An intelligent safety production monitoring system

Abstract

The application relates to the technical field of industrial automation and intelligent safety monitoring, in particular to a safety production intelligent monitoring system; the safety production intelligent monitoring system comprises a multi-dimensional perception fusion module, a dynamic potential energy calculation module, a risk assessment module and an active intervention module; the multi-dimensional perception fusion module constructs a unified space-time reference and synchronously obtains a target posture and equipment working conditions; the dynamic potential energy calculation module deduces a target predicted escape trajectory and an equipment predicted inertia trajectory based on physical laws; the risk assessment module calculates space-time intersection parameters and introduces a physiological response hysteresis parameter to determine a safety buffer time window; the active intervention module calculates the difference between the intersection parameters and the buffer time window, compares the difference with a critical threshold value, generates a forced braking instruction if the difference is less than or equal to the threshold value, and generates an audible and light warning instruction if the difference is greater than the threshold value; the safety production intelligent monitoring system effectively filters invalid alarms through physical critical point judgment logic, eliminates alarm fatigue, forcibly stops before a fatal risk occurs, and blocks inertia mechanical damage.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent safety monitoring technology, specifically to an intelligent monitoring system for safe production. Background Technology

[0002] Intelligent monitoring of safe production refers to the identification and control of potential safety risks in complex industrial scenarios where heavy machinery and personnel work together. Currently, there are three main methods for safety monitoring: physical fence isolation, tag detection based on radio frequency technology, and area intrusion detection based on visual sensors. However, when monitoring safety risks based on existing technologies, on the one hand, it mainly relies on static rules to determine whether personnel have crossed the boundary, and cannot accurately predict whether a collision will occur in the future based on the laws of physical inertia. On the other hand, it lacks a graded intervention mechanism for response delays, which can easily generate invalid alarms and cause alarm fatigue among on-site workers, all of which reduce the actual protective effect of the safety monitoring system. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the present invention provides an intelligent monitoring system for safe production. Specifically, the technical solution of the present invention includes: The multi-dimensional perception fusion module is used to synchronously acquire real-time attitude data of the target object and operating status data of the mobile device through the visual acquisition unit and the device communication interface, and to perform spatial coordinate mapping on the real-time attitude data to generate the target motion vector. The dynamic potential energy calculation module is used to receive the target motion vector and the operating condition data, and to deduce the predicted escape trajectory of the target object and the predicted inertial trajectory of the mobile device based on the preset kinematic prediction model. The risk assessment module is used to construct a spatiotemporal risk model that includes response hysteresis parameters, calculate the spatiotemporal intersection parameters of the predicted escape trajectory and the predicted inertial trajectory within a future time window, and determine a safety buffer time window based on the response hysteresis parameters. The active intervention module is configured to calculate the difference between the spatiotemporal intersection parameter and the safety buffer time window, and compare the difference with a preset critical threshold: if the difference is less than or equal to the critical threshold, a forced braking command is generated as a graded control command; if the difference is greater than the critical threshold, an audible and visual warning command is generated as a graded control command; the active intervention module sends the graded control command to the control system of the mobile device to execute physical actions.

[0004] Preferably, the multi-dimensional perception fusion module includes: The skeleton extraction unit is used to identify key points of the human skeleton in the real-time posture data using a deep learning network, and to calculate the current velocity and orientation of the target object based on the displacement change rate of the key points of the human skeleton. The data alignment unit is used to synchronize the current speed, the orientation and the operating condition data in time based on a unified timestamp protocol to generate a standardized input sequence. The operating condition data includes the current operating speed, acceleration, and preset braking and deceleration curve of the mobile device.

[0005] Preferably, the dynamic potential energy calculation module includes: An inertial extrapolation unit is used to calculate the sweep area of ​​the mobile device under different braking levels based on the operating condition data and the law of physical inertia, and to generate the predicted inertial trajectory. The escape prediction unit is used to calculate the minimum physical displacement path required for the target object to move from its current position to a safe area based on the target motion vector and in combination with preset ergonomic parameters, and to generate the predicted escape trajectory.

[0006] Preferably, the risk assessment module includes: The collision countdown calculation unit is used to calculate the estimated collision time based on the spatial overlap area between the predicted escape trajectory and the predicted inertial trajectory. A limit constraint unit is used to acquire the response hysteresis parameters, which include the perception reaction time and the muscle action delay time. The threshold determination unit is used to compare the remaining time window obtained by subtracting the device braking time from the expected collision time with the sum of the perception reaction time and the muscle action delay time, and generate the spatiotemporal intersection parameter based on the comparison result.

[0007] Preferably, the hierarchical control instructions executed by the active intervention module include: When the spatiotemporal convergence parameters indicate an convergence trend but the difference is greater than the preset critical threshold, an audible and visual warning command is generated. The audible and visual warning command is used to drive the on-site alarm device to provide directional prompts. When the spatiotemporal convergence parameters indicate that the difference is less than or equal to the preset critical threshold, the forced braking command is generated. The forced braking command is used to directly modify the PLC register state of the mobile device to trigger an emergency stop, bypassing the alarm level.

[0008] Preferably, the system further includes: The environment modeling module is used to store static 3D spatial data of the monitored area and define insurmountable physical boundaries; When the dynamic potential energy calculation module deduces the predicted escape trajectory, it uses the physical boundary as a constraint condition and eliminates invalid trajectories that cross the physical boundary.

[0009] Preferably, the preset kinematic prediction model employs a Kalman filter algorithm or a long short-term memory network model to handle nonlinear motion trends; The system also includes an adaptive correction module, which is used to collect the actual avoidance path of the target object after the hierarchical control command is executed, and to use the actual avoidance path to backpropagate and update the weight parameters of the kinematic prediction model.

[0010] Preferably, the multi-dimensional perception fusion module and the dynamic potential energy calculation module are connected through an edge computing gateway, which is used to compress and encrypt the target motion vector for transmission. The active intervention module is connected to the control system of the mobile device via the industrial Ethernet protocol to ensure that the transmission delay of the forced braking command is lower than the preset safe communication delay standard.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This system establishes a unified spatiotemporal benchmark in a human-machine coexistence environment by synchronously acquiring and mapping spatial coordinates of data from the visual acquisition unit and the device communication interface through a multi-dimensional perception fusion module. This enables precise capture of the real-time posture of the target object and the operating conditions of the mobile device, and can transform heterogeneous data sources into standardized motion vectors. By introducing a dynamic potential energy calculation module and a kinematic prediction model, a future trajectory extrapolation mechanism based on physical laws is established, providing dynamic input including inertial trends for subsequent risk assessment. Through the spatiotemporal intersection calculation of predicted escape trajectory and predicted inertial trajectory, the system can effectively solve the problem that traditional technologies cannot identify inertial collision risks by relying solely on static area rules, significantly improving the prediction accuracy of dynamic physical contact and avoiding safety accidents caused by neglecting the device braking distance. 2. This system establishes a risk quantification model that conforms to human physiological characteristics by introducing response hysteresis parameters and constructing a safety buffer time window in the risk assessment module. This achieves a scientific coverage of perception reaction time and muscle action delay, and can accurately calculate the minimum time required for personnel to complete evacuation. Through the application of differential comparison mechanism and hierarchical control strategy in the active intervention module, a dynamic defense line from audible and visual warnings to forced braking is established, providing differentiated handling methods for different hazard levels. By filtering invalid alarms that have a tendency to exceed the limit but still have room for evacuation, the system can effectively eliminate alarm fatigue of on-site workers, while ensuring that control of equipment is directly taken over before the physical critical point of irreversible damage occurs, thus preventing mechanical injury. 3. This system establishes a robust human motion intention recognition mechanism through the application of a skeleton extraction unit and a smoothing filter, achieving suppression of high-frequency jitter noise and smooth calculation of velocity vectors, enabling accurate capture of personnel's orientation and movement trends. Through data alignment and interpolation algorithms, a time synchronization mechanism between visual data and industrial control data is established, resolving time reference deviations caused by inconsistent sampling frequencies of heterogeneous sensors. The environmental modeling module's definition of physical boundaries and invalid trajectory removal logic incorporates static obstacle constraints into the dynamic deduction process, preventing braking delays due to the assumption of false escape spaces, thus improving the robustness of safety assessments in complex industrial scenarios. 4. This system establishes an online optimization mechanism for nonlinear motion trends by combining a long short-term memory network model with an adaptive correction module. It enables reverse updates of the prediction model weights using actual avoidance paths, continuously improving the model's generalization ability in specific operational scenarios. Through data compression and encrypted transmission via an edge computing gateway, it establishes efficient front-end data preprocessing capabilities, reducing network transmission load. Direct connection with the control system via industrial Ethernet protocol ensures that the transmission latency of forced braking commands is below stringent safety standards, achieving millisecond-level system response speed. This ensures effective physical blocking in critical moments, improving the system's real-time performance and reliability. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. Example 1:

[0014] Please see Figure 1 A smart monitoring system for safe production, comprising: The multi-dimensional perception fusion module is used to synchronously acquire the real-time attitude data of the target object and the operating status data of the mobile device through the vision acquisition unit and the device communication interface, and to perform spatial coordinate mapping on the real-time attitude data to generate the target motion vector. The dynamic potential energy calculation module is used to receive the target motion vector and operating condition data, and deduce the predicted escape trajectory of the target object and the predicted inertial trajectory of the mobile device based on the preset kinematic prediction model. The risk assessment module is used to construct a spatiotemporal risk model that includes response hysteresis parameters, calculate the spatiotemporal intersection parameters of the predicted escape trajectory and the predicted inertial trajectory within a future time window, and determine a safe buffer time window based on the response hysteresis parameters. The active intervention module is configured to calculate the difference between the spatiotemporal intersection parameter and the safety buffer time window, and compare the difference with a preset critical threshold: if the difference is less than or equal to the critical threshold, a forced braking command is generated as a graded control command; if the difference is greater than the critical threshold, an audible and visual warning command is generated as a graded control command; the active intervention module sends the graded control command to the control system of the mobile device to execute physical actions.

[0015] This embodiment details the core architecture and execution logic of the system. The system initiates a multi-dimensional perception fusion module, which aims to construct a unified spatiotemporal reference in a human-machine coexistence environment. By connecting an industrial-grade binocular depth camera to an industrial bus interface, it captures the posture of on-site personnel and the operating status of mobile devices such as cranes and AGVs in real time. A dynamic potential energy calculation module intervenes, using physical laws to predict future spatiotemporal states, i.e., based on current momentum and motion trends, calculating possible avoidance paths for personnel and unavoidable inertial sweep paths for equipment. A risk assessment module quantifies the critical point of irreversible damage, introducing a response hysteresis parameter derived from physiological statistics, i.e., the total delay time required from the appearance of a danger signal to the completion of a physical avoidance action. Based on this, an active intervention module executes specific decision logic, calculated using the following formula:

[0016] in, Spatiotemporal convergence parameter, derived from the calculation results of the risk assessment module, physically represents the effective remaining time after deducting the equipment braking time, in seconds; Safety buffer time window, i.e., the total response hysteresis in a physical sense. The parameters are derived from preset physiological response hysteresis parameters. In this embodiment, to ensure the feasibility of the calculation, the perception reaction time is determined based on ISO 13855 safety standards and general population physiological statistics. Set as This time encompasses the time it takes for visual signals to travel to the cerebral cortex and for cognitive processing, including the time it takes for muscle movements to be delayed. Set as This time encompasses the time it takes for nerve impulses to travel to muscle contraction and overcome mechanical inertia; Therefore, a safety buffer window The value is the sum of the two, that is In specific scenarios involving elderly workers, this parameter can be adjusted to... To increase safety redundancy; the physical meaning is the minimum time required for personnel to complete effective evacuation, measured in seconds; Risk margin difference is derived from the difference between the two mentioned above, and its physical meaning is the safety redundancy at the current moment. The preset critical threshold is derived from the system safety redundancy design specifications. Its physical meaning is the minimum time margin required to cover sensor measurement errors, communication network jitter, and fluctuations in the braking friction coefficient. For example, the value is 0.3 seconds to 0.5 seconds, and the unit is seconds. System response With preset critical threshold The comparison results are used to implement hierarchical control: in response to This indicates that the risk margin is insufficient to support a normal response, meaning the remaining time is close to or less than the system's fault tolerance threshold, and the system directly generates a forced braking command; in response to This indicates that although there is a tendency to collide, there is still room for avoidance, and the system generates an audible and visual warning command. In complex industrial scenarios where heavy machinery and personnel work together, this embodiment introduces a dynamic potential energy calculation and difference comparison mechanism to achieve predictive and graded intervention for safety risks. Unlike traditional static rules that only determine whether a person is in the area, this scheme calculates whether irreversible physical contact will occur between a person and equipment in the future. This judgment logic based on physical critical points effectively filters out invalid alarms that cross the boundary but pose no actual collision risk, eliminates alarm fatigue for on-site workers, and ensures that equipment is forced to stop before the physical critical point where a fatal risk occurs, thus preventing mechanical injury accidents caused by inertia.

[0017] Example 2: The multi-dimensional perception fusion module includes: The skeleton extraction unit is used to identify key points of the human skeleton in real-time pose data using a deep learning network, and to calculate the current velocity and orientation of the target object based on the displacement change rate of the key points of the human skeleton. The data alignment unit is used to synchronize the current speed, orientation and operating condition data with the time sequence based on a unified timestamp protocol, and generate a standardized input sequence. The operating condition data includes the mobile device's current operating speed, acceleration, and preset braking and deceleration curves.

[0018] This embodiment further refines the multidimensional perception fusion module, focusing on accurately capturing human motion intentions and eliminating clock drift from heterogeneous data sources; the skeleton extraction unit utilizes a lightweight deep learning network to process visual data and identify the head, left shoulder, etc. right shoulder Key skeletal points such as the pelvis; To address the unavoidable high-frequency jitter noise in the raw visual data and prevent velocity calculation divergence caused by direct differentiation, the system introduces a Savitzky-Golay smoothing filter here; let... For a moment The original skeletal key point coordinates were collected. To obtain the smoothed coordinates, a 5-point cubic polynomial smoothing formula is applied to the coordinate sequence of each skeletal keypoint, as follows:

[0019] in, Specifically, this is represented as a column vector of three-dimensional coordinates of the key point in the camera coordinate system. The above smoothing formula involves scalar operations. In practice, it is applied to the vectors respectively. The three components are independently filtered using this formula to eliminate high-frequency noise along each axis; based on the smoothed key points... Calculate the current velocity vector of the target object based on its continuous displacement; the calculation formula is as follows:

[0020] To fully comply with the orientation calculation limitations in the embodiments, the system further constructs a human body plane normal vector; specifically, it extracts the coordinates of the right shoulder keypoint identified by the deep learning network. Coordinates of key points on the left shoulder Calculate the vector of the shoulder line. Assuming the person is standing upright, project this vector onto a horizontal plane and rotate it by 90 degrees to obtain the normalized human face orientation vector. :

[0021] Among them, matrix This is a two-dimensional rotation matrix, which physically means rotating a vector counterclockwise by 90 degrees in a plane; here, it's intended to rotate the vector parallel to the line connecting the shoulders. The rotation is a normal vector perpendicular to the line connecting the shoulders and pointing forward of the body, thus obtaining the facial orientation; The orientation vector This is used to subsequently determine whether a person is looking at the mobile device, i.e., the angle between their line of sight and the device's position vector, thus affecting the perception reaction time in the risk assessment module. The weight of the value; The specific impact calculation logic is as follows: The system defines an attention deviation penalty factor. For example, take Get the current centroid coordinates of the mobile device. Current centroid coordinates of the target object Calculate the relative position vector between the human and the machine And calculate the vector and the normalized orientation vector. dot product cosine value ; based on Correcting basic sensory reaction time This generates the final sensory reaction time. ,Should The setting value in Example 1 will be used in the safety buffer time window. Calculation:

[0022] To prevent the line of sight from being at a critical angle Small fluctuations nearby caused When the value changes abruptly, the system introduces angle hysteresis comparison logic at this point, that is, when the line of sight deviates from the angle... The penalty is triggered when the system is restored. The punishment will be cancelled in time. The preset anti-shake dead zone angle; in, For example, take a preset penalty factor for attention deviation. This value is based on statistical data from cognitive psychology experiments and represents the physiological characteristic that, under non-gaze conditions, the average cognitive processing time for sudden danger increases by 50%. The formula indicates the system's perceived reaction time when a person's back is to the equipment or their line of sight is off. This will increase significantly, thereby expanding the safety buffer window. ; The data alignment unit addresses the issue of inconsistent sampling frequencies between the vision sensor and the device's PLC. Based on the IEEE 1588PTP protocol, it calculates the current speed... Orientation Synchronize with the operating condition data in time; during this period, the data alignment unit specifically uses the cubic spline interpolation algorithm to process the inter-frequency data. Based on the sampling time of the visual sensor, the operating condition data of the equipment with higher frequency is resampled to ensure that the dimension-aligned personnel posture vector and equipment condition vector are included in each standard time step. Simultaneously, the system retrieves a preset braking deceleration curve, which describes the functional relationship between the speed of the equipment and time decay under emergency braking conditions. This curve is derived from the equipment's factory test data. Finally, a standardized input sequence that has undergone interpolation and alignment processing is generated to provide a time-consistent data source for subsequent prediction models.

[0023] Example 3: The dynamic potential energy calculation module includes: The inertial extrapolation unit is used to calculate the sweep area of ​​the mobile device under different braking levels based on operating condition data and the law of physical inertia, and to generate a predicted inertial trajectory. The escape prediction unit is used to calculate the minimum physical displacement path required for the target object to move from its current position to a safe area based on the target motion vector and in combination with preset ergonomic parameters, and to generate a predicted escape trajectory.

[0024] This embodiment is a refinement of the dynamic potential energy calculation module, aiming to quantify the physical hazard boundary of mechanical equipment and the maximum risk avoidance capability of humans; in order to ensure that the collision detection based on the Separating Axis Theorem (SAT) in the subsequent risk assessment module has the necessary geometric input basis, this embodiment corrects the original scalar area calculation logic and converts it into generating an accurate set of spatial polygons. The inertial extrapolation unit is based on the current velocity in the operating condition data. angular velocity With braking deceleration curve The state sequence of the device's centroid is generated by Euler integral. The position update formula is as follows:

[0025] in, The instantaneous linear velocity at the previous moment, in m / s, is derived from the integral derivation of the braking and deceleration curve. The device heading angle at the previous moment, in rad, is derived from the integral update of the angular velocity; To construct a geometric description of the predicted inertial trajectory Unit calculation for each discrete time point The device outline vertices; let the device length and width be... ,in, and These are preset fixed values, in meters, derived from the equipment manufacturer's specifications or on-site measurement data, for example, the values ​​are taken from... Define the local coordinate vectors of the four vertices relative to the centroid as follows: To clarify the geometric boundaries of collision detection, subscripts are specifically defined here. The correspondence between the equipment's location and its orientation is based on the equipment's direction of travel. The axis, on the left is Establish a local coordinate system along the axes: Corresponding to the left front vertex, Corresponding to the left rear vertex, Corresponding to the right rear vertex, Corresponding to the right front vertex; Using rotation matrix Mapping local coordinates to the world coordinate system yields the vertices. :

[0026] in, Indicates the first The first time step The world coordinate system position vector of each vertex is used to determine the actual space occupied by the device; Final predicted inertial trajectory The Boolean union of the convex hulls of the device profiles at adjacent time steps is constructed to form a continuous sweep polygon band:

[0027] Meanwhile, the escape prediction unit, based on the target motion vector and combined with preset ergonomic parameters, specifies the calculation logic for the minimum physical displacement path; assuming... The current position of the target object. For the current location of the mobile device, the system defines the escape acceleration vector. Its direction is taken as perpendicular to the current velocity vector of the device. And point to the target object The side with the shortest escape direction; to eliminate ambiguity in direction selection, the specific mathematical definition of this direction is as follows: Let the device velocity vector Calculate candidate normal vectors If dot product ,but direction and Same, otherwise with Same; its modulus is set as the human body's ultimate acceleration. This value is derived from the 50th percentile of the "GB / T10000-1988 Chinese Adult Human Dimensions" and related ergonomic and motor function statistics. It represents the instantaneous starting acceleration capability of a typical adult in an emergency and serves as a conservative estimate of the escape performance benchmark for the system. Calculate the centerline of the escape path based on the uniformly accelerated kinematics formula. :

[0028] To transform a one-dimensional path into a two-dimensional region capable of spatial intersection calculations, the system introduces a human safety radius parameter. For example, a value of 0.3 meters is derived from the 95th percentile of adult shoulder width statistics, and the escape trajectory is predicted using Minkowski data. :

[0029] in, The predicted time window length is measured in seconds and typically covers the complete braking cycle required for the device to decelerate from its current speed to zero, for example, 3-5 seconds. This calculation process clarifies the geometric generation method of the trajectory, ensuring that subsequent calculations of spatially overlapping regions are mathematically feasible.

[0030] Example 4: The risk assessment module includes: The collision countdown calculation unit is used to calculate the estimated collision time based on the spatial overlap area between the predicted escape trajectory and the predicted inertial trajectory. Limit constraint unit is used to obtain response hysteresis parameters, which include sensor reaction time and muscle action delay time. The threshold determination unit is used to compare the remaining time window obtained by subtracting the device braking time from the expected collision time with the sum of the perception reaction time and muscle action delay time, and generate spatiotemporal intersection parameters based on the comparison result.

[0031] This embodiment deepens the risk assessment module, focusing on how to accurately determine spatiotemporal collisions between irregular geometric shapes, and corrects the physical calculation model of braking time to conform to the actual deceleration process; the collision countdown calculation unit maps the predicted escape trajectory and the predicted inertial trajectory to the same spatiotemporal coordinate system, and uses the Separating Axis Theorem (SAT) for intersection detection; to clarify the specific calculation steps, this embodiment discloses the internal implementation logic of the CheckOverlap function: defining the discrete time step index. ,in Regarding the first Device trajectory polygon at each time step Safety zone for personnel Conduct testing; In this algorithm, the input variables are defined as follows: According to the formula in Example 3 The calculated first The set of device contour vertices at each time step; Defined as the target object in the first... Predicted location coordinates of time step That is, the geometric center of the safety circle; The aforementioned safe radius for the human body; Candidate Separation Axis Set The generation follows the circle-polygon SAT detection principle: obtaining Add the normal vectors of all edges to the set. ; Get from the center point to The vector of each vertex is added to the set. For sets Each axis in The projection interval of the polygon vertices on this axis is calculated by the dot product operation. Simultaneously calculate the scalar projection of the circle's center onto that axis. Thus, the projection interval of the circular region is obtained. ,Right now If there exists any axis such that or If the two figures are separated on the current axis, it means there is no collision at the current time step; a collision is determined only when the projection intervals on all candidate axes overlap; the estimated collision time is... Defined as the first time step in which overlap is detected:

[0032] in, For time step The polygonal region corresponding to the predicted inertial trajectory of the mobile device at any given time, i.e., the aforementioned The cross-section at that moment; For the same time step The safe circle region corresponding to the predicted escape trajectory of the target object at any given time, i.e., the aforementioned The cross-section at that moment; If there is no overlap, then The limit constraint unit obtains the response hysteresis parameter, i.e., the sensing reaction time. Muscle movement delay time ; Based on this, the threshold determination unit executes rigorous spatiotemporal parameter generation logic; targeting the equipment braking time. In order to correct the deviation in the physical meaning of considering only the velocity scalar integral, this embodiment strictly follows the calculation based on the device's current velocity. Decelerate to a standstill Integrate the physical process; based on the aforementioned braking deceleration curve Here, it is defined as the magnitude of the deceleration, i.e. Based on the assumption of the reversibility of the motion process, the braking process time is equivalent to the time it takes to accelerate from rest to the current speed under the same resistance. Therefore, the following definite integral formula that conforms to the physical process is adopted:

[0033] For ease of calculation in digital systems, the above formula is numerically equivalent to: The system employs discretized inverse summation:

[0034] in, For the velocity slice granularity, for example, take , This represents the total number of slices. For discrete velocity nodes, defined as , from Take to , indicating speed from Gradually accumulate to The process; This represents the instantaneous deceleration value at the corresponding velocity node. To prevent division by zero errors, a minimum deceleration threshold is set, for example, by taking... This calculation, by clearly defining the upper and lower limits of integration and the direction of deceleration, ensures that the braking time calculation conforms to the laws of physics. Unit calculation physical remaining time window To eliminate ambiguity between the internal calculated variables of the module and the system output interface definition, and to meet the requirement of clarity in the embodiment, the system performs explicit parameter mapping here: the calculated physical remaining time window... Directly defined as the spatiotemporal intersection parameter of Example 1 ,Right now ; Will With safety buffer time window That is, the aforementioned total response hysteresis Compare them; if This indicates that the system has entered the critical response region, and the unit is outputting... A high-risk status bit will be added at times; the final output Directly used as the formula in Example 1 The minuend input ensures the consistency of terminology and the integrity of the technical solution from the underlying physical calculations to the top-level risk assessment logic.

[0035] Example 5: The hierarchical control instructions executed by the active intervention module include: When the spatiotemporal convergence parameters indicate an convergence trend but the difference is greater than the preset critical threshold, an audible and visual warning command is generated. The audible and visual warning command is used to drive the on-site alarm device to provide directional prompts. When the spatiotemporal intersection parameters indicate that the difference is less than or equal to a preset critical threshold, a forced braking command is generated. The forced braking command is used to bypass the alarm level and directly modify the PLC register state of the mobile device to trigger an emergency stop.

[0036] This embodiment details the hierarchical execution strategy of the active intervention module. The system continuously monitors the risk margin difference. In response to the decision logic indicating that the difference is greater than the critical threshold, i.e., there is a convergence trend but sufficient time, the system generates an audible and visual warning command, driving the directional acoustic wave detector and strobe light on site to provide a warning, informing the operator that they are approaching a dangerous trajectory. In response to the decision logic indicating that the difference is less than or equal to the critical threshold, i.e., entering an inescapable critical point, the system generates a forced braking command. This command has the highest priority and is sent directly to the equipment's PLC via industrial Ethernet. The specific operation includes setting the emergency stop control bit E-StopBit to active or momentarily setting the speed multiplier register to zero, thereby bypassing the conventional alarm prompt level and directly taking over the control of the equipment. This embodiment achieves graded closed-loop risk management in industrial environments characterized by high noise and easily distracted attention. For low-risk situations, only early warnings are used to maintain production continuity, while for high-risk situations, the subjective reactions of personnel are no longer relied upon, and the equipment is forced to shut down directly by modifying the PLC register status. This direct intervention mechanism, which bypasses alarm levels, effectively prevents alarm failures caused by environmental interference or slow human response, thus constructing the last physical line of defense.

[0037] Example 6: The system also includes: The environment modeling module is used to store static 3D spatial data of the monitored area and define insurmountable physical boundaries; When the dynamic potential energy calculation module predicts the escape trajectory, it uses the physical boundary as a constraint and eliminates invalid trajectories that cross the physical boundary.

[0038] This embodiment introduces an environment modeling module to provide static spatial constraints and specifies the algorithm execution logic for eliminating invalid trajectories. This module stores static three-dimensional spatial data of the monitored area and discretizes insurmountable physical boundaries such as walls, pillars, and guardrails into a two-dimensional static occupancy grid map. ,in For the row and column indexes of the raster; This indicates that the grid is an obstacle. Indicates a free region; To implement the logic of eliminating invalid trajectories in the embodiment, the dynamic potential energy calculation module, when generating predicted escape trajectories, is no longer limited to a single path, but instead generates a set of candidate escape paths. These correspond to the kinematic derivation sequences for moving left, right, backward, and remaining stationary, respectively; for each trajectory in the set... subscript Representing different escape strategy directions, the system checks the path points at each time step. The validity determination function for the state after mapping to a raster map is as follows:

[0039] in, For example, the grid size of the raster map. , This indicates a floor operation, used to convert physical coordinates to raster indexes; For indicator functions; if a certain trajectory's This indicates that the escape direction is blocked by a physical boundary; for example, if a person is squeezed into a corner, the system will remove that trajectory from the candidate set. Remove from the middle; The system checks the filtered candidate set. ;like If the value is not empty, the system executes the maximum-minimum distance optimization algorithm to explicitly define the mathematical definition of maximizing the risk avoidance distance; for each trajectory in the set... Calculate its safety score :

[0040] This formula represents the area swept by people and equipment at any given moment on the trajectory. Minimum Euclidean distance between them; system selection As the final predicted escape trajectory; if If the value is empty, meaning all possible escape routes are blocked by obstacles, the system determines that there is currently no effective escape path and forcibly updates the estimated collision time in the risk assessment module. Set to the minimum value, such as This allows for the triggering of a forced braking command extremely early, before physical contact occurs, preventing braking delays caused by the algorithm's assumption of a false escape space.

[0041] Example 7: The preset kinematic prediction model uses the Kalman filter algorithm or the long short-term memory network model to handle nonlinear motion trends; The system also includes an adaptive correction module, which is used to collect the actual avoidance path of the target object after the execution of the hierarchical control command, and use the actual avoidance path to backpropagate and update the weight parameters of the kinematic prediction model.

[0042] This embodiment focuses on the optimization and self-evolution capabilities of the kinematic prediction model. For nonlinear motion trends, the system employs a Long Short-Term Memory (LSTM) network model. To ensure the reproducibility and real-time performance of this model on edge computing devices with limited industrial computing power, this embodiment discloses a specific lightweight network architecture: the model consists of an input layer, two stacked LSTM hidden layers, and an output layer. The first LSTM layer contains 128 hidden units, responsible for capturing high-frequency posture jitter features, and is configured with a Dropout ratio of 0.2 to prevent overfitting. The aforementioned hyperparameters are empirical values ​​obtained through grid search experiments on a standard human behavior dataset, based on the computing power constraints and real-time requirements of the edge computing gateway, thus balancing model fitting ability and generalization performance. The second LSTM layer contains 64 hidden units, which are responsible for aggregating long-term motion trends. Tanh is used as the state activation function and Sigmoid is used as the gate activation function between layers. The output layer is a fully connected layer without non-linear activation, which directly maps the 64-dimensional hidden state to the predicted coordinate increments of the next K time steps. To process the discrete point sequence output by LSTM To adapt to the collision detection logic based on continuous functions in Example 3, the system introduces a trajectory fitting layer; This layer employs the least squares method, constructing regression equations for the horizontal and vertical coordinate components of the coordinate points in the sequence Y, and utilizing quadratic polynomials.

[0043] Regression analysis was performed on sequence Y to extract the equivalent prediction velocity. With equivalent escape acceleration The system simultaneously calculates the fitted trajectory and the original LSTM prediction sequence. The root mean square error (RMSE); if the RMSE is greater than the preset fidelity threshold, for example... This indicates that the target's trajectory is extremely complex. In this case, we should abandon polynomial fitting and directly use the original discrete sequence. Construct a dynamic bounding box for collision detection; During the risk assessment phase, the system uses the data fitted here. In the formula of alternative embodiment 3 ,use In the substitution formula This achieves a closed-loop fusion of data-driven prediction and physical rule detection; To address the feature tracing problem of input data and ensure the model's spatial generalization ability across different operating regions, this embodiment focuses on the input tensor. A relative coordinate transformation centered on the device was performed; Specifically, the system obtains the centroid coordinates of the mobile device at the current moment. With heading angle For the absolute coordinates of the perceived target Construct normalized features :

[0044] This transformation eliminates the influence of absolute position, allowing the model to focus on learning the human-machine relative motion pattern; the normalized input tensor ,in, The sliding time window length, for example, 50 frames, where the relative velocity vector sequence The calculation also applies a device-centric rotational transformation to ensure physical consistency of the feature space. The calculation formula is as follows:

[0045] in, The velocity vector of the personnel in the world coordinate system. The device velocity vector in the world coordinate system; The feature dimension includes the relative coordinate sequence, the relative velocity vector sequence, and the acceleration sequence of the mobile device; The system introduces an adaptive correction module to construct a time-series replay buffer to address the training latency issue of real-time streaming data; the specific execution logic is as follows: Snapshot storage: at the moment the hierarchical control command is triggered. The system will use the current input feature tensor Freeze and tag the data for storage; result pairing: after the target object completes its avoidance action or the event ends, at any given time... ,in, For a preset prediction step size, such as 30 frames, the system extracts the actual trajectory coordinate sequence within that time period. And invalid frames with a confidence level below 0.5 due to occlusion are removed; Gradient calculation: Freeze the historical input Inputting the LSTM model yields the predicted sequence. ,Will Compared with the actual output observed later A set of training sample pairs is formed, and the loss function is calculated. To match the network architecture described above, the specific loss function used is the weighted mean squared error, as shown in the following formula:

[0046] in, For time decay weights, such as This is used to force the model to pay more attention to the prediction accuracy of recent trajectories; Parameter update: To ensure the real-time stability of the control system, an asynchronous update strategy is adopted. In the background low-priority thread, the error is used to calculate the gradient through the backpropagation time-backtime algorithm (BPTT) to update the weight parameters of the background mirror model. After the error of the shadow model on the validation set converges, the parameters of the main inference model are replaced during the system's idle period. This backtracking matching and asynchronous shadow update mechanism ensures the logical closed loop and code-level reproducibility of the model training process, while avoiding interference from online updates to real-time control tasks.

[0047] Example 8: The multi-dimensional perception fusion module and the dynamic potential energy calculation module are connected through an edge computing gateway, which is used to compress and encrypt the target motion vector data for transmission. The active intervention module connects to the control system of the mobile device via the industrial Ethernet protocol to ensure that the transmission delay of the forced braking command is lower than the preset safe communication delay standard.

[0048] This embodiment details the system's communication architecture and data transmission strategy. An edge computing gateway is deployed between the multi-dimensional perception fusion module and the dynamic potential energy calculation module. This gateway is responsible for preprocessing massive amounts of target motion vector data, including extracting differential data of changes for compression and encapsulating it using encryption algorithms to prevent data tampering or network bandwidth congestion during transmission. The active intervention module establishes a connection with the mobile device's control system using industrial Ethernet protocols such as Time-Sensitive Networking (TSN). This connection is configured as a high-priority channel to ensure that the transmission delay of forced braking commands is strictly below a preset safe communication latency standard. For example, less than 10ms, to meet the requirements of hard real-time in industrial settings; In this embodiment, for security monitoring scenarios that are extremely sensitive to latency, the hardware architecture ensures the system's response speed. The application of edge computing significantly reduces the amount of data transmission, thereby reducing the load and latency of cloud processing. The use of deterministic low-latency industrial Ethernet protocol to connect the control system ensures that the critical command of forced braking can be delivered to the execution end in real time. This millisecond-level response capability is the key to distinguishing ordinary monitoring from security-level monitoring, and directly determines whether the system can effectively save lives in life-or-death moments.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart monitoring system for safe production, characterized in that, The system includes: The multi-dimensional perception fusion module is used to synchronously acquire real-time attitude data of the target object and operating status data of the mobile device through the visual acquisition unit and the device communication interface, and to perform spatial coordinate mapping on the real-time attitude data to generate the target motion vector. The dynamic potential energy calculation module is used to receive the target motion vector and the operating condition data, and to deduce the predicted escape trajectory of the target object and the predicted inertial trajectory of the mobile device based on the preset kinematic prediction model. The risk assessment module is used to construct a spatiotemporal risk model that includes response hysteresis parameters, calculate the spatiotemporal intersection parameters of the predicted escape trajectory and the predicted inertial trajectory within a future time window, and determine a safety buffer time window based on the response hysteresis parameters. The active intervention module is configured to calculate the difference between the spatiotemporal intersection parameter and the safety buffer time window, and compare the difference with a preset critical threshold: if the difference is less than or equal to the critical threshold, a forced braking command is generated as a graded control command; if the difference is greater than the critical threshold, an audible and visual warning command is generated as a graded control command; the active intervention module sends the graded control command to the control system of the mobile device to execute physical actions.

2. The intelligent monitoring system for safe production according to claim 1, characterized in that, The multi-dimensional perception fusion module includes: The skeleton extraction unit is used to identify key points of the human skeleton in the real-time posture data using a deep learning network, and to calculate the current velocity and orientation of the target object based on the displacement change rate of the key points of the human skeleton. The data alignment unit is used to synchronize the current speed, the orientation and the operating condition data in time based on a unified timestamp protocol to generate a standardized input sequence. The operating condition data includes the current operating speed, acceleration, and preset braking and deceleration curve of the mobile device.

3. The intelligent monitoring system for safe production according to claim 1, characterized in that, The dynamic potential energy calculation module includes: An inertial extrapolation unit is used to calculate the sweep area of ​​the mobile device under different braking levels based on the operating condition data and the law of physical inertia, and to generate the predicted inertial trajectory. The escape prediction unit is used to calculate the minimum physical displacement path required for the target object to move from its current position to a safe area based on the target motion vector and in combination with preset ergonomic parameters, and to generate the predicted escape trajectory.

4. The intelligent monitoring system for safe production according to claim 3, characterized in that, The risk assessment module includes: The collision countdown calculation unit is used to calculate the estimated collision time based on the spatial overlap area between the predicted escape trajectory and the predicted inertial trajectory. A limit constraint unit is used to acquire the response hysteresis parameters, which include the perception reaction time and the muscle action delay time. The threshold determination unit is used to compare the remaining time window obtained by subtracting the device braking time from the expected collision time with the sum of the perception reaction time and the muscle action delay time, and generate the spatiotemporal intersection parameter based on the comparison result.

5. A safety production intelligent monitoring system according to claim 1, characterized in that, The hierarchical control instructions executed by the active intervention module include: When the spatiotemporal convergence parameters indicate an convergence trend but the difference is greater than the preset critical threshold, an audible and visual warning command is generated. The audible and visual warning command is used to drive the on-site alarm device to provide directional prompts. When the spatiotemporal convergence parameters indicate that the difference is less than or equal to the preset critical threshold, the forced braking command is generated. The forced braking command is used to directly modify the PLC register state of the mobile device to trigger an emergency stop, bypassing the alarm level.

6. The intelligent monitoring system for safe production according to claim 1, characterized in that, The system also includes: The environment modeling module is used to store static 3D spatial data of the monitored area and define insurmountable physical boundaries; When the dynamic potential energy calculation module deduces the predicted escape trajectory, it uses the physical boundary as a constraint condition and eliminates invalid trajectories that cross the physical boundary.

7. The intelligent monitoring system for safe production according to claim 1, characterized in that, The preset kinematic prediction model uses a Kalman filter algorithm or a long short-term memory network model to handle nonlinear motion trends; The system also includes an adaptive correction module, which is used to collect the actual avoidance path of the target object after the hierarchical control command is executed, and to use the actual avoidance path to backpropagate and update the weight parameters of the kinematic prediction model.

8. A safety production intelligent monitoring system according to claim 1, characterized in that, The multi-dimensional perception fusion module and the dynamic potential energy calculation module are connected through an edge computing gateway, which is used to compress and encrypt the target motion vector for transmission. The active intervention module is connected to the control system of the mobile device via the industrial Ethernet protocol to ensure that the transmission delay of the forced braking command is lower than the preset safe communication delay standard.

Images