Non-sensitive military physical fitness real-time evaluation method based on edge calculation

Abstract

The invention discloses a non-sensitive military physical ability real-time evaluation method based on edge calculation, and aims to solve the problems that an existing evaluation method is subjectively lagged, depends on a fixed site and a fixed network, and is insufficient in tactical action evaluation precision in a complex environment. The method comprises the steps that multi-modal physiological motion data of soldiers are collected in a non-inductive mode through a wearable device, and local preprocessing is carried out; operating the lightweight model by the personal edge computing terminal, identifying military tactical actions in real time and extracting features; and establishing a connection with the edge computing node through a dynamic preferential strategy, and uploading feature data to converge group context information. According to the method, real-time, accurate and non-sensitive evaluation of the whole military training process in a field environment without a stable network is realized, a training evaluation closed loop is formed, and the scientificity and safety of training are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of military training and information technology, and in particular to a non-intrusive real-time assessment method for military physical fitness based on edge computing. Background Technology

[0002] Military physical fitness is the foundation of combat effectiveness. Scientific, accurate, and efficient physical fitness assessment is key to improving training effectiveness, preventing training injuries, and maintaining sustained combat capability. Traditional military physical fitness assessment methods have the following limitations:

[0003] First, the assessment model is outdated and highly subjective. Traditional assessments rely on manual timing, counting, and action evaluation, which not only consumes a lot of manpower but also makes the evaluation criteria susceptible to factors such as the examiner's subjective experience and fatigue level, making it difficult to guarantee fairness and accuracy with a "one-size-fits-all" approach. Although there have been attempts in recent years to introduce technologies such as intelligent robots and machine vision for automated assessments, these systems are usually deployed in fixed assessment venues and are mainly used for the final assessment of standardized courses, failing to cover real-time monitoring and feedback throughout the entire daily training process.

[0004] Secondly, existing technologies struggle to meet the real-time requirements of complex, combat-oriented training scenarios. To acquire richer physiological and motor data, researchers have begun using wearable sensor networks to monitor soldiers. However, uploading all the resulting massive amounts of high-frequency sensor data to a cloud center for processing presents challenges such as unstable network conditions in field training areas, high data transmission latency, and high bandwidth consumption, failing to meet the tactical demands for real-time assessment and immediate feedback. Real-time feedback on soldiers' movement quality and physiological state is crucial for adjusting tactics and preventing injuries when performing complex dynamic tasks such as tactical obstacle crossings and urban warfare simulations, but cloud-based processing models cannot guarantee this low-latency requirement.

[0005] Furthermore, the accuracy of data processing and evaluation models faces challenges. While high-precision data can be acquired in laboratory environments using devices such as optical motion capture, these devices are unsuitable for field training. Wearable devices collect data in complex outdoor environments characterized by high noise levels and diverse patterns. General-purpose activity recognition models often only recognize basic movements like walking, running, and jumping, lacking accuracy in recognizing military-specific tactical movements (such as low-crawl, leaping, and using cover), and making it even more difficult to quantify the standardization and intensity of these movements. In addition, significant individual differences among soldiers make a single model unsuitable for universal application, while running complex evaluation models directly on the sensor end is limited by device computing power and battery life.

[0006] Therefore, the field of military physical fitness assessment urgently needs a technical solution that can achieve real-time, non-intrusive, low-latency, and highly accurate assessment of soldiers' physical fitness throughout the entire training process in complex, unstable network environments during field training. This invention aims to solve the aforementioned problems. Summary of the Invention

[0007] To achieve the above objectives, this invention provides a non-intrusive real-time military physical fitness assessment method based on edge computing, comprising the following steps:

[0008] Step S1: Using a wearable sensing device that integrates a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer, and an optoelectronic heart rate sensor, the soldier's raw acceleration data, angular velocity data, geomagnetic direction data, and heart rate pulse signal are collected at a sampling frequency of not less than 100Hz; the wearable sensing device performs preprocessing on the raw sensing data stream, including amplitude limiting filtering and low-pass filtering, and packages the preprocessed multi-channel sensor data into structured data frames and sends them to the personal edge computing terminal;

[0009] Step S2: The personal edge computing terminal receives structured data frames and inputs multi-sensor data within a fixed time window into a locally deployed lightweight convolutional neural network model for real-time inference to identify military tactical action categories; the personal edge computing terminal simultaneously extracts the temporal and frequency domain features of the data within the time window to generate a local feature dataset containing action category labels and physiological motion feature vectors.

[0010] Step S3: Edge computing nodes deployed in the training site periodically broadcast their own computing load rate, available battery power, and signal reception strength; personal edge computing terminals dynamically select a master edge computing node and establish a connection based on the broadcast information, and upload their local feature datasets to the master edge computing node; the master edge computing node aggregates the data uploaded by all associated personal edge computing terminals to form a group time-series feature dataset;

[0011] Step S4: The main edge computing node loads the personalized physical fitness assessment model; the input of the personalized physical fitness assessment model includes the target soldier's local feature dataset and contextual features extracted from the group's temporal feature dataset; the personalized physical fitness assessment model outputs quantified real-time physical fitness indicators, which include at least instantaneous load intensity, movement standard score, fatigue coefficient and short-term injury risk index.

[0012] Step S5: The main edge computing node sends the real-time physical fitness indicators back to the corresponding personal edge computing terminal, and provides real-time tactile or audiovisual feedback to the soldiers through the personal edge computing terminal; at the same time, the main edge computing node pushes the anonymized aggregated evaluation results of all soldiers to the commander's command terminal.

[0013] Preferably, in step S1, the specific processes of amplitude limiting filtering and low-pass filtering are as follows:

[0014] The microprocessor built into the wearable sensing device first uses a sliding window-based amplitude limiting filtering method to compare the sensor data of the current sampling point with the median of the previous data points in the sliding window. If the difference exceeds a dynamic threshold preset based on the sensor range, the median value is used to replace the current sampling point data to remove impulse noise.

[0015] Next, a fourth-order Butterworth low-pass filter with a cutoff frequency that is adaptively adjusted according to the main movement frequency range of the soldier is used to smooth the acceleration and angular velocity data after amplitude limiting and filtering in order to filter out high-frequency jitter noise.

[0016] The adaptive adjustment process is as follows: the microprocessor calculates the fast Fourier transform spectrum of the acceleration signal mode within the sliding window in real time, takes the frequency value corresponding to the highest energy peak in the spectrum as the main motion frequency, and sets the cutoff frequency of the low-pass filter to a preset multiple of the main motion frequency.

[0017] Preferably, in step S2, the lightweight convolutional neural network model is a compressed and quantized one-dimensional convolutional neural network;

[0018] The categories of military tactical movements include at least stationary movement, walking, running, sprinting, low crawling, side crawling, crouching leap, crawling leap, obstacle crossing, lying down, and using cover;

[0019] The extracted time-domain and frequency-domain features include: the mean and variance of the three-dimensional acceleration vector and magnitude, the integral value of the angular velocity around the coronal and sagittal axes of the human body, the heart rate value, and the slope of the heart rate value relative to the soldier's individual resting heart rate per unit time.

[0020] The individual resting heart rate is obtained by analyzing the soldier's heart rate data during a specific resting period before the start of training.

[0021] Preferably, in step S3, the selection strategy for dynamically selecting the main edge computing node is as follows: the personal edge computing terminal continuously listens to the broadcast information of each edge computing node and constructs a candidate node list that includes signal reception strength, computing load rate, and available battery power.

[0022] For each node in the list, a connection stability score is calculated based on the signal reception strength, and a service sustainability score is calculated based on the load factor and available battery power.

[0023] The connection stability score and the service sustainability score are weighted and summed according to preset weights to obtain a comprehensive optimal score.

[0024] The personal edge computing terminal periodically selects the edge computing node with the highest comprehensive selection score as the main edge computing node, and performs a switch when the current connected node is different from the optimal node.

[0025] Preferably, in step S4, the personalized physical fitness assessment model is a long short-term memory network that incorporates attention mechanisms;

[0026] The contextual features include the group average heart rate of all soldiers in the current training unit, the distribution ratio of the group's main training movements, and the group average movement standardization score.

[0027] The attention mechanism is used to assign different weights to the target soldier's local feature data and different contextual features;

[0028] The instantaneous load intensity output by the model is obtained by mapping the basal metabolic equivalent value, current heart rate value, and soldier's personal maximum heart rate value corresponding to the current action category through a piecewise linear function.

[0029] The action standard score is obtained by calculating the dynamic time warping distance between the feature vector of the current action time window and the feature vector of the pre-stored standard action template for the action category, and mapping the distance to a percentage score.

[0030] The pre-stored standard action template feature vectors are derived from a feature library established after collecting and analyzing standard demonstration actions by professional coaches.

[0031] Preferably, the calculation process of the fatigue coefficient specifically includes:

[0032] The personalized physical fitness assessment model first calculates the heart rate recovery slope based on heart rate data. The heart rate recovery slope is defined as the rate of heart rate decline within a specific time period after the activity is detected to have switched from a high-intensity category to a low-intensity or resting category.

[0033] The model also calculates the degree of motion distortion based on the changing trend of the motion standard score. The degree of motion distortion is defined as the decay ratio of the current motion standard score relative to the historical best score of the motion in this training unit.

[0034] Finally, the heart rate recovery slope and the degree of motion deformation are input into a trained multilayer perceptron model, which outputs a standardized fatigue coefficient ranging from 0 to 1.

[0035] Preferably, the calculation process of the short-term injury risk index specifically includes: the personalized physical fitness assessment model continuously monitors abnormal movement patterns in the local feature dataset, wherein the abnormal movement pattern is defined as a signal segment in which the integral value of angular velocity exceeds the threshold of physiological joint range of motion or the acceleration modulus changes drastically in a very short time.

[0036] The model statistically analyzes the frequency and average intensity of abnormal movement patterns within the current training unit, and combines this with the fatigue coefficient and the instantaneous load intensity to calculate a risk index representing the probability of sports injury occurring in the near future using a logistic regression classifier.

[0037] Preferably, in step S5, the specific rules for the real-time tactile or audiovisual feedback provided by the personal edge computing terminal are as follows: when the instantaneous load intensity exceeds the personalized warning threshold set based on the soldier's historical data, the terminal emits continuous vibration and voice prompts "load too high";

[0038] When the motion accuracy score is lower than the preset passing score, the terminal emits a short vibration and displays an icon indicating "motion distortion".

[0039] When the fatigue coefficient exceeds the preset threshold, the terminal issues a voice suggestion: "Adjust the rhythm."

[0040] When the short-term injury risk index exceeds the high-risk threshold, the terminal issues a strong warning to the soldier, and at the same time, the warning information is pushed to the commander's command terminal with high priority.

[0041] Preferably, in step S5, the anonymized aggregated evaluation results pushed to the commander's command terminal specifically include the real-time average load intensity distribution of different areas in the training ground displayed in the form of a heat map, the anonymized identifiers of high-risk soldiers and their main risk types updated in real time in the form of a list, and the change curve of the group average fatigue coefficient in the entire training unit displayed in the form of a trend graph.

[0042] Preferably, before step S1, there is a model initialization and personalized adaptation stage, the specific steps of which are: when a new soldier uses the system for the first time, he is guided to complete a series of standard military tactical actions under supervision, and corresponding sensor data is collected to fine-tune the parameters of the last fully connected layer related to action classification in the lightweight convolutional neural network model.

[0043] Simultaneously, physiological data of the soldier under resting and extreme exercise conditions are collected to calibrate personalized parameters related to the individual's maximum heart rate, resting heart rate, basal metabolic equivalent, and range of motion in the personalized physical fitness assessment model.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention continuously collects data from soldiers during their natural training state using embedded wearable devices, and performs local real-time action recognition and feature extraction on a personal edge computing terminal, achieving a seamless evaluation process. The key innovation lies in the two-level collaborative processing architecture of "personal terminal-edge node," which deploys the computationally intensive personalized evaluation model at edge nodes close to the training ground. This significantly reduces data transmission latency and bandwidth requirements, reducing the latency of the entire process from data collection and analysis to the generation of evaluation indicators and feedback to the individual soldier's terminal to the second level. This allows soldiers to receive immediate feedback on load and action standard after performing training actions, and commanders can also grasp the overall training situation in real time, forming a real-time closed loop of "monitoring-evaluation-feedback," greatly improving the timeliness and scientific rigor of training guidance.

[0046] 2. This invention abandons the traditional architecture that relies on a stable cloud center. Through the mobile deployment of edge computing nodes and a strategy of dynamic, optimal connection selection by personal terminals based on signal strength and node load, a highly robust distributed processing network is constructed. Even in the event of a public network outage, the system can still operate stably relying on the local area network and edge computing power. Simultaneously, local lightweight model processing reduces the amount of raw data uploaded, and group characteristics are aggregated at edge nodes for contextual analysis, ensuring uninterrupted core evaluation functions in weak network environments and guaranteeing continuous availability and stable output of evaluation results in real tactical training environments.

[0047] 3. This invention does not simply apply a general activity recognition model, but rather customizes a recognition model for military training that includes multiple tactical movement categories. It innovatively incorporates the group training context as one of the input features of the personalized evaluation model. Through attention mechanism fusion analysis, the output indicators such as load intensity and fatigue coefficient are more closely aligned with the actual situation of squad-level collaborative training. In particular, by quantifying movement standardization through a dynamic time warping algorithm and comprehensively calculating a short-term injury risk index by combining abnormal movement pattern frequency, heart rate recovery slope, and other multi-dimensional data, it achieves a deeper, more accurate, and forward-looking assessment, moving beyond simple "whether it was completed" to "how well it was completed" and "whether there is a risk of injury." This provides a direct and quantitative basis for scientific training and training injury prevention. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0050] Figure 2 This is a flowchart illustrating the steps involved in calculating the fatigue coefficient in the method of this invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0052] Please see Figures 1-2 This invention provides a non-intrusive real-time military physical fitness assessment method based on edge computing. Step S1 involves a wearable sensing device continuously collecting raw acceleration, angular velocity, geomagnetic direction, and heart rate data at a sampling frequency of 200 Hz. The device's built-in microprocessor uses a sliding window amplitude-limiting filter to remove impulse noise. Specifically, a window of 50 sampling points is set, and the difference between a new data point and the median value within the window is calculated. If the difference exceeds a dynamic threshold based on 5% of the sensor's range, the median value is used instead. Subsequently, a fourth-order Butterworth low-pass filter is used for smoothing, with the filter cutoff frequency adaptively adjusted according to the main motion frequency calculated in real time. The preprocessed data is packaged into structured data frames containing timestamps and device identifiers and sent to a personal edge computing terminal via Bluetooth Low Energy.

[0053] In step S2, the personal edge computing terminal inputs 9-channel sensor data into a lightweight convolutional neural network model with a time window of 0.5 seconds. This model contains 3 convolutional layers and 2 pooling layers, and outputs tactical action categories such as crouching leaps. Simultaneously, features such as the mean of the acceleration modulus, the integral of the angular velocity, heart rate, and the slope of heart rate rise are extracted to form a local feature dataset.

[0054] In step S3, the edge computing nodes broadcast calculations of load rate, battery level, and signal strength every 5 seconds. Individual terminals calculate a comprehensive score based on signal strength and load, selecting the node with the highest score as the master edge computing node and uploading its local feature dataset. The master node aggregates data from multiple terminals to form a group time-series feature dataset.

[0055] Step S4: The master node runs a long short-term memory network model incorporating an attention mechanism, inputting local characteristics of the target soldiers and contextual features such as average heart rate and movement distribution, and outputs instantaneous load intensity, movement standardization score, fatigue coefficient, and short-term injury risk index. Step S5: The master node distributes the indicators to individual terminals, providing feedback via vibration and voice, while simultaneously pushing a group heatmap and a high-risk list to the commander's terminal. This implementation method achieves real-time and accurate physical fitness assessment in a network-free field environment, forming a closed-loop training and assessment system.

[0056] In one possible implementation, the amplitude limiting and low-pass filtering processes in step S1 are specifically implemented as follows: The microprocessor of the wearable sensing device maintains a sliding window with a length of 50 sampling points, corresponding to 0.25 seconds of data. For each newly acquired acceleration data point, the absolute difference between that point and the median of the previous 49 data points within the sliding window is calculated. A preset dynamic threshold is 5% of the sensor's full range; for example, for a sensor with a range of ±16 times gravitational acceleration, the dynamic threshold is set to 1.6 times gravitational acceleration. If the difference exceeds the dynamic threshold, it is determined to be impulse noise, and the median in the window is used to replace the current point. After amplitude limiting filtering is completed, low-pass filtering is performed.

[0057] The microprocessor performs a Fast Fourier Transform on the acceleration modulus data from the most recent second, identifying the frequency corresponding to the highest energy peak in the spectrum as the primary motion frequency. The cutoff frequency of a fourth-order Butterworth low-pass filter is set to 2.5 times the primary motion frequency; for example, if the primary motion frequency is detected to be 2 Hz, the cutoff frequency is set to 5 Hz. This filter, using its cutoff frequency, smooths the acceleration and angular velocity data. This filtering method effectively removes equipment collision noise and high-frequency jitter, improving data quality and laying the foundation for subsequent motion recognition and feature extraction, thus enhancing the system's robustness in environments with severe motion.

[0058] In one possible implementation, the lightweight convolutional neural network model and feature extraction in step S2 are specifically implemented as follows: The lightweight convolutional neural network deployed on the personal edge computing terminal is a one-dimensional convolutional structure. After pruning and 8-bit integer quantization optimization, the model size is controlled within 2 megabytes. The model takes 100 data points from 9 channels within a 0.5-second time window as input. After processing through 3 convolutional layers with ReLU activation function and 2 max pooling layers, the fully connected layer outputs the probabilities of 12 types of military tactical actions. The action categories include stationary, walking, running, sprinting, low crawling, side crawling, crouching leap, crawling leap, obstacle crossing, lying down, and using cover.

[0059] Model inference takes less than 20 milliseconds. Feature extraction is performed simultaneously, calculating the mean and variance of the three-dimensional acceleration vector and its magnitude to reflect motion intensity and stability. The integral of the angular velocity around the coronal and sagittal axes is calculated to estimate joint range of motion. Heart rate values ​​are read, and their rate of ascent relative to the individual's resting heart rate (obtained by analyzing 5 minutes of sedentary data prior to training) is calculated. The extracted feature vectors are associated with action labels to form a local feature dataset. This implementation achieves efficient and accurate motion recognition and feature generation locally on the terminal, reducing data upload volume and ensuring real-time performance and privacy.

[0060] In one possible implementation, the optimization strategy for dynamically selecting the primary edge computing node in step S3 is specifically implemented as follows: Each edge computing node periodically broadcasts a data packet containing the node identifier, computing load rate, available battery percentage, and received signal strength indicator value. The personal edge computing terminal listens to the broadcast and calculates a connection stability score and a service sustainability score for each candidate node. The connection stability score is linearly mapped to the received signal strength indicator value within the range of 0 to 1; the stronger the signal, the higher the score. The service sustainability score is obtained by calculating a weighted combination of load rate and battery percentage; the lower the load rate and the higher the battery percentage, the higher the score.

[0061] The specific weighting is set at 60% load rate and 40% battery power. The overall selection score is calculated by multiplying the connection stability score by a weight of 0.6 and the service sustainability score by a weight of 0.4. The personal terminal recalculates every 10 seconds and selects the edge computing node with the highest overall score as the primary edge computing node. If the currently connected node is not the optimal node, a smooth switch is performed. This strategy ensures that the personal terminal is always connected to an edge node with stable signal and sufficient computing resources, improving data transmission reliability and overall system stability, and adapting to dynamic environmental changes in the training environment.

[0062] In one possible implementation, the personalized physical fitness assessment model in step S4 is specifically implemented as follows: The personalized physical fitness assessment model loaded by the main edge computing node is a long short-term memory network incorporating an attention mechanism. The model input includes the target soldier's local feature dataset sequence for the most recent 30 seconds, and contextual features extracted from the group temporal feature dataset. The contextual features include the average heart rate of all soldiers in the current training unit, the distribution ratio of the main training movements, and the average movement standard score. The attention mechanism assigns variable weights to different input features; for example, when the movement category is sprinting, the weight of the heart rate feature increases.

[0063] The model outputs instantaneous load intensity, calculated as a piecewise linear function of the current basal metabolic equivalent, real-time heart rate, and individual maximum heart rate. The movement standardization score is obtained by calculating the dynamic time-warped distance between the current movement feature vector and the feature vector of a pre-stored standard movement template, converting the distance into a percentage score. The standard movement template feature vector is derived from a database of demonstration movements by professional instructors. By integrating individual and group information, this model provides a more accurate assessment that closely reflects real-world combat scenarios, enhancing the military specificity and scientific rigor of the assessment results.

[0064] In one possible implementation, the fatigue coefficient calculation process in step S4 is specifically implemented as follows: The personalized physical fitness assessment model first calculates the heart rate recovery slope. When the model detects a switch from a high-intensity category to a low-intensity or resting category, it extracts heart rate data for the following 30 seconds, fits a heart rate decline curve, and calculates the heart rate decline per unit time as the heart rate recovery slope. Simultaneously, the model calculates the degree of movement deformation by comparing the current movement standard score with the soldier's historical best score for the same movement within this training unit, and calculates the percentage of score decay.

[0065] After normalizing the heart rate recovery slope and the degree of movement deformation, the data is input into a multilayer perceptron model with two hidden layers. This multilayer perceptron model is trained using historical training data and outputs a standardized fatigue coefficient between 0 and 1, with the value closer to 1 indicating a higher degree of fatigue. This calculation process comprehensively considers both physiological recovery and movement quality degradation indicators, enabling a quantitative assessment of soldiers' fatigue status and providing a direct basis for preventing overtraining.

[0066] In one possible implementation, the calculation of the short-term injury risk index in step S4 is specifically implemented as follows: The personalized physical fitness assessment model continuously monitors the local feature dataset and defines abnormal movement patterns as signal segments whose angular velocity integral value exceeds the physiological joint range of motion threshold, or whose acceleration modulus changes at a rate exceeding a preset threshold within 10 milliseconds. The physiological joint range of motion threshold is set based on human anatomical data; for example, the upper limit of the knee joint flexion-extension angular velocity integral is 120 degrees per second. The model statistically analyzes the frequency of abnormal movement patterns occurring within the current training unit, i.e., the number of occurrences per minute, and the average intensity, i.e., the mean amplitude of the abnormal signal.

[0067] Abnormal frequency, average intensity, real-time fatigue coefficient, and instantaneous load intensity are used as input features to a logistic regression classifier. This classifier, trained using historical injury data, outputs a short-term injury risk index between 0 and 1, representing the probability of sports injury occurring within the next 15 minutes. This calculation process enables proactive early warning of injury risk based on multi-source data, facilitating timely intervention and reducing the incidence of training injuries.

[0068] In one possible implementation, the specific rules for real-time tactile or audiovisual feedback in step S5 are as follows: The personal edge computing terminal triggers different feedback based on the received real-time physical fitness indicators. A personalized warning threshold is set, for example, the instantaneous load intensity threshold is 80% of the soldier's historical maximum load. When the instantaneous load intensity exceeds the threshold, the terminal generates continuous vibration and plays the voice message "Load Too High". When the action standard score is below 75 points, the terminal generates a short vibration and displays a "Movement Distortion" icon on the screen.

[0069] When the fatigue coefficient exceeds 0.6, the terminal plays a voice prompt suggesting "adjust the pace." When the short-term injury risk index exceeds 0.2, the terminal emits a strong warning vibration and a flashing red icon, simultaneously marking the warning information as high priority and sending it to the commander's terminal. The feedback rules are differentiated according to different indicator levels to ensure soldiers receive clear and timely operational guidance, while immediately reporting high-risk situations to guarantee training safety.

[0070] In one possible implementation, the anonymized aggregated evaluation results pushed to the commander's command terminal in step S5 specifically include the following parts: The main edge computing node generates a heat map of the training ground, dividing the ground into grids, calculating the average instantaneous load intensity of all soldiers within each grid, and visualizing it using a color gradient. A list of high-risk soldiers is generated, containing anonymous identifiers for the soldiers, such as replacing their names with number S01, and listing their main risk types, such as "frequent abnormal movements" or "excessively high fatigue coefficient." A group fatigue trend curve is generated, plotting the change in the average fatigue coefficient of all soldiers with time as the horizontal axis.

[0071] All data is updated every 5 seconds and pushed to the commander's terminal. This aggregated result presents the overall training status in an intuitive way, enabling commanders to quickly grasp the training intensity distribution, identify individuals at risk, and assess overall fatigue progress, supporting scientific training and real-time decision-making.

[0072] In one possible implementation, the model initialization and personalized adaptation phase is specifically implemented as follows: When a new soldier uses the system for the first time, he is guided to complete a set of preset military tactical movements under the supervision of an instructor in a standard training area, including 10 movements such as walking, running, and low crawling. The system collects complete sensor data for each movement and uses this data to fine-tune the fully connected layer parameters of the lightweight convolutional neural network model. The fine-tuning employs transfer learning and backpropagation algorithms to update the weights to adapt to the individual movement characteristics of the soldier.

[0073] Simultaneously, the system collects the soldier's heart rate data after 5 minutes of sitting still, taking the average of the following 3 minutes as the individual's resting heart rate. The system also obtains the soldier's maximum heart rate by reviewing historical physical fitness test records. Furthermore, it collects data on the soldier's joint range of motion under standard movement conditions. These personalized parameters are encrypted and stored on the individual's edge computing terminal and edge computing nodes. This stage ensures that the assessment model matches the soldier's individual physiological characteristics and movement habits, providing a reliable personalized benchmark for subsequent accurate assessments and improving the accuracy and effectiveness of the assessment results.

[0074] Example

[0075] This embodiment uses a comprehensive "squad tactical infiltration and assault" training exercise organized by a certain unit at a field tactical training ground as the application scenario. The training ground simulates a real battlefield environment, including various obstacles such as gravel terrain, trenches, low walls, barbed wire, and log bridges, and lacks stable public network coverage. A participating infantry squad of 12 soldiers must complete a series of tactical maneuvers within a specified time, including rapid marching, crouching and leaping across open ground, crawling low over barbed wire, crossing low walls, and firing from cover.

[0076] 1. System composition and initialization:

[0077] The physical equipment required to implement the method in this embodiment includes:

[0078] Wearable sensing devices: 12 sets, each integrating a nine-axis inertial measurement unit (including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer) and a wrist-worn optical heart rate sensor. The sampling frequency of the devices is set to 200Hz.

[0079] Personal edge computing terminals: 12 units, which are ruggedized military smartphones equipped with high-performance, low-power processors.

[0080] Edge computing nodes: 3 units, using ruggedized vehicle-mounted mobile servers, equipped with edge computing platform software, and mobilely deployed at high points in three different areas of the training ground.

[0081] Commander's command terminal: 1 unit, a ruggedized tablet computer for commander's use.

[0082] Before training begins, perform system initialization and personalized adaptation:

[0083] New recruit Zhang San used the system for the first time. Under the guidance of his instructor, Zhang San performed 10 standard tactical movements in the standard movement assessment area of ​​the training ground, including walking, running, low crawling, and crouching leaps. The system collected complete sensor data for these movements. This data was used to fine-tune the weight parameters of the last fully connected layer related to movement classification in the lightweight convolutional neural network model pre-installed on Zhang San's personal edge computing terminal, making the model more adaptable to Zhang San's personal movement habits. Simultaneously, the system recorded Zhang San's average heart rate after sitting still for 5 minutes as his personal resting heart rate (68 beats / minute) and obtained his personal maximum heart rate of 192 beats / minute through historical physical fitness test data. These personalized physiological parameters were encrypted and stored on his personal terminal and synchronized to the edge computing node for subsequent personalized calculations to evaluate the model.

[0084] 2. Specific implementation of the method and steps:

[0085] Step S1: Non-invasive acquisition and preprocessing of multimodal physiological and motor parameters.

[0086] After the training began, the wearable sensing device worn by Zhang San started continuously collecting raw data at a frequency of 200Hz. The device's built-in microprocessor performed real-time preprocessing on the raw data stream.

[0087] First, amplitude limiting filtering is performed: the microprocessor maintains a sliding window with a length of 50 sampling points (corresponding to 0.25 seconds of data). For a newly acquired acceleration data point, the difference between it and the median of the previous 49 data points within the sliding window is calculated. A dynamic threshold is preset, which is set based on 5% of the accelerometer's range. If the difference exceeds this dynamic threshold, it is determined to be impulse noise, and the median value is used to replace the current acquisition point.

[0088] Then, adaptive low-pass filtering is performed: the microprocessor calculates the vector magnitude of the acceleration data after amplitude-limiting filtering. A Fast Fourier Transform is performed on the acceleration magnitude data (200 points) from the most recent second to obtain a spectrum. The peak with the highest energy in the spectrum is identified, with a corresponding frequency of 2.1Hz, which is Zhang San's current main motion frequency. According to preset rules, the cutoff frequency of the fourth-order Butterworth low-pass filter is set to 2.5 times this main motion frequency, i.e., 5.25Hz. Using this filter with the cutoff frequency, the acceleration and angular velocity data are smoothed, effectively filtering out high-frequency noise caused by muscle tremors and equipment loosening.

[0089] Finally, the microprocessor packages the preprocessed sensor data with timestamps and device IDs, and sends it in real time via Bluetooth Low Energy to Zhang San's ruggedized smartphone (i.e., personal edge computing terminal).

[0090] Step S2: Real-time tactical action recognition and feature extraction based on a personal edge computing terminal. Zhang San's personal edge computing terminal continuously receives Bluetooth data. The application running on the terminal accumulates data (100 sampling points) every 0.5 seconds as a processing window.

[0091] The sensor data (three-axis acceleration, three-axis angular velocity, and three-axis magnetism) from the nine channels within the current window are organized into a 9x100 matrix and input into a lightweight convolutional neural network model that has been fine-tuned. This model employs a one-dimensional convolutional structure, containing three convolutional layers and two pooling layers, finally connected to a fully connected layer and a Softmax classifier. Local inference takes approximately 15 milliseconds, and the model outputs a probability distribution for the action category. It identifies Zhang San's current action as a "bending leap," with a confidence level of 94%.

[0092] Meanwhile, the terminal extracts a set of features from the same window of data: the mean of the magnitude of the vector sum of the triaxial accelerations is 12.5 m / s², and the variance is 8.2; the integral of the angular velocity around the sagittal axis of the body is calculated, and the change in the forward tilt angle of the trunk is approximately 30 degrees; the current heart rate is read as 158 beats / minute; and the slope of the heart rate relative to the individual's resting heart rate (68 beats / minute) is calculated to be 90 beats / minute.

[0093] The terminal associates the action tag "bending forward" with the feature vector [12.5, 8.2, 30, 158, 90] to form a local feature dataset record.

[0094] Step S3: Dynamic selection and data aggregation of collaborative edge computing nodes.

[0095] Three edge computing nodes deployed on-site broadcast a status packet every 5 seconds, including the node ID, current CPU load rate, remaining battery percentage, and signal strength indicators for each terminal. Zhang San's personal terminal listens to these broadcasts. Assume that at a certain moment, the terminal receives information from the three nodes: Node A has a load rate of 30%, battery level of 80%, and signal strength of -65dBm; Node B has a load rate of 70%, battery level of 90%, and signal strength of -70dBm; Node C has a load rate of 10%, battery level of 50%, and signal strength of -55dBm. The terminal executes an optimal selection strategy: First, it calculates the connection stability score for each node, with stronger signal strength resulting in a higher score. The signal strength values ​​are normalized to a score of 0-1, with Node C scoring the highest (-55dBm corresponds to 0.9 points). Second, it calculates the service sustainability score, considering both load rate (lower is better) and battery level (higher is better). Node A scores higher (0.7 points), while Node C scores lower (0.4 points) due to its low battery level. Finally, the comprehensive selection score is calculated according to preset weights (connection stability weight 0.6, service sustainability weight 0.4): Node A: 0.6×0.3+0.7×0.4=0.46; Node C: 0.9×0.6+0.4×0.4=0.7. Therefore, Node C has the highest score. Zhang San's personal terminal chooses to establish and maintain a connection with Node C, and begins uploading the local feature dataset records generated in step S2 to Node C at a frequency of 2 records per second.

[0096] Node C simultaneously receives similar data uploaded from the terminals of four other soldiers. The service program on Node C aggregates this data into a time-series feature dataset organized by soldier ID and timestamp.

[0097] Step S4: Real-time calculation and evaluation of personalized physical fitness indicators based on group context.

[0098] A personalized physical fitness assessment model runs on node C. This model uses Zhang San's local feature data as its core input.

[0099] The model first extracts contextual features from the group time-series feature dataset: the average heart rate of the four soldiers in the same tactical group as Zhang San is calculated to be 162 beats / minute; the main training movements of the group in the past minute are "bending forward" (40%) and "running" (35%); the average movement standard score of the group is calculated to be 78 points.

[0100] Next, the model inputs Zhang San's individual feature vector along with the group context feature vector into its core. This core is a long short-term memory network incorporating an attention mechanism. The attention mechanism assigns weights to different input features; for example, in the current "bending leap" movement, Zhang San's individual angular velocity integral feature (reflecting the amplitude of the movement) and the group's average heart rate (reflecting overall intensity) are given higher weights. The model then performs comprehensive calculations and outputs real-time physical fitness indicators:

[0101] Instantaneous load intensity: Based on the basal metabolic equivalent of "bending forward leap", Zhang San's current heart rate of 158 beats / minute and his personal maximum heart rate of 192 beats / minute, an intensity value of 7.5 (range 0-10) is obtained through a piecewise linear function mapping.

[0102] Action standardization score: Dynamic time warping is performed on Zhang San's current feature vector and the feature vector of the standard action template of "bending forward" (from the professional coach database) to obtain the difference distance, which is then converted into a percentage score, resulting in 85 points.

[0103] Fatigue coefficient: The model retrieved Zhang San's heart rate data from the past 5 minutes and found that during the previous resting interval, his heart rate dropped from 170 beats / minute to 150 beats / minute within 30 seconds, calculating the heart rate recovery slope. Simultaneously, the degree of deformation in his "bending forward leap" movement was calculated: the current score was 85 points, while his best score during this training was 88 points, representing a decay rate of approximately 3.4%. The heart rate recovery slope and the degree of movement deformation were input into a specialized multilayer perceptron model, which outputs a standardized fatigue coefficient of 0.65 (range 0-1, with higher values ​​indicating greater fatigue).

[0104] Short-term injury risk index: Model analysis of Zhang San's local feature data revealed three abnormal movement patterns when his left foot landed within the past two minutes (manifested as a sharp change in vertical acceleration exceeding a threshold within 5 milliseconds). The frequency and intensity of these abnormalities were statistically analyzed, and combined with the current fatigue coefficient of 0.65 and load intensity of 7.5, input into a logistic regression classifier to calculate a short-term injury risk index of 0.15 (range 0-1, representing the probability of risk).

[0105] Step S5: Low-latency feedback and decision support of evaluation results.

[0106] Node C immediately transmits Zhang San's calculated physical fitness indicators (load intensity 7.5, movement standard 85, fatigue coefficient 0.65, risk index 0.15) back to Zhang San's personal terminal. Zhang San's personal terminal triggers feedback according to preset rules: since the fatigue coefficient of 0.65 has exceeded his personal warning threshold of 0.6, the terminal emits a short vibration, displays a yellow exclamation mark icon on the edge of the screen, and simultaneously announces, "Adjusting the pace is recommended."

[0107] Meanwhile, Node C anonymizes and aggregates the assessment results of all 12 soldiers: generating a heat map of the training ground showing the real-time average load intensity of different areas; generating a high-risk list showing that two soldiers (anonymous numbers S05 and S09) have a short-term injury risk index greater than 0.2, with the main risk type being "frequent abnormal movements"; and generating a group fatigue trend curve. This aggregated information is pushed in real time to the commander's ruggedized tablet computer (commander's command terminal). The commander can use the tablet computer to have a global grasp of the training intensity distribution and focus on the status of soldiers S05 and S09, providing individual reminders or adjusting their tasks via radio.

[0108] To demonstrate the effectiveness of the method in this embodiment, it is compared with two existing typical methods under the same training scenario. Comparative Example 1 is the traditional "manual timing and counting evaluation method," and Comparative Example 2 is an automated evaluation method based on "cloud-based central processing," which transmits all sensor data back to a remote cloud server in real time via 4G / 5G network for processing and evaluation.

[0109] Comparison indicators Comparative Example 1: Manual Timekeeping and Counting Evaluation Method Comparative Example 2: Cloud-based Centralized Processing Method Method of this invention Real-time assessment The results are severely delayed, and statistical analysis and subjective scoring can only be performed after the training is completed. The latency is high and highly dependent on network conditions. When the network is unstable, the end-to-end latency exceeds 5 seconds, making it impossible to provide real-time guidance for actions. Extremely low latency (<1 second). Individual action recognition is completed locally on the terminal (millisecond level), and overall evaluation is completed at the edge node (second level), with timely feedback. Assessment of coverage and non-intrusiveness It can only assess timed and countable endpoint results (such as total time, number of rings scored), but cannot assess the quality of the process movements. The assessment process requires close supervision from the examiner, which interferes with training. It can achieve full-process data acquisition, but it relies on a stable high-speed network and cannot work in areas with no or weak network in the wild, where the system "fails". Seamless assessment throughout the entire training process. It operates stably in complex field environments without relying on a fixed network, placing no additional burden on soldiers. Assess accuracy and specificity It is highly subjective, and the standardization and economy of tactical actions cannot be quantitatively evaluated, resulting in low precision. Complex models can be used, but they lack accuracy in recognizing military tactical maneuvers. The models are general and lack personalization, and the evaluation results are out of touch with the needs of military training. It is highly specific to military applications and has high precision. The model is optimized for tactical maneuvers and integrates individual physiological data with group context to output accurate indicators that are relevant to actual combat, such as load intensity, maneuver standard, and tactical fatigue coefficient. System availability and reliability It relies on human intervention and is greatly affected by weather and the examiner's condition, but it is simple and direct. It is highly dependent on cloud networks and computing centers; if the network is interrupted, the system will be paralyzed, and its reliability is poor in field scenarios. High availability and reliability. Utilizing an edge computing architecture, it has low network requirements, allows collaboration between personal terminals and edge nodes, enables some functions to be available offline, and provides strong system robustness. Training and assessment closed-loop support It only provides the final result, lacks real-time feedback during the training process, and cannot support dynamic adjustments. It can provide remote feedback, but it is subject to high latency and may be interrupted, limiting its effectiveness in providing real-time guidance to soldiers; the commander's view is usually lagging. This forms a complete real-time closed loop. It provides layered, real-time, and precise feedback and early warnings to soldiers (for real-time fine-tuning of actions) and commanders (for macro-level decision-making), significantly improving the scientific nature of training.

[0110] The comparison tables above clearly demonstrate that, in the specific application scenario of military field training, the method of this invention significantly outperforms existing traditional methods and technical solutions in multiple key dimensions, including real-time performance, environmental adaptability, assessment accuracy, system reliability, and training and assessment support effectiveness.

[0111] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0112] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A non-intrusive real-time military physical fitness assessment method based on edge computing, characterized in that, Includes the following steps: Step S1: Using a wearable sensing device that integrates a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer, and an optoelectronic heart rate sensor, the soldier's raw acceleration data, angular velocity data, geomagnetic direction data, and heart rate pulse signal are collected at a sampling frequency of not less than 100Hz; the wearable sensing device performs preprocessing on the raw sensing data stream, including amplitude limiting filtering and low-pass filtering, and packages the preprocessed multi-channel sensor data into structured data frames and sends them to the personal edge computing terminal; Step S2: The personal edge computing terminal receives structured data frames and inputs multi-sensor data within a fixed time window into a locally deployed lightweight convolutional neural network model for real-time inference to identify military tactical action categories; the personal edge computing terminal simultaneously extracts the temporal and frequency domain features of the data within the time window to generate a local feature dataset containing action category labels and physiological motion feature vectors. Step S3: Edge computing nodes deployed in the training area periodically broadcast their own computing load rate, available battery power, and signal reception strength; A personal edge computing terminal dynamically selects a master edge computing node and establishes a connection based on broadcast information, and uploads its local feature dataset to the master edge computing node; the master edge computing node aggregates the data uploaded by all associated personal edge computing terminals to form a group time-series feature dataset; Step S4: The main edge computing node loads the personalized physical fitness assessment model; the input of the personalized physical fitness assessment model includes the target soldier's local feature dataset and contextual features extracted from the group's temporal feature dataset; the personalized physical fitness assessment model outputs quantified real-time physical fitness indicators, which include at least instantaneous load intensity, movement standard score, fatigue coefficient and short-term injury risk index. Step S5: The main edge computing node sends the real-time physical fitness indicators back to the corresponding personal edge computing terminal, and provides real-time tactile or audiovisual feedback to the soldiers through the personal edge computing terminal; at the same time, the main edge computing node pushes the anonymized aggregated evaluation results of all soldiers to the commander's command terminal.

2. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S1, the specific processes of amplitude limiting filtering and low-pass filtering are as follows: The microprocessor built into the wearable sensing device first uses a sliding window-based amplitude limiting filtering method to compare the sensor data of the current sampling point with the median of the previous data points in the sliding window. If the difference exceeds a dynamic threshold preset based on the sensor range, the median value is used to replace the current sampling point data to remove impulse noise. Next, a fourth-order Butterworth low-pass filter with a cutoff frequency that is adaptively adjusted according to the main movement frequency range of the soldier is used to smooth the acceleration and angular velocity data after amplitude limiting and filtering in order to filter out high-frequency jitter noise. The adaptive adjustment process is as follows: the microprocessor calculates the fast Fourier transform spectrum of the acceleration signal mode within the sliding window in real time, takes the frequency value corresponding to the highest energy peak in the spectrum as the main motion frequency, and sets the cutoff frequency of the low-pass filter to a preset multiple of the main motion frequency.

3. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S2, the lightweight convolutional neural network model is a compressed and quantized one-dimensional convolutional neural network; The categories of military tactical movements include at least stationary movement, walking, running, sprinting, low crawling, side crawling, crouching leap, crawling leap, obstacle crossing, lying down, and using cover; The extracted time-domain and frequency-domain features include: the mean and variance of the three-dimensional acceleration vector and magnitude, the integral value of the angular velocity around the coronal and sagittal axes of the human body, the heart rate value, and the slope of the heart rate value per unit time relative to the soldier's individual resting heart rate. The individual resting heart rate is obtained by analyzing the soldier's heart rate data during a specific resting period before the start of training.

4. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S3, the selection strategy for dynamically selecting the main edge computing node is as follows: the personal edge computing terminal continuously listens to the broadcast information of each edge computing node and constructs a candidate node list that includes signal reception strength, computing load rate, and available battery power. For each node in the list, a connection stability score is calculated based on the signal reception strength, and a service sustainability score is calculated based on the load factor and available battery power. The connection stability score and the service sustainability score are weighted and summed according to preset weights to obtain a comprehensive optimal score. The personal edge computing terminal periodically selects the edge computing node with the highest comprehensive selection score as the main edge computing node, and performs a switch when the current connected node is different from the optimal node.

5. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S4, the personalized physical fitness assessment model is a long short-term memory network that incorporates attention mechanisms; The contextual features include the group average heart rate of all soldiers in the current training unit, the distribution ratio of the group's main training movements, and the group average movement standardization score. The attention mechanism is used to assign different weights to the target soldier's local feature data and different contextual features; The instantaneous load intensity output by the model is obtained by mapping the basal metabolic equivalent value, current heart rate value, and soldier's personal maximum heart rate value corresponding to the current action category through a piecewise linear function. The action standard score is obtained by calculating the dynamic time warping distance between the feature vector of the current action time window and the feature vector of the pre-stored standard action template for the action category, and mapping the distance to a percentage score. The pre-stored standard action template feature vectors are derived from a feature library established after collecting and analyzing standard demonstration actions by professional coaches.

6. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 5, characterized in that, The calculation process of the fatigue coefficient specifically includes: The personalized physical fitness assessment model first calculates the heart rate recovery slope based on heart rate data. The heart rate recovery slope is defined as the rate of heart rate decline within a specific time period after the activity is detected to have switched from a high-intensity category to a low-intensity or resting category. The model also calculates the degree of motion distortion based on the changing trend of the motion standard score. The degree of motion distortion is defined as the decay ratio of the current motion standard score relative to the historical best score of the motion in this training unit. Finally, the heart rate recovery slope and the degree of motion deformation are input into a trained multilayer perceptron model, which outputs a standardized fatigue coefficient ranging from 0 to 1.

7. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 5, characterized in that, The calculation process of the short-term injury risk index specifically includes: the personalized physical fitness assessment model continuously monitors abnormal movement patterns in the local feature dataset, and the abnormal movement pattern is defined as a signal segment in which the integral value of angular velocity exceeds the threshold of physiological joint range of motion or the acceleration modulus changes drastically in a very short time. The model statistically analyzes the frequency and average intensity of abnormal movement patterns within the current training unit, and combines the fatigue coefficient and the instantaneous load intensity to calculate a risk index representing the probability of sports injury occurring in the near future using a logistic regression classifier.

8. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S5, the specific rules for the real-time tactile or audiovisual feedback provided by the personal edge computing terminal are as follows: when the instantaneous load intensity exceeds the personalized warning threshold set based on the soldier's historical data, the terminal will issue a continuous vibration and a voice prompt "load too high"; When the motion accuracy score is lower than the preset passing score, the terminal emits a short vibration and displays an icon indicating "motion distortion". When the fatigue coefficient exceeds the preset threshold, the terminal issues a voice suggestion: "Adjust the rhythm." When the short-term injury risk index exceeds the high-risk threshold, the terminal issues a strong warning to the soldier, and at the same time, the warning information is pushed to the commander's command terminal with high priority.

9. The non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, In step S5, the anonymized aggregated evaluation results pushed to the commander's command terminal specifically include the real-time average load intensity distribution of different areas in the training ground displayed in the form of a heat map, the anonymized identifiers of high-risk soldiers and their main risk types updated in real time in the form of a list, and the change curve of the group average fatigue coefficient in the entire training unit displayed in the form of a trend graph.

10. A non-intrusive real-time military physical fitness assessment method based on edge computing according to claim 1, characterized in that, Before step S1, there is also a model initialization and personalized adaptation stage. The specific steps are as follows: when a new soldier uses the system for the first time, he is guided to complete a series of standard military tactical actions under supervision, and corresponding sensor data is collected to fine-tune the parameters of the last fully connected layer related to action classification in the lightweight convolutional neural network model. Simultaneously, physiological data of the soldier under resting and extreme exercise conditions are collected to calibrate personalized parameters related to the individual's maximum heart rate, resting heart rate, basal metabolic equivalent, and range of motion in the personalized physical fitness assessment model.

Images