Mobile life detection method, system and device based on transfer learning mechanism and UWB radar

Abstract

The invention discloses a mobile life detection method, system and device based on a transfer learning mechanism and a UWB (Ultra Wideband) radar, and relates to the technical field of radar signal processing and emergency rescue. The method comprises the following steps: generating a large amount of diversified training data with labels through high-fidelity physical simulation, and pre-training a deep neural network model by using the data; when the mobile platform enters a real disaster site, the method adopts a transfer learning technology, and can quickly and finely adjust the pre-training model to adapt to a current special environment only by collecting a very small amount of data of an inlife area on site or performing simple calibration. In this way, high-precision mobile life detection can be achieved, and powerful support is provided for emergency rescue work.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and emergency rescue technology, specifically to a method, system, and device for mobile life detection based on transfer learning mechanism and UWB radar. Background Technology

[0002] In emergency rescue operations following disasters such as earthquakes and landslides, rapid and accurate detection of life signs in buried survivors is crucial. Traditional life detection technologies, such as optical cameras and acoustic / vibration detection, have limitations. Optical cameras are affected by visibility conditions, lighting, and smoke, while acoustic detection requires survivors to speak or actively tap. Ultra-wideband radar technology, due to its strong penetrating power, high range resolution, and sensitivity to minute movements, has become a hot research topic in penetrating life detection. It can penetrate non-metallic media and detect minute movements in the chest cavity through the micro-Doppler effect.

[0003] However, using UWB radar for disaster site detection on mobile platforms faces several technical bottlenecks: First, the environment is complex and highly interfered with; strong clutter in radar echoes and interference from platform movement at disaster sites can drown out vital signs. Second, data is scarce and model generalization ability is poor; disaster scenarios are diverse, making it difficult to obtain large amounts of real data to train AI models. Models trained in a single scenario have reduced generalization ability in real-world scenarios, leading to false alarms and missed alarms. Third, mobile platforms distort signals; platform movement causes frequency shifts and phase distortions in echo signals, making traditional static monitoring algorithms unsuitable. Fourth, real-time requirements are high; rescue missions are urgent, and the system needs to determine signs of life within seconds and quickly screen multiple targets. Moreover, most existing technologies are designed for vital sign monitoring in static, controllable environments, lacking specific optimizations for mobile detection scenarios and extreme environments. For example, the underground information three-dimensional detection method in patent CN112526511A is mostly designed for vital sign monitoring in static, controllable environments, lacking specific optimizations for mobile detection scenarios and extreme environments.

[0004] Therefore, the present invention aims to provide a method, system, and apparatus for mobile life detection based on transfer learning mechanism and UWB radar to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to solve the above-mentioned problems and provide a method, system, and device for mobile life detection based on transfer learning mechanism and UWB radar. It establishes a technical system of simulation-transfer-real-time inference, uses high-fidelity physical simulation to generate a large amount of diverse and labeled training data, and pre-trains a deep neural network model. When the mobile platform enters the real disaster site, it uses transfer learning technology to collect a very small amount of data from inanimate areas or simple calibration data to quickly fine-tune the pre-trained model to adapt to the current special environment, and finally achieves high-precision mobile life detection.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This invention provides a method for detecting mobile life based on transfer learning mechanisms and UWB radar, the method comprising the following steps:

[0008] S1. Mobile platform UWB radar data acquisition and motion compensation:

[0009] UWB radar is mounted on a mobile platform (such as a robot) and continuously collects echo data while moving, performs motion compensation on it, and constructs a radar data cube (slow time × fast time × channel).

[0010] The mobile platform (such as a robot) contains inertial measurement unit data or radar point cloud data, which can perform motion compensation on radar echoes and eliminate phase errors caused by the platform's own movement.

[0011] S2. Clutter suppression and region of interest generation:

[0012] Algorithms such as background cancellation and moving target indication are applied to motion-compensated data to suppress stationary clutter.

[0013] By using constant false alarm rate detection or energy detection, a region of interest containing potential moving targets is generated;

[0014] S3. Separation and feature extraction of vital signs signals:

[0015] Micro-motion signals are separated from the region of interest (ROI), and time-frequency analysis tools such as discrete wavelet transform and VMD are used to extract time-domain, frequency-domain, and nonlinear features (i.e. vital signs signals) related to respiration and heartbeat.

[0016] S4. Life identification based on transfer learning:

[0017] Pre-training phase: On the server side, deep convolutional neural network models or temporal neural network models (such as ResNet and LSTM) are pre-trained using large-scale simulation datasets (including radar echo data and labels of various ruin structures, human postures, vital signs parameters, etc., which are generated by electromagnetic simulation software (such as CST and HFSS) and human physiological models).

[0018] Fine-tuning phase: Simulate the acquisition process of S1-S3 and collect real data from the rescue site. Collect a small amount of real data in the current environment (which may include "negative samples" from inanimate areas and / or simple calibration data from known locations). Use this data to perform domain-adaptive fine-tuning on the aforementioned pre-trained model, allowing the model to quickly adapt to the unique clutter distribution and propagation characteristics of the site.

[0019] Real-time inference: Input the real-time features extracted in step S3 into the fine-tuned model, output the probability of the presence of vital signs, and estimate the respiratory rate and heart rate;

[0020] S5. Multi-target localization and information fusion:

[0021] By combining the ranging and angle measuring capabilities of radar, multiple detected life targets can be located.

[0022] By integrating with SLAM map information, the location and status of life forms are marked on the environmental map of the mobile platform.

[0023] This invention provides a mobile life detection system based on transfer learning mechanism and UWB radar. The system includes a data acquisition module, a preprocessing module, a feature extraction module, a recognition and learning module, a multi-target localization module, and an output terminal.

[0024] The data acquisition module uses UWB radar to collect radar echo data during the movement process;

[0025] The preprocessing module uses an IMU inertial sensor to perform motion compensation on radar echo data to eliminate phase errors; and employs background cancellation and moving target indication algorithms to suppress clutter caused by stationary objects or the environment.

[0026] The feature extraction module generates a region of interest containing potential moving targets through constant false alarm rate detection or energy detection; it separates micro-motion signals from the region of interest and extracts vital signs signals from the micro-motion signals using time-frequency analysis tools;

[0027] The identification and learning module constructs a simulation dataset using electromagnetic simulation software and a human physiological model, and pre-trains the simulation dataset to obtain a pre-trained model; it also collects real data from the rescue site, and the pre-trained model is adaptively fine-tuned based on the real data.

[0028] The multi-target localization module extracts vital signs signals and inputs them into the fine-tuned pre-trained model to calculate the probability of vital signs, respiratory rate, and heart rate, and accurately locates multiple targets through joint processing of multi-sensor data.

[0029] The output terminal outputs the probabilities of vital signs, respiratory rate, and heart rate calculated in the pre-training module.

[0030] A chip implementing the aforementioned mobile life detection method / system based on transfer learning mechanism and UWB radar is disclosed. This chip is a UWB radar core board, which connects an MCU, an RF front-end, an antenna array, an IMU module, a high-performance embedded AI computing card, and various communication interfaces. The UWB radar core board integrates the MCU, RF front-end, and antenna array, providing accurate distance and position measurements while ensuring system stability and reliability. The IMU module outputs high-frequency attitude data for motion compensation, ensuring the radar system can accurately measure distance and position even in dynamic environments. The high-performance embedded AI computing card (e.g., equipped with a GPU / NPU) runs the transfer learning model, significantly improving the radar system's data processing capabilities and intelligence, enabling more efficient target identification and classification. Multiple communication interfaces (including 5G, Wi-Fi, and Ethernet) support high-speed data transmission, enabling collaborative detection with other radar systems or sensor networks. Through these interfaces, the core board can easily communicate with other devices, share data, and work collaboratively, thereby enhancing the overall system performance and reliability.

[0031] Compared with existing technologies, the beneficial effects of this solution are:

[0032] This invention addresses the data scarcity problem by leveraging simulation and transfer learning. It pre-trains models using low-cost, customizable simulation data, then adapts them to real-world scenarios through transfer learning, overcoming the bottleneck of obtaining real-world life data at disaster sites and enhancing model generalization performance and deployment efficiency. Furthermore, the system is optimized end-to-end for mobile platforms, from hardware motion compensation to algorithmic suppression of motion interference, enabling "detection while moving." The pre-trained model boasts excellent feature extraction capabilities, with rapid convergence and minimal data requirements during fine-tuning. The system can complete on-site adaptation within minutes, maintaining over 95% accuracy and an extremely low false alarm rate. The system architecture supports collaborative operation of multiple mobile detection devices across a network, sharing data and model updates, expanding the search range, and improving detection efficiency and reliability. Attached Figure Description

[0033] Figure 1 This is a general flowchart of the method in the embodiments of the present invention;

[0034] Figure 2 This is a schematic diagram illustrating the training and fine-tuning of the transfer learning model in an embodiment of the present invention;

[0035] Figure 3 This is an overall architecture diagram of the system in an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of the UWB radar core board in an embodiment of the present invention;

[0037] Figure 5This is a schematic diagram illustrating the principle of the motion compensation algorithm in this embodiment of the invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be described in further detail below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the embodiments.

[0040] Example 1:

[0041] This invention provides a method for detecting mobile life based on transfer learning mechanisms and UWB radar, the method comprising the following steps:

[0042] S1. Mobile platform UWB radar data acquisition and motion compensation:

[0043] UWB radar is mounted on a mobile platform (such as a robot) and continuously collects echo data while moving, performs motion compensation on it, and constructs a radar data cube (slow time × fast time × channel).

[0044] The mobile platform (such as a robot) contains inertial measurement unit data or radar point cloud data, which can perform motion compensation on radar echoes and eliminate phase errors caused by the platform's own movement.

[0045] S2. Clutter suppression and region of interest generation:

[0046] Algorithms such as background cancellation and moving target indication are applied to motion-compensated data to suppress stationary clutter.

[0047] By using constant false alarm rate detection or energy detection, a region of interest containing potential moving targets is generated;

[0048] S3. Separation and feature extraction of vital signs signals:

[0049] Micro-motion signals are separated from the region of interest (ROI), and time-frequency analysis tools such as discrete wavelet transform and VMD are used to extract vital sign signals;

[0050] S4. Life identification based on transfer learning:

[0051] Pre-training phase: On the server side, deep convolutional neural networks or temporal neural networks (such as ResNet and LSTM) are pre-trained using a large-scale simulation dataset (containing radar echo data and labels of various ruin structures, human postures, vital signs parameters, etc., generated by electromagnetic simulation software (such as CST and HFSS) and human physiological models).

[0052] Fine-tuning phase: Simulate the acquisition process of S1-S3 and collect real data from the rescue site. Collect a small amount of real data in the current environment (which may include "negative samples" from inanimate areas and / or simple calibration data from known locations). Use this data to perform domain-adaptive fine-tuning on the aforementioned pre-trained model, allowing the model to quickly adapt to the unique clutter distribution and propagation characteristics of the site.

[0053] Real-time inference: Input the real-time features extracted in step S3 into the fine-tuned model, output the probability of the presence of vital signs, and estimate the respiratory rate and heart rate;

[0054] S5. Multi-target localization and information fusion:

[0055] By combining the ranging and angle measuring capabilities of radar, multiple detected life targets can be located.

[0056] By integrating with SLAM map information, the location and status of life forms are marked on the environmental map of the mobile platform.

[0057] This invention provides a mobile life detection system based on transfer learning mechanism and UWB radar. The system includes a data acquisition module, a preprocessing module, a feature extraction module, a recognition and learning module, a multi-target localization module, and an output terminal.

[0058] The data acquisition module uses UWB radar to collect radar echo data during the movement process;

[0059] The preprocessing module uses an IMU inertial sensor to perform motion compensation on radar echo data to eliminate phase errors; it also uses background cancellation and moving target indication algorithms to suppress clutter caused by stationary objects or the environment.

[0060] The feature extraction module generates a region of interest containing potential moving targets through constant false alarm rate detection or energy detection; micro-motion signals are separated from the region of interest, and vital signs signals are extracted from the micro-motion signals using time-frequency analysis tools;

[0061] The identification and learning module constructs a simulation dataset using electromagnetic simulation software and a human physiological model, and pre-trains the simulation dataset to learn the general characteristics of identifying vital signs from radar signals, thus obtaining a pre-trained model; real data from the rescue site is collected, and the pre-trained model is adaptively fine-tuned based on the real data;

[0062] The implementation features extracted by S3 are input into the fine-tuned pre-trained model to calculate the probability of vital signs, respiratory rate, and heart rate.

[0063] The multi-target localization module extracts vital signs signals and inputs them into a finely tuned pre-trained model to calculate the probability of vital signs, respiratory rate, and heart rate. It then uses multi-sensor data processing to accurately locate multiple targets. The mobile life detection system connects to an external human-machine interface terminal, which displays the detection results, vital signs data, and a map of the target locations.

[0064] A chip implementing the aforementioned mobile life detection method / system based on transfer learning and UWB radar is disclosed. The core component of this chip is a UWB radar core board, which integrates an MCU, RF front-end, antenna array, IMU module, high-performance embedded AI computing card, and multiple communication interfaces. The UWB radar core board can accurately measure distance and position while ensuring system stability and reliability. The IMU module outputs high-frequency attitude data for motion compensation, ensuring the radar system can accurately measure target position even in dynamic environments. The embedded AI computing card (such as a GPU / NPU) runs the transfer learning model, improving the system's data processing capabilities and intelligence, and effectively identifying and classifying targets. Multiple communication interfaces (such as 5G, Wi-Fi, and Ethernet) support high-speed data transmission, enabling collaborative operation with other radar systems or sensor networks, and enhancing the overall system performance and reliability.

[0065] The MCU (Microcontroller Unit), RF front-end, antenna array, IMU module, embedded AI computing card, and communication interface play crucial roles in the system and work together seamlessly. As the core of the embedded system, the MCU coordinates the work of various modules, executes application logic, and processes data. It exchanges data with other modules through the communication interface, ensuring smooth information flow.

[0066] The radio frequency (RF) front-end is responsible for converting baseband signals into RF signals, or vice versa, the received RF signals. It works in conjunction with the antenna array, adjusting the antenna's phase and amplitude to improve signal transmission efficiency and receiver sensitivity, thereby optimizing communication performance. The MCU manages signal transmission and reception by controlling the operating states of the RF front-end and antenna array, such as frequency band selection and power adjustment.

[0067] The IMU module is used to measure the system's acceleration, angular velocity, and orientation in real time, and is crucial in applications such as positioning, navigation, and attitude control. The MCU processes the data from the IMU to ensure stable system operation in dynamic environments.

[0068] High-performance embedded AI computing cards, equipped with GPUs or dedicated AI accelerators, handle complex AI tasks such as image recognition and path planning. The MCU and AI computing card collaborate via a high-speed communication interface; the MCU schedules tasks, while the AI ​​computing card handles efficient inference computation. Furthermore, data from the IMU module can be further analyzed by the AI ​​computing card, enabling higher-level data processing and intelligent decision-making.

[0069] Multiple communication interfaces ensure that the MCU can exchange data efficiently and in real time with devices such as the RF front-end, IMU module, and AI computing card, playing a crucial role in the system's performance and response speed. Through close cooperation, the detection task was completed.

[0070] When using the device through the above embodiments of the present invention, the mobile detection platform is deployed in complex or dangerous environments (such as earthquake ruins, fire scenes, mine tunnels, etc.) and collects UWB radar echo data in real time during autonomous movement. The IMU module provides high-frequency attitude information simultaneously for motion compensation of the radar signal, effectively eliminating phase errors introduced by platform movement. The compensated radar data is sent to a high-performance embedded AI computing card, combined with a pre-trained and finely tuned transfer learning model, to achieve high-sensitivity identification and classification of weak vital signs, accurately determine whether there are trapped personnel, and simultaneously estimate their respiratory rate and heart rate. At the same time, the system uploads the detection results, target location, and physiological status to the command center in real time through communication interfaces such as 5G, Wi-Fi, or Ethernet, and can work collaboratively with other surrounding sensor nodes to build a multi-source fusion life detection network, significantly improving search and rescue efficiency, coverage, and system robustness.

[0071] Example 2: Implementation of the present invention:

[0072] 1. Preparation Phase: On a cloud server, several types of UWB radar echo data are generated using computer simulation. These data cover different rubble structures, burial depths, and human postures, and are labeled with whether they contain living organisms and their vital signs. A one-dimensional CNN-LSTM hybrid network model is pre-trained using this data.

[0073] 2. On-site deployment: Rescue personnel carried robots equipped with the detection device of this invention into the ruins, and the robots immediately began to move and scan.

[0074] 3. On-site calibration: First, scan an area confirmed to be devoid of life for 2-3 minutes to collect environmental data as a "negative sample"; or place a simple dynamic calibration object (such as a small pendulum swinging) at a known location to collect a small amount of data.

[0075] 4. Rapid model fine-tuning: Using a small amount of collected field data, fine-tuning is performed on the last one or a few layers of the pre-trained CNN-LSTM model. This process can be completed within minutes on the robot's edge computing unit.

[0076] 5. Real-time Life Detection: The fine-tuned model is loaded into the real-time inference engine. The robot continues to move, and the UWB radar collects data in real time. After motion compensation and feature extraction, the data is input into the model for real-time judgment. Once a vital sign signal (such as regular respiratory harmonics) is detected within a certain ROI, the system will immediately issue an audible and visual alarm on the operator's terminal and highlight the target location and estimated vital sign values ​​on the SLAM map.

[0077] 6. Continuous learning: During the detection process, operators can confirm or correct the detection results (e.g., confirm that it is a false alarm). This feedback information can be further used for the online learning of the model, continuously improving its performance in the current scenario.

[0078] Through the above methods, the present invention achieves rapid, accurate, and robust non-contact mobile life detection in complex and unknown disaster site environments.

[0079] The above specific embodiments are merely explanations of the present invention and are not intended to limit the present invention. After reading this specification, those skilled in the art can make modifications to these embodiments without contributing any inventive step, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A mobile life detection method based on transfer learning mechanism and UWB radar, characterized by: The method includes the following steps: S1. Mount the UWB radar on a mobile platform, collect radar echo data collected by the mobile platform, perform motion compensation on the data, and construct a radar data cube. S2. Process the radar data cube to suppress stationary clutter, and generate a region of interest containing potential moving targets through constant false alarm rate detection or energy detection. S3. Separate the micro-motion signal from the region of interest and extract vital signs signals from the micro-motion signal using time-frequency analysis tools; S4. Construct a simulation dataset using electromagnetic simulation software and a human physiological model, and pre-train the simulation dataset. The data collection process (S1-S3) is simulated to collect real data from the rescue site. The pre-trained model is then adaptively fine-tuned based on the real data. The vital signs signals extracted by S3 are input into the fine-tuned pre-trained model to calculate the probability of vital signs, respiratory rate, and heart rate. S5. Combine the calculated ranging and angle measurement capabilities with those of the UWB radar to locate the detected signs of life. By integrating location information with SLAM map information, the location and status of living organisms can be marked on the environmental map of the mobile platform.

2. The mobile life detection method based on transfer learning mechanism and UWB radar as described in claim 1, characterized in that: Motion compensation is performed on the radar echo based on data from the inertial measurement unit of the S1 mobile platform or radar point cloud data.

3. The mobile life detection method based on transfer learning mechanism and UWB radar as described in claim 1, characterized in that: The time-frequency analysis tools in S3 include discrete wavelet transform and VMD.

4. The mobile life detection method based on transfer learning mechanism and UWB radar as described in claim 1, characterized in that: In step S4, a simulation dataset is obtained by inputting a human physiological model and radar echo data into electromagnetic simulation software. The dataset is then pre-trained using a deep convolutional-temporal hybrid network to obtain a pre-trained model. The pre-trained model can be adapted to the clutter distribution and propagation characteristics at the rescue site.

5. The mobile life detection system based on transfer learning mechanism and UWB radar as described in claim 1, characterized in that: The system includes a data acquisition module, a preprocessing module, a feature extraction module, a recognition and learning module, a multi-target localization module, and an output terminal; The data acquisition module uses UWB radar to collect radar echo data during the movement process; The preprocessing module uses an IMU inertial sensor to perform motion compensation on radar echo data to eliminate phase errors; and employs background cancellation and moving target indication algorithms to suppress clutter caused by stationary objects or the environment. The feature extraction module generates a region of interest containing potential moving targets through constant false alarm rate detection or energy detection; it separates micro-motion signals from the region of interest and extracts vital signs signals from the micro-motion signals using time-frequency analysis tools; The identification and learning module constructs a simulation dataset using electromagnetic simulation software and a human physiological model, and pre-trains the simulation dataset to obtain a pre-trained model; it also collects real data from the rescue site, and the pre-trained model is adaptively fine-tuned based on the real data. The multi-target localization module extracts vital signs signals and inputs them into the fine-tuned pre-trained model to calculate the probability of vital signs, respiratory rate, and heart rate, and accurately locates multiple targets through joint processing of multi-sensor data. The output terminal outputs the probabilities of vital signs, respiratory rate, and heart rate calculated in the pre-training module.

6. The mobile life detection system based on transfer learning mechanism and UWB radar as described in claim 5, characterized in that: The mobile life detection system is connected to an external human-machine interaction terminal, and the output terminal is the human-machine interaction terminal, which displays the detection results, vital sign data, and target location map.

7. A chip that implements the above-described method / system for detecting mobile life based on transfer learning mechanism and UWB radar, characterized in that: The chip is a UWB radar core board, which is connected to an MCU, a radio frequency front-end, an antenna array, an IMU module, a high-performance embedded AI computing card, and various communication interfaces.

Images