Lithium ion battery energy storage cabinet fire early warning system based on multi-source data fusion

Abstract

The invention provides a lithium ion battery energy storage cabinet fire early warning system based on multi-source data fusion, and belongs to the technical field of fire early warning. According to the system, through multi-source sensor data fusion and deep learning analysis, the state of a battery in the energy storage cabinet is monitored in an all-around mode, the accuracy and reliability of thermal runaway early warning are improved, and operation safety of energy storage facilities is guaranteed; the system comprises a sensor module, a data acquisition module, an image acquisition module, an edge calculation module, a sound-light alarm module and a remote monitoring module. The lithium ion battery energy storage cabinet fire early warning system has the advantages that the deep learning and multi-source data fusion technology is adopted by the lithium ion battery energy storage cabinet fire early warning system; according to the invention, the arrangement scheme of the industrial camera is innovatively designed; according to the invention, a multi-scale-based early warning evaluation framework is innovatively designed; according to the invention, an edge computing architecture is adopted and an Internet of Things cloud platform is combined, so that an integrated solution from data acquisition, preprocessing and deep learning analysis to remote monitoring is realized.

Description

Technical Field

[0001] The present invention relates to the field of fire warning technology, and in particular to a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion. Background Art

[0002] With the development of science and technology and people's demand for large-capacity mobile energy storage equipment, lithium-ion batteries that convert chemical energy into electrical energy are being used in more and more fields. The scale of energy storage systems and the capacity of lithium-ion batteries are growing rapidly. The electrochemical energy storage industry is also a rapidly growing market. The proposal of the "dual carbon" goal also determines the direction of new power energy storage systems based on new energy. Lithium-ion battery energy storage cabinets use renewable energy generation to store electricity and are an indispensable part of the power grid system. However, high-energy-density lithium-ion batteries are centrally installed in a confined space such as an energy storage cabinet. Once thermal runaway occurs, it will lead to serious safety accidents. Therefore, a reliable early warning system is urgently needed to ensure the safe operation of the energy storage cabinet.

[0003] Therefore, it is necessary to compare and summarize the safety technologies related to lithium-ion battery energy storage cabinets, sort out the existing technical difficulties and key points, especially the fire hazards caused by thermal runaway. It is necessary to develop intelligent early warning technology and look forward to early warning of lithium-ion battery energy storage cabinets, thereby improving the safety of lithium-ion battery energy storage cabinets, realizing dynamic monitoring, and issuing early warnings before a fire occurs in the lithium-ion battery energy storage cabinet, effectively improving the level of safety protection, reducing the probability of fire, and reducing property losses after a fire. Summary of the invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion. The system realizes all-round monitoring of the battery status in the energy storage cabinet through multi-source sensor data fusion and deep learning analysis, improves the accuracy and reliability of thermal runaway warning, and ensures the safe operation of energy storage facilities.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion, including a sensor module, a data acquisition module, an image acquisition module, an edge computing module, an audio-visual alarm module and a remote monitoring module; the sensor module is used to detect and collect temperature data, gas concentration data and smoke concentration data in the lithium-ion battery energy storage cabinet as sensor data, and transmit them to the data acquisition module; the data acquisition module is connected to the edge computing module via Ethernet, and is used to transmit the collected sensor data to the edge computing module via the TCP / IP protocol; the image acquisition module is used to The edge computing module is used to collect image data in the lithium-ion battery energy storage cabinet and transmit it to the edge computing module; the edge computing module is used to receive and synchronize the sensor data from the data acquisition module and the image data from the image acquisition module, that is, the original data, and after standardization preprocessing the original data, analyze it to obtain the lithium-ion battery status warning result, and transmit the original data and the lithium-ion battery status warning result to the Internet of Things cloud platform; the sound and light alarm module is used to trigger different levels of alarm signals according to the lithium-ion battery status warning result output by the edge computing module; the remote monitoring module is used to receive the original data and lithium-ion battery status warning result forwarded by the Internet of Things cloud platform.

[0006] Preferably, the sensor module includes a gas sensor, a smoke sensor, a point temperature sensor and an infrared background temperature sensor; each lithium-ion battery is equipped with a point temperature sensor to measure the temperature of a fixed point of the lithium-ion battery; multiple gas sensors are equipped according to the size of the energy storage cabinet; the gas sensors include a CO sensor, a CO2 sensor, an H2 sensor and a carbon-containing organic matter sensor; the smoke sensor is used to measure the smoke concentration in the energy storage cabinet; the infrared background temperature sensor covers the entire lithium-ion battery cell module in the energy storage cabinet and is used to measure the temperature of the lithium-ion battery cell module.

[0007] Preferably, the data acquisition module is based on a microcontroller unit and an embedded operating system, and is compatible with a variety of standard signals and communication protocols; the data acquisition module is connected to the edge computing module through an Ethernet interface, and communicates through the TCP / IP protocol, and transmits the temperature data, gas concentration data and smoke concentration data in the lithium-ion battery energy storage cabinet detected and collected by the sensor module to the edge computing module in real time for further analysis.

[0008] Preferably, the image acquisition module uses an industrial camera and acquires image data at a frequency of 1 frame per second, and all acquired image data are transmitted to the edge computing module for real-time analysis by the edge computing module;

[0009] The image acquisition module uses an industrial camera and the specific method of acquiring image data at a frequency of 1 frame per second is:

[0010] S1: N industrial cameras are installed on the front of the energy storage cabinet at a 45-degree angle to collect edge contour images of lithium-ion battery cells and identify edge contour distortion of lithium-ion battery cells caused by gas release. When the air density changes or smoke appears, the edge contour of the straight lithium-ion battery cells in the monitoring area will bend or deform.

[0011] S2: An industrial camera is installed at each of the four top corners of the energy storage cabinet, arranged at a 45-degree downward angle to cover all lithium-ion battery cell module areas to monitor the gas release phenomenon of lithium-ion battery cells.

[0012] Preferably, the edge computing module is an edge computer, which has a built-in local communication module, a data storage module, a deep learning module and a remote communication module;

[0013] The local communication module receives and synchronizes sensor data from the data acquisition module and image data from the image acquisition module through a variety of standard interfaces, and provides a unified data access interface for the data storage module and the deep learning module;

[0014] The data storage module is used to store and manage the sensor data and image data received and synchronized by the local communication module; the sensor data is saved in a database format, and the image data is directly stored in a file form and an index record is established; the data storage module performs a first-in-first-out cycle management on the stored sensor data and image data to ensure the rational use of the storage space, and at the same time provides a calling interface for the sensor data and image data to support the data transmission of the remote communication module;

[0015] The remote communication module transmits the original data of the data storage module and the early warning results of the deep learning module to the Internet of Things cloud platform.

[0016] Preferably, the deep learning module is a core component of the edge computing module, including a data preprocessing unit, a feature extraction unit and an early warning analysis unit; the sensor data collected by the data acquisition module and the image data collected by the image acquisition module are analyzed in real time to establish a multi-level early warning judgment mechanism;

[0017] The data preprocessing unit is used to implement data standardization preprocessing, including sensor data preprocessing and image data preprocessing;

[0018] The feature extraction unit includes a parallel LSTM temporal feature extraction network, a CNN spatial feature extraction network and an attention fusion layer;

[0019] The LSTM temporal feature extraction network adopts a cascade structure to extract the temporal features of sensor data, including a warning feature extraction layer and a state evolution prediction layer; the CNN spatial feature extraction network is used to extract the spatial features of the image; the attention fusion layer adaptively fuses the temporal features extracted by the LSTM temporal feature extraction network and the spatial features extracted by the CNN network through a dynamic weight adjustment mechanism to obtain fusion features;

[0020] The early warning analysis unit obtains the predicted value of the lithium-ion battery state through full-connection layer mapping based on the fusion features, and converts the predicted value of the lithium-ion battery state into the probability distribution of four states. Finally, based on the state probability and real-time monitoring sensor data, a multi-level early warning judgment mechanism is established.

[0021] Preferably, the sensor data preprocessing adopts a Z-score standardization method to convert sensor data of different dimensions into the same scale space. The specific formula is:

[0022]

[0023] Where X′ is the standardized sensor data, X is the original sensor data, μ is the average value of the temperature sensor, gas sensor and smoke sensor data, and σ is the standard deviation of the temperature sensor, gas sensor and smoke sensor data;

[0024] Based on formula (1), all temperature data and gas concentration data generated by the CO sensor, CO2 sensor, H2 sensor, carbon-containing organic matter sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, that is, sensor data, are standardized to obtain standardized CO sensor data, CO2 sensor data, H2 sensor data, carbon-containing organic matter sensor data, smoke sensor data, point temperature sensor data and infrared background temperature sensor data, eliminating the dimension effect;

[0025] The image data preprocessing includes image size unification, image tensor conversion and image standardization. The specific method is as follows:

[0026] C1: Image size unification processing: The input original RGB image data is uniformly adjusted to a fixed pixel value to ensure the consistency of the input image size;

[0027] C2: Image tensor conversion: Convert the RGB image data into a tensor format and normalize the pixel values ​​of the RGB image data. The specific conversion formula is:

[0028]

[0029] Among them, pixel normalizedis the pixel value after normalization of RGB image data, pixel value is the original pixel value of the RGB image data, pixel max The maximum value of the pixel value range of the RGB image;

[0030] C3: Image standardization: Standardize the three channels of RGB image data separately to make the distribution of RGB image data more suitable for the training of deep learning modules. The specific formula is:

[0031]

[0032] Among them, channel normalized is the normalized channel data of the three channels of the RGB image data, channel is the original channel data of the three channels of the RGB image data, mean is the global mean of the three channels of the RGB image data, and std is the global standard deviation of the three channels of the RGB image data;

[0033] Based on formulas (2) and (3), each image in the RGB image data is preprocessed to obtain a standardized RGB image data sequence, which is then transmitted to the deep learning module for subsequent deep learning analysis;

[0034] The warning feature extraction layer of the LSTM time series feature extraction network extracts the time series features of the sensor data through a gating mechanism, including an input gate i t 、Forget gate t and output gate o t , the specific formula is:

[0035] i t =α(W i ·[h t-1 ,x t ]+b i ) (4)

[0036] f t =α(W f ·[h t-1 ,x t ]+b f ) (5)

[0037] o t =α(W o ·[h t-1 ,x t ]+b o ) (6)

[0038] Among them, α is the sigmoid activation function, h t-1 is the state information calculated and outputted by the LSTM time series feature extraction network at the previous moment, xt is the input data of the sensor module at the current moment, namely the input data of the gas sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, W i , W f and W o are the weight matrices of the input gate, forget gate, and output gate respectively; b i 、b f and b o They are the bias items of the input gate, forget gate, and output gate respectively;

[0039] Based on the time series features extracted by the gating mechanism, the state evolution prediction layer constructs a double-layer neural network structure to achieve dynamic prediction of the state evolution trend of lithium-ion batteries. The specific formula is:

[0040] h t =tanh(W h ·[h t-1 ,x t ]+b h ) (7)

[0041] s t =α(W s ·h t +b s ) (8)

[0042] Among them, h t is the hidden state vector of the state evolution prediction layer at time t, s t is the lithium-ion battery state evolution prediction result output by the state evolution prediction layer at time t, W h and W s are the weight matrices corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer, and b h and b s They are the bias items corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer respectively;

[0043] At the same time, the change rate of the input data at the current moment is calculated. The specific formula is:

[0044]

[0045] Among them, d t is the rate of change of input data at time t, x t-1 is the sensor data value at time t-1, Δt is the sampling time interval;

[0046] The CNN spatial feature extraction network is used to extract the spatial features of the image. The spatial features of the lithium-ion battery state are extracted through multi-layer convolution operations and pooling operations. The specific formula of the convolution operation is:

[0047] f l =ReLU(W l *f l-1 +b l ) (10)

[0048] Among them, f l is the feature map output of the lth layer, W l is the convolution kernel parameter of the lth layer, * is the convolution operation, f l-1 is the feature map of the l-1th layer, b l is the bias term of the lth layer, and ReLU is the activation function;

[0049] A dynamic weight adjustment mechanism is introduced in the attention fusion layer to adaptively adjust the fusion weights according to the reliability of the LSTM temporal feature extraction network and the CNN spatial feature extraction network:

[0050] r t =α(W r ·[v t ,σ t ]+b r ) (11)

[0051] w t =softmax(W w ·r t +b w ) (12)

[0052] f t =w t [0]·f temporal +w t [1] f spatial (13)

[0053] Among them, r t is the reliability score vector calculated by the attention fusion layer at time t, W r is the weight matrix for calculating the reliability score in the attention fusion layer, v t is the effectiveness index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, σ is the volatility index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, b r Bias term for calculating reliability score in attention fusion layer; w t is the feature fusion weight vector calculated by the attention fusion layer at time t, W w is the parameter matrix for weight calculation in the attention fusion layer, b w is the bias term for weight calculation in the attention fusion layer; f t is the fused feature vector output by the attention fusion layer at time t, f temporalis the time series feature vector output by the LSTM time series feature extraction network, f spatial is the spatial feature vector output by the CNN spatial feature extraction network, w t [0] and w t [1] are the fusion weights corresponding to temporal features and spatial features respectively;

[0054] The early warning analysis unit obtains the predicted value of the lithium-ion battery state through the fully connected layer mapping based on the final fused feature representation f output by the feature fusion layer, and uses the softmax function to convert the predicted value of the lithium-ion battery state into the probability distribution of four states. The specific formula is:

[0055] P = softmax(W fc ·f+b fc ) (14)

[0056] Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, P = [P0, P1, P2, P3] is the state probability distribution, P0, P1, P2, P3 are the normal state probability, warning state probability, initial thermal runaway probability and complete thermal runaway probability respectively;

[0057] Further build a multi-scale early warning assessment framework, including short-term early warning assessment, mid-term early warning assessment and long-term early warning assessment:

[0058] The short-term warning evaluation is based on the state probability sequence p of the most recent T time windows. t-T:t , calculate the state transfer trend of the latest T time windows to capture the rapid change trend of the lithium-ion battery state. The specific formula is:

[0059] δp t =mean(diff(p t-T:t )) (15)

[0060] Among them, p t-T:t is the state probability sequence of the time window from tT to time t, including the four state probabilities [P0, P1, P2, P3] of the most recent T time windows, diff is the difference of the calculated state probability sequence, used to obtain the change in state probability, mean is the average value of the calculated state probability sequence, δp t It is the state transition trend indicator at time t;

[0061] The mid-term early warning assessment: based on the exponentially weighted moving average to calculate the state stability, used to smooth short-term fluctuations, reflecting the stability changes of the state, the formula is as follows:

[0062] S t =λ·p t+(1-λ)·S t-1 (16)

[0063] Among them, S t is the state stability index of the lithium-ion battery at time t, that is, the weighted average result of the probabilities of the four states; p t is the probability distribution vector [P0, P1, P2, P3] of the four states of the lithium-ion battery output by the early warning analysis unit at time t, S t-1 is the state stability index of the lithium-ion battery at time t-1, λ is the smoothing factor, and satisfies 0<λ<1, which is used to control the weight ratio of the current state probability distribution and the historical state stability index;

[0064] The four states are normal state, warning state, initial thermal runaway state and complete thermal runaway state;

[0065] The long-term early warning assessment: constructs normal state boundaries based on historical data statistical characteristics for anomaly detection:

[0066] [μ normal -kσ normal ,μ normal +kσ normal ] (17)

[0067] Among them, μ normal is the mean of historical normal state data, σ normal is the standard deviation of historical normal state data, k is the confidence interval coefficient, which is used to control the width of the boundary range;

[0068] Based on the above multi-scale early warning assessment results, a multi-level early warning judgment mechanism is established. The specific judgment strategies are as follows:

[0069] s1: Normal state: When P0 is the maximum probability value and is greater than the set threshold, and the change rate of all sensor data is lower than their respective thresholds, the lithium-ion battery is judged to be in a normal state;

[0070] s2: Warning state: When P1 is greater than the set threshold, or when any sensor data has an abnormal change rate, the warning state is entered, and the lithium-ion battery is judged to have entered the warning state;

[0071] s3: Initial thermal runaway: When P2 is greater than the set threshold, or when the temperature change rate exceeds the set threshold and the change rate of multiple gas concentrations is abnormal, the lithium-ion battery is judged to have entered the initial thermal runaway state;

[0072] s4: Complete thermal runaway: P3 is greater than the set threshold, or the temperature exceeds the thermal runaway characteristic temperature and the concentrations of multiple gases exceed the set thresholds, and the lithium-ion battery is judged to be in a state of complete thermal runaway; indicating that the battery has entered a state of complete thermal runaway.

[0073] Preferably, the remote communication module uses the MQTT protocol to upload the sensor data, image data and early warning results of the deep learning module to the Internet of Things cloud platform in real time;

[0074] The sound and light alarm module includes an LED indicator light and a high-decibel alarm, which is used to trigger alarm signals of different levels according to the lithium-ion battery status warning results output by the deep learning module.

[0075] Preferably, the remote monitoring module is used to receive the original data of the data storage module and the warning information of the deep learning module forwarded by the Internet of Things cloud platform; the remote monitoring module provides a web and mobile monitoring interface to display the operating status, sensor data, image data and warning information in the lithium-ion battery energy storage cabinet in real time; according to different warning levels, the remote monitoring module automatically triggers a multi-level warning mechanism; at the same time, it supports viewing historical data trends and warning records, which is convenient for managers to analyze and evaluate the system operation status.

[0076] The beneficial effects of adopting the above technical solution are: the lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion provided by the present invention has the following beneficial effects:

[0077] (1) The lithium-ion battery energy storage cabinet fire early warning system provided by the present invention adopts deep learning and multi-source data fusion technology; the time series features of CO, CO2, H2, VOC gas sensor, temperature sensor and smoke sensor data are extracted through the LSTM time series feature extraction network, and the contour distortion features caused by gas release are extracted by the CNN spatial feature extraction network. The attention fusion layer introduces a dynamic full summary adjustment mechanism for multi-modal information fusion; combined with the preset multi-level warning judgment mechanism, it can fully capture the abnormal state of the battery in the energy storage cabinet, significantly improve the accuracy and reliability of thermal runaway warning, and effectively reduce the probability of fire accidents;

[0078] (2) The present invention innovatively designs the layout of industrial cameras; an industrial camera is installed on the front of the energy storage cabinet, and four overhead cameras at the four corners of the top of the energy storage cabinet are used to form a full-scale monitoring of the internal space of the energy storage cabinet; this layout can effectively capture the object contour distortion and smoke diffusion phenomenon during the gas release process, providing an intuitive and reliable visual basis for thermal runaway warning;

[0079] (3) The present invention innovatively designs a multi-scale based early warning assessment framework; the framework captures the rapid change trend of the state through short-term early warning assessment, uses mid-term early warning assessment based on exponentially weighted moving average calculation to reflect the change of state stability, and combines long-term early warning assessment to construct the normal state boundary based on the statistical characteristics of historical data, thereby realizing multi-time scale analysis of the lithium-ion battery state; this multi-level and multi-dimensional early warning mechanism can comprehensively evaluate the evolution of the battery state over time, can not only detect transient anomalies in a timely manner, but also identify slowly accumulated potential risks, effectively reducing the false alarm rate and significantly improving the accuracy and reliability of the early warning;

[0080] (4) The present invention adopts edge computing architecture and combines it with the Internet of Things cloud platform to realize an integrated solution from data acquisition, preprocessing and deep learning analysis to remote monitoring; the data acquisition module based on MCU and embedded operating system provides flexible sensor access capabilities; the edge computing module ensures the real-time performance of the system through a layered data processing mechanism; the remote communication mode using the MQTT protocol realizes reliable data transmission and multi-level warning push; the entire system has strong practicality and scalability, and provides complete technical support for the safe monitoring of lithium-ion battery energy storage facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 A structural diagram of a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion provided in an embodiment of the present invention; DETAILED DESCRIPTION

[0082] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0083] In this embodiment, a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion is provided. Figure 1As shown, it includes a sensor module, a data acquisition module, an image acquisition module, an edge computing module, an audio-visual alarm module and a remote monitoring module; the sensor module is used to detect and collect temperature data, gas concentration data and smoke concentration data in the lithium-ion battery energy storage cabinet as sensor data, and transmit them to the data acquisition module; the data acquisition module is connected to the edge computing module via Ethernet, and is used to transmit the collected sensor data to the edge computing module via the TCP / IP protocol; the image acquisition module is used to collect image data in the lithium-ion battery energy storage cabinet and transmit it to the edge computing module via the USB3.0 interface; the edge computing module is used to receive and synchronize sensor data from the data acquisition module and image data from the image acquisition module, that is, raw data, and after standardization preprocessing the raw data, analyze it to obtain the lithium-ion battery status warning result, and transmit the raw data and the lithium-ion battery status warning result to the Internet of Things cloud platform; the audio-visual alarm module is used to trigger different levels of alarm signals according to the lithium-ion battery status warning result output by the edge computing module; the remote monitoring module is used to receive the raw data and lithium-ion battery status warning result forwarded by the Internet of Things cloud platform.

[0084] In this embodiment, the sensor module includes a gas sensor, a smoke sensor, a point temperature sensor and an infrared background temperature sensor; each lithium-ion battery is equipped with a point temperature sensor to measure the temperature of a fixed point of the lithium-ion battery; multiple gas sensors are equipped according to the size of the energy storage cabinet; the gas sensors include a CO sensor, a CO2 sensor, an H2 sensor and a carbon-containing organic matter (VOC) sensor; the smoke sensor is used to measure the smoke concentration in the energy storage cabinet; the infrared background temperature sensor covers the entire lithium-ion battery cell module in the energy storage cabinet and is used to measure the temperature of the lithium-ion battery cell module.

[0085] The data acquisition module is based on a microcontroller unit (MCU) and an embedded operating system, and is compatible with a variety of standard signals and communication protocols (RS485 / CAN); the data acquisition module is connected to the edge computing module through an Ethernet interface, and communicates through the TCP / IP protocol, and transmits the temperature data, gas concentration data, and smoke concentration data in the lithium-ion battery energy storage cabinet detected and collected by the sensor module to the edge computing module in real time for further analysis.

[0086] The image acquisition module uses an industrial camera and collects image data at a frequency of 1 frame per second. All collected image data are transmitted to the edge computing module through the USB3.0 interface for real-time analysis by the edge computing module;

[0087] The image acquisition module uses an industrial camera and the specific method of acquiring image data at a frequency of 1 frame per second is:

[0088] S1: N industrial cameras are installed on the front of the energy storage cabinet at a 45-degree angle to collect edge contour images of lithium-ion battery cells and identify edge contour distortion of lithium-ion battery cells caused by gas release. When the air density changes or smoke appears, the edge contour of the straight lithium-ion battery cells in the monitoring area will bend or deform.

[0089] S2: An industrial camera is installed at each of the four top corners of the energy storage cabinet, arranged at a 45-degree downward angle to cover all lithium-ion battery cell module areas to monitor the gas release phenomenon of lithium-ion battery cells.

[0090] The edge computing module is an edge computer, which has a built-in local communication module, a data storage module, a deep learning module and a remote communication module;

[0091] The local communication module realizes data collection and interaction within the fire early warning system through a variety of standard interfaces (Ethernet, USB3.0, etc.), is responsible for receiving and synchronizing sensor data from the data acquisition module and image data from the image acquisition module, and provides a unified data access interface for the data storage module and the deep learning module;

[0092] The data storage module is used to store and manage the sensor data and image data received and synchronized by the local communication module; the sensor data is saved in a database format, and the image data is directly stored in a file form and an index record is established; the data storage module performs a first-in-first-out cycle management on the stored sensor data and image data to ensure the rational use of the storage space, and at the same time provides a calling interface for the sensor data and image data to support the data transmission of the remote communication module;

[0093] The remote communication module transmits the original data of the data storage module and the early warning results of the deep learning module to the Internet of Things cloud platform.

[0094] The deep learning module is the core component of the edge computing module, including a data preprocessing unit, a feature extraction unit and an early warning analysis unit; the sensor data collected by the data acquisition module and the image data collected by the image acquisition module are analyzed in real time to establish a multi-level early warning judgment mechanism;

[0095] The data preprocessing unit is used to implement data standardization preprocessing, including sensor data preprocessing and image data preprocessing;

[0096] The feature extraction unit includes a parallel LSTM temporal feature extraction network, a CNN spatial feature extraction network and an attention fusion layer;

[0097] The LSTM temporal feature extraction network adopts a cascade structure to extract the temporal features of sensor data, including a warning feature extraction layer and a state evolution prediction layer; the CNN spatial feature extraction network is used to extract the spatial features of the image; the attention fusion layer adaptively fuses the temporal features extracted by the LSTM temporal feature extraction network and the spatial features extracted by the CNN network through a dynamic weight adjustment mechanism to obtain fusion features;

[0098] The early warning analysis unit obtains the predicted value of the lithium-ion battery state through full-connection layer mapping based on the fusion features, and converts the predicted value of the lithium-ion battery state into the probability distribution of four states. Finally, based on the state probability and real-time monitoring sensor data, a multi-level early warning judgment mechanism is established.

[0099] The sensor data preprocessing adopts the Z-score standardization method to convert sensor data of different dimensions into the same scale space. The specific formula is:

[0100] X′= X-μ (1)

[0101] σ

[0102] Where X′ is the standardized sensor data, X is the original sensor data, μ is the average value of the temperature sensor, gas sensor and smoke sensor data, and σ is the standard deviation of the temperature sensor, gas sensor and smoke sensor data;

[0103] Based on formula (1), all temperature data and gas concentration data generated by the CO sensor, CO2 sensor, H2 sensor, VOC sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, that is, sensor data, are standardized to obtain standardized CO sensor data, CO2 sensor data, H2 sensor data, VOC sensor data, smoke sensor data, point temperature sensor data and infrared background temperature sensor data, eliminating the dimension effect;

[0104] The image data preprocessing includes image size unification, image tensor conversion and image standardization. The specific method is as follows:

[0105] C1: Image size unification processing: The input original RGB image data is uniformly adjusted to a fixed pixel value to ensure the consistency of the input image size;

[0106] C2: Image tensor conversion: Convert the RGB image data into a tensor format and normalize the pixel values ​​of the RGB image data. The specific conversion formula is:

[0107]

[0108] Among them, pixel normalized is the pixel value after normalization of RGB image data, pixel value is the original pixel value of the RGB image data, pixel max The maximum value of the pixel value range of the RGB image;

[0109] In this embodiment, pixel max =255;

[0110] C3: Image standardization: Standardize the three channels of RGB image data separately to make the distribution of RGB image data more suitable for the training of deep learning modules. The specific formula is:

[0111]

[0112] Among them, channel normalized is the normalized channel data of the three channels of the RGB image data, channel is the original channel data of the three channels of the RGB image data, mean is the global mean of the three channels of the RGB image data, and std is the global standard deviation of the three channels of the RGB image data;

[0113] Based on formulas (2) and (3), each image in the RGB image data is preprocessed to obtain a standardized RGB image data sequence, which is then transmitted to the deep learning module for subsequent deep learning analysis;

[0114] The warning feature extraction layer of the LSTM time series feature extraction network extracts the time series features of the sensor data through a gating mechanism, including an input gate i t 、Forget gate t and output gate o t , the specific formula is:

[0115] i t =α(W i ·[h t-1 ,x t ]+b i ) (4)

[0116] f t =α(W f ·[h t-1 ,x t ]+b f ) (5)

[0117] o t =α(W o ·[h t-1 ,x t ]+b o) (6)

[0118] Among them, α is the sigmoid activation function, h t-1 is the state information calculated and outputted by the LSTM time series feature extraction network at the previous moment, x t is the input data of the sensor module at the current moment, namely the input data of the gas sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, W i , W f and W o are the weight matrices of the input gate, forget gate, and output gate respectively; b i , b f and b o They are the bias items of the input gate, forget gate, and output gate respectively;

[0119] Based on the time series features extracted by the gating mechanism, the state evolution prediction layer constructs a double-layer neural network structure to achieve dynamic prediction of the state evolution trend of lithium-ion batteries. The specific formula is:

[0120] h t =tanh(W h ·[h t-1 ,x t ]+b h ) (7)

[0121] s t =α(W s ·h t +b s ) (8)

[0122] Among them, h t is the hidden state vector of the state evolution prediction layer at time t, s t is the lithium-ion battery state evolution prediction result output by the state evolution prediction layer at time t, W h and W s are the weight matrices corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer, and b h and b s They are the bias items corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer. This cascade structure not only captures the current state characteristics of the lithium-ion battery, but also predicts the state evolution trend, providing a more comprehensive basis for early warning analysis.

[0123] At the same time, the change rate of the input data at the current moment is calculated. The specific formula is:

[0124]

[0125] Among them, d t is the rate of change of input data at time t, xt-1 is the sensor data value at time t-1 (i.e., the previous time), and Δt is the sampling time interval;

[0126] The CNN spatial feature extraction network is used to extract the spatial features of the image. The spatial features of the lithium-ion battery state are extracted through multi-layer convolution operations and pooling operations. The specific formula of the convolution operation is:

[0127] f l =ReLU(W l *f l-1 +b l ) (10)

[0128] Among them, f l is the feature map output of the lth layer, W l is the convolution kernel parameter of the lth layer, * is the convolution operation, f l-1 is the feature map of the l-1th layer (the output of the previous layer), b l is the bias term of the lth layer, and ReLU is the activation function. Through this network, the key spatial features of the battery in the thermal runaway process can be effectively identified;

[0129] A dynamic weight adjustment mechanism is introduced in the attention fusion layer to adaptively adjust the fusion weights according to the reliability of the LSTM temporal feature extraction network and the CNN spatial feature extraction network:

[0130] r t =α(W r ·[v t ,σ t ]+b r ) (11)

[0131] w t =softmax(W w ·r t +b w ) (12)

[0132] f t =w t [0]·f temporal +w t [1] f spatial (13)

[0133] Among them, r t is the reliability score vector calculated by the attention fusion layer at time t, W r is the weight matrix for calculating the reliability score in the attention fusion layer, v t is the effectiveness index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, σ is the volatility index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, br Bias term for calculating reliability score in attention fusion layer; w t is the feature fusion weight vector calculated by the attention fusion layer at time t, W w is the parameter matrix for weight calculation in the attention fusion layer, b w is the bias term for weight calculation in the attention fusion layer; f t is the fused feature vector output by the attention fusion layer at time t, f temporal is the time series feature vector output by the LSTM time series feature extraction network, f spatial is the spatial feature vector output by the CNN spatial feature extraction network, w t [0] and w t [1] are the fusion weights corresponding to temporal features and spatial features respectively;

[0134] The early warning analysis unit obtains the predicted value of the lithium-ion battery state through the fully connected layer mapping based on the final fused feature representation f output by the feature fusion layer, and uses the softmax function to convert the predicted value of the lithium-ion battery state into the probability distribution of four states. The specific formula is:

[0135] P = softmax(W fc ·f+b fc ) (14)

[0136] Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, P = [P0, P1, P2, P3] is the state probability distribution, P0, P1, P2, P3 are the normal state probability, warning state probability, initial thermal runaway probability and complete thermal runaway probability respectively;

[0137] Further build a multi-scale early warning assessment framework, including short-term early warning assessment, mid-term early warning assessment and long-term early warning assessment:

[0138] The short-term warning evaluation is based on the state probability sequence p of the most recent T time windows. t-T:t , calculate the state transfer trend of the latest T time windows to capture the rapid change trend of the lithium-ion battery state. The specific formula is:

[0139] δp t =mean(diff(p t-T:t )) (15)

[0140] Among them, p t-T:tis the state probability sequence of the time window from tT to time t, including the four state probabilities [P0, P1, P2, P3] of the most recent T time windows, diff is the difference of the calculated state probability sequence, used to obtain the change in state probability, mean is the average value of the calculated state probability sequence, δp t It is the state transition trend indicator at time t;

[0141] The mid-term early warning assessment: based on the exponentially weighted moving average to calculate the state stability, used to smooth short-term fluctuations, reflecting the stability changes of the state, the formula is as follows:

[0142] S t =λ·p t +(1-λ)·S t-1 (16)

[0143] Among them, S t is the state stability index of the lithium-ion battery at time t, that is, the weighted average result of the probabilities of the four states; p t is the probability distribution vector [P0, P1, P2, P3] of the four states of the lithium-ion battery output by the early warning analysis unit at time t, S t-1 is the state stability index of the lithium-ion battery at time t-1, λ is the smoothing factor, and satisfies 0<λ<1, which is used to control the weight ratio of the current state probability distribution and the historical state stability index;

[0144] The four states are normal state, warning state, initial thermal runaway state and complete thermal runaway state;

[0145] The long-term early warning assessment: constructs normal state boundaries based on historical data statistical characteristics for anomaly detection:

[0146] [μ normal -kσ normal ,μ normal +kσ normal ] (17)

[0147] Among them, μ normal is the mean of historical normal state data, σ normal is the standard deviation of historical normal state data, k is the confidence interval coefficient, which is used to control the width of the boundary range;

[0148] Based on the above multi-scale early warning assessment results, a multi-level early warning judgment mechanism is established. The specific judgment strategies are as follows:

[0149] s1: Normal state: When P0 is the maximum probability value and is greater than the set threshold, and the change rate of all sensor data is lower than their respective thresholds, the lithium-ion battery is judged to be in a normal state;

[0150] s2: Warning state: When P1 is greater than the set threshold, or when any sensor data has an abnormal change rate, the warning state is entered, and the lithium-ion battery is judged to have entered the warning state;

[0151] s3: Initial thermal runaway: When P2 is greater than the set threshold, or when the temperature change rate exceeds the set threshold and the change rate of multiple gas concentrations is abnormal, the lithium-ion battery is judged to have entered the initial thermal runaway state;

[0152] s4: Complete thermal runaway: P3 is greater than the set threshold, or the temperature exceeds the thermal runaway characteristic temperature and the concentrations of multiple gases exceed the set thresholds, and the lithium-ion battery is judged to be in a state of complete thermal runaway; indicating that the battery has entered a state of complete thermal runaway.

[0153] The deep learning module outputs the probability distribution P=[P0, P1, P2, P3] of the lithium-ion battery state based on the analysis results of the sensor data and the image data, which respectively correspond to the probabilities of the normal state, the warning state, the initial thermal runaway state and the complete thermal runaway state; the system performs multi-level warning judgment based on the predicted probability and the sensor data. When it is determined that a warning is needed, the warning information is pushed to the monitoring center through the remote communication module described below, and the on-site sound and light alarm device is triggered to achieve timely and effective safety warning;

[0154] The remote communication module uses the MQTT protocol to upload sensor data, image data and early warning results of the deep learning module to the IoT cloud platform in real time, supports automatic retransmission and local backup mechanism of sensor data and image data, and ensures the integrity of sensor data and image data transmission;

[0155] The sound and light alarm module includes an LED indicator light and a high-decibel alarm, which is used to trigger alarm signals of different levels according to the lithium-ion battery status warning results output by the deep learning module.

[0156] The remote monitoring module is used to receive the original data of the data storage module and the warning information of the deep learning module forwarded by the Internet of Things cloud platform; the remote monitoring module provides a web and mobile monitoring interface to display the operating status, sensor data, image data and warning information in the lithium-ion battery energy storage cabinet in real time; according to different warning levels, the remote monitoring module automatically triggers a multi-level warning mechanism; at the same time, it supports viewing historical data trends and warning records, which is convenient for managers to analyze and evaluate the system operation status.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion, characterized by: It includes a sensor module, a data acquisition module, an image acquisition module, an edge computing module, an audio-visual alarm module and a remote monitoring module; the sensor module is used to detect and acquire temperature data, gas concentration data and smoke concentration data in the lithium-ion battery energy storage cabinet as sensor data, and transmit them to the data acquisition module; the data acquisition module is connected to the edge computing module via Ethernet, and is used to transmit the acquired sensor data to the edge computing module via the TCP / IP protocol; the image acquisition module is used to acquire image data in the lithium-ion battery energy storage cabinet and transmit it to the edge computing module; the edge computing module is used to receive and synchronize sensor data from the data acquisition module and image data from the image acquisition module, i.e., raw data, and after standardization preprocessing the raw data, analyze it to obtain the lithium-ion battery status warning result, and transmit the raw data and the lithium-ion battery status warning result to the Internet of Things cloud platform; The sound and light alarm module is used to trigger alarm signals of different levels according to the lithium-ion battery status warning results output by the edge computing module; the remote monitoring module is used to receive the original data and lithium-ion battery status warning results forwarded by the Internet of Things cloud platform.

2. According to claim 1, a lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion is characterized by: The sensor module includes a gas sensor, a smoke sensor, a point temperature sensor and an infrared background temperature sensor; each lithium-ion battery is equipped with a point temperature sensor to measure the temperature of a fixed point of the lithium-ion battery; multiple gas sensors are equipped according to the size of the energy storage cabinet; the gas sensors include a CO sensor, a CO2 sensor, an H2 sensor and a carbon-containing organic matter sensor; the smoke sensor is used to measure the smoke concentration in the energy storage cabinet; the infrared background temperature sensor covers the entire lithium-ion battery cell module in the energy storage cabinet and is used to measure the temperature of the lithium-ion battery cell module.

3. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 2 is characterized in that: The data acquisition module is based on a microcontroller unit and an embedded operating system, and is compatible with a variety of standard signals and communication protocols; the data acquisition module is connected to the edge computing module through an Ethernet interface, and communicates through the TCP / IP protocol, and transmits the temperature data, gas concentration data, and smoke concentration data in the lithium-ion battery energy storage cabinet detected and collected by the sensor module to the edge computing module in real time for further analysis.

4. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 3 is characterized by: The image acquisition module uses an industrial camera and collects image data at a frequency of 1 frame per second. All collected image data are transmitted to the edge computing module for real-time analysis by the edge computing module; The image acquisition module uses an industrial camera and the specific method of acquiring image data at a frequency of 1 frame per second is: S1: N industrial cameras are installed on the front of the energy storage cabinet at a 45-degree angle to collect edge contour images of lithium-ion battery cells and identify edge contour distortion of lithium-ion battery cells caused by gas release. When the air density changes or smoke appears, the edge contour of the straight lithium-ion battery cells in the monitoring area will bend or deform. S2: An industrial camera is installed at each of the four top corners of the energy storage cabinet, arranged at a 45-degree downward angle to cover all lithium-ion battery cell module areas to monitor the gas release phenomenon of lithium-ion battery cells.

5. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 4 is characterized in that: The edge computing module is an edge computer, which has a built-in local communication module, a data storage module, a deep learning module and a remote communication module; The local communication module receives and synchronizes sensor data from the data acquisition module and image data from the image acquisition module through a variety of standard interfaces, and provides a unified data access interface for the data storage module and the deep learning module; The data storage module is used to store and manage the sensor data and image data received and synchronized by the local communication module; the sensor data is saved in a database format, and the image data is directly stored in a file form and index records are established; The data storage module performs first-in-first-out cyclic management on the stored sensor data and image data to ensure the rational use of storage space. It also provides a calling interface for sensor data and image data to support data transmission of the remote communication module. The remote communication module transmits the original data of the data storage module and the early warning results of the deep learning module to the Internet of Things cloud platform.

6. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 5 is characterized by: The deep learning module is the core component of the edge computing module, including a data preprocessing unit, a feature extraction unit and an early warning analysis unit; the sensor data collected by the data acquisition module and the image data collected by the image acquisition module are analyzed in real time to establish a multi-level early warning judgment mechanism; The data preprocessing unit is used to implement data standardization preprocessing, including sensor data preprocessing and image data preprocessing; The feature extraction unit includes a parallel LSTM temporal feature extraction network, a CNN spatial feature extraction network and an attention fusion layer; The LSTM time series feature extraction network adopts a cascade structure to extract the time series features of sensor data, including a warning feature extraction layer and a state evolution prediction layer; the CNN spatial feature extraction network is used to extract the spatial features of the image; The attention fusion layer adaptively fuses the temporal features extracted by the LSTM temporal feature extraction network and the spatial features extracted by the CNN network through a dynamic weight adjustment mechanism to obtain fused features. The early warning analysis unit obtains the predicted value of the lithium-ion battery state through full-connection layer mapping based on the fusion features, and converts the predicted value of the lithium-ion battery state into the probability distribution of four states. Finally, based on the state probability and real-time monitoring sensor data, a multi-level early warning judgment mechanism is established.

7. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 6 is characterized by: The sensor data preprocessing adopts the Z-score standardization method to convert sensor data of different dimensions into the same scale space. The specific formula is: Where X′ is the standardized sensor data, X is the original sensor data, μ is the average value of the temperature sensor, gas sensor and smoke sensor data, and σ is the standard deviation of the temperature sensor, gas sensor and smoke sensor data; Based on formula (1), all temperature data, gas concentration data and smoke concentration data generated by the CO sensor, CO2 sensor, H2 sensor, carbon-containing organic matter sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, that is, sensor data, are standardized to obtain standardized CO sensor data, CO2 sensor data, H2 sensor data, carbon-containing organic matter sensor data, smoke sensor data, point temperature sensor data and infrared background temperature sensor data, eliminating the dimension effect; The image data preprocessing includes image size unification, image tensor conversion and image standardization. The specific method is as follows: C1: Image size unification processing: The input original RGB image data is uniformly adjusted to a fixed pixel value to ensure the consistency of the input image size; C2: Image tensor conversion: Convert the RGB image data into a tensor format and normalize the pixel values ​​of the RGB image data. The specific conversion formula is: Among them, pixel normalized is the pixel value after normalization of RGB image data, pixel value is the original pixel value of the RGB image data, pixel max The maximum value of the pixel value range of the RGB image; C3: Image standardization: Standardize the three channels of RGB image data separately to make the distribution of RGB image data more suitable for the training of deep learning modules. The specific formula is: Among them, channel normalized is the normalized channel data of the three channels of the RGB image data, channel is the original channel data of the three channels of the RGB image data, mean is the global mean of the three channels of the RGB image data, and std is the global standard deviation of the three channels of the RGB image data; Based on formulas (2) and (3), each image in the RGB image data is preprocessed to obtain a standardized RGB image data sequence, which is then transmitted to the deep learning module for subsequent deep learning analysis.

8. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 7 is characterized by: The warning feature extraction layer of the LSTM time series feature extraction network extracts the time series features of the sensor data through a gating mechanism, including an input gate i t 、Forget gate t and output gate o t , the specific formula is: i t =α(W i ·[h t-1 ,x t ]+b i ) (4) f t =α(W f ·[h t-1 ,x t ]+b f ) (5) o t =α(W o ·[h t-1 ,x t ]+b o ) (6) Among them, α is the sigmoid activation function, h t-1 is the state information calculated and outputted by the LSTM time series feature extraction network at the previous moment, x t is the input data of the sensor module at the current moment, namely the input data of the gas sensor, smoke sensor, point temperature sensor and infrared background temperature sensor, W i , W f and W o are the weight matrices of the input gate, forget gate, and output gate respectively; b i 、b f and b o They are the bias items of the input gate, forget gate, and output gate respectively; Based on the time series features extracted by the gating mechanism, the state evolution prediction layer constructs a double-layer neural network structure to achieve dynamic prediction of the state evolution trend of lithium-ion batteries. The specific formula is: h t =tanh(W h ·[h t-1 ,x t ]+b h ) (7) s t =α(W s ·h t +b s ) (8) Among them, h t is the hidden state vector of the state evolution prediction layer at time t, s t is the lithium-ion battery state evolution prediction result output by the state evolution prediction layer at time t, W h and W s are the weight matrices corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer, and b h and b s They are the bias items corresponding to the hidden state calculation and state prediction calculation in the state evolution prediction layer respectively; At the same time, the change rate of the input data at the current moment is calculated. The specific formula is: Among them, d t is the rate of change of input data at time t, x t-1 is the sensor data value at time t-1, Δt is the sampling time interval; The CNN spatial feature extraction network is used to extract the spatial features of the image. The spatial features of the lithium-ion battery state are extracted through multi-layer convolution operations and pooling operations. The specific formula of the convolution operation is: f l =ReLU(W l *f l-1 +b l ) (10) Among them, f l is the feature map output of the lth layer, W l is the convolution kernel parameter of the lth layer, * is the convolution operation, f l-1 is the feature map of the l-1th layer, b l is the bias term of the lth layer, and ReLU is the activation function; A dynamic weight adjustment mechanism is introduced in the attention fusion layer to adaptively adjust the fusion weights according to the reliability of the LSTM temporal feature extraction network and the CNN spatial feature extraction network: r t =α(W r ·[v t ,s t ]+b r ) (11) w t =softmax(W w ·r t +b w ) (12) f t =w t [0]·f temporal +w t [1]·f spatial (13) Among them, r t is the reliability score vector calculated by the attention fusion layer at time t, W r is the weight matrix for calculating the reliability score in the attention fusion layer, v t is the effectiveness index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, σ is the volatility index vector of the LSTM temporal feature extraction network and the CNN spatial feature extraction network at time t, b r Bias term for calculating reliability score in attention fusion layer; w t is the feature fusion weight vector calculated by the attention fusion layer at time t, W w is the parameter matrix for weight calculation in the attention fusion layer, b w is the bias term for weight calculation in the attention fusion layer; f t is the fused feature vector output by the attention fusion layer at time t, f temporal is the time series feature vector output by the LSTM time series feature extraction network, f spatial is the spatial feature vector output by the CNN spatial feature extraction network, w t [0] and w t [1] are the fusion weights corresponding to temporal features and spatial features respectively; The early warning analysis unit obtains the predicted value of the lithium-ion battery state through the fully connected layer mapping based on the final fused feature representation f output by the feature fusion layer, and uses the softmax function to convert the predicted value of the lithium-ion battery state into the probability distribution of four states. The specific formula is: P=softmax(W fc ·f+b fc ) (14) Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, P = [P0, P1, P2, P3] is the state probability distribution, P0, P1, P2, P3 are the normal state probability, warning state probability, initial thermal runaway probability and complete thermal runaway probability respectively; Further build a multi-scale early warning assessment framework, including short-term early warning assessment, mid-term early warning assessment and long-term early warning assessment: The short-term warning evaluation is based on the state probability sequence p of the most recent T time windows. t-T:t , calculate the state transfer trend of the latest T time windows to capture the rapid change trend of the lithium-ion battery state. The specific formula is: δp t =mean(diff(p t-T:t )) (15) Among them, p t-T:t is the state probability sequence of the time window from tT to time t, including the four state probabilities [P0, P1, P2, P3] of the most recent T time windows, diff is the difference of the calculated state probability sequence, used to obtain the change in state probability, mean is the average value of the calculated state probability sequence, δp t It is the state transition trend indicator at time t; The mid-term early warning assessment: based on the exponentially weighted moving average to calculate the state stability, used to smooth short-term fluctuations, reflecting the stability changes of the state, the formula is as follows: S t =λ·p t +(1-λ)·S t-1 (16) Among them, S t is the state stability index of the lithium-ion battery at time t, that is, the weighted average result of the probabilities of the four states; p t is the probability distribution vector [P0, P1, P2, P3] of the four states of the lithium-ion battery output by the early warning analysis unit at time t, S t-1 is the state stability index of the lithium-ion battery at time t-1, λ is the smoothing factor, and satisfies 0<λ<1, which is used to control the weight ratio of the current state probability distribution and the historical state stability index; The four states are normal state, warning state, initial thermal runaway state and complete thermal runaway state; The long-term early warning assessment: constructs normal state boundaries based on historical data statistical characteristics for anomaly detection: [m normal -kσ normal ,m normal +kσ normal ](17) Among them, μ normal is the mean of historical normal state data, σ normal is the standard deviation of historical normal state data, k is the confidence interval coefficient, which is used to control the width of the boundary range; Based on the above multi-scale early warning assessment results, a multi-level early warning judgment mechanism is established. The specific judgment strategies are as follows: s1: Normal state: When P0 is the maximum probability value and is greater than the set threshold, and the change rate of all sensor data is lower than their respective thresholds, the lithium-ion battery is judged to be in a normal state; s2: Warning state: When P1 is greater than the set threshold, or when any sensor data has an abnormal change rate, the warning state is entered, and the lithium-ion battery is judged to have entered the warning state; s3: Initial thermal runaway: When P2 is greater than the set threshold, or when the temperature change rate exceeds the set threshold and the change rate of multiple gas concentrations is abnormal, the lithium-ion battery is judged to have entered the initial thermal runaway state; s4: Complete thermal runaway: P3 is greater than the set threshold, or the temperature exceeds the thermal runaway characteristic temperature and the concentrations of multiple gases exceed the set thresholds, and the lithium-ion battery is judged to be in a state of complete thermal runaway; indicating that the battery has entered a state of complete thermal runaway.

9. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 8 is characterized by: The remote communication module uses the MQTT protocol to upload sensor data, image data and early warning results of the deep learning module to the Internet of Things cloud platform in real time; The sound and light alarm module includes an LED indicator light and a high-decibel alarm, which is used to trigger alarm signals of different levels according to the lithium-ion battery status warning results output by the deep learning module.

10. The lithium-ion battery energy storage cabinet fire early warning system based on multi-source data fusion according to claim 9, characterized in that: The remote monitoring module is used to receive the original data of the data storage module and the warning information of the deep learning module forwarded by the Internet of Things cloud platform; the remote monitoring module provides a web and mobile monitoring interface to display the operating status, sensor data, image data and warning information in the lithium-ion battery energy storage cabinet in real time; according to different warning levels, the remote monitoring module automatically triggers a multi-level warning mechanism; at the same time, it supports viewing historical data trends and warning records, which is convenient for managers to analyze and evaluate the system operation status.

Images