Field safety monitoring device, system and method

Abstract

The invention provides an on-site safety monitoring device, system and method, a walking mechanism is arranged at the bottom of a base, a monitoring mechanism is arranged at the top of the base, the walking mechanism comprises a wheel type moving assembly composed of a plurality of independent self-driving wheels and a lifting type crawler belt obstacle crossing assembly, the lifting type crawler belt obstacle crossing assembly is arranged in a gap between the base and the ground, and the lifting type crawler belt obstacle crossing assembly is connected with the monitoring mechanism. And the wheel type moving assembly and the base can be supported by being separated from the ground or attached to the ground, so that the obstacle crossing function is achieved. By dynamically adjusting the posture of the chassis, combining a cooperative control strategy of load balancing and lifting height and fusing gravity center modeling and a layered closed-loop algorithm, it is effectively guaranteed that the robot always maintains good balance and anti-overturning capacity in the advancing process, and the operation stability of the robot under complex terrains such as ramps is remarkably improved; therefore, the system adapts to unstructured environments of various industrial sites, on the basis, the system can realize all-directional and dead-corner-free safety monitoring, and the recognition accuracy of abnormal behaviors is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection, and in particular to a field safety monitoring device, system and method. Background Technology

[0002] With the rapid development of industrial automation and intelligence, large industrial sites (such as hydropower stations, thermal power plants, petrochemical plants, mines, and hazardous chemical storage areas) have placed higher demands on the comprehensiveness, real-time performance, and reliability of safety monitoring. Currently, the mainstream safety monitoring methods still rely primarily on fixed surveillance cameras combined with manual inspections, which have many inherent drawbacks: fixed cameras have limited field of view, making it difficult to cover structurally complex areas, resulting in numerous blind spots, and are easily affected by environmental factors such as changes in lighting, rain, fog, and dust, exhibiting poor environmental adaptability; manual inspections are not only costly, inefficient, lack standardized procedures, and are highly subjective, but also pose serious personal safety risks in hazardous environments such as high temperatures, high noise levels, toxic substances, and flammable and explosive materials; while existing mobile monitoring equipment generally has limited functionality, lacking multimodal perception capabilities and on-site intervention methods, making it difficult to achieve adaptive and integrated safety control in dynamic and complex industrial scenarios. Furthermore, mobile robots often tip over when operating on complex terrains such as slopes and steps due to unstable center of gravity control, limiting their practical deployment. Therefore, this paper proposes an on-site safety monitoring device, system, and method to address the above problems. Summary of the Invention

[0003] The main objective of this invention is to provide an on-site safety monitoring device, system, and method to address the problems of blind spots in fixed monitoring systems, low efficiency and high risk of manual inspections, and limited functionality and poor environmental adaptability of existing mobile monitoring equipment in industrial site safety monitoring. To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a field safety monitoring device, system and method, wherein a walking mechanism is provided at the bottom of the base and a monitoring mechanism is provided at the top. The walking mechanism includes a wheeled moving component composed of multiple independent self-driving wheels and a lifting track obstacle-crossing component. The lifting track obstacle-crossing component is placed in the gap between the base and the ground, and can lift the wheeled moving component and the base off or close to the ground to achieve the obstacle-crossing function. The monitoring mechanism includes a rotating base, a human-machine interaction device disposed on the rotating base and which can swing back and forth in the direction of travel, a lifting mechanism disposed on the top of the human-machine interaction device, and a monitoring component disposed on the top of the lifting mechanism and which is lifted and lowered by the lifting mechanism.

[0004] The system includes: a comprehensive management layer: a software platform deployed in the remote monitoring center, which is the command and dispatch center of the system; responsible for macro-level task planning, multi-machine collaborative scheduling, massive data aggregation, storage, analysis and visualization, and providing users with a unified monitoring, alarm and remote control interface; Intelligent robot layer: This is a monitoring device that moves autonomously on-site and serves as the system's task execution terminal. It integrates all physical and functional modules, including drive, perception, control, communication, and response, and is responsible for completing specific patrol, monitoring, early warning, and temporary response tasks in complex environments. AI Recognition Technology Layer: This is the core technology that enables the above two layers; the edge AI part of this layer is embedded in the control module of the intelligent robot layer to realize local real-time analysis of video streams; the other part of the cloud AI is deployed in the comprehensive management layer to perform deeper mining and model iteration on the aggregated multi-source data.

[0005] The method includes: S1. The comprehensive management layer issues inspection task instructions to the intelligent robot layer of the monitoring device. S2. After receiving the task, the monitoring device starts the autonomous navigation process and collects on-site video data through its environmental perception module and monitoring components. S3. The monitoring device's on-site AI image intelligent analysis host computer performs real-time analysis of on-site video data to identify safety risk events. If a safety risk event is identified, the following actions are taken: Local alerts are issued via the interaction and processing module of the monitoring device; The event information is reported to the overall management level through the communication module of the monitoring device; The management team receives and processes reported incident information.

[0006] This invention provides a field safety monitoring device, system, and method. By dynamically adjusting the chassis attitude, combining a load balancing and lifting height collaborative control strategy, and integrating center of gravity modeling and hierarchical closed-loop algorithms, it effectively ensures that the robot maintains good balance and anti-tipping ability during movement, significantly improving its operational stability in complex terrains such as slopes, thus adapting to various unstructured environments in industrial sites. Based on this, the system can achieve comprehensive, blind-spot-free safety monitoring, improve the accuracy of abnormal behavior identification, enhance environmental adaptability to complex working conditions such as lighting, weather, and dust, and support multi-robot collaborative operation, expanding coverage and task flexibility. The system adopts a separate dual-host computer architecture, ensuring high real-time motion control while also considering the efficiency of intelligent analysis. Equipped with multi-mode power supply and multi-link communication mechanisms, it effectively ensures continuous and reliable operation in harsh industrial environments. Attached Figure Description

[0007] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall structural diagram of the monitoring device of the present invention; Figure 2 This is the present invention. Figure 1 Another perspective on the structure diagram; Figure 3 This is a structural diagram of the bottom of the base of the present invention; Figure 4 This is a cross-sectional view of the connection structure between the base and the walking mechanism of the present invention; Figure 5 This is a structural diagram of the walking mechanism of the present invention; Figure 6 This is a structural diagram of the monitoring mechanism of the present invention; Figure 7 This is the overall system architecture diagram of the present invention; Figure 8 This is a diagram showing the module composition of the present invention; Figure 9 This is a system architecture diagram of the remote communication module of the present invention; Figure 10 This is a schematic diagram illustrating the operational relationship of the device of this invention; Figure 11 This is the software functional architecture diagram of the present invention; Figure 12 This is a flowchart of the remote communication process of the present invention.

[0008] In the diagram: Base 1; Extended housing 101; Storage slot 102; Perforation 103; Walking mechanism 2; Independent self-driving wheel 20; Lifting track obstacle crossing component 21; Main frame 210; Lifting telescopic cylinder 211; Side frame 212; Side track unit 213; End track unit 214; Angle adjustment telescopic cylinder 215; Connecting frame 216; Depth camera 3; USB camera 4; Ultrasonic radar 5; LiDAR 6; Rotating base 7; Support seat 701; Rotating seat 702; Rotary drive motor 703; Human-machine interface device 8; Lifting mechanism 9; Monitoring component 10. Detailed Implementation

[0009] Example 1 like Figure 1-6 As shown, a field safety monitoring device includes a base 1, a walking mechanism 2, and a monitoring mechanism.

[0010] The base 1, as the main load-bearing structure, is made of aluminum alloy profiles and stainless steel plates welded or riveted together. It has a cavity inside for installing related equipment, thus effectively controlling the weight of the platform while ensuring the overall structural strength and rigidity. In addition, the top of the base 1 is provided with a flat surface for supporting the equipment.

[0011] In addition, the base 1 has an expansion shell 101 on both the front and back sides. The expansion shell 101 is fixed by welding or riveting, and its bottom is open. The bottom of the base 1 is provided with a storage groove 102 that communicates with the expansion shell 101. The bottom of the base 1 is also provided with multiple through holes 103.

[0012] The walking mechanism 2 includes a wheeled moving assembly consisting of four independent self-driving wheels 20 and a lifting track obstacle crossing assembly 21 consisting of four tracks.

[0013] The four independent self-driving wheels 20 of the wheeled mobile assembly are respectively set at the four corners of the bottom of the base 1. The bottom of the base 1 is supported on the ground by the independent self-driving wheels 20, so that a predetermined gap is formed between the base and the ground. The lifting tracked obstacle crossing assembly 21 is placed in the gap. When it is fully raised, the lifting tracked obstacle crossing assembly 21 can be lifted off the ground, so that the wheeled mobile assembly can travel on flat roads. At the same time, when the lifting tracked obstacle crossing assembly 21 is lowered, it can contact the ground and lift the independent self-driving wheels 20 off the ground, thereby switching to tracked drive mode to perform obstacle crossing operations with high passability.

[0014] In the preferred embodiment, the independent self-driving wheel 20 specifically includes a high-power brushless hub motor, which directly drives a rubber pneumatic tire. The steering of each wheel is controlled by an independent precision servo motor, with a steering angle range of -90° to +90°. By controlling different combinations of rotation speed and steering angle of the four wheels, the platform can achieve various motion modes such as zero-radius stationary rotation, lateral translation, and diagonal movement, and has extremely high flexibility on flat roads.

[0015] The lifting track obstacle crossing assembly 21 includes four lifting telescopic cylinders 211. The lifting telescopic cylinders 211 are fixedly installed on the inner bottom wall of the base 1, and their telescopic ends pass through the corresponding through holes 103 and are hinged to the main frame 210.

[0016] The main frame 210 has symmetrically arranged downward-extending side frames 212 on its left and right sides, and side track units 213 are installed at the bottom of the side frames 212.

[0017] The front and rear sides of the main frame 210 are symmetrically hinged with end track units 214. One end of the end track unit 214 is hinged to the main frame 210, and the other end is fixedly fitted with a connecting frame 216. The connecting frame 216 is a U-shaped frame and is fixedly fitted on the end track unit 214.

[0018] The top of the main frame 210 is symmetrically hinged with two angle-adjusting telescopic cylinders 215. The telescopic cylinders of the angle-adjusting telescopic cylinders 215 are hinged to the corresponding connecting frame 216 and are used to adjust the pitch angle of the end track unit 214 relative to the main frame 210.

[0019] like Figure 4 As shown, when the end track unit 214 is fully tilted up, the angle adjustment telescopic cylinder 215 and the end track unit 214 are placed in the storage slot 102.

[0020] It should be noted that the track unit is specifically a closed-loop walking mechanism composed of a drive wheel, a driven wheel, multiple track rollers, a tensioning mechanism, and flexible tracks arranged around each wheel; wherein, the drive wheel is connected to the drive motor to provide the driving power, and the track adopts a composite structure of engineering plastic matrix and high elastic rubber inlay, which significantly improves the adaptability to uneven ground and the shock absorption performance while ensuring wear resistance and impact resistance.

[0021] In this embodiment, both the lifting telescopic cylinder 211 and the angle adjusting telescopic cylinder 215 are electric push rods, which are coordinated and controlled by a unified control system based on terrain and attitude information.

[0022] With this design, when the device travels to complex obstacle terrain such as stairs, steep slopes, mud, and gravel, it can activate the obstacle-crossing mode. First, the four lifting telescopic cylinders 211 extend synchronously, pushing the main frame 210 and the track unit connected to it downwards. After the side track unit 213 and the end track unit 214 contact the ground and begin to bear weight, they continue to press down, supporting the base 1 together with the independent self-driving wheel 20 as a whole, so that it is off the ground. At the same time, the wheels are slightly lifted or passively followed by the suspension system to avoid interfering with the track operation.

[0023] During obstacle crossing, the end track units 214 on both the front and rear sides provide the main traction and support; the middle side track units 213 fully fit with the steps when climbing stairs to form a continuous support surface, effectively preventing the equipment from getting stuck or overturning due to the center of gravity being suspended. During this process, the angle adjustment telescopic cylinder 215 can finely adjust the pitch angle of the end track units 214 in real time according to the terrain feedback to ensure that they always maintain the best contact posture with the ground, further enhancing obstacle crossing stability and passability.

[0024] Meanwhile, when the equipment travels to uphill or downhill terrain, the lifting telescopic cylinder 211 can be adjusted asymmetrically, such as the front cylinder being slightly longer than the rear cylinder. In conjunction with the angle adjustment mechanism, the pitch posture of the base 1 can be further finely adjusted, thereby dynamically optimizing the overall center of gravity projection position and ensuring that it always falls within the support polygon.

[0025] The monitoring mechanism includes a rotating base 7, a human-machine interaction device 8, a lifting mechanism 9, and a monitoring component 10. The rotating base 7 includes a support 701 that is vertically fixed on the top of the base 1. A rotary drive motor 703 is fixedly installed on one side of the support 701, and a rotating seat 702 is rotatably installed on the other side. The output shaft of the rotary drive motor 703 passes through the support 701 and is connected to the rotating seat 702 for driving the rotating seat 702 to rotate in the front-back direction of the base 1.

[0026] Among them, the rotary drive motor 703 is a motor with a reducer.

[0027] The human-computer interaction device 8 is fixedly installed on the top of the rotating base 702, the lifting mechanism 9 is set on the top of the human-computer interaction device 8, the lifting mechanism 9 is an electric push rod, and the monitoring component 10 is set on the telescopic end of the lifting mechanism 9.

[0028] This design allows for adjustment of different monitoring angles via the rotation adjustment mechanism 8 of the rotating seat 702, the lifting mechanism 9, and the monitoring component 10. Simultaneously, when going up or down slopes, the control system can adjust the pitch angle in real time according to the vehicle's posture: leaning forward to prevent tipping over when climbing stairs and leaning backward to suppress forward movement when moving downhill. This dynamic adjustment of posture, rather than rigid movement, enables stable and safe passage on slopes and standard industrial stairs.

[0029] The basic mechanical platform provided in this embodiment, through its dual-mode walking and attitude adjustment capabilities, provides an indispensable hardware foundation for deploying stable and reliable mobile monitoring devices in complex, unstructured industrial environments.

[0030] Example 2 Further explanation in conjunction with Example 1, such as Figure 7-12 The structure shown is a field safety monitoring system, which is a complete intelligent monitoring system built on the monitoring device of Example 1 based on AI image recognition. The system is divided into the following three functional levels: management and control from top to bottom, and data collection and feedback from bottom to top. The overall management layer is a software platform deployed in the remote monitoring center, serving as the system's command and dispatch hub. It is responsible for macro-level task planning, multi-machine collaborative scheduling, the aggregation, storage, analysis, and visualization of massive amounts of data, and provides users with a unified monitoring, alarm, and remote control interface.

[0031] Intelligent robot layer: This is a monitoring device that moves autonomously on-site and serves as the system's task execution terminal. It integrates all physical and functional modules, including drive, perception, control, communication, and disposal, and is responsible for completing specific patrol, monitoring, early warning, and temporary disposal tasks in complex environments.

[0032] AI Recognition Technology Layer: This is the core technology that enables the above two layers; the edge AI part of this layer is embedded in the control module of the intelligent robot layer to realize local real-time analysis of video streams; the other part of the cloud AI is deployed in the comprehensive management layer to perform deeper mining and model iteration on the aggregated multi-source data.

[0033] The physical structure of the intelligent robot layer consists of: an environmental perception module, a monitoring component 10, a control module, a drive module, a communication module, a power supply and energy management module, and an interaction and processing module.

[0034] The environmental perception module is a key component for the monitoring device's environmental perception and autonomous navigation. It includes various sensors such as LiDAR 6, positioning module, gyroscope, depth camera 3, USB camera 4, and ultrasonic radar 5.

[0035] The lidar 6 adopts a 32-line design and is arranged on the top of the extended housing 101 to ensure an unobstructed field of view in front or behind. The lidar has a ranging capability of 150 meters, a ranging accuracy of ±30mm, a data acquisition rate of up to 320,000 points / second, an angular resolution of 0.3°, and an IP67 protection rating. The lidar is mainly used for environmental modeling, obstacle detection, and autonomous navigation.

[0036] The gyroscope uses a 9-axis high-precision angle sensor and integrates an accelerometer. It has a data update frequency of 100Hz, a resolution of 0.01°, and an orientation accuracy better than 0.1°. The gyroscope can operate normally in a temperature range of -40℃ to 85℃ and has an IP68 waterproof rating, ensuring stable performance in various harsh environments. The gyroscope is mainly used for the precise measurement of the attitude and balance control of the monitoring device.

[0037] The positioning module supports the BeiDou satellite positioning system, with a positioning accuracy better than 100 meters and a data update rate of 10Hz. It is mainly used for the global positioning and navigation of the monitoring device.

[0038] The depth camera 3 is located on the top of the front side of the base 1. It adopts an infrared binocular design with an RGB resolution of 1920×1080, a frame rate of 30fps, a depth field of view of 86°×57°, a depth resolution of 1280×720, and a depth frame rate of 90fps. This camera is mainly used for three-dimensional perception and fine manipulation of close-range objects.

[0039] USB camera 4 is located on the top of the back side of base 1. It has a resolution of 1920×1080 and an RGB frame rate of 30fps, and is mainly used for rear environment monitoring and reversing assistance.

[0040] Eight ultrasonic radars 5 are arranged around the base 1 for near-range obstacle detection and avoidance, and are particularly suitable for transparent objects or environments with insufficient light.

[0041] The monitoring component 10 consists of two independent pan-tilt high-definition cameras, one on each side, with no obstructions on the sides or top. Each camera has an integrated pan-tilt unit and is equipped with two high-beam hybrid fill lights and two low-beam fill lights. Each camera has a CMOS image sensor with a target area size of no less than 1 / 1.8", a pixel count of ≥8 million, and is equipped with a GPU for image processing and AI image recognition. The bitstream resolution of both channels is 2560×1440. Each camera is equipped with a motorized zoom lens with a focal length of no less than 8-32mm, and supports one-key focus and intelligent rain-sensing wiper functions.

[0042] The PTZ camera can be controlled independently, and the position of the PTZ can be adjusted remotely through a browser or client. The dual-lens PTZ rotation angle supports 360° continuous rotation in the horizontal direction and automatic 180° flip in the vertical direction from -30° to 90° for continuous monitoring, with no blind spots.

[0043] The device's own storage capacity includes a 1TB SSD solid-state drive and a 2TB HDD mechanical hard drive, and the integrated gimbal camera has an IP67 protection rating.

[0044] The control module serves as the local decision-making core of the monitoring device, employing a hierarchical computing architecture to ensure real-time response and computational intelligence.

[0045] The upper layer includes a host computer for robot system control and a host computer for on-site AI image intelligent analysis.

[0046] The robot system's control host computer uses a high-performance processor, is equipped with a large capacity of memory and storage space, runs the Ubuntu series operating system and deploys the ROS environment. This host computer is responsible for global coordination, including autonomous navigation, motion control, task scheduling, sensor data fusion and communication with the cloud.

[0047] The on-site AI image intelligent analysis host computer is equipped with a high-performance GPU for executing complex AI algorithm calculations. The host computer performs real-time analysis and processing of the acquired images locally, identifies on-site safety hazards and abnormal situations, and is dedicated to running deep learning algorithms to perform real-time edge computing analysis on surveillance videos.

[0048] The algorithm model consists of an optimized improved YOLOv3 target detection model and a spatiotemporal graph convolutional network ST-GCN behavior recognition model. The improved YOLOv3 model has optimized the feature layer for no less than twelve types of dangerous events, such as safety helmets, seat belts, falls, smoking, and hot work. The ST-GCN model is used to analyze the temporal data of key points of the human skeleton to accurately identify abnormal behaviors such as falls and climbing.

[0049] Specifically, the standard YOLOv3 uses the Feature Pyramid Network (FPN) for three-scale prediction. This embodiment introduces a bottom-up enhancement path based on FPN. Specifically, after upsampling the deep feature map, it is not only concatenated with the shallow feature map, but also an additional 1x1 convolutional layer and skip connections are added to directly inject the fine-grained features of the lower layers into the high-level semantic features. This structure enhances the network's ability to express the features of small targets, making the model more sensitive to small targets such as helmets and tools that are far away or partially occluded.

[0050] Instead of using general datasets, such as COCO's predefined anchor box sizes, we utilized our own scene dataset, including hydropower stations, chemical plants, and mines, to regenerate nine sets of anchor box sizes more suitable for this application scenario using the K-means clustering algorithm. The aspect ratios of the new anchor boxes are more concentrated in typical shapes such as upright figures, circular safety helmets, and long strips of tools, which significantly improves the initial matching accuracy of candidate boxes and accelerates the model convergence speed.

[0051] Building upon existing bounding box regression loss, confidence loss, and classification loss, a variant of Focal Loss is introduced to address the imbalance between positive and negative samples. By reducing the loss weights on easily classified large-area background negative samples, training becomes more focused on difficult-to-classify small targets and targets in complex backgrounds, effectively reducing the false negative rate.

[0052] The spatial graph of the spatiotemporal graph convolutional network ST-GCN is constructed by using 17 skeletal key points of the human body as nodes, extracting them in real time through a lightweight pose estimation model such as OpenPose Lite, and using the physical structure of the human body as edges.

[0053] Temporal graph: Connecting the same key point in consecutive T frames (e.g., T=10) in chronological order forms a temporal edge. To reduce computational load and focus on behavioral abrupt changes, an adaptive temporal sampling strategy is introduced. The sampling frame rate is reduced during periods of calm motion and increased during periods of intense or rapid motion, thereby reducing the overall computational load without losing key motion information.

[0054] To adapt to the computing power of edge computing devices, a separable spatiotemporal graph convolution module was designed. The standard spatiotemporal convolution is decomposed into two steps: first, a spatial dimension graph convolution is performed to capture the spatial relationships between the joints of the human body; then, an independent temporal dimension one-dimensional convolution is performed to capture the temporal evolution of the action. This decomposition significantly reduces the number of parameters and computational load. Simultaneously, in key action classification layers, such as trunk and leg joint connections, an attention mechanism is applied, allowing the network to automatically learn which joint connections are more important when recognizing specific behaviors such as falls and climbing.

[0055] At the end of the network, instead of using a single global pooling for classification, a multi-task recognition head is designed; one branch is used to recognize macroscopic behaviors, such as walking, running, and falling, while the other branch focuses on recognizing subtle hand movements, such as smoking, making phone calls, and violating regulations; this design enables the model to capture both large-scale posture changes and small-scale hand gestures simultaneously, improving the accuracy of recognizing complex behaviors, such as smoking while walking.

[0056] Real-time analysis process: The monitoring video stream is sent to this layer in real time; the improved YOLOv3 model detects static risks frame by frame, such as not wearing a safety helmet or smoke; for the detected personnel, the ST-GCN model analyzes their continuous frame behavior to determine dynamic risks, such as falls or fights; the identification results include event type, location, and confidence level, and an alarm event is generated locally to trigger the interaction and handling module.

[0057] This design fully leverages the advantages of both models: YOLOv3 quickly locates targets and identifies static attributes, while ST-GCN deeply analyzes the dynamic patterns of targets over time. The combination of the two enables accurate and real-time interpretation of complex and dangerous scenarios.

[0058] The lower layer is the real-time control slave device, specifically a highly reliable embedded controller that connects directly to various sensors and actuators, responsible for executing specific control commands issued by the host computer. The slave device uses a real-time operating system and supports multiple communication interfaces such as TTL, RS-485, RS-232, and CAN bus, ensuring compatibility with various hardware devices. The slave device is primarily responsible for motor control, I / O control, and the acquisition of various sensor data and battery power data, and then reports this information to the host computer.

[0059] The host computer and the slave computer communicate via RS232 serial port. A heartbeat mechanism ensures stable and reliable communication and prevents the monitoring device from malfunctioning. The communication messages adopt a unified format, consisting of message header, message type, transmission direction, message length, and message data, to ensure the integrity and accuracy of data transmission.

[0060] The drive module is the walking mechanism 2 in Example 1, which passively receives electrical signals from the real-time control lower-level machine. For example, when the lower-level machine sends a PWM signal for "left front wheel acceleration", the corresponding hub motor driver drives the motor to rotate. The drive module itself does not have any decision-making or calculation capabilities.

[0061] The communication module integrates 5G / 4G, Wi-Fi, and LoRa multi-mode communication devices, serving as a data transmission channel between the monitoring device and the cloud-based integrated management layer. It is controlled by the robot system of the control module and managed by the host computer.

[0062] The interaction and processing module includes two high-sensitivity microphones integrated on the human-computer interaction device 8, a speaker, a remote communication module integrated with a digital signal processor, a projector, an RFID module, a display screen, a strobe light, etc.

[0063] The high-sensitivity microphone has a pickup range covering an area of ​​30 square meters, a sensitivity of -48dB, a signal-to-noise ratio of 80dB, and an output impedance of 600Ω. The microphone adopts an anti-wind noise design and is equipped with a professional noise reduction algorithm, which can effectively filter background noise and extract clear human voices in industrial environments. The microphone shell has an IP67 protection rating, making it suitable for various outdoor environments.

[0064] The speaker has an impedance of 4Ω, a power of 15W, and features a wide frequency response and clear sound performance. It employs directional sound wave technology, enabling sound to be precisely delivered to specific areas and reducing environmental interference.

[0065] The remote communication module integrates a digital signal processor to handle real-time audio signal processing, including echo cancellation, noise suppression, and automatic gain control, ensuring call quality. This module supports full-duplex communication, allowing both parties to speak simultaneously without interruption of sound.

[0066] In practical applications, control center personnel can establish a voice connection with the field monitoring devices via client software to remotely guide on-site work or handle emergencies. The system also supports broadcasting, allowing the control center to issue safety warnings or work instructions to the field, covering an area with a radius of up to 50 meters.

[0067] The projector has an optical resolution of 1920×1080dpi and a brightness of 3000 ANSI lumens, supporting a projection screen range of 30-300 inches. This module is used for on-site guidance, warning, navigation, evacuation instructions, or information display. The projected content, including images, videos, and text, is sent through remote client software, and the projection position is controlled by the host computer controlled by the robot system.

[0068] RFID modules are used for the identification and management of on-site personnel and goods, as well as access control of important areas. They support real-time monitoring and management of workers and vehicles through RFID electronic tags.

[0069] The display screen is an 8-inch high-definition LCD touch screen, which displays the status data and alarm information of the monitoring device locally. The monitoring device can be controlled by operating the page through the touch screen.

[0070] The strobe light is located on the top of base 1 (not shown in the diagram). The robot system's host computer can control the strobe light to trigger an audible and visual alarm based on alarm conditions. In addition, the base 1 is equipped with two front lights and two rear motion status indicator lights; when the monitoring device is in motion, the rear lights display different colors, with red indicating braking and yellow indicating operation.

[0071] The power and energy management module includes a battery, a charging unit, and a power management unit. The battery includes a separate 120Ah drive battery and a 60Ah control battery, both of which are explosion-proof lithium iron phosphate batteries, used to power the drive module and all other modules respectively.

[0072] The charging unit includes a photovoltaic panel, a chassis wireless charging receiver coil, and an external DC power waterproof socket. The photovoltaic panel is located on the top of the base 1.

[0073] The power management unit is responsible for the overall power distribution, charging management, status monitoring and protection. According to the instructions of the robot system control host computer, it can switch the power supply mode, enter a low-power sleep state, and trigger the autonomous charging process when the battery power is lower than the threshold.

[0074] Example 3 Further explanation in conjunction with Examples 1 and 2, such as Figure 1-12 The structure shown illustrates a method for on-site safety monitoring, the method comprising: S1, the comprehensive management layer of the remote monitoring center generates task instructions containing inspection areas, path points, and key monitoring targets based on preset inspection plans or operator instructions; these instructions are sent to the designated on-site monitoring devices, i.e., the intelligent robot layer, through the communication module.

[0075] S2. After receiving the task, the robot system in the control module of the monitoring device starts the autonomous navigation process.

[0076] Specifically: Positioning and mapping: By fusing real-time point cloud data from LiDAR 6, global coordinates from the positioning module, and attitude data from the gyroscope, synchronous positioning and map building are completed locally, achieving autonomous positioning with centimeter-level accuracy.

[0077] Path planning and movement: Global path planning is performed based on the constructed map, and real-time obstacle avoidance is achieved using the dynamic window method; the lower control unit generates specific PWM control signals according to the planned path, driving the drive module to move the monitoring device along the planned route; during the movement, accompanying dynamic attitude adjustment is performed: when the environmental perception module detects a slope or stairs, the robot system controls the upper control unit to coordinate the adjustment of the lifting telescopic cylinder 211 and the rotating base 7 of the monitoring mechanism, adjusting the pitch attitude of the whole machine in real time to maintain the stability of the center of gravity.

[0078] The specific method for maintaining a stable center of gravity is as follows: During the movement of the monitoring device, the environmental perception module continuously measures the real-time pitch angle of the chassis 1 relative to the horizontal plane; at the same time, the data from the depth camera 3 or the lidar 6 is used to perceive the slope of the terrain ahead; the robot system of the control module controls the host computer to integrate this information, accurately calculate the relationship between the current vehicle posture and the terrain, and predict the movement trend.

[0079] The robot system's host computer calculates the overall pitch adjustment and upper load compensation adjustment required to maintain center of gravity stability based on the slope and direction of movement, such as uphill or downhill, according to a pre-set safety algorithm library.

[0080] Uphill driving: The center of gravity tends to shift backward, posing a risk of rollover. The adjustment goal is to shift the vehicle's center of gravity forward.

[0081] Downhill driving conditions: The center of gravity tends to shift forward, posing a risk of tipping over. The adjustment goal is to shift the vehicle's center of gravity backward.

[0082] The pitch angle of the chassis 1 is adjusted asymmetrically by controlling the extension and retraction of the four lifting and telescopic cylinders 211.

[0083] When going uphill: the extension of the two lifting and telescopic cylinders 211 at the front is slightly greater than that at the rear, so that the chassis 1 presents an overall posture of being lower in the front and higher in the rear, which counteracts the backward tilting torque caused by the slope.

[0084] Multi-source data synchronous acquisition: During the movement, the environmental perception module works continuously, and the monitoring component 10 acquires panoramic high-definition video streams according to preset positions or automatic tracking modes; the depth camera 3 and USB camera 4 assist in close-range three-dimensional perception and rear monitoring; the ultrasonic radar 6 rings continuously scan near-field obstacles.

[0085] When going downhill: the extension of the two lifting and telescopic cylinders 211 located at the rear is slightly greater than that at the front, so that the chassis 1 presents an overall posture of being higher in the front and lower in the rear, suppressing the forward rushing tendency when going downhill.

[0086] This adjustment directly changes the relative position of the wheel or track contact point with the vehicle body mass, and is the main means of adjusting the center of gravity.

[0087] Simultaneously, the monitoring mechanism actively balances the system, controlling the rotary drive motor 703 and the lifting mechanism 9 while the walking mechanism is being adjusted.

[0088] When going uphill: drive the rotating seat 702 to rotate backward by an angle, causing the upper loads such as the human-machine interaction device 8 and monitoring components 10 to move backward, generating a compensating torque to counteract the backward tilt. The lifting mechanism 9 can appropriately reduce its height to lower the overall center of gravity.

[0089] When going downhill: drive the rotating seat 702 to rotate forward, so that the upper load moves forward and generates a stabilizing torque. The lifting mechanism 9 can also be adjusted as needed.

[0090] This adjustment, by moving a high-positioned, mass-concentrated monitoring component, efficiently fine-tunes the center of gravity position. It is an auxiliary balancing method, and the two work together to achieve a composite adjustment strategy that combines lower-body adjustment with upper-body balancing.

[0091] During the adjustment process, the actual vehicle pitch angle and angular velocity after adjustment are continuously monitored. The robot system controls the host computer to compare the actual posture with the desired posture, forming a closed-loop control. The output of the lifting telescopic cylinder 211 and the rotary drive motor 703 are fine-tuned to smoothly and accurately maintain the ideal posture. This closed-loop adjustment will continue during the continuous uphill or downhill process to achieve dynamic stability.

[0092] The specific algorithm for maintaining a stable center of gravity is as follows: 1. System state definition and multi-source sensing fusion: System complete state vector definition: ; This vector defines all the key state variables required by the control algorithm; where, : The real-time pitch angle of chassis 1 relative to the horizontal plane, with forward tilt being positive; Chassis pitch rate; The terrain slope ahead is uphill; The real-time length of the four lifting and telescopic cylinders 211; The real-time extension and retraction height of the lifting mechanism 9; The rotation angle of the rotating base 7 is defined as the rotation around the Y-axis of the chassis coordinate system, i.e., the left and right direction. When the rotating base 702 drives the upper mechanism to tilt forward... .

[0093] Multi-source data fusion based on Kalman filtering: ; Using a Kalman filter to process noisy raw sensor data Perform optimal fusion to obtain State estimate at time 1 In the formula, : The optimal estimate of the state vector at time step includes ; : The raw measurement vectors of all sensors at any given time; : The control vector applied to each actuator at all times; : The state estimation error covariance matrix at time 1 is iteratively updated to reflect the estimation confidence level.

[0094] 2. Decision-making based on desired state: Motion condition determination function: ; The speed of the integrated monitoring device Compared with the estimated slope The system intelligently determines the current operating condition; where, : The linear velocity of the monitoring device in the forward direction, with forward movement being positive; Uphill working conditions; Downhill working conditions; : Flat ground or stationary working conditions.

[0095] Desired chassis pitch angle decision: ; The decision determines the target chassis pitch angle required to actively offset the effects of slope; where, : Desired chassis pitch angle; Pitch angle decision gain coefficient, calibrated through stability experiments; : Saturation function, output ; The maximum gradient that the system can traverse; uphill ( )hour, The instruction states that the chassis should be positioned with a lower front and higher rear, shifting the overall center of gravity forward to counteract the tendency to tip over.

[0096] Decision on the desired rotation base angle: ; The target angle of the rotating seat 702 required to move the concentrated mass of the upper part to generate a compensating torque is determined. The desired swivel angle is 702 degrees, which must meet the mechanical limit. ; : Rotation angle decision gain coefficient, which typically satisfies uphill ( )hour, The command rotates the seat 702 forward, causing the center of gravity of the human-machine interface device 8, lifting mechanism 9 and monitoring component 10 to shift forward, thus assisting in preventing rollover; the opposite is true when going downhill.

[0097] Desired lifting mechanism height decision: ; The decision determines the target height of the lifting mechanism required to lower the overall center of gravity and improve stability margin; where, : Expected lifting height; Minimum height permitted for mechanical safety and function; Minimum height permitted for mechanical safety and function; : High decision coefficient.

[0098] Calculation of desired telescopic cylinder length: ; ; ; Desired pitch angle The geometric solution is to obtain the independent length settings for the four lifting and telescopic cylinders; Target length difference between the front and rear cylinder groups; Half of the equivalent lateral distance between the mounting points of the front and rear telescopic cylinders of the chassis is a fixed mechanical parameter; The nominal length of each telescopic cylinder when the chassis is level; : Initial length command for the telescopic cylinder.

[0099] 3. Hierarchical closed-loop tracking control Chassis pitch angle outer loop PID controller: ; Taking the pitch angle tracking error as input, the output is a correction amount for the telescopic cylinder length used to dynamically compensate for model errors and external disturbances; where, : Output of the pitch angle controller; The proportional, integral, and derivative coefficients are obtained through control loop tuning.

[0100] Correction to the final length instruction for the telescopic cylinder: ; ; The correction amount of the outer loop controller With feedforward setting value The data is superimposed to generate the final, precisely-length instruction sent to the underlying driver. .

[0101] The inner loop PID controller for the telescopic cylinder position, for the i-th cylinder: ; The underlying high-bandwidth position controller ensures that a single telescopic cylinder tracks its command length quickly, accurately, and without overshoot. : Drive the first Control signals for the servo motor of the telescopic cylinder; : PID coefficient of the inner ring of the telescopic cylinder.

[0102] Rotary seat angle PID controller: ; Control the rotary drive motor 703 to adjust the angle of the rotating seat. Accurately track expected value In the formula : Control signal for the rotary seat motor.

[0103] Lifting mechanism height PID controller: ; Control the lifting mechanism drive motor to achieve the desired height. Accurately track expected value In the formula, : Control signal for the lifting mechanism motor.

[0104] 4. Real-time center of gravity estimation and safety monitoring Calculation of the center of gravity position of the upper mechanism: ; ; The core formula accurately models the series motion and calculates the position of the composite center of gravity of the monitoring component 10 in the chassis coordinate system {C}, where, The rotation axis center of the rotating base 7 is fixedly installed in the {C} system, assuming... ; Rotation around the Y-axis The rotation matrix of the angle; :when and At that time, the fixed vertical distance from the center of the rotation axis to the initial center of gravity of the upper mechanism; The coordinates of the center of gravity along the new Z-axis in the rotated local coordinate system; this formula fully describes the final effect of a series of motions on the position of the center of gravity.

[0105] Overall center of gravity coordinates and stability criteria: ; ; ; The overall center of gravity of the monitoring device is determined, and its stability is monitored in real time accordingly; where, Chassis components, including the mass and fixed center of gravity coordinates of the running gear; The total mass of the upper mechanism, i.e., the rotating seat and above; The projection coordinates of the overall center of gravity in the front-back direction in the horizontal plane, i.e., the X-axis; The support zone is determined by the extreme positions of the four grounding points in the X direction. : Preset safety margin; This criterion is a core safety constraint, and once violated, the control algorithm will trigger an emergency fault tolerance strategy.

[0106] Actuator output limiting protection: ; Before the control signal is sent to each motor driver, a final amplitude limit is applied to ensure that it does not exceed the actuator's physical capabilities, such as maximum voltage, current, and speed, to prevent equipment damage; where : Limiting function, when Time return ,when Time return Otherwise return .

[0107] This algorithm achieves fast convergence, accurate tracking, and robust stability of the control target, ensuring the safe movement of the monitoring device in complex terrain.

[0108] Using the above method, the monitoring device of the present invention can safely travel on slopes. In the climbing test, the system can always keep the horizontal projection of the center of gravity on the core area of ​​the supporting polygon, which is formed by the ground contact profile of the track or wheel, thus significantly improving the anti-overturning stability.

[0109] S3. During the movement, the environmental perception module and monitoring component 10 work continuously to collect panoramic high-definition video streams; the depth camera 3 and USB camera 4 assist in close-range three-dimensional perception and rear monitoring; the ultrasonic radar 5 continuously scans near-field obstacles.

[0110] The acquired high-definition video stream is sent to the on-site AI image intelligent analysis host computer in real time. The video frames are first input into the improved YOLOv3 model. This model uses optimized feature pyramids and adaptive anchor boxes to quickly and accurately detect targets in each frame of the image, such as personnel, safety helmets, safety belts, smoke, open flames, etc., and determine their compliance status. For each detected personnel, the skeletal key point sequence of multiple consecutive frames is extracted to construct a spatiotemporal map and input into the ST-GCN model. This model accurately analyzes temporal behavior patterns through separable spatiotemporal convolution and attention mechanisms, and identifies dynamic dangerous behaviors such as falls, climbing, running, and smoking. Finally, the on-site AI image intelligent analysis host computer integrates the static state recognition results of the improved YOLOv3 with the dynamic behavior recognition results of ST-GCN, and generates structured safety events with complex semantics through a rule engine. For example, a high-risk event: personnel not wearing safety belts while working at height, with time, location, confidence level, and related evidence, such as screenshots. The robot system controls the host computer to trigger a graded response mechanism based on the risk level of the AI-identified event.

[0111] S4. Tiered Response and Collaborative Handling: The robot system controls the host computer to trigger a tiered response mechanism based on the risk level of the event identified by AI.

[0112] For general risk events, such as not wearing a safety helmet: Local voice prompts: The directional speaker of the control module's command interaction and handling module plays pre-recorded warning voice messages toward the target personnel, such as "Please wear a safety helmet."

[0113] Event logging and reporting: The control module uploads a brief record of the event to the comprehensive management layer for archiving via the communication module.

[0114] For high-risk or emergency events, such as falls, open flames, or calls for help: Local multimodal strong intervention: The control module immediately activates multiple measures: controls the strobe light to flash at high frequency; controls the speaker to play a high-decibel alarm sound; controls the projector to project a red warning circle or danger text on the ground or wall.

[0115] Real-time reporting of emergency information: The main controller pushes alarm information, including event details, precise location of monitoring devices, and related video clips, to the monitoring screen of the integrated management team and the terminals of relevant personnel in real time through the highest priority communication link.

[0116] S5, cloud-based monitoring, decision-making, and remote collaboration: After receiving an alarm message, the management layer initiates a cloud-based collaboration process: the location of the monitoring device is highlighted on the electronic map, and its real-time video stream automatically pops up; personnel in the monitoring center can remotely control the monitoring device's pan-tilt unit to rotate and focus through the platform to confirm the situation on site from multiple angles; It can also communicate with on-site personnel through a remote communication module to provide remote guidance or emergency command; at the same time, it can dispatch other nearby monitoring devices to provide support, generate standardized handling work orders, and record and manage the entire process of the incident in a closed loop.

[0117] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A field safety monitoring device comprising a base (1), characterised in that: The bottom of the base (1) is provided with a walking mechanism (2), and the top is provided with a monitoring mechanism, wherein the walking mechanism (2) comprises a wheeled moving assembly composed of a plurality of independent self-driven wheels (20) and a lifting track obstacle crossing assembly (21), the lifting track obstacle crossing assembly (21) is arranged in the gap between the base (1) and the ground, and can be separated from the ground or attached to the ground to support the wheeled moving assembly and the base (1) to realize the obstacle crossing function, the monitoring mechanism comprises a rotating base (7), a man-machine interaction device (8) arranged on the rotating base (7) and capable of swinging back and forth in the direction of travel, a lifting mechanism (9) arranged on the top of the man-machine interaction device (8), and a monitoring assembly (10) arranged on the top of the lifting mechanism (9) and lifted through the lifting mechanism (9).

2. The device of claim 1, wherein: The base (1) is provided with an expansion shell (101) on both sides of the front and rear, and the bottom is open, the bottom of the base (1) is provided with a receiving groove (102) in communication with the expansion shell (101), and the bottom of the base (1) is further provided with a plurality of perforations (103).

3. The apparatus of claim 2 wherein: The lifting track obstacle crossing assembly (21) comprises four lifting telescopic cylinders (211), the lifting telescopic cylinders (211) are fixedly arranged on the inner bottom wall of the base (1), the telescopic ends of the lifting telescopic cylinders (211) pass through the corresponding perforations (103), and the lifting telescopic cylinders (211) are hingedly connected with a main frame body (210), the left and right sides of the main frame body (210) are symmetrically provided with downwardly extending side frame bodies (212), the bottom ends of the side frame bodies (212) are provided with side track units (213), the front and rear sides of the main frame body (210) are symmetrically hingedly connected with end track units (214), one end of the end track unit (214) is hingedly connected with the main frame body (210), the other end is fixedly sleeved with a connecting frame (216), and the top of the main frame body (210) is symmetrically hingedly connected with two angle adjusting telescopic cylinders (215), the telescopic cylinders of the angle adjusting telescopic cylinders (215) are hingedly connected with the corresponding connecting frames (216), and the angle adjusting telescopic cylinders (215) are used for adjusting the pitch angle of the end track unit (214) relative to the main frame body (210); When the end track unit (214) is completely raised, the angle adjusting telescopic cylinder (215) and the end track unit (214) are arranged in the receiving groove (102).

4. The device of claim 1, wherein: The rotating base (7) comprises a support seat (701) vertically fixed on the top of the base (1), one side of the support seat (701) is fixedly provided with a rotating drive motor (703), the other side is rotatably provided with a rotating seat (702), the output shaft of the rotating drive motor (703) is in transmission connection with the rotating seat (702) through the support seat (701), and the rotating drive motor (703) is used for driving the rotating seat (702) to rotate in the front-rear direction of the base (1), and the man-machine interaction device (8) is fixedly installed on the top of the rotating seat (702).

5. A field safety monitoring system characterized by: The monitoring system comprises: The comprehensive management layer is a software platform deployed in the remote monitoring center, which is the command and dispatch center of the system; responsible for macro task planning, multi-machine cooperative scheduling, massive data aggregation, storage, analysis and visualization, and provides a unified monitoring, alarm and remote control interface to the user; The intelligent robot layer is a monitoring device autonomously moving on site and is a task execution terminal of the system; all physical and functional modules such as driving, sensing, control, communication and disposal are integrated, and the intelligent robot layer is responsible for completing specific patrol, monitoring, early warning and temporary disposal tasks in a complex environment; The AI recognition technology layer is a core technical capability for empowering the above two layers; the edge AI part of the layer is embedded in the control module of the intelligent robot layer to realize local real-time analysis of video streams; the other part of the cloud AI is deployed in the comprehensive management layer to further mine and model iterate the converged multi-source data.

6. The field safety monitoring system of claim 5, wherein: The intelligent robot layer comprises: an environment sensing module comprising a laser radar (6), a positioning module, a gyroscope, a depth camera (3), a USB camera (4), an ultrasonic radar (5) and a monitoring assembly (10) composed of two independent gimbals of high-definition cameras, which are arranged on a base (1); a control module adopting a layered computing architecture, an upper layer comprising a robot system control host computer and an on-site AI image intelligent analysis host computer, and a lower layer being a real-time control slave computer directly connected with various sensors and actuators and responsible for executing specific control instructions issued by the host computer; a driving module, specifically a walking mechanism (2), passively receiving electrical signals from the real-time control slave computer; a communication module integrating 5G / 4G, Wi-Fi and LoRa multi-mode communication devices, serving as a data transmission channel between the monitoring device and the cloud comprehensive management layer and being managed by the robot system control host computer; an interaction and disposal module comprising two high-sensitivity microphones, a loudspeaker, a remote communication module integrated digital signal processor, a projector, an RFID module, a display screen and a flashing light integrated in a man-machine interaction device (8); a power supply and energy management module comprising a battery, a charging unit and a power management unit, the battery comprising a driving battery and a control battery, the charging unit comprising a photovoltaic panel, a bottom disc wireless charging receiving coil and an external DC power supply waterproof socket, the photovoltaic panel being arranged on the top of the base (1), and the power management unit being responsible for whole machine power distribution, charging management, state monitoring and protection; it can switch power supply modes, enter a low-power sleep state and trigger an autonomous charging process when the battery power is lower than a threshold value according to the instructions of the robot system control host computer.

7. The field safety monitoring system of claim 6, wherein: The algorithm model configured in the on-site AI image intelligent analysis host computer is an improved YOLOv3 target detection model and a spatio-temporal graph convolution network ST-GCN behavior recognition model; the improved YOLOv3 model is optimized in feature layer for dangerous event; the ST-GCN model is used for analyzing human skeleton key point time series data to accurately recognize abnormal behavior.

8. The field safety monitoring system of claim 7, wherein: The improved YOLOv3 model and the ST-GCN model are as follows: the standard YOLOv3 adopts an FPN feature pyramid network for three-scale prediction, the improved YOLOv3 model introduces a bottom-up enhancement path based on the FPN; specifically, after sampling the deep feature map, not only is it concatenated with the shallow feature map, but also an additional 1x1 convolution layer and a skip connection are added to directly inject the fine-grained features of the lower layer into the high-level semantic features. Instead of using general data sets, a self-scene data set is used to regenerate 9 groups of anchor box sizes more suitable for the application scenario through the K-means clustering algorithm; On the basis of the original boundary box regression loss, confidence loss and classification loss, the variant of focal loss FocalLoss is introduced, which reduces the loss weight of large area background samples that are easy to classify, so that the training is more focused on small targets and targets in complex backgrounds that are difficult to classify; The spatial graph of the spatio-temporal graph convolutional network ST-GCN: taking the 17 skeletal key points of the human body as nodes, extracting them in real time through a lightweight pose estimation model, and taking the physical structure connection of the human body as edges to construct a spatial graph; Temporal graph: connecting the same key points in consecutive T frames in chronological order to form a time edge, and introducing an adaptive time sampling strategy to reduce the sampling frame rate during the motion flat period and increase the sampling frame rate during the motion intense or rapid change period; The standard spatio-temporal convolution is decomposed into two steps: first, the spatial dimension graph convolution is performed to capture the spatial correlation between the joints of the human body; then, an independent one-dimensional convolution in the time dimension is performed to capture the temporal evolution of the motion; at the same time, an attention mechanism is applied at the key motion classification layer; A multi-task recognition head is designed at the end of the network; one branch is used to identify macro behaviors, and the other branch focuses on identifying fine hand movements; Real-time analysis process: the monitoring video stream is sent into this layer in real time; the improved YOLOv3 model detects static risks frame by frame, and the ST-GCN model analyzes the continuous behavior of the frame to determine dynamic risks; the identification results generate alarm events locally and trigger the interaction and disposal module.

9. The field safety monitoring system of claim 6, wherein: The robot system control host computer is configured with an algorithm for keeping the center of gravity stable, which is specifically: System state definition and multi-source perception fusion: System complete state vector definition: ； This vector defines all the key state quantities required by the control algorithm; in which, : Real-time pitch angle of the chassis (1) relative to the horizontal plane, positive for pitch forward; : Real-time pitch angular velocity of the chassis; : Terrain slope in front, positive for uphill; : Real-time length of the four lifting telescopic cylinders (211); : Real-time telescopic height of the lifting mechanism (9); : Rotation angle of the rotating base (7), which defines the rotation around the Y axis of the chassis coordinate system, i.e. left-right rotation, when the rotating base (702) drives the upper mechanism to pitch forward ; Multi-source data fusion based on Kalman filtering: ； Applying a Kalman filter to raw sensor data containing noise performing optimal fusion to obtain state estimation value at time ; in which, : optimal estimation of state vector at time ; : raw measurement vector of all sensors at time : control amount vector applied to each actuator at time : state estimation error covariance matrix at time, iteratively updated to reflect estimation confidence Expected state decision: Motion working condition determination function: ； The monitoring device advances at a speed The slope is estimated The current working condition is intelligently determined; in the formula, The linear speed of the monitoring device in the advancing direction is monitored, and the advancing is positive; The uphill working condition is determined; The downhill working condition is determined; The flat or static working condition is determined; Expected chassis pitch angle decision: ； The decision determines the target chassis pitch angle required to actively offset the effects of slope; where, : Desired chassis pitch angle; Pitch angle decision gain coefficient, calibrated through stability experiments; : Saturation function, output ; The maximum gradient that the system can traverse; uphill ( )hour, The chassis is positioned with a lower front and higher rear, shifting the overall center of gravity forward to counteract the tendency to tip over. Expected rotating base angle decision: ； Determine the target angle of the rotating seat required to move the concentrated mass of the upper part to generate a compensating torque; The desired rotation angle must meet the mechanical limit. ; : Rotation angle decision gain coefficient, which typically satisfies uphill ( )hour, The command rotates the seat (7) forward, causing the center of gravity of the human-machine interaction device (8), lifting mechanism (9) and monitoring component (10) to shift forward, thus assisting in preventing rollover; the opposite is true when going downhill. Expected lifting mechanism height decision: ； Decision of the target height of the lifting mechanism required to reduce the overall height of the center of gravity to improve the stability margin; wherein, : desired lifting height; : minimum height allowed by mechanical safety and function; : minimum height allowed by mechanical safety and function; : height decision coefficient; Expected telescopic cylinder length solution: ； ； ； Desired pitch angle Geometry solution sets values for the four independent lengths of the lift cylinders; : Target length difference of the front and rear cylinder groups; : Half of the equivalent lateral distance between the mounting points of the front and rear chassis cylinders, a fixed machine parameter; : Nominal length of each cylinder when the chassis is level; : Initial length command for the cylinders; Hierarchical closed-loop tracking control: Chassis pitch angle outer loop PID controller: ； The pitch angle tracking error is inputted, and an extension cylinder length correction quantity for dynamically compensating model error and external disturbance is outputted; in the formula, : output of the pitch angle controller; : proportional, integral, and differential coefficients, obtained by control loop setting Telescopic cylinder final length command correction: ； ； The outer loop controller's correction amount with the feedforward setpoint superimposed, generating the final precise length command sent to the underlying driver ; Telescopic cylinder position inner loop PID controller, i-th cylinder: ； The bottom layer high-bandwidth position controller ensures that the single telescopic cylinder tracks the instruction length quickly, accurately and without overshoot. : control signal for driving the first telescopic cylinder servo motor; : control signal for driving the first telescopic cylinder servo motor; : inner loop PID coefficient of the telescopic cylinder Rotating seat angle PID controller: ； controlling the rotary drive motor (703) to rotate the swivel seat angle accurately tracking the desired value ; wherein : control signal of the rotary seat motor Lifting mechanism height PID controller: ； controlling the lift mechanism drive motor to raise the height accurately tracking the desired value wherein : control signal for the lift mechanism motor Real-time estimation of the center of gravity and safety monitoring: Upper mechanism center of gravity position calculation: ； ； Core formula, accurate modeling of serial kinematics, calculation of the end of the mechanical chain: monitor the position of the synthetic center of gravity of the assembly (10) in the chassis coordinate system {C}, where, : fixed installation position of the center of rotation of the rotating base (7) in the {C} system, assuming ; : rotation matrix around the Y axis by an angle of ; : fixed vertical distance from the center of rotation to the initial center of gravity of the upper mechanism when and ; : coordinate of the center of gravity in the direction of the new Z axis in the rotated local coordinate system; this formula fully describes the final effect of a serial kinematics on the position of the center of gravity; Overall center of gravity coordinates and stability criterion: ； ； ； The overall gravity center of the monitoring device is monitored in real time, and the monitoring is performed according to the overall gravity center, wherein, : chassis part, including the mass of the traveling mechanism and the fixed gravity center coordinate; : the total mass of the upper mechanism, i.e., the rotating seat and the part above the rotating seat; : the projection coordinate of the overall gravity center in the front-rear direction of the horizontal plane, i.e., the X-axis; : the support interval determined by the limit positions of the four grounding points in the X direction; : the preset safety margin; the criterion is a core safety constraint, and once the criterion is violated, the control algorithm will trigger an emergency fault-tolerant strategy. Actuator output amplitude limiting protection: ； Before the control signal is sent to each motor driver, the final amplitude limiting is performed to ensure that it does not exceed the physical capabilities of the actuator, such as maximum voltage, current, speed, etc., to prevent equipment damage; in the formula : limiting function, returning when , returning when , and otherwise returning .

10. A method of on-site safety monitoring, characterized by: The method is applied to the on-site safety monitoring system of any one of claims 5-7, and the method comprises: S1, the comprehensive management layer issues a patrol task instruction to the intelligent robot layer of the monitoring device; S2, after receiving the task, the monitoring device starts the autonomous navigation process and collects on-site video data through its environment perception module and monitoring component (10); S3, the on-site AI image intelligent analysis host computer of the monitoring device performs real-time analysis on the on-site video data, identifies safety risk events, and if a safety risk event is identified, executes: Local warning through the interaction and disposal module of the monitoring device; Reporting the event information to the comprehensive management layer through the communication module of the monitoring device; The comprehensive management layer receives and processes the reported event information.

Images