery intelligent end cover device further includes a single-chip microcomputer control module and a data transmission module. The single-chip microcomputer control module is respectively connected to the status information sampling module and the data transmission module. The single-chip microcomputer receives the characteristic gas content signal, battery voltage, surface temperature information, and stress information collected by the status information sampling module, and sends them to the host computer as sensor data through the data transmission module.

[0017] In the battery thermal runaway warning system described above, the single-chip microcomputer control module includes a microcontroller with an internal integrated analog-to-digital converter to collect various types of data.

[0018] In the battery thermal runaway warning system described above, the gas sensor signal acquisition unit includes several gas sensors, a logarithmic operational amplifier, and an operational amplifier. The gas sensors are connected to the input terminal of the logarithmic operational amplifier. The output terminal of the logarithmic operational amplifier is connected to the positive input terminal of the operational amplifier. The negative input terminal of the operational amplifier is connected to the output terminal and is connected to the single-chip microcomputer control module.

[0019] In the battery thermal runaway warning system described above, the battery voltage acquisition unit includes an MOS transistor and two voltage-dividing resistors. The drain of the MOS transistor is connected in series with the voltage-dividing resistor, the source is connected to the battery power supply, and the gate is connected to the single-chip microcomputer control module; the battery pressure acquisition unit includes a piezoresistive thin film sensor. The MOS transistor is normally closed to prevent current from flowing through the voltage-dividing resistor. During measurement, the single-chip microcomputer control module controls the MOS transistor to conduct, and the battery voltage is connected to the circuit. After being divided by two resistors, it reaches the ADC sampling range and is collected by the ADC. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip microcomputer control module through an IIC interface.

[0020] In the battery thermal runaway warning system described above, the data transmission module is a WIFI module connected to the single-chip microcomputer control module, and communicates with the host computer through wireless signal transmission.

[0021] In the described battery thermal runaway warning system, the gas sensor signal acquisition unit includes gas sensors that operate sequentially. A barrier film that separates them from the ambient gas covers the windproof caps of the gas sensors that are not in operation. A taut rubber band is pasted on the barrier film, and a heating wire is tied to the rubber band. When enabling a gas sensor, the heating wire is energized to burn the rubber band, causing the barrier film to separate from the surface of the windproof cap and exposing the gas sensor to the environment to start working.

[0022] In the described battery thermal runaway warning system, the wireless communication system adopts a hierarchical structure divided into clusters. The node types are divided into cluster head nodes and in-cluster nodes. Multiple in-cluster nodes are connected to the cluster head nodes, and at the same time, multiple cluster head nodes are connected to the upper computer, forming a three-level hierarchy.

[0023] The Bayesian deep learning module based on temporal latent Dirichlet allocation. The temporal latent Dirichlet allocation module receives the input of temporal data within a time window, predicts the latent topics of the time window, and outputs the results to the Bayesian deep learning module, which gives the final prediction result and uncertainty.

[0024] The warning method of the battery thermal runaway warning system includes the following steps:

[0025] The gas sensor signal acquisition unit is arranged near the pressure relief valve of the battery end cover to measure the characteristic gas content signal of the gas released by the battery failure. The battery surface temperature acquisition unit is attached to the surface of the battery end cover to collect the surface temperature information of the single cell. The battery pressure acquisition unit is attached to the surface of the battery end cover to collect the stress information of the single cell. The battery voltage acquisition unit measures the battery voltage data. Among them, the battery voltage data, surface temperature information, stress information, and characteristic gas content signal at the same moment constitute a characteristic vector reflecting the battery state and maintain time alignment.

[0026] The characteristic vector is input into the lithium battery thermal runaway warning model to evaluate the thermal runaway risk. The lithium battery thermal runaway warning model is a Bayesian deep learning model based on temporal latent Dirichlet allocation.

[0027] Compared with the prior art, the beneficial effects brought by the present invention are:

[0028] The present invention adopts multi-parameter fusion for battery thermal runaway state monitoring and early warning. Compared with the existing detection methods based on voltage, temperature, or smoke, etc., it has high accuracy and significantly advances the early warning time, providing more time for fault handling. At the same time, it can analyze the uncertainty of the battery state and improve the early warning accuracy. The battery intelligent end cover device has the advantages of long life, small size, low power consumption, and low cost. It uses wireless self-organizing network connection, has a flexible deployment method, reliable data transmission, and can be applied to large-scale lithium battery energy storage devices with a large number of single cells. The gas sensor signal acquisition circuit of the logarithmic operational amplifier has the characteristics of wide range, high precision, and fast acquisition speed.

[0029] The above description is only an overview of the technical solution of the present invention. In order to make the technical means of the present invention clearer and to the extent that those skilled in the art can implement it according to the content of the specification, and in order to make the above and other purposes, features, and advantages of the present invention more obvious and understandable, the following takes the specific implementation manners of the present invention as examples for illustration. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] By reading the detailed description of the preferred specific implementation manners below, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The drawings in the specification are only for the purpose of showing the preferred implementation manners and are not considered as a limitation of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. And throughout the drawings, the same reference numerals are used to represent the same components.

[0031] In the drawings:

[0032] Figure 1 is a schematic structural diagram of a thermal runaway early warning system provided by an embodiment of the present disclosure;

[0033] Figure 2 is a structural diagram of a battery intelligent end cover device provided by an embodiment of the present disclosure;

[0034] Figure 3 is a circuit schematic diagram of a gas sensor signal acquisition unit in a battery intelligent end cover device provided by an embodiment of the present disclosure;

[0035] Figure 4 is a schematic diagram of a gas barrier method provided by an embodiment of the present disclosure;

[0036] Figure 5 is a circuit schematic diagram of a battery voltage acquisition circuit provided by an embodiment of the present disclosure;

[0037] Figure 6A distributed self-organizing network structure diagram provided by an embodiment of the present disclosure;

[0038] Figure 7 A distributed self-organizing network access flowchart provided by an embodiment of the present disclosure;

[0039] Figure 8 A distributed self-organizing network data transmission flowchart provided by an embodiment of the present disclosure;

[0040] Figure 9 An upper computer software structure diagram provided by an embodiment of the present disclosure;

[0041] Figure 10 A schematic diagram of a battery thermal runaway state monitoring and warning algorithm provided by an embodiment of the present disclosure;

[0042] Figure 11 A battery thermal runaway state monitoring and warning algorithm structure diagram provided by an embodiment of the present disclosure;

[0043] Figure 12 A data graph obtained from a simulated lithium battery thermal runaway experiment conducted by an embodiment of the present disclosure;

[0044] Figure 13 A data graph obtained from another simulated lithium battery thermal runaway experiment conducted by an embodiment of the present disclosure;

[0045] Figure 14 A data graph obtained from a simulated lithium battery leakage experiment conducted by an embodiment of the present disclosure;

[0046] Figure 15 A result graph of a battery thermal runaway state monitoring and warning algorithm provided by an embodiment of the present disclosure.

[0047] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments. Specific Embodiments

[0048] The specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0049] It should be noted that in the specification and claims, certain terms are used to refer to specific components. Those skilled in the art should understand that technicians may use different terms to refer to the same component. The specification and claims do not use the difference in terms as a way to distinguish components, but use the difference in the functions of components as the criterion for distinction. For example, the term "comprising" or "including" mentioned throughout the specification and claims is an open-ended term, so it should be interpreted as "including but not limited to". The subsequent description in the specification is the preferred embodiment for implementing the present invention, but the description is for the purpose of the general principles of the specification and is not used to limit the scope of the present invention. The protection scope of the present invention shall be subject to what is defined by the appended claims.

[0050] For the convenience of understanding the embodiments of the present invention, the following will further explain with several specific embodiments in conjunction with the drawings, and each drawing does not constitute a limitation to the embodiments of the present invention.

[0051] For better understanding, as Figures 1 to 15 shown, a battery thermal runaway warning system includes

[0052] at least one battery intelligent end - cover device, which includes a state information sampling module, and the state information sampling module includes

[0053] a gas sensor signal acquisition unit, which is arranged near the pressure relief valve of the battery end - cover to measure the characteristic gas content signal obtained from the gas released by the battery failure. Among them, the characteristic gas is the gas released by the battery failure and its decomposition products, and is different from the gas components contained in the normal air environment.

[0054] a battery surface temperature acquisition unit, which is attached to the surface of the battery end - cover to acquire the surface temperature information of the single - cell battery.

[0055] a battery pressure acquisition unit, which is attached to the surface of the battery end - cover to acquire the stress information of the single - cell battery.

[0056] a battery voltage acquisition unit, which measures the battery voltage data;

[0057] a host computer, which interacts with and processes data of the battery intelligent end - cover device via a wireless communication system. The host computer includes

[0058] a Bayesian deep - learning module based on temporal latent Dirichlet allocation, which fuses multiple sensor data, extracts features, and establishes a multi - parameter lithium - battery thermal runaway warning model. The model structure and training steps are as follows.

[0059] Step S401: Divide the time - series data into sliding windows (one window per second), and take the data within a 60 - second time window. DAs a sample, the data within each time window is regarded as a "document", and the joint state (i.e., the feature vector) of the sensor variables at each time point is regarded as a "word". First, the four types of sensor data included in the feature vector are normalized according to the following formula to obtain D’ .

[0060]

[0061] where X i，k ’ and x i，k ’ are the sensor data before and after normalization respectively, and X i，max ’ and X i，min ’ are the maximum and minimum values of the sensor data respectively.

[0062] The normalized sensor data is discretized. Each sensor is divided into 10 discrete intervals, and the intervals are encoded using integers to obtain D’’ .

[0063]

[0064] where x i,k ’’ is the D’’ in the i th row and the k th column element, representing the data encoding after discretization. The Floor() function represents rounding down.

[0065] The training set, validation set, and test set are divided in the ratio of 3:1:1.

[0066] Step S402: Train the temporal latent Dirichlet model using the discretized sensor data D’’ as the model input. Initialize the global parameters: the topic-sensor matrix and the state transition matrix.

[0067] Topic-sensor distribution:

[0068] where β k is a probability vector representing the probability of different sensor discrete values under topic k. The dimension is the total number of categories of sensor discrete values, and η is a hyperparameter that controls the sparsity of the initial topic distribution.

[0069] Initial topic distribution:

[0070] where θ1 represents the topic probability distribution of the first time window, and α is a hyperparameter.

[0071] State transition matrix:

[0072]

[0073]

[0074] where π j represents the probability vector of transitioning from the hidden state j to other hidden states, Π is the state transition matrix, and each element π in the matrix j,k represents the probability of transitioning from the hidden state j to the hidden state k, and γ is a hyperparameter.

[0075] Step S403: Sample the hidden state of the current window based on the hidden state and transition matrix of the previous window, and generate the topic probability of the Dirichlet distribution based on the current hidden state.

[0076] For each time window t = 2, …, 60:

[0077] Hidden state of window t:

[0078] where z t represents the hidden state of time window t.

[0079] Topic distribution of window t:

[0080] where θ t is the topic probability distribution of window t, δ zt is a one-hot vector that is 1 only at z t , κ > 0, and α is a hyperparameter that controls the strength of the hidden state on the topic distribution.

[0081] Step S404: Maximize the ELBO objective function, which includes the data likelihood term and the KL divergence between the variational distribution and the prior:

[0082]

[0083] where w is the observed data, z is the hidden state sequence, θ is the window topic distribution sequence, β is the topic-sensor matrix, and Π is the state transition matrix.

[0084] Substitute the joint probability and variational distribution into the above formula, decompose the ELBO, use variational inference, and alternately optimize the hidden state distribution, transition matrix parameters, and topic-sensor matrix to gradually improve the model fitting ability through iterative optimization. Update the parameters according to the following formula:

[0085] Hidden state distribution:

[0086] where φ t,k is the hidden state distribution under the variational distribution, representing the probability that window t is in the hidden state k under the variational distribution, Eq [ ] represents expectation, , which is the assignment of the topic corresponding to the n-th sensor value in window t, , and is the discrete sensor value actually observed.

[0087] Topic-sensor distribution:

[0088]

[0089] In the formula, W k,l is the sensor distribution parameter of topic k under the variational distribution, I() is the indicator function, which is 1 when w t,n equals l and 0 otherwise, where l is the data encoding after discretization of a certain sensor.

[0090] Transition matrix distribution:

[0091]

[0092] In the formula, Ψ j,k is the transition probability function of the hidden state j under the variational distribution.

[0093] Window topic distribution:

[0094]

[0095] In the formula, v t,k is the topic distribution parameter of window t under the variational distribution.

[0096] Step S405: Finally, obtain the latent topic distribution and the topic-sensor association matrix W θ . θ t is a vector with the same dimension as the battery fault state A, that is, the dimension is 5, and its elements represent the probability that the topic of this time window is a certain fault state. W θ is a 5×4 matrix, representing the correlation degree between the topic and a certain sensor. 5 is the dimension of the latent topic distribution θ t dimension, and 4 is the number of sensors.

[0097] Step S406: Train the Bayesian neural network. The network input is composed of two parts spliced together: 1. The latent topic distribution generated by the temporal latent Dirichlet model, 2. The features obtained by weighting the normalized sensor time series data using the topic-sensor association matrix. The specific method is: broadcast the latent topic distribution to each time step; multiply the association vector between the latent topic with the highest probability and the sensor by the normalized sensor time series data points to obtain the weighted features; finally, splice the two to obtain the input data m of the Bayesian neural network. The output of the Bayesian neural network is the predicted state A0, which is a vector with a dimension of 5, and the elements represent the probability of each battery fault state.

[0098]

[0099] where θ t 1×4 is the potential topic distribution after broadcasting, and W θt 1×4 is the association vector between the potential topic with the highest probability and the sensor, D’ 4×60 is the normalized sensor time series data.

[0100] Construct a Bayesian neural network structure: Establish a Bayesian model of the long short-term memory network (LSTM-BDL), set the input unit to 9, the time step to 60, and the number of samples in each batch to 32. The LSTM layer and the Dropout layer are stacked to build a 9-layer network, as shown in Table 1, where the LSTM layer is connected one-to-one with the next Dropout layer, the Dropout layer is fully connected to the next LSTM layer, and the Dropout layer is fully connected to the Dense layer.

[0101] Table 1 Network structure

[0102]

[0103] Step S407: For the LSTM layer, calculate the cell state c t and the hidden state h t updated by the gating mechanism at each time step t:

[0104]

[0105] where W f , W i , W o , W c are the gating weight matrices, and b f , b i , b o , b c are the bias terms, represents element-wise multiplication.

[0106] For the initial parameters, use Xavier initialization to avoid activation value saturation or gradient vanishing:

[0107]

[0108] where n in , n out are the input and output dimensions of the weight matrix.

[0109] Step S408: During each forward propagation round, randomly discard some neurons, and the proportion of discarded neurons is as shown by the dropout rate in Table 1. For the Dropout layer, use the Bernoulli function to randomly generate a probability vector (randomly generate a vector with elements 0 and 1, and the proportion conforms to the dropout rate):

[0110]

[0111] where d i (l) is an element in the probability vector d (l) generated by the Bernoulli function, which conforms to the binomial distribution, and p is 1 minus the dropout rate.

[0112] Multiply the input of the Dropout layer by d (l) to obtain the input after deletion:

[0113]

[0114] where , are the inputs before and after deletion respectively.

[0115] After that, calculate the output:

[0116]

[0117] where is the output of this layer, f() is the activation function, W and b are the weight vector and bias value, represents element-wise multiplication.

[0118] Step S409: Calculate the loss function:

[0119]

[0120] where is the true label; is the probability predicted by the model.

[0121] Step S410: Through backpropagation through time, calculate the gradient and take the derivative of the parameter θ = {W f,k , W i,k , W o,k , W c,k , b f,k , b i,k , b o,k , b c,k , W j , b j}:

[0122]

[0123] To avoid gradient explosion, clip the gradients:

[0124]

[0125] Step S411: Update the parameters using the Adam algorithm:

[0126]

[0127] In the formula, η is the learning rate.

[0128] If the model does not converge, replace it with different optimizers such as SGD, RMSprop, etc., and appropriately increase the number of training epochs. At the same time, to avoid overfitting, use the early stopping strategy. When the loss on the validation set no longer decreases, stop the training.

[0129] The thermal runaway warning unit is connected to the Bayesian deep learning module based on temporal latent Dirichlet allocation to warn of thermal runaway based on the model results. The thermal runaway warning unit sends the sensor data collected within a certain time window to the Bayesian deep learning module based on temporal latent Dirichlet allocation and receives the returned results from the module. According to the returned results, it judges whether to give a warning: when the returned result is "normal" and the uncertainty is less than 0.5, it is considered that the battery is normal and no warning is given; when the returned result is "overheat", "overcharge", "thermal runaway", etc., or the result is "normal", but the uncertainty is greater than 0.5, an alarm is given immediately.

[0130] In a preferred embodiment of the battery thermal runaway warning system, the battery intelligent end cover device further includes a single-chip microcomputer control module and a data transmission module. The single-chip microcomputer control module is respectively connected to the status information sampling module and the data transmission module. The single-chip microcomputer receives the characteristic gas content signal, battery voltage, surface temperature information, and stress information collected by the status information sampling module and sends them to the upper computer as sensor data through the data transmission module.

[0131] In a preferred embodiment of the battery thermal runaway warning system, the single-chip microcomputer control module includes a microcontroller with an internal integrated analog-to-digital converter to collect various types of data.

[0132] In a preferred embodiment of the battery thermal runaway warning system, the gas sensor signal acquisition unit includes several gas sensors, a logarithmic operational amplifier, and an operational amplifier. The gas sensors are connected to the input end of the logarithmic operational amplifier. The output end of the logarithmic operational amplifier is connected to the positive input end of the operational amplifier. The negative input end of the operational amplifier is connected to the output end and is connected to the single-chip microcomputer control module.

[0133] In a preferred embodiment of the battery thermal runaway warning system, the battery voltage acquisition unit includes an MOS transistor and two voltage-dividing resistors. The drain of the MOS transistor is connected in series with the voltage-dividing resistor, the source is connected to the battery power supply, and the gate is connected to the single-chip microcomputer control module. The battery pressure acquisition unit includes a piezoresistive thin film sensor. The MOS transistor is normally closed to prevent current from flowing through the voltage-dividing resistor. During measurement, the single-chip microcomputer control module controls the MOS transistor to conduct, and the battery voltage is connected to the circuit. After being divided by the two resistors, it is brought within the ADC sampling range and collected by the ADC. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip microcomputer control module through the IIC interface.

[0134] In a preferred embodiment of the battery thermal runaway warning system, the data transmission module is a WIFI module connected to the single-chip microcomputer control module, and communicates with the upper computer through wireless signal transmission.

[0135] In a preferred embodiment of the battery thermal runaway warning system, the gas sensor signal acquisition unit includes gas sensors that work sequentially. A barrier film that separates it from the ambient gas is covered on the windproof cap surface of the gas sensor that is not working. A taut rubber band is pasted on the barrier film, and a heating wire is tied to the rubber band. When the gas sensor is enabled, the heating wire is powered on to burn the rubber band, and the barrier film is detached from the windproof cap surface to expose the gas sensor to the environment to start working.

[0136] In a preferred embodiment of the battery thermal runaway warning system, the wireless communication system adopts a hierarchical structure divided into clusters. The node types are divided into cluster head nodes and in-cluster nodes. Multiple in-cluster nodes are connected to the cluster head node, and at the same time, multiple cluster head nodes are connected to the upper computer to form a three-level hierarchy.

[0137] In a preferred embodiment of the battery thermal runaway warning system, the Bayesian deep learning module based on temporal latent Dirichlet allocation includes a temporal latent Dirichlet allocation model unit, which analyzes the temporal changes of sensor data to identify potential degradation trends and thermal runaway risks.

[0138] The warning method of the battery thermal runaway warning system includes the following steps,

[0139] The gas sensor signal acquisition unit is arranged near the pressure relief valve of the battery end cover to measure the characteristic gas content signal of the gas released by the battery failure. The battery surface temperature acquisition unit is attached to the surface of the battery end cover to collect the surface temperature information of the single cell. The battery pressure acquisition unit is attached to the surface of the battery end cover to collect the stress information of the single cell. The battery voltage acquisition unit measures the battery voltage data. Among them, the battery voltage data, surface temperature information, stress information, and characteristic gas content signal at the same moment constitute a characteristic vector reflecting the battery state and maintain time alignment;

[0140] Input the feature vector into the lithium battery thermal runaway warning model to evaluate the thermal runaway risk. The lithium battery thermal runaway warning model is a Bayesian deep learning model based on temporal latent Dirichlet allocation. Among them, the model accepts the outputs of multiple sensors within a certain time window, uses the temporal latent Dirichlet allocation model to predict the latent topics of the time window, and outputs them to the Bayesian deep learning model. The temporal data is processed through the LSTM layer, and the model structure is adjusted through the dropout layer to construct a Bayesianized model to evaluate the thermal runaway risk.

[0141] In one embodiment, the shape of the circuit board of the battery state monitoring device is designed to match the battery end cap. On the premise of not affecting the function of the end cap itself, the battery state detection is realized, so as to develop a battery intelligent end cap of the battery state monitoring device integrating multi-dimensional information such as gas sensing and voltage. The gas sensor is arranged near the pressure relief valve of the end cap to quickly respond to the gas released by the battery failure; the temperature and stress detection part is closely attached to the surface of the battery end cap to accurately collect the state information of the single cell.

[0142] The single-chip microcomputer control module is used to control the operating state of the device, collect the state information of each battery, and communicate with the host computer. The state information sampling module is used to collect the signal of the gas content of the battery fault characteristics and the voltage, surface temperature, and stress information of the battery. The data transmission module is used to send the data collected by the device to the host computer for analysis and storage. The main feature of the battery state monitoring device is that the state information sampling module is connected to the single-chip microcomputer control module, and the single-chip microcomputer control module is connected to the data transmission module. The single-chip microcomputer control module includes a microcontroller, and a high-precision analog-to-digital converter is integrated inside the microcontroller for accurately collecting various types of data. The state information sampling module includes a gas sensor signal acquisition unit, a battery voltage acquisition unit, and a battery surface temperature and stress acquisition unit. The gas sensor signal acquisition unit includes several semiconductor gas sensors, a logarithmic operational amplifier, and an operational amplifier; the gas sensor is connected to the input end of the logarithmic operational amplifier, the output end of the logarithmic operational amplifier is connected to the positive input end of the operational amplifier, the negative input end of the operational amplifier is connected to the output end, and is connected to the single-chip microcomputer control module. The battery voltage acquisition unit includes a MOS tube and two voltage-dividing resistors. The drain of the MOS tube is connected in series with the voltage-dividing resistor, the source is connected to the battery power supply, and the gate is connected to the single-chip microcomputer control module. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip microcomputer control module through an IIC interface, and the stress is collected by a piezoresistive thin film sensor. The MOS tube is normally closed to avoid current flowing through the voltage-dividing resistor, resulting in additional power consumption. During measurement, the single-chip microcomputer control module controls the MOS tube to conduct, and the battery voltage is connected to the circuit. After being divided by two resistors, it reaches the ADC sampling range and is collected by the ADC. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip microcomputer control module through an IIC interface. The data transmission module is a WIFI module, which is connected to the single-chip microcomputer control module and communicates with the host computer through wireless signal transmission. Thus, the complex wiring harness connection in the wired-connected battery management system is avoided.

[0143] Since the lifespan of semiconductor gas sensors is relatively short, generally only 4 to 5 years, while the battery lifespan can be up to 10 years or more. To solve the contradiction of the mismatched lifespans between the two, multiple sensors work sequentially to extend the device's working lifespan. To protect the sensors that are not working from environmental gas poisoning, a gas diaphragm is used to separate them from the environmental gas. The said barrier film is electrostatically adsorbed on the surface of the sensor's wind cap. A taut rubber band is pasted on the film, with both ends of the rubber band fixed to the device, and a heating wire is tied to the rubber band. When the sensor starts to work, the heating wire is powered on, and its temperature rises, burning the rubber band. When the rubber band is not burned, the barrier film is under the simultaneous action of the pulling forces at both ends, and the pulling forces at both ends cancel each other out; when one end of the rubber band is burned, the barrier film is only under the pulling force at one end and separates from the wind cap, exposing the sensor to the environment and starting to work normally.

[0144] The said wireless communication system classifies node types into cluster head nodes and in-cluster nodes, and the network is established through the following process. Multiple battery intelligent end cap devices integrated with gas sensors are automatically interconnected with other surrounding devices through self-organization via WIFI, and a cluster head node is selected and generated according to the access algorithm. The cluster head nodes are interconnected and connected to the terminal equipped with the upper computer software, forming a wireless communication system with a hierarchical structure. The said access algorithm is described as follows: when a new battery intelligent end cap device starts to work, it will automatically detect whether there is a cluster head node around. If there is one and the cluster head node has not reached the access capacity limit, it will connect to this cluster head node and send its own number and address information to the cluster head node for storage. If there is no cluster head node or the existing cluster head nodes have reached the access capacity limit, it will become a new cluster head node and connect to the terminal equipped with the upper computer software. When a new cluster head node is established or a new node accesses, this cluster head node will update the connection information to the terminal. The said connection information is the cluster head node number and address, as well as the number and address of the newly connected node.

[0145] When the terminal requests data from a specific node, it sends a message to the cluster head node connected to this node, and the cluster head node transmits the information to the target node. When the target node feeds back data, it sends the message to the terminal equipped with the upper computer software along the reverse path of the above. When traversing all nodes to request data, the terminal sends messages to all cluster head nodes and then waits for a reply. After receiving the instruction, the cluster head node, according to the order of the connection information table, first sends a message to the first node and waits for a reply. If a reply is received, the data is saved; if no reply is received within the specified time, only the node information is recorded. Then the above process is repeated for the next node until all nodes connected to this cluster head node are traversed. After completing the traversal of the connected nodes, the cluster head node sends the aggregated data and the data collected by itself to the terminal.

[0146] The host computer software running on the terminal has the functions of summarizing the data of each sensor, executing the battery thermal runaway state monitoring and warning algorithm for multi-parameter fusion, and displaying the data and battery state information. A Bayesian deep learning module based on temporal latent Dirichlet allocation is adopted to fuse various sensor data, extract comprehensive features, and establish a lithium battery thermal runaway warning model that integrates multiple parameters such as gas, voltage, and temperature. First, by Bayesianizing the model parameters, the cognitive uncertainty of the model in different states is quantified, and at the same time, an accidental node layer is added to the last layer of the model to quantify the uncertainty related to the data. Secondly, for the measurement parameters such as gas, voltage, and temperature, a temporal latent Dirichlet allocation (TLDA) model is introduced to analyze the temporal changes of sensor data and identify potential degradation trends and thermal runaway risks. Finally, the effectiveness of the model in predicting the accuracy, response time, and uncertainty estimation of lithium battery thermal runaway is evaluated.

[0147] In one embodiment, the lithium battery thermal runaway warning includes the following steps:

[0148] Step S100: The battery intelligent end cover device collects the battery voltage Vol, surface temperature Temp, pressure Pres, and gas sensor resistance Res data as feature vectors, and the sampling frequency is 1 Hz.

[0149] Step S200: For each physical quantity (denoted as X, where X can take Vol, Temp, Pres, and Res), the following formula is used to perform low-pass filtering using a Butterworth filter:

[0150]

[0151] In the formula, X j is the data in the previous j seconds, n is the filter order, which can be taken as 11, and h(j) is the filter coefficient that can be generated by relevant software.

[0152] Step S300: Take the filtered data within a 60-second time window before time t as the model input D:

[0153]

[0154] In the formula, Vol, Temp, Pres, and Res are all 1×60 tensors.

[0155] Step S400: Use the input data D to calculate the output of the lithium battery thermal runaway warning model multiple times, take the mean of multiple prediction results, and calculate the information entropy to obtain the predicted battery fault state A (normal, overheating, overcharging, electrolyte leakage, thermal runaway) and the uncertainty measure Q, where Q ∈ (0, 1).

[0156] Step S500: Determine the battery status. If an abnormality occurs, issue a high-level alarm; if it is "normal", determine whether the uncertainty measure Q is greater than 0.5. If so, issue a low-level alarm. Then, repeat Step S200.

[0157] In the method described above, the method for constructing the lithium battery thermal runaway early warning model described in Step S400 includes the following steps:

[0158] Step S401: Divide the time series data into sliding windows (one window per second), and take the data within a 60-second time window D as a sample. The data within each time window is regarded as a "document", and the joint state (i.e., the feature vector) of the sensor variables at each time point is regarded as a "word". First, normalize the four types of sensor data included in the feature vector according to the following formula to obtain D’ .

[0159]

[0160] In the formula, X i，k ’ , x i，k ’ are the sensor data before and after normalization respectively, X i，max ’ , X i，min ’ are the maximum and minimum values of the sensor data respectively.

[0161] Discretize the normalized sensor data. Each sensor is divided into 10 discrete intervals, and integer encoding is used for the intervals to obtain D’’ .

[0162]

[0163] In the formula, x i,k ’’ is D’’ in the i th row and the k th column element, representing the data encoding after discretization. The Floor() function represents rounding down.

[0164] Divide the training set, validation set, and test set in a ratio of 3:1:1.

[0165] Step S402: Train the temporal latent Dirichlet model, using the discretized sensor data D’’ as the model input. Initialize the global parameters: the topic-sensor matrix and the state transition matrix.

[0166] Topic-sensor distribution:

[0167] where, β k is a probability vector, representing the probabilities of different sensor discrete values under topic k, with the dimension being the total number of categories of sensor discrete values, and η is a hyperparameter that controls the sparsity of the initial topic distribution.

[0168] Initial topic distribution:

[0169] where, θ1 represents the topic probability distribution of the first time window, and α is a hyperparameter.

[0170] State transition matrix:

[0171]

[0172]

[0173] where, π j represents the probability vector of transitioning from hidden state j to other hidden states, Π is the state transition matrix, and each element π j,k in the matrix represents the probability of transitioning from hidden state j to hidden state k, and γ is a hyperparameter.

[0174] Step S403: Sample the hidden state of the current window based on the hidden state of the previous window and the transition matrix, and generate the topic probability of the Dirichlet distribution based on the current hidden state.

[0175] For each time window t = 2, …, 60:

[0176] Hidden state of window t:

[0177] where, z t represents the hidden state of time window t.

[0178] Topic distribution of window t:

[0179] where, θ t is the topic probability distribution of window t, δ zt is a one-hot vector, which is 1 only at z t , κ > 0, and α is a hyperparameter that controls the strength of the hidden state on the topic distribution.

[0180] Step S404: Maximize the ELBO objective function, which includes the data likelihood term and the KL divergence between the variational distribution and the prior:

[0181]

[0182] where, w is the observed data, z is the hidden state sequence, θ is the window topic distribution sequence, β is the topic-sensor matrix, and Π is the state transition matrix.

[0183] Substitute the joint probability and variational distribution into the above formula, decompose the ELBO, and use variational inference. By alternately optimizing the hidden state distribution, transition matrix parameters, and topic-sensor matrix, the model fitting ability is gradually improved through iterative optimization. Update the parameters iteratively according to the following formula:

[0184] Hidden state distribution:

[0185] In the formula, φ t,k is the hidden state distribution under the variational distribution, representing the probability that window t is in hidden state k under the variational distribution, and E q [ ] represents the expectation, , is the topic assignment corresponding to the nth sensor value in window t, , is the actually observed discrete sensor value.

[0186] Topic-sensor distribution:

[0187]

[0188] In the formula, W k,l is the sensor distribution parameter of topic k under the variational distribution, and I() is the indicator function, which is 1 when w t,n is equal to l and 0 otherwise, where l is the data encoding after discretization of a certain sensor.

[0189] Transition matrix distribution:

[0190]

[0191] In the formula, Ψ j,k is the transition probability function of hidden state j under the variational distribution.

[0192] Window topic distribution:

[0193]

[0194] In the formula, v t,k is the topic distribution parameter of window t under the variational distribution.

[0195] Step S405: Finally, obtain the latent topic distribution and the topic-sensor association matrix W θ . θ t is a vector with the same dimension as the battery failure state A, that is, the dimension is 5, and its elements represent the probability that the topic of this time window is a certain failure state. W θ is a 5×4 matrix, representing the correlation degree between the topic and a certain sensor. 5 is the dimension of the latent topic distribution θ t dimension, and 4 is the number of sensors.

[0196] Step S406: Train a Bayesian neural network. The network input is composed of two concatenated parts: 1. The latent topic distribution generated by the temporal latent Dirichlet model; 2. The features obtained by weighting the normalized sensor time series data using the topic-sensor association matrix. The specific method is as follows: Broadcast the latent topic distribution to each time step; Multiply the association vector of the latent topic with the highest probability and the sensor by the normalized sensor time series data points to obtain the weighted features; Finally, concatenate the two to obtain the input data m of the Bayesian neural network. The output of the Bayesian neural network is the predicted state A 0 , which is a vector with a dimension of 5, where the elements represent the probabilities of each battery failure state.

[0197]

[0198] In the formula, θ t 1×4 is the broadcast latent topic distribution, W θt 1×4 is the association vector of the latent topic with the highest probability and the sensor, D’ 4×60 is the normalized sensor time series data.

[0199] Construct the Bayesian neural network structure: Establish a Bayesian model of the long short-term memory network (LSTM-BDL). The input unit is set to 9, the time step is 60, and the number of samples in each batch is 32. The LSTM layer and the Dropout layer are stacked to build a 9-layer network, as shown in Table 1, where the LSTM layer is connected one-to-one with the next Dropout layer, the Dropout layer is fully connected to the next LSTM layer, and the Dropout layer is fully connected to the Dense layer.

[0200] Step S407: For the LSTM layer, calculate the cell state c t and the hidden state h t at each time step t through the gating mechanism:

[0201]

[0202] In the formula, W f , W i , W o , W c are the gating weight matrices, b f , b i , b o , b c are the bias terms, represents element-wise multiplication.

[0203] For the initial parameters, use Xavier initialization to avoid activation value saturation or gradient vanishing:

[0204]

[0205] where n in , n out are the input and output dimensions of the weight matrix.

[0206] Step S408: During each round of forward propagation, randomly discard some neurons. The proportion of discarded neurons is shown as the dropout rate in Table 1. For the Dropout layer, use the Bernoulli function to randomly generate a probability vector (a vector with randomly generated elements 0 and 1, with the proportion conforming to the dropout rate):

[0207]

[0208] where d i (l) is an element in the probability vector d (l) generated by the Bernoulli function, which conforms to the binomial distribution, and p is 1 minus the dropout rate.

[0209] The input of the Dropout layer is multiplied by d (l) to obtain the input after deletion:

[0210]

[0211] where , are the input before and after deletion respectively.

[0212] After that, calculate the output:

[0213]

[0214] where, is the output of this layer, f() is the activation function, W and b are the weight vector and bias value, represents element-wise multiplication.

[0215] Step S409: Calculate the loss function:

[0216]

[0217] where, is the true label; is the probability predicted by the model.

[0218] Step S410: Through backpropagation through time, calculate the gradient and update the parameters θ = {W f,k , W i,k , W o,k , W c,k , bf,k , b i,k , b o,k , b c,k , W j , b j Derivation:

[0219]

[0220] To avoid gradient explosion, clip the gradient:

[0221]

[0222] Step S411: Update the parameters using the Adam algorithm:

[0223]

[0224] In the formula, η is the learning rate.

[0225] If the model does not converge, replace it with different optimizers such as SGD, RMSprop, etc., and appropriately increase the number of training epochs; at the same time, to avoid overfitting, use the early stopping strategy and stop training when the loss on the validation set no longer decreases.

[0226] In one embodiment, as Figures 1 to 11 shown, the structural diagram of the battery thermal runaway warning system is as Figure 1 shown. The structural diagram of the battery intelligent end cover device integrated with a gas sensor is as Figure 2 shown, including a status information sampling module, a data transmission module, and a single-chip microcomputer control module. The status information sampling module includes a gas sensor signal acquisition unit, a battery voltage acquisition unit, a battery surface temperature and pressure acquisition unit.

[0227] The gas sensor signal acquisition unit includes two semiconductor gas sensors and a signal acquisition circuit. The two semiconductor gas sensors respectively detect combustible gas and battery electrolyte vapor. Since the resistance difference between the two gas sensors is large and the resistance change range during response is large, the commonly used resistance voltage division method is difficult to meet the requirements of measurement accuracy and range. Therefore, the structural diagram of the gas sensor signal acquisition circuit adopted is as Figure 3As shown in the figure, in the figure, 1 to 17 are the pins of the logarithmic operational amplifier TPA8304. Among them, pin 1 is the NC pin and is not connected to the inside of the chip; pins 2 and 4 are the VSUM pins, and the two are connected to protect the current input; pin 3 is the input pin and is connected to the sensor signal terminal; pin 5 is the photodiode laser output bias pin and is not used in this embodiment; pin 6 is the reference voltage output pin and is not used in this embodiment; pin 7 is the analog ground pin and is connected to the analog ground of the circuit board; pin 8 is the logarithmic voltage output pin and is connected to pin 9; pin 9 is the positive input terminal of the operational amplifier inside the chip and amplifies the voltage output from pin 8; pins 10 and 12 are the positive power supply of the chip and are connected to the 5V power supply; pin 11 is the voltage output pin and is connected to the output terminal of the operational amplifier inside the chip; pin 13 is the negative input terminal of the operational amplifier inside the chip, and the amplification factor is adjusted by the resistor connected to it; pins 14 and 15 are the power ground of the chip and are connected to the analog ground of the circuit board; pin 16 is the power-down control pin and is connected to the analog ground of the circuit board to keep the chip working all the time; pin 17 is the heat sink and is connected to pins 2 and 4 to reduce the leakage of the input current. R1 and R2 are voltage-dividing resistors. R1 is connected between pin 13 and pin 14, and R2 is connected between pin 11 and pin 13 to amplify the output of the logarithmic operational amplifier in the same direction by 2 times. R3 is a current sampling resistor and is connected to the sensor heating end to monitor the sensor heating current; the sensor signal terminal is connected in series with the protection resistor R4 and then connected to the input terminal of the logarithmic operational amplifier, that is, pin 3. The output terminal 11 of the logarithmic operational amplifier is connected to the single-chip microcomputer control module. The logarithmic operation amplifier chip performs a logarithmic operation on the current signal with a very large change range flowing through the sensor to obtain a voltage signal that changes in the same order of magnitude. The analog signal is converted into a digital signal by the ADC, and an inverse operation is performed in the single-chip microcomputer control module to obtain the original current signal and calculate the sensor resistance.

[0228] To extend the working life of the device, two semiconductor gas sensors are used for each type, and the sensors work in sequence. To protect the sensors put into use later from environmental gas poisoning, a gas diaphragm is used to separate them from the environmental gas. As Figure 4 shown, the barrier film is electrostatically adsorbed on the surface of the sensor wind cap. A tight rubber band is pasted on the film, and both ends of the rubber band are fixed on the device, and a heating wire is tied to the rubber band. When the first sensor reaches its expected life, the heating wire is energized, and the temperature of the heating wire rises, burning the rubber band. When the rubber band is not burned, the barrier film is under the simultaneous action of the pulling forces at both ends, and the pulling forces at both ends cancel each other out; when one end of the rubber band is burned, the barrier film is only under the pulling force at one end and separates from the wind cap, exposing the second sensor to the environment and starting to work normally.

[0229] The battery voltage acquisition unit is as Figure 5As shown in the figure, it includes a MOS transistor and two voltage-dividing resistors. The drain of the MOS transistor is connected in series with the voltage-dividing resistor, the source is connected to the battery power supply, and the gate is connected to the single-chip microcomputer control module. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip microcomputer control module through the IIC interface. The MOS transistor is normally closed to avoid current flowing through the voltage-dividing resistor, resulting in additional power consumption. During measurement, the single-chip microcomputer control module controls the MOS transistor to conduct, and the battery voltage is connected to the circuit. After being divided by the two resistors, it reaches within the ADC sampling range and is collected by the ADC. The data transmission module uses a WIFI module with the model number E103-W05B, which is connected to the single-chip microcomputer control module through the serial port to realize wireless communication between the device and the upper computer. The single-chip microcomputer control module includes a microcontroller, and an analog-to-digital converter is integrated inside the microcontroller.

[0230] In a lithium battery energy storage device composed of a large number of single battery packs, an integrated gas sensor battery intelligent end cap device is arranged for each battery; the intelligent end cap device is automatically interconnected with other surrounding devices through WIFI in an ad hoc manner, and a cluster head node is selected according to the access algorithm. The cluster head nodes are interconnected and connected to the terminal equipped with the upper computer software to form a hierarchical wireless communication system as Figure 6 shown.

[0231] As Figure 7 shown, when a new battery intelligent end cap device starts to work, it will automatically detect whether there is a cluster head node around. If there is and the cluster head node has not reached the access capacity limit, it will connect to the cluster head node and send its own number and address information to the cluster head node for storage. If there is no cluster head node or the existing cluster head node has reached the access capacity limit, it will become a new cluster head node and connect to the terminal equipped with the upper computer software. When a new cluster head node is established or a new node is connected, the cluster head node will update the connection information to the terminal. The connection information is the cluster head node number and address, as well as the number and address of the newly connected node.

[0232] The flowchart of the terminal requesting data from a specific node is as Figure 8As shown in the figure, a message is sent to the cluster head node connected to the node, and the cluster head node transmits the information to the target node. When the target node feeds back data, the message is passed to the terminal equipped with the upper computer software along the reverse path of the above. When traversing all nodes to request data, the terminal sends a message to all cluster head nodes and then waits for a reply. After receiving the instruction, the cluster head node, according to the order of the connection information table, first sends a message to the first node and waits for a reply. If a reply is received, the data is saved; if no reply is received within the specified time, only the node information is recorded. Then the above process is repeated for the next node until all nodes connected to the cluster head node are traversed. After completing the traversal of the connected nodes, the cluster head node sends the collected data and the data collected by itself to the terminal.

[0233] The upper computer has the functions of summarizing the data of each sensor, executing the battery thermal runaway state monitoring and early warning algorithm, the temperature drift correction algorithm, and displaying the data and battery state information. The software structure of the upper computer is as Figure 9 shown.

[0234] The early warning algorithm adopts Bayesian deep learning based on temporal latent Dirichlet allocation, fuses various sensor data, extracts comprehensive features, and establishes a lithium battery thermal runaway early warning model integrating multiple parameters such as gas, voltage, and temperature. The schematic diagram of the algorithm is as Figure 10 shown, and the structure diagram is as Figure 11 shown.

[0235] The multi-sensor parameter fusion method is as follows:

[0236] 1. Use multiple sensors to measure the data of the observed target;

[0237] 2. The data within each time window is regarded as a "document", and the joint state of the sensor variables at each time point is regarded as a "word". The temporal latent Dirichlet model is used to obtain the latent topic distribution of the document and the topic-sensor association matrix.

[0238] In another embodiment, in order to verify the actual effect of the battery intelligent end cover device integrated with the gas sensor and the thermal runaway early warning system on the lithium battery thermal runaway early warning, two lithium battery overcharging experiments as described below were carried out.

[0239] The conditions of the two experiments were the same. The experimental objects were 5 series-connected battery cells with a capacity of 628 Ah. Before the start of the experiment, the battery cells were pre-charged to keep them fully charged. Thermocouples were attached to multiple places on the surface of the battery cells to detect the surface temperature and determine the moment of thermal runaway. In addition, after the battery cells were connected in series, they were led out by a cable to apply current for overcharge experiments. The whole battery cell was placed in a metal shell. At the same time, the battery intelligent end cap device integrated with a gas sensor was placed above the battery cells for monitoring. During the experiment, a constant current overcharge experiment with a current of 314 A (0.5C) was carried out on the 5 series-connected battery cells until thermal runaway occurred in all the battery cells.

[0240] The results of the first experiment were as Figure 12 shown. At 14:55:12, the device detected the response of the combustible gas sensor and started to send a warning signal. 21 s later, the pressure relief valve of the first battery cell opened and gas was ejected. 7 s later, the electrolyte vapor sensor also started to respond. At 15:04:05, the first battery cell had a thermal runaway. At this time, 533 s had passed since the device sent the warning signal.

[0241] The results of the second experiment were as Figure 13 shown. At 17:33:18, the pressure relief valve of the first battery cell opened and gas was ejected. Subsequently, the combustible gas sensor immediately responded. 10 s later, the electrolyte vapor sensor also started to respond. At 17:41:52, the first battery cell had a thermal runaway. At this time, 514 s had passed since the device sent the warning signal.

[0242] It can be seen from the two experiments that the warning times of the gas sensor in the device for the thermal runaway fault of the lithium battery were 533 s and 514 s respectively, and it had excellent detection performance.

[0243] In another embodiment, in order to verify the actual effect of the battery intelligent end cap device integrated with the gas sensor and the thermal runaway warning system on the detection of lithium battery leakage, the following simulated leakage experiment was carried out.

[0244] The experiment was carried out in a cube acrylic chamber with a side length of 50 cm. A glass slide was placed at the center of the bottom of the chamber, and the gas detection device was placed on the side wall at a height of about 15 cm from the bottom surface. During the experiment, a pipette was used to suck 10 μL of anhydrous dimethyl carbonate (DMC) liquid and drop it at the center of the glass slide, and the response of the sensor was observed. After the 10 μL of DMC liquid was completely evaporated and evenly diffused in the chamber, the concentration of DMC gas in the chamber was about 23.3 ppm.

[0245] The results of the first experiment were as Figure 14As shown, at 19:04:30, DMC liquid was dropped; about 34 s later, two DMC sensors responded successively; 179 s after the liquid drop was dropped, DMC was completely evaporated, and at this time, the measured DMC concentration was about 26 ppm. Therefore, the device can realize timely early warning of the micro liquid leakage condition of the lithium battery, and the early warning time is about 34 s.

[0246] In another embodiment, in order to verify the effect of the lithium battery thermal runaway early warning model, thermal runaway experiments were carried out under conditions such as overcharge and overheat, sensor data was collected, and the data within a 60-s time window during the experiment was taken as a sample, and the lithium battery thermal runaway early warning model was used to predict the battery state. Some data curves, model outputs, and actual states are as Figure 15 shown. It can be seen that the model accurately predicts the battery state.

[0247] It should be specifically noted that when the characteristic gas content signal is 0, that is, the battery does not generate characteristic gas. At this time, the characteristic gas content signal, as part of the data or information of the characteristic vector, the present invention can still effectively conduct early warning of battery thermal runaway. The root cause is that: the characteristic vector includes various data or information. When one of the data or information appears to be normal on the surface, one or several other data or information may still reach near the critical point. At this time, the present invention may still make a thermal runaway early warning with engineering significance. Therefore, the present invention is not limited to the state where one or several data or information in the characteristic vector are normal or the so-called 0 state.

[0248] The basic principle of the present application has been described above in combination with specific embodiments. However, it should be pointed out that the advantages, advantages, effects, etc. mentioned in the present application are only examples and not limitations. It cannot be considered that these advantages, advantages, effects, etc. are essential for each embodiment of the present application. In addition, the above-disclosed specific details are only for the purpose of illustration and easy understanding, rather than limitations. The above details do not limit the present application to necessarily adopt the above specific details to implement.

[0249] The above description has been given for purposes of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the form disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions, and sub-combinations thereof.

Claims

1. A battery thermal runaway warning system, characterized in that: These include, At least one battery intelligent end cap device, comprising a state information sampling module, wherein the state information sampling module comprises: The gas sensor signal acquisition unit is arranged near the pressure relief valve of the battery end cover to measure the gas released by the battery failure to obtain a characteristic gas content signal, wherein the characteristic gas is the gas released by the battery failure and its decomposition, and is different from the gas components contained in the normal air environment. The battery surface temperature collection unit is attached to the surface of the battery end cover to collect the surface temperature information of the single cell. The battery pressure collection unit is attached to the surface of the battery end cover to collect the stress information of the single cell. A battery voltage acquisition unit, which measures battery voltage data; A host computer, which interacts and processes data with the battery intelligent end cover device via a wireless communication system, and the host computer includes: A Bayesian deep learning module based on time series potential Dirichlet allocation, the Bayesian deep learning module includes a time series potential Dirichlet allocation model, the time series potential Dirichlet allocation model accepts time series data input in a time window, predicts the potential topic of the time window, and outputs the result to the Bayesian deep learning module, and the Bayesian deep learning module gives the final prediction result and uncertainty; A thermal runaway warning unit is connected to the Bayesian deep learning based on time-series potential Dirichlet allocation to warn of thermal runaway based on model results.

2. The battery thermal runaway warning system according to claim 1, characterized in that: The battery intelligent end cover device also includes a single-chip microcomputer control module and a data transmission module. The single-chip microcomputer control module is connected to the status information sampling module and the data transmission module respectively. The single-chip microcomputer receives the characteristic gas content signal, battery voltage, surface temperature information, and stress information collected by the status information sampling module, and sends them to the host computer through the data transmission module as sensor data.

3. The battery thermal runaway warning system according to claim 2, characterized in that: The single-chip control module includes a microcontroller with an internal integrated analog-to-digital converter to collect various types of data.

4. The battery thermal runaway warning system according to claim 2, characterized in that: The gas sensor signal acquisition unit includes several gas sensors, a logarithmic operational amplifier and an operational amplifier. The gas sensor is connected to the input end of the logarithmic operational amplifier, the output end of the logarithmic operational amplifier is connected to the positive input end of the operational amplifier, the reverse input end of the operational amplifier is connected to the output end, and is connected to the single-chip control module.

5. The battery thermal runaway warning system according to claim 2, characterized in that: The battery voltage acquisition unit includes a MOS tube and two voltage-dividing resistors, wherein the drain of the MOS tube is connected in series with the voltage-dividing resistor, the source is connected to the battery power supply, and the gate is connected to the single-chip control module; the battery pressure acquisition unit includes a pressure-sensitive film sensor, the MOS tube is normally closed to prevent current from flowing through the voltage-dividing resistor. During measurement, the single-chip control module controls the MOS tube to be turned on, the battery voltage is connected to the circuit, and after being divided by the two resistors, it reaches the ADC sampling range and is acquired by the ADC. The battery surface temperature acquisition unit is a temperature and humidity sensor, which is connected to the single-chip control module through an IIC interface.

6. The battery thermal runaway warning system according to claim 2, characterized in that: The data transmission module is a WIFI module connected to the single-chip control module, and communicates with the host computer through wireless signal transmission.

7. The battery thermal runaway warning system according to claim 1, characterized in that: The gas sensor signal acquisition unit includes a plurality of gas sensors that work in sequence. The surface of the windproof cap of the non-working gas sensor is covered with a barrier film that separates it from the ambient gas. A tight rubber band is adhered to the barrier film, and a heating wire is tied to the rubber band. When the gas sensor is activated, the heating wire is energized to burn the rubber band, and the barrier film is detached from the surface of the windproof cap to expose the gas sensor to the environment and start working.

8. The battery thermal runaway warning system according to claim 1, characterized in that: The wireless communication system adopts a hierarchical structure divided by clusters, and divides the node types into cluster head nodes and intra-cluster nodes. Multiple intra-cluster nodes are connected to the cluster head node, and multiple cluster head nodes are connected to the host computer, forming a three-level hierarchy.

9. The battery thermal runaway warning system according to claim 1, characterized in that: The Bayesian deep learning based on time series potential Dirichlet allocation includes introducing a time series potential Dirichlet allocation model unit to analyze the time series changes of sensor data to identify potential degradation trends and thermal runaway risks.

10. The early warning method of the battery thermal runaway early warning system according to any one of claims 1 to 9, characterized in that: It includes the following steps, The gas sensor signal acquisition unit is arranged near the pressure relief valve of the battery end cover to measure the gas released by the battery failure to obtain a characteristic gas content signal. The battery surface temperature acquisition unit is attached to the surface of the battery end cover to collect the surface temperature information of the single battery cell. The battery pressure acquisition unit is attached to the surface of the battery end cover to collect the stress information of the single battery cell. The battery voltage acquisition unit measures the battery voltage data, wherein the battery voltage data, surface temperature information, stress information, and characteristic gas content signal at the same time constitute a characteristic vector reflecting the battery state and keep time alignment; The feature vector is input into a lithium battery thermal runaway warning model to evaluate the thermal runaway risk. The lithium battery thermal runaway warning model is a Bayesian deep learning model based on time series potential Dirichlet allocation.

Images