Electric vehicle single lithium battery home and elevator intelligent early warning management method and system based on multi-modal recognition

Abstract

The application discloses a kind of based on multi-modal identification's electric battery car single lithium battery enters house enters elevator intelligent early warning management and control method and system, method includes: synchronous acquisition image data, electromagnetic field intensity data and object density data, generate multi-source perception data flow;Multi-source perception data flow is carried out space closed attribute identification and position information joint discrimination, generate scene type label;According to scene type label, start multi-modal AI fusion identification in closed area, generate single lithium battery target recognition result containing target category and confidence;Single lithium battery target recognition result is associated and rechecked between the multiple point perception terminal of unit entrance, elevator car and house corridor, generate full-link risk determination result;According to full-link risk determination result, execute hierarchical disposal strategy, generate closed-loop control record.Utilize the embodiment of the application, can improve the accuracy of electric battery car enters house enters elevator risk identification and the intelligent level of management and control.

Description

Technical Field

[0001] This invention belongs to the field of security technology, specifically a method and system for intelligent early warning and control of single lithium batteries of electric vehicles entering homes and elevators based on multimodal recognition. Background Technology

[0002] Lithium-ion batteries for electric bicycles are widely used due to their high energy density and long cycle life, but they pose a significant risk of thermal runaway. Currently, the control measures for electric bicycles entering homes and elevators mainly rely on ordinary video surveillance or vehicle-blocking sensors inside elevator cars. The former relies solely on visible light images for shape identification, making it easily evaded by raincoats, angle changes, or carrying obstructions. The latter often uses weighing or infrared beam detection, resulting in a high false alarm rate and an inability to distinguish electric bicycles from other metal objects, leading to numerous invalid alarms. More critically, existing solutions lack the ability to accurately identify individual lithium batteries, making it difficult to differentiate batteries from the entire vehicle or similar items. Furthermore, they lack a cross-building verification mechanism (unit entrance, elevator car, entrance corridor), meaning a single misjudgment can trigger mandatory intervention, disrupting residents' normal travel. In addition, existing systems are lacking in multi-dimensional sensing fusion of electromagnetic field characteristics and physical density characteristics in enclosed spaces. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent early warning and control of electric vehicle single-cell lithium batteries entering homes and elevators based on multimodal recognition, so as to overcome the shortcomings of the existing technology and improve the accuracy of risk identification and intelligent control of electric vehicle entry into homes and elevators.

[0004] One embodiment of this application provides a method for intelligent early warning and control of single-cell lithium batteries for electric vehicles entering homes and elevators based on multimodal recognition. The method includes: By simultaneously collecting image data, electromagnetic field strength data, and object density data through integrated sensing terminals deployed at building security points, a multi-source sensing data stream is generated. The multi-source sensing data stream is subjected to spatial closure attribute recognition and location information joint discrimination to distinguish between outdoor open areas and elevator and entrance closed areas, and scene type labels are generated. Based on the scene type label, multimodal AI fusion recognition is initiated in the enclosed area. Based on the pre-trained electric vehicle single lithium battery exclusive recognition model, the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features are fused and compared in multiple dimensions. The interference of similar items is eliminated through the differential feature library to generate single lithium battery target recognition results containing target category and confidence level. The target identification results of the single lithium battery are correlated and verified among multiple sensing terminals at the unit entrance, elevator car and entrance corridor to generate a full-link risk assessment result. Based on the full-link risk assessment results, a tiered handling strategy is implemented. For confirmed risk-carrying events, voice prompts, sound and light warnings, or elevator control locking actions are triggered, and the captured data and warning information are pushed to the back-end management terminal to generate a closed-loop management record.

[0005] Another embodiment of this application provides an intelligent early warning and control system for electric vehicle single-cell lithium batteries entering homes and elevators based on multimodal recognition, the system comprising: The data acquisition module is used to simultaneously acquire image data, electromagnetic field strength data, and object density data through integrated sensing terminals deployed at building security points, and generate multi-source sensing data streams. The differentiation module is used to perform spatial closure attribute recognition and location information joint discrimination on the multi-source sensing data stream, distinguish outdoor open areas from elevator and enclosed entrance areas, and generate scene type labels. The identification module is used to initiate multimodal AI fusion identification in a closed area based on the scene type label. Based on the pre-trained electric vehicle single lithium battery exclusive identification model, it performs multi-dimensional fusion comparison of the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features, and removes interference from similar items through a differentiated feature library to generate single lithium battery target identification results containing target category and confidence level. The verification module is used to correlate and verify the target identification results of the single lithium battery in multiple sensing terminals at the unit entrance, elevator car and entrance corridor, and generate a full-link risk assessment result. The early warning module is used to execute a graded handling strategy based on the full-link risk assessment results. It triggers voice prompts, sound and light warnings, or elevator control locking actions for confirmed risk-carrying events, and pushes the captured data and early warning information to the back-end management terminal to generate a closed-loop management record.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides an intelligent early warning and control method for electric vehicle single lithium battery entering homes and elevators based on multimodal recognition, which can improve the accuracy of risk identification and the level of intelligent control of electric vehicle entering homes and elevators. Attached Figure Description

[0009] Figure 1Hardware structure block diagram of a computer terminal for a smart early warning and control method for electric vehicle single lithium battery entering homes and elevators based on multimodal recognition, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating an intelligent early warning and control method for electric vehicle single-cell lithium batteries entering homes and elevators based on multimodal recognition, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent early warning and control system for electric vehicle single-cell lithium batteries entering homes and elevators, based on multimodal recognition, provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] This invention first provides a method for intelligent early warning and control of single-cell lithium batteries of electric vehicles entering homes and elevators based on multimodal recognition. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a multimodal recognition-based intelligent early warning and control method for electric vehicle single-cell lithium batteries entering homes and elevators, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The present invention provides an intelligent early warning and control method for single-cell lithium batteries of electric vehicles entering homes and elevators based on multimodal recognition, which may include the following steps: S201, through the integrated sensing terminal deployed at building security points, synchronously collects image data, electromagnetic field strength data and object density data, and generates multi-source sensing data stream; Specifically, integrated sensing terminals can be installed at the entrance of building units, the top of elevator cars, and the side walls of entrance corridors. The terminals integrate high-definition cameras, electromagnetic field sensors, and density sensors, and complete the clock synchronization calibration of each sensor to generate a cluster of sensing terminals that are deployed and synchronized. The core of this step is to complete the point-to-point deployment of multimodal sensing hardware and unify the clock reference across sensors, build a sensing network covering the entire link of public access in buildings, and ensure the homogeneity and temporal consistency of image, electromagnetic, and density data from the hardware level, so as to provide a stable acquisition foundation for subsequent multi-source data fusion and target recognition.

[0014] The integrated sensing terminal is encapsulated in a single waterproof and dustproof shell, adaptable to the installation environment of public areas in buildings. It integrates three types of core sensors, sharing the same main control processing unit, each performing its specific function while working collaboratively. The deployment locations of the three types of monitoring points have been optimized for specific scenarios, achieving full-link coverage without blind spots. The unit entrance is the first checkpoint for electric vehicles entering the building. The terminal is installed 2.2 meters above the unit door access control, with the lens tilted vertically downwards at a 15-degree angle, completely covering a passage area 3.5 meters wide and 6 meters deep, capturing all personnel and items entering the building. The elevator car is the core control area. The terminal is installed in the center of the top of the car, with the lens tilted vertically downwards, completely covering the entire interior space of the car, monitoring all items entering the elevator in real time. The entrance corridor is the last control point before entering the apartment. One terminal is installed 1.8 meters high on the side wall of the elevator lobby on each floor, covering the corridor passage area and the area in front of each apartment door, achieving full-link sensing coverage from entering the unit to entering the apartment.

[0015] Each of the three types of sensors has a clearly defined function and parameter specifications. The high-definition camera uses a visible light imaging module with a resolution of 1920×1080 and a frame rate of 5 frames per second, balancing image clarity with computational power consumption. It can clearly capture visual information such as the shape outline, surface markings, and structural features of objects, and has built-in infrared supplementary lighting for nighttime operation, adapting to 24-hour continuous data acquisition needs. The electromagnetic field sensor uses a wide-band magnetic sensing module, covering a detection frequency range of 0.1MHz to 100MHz, with a sensitivity of up to 0.01μT. It can accurately capture weak electromagnetic radiation signals in space, corresponding to the specific electromagnetic characteristics generated by lithium battery protection boards and cell self-discharge, effectively distinguishing ordinary metal objects from charged energy storage devices. The density sensor uses a frequency-modulated continuous wave millimeter-wave radar module, operating at a frequency of 24GHz. By emitting millimeter waves and receiving the energy attenuation characteristics of the reflected echoes from objects, it calculates the average volume density of objects within the detection area, expressed in g / cm³. 3 Density detection accuracy can reach 0.05 g / cm³. 3 It can distinguish high-density energy storage items such as lithium batteries from low-density items such as foam and plastic bags.

[0016] Clock synchronization calibration is divided into two levels: global terminal synchronization and single-terminal internal synchronization. All terminals are connected to the building's local area network and use the Network Time Protocol (NAT) for unified clock calibration. The calibration cycle is automatically performed every hour to ensure that the system time deviation of all terminals is less than 1ms. The three types of sensors within a single terminal achieve synchronous data acquisition through hardware trigger signals. The terminal's main control unit issues a unified acquisition trigger pulse, and the camera, electromagnetic field sensor, and density sensor simultaneously respond to the trigger signal to complete data acquisition, eliminating the time difference between multiple sensors within a single terminal at the hardware level. After the installation, hardware debugging, and global clock calibration of all terminals are completed, all terminals form an interconnected and clock-unified sensing terminal cluster. Each terminal has independent data acquisition, preliminary preprocessing, and data uploading capabilities, providing hardware support for the continuous generation of multi-source sensing data.

[0017] Each sensing terminal is activated to synchronously collect raw image data, electromagnetic field strength data, and object density data according to a unified sampling period. Each frame of data is automatically associated with the collection timestamp and device number to generate a multi-source raw sensing dataset. The core of this step is to initiate parallel data acquisition by the full-scale sensing terminals according to a unified time-series benchmark, acquire three types of raw modal data and add unique identifiers to form a raw data set with spatiotemporal labels, ensuring that the data is traceable and matchable, and providing raw input for subsequent spatiotemporal alignment processing.

[0018] The unified sampling period is set to 200ms, meaning 5 complete multimodal data acquisitions are completed per second. This sampling frequency has been optimized for specific scenarios, meeting the continuous trajectory capture requirements of electric vehicles and personnel movement without causing target loss due to excessively long sampling intervals, nor causing excessive computing power waste and storage pressure due to excessively high sampling frequencies. All terminals strictly adhere to this sampling period, synchronously triggering acquisition actions under the drive of a unified clock, ensuring that all prevention and control points within the building complete data acquisition synchronously at the same time, without any timing misalignment between points.

[0019] The raw image data is in visible light RGB format. Each frame contains visual information of all people and objects within the channel. The camera automatically adjusts exposure parameters and white balance during acquisition to adapt to changes in lighting in corridors and elevators, ensuring clear imaging even in low-light environments. Infrared supplementary lighting automatically activates at night, maintaining continuous 24-hour acquisition. Electromagnetic field strength data is spatial full-domain electromagnetic intensity distribution data. The sensor completes a full-frequency scan in a single acquisition, outputting electromagnetic field strength values ​​corresponding to different frequency points, fully recording the electromagnetic radiation characteristics within the detection area. It can distinguish specific harmonic components generated during lithium battery operation from environmental background electromagnetic noise. Object density data is spatial density distribution data within the radar detection area. The radar outputs object density values ​​corresponding to different distances and angles, allowing for the location of objects in space and the calculation of their overall average density, distinguishing the properties of objects made of different materials.

[0020] After each data acquisition operation, the terminal control unit automatically adds a unified acquisition timestamp and device number to the three types of data in that frame. The timestamp is accurate to the millisecond level and is formatted as year, month, day, hour, minute, second, and millisecond, accurately recording the precise moment the data was generated. The device number is a unique identification code for each terminal, containing location type and sequence information. For example, the prefix for the terminal number at the unit entrance is DY, for the elevator car it's DT, and for the entrance corridor it's DL, with a suffix of a numerical sequence number. The physical location corresponding to the data can be directly located through the number. The three types of data share the same timestamp and device number, naturally possessing a common source association attribute.

[0021] The raw, labeled data continuously collected by all terminals is aggregated to the edge processing node via the local area network to form a multi-source raw perception dataset containing all locations, all modalities, and all time periods. This dataset is the raw collection result without filtering or processing, and completely preserves all perception information in the environment, providing the most original data support for subsequent spatiotemporal alignment and feature extraction.

[0022] The multi-source raw sensing dataset is time-stamp aligned and spatial coordinates are normalized. Images, electromagnetic fields and density data collected at the same time are bound into a single sensing recording unit to generate spatiotemporally aligned multimodal sensing records. The core of this step is to calibrate and align the multi-source raw data from both temporal and spatial dimensions, eliminate temporal deviations and spatial coordinate differences between different sensors, and bind the originally independent three types of modal data into a spatiotemporally unified minimum sensing unit. This ensures that when multimodal features are fused subsequently, the visual, electromagnetic, and density features of the same target completely correspond to the same time and the same spatial location, avoiding recognition errors caused by feature misalignment.

[0023] Timestamp alignment is achieved using a benchmark interpolation method. Due to slight differences in the response delays of the three types of sensors, even with hardware synchronization, the actual acquisition completion time will still have a sub-millisecond deviation. Furthermore, the data transmission delay at different terminal locations also varies slightly. Therefore, software-level timestamp alignment is necessary. Alignment is based on the image data acquisition timestamp. A linear interpolation algorithm is used to resample the electromagnetic field and density data in time sequence. For electromagnetic and density data at two adjacent sampling times, linear interpolation is used to calculate the corresponding values ​​at the image acquisition time, ensuring that the time deviation of the three types of data after final alignment is less than 0.5ms, fully meeting the time accuracy requirements for real-time recognition. For cross-terminal data, a globally unified clock is also used as a benchmark, and the sensing data from different locations at the same time are matched according to the timestamp to achieve global time sequence alignment.

[0024] Spatial coordinate normalization is achieved through a unified world coordinate system. The installation positions, detection angles, and detection ranges of the three types of sensors differ, resulting in inconsistent coordinate positions for the same object in the raw data from the three types of sensors. Therefore, a unified world coordinate system is needed to normalize the spatial coordinates. The unified coordinate system uses the ground point directly below each terminal as the origin, with the x-axis extending along the channel width, the y-axis along the channel depth, and the z-axis perpendicular to the ground upwards. The unit is meters. For high-definition cameras, the intrinsic and extrinsic parameters are obtained in advance using a checkerboard calibration method, establishing a mapping relationship between pixel coordinates and world coordinates, converting any pixel in the image into real spatial coordinates. For millimeter-wave density sensors, the radar's installation height, elevation angle, and horizontal deflection angle are calibrated in advance, establishing a conversion formula between radar polar coordinates and world coordinates, mapping the target position detected by the radar to the unified coordinate system. For electromagnetic field sensors, the coordinate mapping of their spatial detection range is determined through multi-point calibration, mapping the electromagnetic intensity distribution to a unified spatial coordinate system.

[0025] After completing temporal alignment and spatial coordinate normalization, image data, electromagnetic field intensity spatial distribution data, and object density spatial distribution data from the same time and terminal are bound together and encapsulated into a single independent sensing recording unit. Each sensing recording unit has a unique spatiotemporal identifier, containing complete data of the three modalities and a unified coordinate mapping relationship. The data correspond one-to-one, with no misalignment or deviation. All bound sensing recording units together form a spatiotemporally aligned multimodal sensing record. This record eliminates the system errors of multiple sensors, achieves dual unification of the temporal and spatial dimensions, and provides standardized input data for subsequent scene classification and multimodal target recognition. It can be directly input into the recognition model for feature extraction and analysis.

[0026] The spatiotemporally aligned multimodal sensing records are encapsulated according to time series, data integrity verification identifiers are added, and a continuously updated multi-source sensing data stream is generated.

[0027] The core of this step is to standardize and stream the discrete multimodal sensing records, add a data verification mechanism to ensure transmission reliability, and form a continuous, stable, and real-time readable data stream. This provides continuous data input for subsequent identification, judgment, and handling throughout the entire process, supporting the real-time operation of the entire early warning and control system.

[0028] All spatiotemporally aligned sensing recording units are arranged strictly in chronological order of acquisition time, with records having earlier timestamps appearing first, forming a continuous time sequence. Encapsulation uses a standardized data frame format, with each frame containing, in sequence, a frame header identifier, frame sequence number, timestamp, device number, data length, three types of sensing data body, checksum, and frame tail identifier. The frame sequence number is a continuously incrementing integer; it is automatically incremented by 1 each time a new sensing recording unit is generated, used to determine if frame loss occurs during transmission. The data length field clearly indicates the total number of bytes in the frame, facilitating rapid parsing and reading by the receiving end.

[0029] Data integrity verification employs a 32-bit cyclic redundancy check (CRC) mechanism, with a corresponding checksum appended to the end of each data frame as an integrity verification identifier. The checksum is generated by the sending end using a standard algorithm based on all data content within the frame. Upon receiving the data, the receiving end recalculates the checksum using the same algorithm and compares it with the checksum carried within the frame. If the two calculations are identical, it indicates that the data frame has not been damaged, lost, or tampered with during transmission, and the data is complete and valid. If the results are inconsistent, the data frame is deemed invalid and discarded. The corresponding sensing data is then supplemented by linear interpolation using adjacent valid frames, preventing single-frame errors from affecting the continuity of subsequent recognition results. The introduction of this verification mechanism effectively shields against data errors caused by network transmission fluctuations, ensuring the reliability and stability of the data stream.

[0030] The encapsulated and verified continuous time-series data forms a continuously updated multi-source sensing data stream. This data stream employs a streaming transmission mode, collecting, processing, and transmitting data simultaneously. Data is pushed to backend processing nodes immediately after generation, with an end-to-end transmission latency of less than 50ms, fully meeting the response speed requirements for real-time early warning. The data stream is characterized by continuous generation, temporal continuity, multimodal binding, and traceability and verifiability, covering all monitoring points across all time periods. It serves as the data input foundation for the entire intelligent early warning and control system. Subsequent processing steps, such as spatial scene recognition, multimodal target recognition, and end-to-end risk assessment, are all based on this multi-source sensing data stream. The stability, accuracy, and real-time performance of the data stream directly determine the operational effectiveness of the entire control system.

[0031] S202, perform spatial closure attribute recognition and location information joint discrimination on the multi-source sensing data stream, distinguish between outdoor open areas and elevator and entrance closed areas, and generate scene type labels; Specifically, the location coordinates and spatial environment features of the current sensing terminal can be extracted from the multi-source sensing data stream to generate a location and environment feature vector; The core of this step is to decouple spatial positioning information from environmental physical features from structured multi-source sensing data streams, completing the transformation of raw sensing data into feature vectors. This provides standardized input data for subsequent spatial classifiers, ensuring that classification decisions have a unified numerical benchmark. The specific implementation method is as follows: The location coordinates are fixed spatial coordinates pre-calibrated for each integrated sensing terminal before it leaves the factory. A local 3D Cartesian coordinate system is used, with the origin set at the center point of the building unit entrance. The x-axis extends horizontally along the building, the y-axis extends vertically along the building, and the z-axis corresponds to the floor height vertically. The unit is uniformly meters, and the coordinate accuracy is controlled within 0.1 meters. Each terminal has a unique device number bound to its own 3D coordinates, stored in the terminal's local configuration file. During data collection, each sensing record is automatically embedded with a timestamp, allowing the precise spatial location of the terminal to be obtained without real-time dynamic positioning. In the example, the coordinates of the sensing terminal on the top of the elevator car are (12.5, 8.3, 12.0), corresponding to the elevator shaft area on the 3rd floor; the coordinates of the terminal at the unit entrance are (12.5, 2.1, 0.0), corresponding to the outdoor open area on the first floor; and the coordinates of the terminal on the side wall of the entrance corridor are (10.2, 15.6, 12.0), corresponding to the enclosed corridor area outside the resident's door on the 3rd floor.

[0032] The spatial environment characteristics are extracted from three types of data in the multi-source sensing data stream: image, electromagnetic field, and density. These characteristics comprise six core dimensions: illumination uniformity, equivalent reverberation time, electromagnetic field reflection coefficient, mean ambient background density, illumination fluctuation amplitude, and satellite signal penetration loss. Illumination uniformity is calculated from image data, ranging from 0 to 1. A value closer to 1 indicates a more uniform illumination distribution within the space. Illumination uniformity in enclosed artificially lit spaces is generally higher than 0.85, while the uniformity in open natural light environments is typically lower than 0.6. Equivalent reverberation time is indirectly calculated using the echo characteristics of electromagnetic field sensors, measured in seconds. Enclosed spaces experience more sound wave reflections, resulting in reverberation times ranging from 0.6 to 1.2 seconds, while open spaces typically have reverberation times lower than 0.2 seconds. The electromagnetic field reflection coefficient characterizes the ability of a spatial boundary to reflect electromagnetic waves. The reflection coefficient of a metal elevator car can reach 0.9, while that of a concrete-walled corridor is approximately 0.6, and that of open areas is lower than 0.3. The other three dimensions correspond to environmental stability and signal attenuation characteristics, respectively, which together characterize the physical enclosure properties of the space.

[0033] The three-dimensional position coordinates and six-dimensional environmental features are sequentially concatenated to form a nine-dimensional position and environment feature vector. Each dimension of the vector is normalized, with values ​​uniformly mapped to the interval between 0 and 1, eliminating the interference of different units on the classification results. In the example, the feature vector corresponding to the elevator car terminal is (0.52, 0.35, 0.40, 0.92, 0.85, 0.88, 0.76, 0.12, 0.94), where the first three dimensions are the normalized position coordinates and the last six dimensions are the normalized environmental features, fully carrying the spatial location and environmental physical attribute information of the terminal.

[0034] The location and environmental feature vectors are input into a pre-trained spatial classifier. By judging whether the location coordinates are located in the elevator shaft or the area inside the entrance door, and whether the environmental features conform to the sound and light propagation law of a closed space, a spatial closure attribute probability value is generated. The core of this step is to rely on a pre-trained lightweight spatial classification model, which integrates prior location information and environmental physical features for dual judgment, outputting a continuous probability value in the range of 0 to 1. This quantitatively represents the degree of closure of the current space, providing a continuous numerical basis for subsequent threshold determination. The specific implementation method is as follows: The pre-trained spatial classifier employs a lightweight gradient boosting tree model. The training dataset covers tens of thousands of feature samples across three scenarios: open areas, semi-open transition areas, and fully enclosed areas, all under different building structures. During training, differentiated weights are assigned to location features and environmental features, with environmental features accounting for 65% and location features for 35%. This approach leverages prior constraints based on location coordinates to narrow the judgment range while avoiding the limitation of pure location-based judgments failing to adapt to differences in building structures. The classifier outputs a spatial closure probability value P, ranging from 0 to 1. A P value closer to 1 indicates a higher degree of spatial closure, while a P value closer to 0 indicates a higher degree of spatial openness.

[0035] The location coordinate determination logic is based on preset electronic fence rules. The system pre-marks the coordinate range of all elevator shafts and entrance door areas in the building coordinate system, forming a set of enclosed space electronic fences. When the location coordinates in the input feature vector fall within any enclosed electronic fence, the location dimension is determined to meet the enclosed condition, and the corresponding location sub-probability P_pos is assigned a value of 0.8; if the coordinates fall within the electronic fence of the semi-open transition area, P_pos is assigned a value of 0.4; if the coordinates fall within the outdoor open area fence, P_pos is assigned a value of 0.1. In the example, the elevator car terminal coordinates fall completely within the elevator shaft electronic fence range, so P_pos is 0.8; the unit lobby terminal coordinates fall within the semi-open transition area fence, so P_pos is 0.4; and the outdoor unit entrance terminal coordinates fall within the open area fence, so P_pos is 0.1.

[0036] The environmental feature determination logic is based on the physical laws of sound and light propagation in enclosed spaces. Six-dimensional environmental features are compared with preset standard feature thresholds for enclosed spaces. The matching degree of each dimension is calculated independently, and then a weighted sum is used to obtain the environmental sub-probability P_env. The standard feature thresholds for enclosed spaces are: illumination uniformity greater than 0.8, reverberation time greater than 0.5 seconds, electromagnetic field reflection coefficient greater than 0.7, background density mean stable within a fixed range, illumination fluctuation amplitude less than 0.2, and signal penetration loss greater than 20dB. Each dimension's matching degree is scored out of 1 point, and the weighted sum yields an environmental sub-probability ranging from 0 to 1. In the example, the matching degrees for the six environmental dimensions inside the elevator car are 0.95, 0.92, 0.90, 0.88, 0.94, and 0.96, respectively, resulting in a weighted P_env value of 0.92. The matching degree of the entrance corridor is slightly lower, with a P_env value of 0.82. The matching degree of the outdoor open area is low, with a P_env value of 0.15.

[0037] The final spatial closure probability value is obtained by fusing the location sub-probability and the environment sub-probability according to weights. The fusion formula is P = w_pos × P_pos + w_env × P_env, where w_pos is the location weight with a value of 0.35 and w_env is the environment weight with a value of 0.65. Substituting the example values ​​for the elevator car, we get P = 0.35 × 0.8 + 0.65 × 0.92 = 0.878. This value is in the high range, indicating that the current spatial closure attribute is extremely strong. Similarly, the probability value for the entrance corridor is approximately 0.80, for the unit lobby it is approximately 0.54, and for the outdoor entrance it is approximately 0.13. The probability values ​​for different scenarios show a clear gradient distinction, providing clear numerical boundaries for subsequent threshold determination.

[0038] When the probability value of the spatial enclosure attribute is greater than the enclosure threshold, the floor change data in the joint location information and the navigation signal strength are cross-validated to exclude transition areas such as semi-open corridors or lobbies, and generate a preliminary judgment result of the scene type. The core of this step is to set a hard threshold to filter out high-probability closed candidates, and then introduce two independent verification features for cross-validation to eliminate semi-open scenes in the boundary region, avoiding misjudgments caused by similar environmental features and improving the accuracy of scene classification. The specific implementation method is as follows: The closure threshold is a critical value that distinguishes between open and closed scenarios. Based on engineering measurement data and the target for optimizing the false positive rate, the closure threshold T is set to 0.70. When the probability value P of the spatial closure attribute is greater than or equal to 0.70, the current scenario is determined to be a candidate for closure and proceeds to the subsequent cross-validation stage. When the P value is less than 0.70, it is directly determined to be an outdoor open or semi-open area, without initiating the subsequent recognition process. This reduces invalid calculations and improves terminal operating efficiency. In the example, the P values ​​for the elevator car (0.878) and the entrance corridor (0.80) are both greater than 0.70, thus proceeding to cross-validation; the P values ​​for the unit lobby (0.54) and the outdoor entrance (0.13) are both less than 0.70, thus directly determining it as an open area.

[0039] The first cross-validation dimension is floor change data, sourced from a high-precision barometric altimeter built into the sensing terminal, with a sampling frequency of 1Hz. This data can be indirectly used to calculate the current floor level, achieving a floor resolution accuracy of 0.1 floors. When the elevator car is in operation, the floor level changes continuously and linearly over time, with a change greater than 0.3 floors per 10 seconds. In static areas such as entrance corridors and lobbies, the floor level remains stable over a long period, with a change less than 0.05 floors per 10 seconds. By calculating the floor fluctuation variance over 10 consecutive seconds, the dynamically operating elevator space can be distinguished from the static corridor and lobby spaces. In the example, the elevator car in operation rises from floor 2.1 to floor 3.2 within 10 seconds, a change of 1.1 floors with a variance of 0.12, which conforms to the dynamic characteristics of an elevator; the stationary entrance corridor remains stable at floor 3.0 within 10 seconds, a change of 0.02 floors with a variance of 0.001, which conforms to the characteristics of a static enclosed corridor; the unit lobby floor is also stable, with a change of 0.03 floors, which conforms to the characteristics of a static semi-open corridor.

[0040] The second cross-validation dimension is navigation signal strength, specifically the received power of satellite navigation signals, measured in dBm. A higher value indicates a stronger signal. Enclosed elevator cars and entrance corridors are located inside buildings, where walls and metal structures severely obstruct satellite signals, resulting in received power typically below -110 dBm, and even below -120 dBm inside the elevator car. Semi-open unit lobbies, with entrances and exits connecting to the outside, allow partial signal penetration, with received power typically between -70 dBm and -90 dBm. Signal strength in open outdoor areas exceeds -60 dBm. Signal strength thresholds can further distinguish between enclosed and semi-open transitional spaces within a building. In the example, the navigation signal strength inside the elevator car is -126 dBm, far below the enclosed threshold of -110 dBm; the entrance corridor signal strength is -115 dBm, also below the threshold; and the unit lobby signal strength is -82 dBm, above the threshold, classifying it as a semi-open area.

[0041] Cross-validation employs a dual-condition judgment rule: only when both floor change characteristics and navigation signal strength conform to the characteristics of a confined space is it ultimately determined to be an elevator and entrance confined area. If either condition is not met, it is judged as a semi-open or open area and is excluded. In the example, the elevator car meets both conditions and is initially judged as a confined area; the entrance corridor has stable floors and weak signal, both conditions conforming to the static confined characteristics, and is initially judged as a confined area; the unit lobby has stable floors but high signal strength, not meeting the second condition, and is excluded, judged as a semi-open area. After validating all candidate scenarios, the preliminary judgment result of the scenario type is output, clearly indicating whether the current scenario belongs to an outdoor open area or an elevator and entrance confined area.

[0042] The preliminary scene type determination results are processed by time-series smoothing to eliminate single-frame misjudgment and edge jitter, and finally generate scene type labels for outdoor open areas or elevator and enclosed entrance areas corresponding to the current perception cycle.

[0043] The core of this step is to filter out instantaneous misjudgments caused by fluctuations in single-frame data and jitter in boundary areas through continuous frame statistical verification in the time domain, so as to keep the scene label output continuous and stable and avoid frequent switching that triggers unnecessary recognition processes. The specific implementation method is as follows: The temporal smoothing process employs a sliding window majority voting method. The sliding window length is set to 5 consecutive sensing cycles, with each sensing cycle lasting 1 second. This means the window covers the preliminary judgment results of the past 5 frames within the last 5 seconds. The voting rules are as follows: if more than 3 frames within the window indicate a closed area, the final judgment result for the current frame is the elevator and entrance closed area; if more than 3 frames within the window indicate an open area, the final judgment result is the outdoor open area; if the number of frames in both categories is equal, the judgment result of the previous frame is retained to maintain output continuity. This majority voting mechanism effectively offsets misjudgments caused by single-frame data anomalies. For example, during the moment the elevator doors open and close, the navigation signal briefly strengthens, and environmental characteristics fluctuate, potentially leading to a single frame being judged as an open area while the other 4 frames before and after are closed. After voting, the closed label is still output, eliminating edge jitter.

[0044] In the example, at the moment the elevator reaches the 3rd floor and the door opens, the initial judgment result of the 3rd frame changes to an open area due to signal fluctuations. The remaining 1st, 2nd, 4th, and 5th frames are all closed areas. Four out of the five frames are closed, exceeding the 3-frame threshold. Therefore, the smoothed current frame still outputs the closed label, and the instantaneous fluctuations of the door opening and closing are completely filtered out. Another example is the transition phase of a person carrying a target from outdoors into a unit lobby. The first three frames are open, and the last two frames are semi-open. After voting, the open label is still output until three or more consecutive frames are determined to be closed before the label changes, avoiding frequent scene label jumps.

[0045] In addition to discrete label voting, the probability values ​​of spatial closure attributes are also processed by first-order inertial filtering. The filtering formula is P_smooth = α × P_current + (1-α) × P_last, where α is the filtering coefficient with a value of 0.3, P_current is the original probability value of the current frame, and P_last is the smoothed probability value of the previous frame. Low-pass filtering further smooths the probability curve, allowing the probability values ​​to transition slowly without abrupt jumps. This, combined with the voting method, forms a dual smoothing mechanism, further improving output stability.

[0046] After time-series smoothing, standardized scene type labels are generated. These labels are divided into two categories: outdoor open areas and elevator / entrance enclosed areas. Each label includes a corresponding smoothed confidence value and timestamp, which are synchronously transmitted to subsequent modules along with the multi-source sensing data stream. When the label is for an elevator / entrance enclosed area, the subsequent multimodal AI fusion recognition process is automatically triggered. When the label is for an outdoor open area, a low-power standby state is maintained, continuously collecting data without activating the recognition algorithm, balancing detection accuracy and terminal power consumption. The final output scene labels are continuous and stable, with a single-frame false positive rate of less than 0.5% and boundary area jitter lasting no more than 2 seconds, fully meeting the real-time and accuracy requirements of building security.

[0047] S203, according to the scene type label, initiate multimodal AI fusion recognition in the enclosed area. Based on the pre-trained electric vehicle single lithium battery exclusive recognition model, perform multi-dimensional fusion comparison of the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features, and remove interference from similar items through the differentiated feature library to generate single lithium battery target recognition results containing target category and confidence level. Specifically, when the scene type label is determined to be an elevator car or a closed area at the entrance, the multimodal AI fusion recognition engine is triggered to enter the working state, extracting visual appearance features, including battery size, electrode arrangement and shell markings, from the corresponding time period of the multi-source perception data stream, and generating a visual feature vector. The core of this step is to enable the recognition engine to be triggered on demand based on the scene classification results. High-precision multimodal recognition is only activated in enclosed and controlled areas such as elevators and entrances, while maintaining a low-power standby state in open outdoor areas. This saves terminal computing resources, reduces the probability of false alarms, and ensures the recognition sensitivity of the core controlled areas, providing standardized visual feature input for subsequent multi-dimensional feature fusion.

[0048] The specific implementation is as follows: The triggering logic of the multimodal AI fusion recognition engine is driven by the scene type label output by the front-end spatial classification module. When the label is stable for two consecutive perception cycles as an elevator car or a closed area at the entrance, the engine automatically wakes up without any manual intervention. The wake-up response latency is less than 100ms, ensuring that detection starts immediately after the target enters the closed area. The visual features are collected from the high-definition camera built into the integrated perception terminal, with a resolution of 1920×1080 and a frame rate of 15 frames per second. The acquired raw images are first preprocessed, including denoising, distortion correction, and grayscale normalization, to eliminate feature deviations caused by changes in lighting and lens distortion. The preprocessed images are then input into a deep feature extraction network to extract the appearance features of the target layer by layer.

[0049] The three extracted core visual appearance features each have a clear discriminative function. The first category is the shape and size feature. By combining monocular ranging with pixel ratio, the actual length, width, and height of the target are calculated, accurate to the centimeter. The typical shape of a single lithium battery cell in an electric bicycle is a regular cuboid, with common size ranges of 250mm to 350mm in length, 120mm to 200mm in width, and 100mm to 180mm in height. The length-width-height ratio is usually stable at around 1.5:1:0.8. This feature can preliminarily distinguish similar items such as round batteries, small power banks, and large energy storage batteries. The second category is the electrode arrangement feature. The position, number, and spacing of the metal electrodes in the top area of ​​the target are extracted. The positive and negative electrodes of an electric bicycle lithium battery are usually symmetrically distributed at both ends of the top surface. The electrode size is relatively large and the spacing is fixed, which is different from the micro interface of a power bank and the side bolt electrodes of a lead-acid battery. The third category is the shell marking features, which extracts the texture and color features of warning signs, brand codes, and parameter labels on the shell surface. Lithium battery shells generally have yellow high-voltage warning signs, flame-retardant signs, and capacity parameter labels, and the texture features are unique.

[0050] The three types of features mentioned above are numerically normalized, with all feature values ​​mapped to the range of 0 to 1. These values ​​are then concatenated to form a 128-dimensional visual feature vector, where each dimension corresponds to a matching degree value for a specific visual feature. In the example, the matching degree for the current target's external dimensions is 0.83, the matching degree for the electrode arrangement is 0.78, the matching degree for the casing markings is 0.72, and the overall matching degree between the visual features and the standard lithium battery template is 0.78. The generated visual feature vector can be directly input into the subsequent fusion recognition model.

[0051] Simultaneously extract frequency domain features of electromagnetic field intensity data from multi-source sensing data streams, obtain characteristic harmonic components under lithium battery charging and discharging states, calculate the average density value of objects from density data, and generate a fusion vector of electromagnetic features and density features. The core of this step is to simultaneously extract two types of non-visual features: electromagnetic and physical density. This compensates for the shortcomings of visual features, which are easily affected by occlusion and lighting. By utilizing the unique electromagnetic radiation and density characteristics of lithium batteries, a differentiated recognition dimension is constructed, which complements the visual features and provides multi-dimensional feature support for multimodal fusion.

[0052] The specific implementation is as follows: The electromagnetic field data comes from the electromagnetic field sensor built into the integrated sensing terminal. The sensor's frequency detection range is 1kHz to 100kHz, and the sampling frequency is 200kHz, which can accurately capture weak electromagnetic radiation signals. The raw time-domain electromagnetic field data is first processed by bandpass filtering to remove background noise such as 50Hz power frequency interference and broadcast signals in the environment. Then, it is converted into frequency domain data through fast Fourier transform, and the signal amplitude at each frequency point is extracted to form a spectrum distribution curve. The battery management system protection board inside the lithium battery generates switching harmonics of a fixed frequency when it is working. The typical characteristic frequency is concentrated in the range of 25kHz to 40kHz. The characteristic harmonics in this frequency band are the core electromagnetic identifier that distinguishes lithium batteries from other metal objects and ordinary batteries. Even when the battery is in a static state, the standby power consumption of the protection board will generate weak characteristic harmonics, which can be stably captured by the sensor. In the example, the frequency domain data corresponding to the current target shows a significant peak at the 32kHz frequency point, with an amplitude of 0.13mT. The matching degree with the standard lithium battery electromagnetic characteristic template reaches 0.89, indicating the existence of lithium battery-specific electromagnetic characteristics.

[0053] Object density data is acquired by a millimeter-wave radar density sensor built into the terminal. The sensor transmits frequency-modulated continuous waves, receives the echo signal reflected from the target, calculates the target's volume and equivalent mass, and then derives the average density value. The density detection accuracy can reach 0.1 g / cm³. 3 The detection distance covers a range of 0.5m to 5m. The average density of a single lithium battery cell in an electric vehicle is typically 1.2g / cm³. 3 Up to 1.8 g / cm 3 The density is significantly higher than that of empty cardboard boxes and plastic suitcases, but lower than that of solid metal blocks and lead-acid batteries, making it an important physical characteristic for distinguishing similar items. In the example, the average density of the currently detected target is 1.54 g / cm³. 3 It falls within the standard density range of lithium batteries, with a density feature matching degree of 0.86.

[0054] The electromagnetic feature parameters—spectral distribution, frequency matching degree, and signal amplitude—are normalized along with the density feature parameters—average density value and density distribution uniformity—and then concatenated to form a 64-dimensional electromagnetic and density feature fusion vector. During the fusion process, the physical independence of the two types of features is preserved; only dimensional concatenation is performed, without weighted calculations, ensuring that the original feature information is completely transmitted to the subsequent recognition model and avoiding information loss.

[0055] The fused vectors of visual features, electromagnetic features, and density features are input into a pre-trained recognition model specifically for single lithium batteries in electric vehicles. The contribution weight of each modality feature to the target discrimination is calculated through a multimodal feature cross-attention mechanism, and then compared and eliminated with similar item features in the differential feature library to generate a fused recognition score. The core of this step is to rely on a specially trained recognition model, dynamically allocate the weights of each modality feature through a cross-attention mechanism, adaptively adapt the feature reliability under different scenarios, and filter out interference from easily confused items through a differentiated feature library, thereby significantly reducing the probability of misidentification and outputting a fusion recognition score that integrates multi-dimensional information, providing a quantitative basis for the final target determination.

[0056] The specific implementation is as follows: The dedicated identification model for single lithium batteries in electric bicycles is a specialized classification model pre-trained based on a deep learning framework. The training dataset contains tens of thousands of samples of lithium batteries from electric bicycles of different brands, capacities, and appearances, covering scenes with different lighting conditions, angles, and degrees of occlusion. Corresponding electromagnetic and density feature data are also collected to ensure that the model learns multi-dimensional essential features of the target, rather than just single appearance features. During model training, a multimodal joint loss function is used, simultaneously constraining the feature extraction accuracy of the three modalities. The overall recognition accuracy can reach over 98%, with a false positive rate of less than 1%.

[0057] The core function of the multimodal feature cross-attention mechanism is to dynamically adjust the weight ratios of visual, electromagnetic, and density features based on the feature quality of the current scene, avoiding recognition errors caused by the failure of a single modality. The logic for weight calculation is as follows: first, the confidence score of each modality feature is calculated separately. The confidence score of visual features is determined by the target occlusion ratio and the uniformity of illumination; the more occlusion and the worse the illumination, the lower the confidence score. The confidence score of electromagnetic features is determined by the intensity of environmental background noise; the stronger the background noise, the lower the confidence score. The confidence score of density features is determined by the target distance and the degree of overlap between multiple objects; the farther the distance and the more overlap, the lower the confidence score. Then, the corresponding weights are calculated based on the confidence scores of each modality. The weight calculation formula is w_i = c_i / (c_v + c_e + c_d), where w_i is the weight of the i-th type of feature, c_i is the confidence score of the corresponding feature, and c_v, c_e, and c_d are the confidence scores of the visual, electromagnetic, and density types of features, respectively. The sum of all weights is 1 to ensure weight normalization. In the example, the target in the current scene has 30% area occlusion. The confidence score of the visual feature c_v is 0.6, the ambient electromagnetic background noise is low, the confidence score of the electromagnetic feature c_e is 0.9, the target has no overlap, and the confidence score of the density feature c_d is 0.8. Substituting these values ​​into the formula, the calculated visual weight w_v is 0.26, the electromagnetic weight w_e is 0.39, and the density weight w_d is 0.35. The electromagnetic and density features have higher weights, which is suitable for the recognition needs in occluded scenes.

[0058] The differentiated feature library pre-stores multimodal feature data for over ten items easily confused with lithium batteries, including common items such as lead-acid batteries, power banks, laptops, metal toolboxes, plastic storage boxes, and bottled drinking water. Each item corresponds to three types of feature templates: visual, electromagnetic, and density. The comparison and elimination logic is as follows: the multimodal features of the current target are compared with the features of each similar item in the library to calculate the cosine similarity. When the similarity is greater than 0.7, the current target is considered to have a high degree of feature overlap with that similar item and needs to be further eliminated; when the similarity is less than 0.7, it is determined that it does not belong to that category of similar items. In the example, the feature similarity between the current target and a lead-acid battery is 0.32, with a power bank is 0.58, and with a metal toolbox is 0.41, all of which are below the threshold of 0.7, successfully eliminating interference from all similar items and eliminating the risk of mismatch.

[0059] The final fusion recognition score is calculated using the formula S = w_v * S_v + w_e * S_e + w_d * S_d, where S is the final fusion recognition score, ranging from 0 to 1, with higher values ​​indicating a greater probability of identifying the target; S_v is the visual feature matching score, S_e is the electromagnetic feature matching score, and S_d is the density feature matching score. Substituting the weights and modal scores from the example, S_v is 0.78, S_e is 0.89, and S_d is 0.86, resulting in a fusion recognition score S of 0.85. This score integrates the matching results of the three types of features with a reliability weight, making it more accurate than single-modal recognition.

[0060] Based on the fusion recognition score and the preset multi-level confidence threshold, it is determined whether the current target is a single lithium battery cell of an electric vehicle and its carrying method, and finally a single lithium battery cell target recognition result containing target category label and confidence value is generated.

[0061] The core of this step is to classify and determine the fusion score by using multi-level confidence thresholds, clarify the target attributes and confidence levels, and identify the way the target is carried. This provides a clear basis for subsequent end-to-end review and graded handling, ensuring that the output identification results are traceable and implementable.

[0062] The specific implementation is as follows: The preset multi-level confidence thresholds are divided into three levels: low confidence threshold (0.6), medium confidence threshold (0.75), and high confidence threshold (0.85). The threshold division is based on statistical results from a large amount of actual test data, balancing recognition sensitivity and accuracy to avoid missed and false alarms. The judgment rule is as follows: when the fusion recognition score is greater than or equal to 0.85, it is judged as a high-confidence target, confirming it as a single lithium battery cell for an electric vehicle, and subsequent warning logic can be directly triggered; when the score is between 0.75 and 0.85, it is judged as a medium-confidence suspected target, requiring further verification using multiple consecutive frames of time-series data. If the score is higher than 0.75 for three consecutive frames, it is upgraded to high confidence; otherwise, it is downgraded to a non-target; when the score is between 0.6 and 0.75, it is judged as a low-confidence target awaiting confirmation, only recorded without triggering a warning, and continuously tracked and observed; when the score is lower than 0.6, it is judged as a non-target item, and no warning processing is performed.

[0063] After confirming the target as a lithium battery, the specific carrying method is determined by the positional relationship and size ratio between the human body silhouette and the battery target in visual features. Common carrying methods include hand-held, shoulder bag, rolling suitcase, shopping bag, and trolley transport. The determination logic is as follows: if the battery target is located in the hand area and is relatively small, it is determined to be hand-held; if the battery target is located on the side of the human torso and overlaps with the silhouette of a backpack, it is determined to be backpack-carried; if the battery target is located in front of or to the side of the human body and matches the silhouette of a rolling suitcase, it is determined to be suitcase-carried. In the example, the current target fusion score is 0.85, reaching the high confidence threshold, and it is determined to be a confirmed single lithium battery for an electric vehicle, and the carrying method is hand-held.

[0064] The target identification results for individual lithium batteries adopt a structured data format, containing seven categories of information: target category label, confidence score, target location coordinates, carrying method, matching scores for each modality, timestamp, and device number. The data format is standardized and can be directly passed to subsequent cross-site verification and processing modules. To further improve the reliability of the results, a temporal smoothing verification is performed before outputting the single-frame identification result. Only when the target is identified as the same type for three consecutive sensing cycles will the final identification result be output, eliminating the impact of occasional misjudgments in a single frame.

[0065] S204, The target identification result of the single lithium battery is correlated and verified among the multi-point sensing terminals at the unit entrance, elevator car and entrance corridor to generate the full-link risk judgment result. Specifically, the single lithium battery target identification results generated by the unit entrance sensing terminal can be carried with timestamps and spatial coordinates and broadcast to the associated sensing terminals in the elevator car and entrance corridor through the local area communication network to generate a cross-point waiting queue for verification. The core of this step is to achieve cross-terminal synchronization of recognition results and distribution of tasks to be reviewed, and to connect the information links of the three types of sensing nodes: unit entrance, elevator, and corridor. This provides a unified target benchmark and time sequence anchor for subsequent segmented review, avoiding the problems of target misalignment and duplicate judgment that occur when multiple points are identified independently. The specific implementation method is as follows: The local area communication network adopts a dual-link backup of wired Ethernet and wireless LAN, with single data transmission latency controlled within 200ms and dual-link switching time less than 50ms, ensuring the real-time performance of identification results and avoiding the lag problem where the target has moved to the next area but the verification information has not yet been delivered. The single lithium battery target identification result generated by the unit entrance sensing terminal includes not only the target category label and confidence value, but also three key related information items: high-precision acquisition timestamp, spatial coordinate information, and target appearance feature summary. The timestamp uses millisecond precision, formatted as year, month, day, hour, minute, second, and millisecond, for subsequent time alignment of multiple terminals; the spatial coordinates are two-dimensional plane coordinates based on the building's indoor coordinate system, in meters, with an accuracy of 0.1 meters, marking the target's specific location in the unit entrance area; the target appearance feature summary includes the target's dimensions, shell color, and electromagnetic feature hash value, for subsequent terminals to quickly match the corresponding target, avoiding target confusion when multiple people pass through simultaneously.

[0066] The broadcast mechanism uses targeted multicast, sending identification results only to elevator car terminals and entrance corridor terminals within the same unit. This avoids consuming the entire building's network bandwidth and prevents irrelevant terminals from receiving redundant data. Each record awaiting review is assigned a unique event ID, serving as a unique identifier for end-to-end tracking of that target. Subsequent review results from all terminals are bound to the same event ID, ensuring end-to-end data traceability. The queue for review is arranged in chronological order, with newly arriving identification results appended to the end of the queue and completed results automatically removed. The maximum queue buffer capacity is 50 records, sufficient to handle scenarios with multiple people passing through simultaneously during peak hours.

[0067] In the example, the unit entrance terminal detected a single lithium battery target at 08:22:15:320 on June 15, 2024, with a confidence level of 0.92. The spatial coordinates were (3.2, 1.5) meters inside the unit entrance gate. The target's feature summary was a rectangular silver shell, dimensions of 180mm × 70mm × 120mm, and an electromagnetic feature hash value of H_240615082215. The terminal assigned a unique ID to this event as EV_20240615_0822_0037, and then packaged all the information and broadcast it through the local area network to the elevator car terminal of unit 1 and the entrance corridor terminals on floors 1 to 6. After each terminal received the data and completed the integrity verification, it stored the record in its local cross-point pending verification queue, awaiting subsequent matching and detection.

[0068] The elevator car sensing terminal performs time-series matching and trajectory continuity verification on the data it collects based on the target characteristics in the queue to be verified, calculates the changes in the carrying state of the same target after entering the elevator, and generates the elevator segment verification result. The core of this step is to rely on the local sensing capabilities of the elevator car terminal to perform secondary identification and status verification on the targets in the verification queue, verifying whether the targets have actually entered the elevator. At the same time, it monitors the changes in the carrying status of the targets during elevator operation, eliminates false identification results at the unit entrance, improves the overall identification accuracy, and provides objective basis for risk level upgrades. The specific implementation method is as follows: After receiving the queue of items to be verified, the elevator car sensing terminal first predicts the time window for the target to arrive at the elevator based on the timestamp. The average walking time from the unit entrance to the elevator car is 8 to 15 seconds. Therefore, the terminal opens a 30-second target detection window starting from the timestamp of the record to be verified. Within this window, it focuses on matching targets with corresponding features, reducing unnecessary computation and improving detection response speed. The temporal matching uses a feature similarity weighted algorithm. The matching dimensions include visual appearance similarity, electromagnetic feature matching degree, and density feature matching degree, with weights of 0.4, 0.35, and 0.25 respectively. A weighted calculation yields the total matching score, and items with a total score greater than 0.85 are considered the same target. During the matching process, items with similar appearance but inconsistent electromagnetic and density features are automatically excluded, such as ordinary power banks and laptop power adapters, which are easily confused, reducing the probability of false matching.

[0069] Trajectory continuity verification is used to exclude non-entering targets that briefly enter and then leave. The verification logic is as follows: if a target first appears in the elevator car door area, then gradually moves into the car, and finally stays for more than 5 seconds before the elevator door closes, it is determined that the target has actually entered the elevator. If the target only stays at the door without entering the car, or enters and then leaves immediately, it is determined that it has not entered, and the record to be reviewed is marked as invalid, terminating subsequent tracking. Carrying status change detection is used to distinguish between different carrying methods such as hand-carrying, carrying on the back, and placing on a trolley. It also monitors for abnormal states such as battery disassembly, exposed or damaged casing, etc., and makes judgments based on a combination of changes in visual contour features and changes in electromagnetic intensity. In the example, the target in the queue to be reviewed was in a hand-carrying state at the entrance. After entering the elevator and placing it on the ground, the electromagnetic intensity increased by 15% due to the closer proximity to the sensor. The terminal recorded this state change and marked it as a change from hand-carrying to placing on the ground, with no exposed abnormalities.

[0070] After matching and verification are completed, the elevator car terminal generates an elevator segment verification result, which includes eight items: event ID, target matching confirmation status, elevator entry timestamp, elevator number, current floor, change in carrying status, and verification confidence level. If the match is successful and the target is confirmed to have entered the elevator, it is marked as verification passed, and the risk continues; if no matching target is found, it is marked as verification failed, entry misjudgment, and is simultaneously fed back to the unit entry terminal for iterative optimization of the recognition model. In the example, the elevator terminal matched the corresponding target at 08:22:28, confirmed entry into elevator number 1, currently located on the 1st floor, carrying status changed from hand-held to placed, verification confidence level of 0.94, generated an elevator segment verification result, and broadcast it to all associated terminals.

[0071] After receiving the elevator section verification result, the entrance corridor sensing terminal detects whether the target has entered the entrance corridor after exiting the elevator. It uses face or human body tracking tags between multiple terminals to associate the target's affiliation and generates the entrance section verification result. The core of this step is to complete the third round of identification and verification, as well as the association of personnel identity, as the target leaves the elevator and enters the floor corridor. This confirms whether the target is moving towards the direction of the resident's entrance, pinpoints the specific location of the risk target, and achieves accurate tracking of risk from public areas to before the resident enters the home. This provides direct evidence for high-risk assessment. The specific implementation method is as follows: The entrance corridor sensing terminals are distributed at the connection between the elevator lobby and the corridor on each floor, covering the entire passageway from the elevator exit to the entrance door of each resident. Each terminal covers a corridor length of 15 meters, with adjacent terminals overlapping by 2 meters to ensure no blind spots. After receiving the elevator section verification result, the terminal first predicts the target floor based on the elevator's current floor and direction of travel, and preemptively activates the full-power detection mode of the corresponding floor's corridor terminals. Terminals on non-target floors remain in low-power standby mode to reduce overall computing power consumption and equipment operating wear.

[0072] The detection logic after a target exits the elevator is as follows: When the elevator door opens, the corridor terminal continuously scans the exit area, matching the target's appearance and electromagnetic features to confirm whether the target has exited the elevator and entered the corridor area. To avoid target confusion caused by multiple people exiting the elevator simultaneously on the same floor, a human tracking identifier is used for target attribution association. The human tracking identifier generates a unique personnel feature ID based on human contour, clothing features, and gait features. This ID is initially collected and bound to the corresponding lithium battery target at the unit entrance. During verification inside the elevator, the personnel feature ID is updated synchronously. During corridor detection, the personnel ID is directly matched, achieving full-process tracking of the bound person and completely solving the matching misalignment problem when multiple targets are running in parallel.

[0073] After the target is associated with the property, the terminal continuously tracks the target's movement trajectory in the corridor and determines whether its movement direction is towards the resident's entrance door. If the target moves along the corridor towards the resident's entrance door and stays in front of a resident's door for more than 3 seconds, it is determined that there is a tendency to enter the resident's home and is marked as a high-risk pre-entry state. If the target turns in the corridor to non-entry areas such as fire exits or stairwells, or directly takes the elevator back downstairs, it is determined that there is no risk of entering the resident's home, and the risk level is downgraded accordingly.

[0074] In the example, after elevator number 1 reaches the 3rd floor, the terminal in the 3rd-floor corridor detects a target with the corresponding personnel characteristic ID exiting the elevator, carrying a lithium battery, and moving along the corridor towards the door of apartment 302 at a speed of 1.2 m / s. The target remains at the door of apartment 302 for 4 seconds, confirming a potential for trespassing. The terminal generates a verification result for the trespassing segment, binds the event ID, labels the target's floor as the 3rd floor, the target's location as in front of apartment 302, the risk tendency as high-risk trespassing, and the verification confidence level as 0.91. The result is then synchronized to all terminals and the backend system across the entire chain.

[0075] The initial identification results of the fusion unit entrance, the verification results of the elevator section, and the verification results of the entrance section are spliced ​​together according to the spatial order to generate a full-link risk assessment result that includes the target's identity, movement trajectory, and current risk status.

[0076] The core of this step is to integrate the three-stage identification and verification results, piece together the complete target movement trajectory according to the spatial progression, uniformly assess the risk status of the entire process, eliminate the limitations and occasional misjudgments of single-point identification, and output the final end-to-end risk assessment result that can be directly used for handling decisions, providing a reliable basis for subsequent hierarchical management and control. The specific implementation method is as follows: The trajectory stitching uses the event ID as a unique index, pulling three types of data from the building's backend data scheduling center: initial identification results of the unit entrance, elevator section verification results, and apartment entrance section verification results. The data is then aligned temporally and stitched together according to the spatial sequence of unit entrance, elevator car, and apartment entrance corridor. The stitching process employs spatiotemporal consistency verification to check if the arrival time difference between two adjacent points conforms to the normal walking speed range of 0.8 to 1.5 m / s. If the average speed corresponding to the time difference exceeds this range, the trajectory stitching is considered abnormal, requiring re-matching of the target to avoid incorrectly stitching trajectories of different targets. The verification process also incorporates floor change information to verify the matching degree between elevator runtime and floor differences, further improving stitching accuracy.

[0077] In the example, the unit entrance detection time is 08:22:15, the elevator entrance matching time is 08:22:28, the straight-line distance between the two points is 12 meters, and the calculated average speed is approximately 0.92 m / s, which is consistent with normal walking speed; the elevator takes 42 seconds to travel from the 1st floor to the 3rd floor, and the arrival time is 08:23:10. The corridor detects the target at 08:23:14, with an interval of 4 seconds, which is consistent with the normal time from the elevator door opening to exiting the car. All spatiotemporal consistency checks pass, confirming that the three trajectory segments belong to the same target.

[0078] After trajectory stitching is completed, a complete target movement trajectory is generated, including a time series and a corresponding spatial coordinate series, which can intuitively reconstruct the entire path of the target from entering the unit to reaching the resident's door. Subsequently, a comprehensive risk assessment is performed, with the following rules: if only the unit entrance is successfully identified, and no target is matched in the elevator or entrance segment, it is considered risk-free, and the event is directly archived; if both the unit entrance and elevator segment are successfully identified, but no match is found in the entrance segment, it is considered a general risk, and the target is being carried in the building's common areas; if all three segments are successfully identified and the target shows a tendency to enter the resident's home, it is considered a high-risk event, and the target is about to enter the resident's home.

[0079] The final end-to-end risk assessment result includes six core components: unique event ID, target identity identifier, personnel feature ID, complete movement trajectory, current spatial node, and current risk level. The result also includes the verification confidence level and feature matching degree for each segment, facilitating manual verification by backend administrators. In the example, the final end-to-end result is: Event ID EV_20240615_0822_0037, target is a single lithium battery cell in an electric vehicle, personnel ID P_3_0615_012, trajectory is unit entrance, elevator 1, 3rd floor corridor in front of door 302, current node is the entrance corridor, and risk level is high-risk. This result is directly input into the subsequent tiered response module to drive corresponding early warning and control actions.

[0080] S205, based on the full-link risk assessment results, execute the graded handling strategy, trigger voice prompts, sound and light warnings or elevator control locking actions for confirmed risk events, and push the captured data and warning information to the background management terminal to generate closed-loop management records.

[0081] Specifically, the target's trajectory and current spatial node can be extracted from the end-to-end risk assessment results. When the target only appears in the unit entrance area, it is judged as a suspected risk; when the target enters the elevator car, it is judged as a confirmed risk; and when the target enters the entrance corridor, it is judged as a high-risk risk, generating a spatial location-driven risk level classification result. The core of this step is to establish a three-level risk classification mechanism based on spatial progression. The classification is based on the enclosedness of the target's location and the degree of fire spread hazard. Continuous trajectory data is mapped to discrete risk levels, providing clear triggering conditions for subsequent graded response and avoiding misjudgments and over-intervention caused by a single identification result. The specific implementation method is as follows: First, the spatiotemporal trajectory sequence of the target is analyzed from the end-to-end risk assessment results. The sequence includes the spatial coordinates, location, identification confidence level, and timestamp of each sampling moment. The trajectory strictly follows the spatial order of unit entrance, elevator car, and entrance corridor, with the corresponding risk level gradually increasing. The hierarchical assessment is driven by spatial location as the core, while incorporating temporal continuity rules to eliminate boundary jitter. The corresponding level assessment is triggered only after the target has continuously stayed in the corresponding area for more than 3 sampling cycles, i.e., 1.5 seconds, to avoid level jumps caused by the target briefly passing through the boundary area.

[0082] The criteria for determining a suspected risk are that the target is only identified by the sensing terminal at the unit entrance and has not yet entered the enclosed area inside the building. In this case, there is a possibility that the target is merely passing by and has not brought the battery into the building, which represents the lowest risk level. In the example, the target is identified as a single lithium battery cell of an electric vehicle at the unit entrance, with a fusion confidence level of 0.82, a dwell time of 8 seconds, and no movement towards the elevator. In this case, it is determined to be a suspected risk, corresponding to a risk level code of Level 1.

[0083] The criteria for confirming risk are: after the target enters the building through the unit entrance, it is verified and identified by the sensing terminal inside the elevator car, and the target is within the enclosed space of the elevator. At this point, the lithium battery has entered a densely populated vertical transportation area. Thermal runaway would directly threaten the safety of people inside the elevator, significantly increasing the difficulty of escape, thus classifying the risk level as medium. In the example, the target enters the elevator car 25 seconds after being identified at the unit entrance, the elevator verification confidence level is 0.95, and the target is located in the center of the car. At this point, the risk is confirmed, corresponding to a risk level code of 2.

[0084] The criteria for determining a high-risk situation are: after leaving the elevator, the target enters the entrance corridor area and continues to move towards the resident's door. At this point, the lithium battery is close to the resident's indoor space, which is a highly enclosed living area. Once a fire breaks out, it will endanger the lives and property of the residents, and the rescue will be the most difficult, resulting in the highest risk level. In the example, after leaving the elevator, the target entered the 12th-floor entrance corridor, walked 15 meters along the corridor, and stopped at the door of resident 1203 for more than 10 seconds. This is considered a high-risk situation, corresponding to a risk level code of 3.

[0085] The final risk level classification result includes five core fields: target unique identifier, current risk level, spatial node, entry time, and confidence level. The field format is standardized and can be directly read and called by the subsequent handling module, serving as the sole basis for matching the classification and handling strategy.

[0086] Based on the risk level classification results, the corresponding handling actions are matched. If the risk is suspected to be triggered, the built-in speaker of the terminal plays a gentle voice prompt. If the risk is confirmed to be triggered, the sound and light alarm in the elevator car flashes continuously and plays a voice warning to leave. If the risk is high, a lock signal is sent to the elevator controller to prevent the elevator door from closing and running, and a graded handling instruction sequence is generated. The core of this step is to establish a one-to-one mapping relationship between risk levels and response actions. Following the principle of matching intervention intensity with risk level, the response is escalated step-by-step from mild warnings to mandatory control measures. This ensures effective control in high-risk scenarios while avoiding unnecessary disruption to residents in low-risk scenarios, achieving precise and tiered response. The specific implementation method is as follows: First, a mapping rule for handling actions is established. Different risk levels correspond to different intervention intensities, execution devices, and action parameters. All handling instructions are executed locally on the terminal first, without waiting for background instructions, ensuring a response latency of less than 200 milliseconds to meet real-time management requirements. Instruction execution has priority logic, with higher-level instructions automatically overriding lower-level instructions to avoid multiple action conflicts. At the same time, all action parameters can be remotely configured through the backend to adapt to the management requirements of different buildings.

[0087] Suspected risks trigger Level 1 intervention, implemented by the integrated sensing terminal at the unit entrance, using a gentle voice prompt. The voice content is pre-recorded as a standardized prompt: "For your safety and the safety of others, please do not bring electric bicycle lithium batteries indoors." The voice volume is set to 60 decibels, a level that clearly conveys the message without disturbing nearby residents. The voice message is played twice in a loop, for a total duration of approximately 12 seconds, and automatically stops after completion. If the target remains in the area, the prompt is triggered again every 30 seconds. In the example, after the unit entrance terminal identifies a suspected risk, it immediately triggers the local voice module, starting to play the prompt voice within 0.15 seconds, without requiring background intervention, achieving a millisecond-level response.

[0088] The confirmed risk corresponds to a Level 2 intervention, implemented by an integrated sensing terminal and an audible and visual alarm within the elevator car. The alarm uses a combination of flashing red LEDs and a high-volume voice warning. The LED flashing frequency is set to 2Hz (twice per second) for a strong visual warning. The voice warning reads, "Electric vehicle lithium batteries are prohibited from entering the elevator. Please take your batteries and leave the elevator immediately. Thank you for your cooperation," with a volume set to 80 decibels, ensuring clear coverage within the enclosed space of the elevator car. The warning continues until the target leaves the elevator car. If the target remains for more than 30 seconds, the warning intensity is automatically escalated, increasing the LED flashing frequency to 3Hz and the volume to 85 decibels. In the example, after the elevator car terminal confirms the risk, it simultaneously triggers flashing lights and a voice warning, allowing passengers inside the car to immediately receive the warning and proactively advise the person carrying the battery to leave.

[0089] High-risk situations require a three-level intervention, implemented by the elevator main controller and the car's audio-visual equipment. The action involves elevator control locking with enhanced warning. The locking signal is sent directly to the elevator main control board via a hard-wired interface. This signal has higher priority than the elevator door operator's closing command. Once triggered, the elevator remains open and cannot perform closing or running actions until the risk is eliminated or security personnel manually reset it. Simultaneously with locking, the audio-visual alarm maintains its highest intensity warning, continuously playing eviction warning messages and displaying warning lights. The elevator control locking logic has safety redundancy, triggering only when the elevator is level and the door is open. If the elevator is running, it waits for the next level and door opening before triggering, preventing accidents caused by locking during operation.

[0090] All handling actions are encapsulated into a time-sequential hierarchical handling instruction sequence according to the triggering order and execution logic. Each instruction in the sequence includes the execution device number, action type, parameter configuration, triggering time, and duration. Instructions take effect level by level according to risk level, and the entire process is traceable and configurable.

[0091] While executing the tiered response command sequence, the system extracts captured images and electromagnetic feature data before and after the target appears from the multi-source sensing data stream, packages the early warning information, including time, location, risk level, and response actions, into structured data, and generates an early warning data package to be pushed. The core of this step is to simultaneously capture and package evidence data upon triggering the action, preserving complete evidence for event tracing. At the same time, it converts unstructured perceived data into standardized structured data, ensuring consistency in subsequent backend transmission and storage, and providing traceable data support for closed-loop management. The specific implementation method is as follows: The capture data was extracted following the principles of front-to-back coverage and highlighting key points. Image data was captured from continuous footage spanning 5 seconds before and 10 seconds after the target's initial identification, from which three keyframes were extracted: the frame of the target's initial entry into the frame, the frame showing the clearest frontal features of the target, and the frame of the moment the action was triggered. The keyframes maintained the original 1080P resolution, clearly presenting the target's appearance, carrying method, and surrounding environment. A timestamp and device serial number watermark were automatically overlaid to ensure the evidence's immutability. Electromagnetic feature data captured the complete electromagnetic field waveform during the target's presence, extracting three core parameters: characteristic harmonic frequency, peak intensity, and duration. Density data captured the peak and average density values ​​during the target's presence. These two types of non-visual data supplement the image evidence, helping to eliminate interference from similar objects during subsequent verification and enhancing the persuasiveness of the evidence.

[0092] The early warning information employs a structured field design, comprising six core fields: The first is time information, including the event trigger timestamp and corresponding date accurate to milliseconds; the second is location information, including the building number, unit number, specific control point number, and spatial location description; the third is risk level information, indicating the currently determined risk level code and textual description, along with a comprehensive confidence score for target identification; the fourth is response action information, recording all triggered response action types, trigger times, execution status, and duration; the fifth is target characteristic information, briefly describing the target's appearance, dimensions, carrying method, and electromagnetic signature summary; and the sixth is equipment information, recording the terminal device number and operating status for data collection. All fields use a unified data format, with numeric fields retaining two decimal places and text fields using fixed-length encoding to ensure standardized data packet format and compatibility with automatic parsing by the backend management system.

[0093] During data packaging, captured images and electromagnetic and density feature data are attached as appendices, while structured early warning information is included as the main text. This is encapsulated using a standard transmission protocol, and a data integrity check code is added. This check code is generated by hashing the entire dataset and is used to verify whether data has been lost or tampered with during transmission. The final generated early warning data packet size is controlled within 2MB to adapt to the bandwidth conditions of the building's local area network, ensuring that uploads can be completed within 1 second without causing transmission delays due to excessive data volume.

[0094] The data packets to be pushed as early warnings are uploaded to the property management back-end terminal via wireless network. The back-end automatically generates a closed-loop control record containing the event number, handling time, operation record, and review status, completing the entire closed loop from risk discovery to handling feedback.

[0095] The core of this step is to synchronize the entire process data of front-end perception, identification, and handling to the back-end management terminal, forming standardized and traceable control records. At the same time, a manual review mechanism is established to achieve linkage between automatic front-end handling and manual back-end supervision, forming a complete closed loop from risk discovery, intervention and handling to record archiving, thereby improving the manageability and optimizability of the entire control system. The specific implementation method is as follows: Data upload employs a dual-link transmission mechanism, prioritizing the building's local area network (LAN) and backing up with a 4G wireless network. Under normal conditions, data packets are transmitted via the building's internal WiFi network, with a transmission latency of less than 500 milliseconds. In the event of a LAN failure, the system automatically switches to the 4G backup link, ensuring 100% delivery of alert information and preventing missed events due to network interruptions. Upon receiving the data packet, the backend management terminal first verifies data integrity using a checksum. After successful verification, it automatically parses all structured fields, extracts the corresponding image and feature data, stores it in the local database, and automatically generates a unique event number. The event number uses a "date + daily serial number" encoding rule. In the example, the 37th alert event on June 15, 2024, is numbered 202406150037. This globally unique number serves as a unique identifier for subsequent event retrieval and tracing.

[0096] The closed-loop management and control record comprises four core modules. The first module is the basic event information, which corresponds to the time, location, risk level, and target characteristics in the warning data package, fully reconstructing the basic situation of the event. The second module is the handling execution record, which automatically records all handling actions triggered by the front end, the trigger time, execution status, and end time, while also recording whether the equipment is operating normally, ensuring that the handling process is fully traceable. The third module is the manual operation record, which provides an entry point for security personnel to review the operation. Security personnel can view the captured evidence and real-time footage, mark the event as "verified and valid" after confirming that it is true, and mark it as "false alarm" if it is a false alarm and fill in the reason for the false alarm. All operations are automatically recorded with the operator and operation time. The fourth module is the review status indicator, which is divided into three categories: pending review, verified, and false alarm, intuitively presenting the progress of the event handling.

[0097] The core of the closed-loop mechanism lies in the full-process tracking of event status. When an event is generated, it is automatically marked as pending review. A pop-up window in the backend reminds security personnel to complete the review within 15 minutes. After the review is completed, the corresponding status is updated. If the review is not completed within the time limit, the reminder is automatically escalated and pushed to the property management personnel's account. For confirmed risk events, subsequent handling results and rectification records can be supplemented to form a complete event closed loop. All closed-loop control records are retained in the backend database for 90 days, supporting multi-dimensional retrieval and statistics by time, location, and risk level. This can be used to analyze high-incidence periods and locations, providing data support for optimizing prevention and control layout and adjusting control strategies. In the example, for the high-risk warning event numbered 202406150037, the backend receives the data packet 0.8 seconds after triggering, automatically generates a control record, and security personnel complete the review 8 minutes later, marking it as verified and valid, while also noting that residents have been contacted and persuaded to leave. The entire process from event discovery to archiving is traceable, forming a complete control closed loop.

[0098] Another embodiment of the present invention provides an intelligent early warning and control system for electric vehicle single-cell lithium batteries entering homes and elevators based on multimodal recognition. See [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to simultaneously acquire image data, electromagnetic field strength data and object density data through integrated sensing terminals deployed at building security points, and generate multi-source sensing data streams; The differentiation module 302 is used to perform spatial closure attribute recognition and location information joint discrimination on the multi-source sensing data stream, distinguish outdoor open areas from elevator and enclosed entrance areas, and generate scene type labels. The identification module 303 is used to initiate multimodal AI fusion identification in a closed area according to the scene type label. Based on the pre-trained electric vehicle single lithium battery exclusive identification model, it performs multi-dimensional fusion comparison of the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features, and removes interference from similar items through the differential feature library to generate single lithium battery target identification results containing target category and confidence level. The verification module 304 is used to correlate and verify the target identification results of the single lithium battery in multiple sensing terminals at the unit entrance, elevator car and entrance corridor, and generate a full-link risk assessment result. The early warning module 305 is used to execute a graded handling strategy based on the full-link risk assessment results, trigger voice prompts, sound and light warnings or elevator control locking actions for confirmed risk events, and push the captured data and early warning information to the background management terminal to generate a closed-loop management record.

[0099] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0100] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0101] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0102] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for intelligent early warning and control of single-cell lithium batteries for electric vehicles entering homes and elevators based on multimodal recognition, characterized in that, The method includes: By simultaneously collecting image data, electromagnetic field strength data, and object density data through integrated sensing terminals deployed at building security points, a multi-source sensing data stream is generated. The multi-source sensing data stream is subjected to spatial closure attribute recognition and location information joint discrimination to distinguish between outdoor open areas and elevator and entrance closed areas, and scene type labels are generated. Based on the scene type label, multimodal AI fusion recognition is initiated in the enclosed area. Based on the pre-trained electric vehicle single lithium battery exclusive recognition model, the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features are fused and compared in multiple dimensions. The interference of similar items is eliminated through the differential feature library to generate single lithium battery target recognition results containing target category and confidence level. The target identification results of the single lithium battery are correlated and verified among multiple sensing terminals at the unit entrance, elevator car and entrance corridor to generate a full-link risk assessment result. Based on the full-link risk assessment results, a tiered handling strategy is implemented. For confirmed risk-carrying events, voice prompts, sound and light warnings, or elevator control locking actions are triggered, and the captured data and warning information are pushed to the back-end management terminal to generate a closed-loop management record.

2. The method according to claim 1, characterized in that, The integrated sensing terminals deployed at building security checkpoints simultaneously collect image data, electromagnetic field strength data, and object density data to generate a multi-source sensing data stream, including: An integrated sensing terminal is installed at the entrance of the building unit, the top of the elevator car, and the side wall of the entrance corridor. The terminal integrates a high-definition camera, an electromagnetic field sensor, and a density sensor, and completes the clock synchronization calibration of each sensor to generate a cluster of sensing terminals that are deployed and synchronized. Each sensing terminal is activated to synchronously collect raw image data, electromagnetic field strength data, and object density data according to a unified sampling period. Each frame of data is automatically associated with the collection timestamp and device number to generate a multi-source raw sensing dataset. The multi-source raw sensing dataset is time-stamp aligned and spatial coordinates are normalized. Images, electromagnetic fields and density data collected at the same time are bound into a single sensing recording unit to generate spatiotemporally aligned multimodal sensing records. The spatiotemporally aligned multimodal sensing records are encapsulated according to time series, data integrity verification identifiers are added, and a continuously updated multi-source sensing data stream is generated.

3. The method according to claim 2, characterized in that, The process of identifying spatial closure attributes and jointly determining location information from the multi-source sensing data stream, distinguishing between open outdoor areas and enclosed areas such as elevators and entrances, and generating scene type labels includes: Extract the current sensing terminal's location coordinates and spatial environment features from the multi-source sensing data stream to generate a location and environment feature vector; The location and environmental feature vectors are input into a pre-trained spatial classifier. By judging whether the location coordinates are located in the elevator shaft or the area inside the entrance door, and whether the environmental features conform to the sound and light propagation law of a closed space, a spatial closure attribute probability value is generated. When the probability value of the spatial enclosure attribute is greater than the enclosure threshold, the floor change data in the joint location information and the navigation signal strength are cross-validated to exclude transition areas such as semi-open corridors or lobbies, and generate a preliminary judgment result of the scene type. The preliminary scene type determination results are processed by time-series smoothing to eliminate single-frame misjudgment and edge jitter, and finally generate scene type labels for outdoor open areas or elevator and enclosed entrance areas corresponding to the current perception cycle.

4. The method according to claim 3, characterized in that, The process involves initiating multimodal AI fusion recognition within a confined area based on the scene type label. Using a pre-trained, dedicated recognition model for individual lithium batteries in electric vehicles, the collected visual appearance features, lithium battery-specific electromagnetic parameter features, and physical density features are fused and compared across multiple dimensions. Furthermore, a differentiated feature library is used to eliminate interference from similar objects, generating a single-cell lithium battery target recognition result that includes the target category and confidence level. This includes: When the scene type label is determined to be an elevator car or a closed area at the entrance, the multimodal AI fusion recognition engine is triggered to enter the working state, extracting visual appearance features, including battery size, electrode arrangement and shell markings, from the corresponding time period of the multi-source perception data stream, and generating a visual feature vector. Simultaneously extract frequency domain features of electromagnetic field intensity data from multi-source sensing data streams, obtain characteristic harmonic components under lithium battery charging and discharging states, calculate the average density value of objects from density data, and generate a fusion vector of electromagnetic features and density features. The fused vectors of visual features, electromagnetic features, and density features are input into a pre-trained recognition model specifically for single lithium batteries in electric vehicles. The contribution weight of each modality feature to the target discrimination is calculated through a multimodal feature cross-attention mechanism, and then compared and eliminated with similar item features in the differential feature library to generate a fused recognition score. Based on the fusion recognition score and the preset multi-level confidence threshold, it is determined whether the current target is a single lithium battery cell of an electric vehicle and its carrying method, and finally a single lithium battery cell target recognition result containing target category label and confidence value is generated.

5. The method according to claim 4, characterized in that, The process of correlating and verifying the single-cell lithium battery target identification results across multiple sensing terminals at the unit entrance, elevator car, and entrance corridor to generate a full-link risk assessment result includes: The single lithium battery target identification results generated by the unit entrance sensing terminal are carried with timestamps and spatial coordinates and broadcast to the associated sensing terminals in the elevator car and entrance corridor through the local area communication network to generate a cross-point waiting queue for verification. The elevator car sensing terminal performs time-series matching and trajectory continuity verification on the data it collects based on the target characteristics in the queue to be verified, calculates the changes in the carrying state of the same target after entering the elevator, and generates the elevator segment verification result. After receiving the elevator section verification result, the entrance corridor sensing terminal detects whether the target has entered the entrance corridor after exiting the elevator. It uses face or human body tracking tags between multiple terminals to associate the target's affiliation and generates the entrance section verification result. The initial identification results of the fusion unit entrance, the verification results of the elevator section, and the verification results of the entrance section are spliced ​​together according to the spatial order to generate a full-link risk assessment result that includes the target's identity, movement trajectory, and current risk status.

6. The method according to claim 5, characterized in that, The step involves executing a tiered response strategy based on the end-to-end risk assessment results. For confirmed risk-carrying events, this includes triggering voice prompts, audible and visual warnings, or elevator control locking actions. The captured data and warning information are then pushed to the backend management terminal, generating a closed-loop management record, including: The target's trajectory and current spatial node are extracted from the end-to-end risk assessment results. When the target only appears in the unit entrance area, it is judged as a suspected risk; when the target enters the elevator car, it is judged as a confirmed risk; and when the target enters the entrance corridor, it is judged as a high-risk risk. A spatial location-driven risk level classification result is generated. Based on the risk level classification results, the corresponding handling actions are matched. If the risk is suspected to be triggered, the built-in speaker of the terminal plays a gentle voice prompt. If the risk is confirmed to be triggered, the sound and light alarm in the elevator car flashes continuously and plays a voice warning to leave. If the risk is high, a lock signal is sent to the elevator controller to prevent the elevator door from closing and running, and a graded handling instruction sequence is generated. While executing the tiered response command sequence, the system extracts captured images and electromagnetic feature data before and after the target appears from the multi-source sensing data stream, packages the early warning information, including time, location, risk level, and response actions, into structured data, and generates an early warning data package to be pushed. The data packets to be pushed as early warnings are uploaded to the property management back-end terminal via wireless network. The back-end automatically generates a closed-loop control record containing the event number, handling time, operation record, and review status, completing the entire closed loop from risk discovery to handling feedback.

7. A smart early warning and control system for single-cell lithium batteries of electric vehicles entering homes and elevators based on multimodal recognition, characterized in that, The system includes: The data acquisition module is used to simultaneously acquire image data, electromagnetic field strength data, and object density data through integrated sensing terminals deployed at building security points, and generate multi-source sensing data streams. The differentiation module is used to perform spatial closure attribute recognition and location information joint discrimination on the multi-source sensing data stream, distinguish outdoor open areas from elevator and enclosed entrance areas, and generate scene type labels. The identification module is used to initiate multimodal AI fusion identification in a closed area based on the scene type label. Based on the pre-trained electric vehicle single lithium battery exclusive identification model, it performs multi-dimensional fusion comparison of the collected visual appearance features, lithium battery exclusive electromagnetic parameter features and physical density features, and removes interference from similar items through a differentiated feature library to generate single lithium battery target identification results containing target category and confidence level. The verification module is used to correlate and verify the target identification results of the single lithium battery in multiple sensing terminals at the unit entrance, elevator car and entrance corridor, and generate a full-link risk assessment result. The early warning module is used to execute a graded handling strategy based on the full-link risk assessment results. It triggers voice prompts, sound and light warnings, or elevator control locking actions for confirmed risk-carrying events, and pushes the captured data and early warning information to the back-end management terminal to generate a closed-loop management record.

8. The system according to claim 7, characterized in that, The acquisition module is specifically used for: An integrated sensing terminal is installed at the entrance of the building unit, the top of the elevator car, and the side wall of the entrance corridor. The terminal integrates a high-definition camera, an electromagnetic field sensor, and a density sensor, and completes the clock synchronization calibration of each sensor to generate a cluster of sensing terminals that are deployed and synchronized. Each sensing terminal is activated to synchronously collect raw image data, electromagnetic field strength data, and object density data according to a unified sampling period. Each frame of data is automatically associated with the collection timestamp and device number to generate a multi-source raw sensing dataset. The multi-source raw sensing dataset is time-stamp aligned and spatial coordinates are normalized. Images, electromagnetic fields and density data collected at the same time are bound into a single sensing recording unit to generate spatiotemporally aligned multimodal sensing records. The spatiotemporally aligned multimodal sensing records are encapsulated according to time series, data integrity verification identifiers are added, and a continuously updated multi-source sensing data stream is generated.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images