XGBoost-based high-altitude rescue sign detection method and device

Abstract

The invention provides a high-altitude rescue sign detection method and device based on XGBoost, and the method comprises the steps: obtaining multi-source data of a rescue worker in an operation process, and the multi-source data comprises personal data, environment data and physiological detection data; aggregating the multi-source data to generate aggregated feature data; storing the aggregated feature data as sparse matrix data; inputting the sparse matrix data into a pre-trained XGBoost model, and generating a health risk prediction result; and the health risk prediction result is transmitted to a command end system, and the command end system triggers a rescue instruction according to the health risk prediction result. According to the method, the health risk prediction is carried out by aggregating the multi-source data and using the pre-trained XGBoost model, so that the intelligent assessment capability of the health risk of the rescue personnel in the high-altitude rescue scene is improved, the life safety of the rescue personnel is effectively guaranteed, and the defect of imperfect decision logic in the prior art is overcome.

Description

Technical Field

[0001] The present application relates to the field of high-altitude rescue, and in particular to a method and device for detecting vital signs of high-altitude rescue based on XGBoost. Background Art

[0002] my country's high-altitude areas have harsh environmental problems such as low temperature, low pressure, low oxygen and prone to disasters. High-altitude rescue workers are prone to diseases and psychological problems during their operations. The existing vital sign detection system mainly focuses on physiological indicator detection, and has not established a deep connection between changes in physiological indicators, environmental factors, and health risks. The rescue decision-making logic needs to be optimized. Summary of the Invention

[0003] The purpose of this application is to overcome the defects in the above-mentioned prior art and provide a high-altitude rescue vital sign detection method and device based on XGBoost.

[0004] This application provides a high-altitude rescue sign detection method based on XGBoost, including: Acquire multi-source data of rescue personnel during operations, including personal data, environmental data, and physiological test data; Aggregating the multi-source data to generate aggregated feature data; storing the aggregated feature data as sparse matrix data; Inputting the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; The health risk prediction result is transmitted to the command end system, and the command end system triggers a rescue instruction according to the health risk prediction result.

[0005] Optionally, the pre-trained XGBoost model is a TensorFlow-Lite format model and is deployed on the head-mounted display Android system.

[0006] Optionally, the XGBoost model includes: When processing missing values, the feature values of normal samples are traversed and missing samples are divided, and the splitting direction is selected according to the optimal loss value of the normal samples and the missing samples.

[0007] Optionally, the environmental data is collected once every 3 minutes through a temperature sensor and a pressure sensor; the physiological detection data is collected once every 5 seconds through a smart bracelet and a head-mounted display movement sensor.

[0008] Optionally, it also includes: The multi-source data is transmitted via an ad hoc network, which includes a UAV base station and a BeiDou / GPS positioning system; The pre-trained XGBoost model is regularly updated based on the aggregated feature data from the head-mounted display to the server.

[0009] This application also provides a high-altitude rescue vital sign detection device based on XGBoost, comprising: An acquisition module is used to acquire multi-source data of rescue personnel during their operations, including personal data, environmental data, and physiological detection data; an aggregation module, aggregating the multi-source data to generate aggregated feature data; A storage module, storing the aggregated feature data as sparse matrix data; A prediction module, which inputs the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; The trigger module transmits the health risk prediction result to the command end system, and the command end system triggers the rescue command according to the health risk prediction result.

[0010] Optionally, the pre-trained XGBoost model is a TensorFlow-Lite format model and is deployed on the head-mounted display Android system.

[0011] Optionally, the XGBoost model includes: When processing missing values, the feature values of normal samples are traversed and missing samples are divided, and the splitting direction is selected according to the optimal loss value of the normal samples and the missing samples.

[0012] Optionally, the environmental data is collected once every 3 minutes through a temperature sensor and a pressure sensor; the physiological detection data is collected once every 5 seconds through a smart bracelet and a head-mounted display movement sensor.

[0013] Optionally, it also includes: The multi-source data is transmitted via an ad hoc network, which includes a UAV base station and a BeiDou / GPS positioning system; The pre-trained XGBoost model is regularly updated based on the aggregated feature data from the head-mounted display to the server.

[0014] The beneficial effects of this application are: The present application provides a high-altitude rescue vital sign detection method based on XGBoost, comprising: obtaining multi-source data of rescue personnel during the operation process, wherein the multi-source data includes personal data, environmental data and physiological detection data; aggregating the multi-source data to generate aggregated feature data; storing the aggregated feature data as sparse matrix data; inputting the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; transmitting the health risk prediction result to the command end system, and the command end system triggers a rescue instruction according to the health risk prediction result. The present application improves the intelligent assessment capability of the health risks of rescue personnel in high-altitude rescue scenarios by aggregating multi-source data and using a pre-trained XGBoost model to predict health risks, effectively ensuring the life safety of rescue personnel, and overcoming the defects of imperfect decision-making logic in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of the high-altitude rescue vital sign detection process based on XGBoost in this application. DETAILED DESCRIPTION

[0016] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that various forms of implementation of the present disclosure are not limited to the embodiments set forth herein. Rather, the embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0017] In this application, the research on the rescue personnel operation support system in high-altitude environment is carried out by having the rescue personnel wear smart devices to collect the specified physiological indicator data and then transmit the data back to the head-mounted smart display device.

[0018] The headset aggregates the feature data detected by the device, the entered personal feature data, and the environmental feature data to form online feature data, and scores it using the XGBoost algorithm model deployed on the headset. The scoring results are synchronized to the command end for result notification, providing a basis for decision-making for rescue personnel and command personnel to take measures.

[0019] At the same time, the feature data is stored in the database on the headset side, and the data is exported after the rescue workers return to the base. It is used as a data set for model training to support iterative optimization of the model.

[0020] Please refer to Figure 1 As shown, the present application provides a high-altitude rescue vital sign detection method based on XGBoost, comprising: S101. Acquire multi-source data of rescue personnel during operations, wherein the multi-source data includes personal data, environmental data, and physiological test data; Build a high-altitude ad hoc network and connect the command system and headset-side equipment to the ad hoc network to ensure signal coverage and network connectivity within the rescue area. At the same time, verify the headset-side positioning function, allowing the location, number, name, and other information of the rescue personnel to be observed from the command end display screen, facilitating real-time command. Verify the connection between the headset and smart wearable devices, including smart bracelets and smart anklets, as well as temperature and air pressure sensors worn on the side of clothing. Each smart device is connected to the headset via Bluetooth.

[0021] Communication is limited at high altitudes, preventing cloud-based models from responding in real time. Local deployment of a complete model is limited by the computing power of the headset. In this application, the XGBoost model is ranked by feature importance based on plateau pathology characteristics, and low-altitude irrelevant features (such as humidity-related branches) are pruned to compress the model size to less than 200KB. Communication areas are dynamically divided using Beidou / GPS positioning. Directional data relay from the drone base station to the command center is triggered only when the model predicts a risk level ≥ a threshold, avoiding channel congestion.

[0022] This application enables the head-mounted display to complete real-time predictions in a short time, meeting the low power consumption and high real-time requirements of extreme environments above 5,000 meters above sea level.

[0023] The command side maintains the personal data of the rescue workers in the early stage, and imports the corresponding wearer's personal information into the head-mounted display device when the mission begins, such as the wearer Zhang San, 35 years old, etc. The head-mounted display device obtains the current ambient temperature and air pressure information through the temperature and air pressure sensors worn on the outside of the clothing. The collection interval is once every 3 minutes. After the information is collected, it is stored in the LruCache cache; The head-mounted display device collects the wearer's personal physiological information through smart bracelets and smart anklets, including electrocardiogram, heart rate, blood oxygen saturation, body temperature and other data. The head-mounted display device calls on its own mobile sensor device to obtain the wearer's linear acceleration and pedometer related information; Collecting raw electrocardiogram (ECG) and heart rate information requires collecting 7-10 seconds of continuous data, from which we extract 5 seconds of data. Because blood oxygen saturation and heart rate are closely related, we measure data from both within 5 seconds. After collection, we take the lowest blood oxygen content value as the data sample, such as a heart rate of 100 and a blood oxygen saturation of 90%.

[0024] At high altitudes, environmental parameters change slowly, while physiological indicators can change suddenly. Traditional methods use fixed-frequency acquisition, resulting in redundant data (environmental) or missed detection of critical signals (physiological).

[0025] Based on a high-altitude physical model of air pressure change, this application derives a three-minute effective monitoring cycle for environmental parameters, avoiding the power consumption of frequent sensor wake-ups. Furthermore, incorporating the pathological characteristics of acute mountain sickness (e.g., a drop in blood oxygen saturation of ≥5% within 5 seconds is considered a dangerous threshold), a 5-second sliding window is set to calculate indicator extremes to capture transient anomalies. By dynamically matching differentiated acquisition frequencies with time windows, this approach reduces overall power consumption while ensuring millisecond-level response to high-altitude hypoxia.

[0026] S102, aggregating the multi-source data to generate aggregated feature data; After collecting all the data, the data format is converted on the headset side, and the features are extracted into a sparse matrix and predicted and scored. After scoring, the results are synchronized to the command side system so that the command personnel can obtain the status information of the rescue workers as soon as possible.

[0027] Among them, "aggregating feature data" includes: converting raw sensor data into numerical features; extracting key indicators from the raw data; and storing the features in a sparse format.

[0028] For example: Rescuer Zhang San is performing a mission at an altitude of 4,500 meters.

[0029] Original data input: Personal data: Age 35, male, plateau acclimatization index 0.72 (preset parameters); Environmental data (collected at 10:00:00): Temperature -15°C, Air pressure 57.2 kPa; Physiological data (collected at 10:00:05): Smart wristband: Heart rate 112 bpm, Blood oxygen saturation 88%; Head-mounted display accelerometer: X / Y / Z axis acceleration standard deviation 0.3 g; Pedometer: 0 steps in 5 seconds (stationary state) Aggregation feature generation logic: Time series alignment: 10:00:05 is used as the base timestamp, and environmental data continues to use the value of 10:00:00; Sliding window calculation: Heart rate takes the average of the last 5 seconds: (110+115+112+109+112) / 5=111.6 → integerized to 112 bpm; Blood oxygen saturation takes the lowest value within 5 seconds: 88% Motion feature extraction: Acceleration standard deviation 0.3g → Exercise intensity level 2 (0.2-0.5g is preset as light activity); Continuous inactivity time: 5 seconds (triggering the sedentary warning flag) Plateau adaptation compensation calculation: Comprehensive formula: Risk factor = (blood oxygen difference × 0.5 + heart rate increase × 0.3) / plateau acclimatization index; substitute the data: ( (100%-88%) × 0.5 + (112-80) × 0.3 ) / 0.72 = 9.2.

[0030] S103, storing the aggregated feature data as sparse matrix data; After the HMD-side model performs scoring processing, the aggregated data that has undergone feature extraction is stored in the HMD-side SQLiteDatabase database as a dataset for subsequent training or verification; Extract various features into a sparse matrix, including: For time series data such as heart rate and blood oxygen saturation, the average value of the sliding window in the last 5 seconds is taken as the feature value; for linear acceleration data, the standard deviation and maximum value are taken as motion state features; All sensor data are aligned with the 1-second timestamp based on the HMD device clock to generate aggregate features.

[0031] For example: input aggregated feature data: the aggregated data at 10:00:05 in the above step S103.

[0032] Sparse matrix conversion rules: Column definitions are shown in Table 1: Table 1

[0033] Conversion process: Age 35 (within the typical range of 25-45) → Not stored; Plateau Acclimation Index 0.72 → Store 0.72 in Column 1; Temperature -15°C (1°C change from the previous -16°C) → Store -15 in Column 2; Heart Rate 112 (≥100) → Store 112 in Column 3; Blood Oxygen 88 (≤90%) → Store 88 in Column 4; Exercise Intensity 2 (non-zero) → Store 2 in Column 5.

[0034] In high-altitude rescue scenarios, sensors are prone to failure due to low temperatures or vibration, resulting in data loss. Traditional sparse matrix storage requires manual interpolation of missing values, introducing bias.

[0035] This application does not handle missing values during data processing.

[0036] Compared with the problem of sample bias that may occur when manually filling missing values used in the GBDT algorithm, this application prioritizes normal samples during processing in XGBoost, attempts to divide missing samples when traversing normal eigenvalues, and then selects the split point with the optimal loss value.

[0037] The model is trained on the server and verified after completion. After verification, the obtained trained model is converted into a TensorFlow-Lite format model.

[0038] During XGBoost model training, this application assigns missing samples to the subtree with the fastest loss function decrease (rather than default filling), enabling the model to directly learn missing patterns in high-altitude scenarios (for example, when a temperature sensor fails at low temperatures, the weight of the blood oxygen oscillator automatically increases). During storage, only non-zero features and timestamp offsets are retained, and memory mapping technology on the headset enables real-time matrix updates, avoiding the static storage limitations of traditional sparse formats. This collaborative design of the model and storage format ensures that the system maintains high prediction accuracy even with high data missingness.

[0039] For example, rescue workers are working in high-altitude areas, and the sensor collects the following three pieces of data: Sample A: Blood oxygen 85% (low), missing temperature → high risk Sample B: Blood oxygen 92% (normal), temperature 5°C → Low risk Sample C: Blood oxygen 88% (low), missing temperature → high risk Do not fill missing values and process the original data directly: Normal sample B has a temperature of 5°C. When splitting, it is first divided into the left subtree (temperature ≤ 5°C). Attempt 1: Split A / C into the left subtree (on the same side as B) → The left subtree contains 1 low-risk (B) + 2 high-risk (A / C), resulting in a high loss value. Attempt 2: Split A / C into the right subtree → The right subtree has only two high-risk (A / C) and a low loss value Select the optimal split: Use attempt 2 and mark the right subtree as a high-risk area.

[0040] Finally, the left subtree: sample B (low risk, normal temperature); the right subtree: samples A / C (high risk, low blood oxygen and missing temperature).

[0041] S104, inputting the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; After collecting all the data, the data format is converted on the headset side, and the features are extracted into a sparse matrix and predicted and scored. The offline training system aggregates and preprocesses the stored data, then performs model training. After training, the XGBoost algorithm model is obtained and the model is verified and optimized. After the model is verified and exported to binary format, use tf.lite.TFLiteConverter.from_saved_model() to convert the model to the TensorFlow-Lite format that is compatible with the Android system on the headset. Deploy the XGBoost model to the head-mounted display device to support the scoring prediction business function; The model is continuously optimized based on the data sets collected during training or testing. Model training and deployment prediction are performed separately, so the headset-side model needs to be updated regularly based on the optimization results.

[0042] S105: Transmitting the health risk prediction result to a command-end system, and the command-end system triggering a rescue command according to the health risk prediction result.

[0043] After scoring, the results are synchronized to the command system, allowing the command personnel to obtain the status information of the rescue workers as soon as possible; The command side obtains abnormal information about the rescue personnel. For example, if it is found that the rescue worker Zhang San is in a state of hypoxia and stress, it will immediately command relevant personnel to take assistance measures.

[0044] After the training is completed and the rescue workers return to the base, the data on the headset side will be imported into the server-side file system.

[0045] This application also provides a high-altitude rescue vital sign detection device based on XGBoost, comprising: An acquisition module is used to acquire multi-source data of rescue personnel during their operations, including personal data, environmental data, and physiological detection data; an aggregation module, aggregating the multi-source data to generate aggregated feature data; A storage module, storing the aggregated feature data as sparse matrix data; A prediction module, which inputs the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; The trigger module transmits the health risk prediction result to the command end system, and the command end system triggers the rescue command according to the health risk prediction result.

[0046] Furthermore, the pre-trained XGBoost model is a TensorFlow-Lite format model and is deployed on the Android system of the head-mounted display.

[0047] Furthermore, the XGBoost model includes: When processing missing values, the feature values of normal samples are traversed and missing samples are divided, and the splitting direction is selected according to the optimal loss value of the normal samples and the missing samples.

[0048] Furthermore, the environmental data is collected every 3 minutes through temperature sensors and pressure sensors; the physiological detection data is collected every 5 seconds through smart bracelets and head-mounted display motion sensors.

[0049] Furthermore, it also includes: The multi-source data is transmitted via an ad hoc network, which includes a UAV base station and a BeiDou / GPS positioning system; The pre-trained XGBoost model is regularly updated based on the aggregated feature data from the head-mounted display to the server.

[0050] The above description of the embodiments is intended to facilitate understanding and application of the present invention by those skilled in the art. It will be readily apparent to those skilled in the art that various modifications to the above embodiments can be made, and the general principles described herein can be applied to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the present disclosure are intended to fall within the scope of protection of the present invention.

Claims

1. A high-altitude rescue vital sign detection method based on XGBoost, characterized in that: include: Acquire multi-source data of rescue personnel during operations, including personal data, environmental data, and physiological test data; Aggregating the multi-source data to generate aggregated feature data; storing the aggregated feature data as sparse matrix data; Inputting the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; Transmitting the health risk prediction result to a command-end system, wherein the command-end system triggers a rescue command according to the health risk prediction result; The blood oxygen saturation in the aggregated feature data is the lowest value within 5 seconds; The sparse matrix data only stores eigenvalues that are outside a preset range.

2. The high-altitude rescue vital sign detection method based on XGBoost according to claim 1, characterized in that: The pre-trained XGBoost model is a TensorFlow-Lite format model and is deployed on the Android system of the head-mounted display.

3. The high altitude rescue vital sign detection method based on XGBoost according to claim 1, characterized in that: The XGBoost model includes: When processing missing values, the feature values of normal samples are traversed and missing samples are divided, and the splitting direction is selected according to the optimal loss value of the normal samples and the missing samples.

4. The high-altitude rescue vital sign detection method based on XGBoost according to claim 1, characterized in that: The environmental data is collected every 3 minutes through temperature sensors and pressure sensors; the physiological detection data is collected every 5 seconds through smart bracelets and head-mounted display motion sensors.

5. The high altitude rescue vital sign detection method based on XGBoost according to claim 1, characterized in that: Also includes: The multi-source data is transmitted via an ad hoc network, which includes a UAV base station and a BeiDou / GPS positioning system; The pre-trained XGBoost model is regularly updated based on the aggregated feature data from the head-mounted display to the server.

6. A high altitude rescue vital sign detection device based on XGBoost, characterized in that: include: An acquisition module is used to acquire multi-source data of rescue personnel during their operations, including personal data, environmental data, and physiological detection data; an aggregation module, aggregating the multi-source data to generate aggregated feature data; A storage module, storing the aggregated feature data as sparse matrix data; A prediction module, which inputs the sparse matrix data into a pre-trained XGBoost model to generate a health risk prediction result; A trigger module transmits the health risk prediction result to a command-end system, and the command-end system triggers a rescue command according to the health risk prediction result; The blood oxygen saturation in the aggregated feature data is the lowest value within 5 seconds; The sparse matrix data only stores eigenvalues that are outside a preset range.

7. The high altitude rescue vital sign detection device based on XGBoost according to claim 6, characterized in that: The pre-trained XGBoost model is a TensorFlow-Lite format model and is deployed on the Android system of the head-mounted display.

8. The high altitude rescue vital sign detection device based on XGBoost according to claim 6, characterized in that: The XGBoost model includes: When processing missing values, the feature values of normal samples are traversed and missing samples are divided, and the splitting direction is selected according to the optimal loss value of the normal samples and the missing samples.

9. The high altitude rescue vital sign detection device based on XGBoost according to claim 6, characterized in that: The environmental data is collected every 3 minutes through temperature sensors and pressure sensors; the physiological detection data is collected every 5 seconds through smart bracelets and head-mounted display motion sensors.

10. The high altitude rescue vital sign detection device based on XGBoost according to claim 6, characterized in that: Also includes: The multi-source data is transmitted via an ad hoc network, which includes a UAV base station and a BeiDou / GPS positioning system; The pre-trained XGBoost model is regularly updated based on the aggregated feature data from the head-mounted display to the server.

Images