A tunnel emergency state vehicle running state intelligent perception method and system

Abstract

The application relates to the technical field of tunnel emergency sensing, and discloses a tunnel emergency state vehicle running state intelligent sensing method and system, wherein the method comprises the following steps: collecting communication signals and detecting signal interruption events; extracting signal interruption mode feature vectors and classifying interruption causes; determining a communication sensing blind area boundary and inferring a trapped vehicle position; obtaining a real-time temperature field and generating an online sound field propagation model; performing spatially guided acoustic focusing collection on the communication sensing blind area; performing time-frequency analysis on the enhanced sound pressure signals and detecting candidate sound source events; performing sound source position inversion based on the online sound field propagation model and fusing position estimation; judging trapped vehicle running state labels based on sound source category labels; and generating a full-tunnel vehicle running state sensing report.

Description

Technical Field

[0001] This invention relates to the field of tunnel emergency sensing technology, specifically to a method and system for intelligent sensing of vehicle operating status under tunnel emergency conditions. Background Technology

[0002] In the event of an emergency such as a fire or explosion inside the tunnel, the high temperature heat wave or blast shock wave in the core area of ​​the accident may cause damage to the hardware of the vehicle communication module or power failure of the vehicle's electrical system, causing the vehicle to stop transmitting wireless communication signals and creating a communication and sensing blind spot.

[0003] Existing vehicle operation status perception methods based on communication signal monitoring can only mark lost vehicles as "signal lost" in the above scenario, and cannot obtain their subsequent location and operation status information. The distributed microphone array and temperature sensor deployed in the tunnel can still function normally after the accident. Although the damaged vehicle has lost its communication capability, it may still generate passive acoustic signals such as metallic thermal deformation sound, leakage fluid evaporation sound, or the cries for help from trapped personnel, which provides the conditions for sensing the status of trapped vehicles through acoustic means.

[0004] However, existing methods have the following problems: First, without knowing the precise range of the communication blind zone and the approximate location of the trapped vehicle, performing indiscriminate acoustic scanning of the entire tunnel is extremely inefficient, and the probability of detecting weak sound sources is insufficient; Second, the high temperature of the fire creates a non-uniform temperature gradient field in the tunnel, causing drastic changes in the spatial distribution of sound velocity and bending and refraction of the sound wave propagation path. The offline acoustic model based on the assumption of a uniform sound field is seriously mismatched with the actual propagation environment, resulting in a significant deterioration in the accuracy of sound source localization. Summary of the Invention

[0005] This invention provides a method and system for intelligent sensing of vehicle operating status in tunnel emergency situations, solving the technical problems in related technologies such as communication interruption leading to unknown vehicle location, temperature gradient causing bending of sound wave propagation path leading to inaccurate acoustic sensing, and inability to effectively determine the operating status of trapped vehicles in tunnel emergency scenarios.

[0006] This invention discloses an intelligent perception method for vehicle operation status in tunnel emergency situations, comprising: monitoring vehicle communication signals through distributed wireless signal receiving nodes, detecting signal interruption events, extracting signal interruption mode feature vectors for each signal interruption event, inputting the signal interruption mode feature vectors into a pre-trained interruption cause classifier, outputting interruption cause category labels, and filtering accident-related signal interruption events; Based on the last known location coordinates of the accident-related signal interruption event, the spatial boundary range of the communication sensing blind zone is determined by spatial kernel density estimation, and kinematic extrapolation is performed based on the last known location and last known velocity vector of each accident-related interrupted vehicle to generate the inferred location coordinates and location uncertainty ellipse of each trapped vehicle. The real-time temperature values ​​of each measuring point of the distributed temperature sensor are obtained, a three-dimensional temperature field distribution is generated and converted into a three-dimensional sound velocity field distribution, and an online sound field propagation model is generated based on the three-dimensional sound velocity field distribution using a ray tracing algorithm. Based on the spatial boundary range and the inferred position coordinates, spatial domain beam focusing processing is performed on the distributed microphone array, and delay compensation is performed using the curved propagation path parameters in the online sound field propagation model to extract the enhanced sound pressure signal in each focusing direction. Time-frequency analysis is performed on the enhanced sound pressure signal to detect candidate sound source events and classify them to obtain sound source category labels. Based on the sound source category labels, the operating status label of the trapped vehicle is determined.

[0007] Furthermore, the signal interruption mode feature vector is composed of the slope of the received signal strength change, the amplitude of the carrier frequency offset change, and the decreasing pattern of the number of detectable nodes in the last few time slots before each signal interruption event; Z-score normalization is performed on each feature before combination.

[0008] Furthermore, the interruption cause classifier is based on the random forest algorithm, taking the signal interruption mode feature vector as input and the interruption cause category label as output, and is trained using pre-labeled historical signal interruption samples; the interruption cause category label includes normal departure, communication module natural failure, high temperature progressive signal attenuation interruption, and explosion impact instantaneous signal interruption; the accident-related signal interruption events are signal interruption events classified as high temperature progressive signal attenuation interruption and explosion impact instantaneous signal interruption.

[0009] Furthermore, the kinematic extrapolation, combined with the interruption cause category label, constrains the motion state, wherein: for high-temperature progressive signal attenuation interruption, the velocity attenuation rate used in the kinematic extrapolation is greater than the velocity attenuation rate of the instantaneous impact signal interruption; for explosive impact instantaneous signal interruption, the distance of the kinematic extrapolation is calculated based on the last known velocity and a preset collision deceleration model; the major axis of the position uncertainty ellipse is along the longitudinal direction of the tunnel, and the minor axis is along the transverse direction of the tunnel, and the size of the position uncertainty ellipse is positively correlated with the duration elapsed after the interruption.

[0010] Furthermore, the three-dimensional temperature field distribution is converted into a three-dimensional sound velocity field distribution, including: converting the Celsius temperature of each measuring point into an absolute temperature; based on the physical relationship that the sound velocity is proportional to the square root of the absolute temperature, converting the absolute temperature of each measuring point into a corresponding sound velocity value to obtain the three-dimensional sound velocity field distribution; the sound ray tracking algorithm takes the three-dimensional sound velocity field distribution, the sound source position coordinates, and the microphone array element position coordinates as inputs, and gradually advances the sound ray trajectory according to the refraction effect of the sound velocity gradient on the sound wave ray, outputting the sound wave bending propagation path, propagation delay, and path attenuation.

[0011] Furthermore, whenever the distributed temperature sensor updates the temperature value, the generation of the three-dimensional temperature field distribution, the conversion of the three-dimensional sound velocity field distribution, and the sound ray tracing calculation are re-executed to update the propagation path parameters and time delay parameters in the online sound field propagation model; the update period of the online sound field propagation model is consistent with the sampling period of the distributed temperature sensor.

[0012] Furthermore, the spatial domain beam focusing process includes: pointing the main lobe of the beam sequentially to the coordinate directions of each inferred position, and using the position uncertainty ellipse to set the beam width coverage range; applying delay compensation and weighted summation operations to the multi-channel sound pressure signal, wherein the delay compensation uses the propagation delay corresponding to the curved propagation path at the inferred position in the online sound field propagation model as the delay compensation amount for each array element.

[0013] Furthermore, time-frequency analysis is performed on the enhanced sound pressure signal to detect candidate sound source events and classify them to obtain sound source category labels. This includes: performing continuous wavelet transform on the enhanced sound pressure signal in each focusing direction to obtain the time-frequency energy distribution; detecting candidate sound source events in the energy accumulation region of the time-frequency energy distribution whose duration exceeds the shortest duration threshold; extracting the center frequency, bandwidth, duration, and energy change pattern of each candidate sound source event as feature parameters, inputting them into a pre-trained sound source type classifier, and outputting sound source category labels; the sound source category labels include metal thermal deformation sound, fluid evaporation sound, and human voice.

[0014] Furthermore, the determination of the operating status label of the trapped vehicle based on the sound source category label includes: for vehicle-related acoustic events, using the cross-correlation time delay difference between multi-channel enhanced sound pressure signals, combined with the bending propagation path parameters in the online sound field propagation model, performing sound source location inversion, and using the inferred location coordinates as the initial value for optimization solution to generate acoustic detection location coordinates; when the deviation distance between the acoustic detection location coordinates and the inferred location coordinates is within the range of the location uncertainty ellipse, the location estimate of the trapped vehicle is updated to the acoustic detection location coordinates; the detection of human voices is marked as a state of survival, the detection of metal thermal deformation sound is marked as a state of high temperature damage, and the detection of fluid evaporation sound is marked as a state of liquid leakage risk; when multiple sound source category labels are detected at the inferred location of the same trapped vehicle, the main operating status label is determined according to the priority order of survival status, liquid leakage risk status, and high temperature damage status.

[0015] This invention provides an intelligent perception system for vehicle operation status in tunnel emergency situations, comprising: a signal interruption detection module for monitoring vehicle communication signals through distributed wireless signal receiving nodes and detecting signal interruption events; an interruption cause classification module for extracting signal interruption pattern feature vectors for each signal interruption event, inputting the signal interruption pattern feature vectors into a pre-trained interruption cause classifier to output interruption cause category labels, and filtering accident-related signal interruption events; a blind zone inference module for determining the spatial boundary range of the communication perception blind zone based on the last known position coordinates of accident-related signal interruption events, and performing kinematic extrapolation to generate the inferred position coordinates and position uncertainty ellipse of each trapped vehicle; and a sound field modeling module. The system is used to acquire real-time temperature values ​​from distributed temperature sensors, generate a three-dimensional temperature field distribution, and convert it into a three-dimensional sound velocity field distribution. Based on the three-dimensional sound velocity field distribution, an online sound field propagation model is generated using a ray tracing algorithm. The acoustic focusing acquisition module is used to perform spatial domain beam focusing processing on the distributed microphone array based on the spatial boundary range and inferred position coordinates. Delay compensation is performed using the curved propagation path parameters in the online sound field propagation model, and enhanced sound pressure signals in each focusing direction are extracted. The sound source detection and status determination module is used to perform time-frequency analysis on the enhanced sound pressure signals, detect candidate sound source events and classify them to obtain sound source category labels, and determine the operating status label of the trapped vehicle based on the sound source category labels.

[0016] This invention extracts a signal interruption pattern feature vector composed of the slope of the received signal strength change, the amplitude of the carrier frequency offset abrupt change, and the decreasing pattern of the number of detectable nodes. An interruption cause classifier then distinguishes between accident-related and non-accident-related interruptions, transforming the "signal loss" event, which is simply discarded in traditional methods, into prior information about the accident's impact range and the spatial distribution of trapped vehicles. Based on this prior information, the boundaries of the communication sensing blind zone and the inferred location coordinates of each trapped vehicle are determined. This enables acoustic sensing to perform directional acquisition in a spatial domain beam-focusing manner, rather than performing indiscriminate scanning of the entire tunnel, thereby improving the detection capability of weak sound sources.

[0017] This invention generates an online sound field propagation model using a sound ray tracing algorithm driven by real-time three-dimensional temperature field distribution. It replaces the straight-line propagation parameters under the assumption of a uniform sound field with curved propagation path parameters, correcting the influence of spatial non-uniformity of sound velocity caused by high temperature in fire on the sound wave propagation path. This allows delay compensation and sound source location inversion in spatial domain beam focusing to be performed based on path parameters that match the actual propagation environment, overcoming the problem of deteriorated positioning accuracy caused by mismatch of offline acoustic models under high temperature environments.

[0018] This invention uses the inferred location coordinates as the initial value for the sound source location inversion iteration, constrains the search range of nonlinear optimization to the possible area of ​​the trapped vehicle, reduces the probability of convergence to local extrema and reduces the number of iterations, and realizes the synergistic fusion of spatial prior information of communication perception and real-time detection capability of acoustic perception in the accident blind zone scenario, making it possible to perceive the vehicle operation status of the entire area in tunnel emergency state. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent perception method for vehicle operating status in tunnel emergency situations provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the comparison of signal interruption mode feature vectors (signal strength change slope and frequency offset abrupt change amplitude) provided in an embodiment of the present invention; Figure 3 This is a schematic diagram showing the longitudinal distribution of the temperature field and sound velocity field at a representative measuring point provided in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the major and minor axes of the ellipse representing the uncertainty of the location of a trapped vehicle, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the distribution of the duration and center frequency of candidate sound source events provided in an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the distance of the acoustic detection position and the inferred position provided in an embodiment of the present invention; Figure 7This is a schematic diagram comparing the propagation delay of a curved path with the assumed delay of a uniform sound field (from each inferred position to the nearest array) provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the number of silent time slots and signal interruption time sequence for each trapped vehicle provided in an embodiment of the present invention; Figure 9 This is a schematic diagram showing the longitudinal distribution of the last known speed and inferred position of the trapped vehicle provided in an embodiment of the present invention. Detailed Implementation

[0020] In the event of an emergency such as a fire or explosion in a tunnel, the high temperature heat wave or blast shock wave in the core area of ​​the accident can damage the hardware of the vehicle's communication module or cause a power outage in the vehicle's electrical system, preventing the vehicle from transmitting any wireless communication signals and creating a communication blind spot. Existing vehicle operation status perception methods based on communication signal monitoring can only mark lost vehicles as "signal lost" in this scenario, unable to obtain their subsequent location and status information. Meanwhile, the distributed microphone arrays and temperature sensors deployed within the tunnel can still function normally. Although damaged vehicles in the accident area have lost their communication capabilities, they may still generate passive acoustic signals such as metallic thermal deformation sounds, leaking fluid evaporation sounds, or cries for help from trapped personnel. However, without knowing the precise range of the communication blind spot and the approximate location of trapped vehicles, performing indiscriminate acoustic scanning of the entire tunnel is extremely inefficient, and the probability of detecting weak sound sources is insufficient. Furthermore, the high temperature of a fire creates a non-uniform temperature gradient field within the tunnel, causing drastic changes in the spatial distribution of sound velocity and bending and refraction of sound wave propagation paths. This results in a severe mismatch between offline acoustic models based on the assumption of a uniform sound field and the actual propagation environment, significantly degrading the accuracy of sound source localization.

[0021] In this embodiment, distributed wireless signal receiving nodes, distributed microphone arrays, and distributed temperature sensors are pre-deployed within the tunnel. The distributed wireless signal receiving nodes are spaced along the longitudinal direction of the tunnel to monitor wireless communication signals transmitted by the vehicle-mounted communication module. The distributed microphone arrays are spaced along the tunnel walls, each array containing multiple elements, to collect sound pressure signals within the tunnel. The distributed temperature sensors are deployed along the longitudinal and cross-sectional directions of the tunnel to collect temperature values ​​at various measuring points within the tunnel in real time.

[0022] At least one embodiment of the present invention discloses a method for intelligent sensing of vehicle operating status in tunnel emergency situations, such as... Figure 1 As shown, it includes the following steps: Step S100: Acquire communication signals and detect signal interruption events.

[0023] Distributed wireless signal receiving nodes within the tunnel continuously monitor vehicle-mounted communication signals, recording signal fingerprint timing data for each tracked signal source. When a signal source is not detected by any receiving node for a preset number of consecutive sampling time slots, it is marked as a signal interruption event. The signal interruption timestamp and the location coordinates of the last time it was detected are recorded to generate a list of signal interruption events.

[0024] It should be noted that the aforementioned preset number of sampling time slots refers to a continuous silent judgment window predetermined based on the normal fluctuation characteristics of communication signals within the tunnel. In one embodiment, the preset number of sampling time slots can be set to 3 to 8 sampling time slots to achieve a balance between false positive rate and response delay.

[0025] Step S200: Extract the feature vector of the signal interruption mode and classify the interruption cause.

[0026] The slope of the received signal strength change, the amplitude of the carrier frequency offset abrupt change, and the decreasing pattern of the number of detectable nodes are extracted from the last few time slots before each signal interruption event, and combined into a signal interruption mode feature vector. Since these three types of features have different dimensions, Z-score standardization is performed on each feature before combination to eliminate the impact of dimensional differences on subsequent classification operations.

[0027] The feature vectors of each signal interruption mode are input into an interruption cause classifier pre-trained using the random forest algorithm, which outputs an interruption cause category label. The random forest algorithm is trained using pre-labeled historical signal interruption samples, with the signal interruption mode feature vectors as input and the interruption cause category label as output. The interruption cause category label includes normal departure, natural failure of the communication module, high-temperature progressive signal attenuation interruption, and instantaneous signal interruption due to explosion impact. Accident-related signal interruption events classified as high-temperature progressive interruption and instantaneous impact interruption are filtered and retained.

[0028] It should be noted that in the above signal interruption mode feature vectors, the slope of the received signal strength change reflects the rate of signal attenuation. The signal strength corresponding to a gradual high-temperature interruption typically shows a continuous downward trend, while the signal strength corresponding to an instantaneous impact interruption exhibits a sudden drop. The amplitude of the carrier frequency offset abrupt change reflects whether the vehicle was subjected to a physical impact before the interruption, causing instantaneous displacement. The decreasing pattern of the number of detectable nodes reflects the shrinking pattern of signal coverage. The combination of these three factors can distinguish between signal interruptions caused by accidents and those not caused by accidents.

[0029] Step S300: Determine the boundary of the communication perception blind zone and infer the location of the trapped vehicle.

[0030] Using the last known location coordinates of the accident-related signal interruption event as sample points, spatial kernel density estimation is used to calculate the spatial probability density distribution of the accident-affected area. The probability density is set to a preset density threshold. The continuous area is defined as the spatial boundary range of the communication sensing blind zone, in which The parameters are pre-defined based on the tunnel length and the expected maximum impact range of an accident.

[0031] Kinematic extrapolation is performed based on the last known position and last known velocity vector of each accident-related interrupted vehicle. Combined with the constraints on the motion state by the interruption cause category label, the inferred position coordinates and position uncertainty ellipse of each trapped vehicle are generated.

[0032] It should be noted that the constraints on the motion state imposed by the above-mentioned interruption cause category labels refer to different assumptions about the subsequent vehicle motion corresponding to different interruption cause category labels. For gradual high-temperature interruptions, the vehicle is usually already in a deceleration or stop state during signal attenuation, resulting in a larger velocity decay rate and a shorter extrapolation distance due to kinematic extrapolation. For instantaneous impact interruptions, the vehicle may still retain a certain inertial velocity at the moment of interruption, and the kinematic extrapolation distance is calculated based on the last known velocity and a preset collision deceleration model. The major axis of the position uncertainty ellipse is along the longitudinal direction of the tunnel, and the minor axis is along the transverse direction of the tunnel. The size of the position uncertainty ellipse is positively correlated with the duration of the interruption and the uncertainty of the motion state constraints.

[0033] Step S400: Obtain the real-time temperature field and generate an online sound field propagation model.

[0034] Real-time temperature values ​​from various measuring points of a distributed temperature sensor within the tunnel are acquired, and a three-dimensional temperature field distribution within the tunnel is generated using spatial interpolation methods. Based on the physical relationship between air speed of sound and temperature, the three-dimensional temperature field distribution is converted into a three-dimensional sound speed field distribution; specifically, the air speed of sound... With absolute temperature The following conditions must be met: in, The sound velocity at each measuring point is... For reference absolute temperature The speed of sound at that time The absolute temperature corresponding to each measuring point. For absolute temperature reference, the temperature values ​​collected by the distributed temperature sensors at each measuring point are in Celsius and must be converted to absolute temperature before being substituted into the above formula. In fire scenarios, the temperature distribution within tunnels exhibits strong spatial non-uniformity; the temperature in the core area of ​​the accident can be much higher than that in the surrounding areas, resulting in a significant spatial gradient in the three-dimensional sound velocity field distribution.

[0035] Based on the three-dimensional sound velocity field distribution, a ray tracing algorithm is used to calculate the sound wave bending propagation path, propagation delay, and path attenuation from each inferred location coordinate to each microphone element within the communication sensing blind zone, generating an online sound field propagation model adapted to the current non-uniform temperature field. The ray tracing algorithm takes the three-dimensional sound velocity field distribution, sound source location coordinates, and microphone element location coordinates as inputs, and outputs the sound wave propagation path, propagation delay, and path attenuation. It gradually advances the sound ray trajectory based on the refraction effect of the sound velocity gradient on the sound wave ray, reflecting the bending characteristics of the sound wave path under a non-uniform temperature field.

[0036] In this embodiment, to enable the online sound field propagation model to adapt to the dynamic changes in the temperature field during the ongoing development of a fire, the three-dimensional temperature field distribution generation, three-dimensional sound velocity field distribution conversion, and ray tracing calculation are re-executed whenever the distributed temperature sensor updates its temperature value, updating the propagation path parameters and time delay parameters in the online sound field propagation model. The update cycle of the online sound field propagation model is consistent with the sampling cycle of the distributed temperature sensor, allowing the online sound field propagation model to continuously track changes in the actual temperature environment.

[0037] Step S500: Perform spatially guided acoustic focusing acquisition in the communication sensing blind zone.

[0038] Multi-channel sound pressure signals are synchronously acquired using a distributed microphone array. Based on the spatial boundary of the communication perception blind zone and the inferred position coordinates of each stranded vehicle, spatial domain beam focusing processing is performed on the distributed microphone array. The main lobe of the beam is sequentially pointed to the directions of each inferred position coordinate, and the beamwidth coverage range is set using the position uncertainty ellipse to extract the enhanced sound pressure signals in each focusing direction.

[0039] It should be noted that the aforementioned spatial domain beamforming processing refers to applying corresponding delay compensation and weighted summation operations to the multi-channel sound pressure signals based on the spatial geometric relationship between each microphone array element and the target direction. This enhances the coherent superposition of signals from the target direction while suppressing the incoherent superposition of interference signals from other directions. The beamwidth is determined by the spatial range of the position uncertainty ellipse; a larger position uncertainty ellipse results in a wider beamwidth to ensure coverage of areas where vehicles may be present.

[0040] In this embodiment, when performing spatial domain beam focusing processing, the curved propagation path parameters at the inferred location in the online sound field propagation model are used to replace the delay compensation parameters under the straight propagation assumption. Since the high temperature of a fire causes the actual propagation path of the sound wave to bend, using the propagation delay corresponding to the curved path as the delay compensation amount for each array element ensures that the beam focusing direction is consistent with the actual arrival direction of the sound wave, avoiding beam pointing deviation due to propagation path mismatch.

[0041] Step S600: Perform time-frequency analysis on the enhanced sound pressure signal and detect candidate sound source events.

[0042] Continuous wavelet transform time-frequency analysis is performed on the enhanced sound pressure signals from each focusing direction to obtain the time-frequency energy distribution. The duration of energy accumulation regions exceeding the minimum duration threshold is detected within the time-frequency energy distribution. Candidate sound source events, among which Pre-set based on the shortest duration characteristics of the target sound source type to eliminate transient noise interference.

[0043] The center frequency, bandwidth, duration, and energy variation pattern of each candidate sound source event are extracted as feature parameters and input into a pre-trained sound source type classifier, which outputs a sound source category label. The sound source type classifier takes the center frequency, bandwidth, duration, and energy variation pattern of candidate sound source events as input and the sound source category label as output, and is trained using pre-labeled acoustic samples. The output layer of the sound source type classifier is a fully connected layer with the number of nodes matching the number of sound source categories. It uses a softmax activation function to output the probability distribution of each category, where softmax represents a normalized exponential function. A cross-entropy loss function is used during training. The sound source category labels correspond to three categories: metallic thermal deformation sound, fluid evaporation sound, and human voice. Vehicle-related acoustic events classified into these three categories are retained.

[0044] It should be noted that the sound of metal thermal deformation is usually manifested as short, high-frequency pulses, the sound of fluid evaporation is manifested as continuous broadband noise in the mid-frequency range, and the human voice is manifested as a narrow-band signal with a fundamental frequency and harmonic structure. The three types of sound sources have distinguishable characteristics in terms of center frequency, bandwidth, duration, and energy change patterns.

[0045] Step S700: Perform sound source location inversion and fuse location estimates based on the online sound field propagation model.

[0046] For vehicle-related acoustic events, the cross-correlation time delay difference between multi-channel enhanced sound pressure signals is utilized, combined with the curved propagation path parameters at the inferred location in the online sound field propagation model, to perform sound source location inversion. The inferred location coordinates are used as initial values ​​for optimization to generate the acoustic detection location coordinates.

[0047] It should be noted that the above optimization solution using the inferred location coordinates as the initial value for iteration refers to the fact that the sound source location inversion problem is a nonlinear optimization problem under a non-uniform sound velocity field, requiring an initial solution for iterative search. The inferred location coordinates already contain the approximate location information of the trapped vehicle. Using the inferred location coordinates as the initial value for iteration can limit the search range to the vicinity of the actual location, reduce the number of iterations, and lower the probability of convergence to a local extremum.

[0048] Calculate the deviation distance between the acoustic detection location coordinates and the inferred location coordinates. When the deviation distance is within the range of the location uncertainty ellipse, associate the acoustic detection location coordinates with the inferred location coordinates and update the position estimate of the stranded vehicle to the acoustic detection location coordinates.

[0049] In this embodiment, when the deviation between the acoustic detection location coordinates and the inferred location coordinates exceeds the range of the position uncertainty ellipse, it is determined that the acoustic event may originate from a non-tracked target. At this time, the acoustic event is marked as an unrelated acoustic event, and the acoustic detection location coordinates and sound source category label of the unrelated acoustic event are retained as candidate records for potential unknown trapped targets and included in subsequent outputs.

[0050] Step S800: Determine the operating status label of the trapped vehicle based on the sound source category label.

[0051] For vehicles identified as trapped acoustically, their operational status is determined based on sound source category labels. Detecting human voices indicates the presence of surviving personnel; detecting metallic thermal deformation sounds indicates high-temperature damage; and detecting fluid evaporation sounds indicates a risk of liquid leakage.

[0052] When multiple sound source category tags are detected at the inferred location of the same trapped vehicle, the main operating status tag is determined in the following priority order: first, the status of personnel survival; second, the status of liquid leakage risk; and third, the status of high temperature damage. All detected sound source category tags are added as auxiliary information.

[0053] For stranded vehicles where no acoustic event was detected at the inferred location, they are marked as communication interruption pending confirmation, and the inferred location coordinates and interruption reason category label information are retained.

[0054] Step S900: Generate a report on the perception of vehicle operation status throughout the tunnel.

[0055] The boundaries of the communication sensing blind spots, the acoustic detection location coordinates of each trapped vehicle, the operating status tags, and the sensing source markers are summarized and arranged according to the tunnel mileage.

[0056] In this embodiment, to support the priority allocation of emergency rescue resources, a rescue urgency level is generated for each stranded vehicle based on its operational status tag. The highest urgency level corresponds to the status of having personnel alive, the second highest urgency level corresponds to the status of liquid leakage risk, the medium urgency level corresponds to the status of high temperature damage, and the basic urgency level corresponds to the status of communication interruption pending confirmation.

[0057] The system combines the operational status information of vehicles trapped within the communication blind zone with the real-time location and speed information of vehicles outside the communication blind zone that are still communicating normally, and generates a vehicle operational status perception report that covers the entire tunnel area.

[0058] This implementation extracts a signal interruption pattern feature vector from signal interruption events, consisting of the slope of the received signal strength change, the amplitude of the carrier frequency offset abrupt change, and the decreasing pattern of the number of detectable nodes. An interruption cause classifier then distinguishes between accident-related and non-accident-related interruptions, transforming the "signal loss" event, which is simply discarded in traditional methods, into prior information about the accident's impact range and the spatial distribution of trapped vehicles. Based on this prior information, the boundaries of the communication sensing blind zone and the inferred location coordinates of each trapped vehicle are determined. This allows acoustic sensing to perform directional acquisition in a spatial domain beam-focusing manner, rather than performing indiscriminate scanning of the entire tunnel, thereby improving the detection capability of weak sound sources. Simultaneously, this implementation uses a real-time three-dimensional temperature field distribution-driven ray-tracing algorithm to generate an online sound field propagation model. By replacing the straight-line propagation parameters under the assumption of a uniform sound field with curved propagation path parameters, it corrects the influence of the spatial non-uniformity of sound velocity caused by high temperatures during fires on the bending of the sound wave propagation path. This ensures that delay compensation and sound source location inversion in spatial domain beam-focusing processing are performed based on path parameters that match the actual propagation environment, overcoming the problem of degraded positioning accuracy due to mismatch in offline acoustic models under high-temperature environments. Furthermore, by using the inferred location coordinates as the initial value for the sound source location inversion iteration, the search range of nonlinear optimization is constrained to the possible area of ​​the trapped vehicle, which reduces the probability of convergence to local extrema and reduces the number of iterations. This enables the synergistic fusion of the spatial prior guidance capability of communication perception and the real-time detection capability of acoustic perception in the accident blind zone scenario, making it possible to perceive the vehicle operation status across the entire area in tunnel emergency situations.

[0059] A 3800-meter-long, two-way, four-lane highway tunnel (TUN-07) experienced a fuel fire caused by a rear-end collision at 14:23 on a certain day in 20XX, in the section between kilometer markers K1+640 and K1+820. The tunnel was pre-positioned with 19 wireless signal receiving nodes (N01 to N19) every 200 meters longitudinally, and a microphone array (8 elements per array) every 100 meters along the tunnel walls. Temperature sensors were installed at points every 50 meters longitudinally and at the top, bottom, left, and right sides of the cross-section. Before the accident, the system was normally tracking the communication signals of 14 vehicles traveling within the tunnel.

[0060] like Figures 2-9 As shown, it includes the following steps: The communication signal receiving node continuously records the signal fingerprint timing data of each vehicle. Taking a sampling time slot of 2 seconds and a silence judgment window of 5 time slots as an example, the system successively detected multiple signal sources remaining silent for more than 5 consecutive time slots between 14:23 and 14:25, triggering signal interruption event markers. The table below summarizes the original records of the signal interruption events identified in this incident.

[0061] Table 1 Original Records of Signal Interruption Events: For the four interruption events mentioned above, the system extracts the slope of the received signal strength change, the amplitude of the carrier frequency offset change, and the decreasing pattern of the number of detectable nodes in the last five time slots before the interruption. These are combined into a signal interruption mode feature vector, which is then input into a random forest classifier after Z-score standardization.

[0062] Table 2 Feature vectors and classification results of signal interruption modes: All four vehicles were classified as accident-related interruptions (progressive high-temperature interruptions or instantaneous impact interruptions) and entered the subsequent processing procedure. VH-07 and VH-11 correspond to progressive high-temperature attenuation, with small absolute values ​​of signal strength slope and low frequency offset abrupt changes, consistent with the attenuation characteristics of communication modules gradually failing in high-temperature environments. VH-03 and VH-09 experienced a sharp drop in signal strength and significant frequency offset abrupt changes, consistent with the characteristics of instantaneous damage to onboard equipment caused by an explosion impact.

[0063] Spatial kernel density estimation was performed using the last known location coordinates of the four vehicles as sample points. The resulting probability density exceeded a preset density threshold. The continuous area was identified as a communication sensing blind zone, with a corresponding mileage range of K1+635 to K1+825 and a length of approximately 190 meters.

[0064] When performing kinematic extrapolation for each vehicle, a larger velocity decay rate was used for vehicles with gradual high-temperature interruption (VH-07, VH-11) to infer that they had stopped within a short distance after the interruption; for vehicles with instantaneous impact interruption (VH-03, VH-09), the extrapolation distance was calculated based on the last known velocity and the collision deceleration model. The table below gives the inferred position coordinates and position uncertainty ellipse parameters for each trapped vehicle.

[0065] Table 3 Ellipse of Estimated Location and Uncertainty of Trapped Vehicles: The major axis of the uncertainty ellipse for vehicles that are interrupted by impact (VH-03, VH-09) is greater than that for vehicles that are interrupted by temperature asymptotic impact, because the estimation uncertainty of the inertial motion distance of the vehicle after impact is higher.

[0066] The distributed temperature sensors updated the temperature values ​​at each measuring point at 14:25. The temperature in the core accident area, from K1+640 to K1+820, reached a maximum of 623 degrees Celsius, while the temperature in the surrounding areas gradually decreased to the ambient temperature of approximately 28 degrees Celsius with increasing distance.

[0067] After generating a three-dimensional temperature field through spatial interpolation, the formula is used. The temperature field is converted into a three-dimensional sound velocity field, where m / s is the reference speed of sound K is the reference temperature. For the absolute temperature of the measuring point, To correspond to the speed of sound, taking the 623-degree Celsius measuring point in the core area of ​​the accident as an example, the absolute temperature is converted. K corresponds to approximately the speed of sound. The speed of sound is m / s, which is much higher than the approximately 347 m / s at the tunnel entrance at 28 degrees Celsius. The table below shows the temperature and sound velocity comparison at representative measuring points.

[0068] Table 4 Comparison of temperature field and sound velocity field at representative measuring points: The sound speed peaks near K1+740, forming a significant gradient with the surrounding regions. Based on this three-dimensional sound speed field, the ray tracing algorithm calculates the curved propagation path, propagation delay, and path attenuation from each inferred position to each microphone element, generating an online sound field propagation model. Taking the inferred position VH-07 (K1+718) to a certain element of the nearest microphone array (located at K1+700) as an example, the curved path propagation delay is approximately 0.041 seconds, while the straight-line propagation delay under the uniform sound field assumption (calculated based on an ambient sound speed of 347 m / s at 28 degrees Celsius) is approximately 0.052 seconds, a difference of 21%. This indicates that using the offline uniform sound field model would introduce significant delay compensation errors.

[0069] Based on the communication sensing blind zone (K1+635 to K1+825) and the inferred positions of four vehicles, spatial domain beam focusing processing is performed on the microphone array within this area. The main lobe of the beam points sequentially to the inferred positions VH-03, VH-07, VH-11, and VH-09. The delay compensation for each focusing direction uses the propagation delay of the corresponding curved path in the online sound field propagation model, rather than the value assumed for straight-line propagation. When the inferred position VH-07 is taken as the focusing direction, the delay compensation for each of the eight array elements is set according to the curved path delay, and the weighted sum is used to output the enhanced sound pressure signal in that direction. The enhanced sound pressure signals in each focusing direction are then entered into the time-frequency analysis process.

[0070] Perform continuous wavelet transform on the enhanced sound pressure signals in each focusing direction, and detect durations exceeding the shortest duration threshold in the time-frequency energy distribution. The energy concentration region (set to 0.3 seconds) is used to extract the center frequency, bandwidth, duration, and energy variation pattern of candidate sound source events, which are then input into the sound source type classifier. The table below shows the candidate sound source events detected in each focusing direction and their classification results.

[0071] Table 5. Characteristics and classification results of candidate sound source events: The duration of the first candidate event in the VH-03 direction was 0.12 seconds, which did not reach the target. A narrowband continuous signal with a center frequency of 180Hz and a bandwidth of 60Hz was detected in the VH-07 direction. It has fundamental frequency and harmonic structure characteristics and is classified as human voice, indicating that there may be people alive in the vehicle.

[0072] For the aforementioned vehicle-related acoustic events, the sound source location was inverted using the cross-correlation time delay difference of multi-channel enhanced sound pressure signals combined with an online sound field propagation model. The inferred location coordinates were used as the initial values ​​for optimization. The table below shows the deviation and correlation results between the acoustic detection location coordinates and the inferred location coordinates.

[0073] Table 6. Deviation between acoustic detection location and inferred location, and fusion results: The deviations between the acoustic detection positions and the inferred positions of all four vehicles were within the range of the major axis of their respective uncertainty ellipses, and the correlation was successfully established, updating the position estimates to the acoustic detection positions.

[0074] Based on the sound source category labels detected by each vehicle, the operational status label is determined according to the priority rule of prioritizing the presence of surviving personnel, and a rescue urgency level is generated. VH-07 detected human voices and was directly marked as a surviving personnel status, corresponding to the highest urgency level. VH-03 and VH-09 detected fluid evaporation sounds and were marked as a liquid leakage risk status, corresponding to the second highest urgency level. VH-11 detected metallic thermal deformation sounds and was marked as a high-temperature damage status, corresponding to a medium urgency level.

[0075] The system merges the boundary of the communication perception blind zone (K1+635 to K1+825), the acoustic detection location of the four trapped vehicles, their operating status tags, and the urgency level of the rescue, with the real-time location and speed information of the ten vehicles that are still communicating normally outside the blind zone, to generate a vehicle operating status perception report covering the entire TUN-07 tunnel.

[0076] The logical chain of the entire data flow is as follows: the signal fingerprint time-series data collected by the communication signal receiving node triggers an interruption event (step S100); the three-dimensional feature vector of the interruption event is used by a classifier to distinguish 4 accident-related vehicles (step S200); the last known position coordinates of the 4 vehicles are used to determine the blind zone boundary through kernel density estimation and generate the inferred position through kinematic extrapolation (step S300); the real-time data from the temperature sensor drives the sound ray tracking algorithm to generate curved path parameters (step S400); the curved path parameters replace the straight propagation assumption for beam focusing delay compensation (step S500); the enhanced sound pressure signal is used for time-frequency analysis and a classifier to identify vehicle-related acoustic events (step S600); the cross-correlation time delay difference is combined with the curved path parameters to invert the acoustic detection position and associate and fuse it with the inferred position (step S700); the sound source category label is converted into an operating status label and urgency level (step S800); and finally, a full tunnel perception report is summarized and output (step S900). The inferred location generated during the communication sensing phase is integrated throughout the entire process of acoustic sensing beam focusing, location inversion, and correlation fusion. The data from the two sensing methods are fused at the spatial level in step S700, enabling continuous perception of the location and status of the trapped vehicle even under conditions of complete communication interruption.

[0077] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligently sensing the running state of a vehicle in a tunnel emergency state, characterized in that, include: By monitoring vehicle communication signals through distributed wireless signal receiving nodes, signal interruption events are detected, signal interruption pattern feature vectors of each signal interruption event are extracted, and the signal interruption pattern feature vectors are input into a pre-trained interruption cause classifier to output interruption cause category labels and filter accident-related signal interruption events. Based on the last known location coordinates of the accident-related signal interruption event, the spatial boundary range of the communication sensing blind zone is determined by spatial kernel density estimation, and kinematic extrapolation is performed based on the last known location and last known velocity vector of each accident-related interrupted vehicle to generate the inferred location coordinates and location uncertainty ellipse of each trapped vehicle. The real-time temperature values ​​of each measuring point of the distributed temperature sensor are obtained, a three-dimensional temperature field distribution is generated and converted into a three-dimensional sound velocity field distribution, and an online sound field propagation model is generated based on the three-dimensional sound velocity field distribution using a ray tracing algorithm. Based on the spatial boundary range and the inferred position coordinates, spatial domain beam focusing processing is performed on the distributed microphone array, and delay compensation is performed using the curved propagation path parameters in the online sound field propagation model to extract the enhanced sound pressure signal in each focusing direction. Time-frequency analysis is performed on the enhanced sound pressure signal to detect candidate sound source events and classify them to obtain sound source category labels. Based on the sound source category labels, the operating status label of the trapped vehicle is determined. 2.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, The signal interruption mode feature vector is composed of the slope of the received signal strength change, the amplitude of the carrier frequency offset change, and the decreasing pattern of the number of detectable nodes in the last few time slots before each signal interruption event; Z-score normalization is performed on each feature before combination. 3.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, The interruption cause classifier is based on the random forest algorithm, taking the signal interruption mode feature vector as input and the interruption cause category label as output, and is trained using pre-labeled historical signal interruption samples. The interruption cause category label includes normal departure, communication module natural failure, high temperature progressive signal attenuation interruption, and explosion impact instantaneous signal interruption. Accident-related signal interruption events are signal interruption events classified as high temperature progressive signal attenuation interruption and explosion impact instantaneous signal interruption. 4.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, The kinematic extrapolation, combined with the interruption cause category label, constrains the motion state, wherein: for high-temperature progressive signal attenuation interruption, the velocity attenuation rate used in the kinematic extrapolation is greater than the velocity attenuation rate of the instantaneous impact signal interruption; for explosive impact instantaneous signal interruption, the distance of the kinematic extrapolation is calculated based on the last known velocity and the preset collision deceleration model; the major axis of the position uncertainty ellipse is along the longitudinal direction of the tunnel, and the minor axis is along the transverse direction of the tunnel, and the size of the position uncertainty ellipse is positively correlated with the elapsed time after the interruption. 5.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, The process of converting a three-dimensional temperature field distribution into a three-dimensional sound velocity field distribution includes: converting the Celsius temperature at each measuring point into an absolute temperature; based on the physical relationship that the speed of sound is proportional to the square root of the absolute temperature, converting the absolute temperature at each measuring point into a corresponding speed of sound value to obtain the three-dimensional sound velocity field distribution; the sound ray tracing algorithm takes the three-dimensional sound velocity field distribution, the coordinates of the sound source position, and the coordinates of the microphone array element position as input, and gradually advances the sound ray trajectory according to the refraction effect of the sound velocity gradient on the sound wave ray, outputting the sound wave bending propagation path, propagation delay, and path attenuation. 6.The tunnel emergency state vehicle running state intelligent sensing method according to claim 5, characterized in that, Whenever the distributed temperature sensor updates the temperature value, the generation of the three-dimensional temperature field distribution, the conversion of the three-dimensional sound velocity field distribution, and the sound ray tracing calculation are re-executed to update the propagation path parameters and time delay parameters in the online sound field propagation model; the update period of the online sound field propagation model is consistent with the sampling period of the distributed temperature sensor. 7.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, The spatial domain beam focusing process includes: pointing the main lobe of the beam sequentially to the coordinate directions of each inferred position, and using the position uncertainty ellipse to set the beam width coverage range; applying delay compensation and weighted summation operations to the multi-channel sound pressure signal, wherein the delay compensation uses the propagation delay corresponding to the curved propagation path at the inferred position in the online sound field propagation model as the delay compensation amount for each array element. 8.The tunnel emergency state vehicle running state intelligent sensing method according to claim 1, characterized in that, Time-frequency analysis is performed on the enhanced sound pressure signal to detect candidate sound source events and classify them to obtain sound source category labels. This includes: performing continuous wavelet transform on the enhanced sound pressure signal in each focusing direction to obtain the time-frequency energy distribution; detecting candidate sound source events whose duration exceeds the shortest duration threshold in the energy accumulation region of the time-frequency energy distribution; extracting the center frequency, bandwidth, duration, and energy change pattern of each candidate sound source event as feature parameters, inputting them into a pre-trained sound source type classifier, and outputting sound source category labels; the sound source category labels include metal thermal deformation sound, fluid evaporation sound, and human voice. 9.The tunnel emergency state vehicle running state intelligent sensing method according to claim 8, characterized in that, The method for determining the operating status label of a trapped vehicle based on sound source category labels includes: for vehicle-related acoustic events, using the cross-correlation time delay difference between multi-channel enhanced sound pressure signals, combined with the bending propagation path parameters in the online sound field propagation model, performing sound source location inversion, and using the inferred location coordinates as the initial value for optimization solution to generate acoustic detection location coordinates; when the deviation distance between the acoustic detection location coordinates and the inferred location coordinates is within the range of the location uncertainty ellipse, the location estimate of the trapped vehicle is updated to the acoustic detection location coordinates; the detection of human voices is marked as a state of survival, the detection of metal thermal deformation sound is marked as a state of high temperature damage, and the detection of fluid evaporation sound is marked as a state of liquid leakage risk; when multiple sound source category labels are detected at the inferred location of the same trapped vehicle, the main operating status label is determined according to the priority order of survival status, liquid leakage risk status, and high temperature damage status.

10. A vehicle operation status intelligent sensing system under tunnel emergency conditions, used to execute the vehicle operation status intelligent sensing method under tunnel emergency conditions as described in any one of claims 1 to 9, characterized in that, include: The signal interruption detection module is used to monitor vehicle communication signals through distributed wireless signal receiving nodes and detect signal interruption events. The interruption cause classification module is used to extract the signal interruption mode feature vector of each signal interruption event, input the signal interruption mode feature vector into the pre-trained interruption cause classifier to output the interruption cause category label, and filter the accident-related signal interruption events; the blind zone inference module is used to determine the spatial boundary range of the communication perception blind zone based on the last known position coordinates of the accident-related signal interruption events, and perform kinematic extrapolation to generate the inferred position coordinates and position uncertainty ellipse of each trapped vehicle. The sound field modeling module is used to acquire real-time temperature values ​​from distributed temperature sensors, generate a three-dimensional temperature field distribution, and convert it into a three-dimensional sound velocity field distribution. Based on the three-dimensional sound velocity field distribution, an online sound field propagation model is generated using a ray tracing algorithm. The acoustic focusing acquisition module is used to perform spatial domain beam focusing processing on the distributed microphone array based on the spatial boundary range and inferred position coordinates. Delay compensation is performed using the curved propagation path parameters in the online sound field propagation model, and enhanced sound pressure signals in each focusing direction are extracted. The sound source detection and status determination module is used to perform time-frequency analysis on the enhanced sound pressure signals, detect candidate sound source events, classify and obtain sound source category labels, and determine the operating status label of the trapped vehicle based on the sound source category labels.

Images