Strip mine personnel falling detection method based on Markov random field
By using a three-axis gyroscope and a Markov random field model in the open-pit mining environment, the impact intensity and rollover degree features were extracted, which solved the false alarm problem of drop detection in open-pit mining operations, improved detection accuracy and reduced the false alarm rate, and achieved higher safety and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CCTEG SHENYANG ENG CO
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-19
AI Technical Summary
In existing open-pit mine operations, drop detection technology based on triaxial accelerometers has a high false alarm rate and is difficult to effectively distinguish between real falls and normal operations, resulting in insufficient system reliability and trustworthiness.
Using a three-axis gyroscope as the main sensor, the system extracts impact intensity and rollover characteristics, combines them with a Markov random field model, and identifies the state of a person's fall. The impact intensity feature sequence and rollover feature sequence are input into the Markov random field motion recognition model to calculate the fall probability and determine the fall event.
It improved the accuracy of drop detection by about 15% and reduced the false alarm rate by more than 60%, achieving higher safety and reliability in open-pit mining environments.
Smart Images
Figure CN122064940A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of human behavior recognition and safety monitoring technology, and specifically relates to a method for detecting personnel falls in open-pit mines based on Markov random fields. Background Technology
[0002] In the field of open-pit mine safety production, falls are one of the main risks leading to serious personal injury. The open-pit mine working environment has its unique complexity: the working area is vast, the terrain is undulating, and there are deep pits and steep slopes ranging from tens to hundreds of meters deep; large mining equipment (such as electric shovels and mining trucks) operate intensively, and the vehicles generate continuous and severe vibrations on the rugged road surface; at the same time, blasting, rock crushing, equipment operation and other production activities can generate strong linear impacts.
[0003] Currently, mainstream wearable fall detection technologies generally rely on triaxial accelerometers as the core sensor, judging the impact by detecting threshold abrupt changes in resultant acceleration. However, in the specific scenario of open-pit mines, this approach has a fundamental flaw. Accelerometers measure linear acceleration, making them highly sensitive to vibrations and impacts in any direction. This means that the acceleration signals generated by normal work activities such as the bumps while riding in a mining truck, the physical impact while operating tools, and walking on sloping gravel roads are highly similar to the initial characteristics of a real fall, leading to a persistently high false alarm rate. Frequent false alarms not only cause "alarm fatigue," but more seriously, they may cause workers to lose trust in the safety system, thus ignoring warnings when real danger occurs.
[0004] It is worth noting that the essential characteristic of real falls in open-pit mining environments (such as slips, trips, and falls from equipment) is not simply weightlessness or impact, but rather an uncontrolled roll and rotation of the body. This crucial state cannot be effectively captured by accelerometers, which primarily sense linear motion, but is the core measurement target of gyroscopes (angular velocity sensors). By measuring the angular velocity of rotation around an axis, gyroscopes can directly and accurately characterize the body's rolling motion. More importantly, vehicle vibrations and most operational impacts primarily manifest as linear motion, almost entirely lacking sustained, large-amplitude spatial rolls of the body. Therefore, gyroscope-based detection strategies can fundamentally and effectively distinguish between "vigorous but normal" work activities and "uncontrolled" falls.
[0005] While some existing solutions attempt to incorporate gyroscopes, most still relegate them to a secondary role of auxiliary and calibration functions, failing to fully leverage their unique advantages as the primary sensor in resisting vibration interference and identifying attitude loss of control. Therefore, given the unique operating environment of open-pit mines, there is an urgent need for a new drop detection method centered on gyroscopes to address the false alarm problem at the sensing principle level and effectively improve the reliability and trustworthiness of the safety assurance system. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method for detecting falls in open-pit mines based on Markov random fields.
[0007] The technical solution adopted in this invention is: a method for detecting falls in open-pit mines based on Markov random fields, the key technical points of which include the following steps:
[0008] The process involves obtaining raw time-series data from a three-axis gyroscope and a three-axis accelerometer; extracting impact intensity and rollover degree feature sequences from the gyroscope time-series data; inputting the impact intensity and rollover degree feature sequences into a motion recognition model based on a Markov random field; wherein the motion model is trained by learning normal, unbalanced, falling, and tripping motion sequences, and is used to calculate the probability that the current motion sequence belongs to a fall state; by inputting data into the trained Markov random field model, the probability that the current motion sequence belongs to a fall state is calculated, and when the probability exceeds a preset threshold, it is determined that a fall event has occurred.
[0009] In the above scheme, the impact strength characteristic is obtained by calculating the rate of change of the sum of the angular velocity moduli of the three-axis gyroscopes within a unit time window.
[0010] In the above scheme, the rate of change of the impact strength characteristic is obtained in the following ways, specifically including:
[0011] Based on the angular velocity readings of the three-axis gyroscope at each sampling moment, the angular velocity modulus at that moment is calculated; within a preset sliding time window, the angular velocity modulus is summed; the modulus and the rate of change between adjacent time windows are calculated, and this rate of change is used as the impact strength characteristic.
[0012] In the above scheme, the degree of flipping is obtained by calculating the integral of the angular velocity modulus within a time window.
[0013] The process of obtaining the integral in the above scheme includes:
[0014] Based on the angular velocity reading of the three-axis gyroscope at each sampling moment, the angular velocity modulus at that moment is calculated; within a preset time window used to characterize the body rollover process, the angular velocity modulus of multiple consecutive sampling moments is accumulated; the accumulated result is used as the rollover degree feature value at the current moment, and the rollover degree feature sequence is formed in chronological order.
[0015] In the above scheme, the establishment of the action transfer model includes the following steps:
[0016] Training sequences containing four action states—normal, unbalanced, falling, and knocked down—are collected. For the observed feature vector at each time step, the degree of matching between it and each action state is calculated using the node potential function. The transition probability between action states at adjacent time steps is characterized using the edge potential function. Based on the node potential function and the edge potential function, a joint probability distribution model of the state sequence is established, and the model parameters are learned through training data, so that the model can output the confidence level of each action state based on the input feature sequence.
[0017] In the above scheme, the probability of the fall state is a comprehensive probability calculated based on the probability of the current moment belonging to the fall-related state output by the action recognition model, combined with the state transition probability from the previous state to the current state.
[0018] In the above scheme, the method for determining the previous state is as follows: based on the impact intensity feature and the flipping degree feature sequence, calculate the probability of the corresponding feature, and take the state corresponding to the maximum value as the previous state.
[0019] The beneficial effects of this invention are as follows: Compared with existing technologies, the open-pit mine personnel fall detection method based on Markov random fields proposed in this invention uses a gyroscope as the main information source for fall judgment and innovatively proposes two feature quantities, "impact intensity" and "degree of rollover," which can more accurately capture the essential characteristics of body rollover and weightlessness during a fall. Experimental verification shows that the accuracy rate is improved by approximately 15%. By using Markov random field theory to fit the action state transition process, the action pattern can be understood from a holistic temporal perspective, effectively distinguishing real falls from similar daily activities such as sitting down or jumping, thus reducing the false positive rate. The algorithm has a moderate computational load, can run in real time on embedded devices, and is suitable for various application scenarios such as mine safety individual equipment, exhibiting good real-time performance and versatility. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a flowchart illustrating the overall process of the open-pit mine personnel fall detection method based on Markov random fields in this embodiment of the invention.
[0022] Figure 2 This is a schematic diagram illustrating the drop detection performance analysis of an embodiment of the present invention;
[0023] Figure 3 This is a schematic diagram illustrating the performance improvement analysis of an embodiment of the present invention. Detailed Implementation
[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figures 1-3 The present invention will be further described in detail below with reference to specific embodiments.
[0025] This embodiment provides a method for detecting personnel falls based on time-series data, implemented using the following units, specifically including:
[0026] An inertial measurement unit (IMU), integrated into the wearable device, is used to acquire raw timing data from a three-axis gyroscope and a three-axis accelerometer. The three-axis accelerometer is model MPU6050, and the three-axis gyroscope is model PDVG6100.
[0027] The processing unit is used to extract impact intensity features and rollover degree features from the time-series data and input them into a Markov random field model for drop detection. Its model is a Raspberry Pi 4B.
[0028] An alarm unit, used to activate an alarm and send alarm information upon detection of a fall, includes:
[0029] Local sound and light alarm: Integrated into the shoulder of the vest, it emits a high-frequency buzzer and flashes a bright light when a fall is detected.
[0030] Wireless communication module: It has LoRa and 4G dual-mode communication capabilities, which are used to send alarm information including personnel ID, geographical location and timestamp to the mine safety production command center in real time when a fall occurs.
[0031] The vest also includes a power supply unit, with a built-in high-capacity rechargeable lithium battery pack that can provide stable power for the entire system for more than 8 hours.
[0032] The processing unit in this embodiment is equipped with a fall detection method for open-pit miners based on Markov random fields, which specifically includes the following process:
[0033] S1: Data Acquisition Steps: Raw timing data from a three-axis gyroscope and a three-axis accelerometer are continuously acquired at 50Hz using an inertial measurement unit (IMU) worn on a smart vest. To eliminate high-frequency noise interference, the raw angular velocity and acceleration data are first subjected to a low-pass digital filter with a cutoff frequency of 20Hz.
[0034] S2: Feature extraction step: Process the gyroscope time-series data and calculate two core feature quantities:
[0035] S2.1: Calculate the angular velocity modulus.
[0036] For each sampling moment, the angular velocity modulus at that moment can be calculated based on the angular velocity reading of the three-axis gyroscope. In this embodiment, the instantaneous rotation intensity is characterized by the angular velocity modulus.
[0037] This embodiment calculates the angular velocity modulus G of a three-axis gyroscope. mag (t), the formula is:
[0038]
[0039] In the formula, G x (t),G y (t),G z (t) represent the angular velocity readings of the three-axis gyroscope at time t.
[0040] S2.2: Calculation of Impact Intensity Sequence: This characteristic quantity characterizes the severity of body posture changes by calculating the rate of change of the sum of the angular velocity moduli of the three-axis gyroscopes within a unit time window. The formula for calculating the impact intensity I(t) is:
[0041]
[0042] In the formula, Δt is the sampling time interval, and G mag (t) Angular velocity modulus of a three-axis gyroscope.
[0043] Set a sliding time window with a width of 100ms, such as corresponding to 5 sampling points. Within the window, sum the angular velocity modulus of consecutive sampling points, calculate the rate of change of the modulus between adjacent windows, i.e., the first-order difference, as the impact intensity value at the current moment; this process is carried out continuously to form an impact intensity feature sequence.
[0044] S2.3: Calculate the flipping degree sequence.
[0045] This characteristic quantity characterizes the cumulative roll angle of the body in space by calculating the integral of the angular velocity modulus within a time window. The calculation formula is as follows:
[0046]
[0047] Where N is the number of sampling points within the time window, G mag (i) Angular velocity modulus of a three-axis gyroscope.
[0048] This embodiment uses a relatively long sliding time window with a width of 500ms, such as corresponding to 25 sampling points, to cover a complete flipping process. Within the window, the angular velocity modulus of consecutive sampling points is accumulated, i.e., numerically integrated. The accumulated result is used as the flipping degree value at the current moment. This process is repeated continuously to form a flipping degree feature sequence.
[0049] S3 Action State Transition Process Fitting and Judgment Steps.
[0050] The extracted impact intensity and rollover degree feature sequences are input into an action state transition process model based on a Markov random field.
[0051] The Markov random field model in this embodiment is defined as follows:
[0052] Let the observation sequence be O = {O1, O2, ..., O...} T}, where O t = (I(t), R(t)) is the feature vector at time t, where I(t) and R(t) are the impact intensity and the degree of overturning, respectively.
[0053] The hidden state sequence is S = {S1, S2, ..., S} T The states include four types: normal, unbalanced, falling, and falling to the ground.
[0054] According to Markov random field theory, the joint probability distribution of the state sequence is:
[0055]
[0056] Where: ψ t (S t O t Let S be the nodal potential function, representing the state S at time t. t With the observed value O t Compatibility; ψ t，t+1 (S t S t+1 ) is the edge potential function, representing the transition probability between adjacent states; Z is the normalization constant; E represents the set of edges at adjacent time points in the graph.
[0057] This model learns from a large number of fall and non-fall action sequences to fit the continuous motion state transition process of a person and calculate the probability that the current action sequence belongs to a fall state. When this probability exceeds a preset threshold, it is determined that a fall event has occurred. For example, labeled data covering scenarios such as falling, walking, running, jumping, sitting, getting on and off vehicles constitute the training set. In this embodiment, the dataset contains 1200 samples, of which 200 are simulated falls, including simulated slips, trips, and falls from platforms, and 1000 are daily activities, including walking, running, jumping, sitting, getting on and off mine carts, and tool operations.
[0058] Each set of data contains the feature sequence extracted in step S2 and its corresponding true state sequence label; using the training set, two key functions in the Markov random field model are learned through methods such as maximum likelihood estimation, as follows:
[0059] Node potential function: Represents the degree of matching between the observed feature vector and each hidden state (normal, unbalanced, falling, knocked down) at a certain moment.
[0060] Side potential function: Represents the probability of transitions between four action states over time (e.g., the probability of transitioning from "normal" to "unbalanced").
[0061] After training, the model is able to understand the temporal evolution of action states.
[0062] Online detection is performed using the model trained in this embodiment:
[0063] The model inputs a 2-second feature sequence acquired in real time. The model calculates the joint probability distribution and outputs the confidence probability that the current sequence belongs to the "falling" or "falling" state.
[0064] The decision threshold is set to 0.75. When the confidence probability is higher than 0.75, the processing unit determines that a drop event has occurred.
[0065] S4 alarm execution steps.
[0066] Once a fall is detected, the processing unit immediately performs the following operations: It activates the local audible and visual alarm to alert on-site personnel. It also uploads the emergency alarm information (including personnel ID, GPS location, and time) to the monitoring center via the wireless communication module. Verification of the technical effectiveness of this embodiment:
[0067] like Figure 2 As shown in the comparative experiments, the method of this embodiment significantly outperforms traditional methods in terms of detection accuracy. On a test set containing 1200 samples, the method achieves an accuracy of 98.5%, which is approximately 15 percentage points higher than the accuracy of methods based on accelerometer thresholds. Particularly in distinguishing between drop and bumpy road scenarios, the false alarm rate is reduced by more than 60%.
[0068] like Figure 3 As shown, the method used in this embodiment improves accuracy by approximately 15 percentage points. In particular, when distinguishing between real falls and interference scenarios such as mine car bumps and tool impacts, the false alarm rate is reduced by more than 60%, with an average of less than one false alarm per day, significantly solving the problem of frequent false alarms and demonstrating the superiority of the time-series model strategy.
[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for detecting falls in open-pit mines based on Markov random fields, characterized in that: Includes the following steps: Obtain raw time-series data from a three-axis gyroscope and a three-axis accelerometer; extract impact intensity and rollover degree feature sequences from the gyroscope time-series data; The impact intensity feature sequence and the rollover degree feature sequence are input into the action recognition model constructed based on Markov random fields. The action model is trained by learning normal, unbalanced, falling, and falling action sequences, and is used to calculate the probability that the current action sequence belongs to a fall state. The probability that the current action sequence belongs to a fall state is calculated by inputting data into the trained Markov random field model. When the probability exceeds a preset threshold, it is determined that a fall event has occurred.
2. The method for detecting falls in open-pit mines based on Markov random fields according to claim 1, characterized in that: The impact strength characteristic is obtained by calculating the rate of change of the sum of the angular velocity moduli of the three-axis gyroscopes within a unit time window.
3. The method for detecting falls in open-pit mines based on Markov random fields according to claim 2, characterized in that: The rate of change of the impact strength characteristic is obtained through the following methods, specifically including: Based on the angular velocity readings of the three-axis gyroscope at each sampling moment, the angular velocity modulus at that moment is calculated; within a preset sliding time window, the angular velocity modulus is summed; the modulus and the rate of change between adjacent time windows are calculated, and this rate of change is used as the impact strength characteristic.
4. The method for detecting falls in open-pit mines based on Markov random fields according to claim 1, characterized in that, The degree of flipping is obtained by calculating the integral of the angular velocity modulus within a time window.
5. The method for detecting falls in open-pit mines based on Markov random fields according to claim 4, characterized in that, The process of obtaining the points described includes: Based on the angular velocity reading of the three-axis gyroscope at each sampling moment, the angular velocity modulus at that moment is calculated; within a preset time window used to characterize the body rollover process, the angular velocity modulus of multiple consecutive sampling moments is accumulated; the accumulated result is used as the rollover degree feature value at the current moment, and the rollover degree feature sequence is formed in chronological order.
6. The method for detecting falls in open-pit mines based on Markov random fields according to claim 1, characterized in that, The establishment of the action transfer model includes the following steps: Training sequences containing four action states—normal, unbalanced, falling, and knocked down—are collected. For the observed feature vector at each time step, the degree of matching between it and each action state is calculated using the node potential function. The transition probability between action states at adjacent time steps is characterized using the edge potential function. Based on the node potential function and the edge potential function, a joint probability distribution model of the state sequence is established, and the model parameters are learned through training data, so that the model can output the confidence level of each action state based on the input feature sequence.
7. The method for detecting falls in open-pit mines based on Markov random fields according to claim 1, characterized in that: The probability of the fall state is a comprehensive probability calculated based on the probability of the current moment belonging to a fall-related state output by the action recognition model, combined with the state transition probability from the previous state to the current state.
8. The method for detecting falls in open-pit mines based on Markov random fields according to claim 7, characterized in that: The method for determining the previous state is as follows: based on the impact intensity feature and the flipping degree feature sequence, calculate the probability of the corresponding feature, and take the state corresponding to the maximum value as the previous state.