A parking lot emergency warning method and system based on multi-point edge computing

Abstract

The present invention discloses a method and system for early warning of parking lot emergencies based on multi-point edge computing. First, multi-modal environmental data and video streams are collected in different areas of the parking lot to obtain the collected data. Then, the collected data is time-synchronized and dynamically threshold-filtered to obtain the pre-processed data. Next, the spatio-temporal feature alignment and noise filtering are performed on the pre-processed data to obtain a three-dimensional environmental situation map. The present invention realizes the functions of real-time monitoring, analysis, early warning and response to emergencies such as fires, gas leaks and equipment failures in the parking lot. Moreover, through distributed data processing and collaborative analysis, the safety of the parking lot can be significantly improved and the emergency response strategy can be optimized. At the same time, through multi-modal data fusion and dynamic threshold collaborative analysis, the false alarm rate of emergencies can be reduced, not only shortening the average response time, but also significantly improving the early warning accuracy and timeliness, and is suitable for being widely promoted and used.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent security, and particularly relates to a method and system for early warning of parking lot emergencies based on multi-point edge computing. Background Art

[0002] With the acceleration of the urbanization process, parking lots, as important public facilities, especially in commercial areas, transportation hubs, residential communities, etc., undertake important functions of traffic guidance and storage. However, parking lots, as enclosed or semi-enclosed spaces, often face a series of safety hazards, such as: fires, gas leaks, equipment failures, and environmental anomaly emergencies.

[0003] Existing parking lot security monitoring systems generally adopt a centralized monitoring architecture, resulting in a high dependence on a central server for data processing and analysis. Moreover, this centralized architecture faces performance bottlenecks. Since the parking lot security monitoring system relies on the central server to process global data, the data collected by sensors needs to be transmitted through the network to a remote server, which not only leads to a high response delay but also makes the average response time often exceed 2 seconds in case of emergencies such as fires, making it difficult to meet the real-time security requirements of parking lots; at the same time, the centralized system also has the risk of single-point failure. Once the central server crashes, the entire monitoring system will fail, resulting in the inability to guarantee the security of the parking lot; the data analysis ability of existing parking lot monitoring systems is insufficient, especially in static threshold detection; most existing parking lot monitoring systems use fixed thresholds to judge anomalies, but the environment of parking lots is affected by factors such as day-night temperature difference and vehicle exhaust, resulting in a false alarm rate as high as 18% - 25%;

[0004] Currently, existing parking lot security monitoring systems generally lack the ability of spatio-temporal fusion analysis of multi-source heterogeneous data, such as environmental parameters, videos, and equipment status, resulting in the inability to comprehensively grasp the environmental situation of parking lots, and the emergency response strategy is single and cannot achieve the collaborative optimization of hierarchical early warning and dynamic path planning, resulting in easy lag in the response to emergencies. At the same time, the utilization efficiency of edge computing resources is low and it cannot support multi-model collaborative reasoning and incremental learning, severely restricting flexibility and adaptability; therefore, it is necessary to design a method and system for early warning of parking lot emergencies based on multi-point edge computing. Summary of the Invention

[0005] The object of the present invention is to overcome the deficiencies of the prior art and, in order to better and effectively solve the problems that the existing parking lot safety monitoring system generally lacks the spatio-temporal fusion analysis ability for multi-source heterogeneous data, such as environmental parameters, videos, and equipment status, resulting in the inability to comprehensively grasp the environmental situation of the parking lot, and the single emergency response strategy and the inability to achieve the collaborative optimization of hierarchical early warning and dynamic path planning, thus leading to the lag in the response to emergencies, and at the same time, the low utilization efficiency of edge computing resources and the inability to support multi-model collaborative reasoning and incremental learning, which seriously limit the flexibility and adaptability. The present invention provides a method and system for early warning of parking lot emergencies based on multi-point edge computing, which realizes the functions of real-time monitoring, analysis, early warning, and response to fire, gas leakage, and equipment failure emergencies in the parking lot, and can significantly improve the safety of the parking lot and optimize the emergency response strategy through distributed data processing and collaborative analysis. At the same time, through the collaborative analysis of multi-modal data fusion and dynamic thresholds, the false alarm rate of emergencies can be reduced, not only shortening the average response time, but also significantly improving the early warning accuracy and timeliness.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for early warning of parking lot emergencies based on multi-point edge computing, comprising the following steps:

[0008] Step A: Collect multi-modal environmental data and video streams in different areas of the parking lot and obtain the collected data.

[0009] Step B: Perform time synchronization and dynamic threshold filtering on the collected data and obtain the preprocessed data.

[0010] Step C: Align the spatio-temporal features of the preprocessed data and filter out the noise to obtain a three-dimensional environmental situation map.

[0011] Step D: Determine the type of emergency and the early warning level by using multi-model collaborative analysis according to the three-dimensional environmental situation map.

[0012] Step E: Trigger a differentiated emergency response strategy according to the early warning level to complete the early warning operation of parking lot emergencies.

[0013] Step F: Compress the updated model issued by using the knowledge distillation method and retain the key weights when the model is updated.

[0014] In the aforementioned method for early warning of parking lot emergencies based on multi-point edge computing, in Step A, multi-modal environmental data and video streams are collected in different areas of the parking lot and the collected data is obtained. Specifically, the collection of multi-modal environmental data and video streams is to collect the corresponding data by using a composite temperature and humidity sensor, a gas concentration sensor, a vibration perception sensor, and a video collector.

[0015] The composite temperature and humidity sensor uses a PT100 platinum resistor and a capacitive humidity sensor to measure and collect temperature and humidity respectively;

[0016] The gas concentration sensor uses an electro-chemical sensor and an infrared absorption sensor to detect carbon monoxide concentration and hydrogen concentration respectively. The gas concentration sensor adopts a dual-threshold trigger mechanism combining a carbon monoxide concentration threshold and a hydrogen concentration threshold;

[0017] The vibration sensing sensor uses a triaxial accelerometer to detect vehicle collisions and abnormal equipment vibrations. The vibration sensing sensor can distinguish normal vehicle driving and structural vibration events by combining frequency domain analysis FFT;

[0018] The video collector adopts a multi-camera collaborative architecture combining a main camera and a secondary camera and is equipped with a video acquisition trigger mechanism. Among them, the main camera is a wide-angle lens, and the secondary camera is an infrared thermal imaging camera. The video acquisition trigger mechanism is specifically to switch the video collector to the 60fps frame rate mode when the environmental sensor detects an abnormality.

[0019] The aforementioned method for warning parking lot emergencies based on multi-point edge computing, step B, performs time synchronization and dynamic threshold filtering on the collected data and obtains preprocessed data. The specific process of time synchronization and dynamic threshold filtering is to process the collected data using a spatio-temporal alignment module, a noise filtering module, and a dynamic threshold calculation module;

[0020] The spatio-temporal alignment module is used to add a unified time stamp to the data of the composite temperature and humidity sensor, the gas concentration sensor, the vibration sensing sensor, and the video collector using a GPS module and the NTP protocol, and use the interpolation method to compensate for the data time sequence deviation caused by transmission delay;

[0021] The noise filtering module is used to denoise the temperature data using wavelet transform and suppress the environmental instantaneous fluctuation interference of the gas concentration data using Kalman filter;

[0022] The dynamic threshold calculation module is used to statistically calculate the mean value μ and standard deviation σ of the environmental parameters based on a sliding window, and then calculate the dynamically adjusted abnormal determination threshold μ±3σ, while superimposing a day-night cycle correction coefficient.

[0023] The aforementioned method for warning parking lot emergencies based on multi-point edge computing, step C, performs spatio-temporal feature alignment and noise filtering on the preprocessed data and obtains a three-dimensional environmental situation map. The specific steps are as follows

[0024] Step C1, add a time stamp to the preprocessed data and synchronize it to a unified time axis. Specifically, the IEEE 1588 time protocol is used to achieve a synchronization accuracy of ±1ms in the local area network, as shown in formula (1),

[0025] (1)

[0026] Among them, is the standard timestamp after time and space calibration, is the original acquisition timestamp, is the time offset calculated by using the master-slave clock exchange message, d is the physical transmission distance, and c is the signal propagation speed;

[0027] Step C2, adopt Kalman filtering to eliminate sensor noise. Specifically, perform fifth-order filtering iteration on the preprocessed data, as shown in formula (2),

[0028] (2)

[0029] Among them, and are the state estimation vectors at time k and time k-1 respectively, is the Kalman gain, is the state transition matrix, is the actual observation value at the kth time, is the observation matrix;

[0030] Step C3, generate a three-dimensional environmental parameter distribution model based on the spatial interpolation algorithm. Among them, the spatial variogram in the spatial interpolation algorithm adopts the spherical model, as shown in formula (3),

[0031] (3)

[0032] Among them, h is the spatial lag distance, is the nugget effect value, C is the sill value, and a is the range;

[0033] Step C4, superimpose the video stream data and the three-dimensional model to generate a three-dimensional environmental situation map. Among them, superimposing the video stream data and the three-dimensional model is specifically to perform coordinate mapping by using affine transformation.

[0034] For the aforementioned method for warning of parking lot emergencies based on multi-point edge computing, step D, determine the type and warning level of emergencies by using multi-model collaborative analysis according to the three-dimensional environmental situation map. Specifically, use the fire recognition model, gas leakage classification model, and equipment failure detection model to determine the type and warning level of emergencies through the multi-model voting mechanism. Among them, the multi-model voting mechanism is specifically to set a confidence threshold and set weights for the fire recognition model, gas leakage classification model, and equipment failure detection model.

[0035] The aforementioned method for early warning of parking lot emergencies based on multi-point edge computing. The fire recognition model in step D specifically embeds the end of the Backbone of the YOLOv5 model into the CBAM attention module and sets the channel compression ratio to 16:1;

[0036] The input features of the gas leakage classification model in step D include gas concentration gradient features, humidity change rate features, and vibration spectrum features;

[0037] The gas concentration gradient feature is used to judge the leakage diffusion trend, and the calculation process of the gas concentration gradient feature is shown in formula (4),

[0038] (4)

[0039] where, is the gas concentration gradient, is the gas concentration at the current moment, is the historical concentration value 10 seconds ago;

[0040] The calculation process of the humidity change rate feature is shown in formula (5),

[0041] (5)

[0042] where, is the humidity change rate, is the current relative humidity value, is the time interval, is the current humidity value, is the humidity value 60 seconds ago;

[0043] The vibration spectrum feature is used to distinguish leakage from normal vibration by using the ratio of high-frequency energy to low-frequency energy, and the calculation process of the vibration spectrum feature is shown in formula (6),

[0044] (6)

[0045] where, is the vibration energy ratio, P(f) is the power spectral density of the vibration signal. The numerator integration interval 20Hz - 50Hz is the high-frequency feature band of gas leakage, and the denominator integration interval 5Hz - 15Hz is the low-frequency feature band of equipment mechanical vibration.

[0046] The above-mentioned parking lot emergency warning method based on multi-point edge computing, step E, triggers a differentiated emergency response strategy according to the warning level to complete the parking lot emergency warning operation, which specifically includes a local response module, a global evacuation module, and a response strategy selection module. The local response module is used to control the start and stop of the sprinkler equipment and the ventilation equipment. The global evacuation module is used to dynamically adjust the LED guiding screen and the in-vehicle navigation information. The response strategy selection module is used to allocate the response action priority according to the warning level and real-time data.

[0047] The above-mentioned parking lot emergency warning method based on multi-point edge computing, step F, uses the knowledge distillation method to compress the downloaded updated model and retain the key weights when the model is updated. The specific steps are as follows:

[0048] Step F1, use the knowledge distillation method to compress the downloaded updated model. The input of the knowledge distillation method is the updated model package, and the output of the knowledge distillation method is the student model adapted to the hardware resources. The dynamic distillation loss function of the knowledge distillation method is shown in formula (7).

[0049] (7)

[0050] Among them, is the total objective function, is the cross-entropy loss, y is the true label, is the data of the input updated model package, is the teacher model, is the knowledge distillation loss;

[0051] Step F2, retain the key weights when the model is updated. Specifically, use the elastic weight screening method to calculate the diagonal elements of the Fisher information matrix of the old model parameters and generate a weight importance mask based on the median. The key weights adopt regularization to constrain the change range during the optimization process, and the key weights are shown in formula (8).

[0052] (8)

[0053] Among them, is the median in the dataset F;

[0054] Step F3, construct the trigger condition and update process of the updated model. The specific steps are as follows:

[0055] Step F31, construct the trigger condition for updating the model. The trigger condition for updating the model is determined by the change in the accuracy of the validation set. Specifically, when the accuracy of the new model on the local validation set drops by more than the set probability compared to the current model, the update process is initiated.

[0056] Step F32, construct the update process for updating the model. The update process for updating the model uses the block update method to update in stages according to the priority order of the fully connected layer, the deep convolutional layer, and the shallow convolutional layer, and combines quantization and weight sparsification to compress the memory occupancy of the model. The quantization formula is as shown in formula (9).

[0057] (9)

[0058] Where, is the quantized weight, is to convert the floating-point result to the closest integer, is the original weight matrix, is the original weight mean, is the original weight standard deviation.

[0059] A parking lot emergency warning system based on multi-point edge computing, including a data collection unit, a data preprocessing unit, a multi-model collaborative analysis unit, an emergency response unit, and a model update unit. The data collection unit is used to collect multi-modal environmental data and video streams in different areas of the parking lot and obtain the collected data; the data preprocessing unit is used to perform time synchronization and dynamic threshold filtering on the collected data and obtain the preprocessed data; the data fusion unit is used to perform spatio-temporal feature alignment and noise filtering on the preprocessed data and obtain a three-dimensional environmental situation map; the multi-model collaborative analysis unit is used to determine the type of emergency and the warning level by using multi-model collaborative analysis according to the three-dimensional environmental situation map; the emergency response unit is used to trigger a differential emergency response strategy according to the warning level and complete the parking lot emergency warning operation; the model update unit is used to compress the downloaded updated model by using the knowledge distillation method and retain the key weights when the model is updated.

[0060] For the aforementioned parking lot emergency warning system based on multi-point edge computing, the data collection unit, the data preprocessing unit, and the model update unit are set in distributed edge nodes, the multi-model collaborative analysis unit and the emergency response unit are set in the central server, and the distributed edge nodes and the central server are connected by 5G-V2X communication.

[0061] The beneficial effects of the present invention are:

[0062] (1). A method and system for warning of parking lot emergencies based on multi-point edge computing. First, multi-modal environmental data and video streams are collected in different areas of the parking lot to obtain the collected data. Then, the collected data is subjected to time synchronization and dynamic threshold filtering to obtain preprocessed data. Next, the preprocessed data is subjected to spatio-temporal feature alignment and noise filtering to obtain a three-dimensional environmental situation map. Then, based on the three-dimensional environmental situation map, multi-model collaborative analysis is used to determine the type of emergency and the warning level. Subsequently, a differentiated emergency response strategy is triggered according to the warning level to complete the warning operation of parking lot emergencies. Finally, the knowledge distillation method is used to compress the updated model issued and retain the key weights during model update; effectively realizing that the method and system for warning of parking lot emergencies have the function of constructing an adaptive anomaly detection system by using a dynamic threshold calculation mechanism and spatio-temporal correlation analysis of environmental parameters. It not only overcomes the contradiction between sensitivity and false judgment of traditional fixed threshold systems in complex parking lot environments, but also significantly improves the ability to capture hidden safety hazards. It also realizes the function of forming a cross-verification type emergency judgment logic by combining visual features, environmental parameters and equipment status according to multi-model collaborative analysis and voting decision-making mechanism, greatly improving the robustness of warning decision-making in complex scenarios. It also realizes the function of establishing a hierarchical data processing pipeline and ensuring the integrity of key information under the condition of local preprocessing and feature extraction at the edge node. It not only enables refined data transmission, but also constructs a low-latency and highly reliable emergency response channel. And the model optimization architecture based on knowledge distillation can break through the computing power limit of edge devices while maintaining the analysis accuracy and realize the seamless deployment of lightweight AI models on heterogeneous nodes, thus forming a flexibly scalable intelligent perception network, and further supporting the plug-and-play and collaborative networking of edge nodes in parking lots of different scales. This can achieve the adaptive matching of the scale of the parking lot emergency warning system and the scene requirements through a dynamic resource scheduling mechanism.

[0063] (2). Through the collaborative analysis of multi-modal data fusion and dynamic thresholds, the present invention can reduce the false alarm rate of emergencies, not only shortening the average response time, but also significantly improving the warning accuracy and timeliness. At the same time, it also realizes the adaptive continuous learning of edge nodes and reduces the maintenance cost of the system.

[0064] (3) The present invention utilizes multiple sensor data and intelligent algorithms for real-time monitoring, analysis, and early warning, and combines a hierarchical early warning architecture with edge-cloud collaboration to achieve rapid and accurate disposal of complex emergencies in the parking lot. This multi-node data fusion, dynamic model compression, and conflict resolution mechanism greatly improves the safety and management efficiency of the parking lot, reduces the impact of emergencies on property and personnel, and the present invention realizes the functions of real-time monitoring, analysis, early warning, and response to fires, gas leaks, and equipment failures in the parking lot. Moreover, through distributed data processing and collaborative analysis, it can significantly improve the safety of the parking lot and optimize the emergency response strategy. Description of the Drawings

[0065] Figure 1 is the overall flowchart of a method for early warning of emergencies in a parking lot based on multi-point edge computing according to the present invention;

[0066] Figure 2 is a schematic diagram of the data acquisition principle of the data acquisition unit of the present invention;

[0067] Figure 3 is a schematic diagram of the operation principle of the central server of the present invention. Specific Embodiments

[0068] The present invention will be further described below in conjunction with the drawings of the specification.

[0069] As Figure 1 shown, a method for early warning of emergencies in a parking lot based on multi-point edge computing according to the present invention includes the following steps

[0070] As Figure 2 shown, in step A, multi-modal environmental data and video streams are collected in different areas of the parking lot to obtain the collected data. Specifically, the multi-modal environmental data and video streams are collected by using a composite temperature and humidity sensor, a gas concentration sensor, a vibration sensing sensor, and a video collector to collect the corresponding data;

[0071] The composite temperature and humidity sensor uses a PT100 platinum resistance and a capacitive humidity sensor to measure and collect temperature and humidity respectively;

[0072] Among them, the temperature measurement range of the PT100 platinum resistance is -20°C to 120°C, the capacitive humidity sensor uses the HS1101LF model and is combined with an RC oscillation circuit to measure in the range of 0-100%RH, and the temperature compensation response time of the capacitive humidity sensor is less than 200ms.

[0073] The gas concentration sensor uses an electro-chemical sensor and an infrared absorption sensor to detect carbon monoxide concentration and hydrogen concentration respectively, and the gas concentration sensor adopts a dual-threshold trigger mechanism combining a carbon monoxide concentration threshold and a hydrogen concentration threshold;

[0074] Among them, the electro-chemical sensor uses a TGS5042 type CO sensor. The detection range of the TGS5042 type CO sensor is 0 - 1000 ppm, and it is set to trigger the local audible and visual alarm for 30 ppm continuously for 10 seconds. When the concentration reaches 200 ppm, the cloud emergency response is directly activated; the infrared absorption sensor uses an IR7000 type sensor, and the IR7000 type sensor sets 4% LEL as 20% of the lower explosion limit safety threshold. If triggered, the exhaust system is immediately started;

[0075] The vibration sensing sensor uses a three-axis accelerometer to detect vehicle collisions and abnormal vibrations of the equipment, and the vibration sensing sensor can distinguish normal vehicle driving and structural vibration events by combining frequency domain analysis FFT;

[0076] Among them, the vibration sensing sensor uses an ADXL357 three-axis MEMS accelerometer. The measurement range of the ADXL357 three-axis MEMS accelerometer is set to ±16g, and the signal-to-noise ratio can reach at least 80 dB at a sampling frequency of 100 Hz; the frequency domain analysis FFT performs a 1024-point FFT operation and demarcates 5 Hz - 15 Hz as the vehicle driving characteristic frequency band, and 20 Hz - 50 Hz as the structural abnormality determination frequency band.

[0077] The video collector adopts a multi-camera cooperative architecture combining a main camera and a secondary camera and is equipped with a video collection trigger mechanism. Among them, the main camera is a wide-angle lens, and the secondary camera is an infrared thermal imaging camera. The video collection trigger mechanism is specifically to switch the video collector to the 60 fps frame rate mode when the environmental sensor detects an abnormality.

[0078] Among them, the main camera uses an IMX307 sensor with a 120° wide-angle lens. The main camera outputs a H.264 encoded video stream of 1080P@30fps when the ambient illumination is greater than 10 lux; the secondary camera uses a FLIR Lepton3.5 component. The secondary camera is automatically activated when the ambient illumination is lower than 5 lux or the temperature rise rate of the monitoring area exceeds 2℃ / s. The thermal sensitivity of the secondary camera is 50 mK level; the intelligent trigger mechanism activates the high frame rate mode with the abnormal signal of the environmental sensor. The high frame rate mode is specifically to automatically adjust the video resolution to 720P to maintain a 60 fps acquisition rate. Motion detection uses a three-frame difference algorithm and sets a 15% pixel change threshold, and the key frame extraction interval is less than 200 ms. The video clips of abnormal events are intercepted according to 5 seconds before and after, and the volume of a single clip is compressed to less than 8 MB after H.265 encoding.

[0079] Step B: Perform time synchronization and dynamic threshold filtering on the collected data to obtain preprocessed data. The specific process of time synchronization and dynamic threshold filtering is to process the collected data using a spatio-temporal alignment module, a noise filtering module, and a dynamic threshold calculation module;

[0080] The spatio-temporal alignment module is used to add a unified timestamp to the data of the composite temperature and humidity sensor, the gas concentration sensor, the vibration perception sensor, and the video collector using a GPS module and the NTP protocol, and use the interpolation method to compensate for the data timing deviation caused by transmission delay;

[0081] Among them, the spatio-temporal alignment module uses the GPS second pulse to trigger the timestamp writing and controls the NTP synchronization accuracy within the range of ±10 ms. For the maximum transmission delay of 500 ms, the cubic spline interpolation algorithm is used for timing compensation and a data buffer with 120 sampling points is set;

[0082] The noise filtering module is used to denoise the temperature data using wavelet transform and suppress the environmental instantaneous fluctuation interference of the gas concentration data using Kalman filtering;

[0083] Among them, the noise filtering module performs five-layer decomposition processing of the db4 wavelet on the temperature data and retains the approximation coefficient a5 for reconstruction. At the same time, the gas concentration data applies Kalman filtering with parameters Q = 0.01 and R = 0.1, and the initial covariance matrix is set to P0 = 1;

[0084] The dynamic threshold calculation module is used to statistically calculate the mean μ and standard deviation σ of the environmental parameters based on a sliding window, and then calculate the dynamically adjusted anomaly determination threshold μ ± 3σ, while superimposing the day-night cycle correction coefficient.

[0085] Among them, the dynamic threshold calculation module uses a 30-minute sliding window of 1800 sampling points and calculates the mean μ and standard deviation σ of the environmental parameters in real time based on the Welford algorithm. After superimposing the day-night correction coefficient, κ = 1.0 is taken from 06:00 to 18:00 during the day, and κ = 1.2 is adjusted from 18:00 to 06:00 at night. Anomaly determination requires that 5 consecutive sampling points exceed the range of μ ± 3σ, thus achieving a false alarm suppression rate of 99.7%.

[0086] Step C: Perform spatio-temporal feature alignment and noise filtering on the preprocessed data to obtain a three-dimensional environmental situation map. The specific steps are as follows

[0087] Step C1: Add a timestamp to the preprocessed data and synchronize it to a unified time axis. Specifically, the IEEE 1588 time protocol is used to achieve a synchronization accuracy of ±1 ms within the local area network, as shown in formula (1),

[0088] (1)

[0089] Among them, is the standard timestamp after spatio-temporal calibration, is the original acquisition timestamp, is the time offset calculated by using the master-slave clock exchange message, d is the physical transmission distance, and c is the signal propagation speed;

[0090] Specifically, add a 0.1s timestamp resolution to the composite temperature and humidity sensor data, and use a 33ms timestamp interval for the video collector data.

[0091] Step C2, adopt Kalman filtering to eliminate sensor noise. Specifically, perform a fifth-order filtering iteration on the preprocessed data, as shown in formula (2),

[0092] (2)

[0093] Among them, and are the state estimation vectors at time k and time k-1 respectively, is the Kalman gain, is the state transition matrix, is the actual observation value at the kth time, is the observation matrix;

[0094] Step C3, generate a three-dimensional environmental parameter distribution model based on the spatial interpolation algorithm. Among them, the spherical model is adopted for the spatial variogram in the spatial interpolation algorithm, as shown in formula (3),

[0095] (3)

[0096] Among them, h is the spatial lag distance, is the nugget effect value, C is the sill value, and a is the range;

[0097] Step C4, superimpose the video stream data on the three-dimensional model to generate a three-dimensional environmental situation map. Among them, superimposing the video stream data on the three-dimensional model is specifically to perform coordinate mapping by using affine transformation;

[0098] Among them, the parameters of the affine transformation matrix are .

[0099] Such as Figure 3As shown in the figure, in step D, the multi-model collaborative analysis is adopted according to the three-dimensional environmental situation map to determine the type of emergency and the early warning level. Specifically, the fire recognition model, the gas leakage classification model, and the equipment failure detection model are used to determine the type of emergency and the early warning level through the multi-model voting mechanism. The multi-model voting mechanism specifically sets the confidence threshold and assigns weights to the fire recognition model, the gas leakage classification model, and the equipment failure detection model. The fire recognition model specifically embeds the end of the Backbone of the YOLOv5 model into the CBAM attention module and sets the channel compression ratio to 16:1. The input features of the gas leakage classification model include the gas concentration gradient feature, the humidity change rate feature, and the vibration spectrum feature.

[0100] Among them, the multi-model voting mechanism sets the confidence threshold such that when the confidence levels output by two models both exceed 75%, it is determined that the early warning level is raised by one level. The weight distribution scheme is that the fire recognition model accounts for 45%, the gas leakage classification model accounts for 35%, and the equipment failure detection model accounts for 20%.

[0101] The gas concentration gradient feature is used to judge the leakage diffusion trend, and the calculation process of the gas concentration gradient feature is shown in formula (4).

[0102] (4)

[0103] Where is the gas concentration gradient, is the gas concentration at the current moment, is the historical concentration value 10 seconds ago;

[0104] The calculation process of the humidity change rate feature is shown in formula (5).

[0105] (5)

[0106] Where is the humidity change rate, is the current relative humidity value, is the time interval, is the current humidity value, is the humidity value 60 seconds ago;

[0107] The vibration spectrum feature is used to distinguish leakage from normal vibration by using the ratio of high-frequency energy to low-frequency energy, and the calculation process of the vibration spectrum feature is shown in formula (6).

[0108] (6)

[0109] Where Let ER be the vibration energy ratio, and P(f) be the power spectral density of the vibration signal. The integration interval of the numerator, 20 Hz - 50 Hz, is the high-frequency characteristic frequency band of gas leakage, and the integration interval of the denominator, 5 Hz - 15 Hz, is the low-frequency characteristic frequency band of equipment mechanical vibration;

[0110] When the set threshold ER ≥ 3.0, it is determined as a leakage.

[0111] Step E: Trigger a differentiated emergency response strategy according to the warning level to complete the early warning operation for parking lot emergencies, which specifically includes a local response module, a global evacuation module, and a response strategy selection module. The local response module is used to control the start and stop of the sprinkler equipment and the ventilation equipment. The global evacuation module is used to dynamically adjust the LED guiding screen and in-vehicle navigation information. The response strategy selection module is used to allocate the priority of response actions according to the warning level and real-time data;

[0112] Among them, the control parameter setting of the local response module includes that the sprinkler system starts with a 3 s delay after being triggered, the water pressure is maintained at 0.35 ± 0.05 MPa, and the wind speed of the ventilation equipment is adjusted in three levels. The wind speed increases by one level for every 50 ppm increase in the CO concentration, and the maximum wind speed is 12 m / s;

[0113] The specific priority order is: fire event > gas leakage > equipment failure.

[0114] Step F: Use the knowledge distillation method to compress the downloaded updated model and retain the key weights during model update. The specific steps are as follows,

[0115] Step F1: Compress the downloaded updated model using the knowledge distillation method. The input of the knowledge distillation method is the updated model package, and the output of the knowledge distillation method is the student model adapted to the hardware resources , and the dynamic distillation loss function of the knowledge distillation method is shown in formula (7),

[0116] (7)

[0117] Among them, is the total objective function, is the cross-entropy loss, y is the true label, is the data of the downloaded updated model package input, is the teacher model, is the knowledge distillation loss;

[0118] Step F2: Retain the key weights during model update. Specifically, the diagonal elements of the Fisher information matrix of the old model parameters are calculated using the elastic weight screening method and a weight importance mask is generated based on the median. The key weights adopt regularization to constrain the change amplitude during the optimization process, and the key weights As shown in formula (8),

[0119] (8)

[0120] wherein, is the median in the dataset F;

[0121] Step F3, construct the trigger condition and update process of the updated model. The specific steps are as follows.

[0122] Step F31, construct the trigger condition of the updated model. The trigger condition of the updated model is determined by the change in the accuracy of the validation set. Specifically, if the accuracy of the new model on the local validation set drops by more than the set probability compared to the current model, the update process is initiated.

[0123] Step F32, construct the update process of the updated model. The update process of the updated model uses the block update method to update in stages according to the priority order of the fully connected layer, deep convolutional layer, and shallow convolutional layer, and combines quantization and weight sparsification to compress the memory occupancy of the model. The quantization formula is as shown in formula (9).

[0124] (9)

[0125] wherein, is the quantized weight, is to convert the floating-point result to the closest integer, is the original weight matrix, is the original weight mean, is the original weight standard deviation.

[0126] A parking lot emergency warning system based on multi-point edge computing includes a data acquisition unit, a data preprocessing unit, a multi-model collaborative analysis unit, an emergency response unit, and a model update unit. The data acquisition unit is used to collect multi-modal environmental data and video streams in different areas of the parking lot and obtain the collected data; the data preprocessing unit is used to perform time synchronization and dynamic threshold filtering on the collected data and obtain the preprocessed data; the data fusion unit is used to perform spatio-temporal feature alignment and noise filtering on the preprocessed data and obtain a three-dimensional environmental situation map; the multi-model collaborative analysis unit is used to determine the type and warning level of the emergency event by multi-model collaborative analysis according to the three-dimensional environmental situation map; the emergency response unit is used to trigger a differentiated emergency response strategy according to the warning level and complete the parking lot emergency warning operation; the model update unit is used to compress the downloaded updated model by the knowledge distillation method and retain the key weights during model update.

[0127] Specifically, the data acquisition unit, the data preprocessing unit, and the model update unit are set in the distributed edge node, the multi-model collaborative analysis unit and the emergency response unit are set in the central server, and the distributed edge node and the central server are communicatively connected by 5G-V2X.

[0128] In summary, for the parking lot emergency warning method and system based on multi-point edge computing of the present invention, first, multi-modal environmental data and video streams are collected in different areas of the parking lot to obtain the collected data. Then, the collected data is time-synchronized and dynamically threshold-filtered to obtain the preprocessed data. Next, the spatio-temporal features of the preprocessed data are aligned and noise-filtered to obtain a three-dimensional environmental situation map. Then, based on the three-dimensional environmental situation map, multi-model collaborative analysis is used to determine the type of emergency and the warning level. Subsequently, according to the warning level, a differentiated emergency response strategy is triggered to complete the parking lot emergency warning operation. Finally, the knowledge distillation method is used to compress the updated model sent down and retain the key weights during model update; effectively realizing that the parking lot emergency warning method and system have the function of constructing an adaptive anomaly detection system by using a dynamic threshold calculation mechanism and spatio-temporal correlation analysis of environmental parameters. It not only overcomes the contradiction between the sensitivity and false judgment of traditional fixed-threshold systems in complex parking lot environments, but also significantly improves the ability to capture hidden safety hazards. It also realizes the function of forming a cross-verification type emergency judgment logic by combining visual features, environmental parameters, and equipment status according to multi-model collaborative analysis and voting decision-making mechanism, greatly improving the robustness of warning decision-making in complex scenarios. It also realizes the function of establishing a hierarchical data processing pipeline and ensuring the integrity of key information under the condition of local preprocessing and feature extraction at the edge node. It not only enables refined data transmission, but also constructs a low-latency and highly reliable emergency response channel. And the model optimization architecture based on knowledge distillation can break through the computing power limitation of edge devices while maintaining the analysis accuracy and realize the seamless deployment of lightweight AI models on heterogeneous nodes, thus forming a elastically scalable intelligent perception network, and further supporting the plug-and-play and collaborative networking of edge nodes in parking lots of different scales. This can achieve the adaptive matching of the scale of the parking lot emergency warning system and the scene requirements through a dynamic resource scheduling mechanism; the present invention can reduce the false alarm rate of emergencies through multi-modal data fusion and dynamic threshold collaborative analysis, not only shortening the average response time, but also significantly improving the warning accuracy and timeliness. At the same time, it also realizes the adaptive continuous learning of edge nodes and reduces the system maintenance cost; the present invention uses a variety of sensor data and intelligent algorithms for real-time monitoring, analysis and warning, and combines the hierarchical warning architecture of edge-cloud collaboration to achieve the rapid and accurate disposal of complex emergencies in the parking lot. This greatly improves the safety and management efficiency of the parking lot through multi-node data fusion, dynamic model compression and conflict resolution mechanism, reducing the impact of emergencies on property and personnel. The present invention realizes the function of being able to monitor, analyze, warn and respond to emergencies such as fires, gas leaks and equipment failures in the parking lot in real time, and can significantly improve the safety of the parking lot and optimize the emergency response strategy through distributed data processing and collaborative analysis.

[0129] The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of parking lot emergencies based on multi-point edge computing, characterized in that: Including the following steps, Step A: Collect multi-modal environmental data and video streams in different areas of the parking lot and obtain the collected data. Specifically, collecting multi-modal environmental data and video streams is to use a composite temperature and humidity sensor, a gas concentration sensor, a vibration sensing sensor, and a video collector to collect corresponding data; The composite temperature and humidity sensor uses a PT100 platinum resistance and a capacitive humidity sensor to measure and collect temperature and humidity respectively; The gas concentration sensor uses an electro-chemical sensor and an infrared absorption sensor to detect carbon monoxide concentration and hydrogen concentration respectively. The gas concentration sensor uses a dual-threshold trigger mechanism combining a carbon monoxide concentration threshold and a hydrogen concentration threshold; The vibration sensing sensor uses a three-axis accelerometer to detect vehicle collisions and abnormal vibrations of equipment. The vibration sensing sensor can combine frequency domain analysis FFT to distinguish normal vehicle driving from structural vibration events; The video collector uses a multi-camera collaborative architecture combining a main camera and an auxiliary camera and is equipped with a video collection trigger mechanism. Among them, the main camera is a wide-angle lens, and the auxiliary camera is an infrared thermal imaging camera. The video collection trigger mechanism is specifically to switch the video collector to the 60fps frame rate mode when the environmental sensor detects an abnormality; Step B: Perform time synchronization and dynamic threshold filtering on the collected data and obtain preprocessed data. The specific process of time synchronization and dynamic threshold filtering is to use a spatio-temporal alignment module, a noise filtering module, and a dynamic threshold calculation module to process the collected data; The spatio-temporal alignment module is used to add a unified timestamp to the composite temperature and humidity sensor data, gas concentration sensor data, vibration sensing sensor data, and video collector data using a GPS module and the NTP protocol, and use interpolation to compensate for the data timing deviation caused by transmission delay; The noise filtering module is used to denoise the temperature data using wavelet transform and suppress environmental instantaneous fluctuation interference for the gas concentration data using Kalman filtering; The dynamic threshold calculation module is used to calculate the mean value μ and standard deviation σ of environmental parameters based on sliding window statistics, and then calculate the dynamically adjusted abnormal determination threshold μ±3σ, while superimposing a day-night cycle correction coefficient; Step C: Perform spatio-temporal feature alignment and noise filtering on the preprocessed data and obtain a three-dimensional environmental situation map. The specific steps are as follows Step C1: Add a timestamp to the preprocessed data and synchronize it to a unified time axis. Specifically, use the IEEE 1588 time protocol to achieve a synchronization accuracy of ±1ms in the local area network, as shown in formula (1); Among them, ts ync is the standard timestamp after time and space calibration, t raw is the original acquisition timestamp, Δt offset is the time offset calculated by using the master-slave clock exchange message, d is the physical transmission distance, and c is the signal propagation speed; Step C2: Use Kalman filtering to eliminate sensor noise. Specifically, perform a fifth-order filtering iteration on the preprocessed data, as shown in formula (2); wherein, and are the state estimation vectors at time k and time k-1 respectively, F k is the Kalman gain, K k is the state transition matrix, z k is the actual observation value at the k-th time, H k is the observation matrix; Step C3: Generate a three-dimensional environmental parameter distribution model based on a spatial interpolation algorithm. Among them, the spatial variogram in the spatial interpolation algorithm uses a spherical model, as shown in formula (3); where h is the spatial lag distance, C0 is the nugget effect value, C is the sill value, and a is the range; Step C4: Superimpose the video stream data on the 3D model to generate a 3D environmental situation map. Specifically, the superimposition of the video stream data on the 3D model is performed by using affine transformation for coordinate mapping; Step D: Use multi-model collaborative analysis based on the 3D environmental situation map to determine the type of emergency and the warning level. Specifically, a fire recognition model, a gas leakage classification model, and an equipment failure detection model are used to determine the type of emergency and the warning level through a multi-model voting mechanism. The multi-model voting mechanism specifically sets a confidence threshold and assigns weights to the fire recognition model, the gas leakage classification model, and the equipment failure detection model; The fire recognition model in Step D specifically embeds the end of the Backbone of the YOLOv5 model into the CBAM attention module and sets the channel compression ratio to 16:1; The input features of the gas leakage classification model in Step D include gas concentration gradient features, humidity change rate features, and vibration spectrum features; The gas concentration gradient feature is used to judge the leakage diffusion trend, and the calculation process of the gas concentration gradient feature is shown in Formula (4); Among them, is the gas concentration gradient, C t is the gas concentration at the current moment, C t-10s is the historical concentration value 10 seconds ago; The calculation process of the humidity change rate feature is shown in Formula (5); Among them, is the humidity change rate, ΔRH is the current relative humidity value, Δt is the time interval, and RH t is the current humidity value, and RH t-60s is the humidity value 60 seconds ago; The vibration spectrum feature is used to distinguish leakage from normal vibration by using the ratio of high-frequency energy to low-frequency energy, and the calculation process of the vibration spectrum feature is shown in Formula (6); Among them, ER is the vibration energy ratio, P(f) is the power spectral density of the vibration signal. The molecular integration interval of 20Hz - 50Hz is the high-frequency feature band of gas leakage, and the denominator integration interval of 5Hz - 15Hz is the low-frequency feature band of equipment mechanical vibration; Step E: Trigger a differentiated emergency response strategy according to the warning level to complete the warning operation for parking lot emergencies. It specifically includes a local response module, a global evacuation module, and a response strategy selection module. The local response module is used to control the start and stop of sprinkler equipment and ventilation equipment. The global evacuation module is used to dynamically adjust the LED guidance screen and in-vehicle navigation information. The response strategy selection module is used to allocate the priority of response actions according to the warning level and real-time data; Step F: Use the knowledge distillation method to compress the downloaded updated model and retain the key weights during model update. The specific steps are as follows: Step F1, compress the downloaded updated model using the knowledge distillation method. The input of the knowledge distillation method is the updated model package, and the output of the knowledge distillation method is the student model f adapted to the hardware resources student , and the dynamic distillation loss function of the knowledge distillation method is shown in formula (7). Among them, is the total objective function, is the cross-entropy loss, y is the true label, x is the input updated model package data, and f teacher is the teacher model, is the knowledge distillation loss; Step F2, retain the key weights during model update, specifically, calculate the diagonal elements F of the Fisher information matrix of the old model parameters using the Elastic Weight Consolidation method i and generate a weight importance mask based on the median, the key weight F i adopt regularization to constrain the change amplitude during the optimization process, and the key weight F i as shown in formula (8) F i ≥ median(F) (8) Among them, median(F) is the median in the dataset F; Step F3: Construct the trigger condition and update process of the updated model. The specific steps are as follows: Step F31: Construct the trigger condition of the updated model. The trigger condition of the updated model is determined by the change in the accuracy of the validation set. Specifically, if the accuracy of the new model on the local validation set drops by more than the set probability compared to the current model, the update process is started; Step F32: Construct the update process of the updated model. The update process of the updated model uses the block update method to update in stages according to the priority order of the fully connected layer, the deep convolutional layer, and the shallow convolutional layer, and combines quantization and weight sparsification to compress the memory occupancy of the model. The quantization formula is shown in Formula (9); Among them, W quant is the quantized weight, round is to convert the floating-point result to the closest integer, W is the original weight matrix, μ W is the original weight mean, σ W is the original weight standard deviation; The specific application system of the parking lot emergency event warning method based on multi-point edge computing includes a data acquisition unit, a data preprocessing unit, a multi-model collaborative analysis unit, an emergency response unit, and a model update unit. The data acquisition unit is used to collect multi-modal environmental data and video streams in different areas of the parking lot and obtain the collected data; The data preprocessing unit is used to perform time synchronization and dynamic threshold filtering on the collected data and obtain the preprocessed data; The data fusion unit is used to perform spatio-temporal feature alignment and noise filtering on the preprocessed data and obtain a three-dimensional environmental situation map; The multi-model collaborative analysis unit is used to determine the type of emergency event and the warning level by using multi-model collaborative analysis according to the three-dimensional environmental situation map; The emergency response unit is used to trigger a differentiated emergency response strategy according to the warning level and complete the parking lot emergency event warning operation; The model update unit is used to compress the updated model sent down by using the knowledge distillation method and retain the key weights during model update; The data acquisition unit, the data preprocessing unit, and the model update unit are set in the distributed edge nodes, the multi-model collaborative analysis unit and the emergency response unit are set in the central server, and the distributed edge nodes and the central server are connected by 5G-V2X communication.

Images