An intelligent safety early warning system for a production workshop based on an internet of things

Abstract

The application discloses a production workshop intelligent safety early warning system based on Internet of Things, and relates to the technical field of industrial safety early warning.The edge computing layer is provided with differential heterogeneous edge computing nodes, the heterogeneous edge computing nodes comprise visual edge nodes, equipment edge nodes and distributed sensing edge nodes, and the edge computing layer performs local fusion and reinforcement on the extracted local space-time features through a space-time attention mechanism.The application has the advantages that three types of multi-modal sensing array data are fused, the full-dimension risks of equipment internal microscopic hidden dangers, equipment operation macroscopic states and environments and personnel are deeply covered, the problem of split monitoring dimensions and isolated alarms in the prior art is solved through graph network fusion, the blind area of workshop monitoring is eliminated, and the risk identification coverage is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of industrial safety early warning technology, specifically to an intelligent safety early warning system for production workshops based on the Internet of Things. Background Technology

[0002] With the advancement of Industry 4.0 and intelligent manufacturing, the safety management and control of production workshops has evolved from traditional manual inspections and single threshold alarms to intelligent and comprehensive early warning systems. Currently, there are relevant technical solutions that have been optimized for workshop safety early warning. For example, equipment monitoring solutions can achieve multi-parameter monitoring of equipment vibration, temperature, and current through edge, middleware, and cloud architecture. Specific equipment safety solutions deploy distributed acoustic emission sensor networks for high-risk production equipment to predict local transient risks; Visual early warning solutions use multispectral visual sensors to monitor for violations by personnel and abnormalities in the environment.

[0003] In existing technologies, such as Chinese Patent Publication No. CN111833772A (A Workshop Safety Monitoring and Early Warning System Based on the Internet of Things), a monitoring method is disclosed that collects equipment operating parameters at the edge and works in conjunction with a cloud platform. This solution solves the problems of equipment status monitoring and data uploading to the cloud to a certain extent. However, its technical solution and other common solutions in the industry generally suffer from the following common defects: 1. Fragmented monitoring dimensions: Existing solutions either focus only on monitoring the external operating status of equipment or only on visual scene monitoring. They cannot deeply integrate multimodal information such as hidden sensors inside the equipment, external parameters of equipment operation, and visual data of the environment. This results in a large number of monitoring blind spots and cannot cover the intertwined risks of equipment, environment, and personnel in the workshop.

[0004] 2. Poor adaptability of wireless communication: There are a large number of metal equipment and strong electromagnetic interference in the production workshop. The single wireless networking scheme in the existing patent literature cannot adapt to the heterogeneous transmission requirements of different sensors: high-frequency vision and device sensing data require low latency and large bandwidth, and low-power distributed sensing requires wide coverage. The single networking and lack of anti-interference channel dynamic switching mechanism lead to frequent data packet loss and a surge in transmission delay in strong interference environment, which seriously affects the real-time performance of early warning.

[0005] 3. Lack of edge computing and causal reasoning architecture: Existing edge computing solutions only perform simple threshold alarms or local preliminary processing for single types of data, and cannot adapt to the heterogeneous processing needs of multimodal data; At the same time, the lack of a causal reasoning mechanism for cross-modal data makes it impossible to fundamentally distinguish between the direct causes and indirect effects in associated alarm signals, resulting in wasted system resources and a persistently high false alarm rate.

[0006] 4. Limitations in linkage control capabilities: Existing solutions often only target parameter adjustments for a single device, lacking a cross-system and cross-device full-domain linkage control mechanism. When complex risks occur, they cannot simultaneously link the device control base, environmental ventilation, area access control, and emergency response system to proactively eliminate the risks. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the present invention aims to provide an intelligent safety early warning system for production workshops based on the Internet of Things (IoT). This system integrates multimodal sensor data to construct a highly reliable heterogeneous wireless communication network. Relying on a heterogeneous edge computing architecture and a cross-modal causal graph network analysis model, it achieves low-latency, high-accuracy hierarchical intelligent early warning and full-domain cross-system linkage control.

[0008] To achieve the above objectives, the present invention adopts the following core technical solution: An intelligent safety early warning system for a production workshop based on the Internet of Things is characterized in that the system comprises a perception layer, a communication layer, an edge computing layer, and a middleware application layer that are connected in a sequential manner from bottom to top. The perception layer uses a multimodal intelligent sensor array to build a global perception network in the production workshop. The multimodal intelligent sensor array includes internal sensing nodes, equipment operation sensing nodes, and multispectral visual sensing nodes. The internal sensing nodes of the device are used to collect transient local latent risk data inside the device. The device operation sensor node is used to collect external operating status data of the device; The multispectral visual sensing nodes are deployed in key areas of the workshop to simultaneously collect visible light, thermal infrared, and depth image data to monitor personnel and environmental anomalies. The communication layer constructs a heterogeneous wireless communication network based on the integration of 5G, LoRa, and Wi-Fi 6 networks; The data from the multispectral visual nodes is transmitted via a Wi-Fi 6 network. Data from the device's operating sensor nodes is transmitted via a dedicated 5G network. Data from the internal sensor nodes of the device is transmitted via a LoRa network; Furthermore, the communication layer incorporates an adaptive dynamic channel allocation mechanism based on a channel quality assessment model, which calculates the wireless communication channel quality assessment index. And set the quality threshold limit for triggering dynamic channel switching. Real-time switching of transmission channels; The edge computing layer deploys differentiated heterogeneous edge computing nodes, including: visual edge nodes, used to run lightweight edge feature extraction mapping functions. Complete feature extraction from multimodal images; Equipment edge nodes are used to run anomaly detection models to extract temporal features of vibration and current data; Distributed sensing edge nodes, deployed at the workshop edge gateway, are used to extract high-frequency temporal feature vectors of acoustic emission signals. ; The edge computing layer enhances the extracted local spatiotemporal features through a spatiotemporal attention mechanism; The middleware application layer includes a multimodal spatiotemporal risk fusion assessment module and a cross-domain linkage hierarchical early warning module; The multimodal spatiotemporal risk fusion assessment module obtains the device's internal feature vector uploaded by the edge computing layer. Equipment operation feature vector and visual feature vectors And construct a multimodal causal graph model. The overall early warning risk probability value of the system is calculated and output through a graph convolutional network model. Risk level continuous evaluation score ; The cross-domain linkage hierarchical early warning module is based on The scope is mapped to a four-level risk state, and control commands are issued downwards to execute cross-system coordinated interventions, including equipment adjustment, access control locking, alarms, and emergency shutdowns.

[0009] As a further supplement to the inventive technical solution of this invention, in the communication layer, the wireless communication channel quality assessment index... The calculation formula is as follows: ; The meanings of all parameters in the formula are explained below: The signal-to-noise ratio parameter is measured at the communication receiver. The signal multipath reflection interference intensity of metal equipment in the workshop; The noise floor energy intensity under strong electromagnetic interference environment; This represents the available bandwidth transmission rate of the current wireless channel. This refers to the average end-to-end transmission delay of data packets. This is the signal-to-noise ratio gain coefficient; This is the penalty coefficient for signal reflection interference; This is the electromagnetic interference penalty coefficient; These are the standardized weighting coefficients for bandwidth rate in the channel quality assessment model; These are the standardized weighting coefficients for delay time in the channel quality assessment model.

[0010] When satisfied At that time, the system is within the given decision time window length of the early warning system. The system automatically switches the communication link to a spare frequency band.

[0011] In the edge computing layer, the spatiotemporal attention mechanism is used to enhance the spatiotemporal correlation between acoustic emission signals and vibration signals, and the spatiotemporal attention weight matrix it generates... The calculation formula is: ; The meanings of all parameters in the formula are explained below: The query vector in the spatiotemporal attention mechanism is obtained by linear mapping from the input feature sequence. This refers to the key vector in the spatiotemporal attention mechanism; This refers to the value vector in the spatiotemporal attention mechanism; Key vector Feature mapping dimension parameters; This indicates a transpose operation on the matrix formed by the key vectors; This is the softmax activation function used for normalization.

[0012] In the application layer of the middle platform, a multimodal causal relationship graph model Includes a set of node states With the set of directed edges Node state set It consists of the mapped multimodal feature matrix. The convolutional network model in the figure is in the first... +1 layer hidden node state matrix The updated formula is: ; The final risk level continuous evaluation score The calculation formula is: ; The meanings of all parameters in the formula are explained below: This is the adjacency matrix of the causal graph model, representing the strength of causal connections between multimodal nodes; This is the degree matrix corresponding to the adjacency matrix; For graph convolutional network models in the 1st... The hidden node state matrix of the layer, when =0 represents the initial features of the input multimodal nodes; For graph convolutional network models in the 1st... The learnable network weight parameter matrix of the layer; ReLU activation function is used for nonlinear activation of the model; These are the weighting coefficients of the device's internal feature vectors in the causal graph; These are the weighting coefficients of the device's operational feature vectors in the causal graph; These are the weight coefficients of the visual feature vectors in the causal graph; This represents the norm operation performed on the eigenvector; This is the output feature matrix of the last layer of the graph convolutional network model; This is the mapping function for a fully connected multilayer perceptron network.

[0013] Through the above network processing, the final output is a continuous risk level assessment score. The range is normalized to 0 to 100 when determining When the preset safety threshold is exceeded, different levels of linkage response are triggered.

[0014] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. Comprehensive monitoring without blind spots: It integrates data from three types of multimodal sensor arrays to deeply cover microscopic hidden dangers inside equipment, macroscopic operating status of equipment, and all-dimensional risks of the environment and personnel. Through graph network fusion, it solves the problems of fragmented monitoring dimensions and isolated alarms in existing technologies, eliminates blind spots in workshop monitoring, and significantly improves the risk identification coverage.

[0015] 2. Highly Reliable Heterogeneous Wireless Communication: In response to the complex interference in the workshop, the innovative heterogeneous wireless network channel quality assessment model and dynamic switching mechanism adapt to the high bandwidth and low power consumption requirements of different frequency bands. It can effectively resist multipath reflection and electromagnetic noise, and ensure extremely low data packet loss rate and significantly reduced transmission delay under strong interference conditions, thus solidifying the communication foundation for real-time early warning.

[0016] 3. Low latency and high accuracy early warning: By extracting features from heterogeneous edge computing nodes, the pressure on network bandwidth is greatly reduced. The application layer of the middle platform uses graph convolutional networks for cross-modal causal inference, which successfully removes indirect features caused by interference, shortens the early warning response time to the millisecond level, and significantly reduces the false alarm rate of traditional solutions.

[0017] 4. Proactive and coordinated control: Breaking down information silos, the cross-system coordinated control mechanism built on a four-level early warning score enables multi-device and multi-subsystem collaborative intervention after a risk is triggered, upgrading the passive alarm of risk handling to a proactive elimination closed loop, effectively preventing the occurrence of production safety accidents. Attached Figure Description

[0018] Figure 1 This is a diagram of the overall system architecture of the present invention. Figure 2 The flowchart of the adaptive dynamic channel allocation mechanism of the communication layer of the present invention is as follows. Figure 3 This is a flowchart illustrating the data flow and processing of the multimodal spatiotemporal risk fusion graph convolutional network in the application layer of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the composition, working principle, and implementation steps of the Internet of Things-based intelligent safety early warning system for production workshops will be described in great detail and in complete detail below, in conjunction with specific embodiments.

[0020] It should be understood that the specific embodiments described herein are merely for explaining the present invention, and the descriptions of various technical parameters and algorithm flows are to ensure that those skilled in the art can fully understand and implement the system. Any subsequent non-substantial substitutions or adjustments shall not exceed the scope of this specification.

[0021] Please see the appendix Figure 1 - Appendix Figure 3 The present invention provides an intelligent safety early warning system for production workshops based on the Internet of Things, which is logically divided into a four-layer hardware and software collaborative system: perception layer, communication layer, edge computing layer and middle platform application layer.

[0022] In the hardware deployment and physical structure design of the sensing layer, considering the harsh conditions of industrial production environments, the sensors are strictly divided into a three-dimensional acquisition array structure: (1) Internal sensing nodes of the equipment: These nodes are mainly deployed inside the cavity where there is high pressure, high temperature or high speed physical collision reaction, or attached to the core load-bearing components. These nodes usually use industrial-grade acoustic emission sensors (the acquisition frequency band is usually set between 50kHz and 1MHz to capture weak high-frequency sound waves generated by metal microcracks or internal abnormal friction), miniature pressure sensors (using MEMS technology, with metal isolation diaphragm structure to resist corrosion and high temperature) and miniature contact temperature probes.

[0023] (2) Equipment operation sensing nodes: deployed in key external locations such as equipment shell, motor bearing housing, inlet and outlet. Typical devices include triaxial piezoelectric accelerometer (used to acquire triaxial vibration signals in a wide frequency range), Hall current sensor (non-contact clamped on the power supply bus) and global ambient temperature probe.

[0024] (3) Multispectral vision sensing nodes: Multispectral camera arrays are suspended at key viewing positions such as the top ceiling of the workshop, the entrance of confined space around hazardous equipment, and the material stacking area. The multispectral vision device integrates a high-resolution visible light sensor, an uncooled infrared focal plane thermal imaging detector (response band of 8-14μm), and a time-of-flight (ToF) or binocular stereo vision lens for depth feature acquisition. This array can capture the two-dimensional and three-dimensional shape and temperature field distribution of the scene in all weather conditions, unaffected by workshop dust or drastic changes in lighting. All of the above sensing nodes integrate communication module chips that conform to the Industrial Internet of Things standard at the hardware level.

[0025] In the construction of the communication layer and the protocol interaction mechanism, the system breaks through the pain point of congestion caused by the use of a single frequency band in traditional workshops, and builds a heterogeneous wireless converged network of 5G+LoRa+Wi-Fi 6: (1) For the high-definition video stream and thermal imaging spectrum generated by the multispectral visual sensing node, the data has the characteristics of "large bandwidth and low latency". The system adopts wireless network transmission based on Wi-Fi 6 (802.11ax) protocol. The OFDMA (Orthogonal Frequency Division Multiple Access) and MU-MIMO (Multi-User Multiple Input Multiple Output) technologies adopted by Wi-Fi 6 enable multiple cameras to upload data simultaneously in the same space without causing serious channel contention.

[0026] (2) For equipment operation sensing nodes (such as vibration and current), the data is uploaded periodically or in bursts triggered by events. The data throughput is moderate but extremely sensitive to packet loss. The system adopts a 5G standalone (SA) private network deployed inside the workshop and uses the URLLC (ultra-reliable low-latency communication) feature of 5G to ensure the data with control loop requirements.

[0027] (3) For the internal sensing nodes of the equipment, due to their large number, wide distribution and often deep metal shielding, the system adopts the LoRaWAN low power wide area network protocol and connects to the edge gateway at the top of the workshop through a star topology architecture. LoRa's spread spectrum modulation technology gives it excellent penetration capability and extremely low radio frequency power consumption.

[0028] More importantly, the communication layer embeds a channel quality assessment mechanism in the gateway's underlying driver. Movement of large metal equipment or starting and stopping of high-power frequency converters within the workshop can cause sudden changes in the channel environment. The system gateway reads the physical layer parameters of the RF underlying layer at a set sampling frequency: received signal strength indication, background noise level, and number of packet loss retransmissions, using the formulas described in the manual. Gateway microprocessor real-time computing The value is set by the configured software-defined networking (SDN) controller to dynamically switch over a threshold value. Once continuous Within the period When the index falls below the threshold, the SDN controller immediately sends a control message to force the communication link between the sensing node and the gateway to hop to a pre-scanned silent idle channel, thereby cutting off the physical layer link fading caused by multipath reflection or electromagnetic interference.

[0029] In the logical processing mechanism of the edge computing layer, the system pushes the massive computing power demand that was originally all handled in the cloud down to the workshop site. The edge nodes can be physically represented as ruggedized industrial computing boxes equipped with NPUs (embedded neural network processors), performing differentiated processing on different sensor data: (1) The visual edge node is directly connected to the Wi-Fi 6 receiver and internally runs a mapping function based on a lightweight network architecture (such as a pruned MobileNet series network). Instead of uploading the original video stream, the nodes extract high-order semantic features frame by frame, such as the coordinates of key points on the skeleton of the person and the boundaries of abnormal high-temperature connected domains in thermal imaging images, which greatly reduces the bandwidth usage of the backbone network.

[0030] (2) The edge nodes of the equipment are connected to the 5G receiving module. The autoregressive model or support vector machine is used to detect anomalies in the operating characteristics of the equipment. For example, the waveform envelope of the current data is calculated in the time domain, or the frequency band energy ratio is extracted after the vibration signal is subjected to fast Fourier transform (FFT).

[0031] (3) The distributed sensing edge node acts as a coprocessor for the LoRa gateway. After receiving the timing waveform, it adopts a spatiotemporal attention mechanism to overcome the spatiotemporal drift of acoustic emission and vibration signals, using the formula The attention weight matrix is ​​calculated to achieve time-series focusing and amplification of weak precursor anomalous signals.

[0032] In the data aggregation and decision-making mechanism of the middle platform application layer, a high-performance computing cluster based on containerization technology is deployed. The middle platform receives three modal dimensionality reduction features uploaded via edge nodes, namely the device's internal feature vectors. Equipment operation feature vector and visual feature vectors First, the graph neural network engine of the middle platform initializes the multimodal causal graph model. Each modality's data features and its sub-features form the set of nodes in the graph network. And by utilizing prior knowledge of the temporal order and physical correlation of each modality, an initial set of directed edges is constructed. and its corresponding adjacency matrix .

[0033] The multi-layer iterative computation of graph convolutional networks follows the formula... The deep diffusion and causal relationships of execution information can be learned through the network's weight parameter matrix. The continuous updating filters out superficial correlations (such as the normal phenomenon of a slight increase in equipment casing temperature due to an increase in ambient temperature), and uncovers deeper causal relationships (such as increased internal friction leading to increased high-frequency acoustic emission energy, which in turn causes current fluctuations and localized high brightness in thermal infrared images), ultimately through formulas. The risk level continuous assessment score, accurate to the decimal point, is calculated. (100 marks)

[0034] The final application layer is based on Threshold segmentation (e.g., [0,25) for low risk, [25,60) for medium risk, [60,85) for high risk, and [85,100] for emergency risk) converts the warning signal into control commands via industrial Ethernet (e.g., PROFINET or EtherCAT), and sends them to programmable logic controllers (PLCs) and distributed control systems (DCS) to execute precise millisecond-level equipment linkage closed-loop control.

[0035] The following two specific implementation examples from different industrial scenarios further illustrate the detailed implementation of the above system: Example 1: Application in high-risk powder production workshops This embodiment takes a large neodymium iron boron powder production workshop as an example. Neodymium iron boron powder is an extremely flammable and explosive substance. The core equipment in its preparation process is the air jet mill, which uses high-speed inert airflow to make the material collide and crush with each other. If there is abnormal feeding or wear of the air jet mill lining, the material may experience abnormal high-intensity mechanical friction in some areas, generating hidden hot spots. Once localized heat buildup triggers a tiny spontaneous combustion reaction in the powder, it can lead to a catastrophic dust explosion. Such transient hotspot risks are difficult for traditional external temperature control systems to detect in their early stages.

[0036] In the aforementioned NdFeB powder production workshop, the specific physical deployment and parameter configuration of the system are as follows: First, a sensing layer is configured. In the main cavity of the air jet mill, below the classifier wheel, and other hidden and easily worn areas, eight acoustic emission micro sensors are distributed in a distributed manner using a corrosion-resistant threaded hard connection method. The center response frequency of these acoustic emission sensors is set to 150kHz, which is used to highly sensitively capture the high-frequency elastic wave strain energy released instantaneously by abnormal material impact or metal friction. Meanwhile, a miniature pressure sensor is installed inside the cavity to obtain the rate of pressure change within the cavity. To prevent localized airflow blockage, the internal sensing nodes of these devices are connected to an internal LoRa transmitter module, which transmits data outward through the cavity via a 470MHz frequency band.

[0037] Meanwhile, piezoelectric vibration sensors with magnetic mounting bases were deployed on the external main bearing housing, external feed pipe section, and exhaust fan surface of the air jet mill to measure the root mean square value of the three-phase operating current of the motor. Current transformers and monitoring of local surface temperature data of external equipment The contact-type thermal resistors, these external operating sensor nodes are connected to the workshop's 5G private network via 5G CPE terminals.

[0038] To monitor the overall macroscopic environment of the workshop, four multispectral cameras were evenly deployed on the steel truss at the top of the workshop. Each camera includes a 4K visible light lens and a high-sensitivity thermal infrared lens. The visual data is transmitted back in real time through Wi-Fi 6 wireless access points (APs) installed in the workshop.

[0039] Because the air jet mill contains a high-speed rotating, high-power variable frequency motor, and the workshop is densely packed with stainless steel pipelines, it generates extremely strong electromagnetic interference (leading to...). (Indicators surge) and severe radio frequency signal multipath reflection (leading to) (With persistently high performance indicators), during system operation, the LoRa gateway edge node continuously assesses the channel status. During one particular crushing shift, the movement of a large gantry crane caused severe reflection interference, and the gateway measured... A landslide has occurred; calculations are performed using the parameters to assess the situation: assuming... The weight is configured as 1.2. It is 0.8. The wireless communication channel quality assessment index is 1.5, calculated using the formula. Value within the judgment time window length of the early warning system (Set to 100ms) It drops sharply to below the quality threshold. The system triggers the adaptive dynamic mechanism of the communication layer. The control module immediately broadcasts a frequency hopping command to all LoRa nodes, smoothly switching the communication frequency band from the congested sub-channel to the backup spread spectrum channel with strong anti-interference capabilities. This ensures that the packet loss rate of subsequent acoustic emission feature vector data transmission is less than 0.1%, perfectly solving the communication stability problem under strong interference environment.

[0040] When the data converges at the edge layer, the preprocessing module in the edge computing box begins to work. For acoustic emission data, a spatiotemporal attention mechanism is used to map the input feature tensor to... , and Calculate the spatiotemporal attention weight matrix By fusing time-series signals from eight spatially distributed sensors, highly representative high-frequency time-series feature vectors were accurately extracted. The visual edge nodes utilize lightweight mapping functions. In real time, thermal image streams were used to select temperature-rich regions that abruptly increased relative to the ambient background, and visual feature vectors were constructed. .

[0041] After all edge nodes complete millisecond-level feature dimensionality reduction locally, (mainly by) and constitute), (Depend on (and vibration signal frequency domain characteristics, etc.) and These feature vectors are simultaneously sent to the application layer of the middle platform, which then constructs a causal graph based on them. Directed connections are established between multimodal nodes. For example, the knowledge base assigns a surge in internal acoustic emission as the "cause," with directed edges pointing to the "effects" of abnormal current fluctuations and increased external infrared temperature, through formulas. After performing multi-layer graph convolution operations, the fused multimodal feature matrix is ​​fed into a fully connected multilayer perceptron network. middle.

[0042] During the actual measurement, when a certain area inside the air jet mill experienced very early high-speed dry friction due to material accumulation, the acoustic emission node first captured the characteristic abrupt change inside. However, at this time, the external vibration and current only fluctuated very weakly, and traditional solutions could not identify this as a risk. The causal graph convolutional network of this system accurately determined that this was not background noise, but the early stage of internal hotspot generation. Subsequently, the thermal infrared mode of the multispectral camera captured a tiny abnormal temperature rise of 0.5 degrees Celsius on a specific outer surface of the equipment after a very short time. (The feature weights change accordingly through the system's causal mechanism). Because a strong correlation between the internal source and the external manifestation is obtained, the risk level continuous evaluation score is calculated by substituting these values ​​into the overall assessment formula. The score quickly climbed to 82 points, which was strictly judged by the system as "high risk". The overall calculation and judgment response time was extremely short, and the system could detect transient hotspot risks 200ms in advance.

[0043] Subsequently, the linkage module of the application layer of the middle platform automatically triggered cross-system collaborative intervention: the system instantly sent PID control parameters to the electromagnetic actuator of the air intake valve of the air mill through the bus, reducing the high-pressure air intake by 40% to reduce the crushing intensity; At the same time, a start command is issued to the environmental control subsystem, which is linked to the explosion-proof ventilation system to start at maximum power to enhance local exhaust heat dissipation; A pulse signal is sent to the security system to forcibly cut off the power supply to the magnetic lock of the entrance access control of the crushing room to lock the area access control, activate the local workshop's audible and visual alarm, and strictly prohibit any personnel from entering. After long-term actual operation, this early warning mechanism, based on causal reasoning, has successfully eliminated the misleading effects of vibrations from external large equipment. The overall false alarm rate of the system is only 2.1%, far lower than the 12.5% ​​of the traditional threshold scheme used before the workshop was renovated. It has successfully avoided many spontaneous combustion or dust explosion accidents that could have been caused by the accumulation of transient hot spots, and achieved true proactive risk elimination.

[0044] Example 2: Application in a machining workshop This embodiment further illustrates the implementation of the system of the present invention in an automotive parts machining workshop. The workshop is equipped with a large number of five-axis linkage CNC machine tools and automated robotic arms. The core safety risks in the machining process include two dimensions: First, fatigue spalling or transient jamming of the CNC machine tool spindle bearing or feed screw due to long-term heavy-load cutting can lead to scrapped parts or broken tools that fly out and injure people. Second, serious personal injury accidents caused by operators in the workshop violating safety protection areas and intruding into the working envelope of a high-speed robotic arm or machine tool.

[0045] To address this complex risk scenario, the system of this invention has been specifically deployed and configured as follows: In terms of the sensing layer, piezoelectric contact acoustic emission sensors and miniature hydraulic sensors are deployed in close contact with the spindle housing and the sidewalls of the oil cooling channel cavity near the high-speed bearing in the CNC machine tool. This is to capture the release of high-frequency stress waves caused by the early spalling of bearing ball material. Vibration velocity RMS values ​​are also deployed on the machine tool bed and outside the motor housing. Sensors and three-phase power supply current The data acquisition system enables multi-level monitoring of both internal and operational characteristics of the equipment. Furthermore, six multispectral cameras are suspended high on a pan-tilt-zoom (PTZ) unit in key passageways, blind spots of robotic arm enclosures, and gaps between multiple machine tools throughout the processing workshop. These cameras not only monitor visible light but also utilize integrated depth cameras to acquire 3D point cloud distance feature tensors. These depth image data can accurately reconstruct the 3D skeleton trajectory of a person and accurately assess the three-dimensional straight-line distance between the human body and dangerous moving parts.

[0046] The communication layer also adopts a heterogeneous networking mechanism. Because machine tools generate a lot of sparks and radio frequency noise when cutting metal, and large metal cranes are constantly moving in the factory area, traditional 2.4GHz Wi-Fi signals are easily blocked or overwhelmed by electromagnetic interference. Therefore, during implementation, the system adopts a mechanism of mixed coverage of 5G private network and Wi-Fi 6 AP. The visual nodes push a large amount of fused video stream containing distance information and image information to the middle platform at high speed. Vibration and current data of the equipment are transmitted with extremely low latency. The 5G channel reports periodically; if the edge gateway calculates the signal-to-noise ratio parameter... If attenuation occurs and channel impairment is determined through an evaluation formula, more idle bandwidth will be automatically scheduled and allocated for transmission rate. Ensure the complete delivery of visual violation characteristics and high-frequency equipment fault characteristics data to critical affected nodes.

[0047] During operation, when the machine tool is cutting metal at high speed, the edge computing device uses an anomaly detection model in real time to analyze and extract the high-frequency envelope of acoustic emission and the effective value of external broadband vibration velocity. Since cutting itself is a high-energy physical process, traditional simple threshold judgment is prone to misreporting normal heavy cutting conditions as equipment failures. However, in the system of this invention, edge computing nodes compress features into vectors and upload them, and the multimodal spatiotemporal risk fusion assessment module of the middle platform is activated. The module constructs a complete causal graph from vibration, acoustic emission, electrical characteristics, and personnel trajectories and environmental behavior features identified through a visual network. For example, in a node cluster, at a certain point in time, the system detects a sharp increase in the high-frequency acoustic emission characteristics inside a machine tool, but at the same time, the convolutional network in the middle platform graph... Through reasoning, it was discovered that this sudden increase in characteristic was accompanied not only by external vibrations. The increase in [something] corresponds to the increase in spindle current. The regular pulse rises, and the visual characteristics The analysis indicates that the current roughing program is executing a pre-set tool entry action. In this case, the graph network automatically adjusts its internal causal transmission coefficients, determining that the phenomenon is a normal cutting load change under the control of the operation command, and finally calculates the output. The score remains in the low-risk range, and the system only pushes normal status information, perfectly suppressing false alarms.

[0048] However, in another real-world early warning scenario, the visual node detected that an employee was not wearing a compliant safety helmet, and their 3D skeletal trajectory was detected using depth features. The calculation determined that the person had suddenly crossed the yellow safety warning line marked on the ground due to operational error and directly entered the dangerous area of ​​the high-speed machine tool chuck. At this time, the lightweight target detection and distance calculation algorithm running on the visual edge node output an abnormal person violation boundary feature vector with extremely high confidence within a few milliseconds.

[0049] Almost simultaneously, due to unintentional interference from the operator (such as slight loosening of the workpiece due to human contact), the edge node of the equipment detected nonlinear high-frequency oscillations in the machine tool vibration characteristics that did not conform to the current cutting cycle, and the current... The data exhibits irregular fluctuations.

[0050] These two seemingly independent modal feature sources ( This indicates unauthorized intrusion by personnel and The transient abnormal vibration of the device (represented by the model) is instantaneously injected into the graph network model in the middle platform. Through iterative formulas between network layers, the model is assigned a visual trajectory feature vector. With extremely high causal weights, the middleware logic maps functions through a multilayer perceptron fully connected network in less than 100 milliseconds. Calculation The score is 94 points, which determines that the current situation belongs to the "emergency risk" level, which may lead to serious injury or even death.

[0051] Once the highest-level warning is triggered, the cross-domain linkage control mechanism is immediately activated. The system does not require any manual confirmation; the command goes directly to the lowest-level control hardwired loop. The system's main control unit urgently sends a high-priority PROFINET data frame containing an emergency stop register to the CNC system of the machine tool in the area of ​​unauthorized intrusion. Upon receiving the command, the machine tool's feed axis servo motors and spindle drivers execute a level 0 safety stop mechanism (cutting off motor power and applying mechanical brakes). The spindle is forcibly stopped at a high speed of 8000 rpm in less than 0.3 seconds. At the same time, the linkage system locks the protective door of the operating panel where personnel attempt to approach further. The remote monitoring center screen of the command safety management platform is forced to display a red warning pop-up window for that view and activates an ultra-high decibel alarm and warning flashing lights on site.

[0052] Extensive field data analysis and testing have verified that the system achieves a 98.2% accuracy rate in predicting machine tool internal faults. Furthermore, for transient human-machine interface risks caused by personnel misconduct, the entire response loop—from front-end image acquisition, feature extraction, multimodal causal graph evaluation and decision-making to the final forced shutdown of the machine tool—is strictly controlled within 0.5 seconds. This demonstrates that the cross-modal causal reasoning and intelligent control strategy, from passive alarm to proactive multi-level response, of this invention completely overturns the previous isolated security camera monitoring model, powerfully and effectively reducing the incidence of serious safety accidents in machining workshops through IoT technology.

[0053] In summary, this invention, by constructing a converged communication network with heterogeneous sensing hardware, adaptive channel adjustment, an edge computing system with feature mapping and attention enhancement, and a middleware processing architecture with cross-modal causal graph network analysis, thoroughly solves the problems of slow multi-source risk perception, poor anti-interference communication capabilities, and inability to implement effective full-domain linkage in response to sudden hidden dangers in complex production workshop environments. The above are only specific preferred embodiments of this invention, but the core inventive concept of this invention is not limited thereto. Any equivalent substitutions or functional changes made by those skilled in the art within the scope of the technology disclosed in this invention by using formulas, methods, and networking means should be included within the scope of technical protection claimed in this specification.

Claims

1. An intelligent safety early warning system for production workshops based on the Internet of Things, characterized in that, The system comprises a perception layer, a communication layer, an edge computing layer, and a middleware application layer that are interconnected from bottom to top. The perception layer uses a multimodal intelligent sensor array to build a global perception network in the production workshop. The multimodal intelligent sensor array includes internal sensing nodes, equipment operation sensing nodes, and multispectral visual sensing nodes. The communication layer constructs a heterogeneous wireless communication network based on the convergence of 5G, LoRa and Wi-Fi 6 networks, and the communication layer has a built-in adaptive dynamic channel allocation mechanism based on the channel quality assessment model to switch transmission channels in real time. The edge computing layer deploys differentiated heterogeneous edge computing nodes, including visual edge nodes, device edge nodes, and distributed sensing edge nodes. The edge computing layer uses a spatiotemporal attention mechanism to locally fuse and enhance the extracted local spatiotemporal features. The middleware application layer includes a multimodal spatiotemporal risk fusion assessment module and a cross-domain linkage hierarchical early warning module; The multimodal spatiotemporal risk fusion assessment module obtains the device's internal feature vector uploaded by the edge computing layer. Equipment operation feature vector and visual feature vectors And construct a multimodal causal graph model. The continuous risk level assessment score is calculated and output through a graph convolutional network model. ; The cross-domain linkage hierarchical early warning module is based on The scope is mapped to a level four risk state, and control commands are issued to execute cross-system coordinated intervention.

2. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, The internal sensing nodes of the device are used to collect transient local latent risk data inside the device; The device operation sensor node is used to collect external operating status data of the device; The multispectral visual sensing nodes are deployed in key areas of the workshop to simultaneously collect visible light, thermal infrared, and depth image data to monitor personnel and environmental anomalies.

3. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, In the heterogeneous wireless communication network, the data of the multispectral visual sensing nodes is transmitted through the Wi-Fi 6 network, the data of the device-operated sensing nodes is transmitted through the 5G private network, and the data of the sensing nodes inside the device is transmitted through the LoRa network.

4. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, The adaptive dynamic channel allocation mechanism calculates a wireless communication channel quality assessment index. And set the quality threshold limit for triggering dynamic channel switching. Real-time switching of transmission channels; The wireless communication channel quality assessment index The calculation formula is: ； in, The signal-to-noise ratio parameter is measured at the communication receiver. The signal multipath reflection interference intensity of metal equipment in the workshop; The noise floor energy intensity under strong electromagnetic interference environment; This represents the available bandwidth transmission rate of the current wireless channel. This refers to the average end-to-end transmission delay of data packets. This is the signal-to-noise ratio gain coefficient; This is the penalty coefficient for signal reflection interference; This is the electromagnetic interference penalty coefficient; These are the standardized weighting coefficients for bandwidth rate in the channel quality assessment model; These are the standardized weighting coefficients for delay time in the channel quality assessment model.

5. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 4, characterized in that, When satisfied At that time, the system is within the given decision time window length of the early warning system. The system automatically switches the communication link to a spare frequency band.

6. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, The visual edge nodes are used to run a lightweight edge feature extraction mapping function. Complete feature extraction from multimodal images; The device edge nodes are used to run an anomaly detection model to extract the temporal features of vibration and current data; The distributed sensing edge nodes are deployed at the workshop edge gateway and are used to extract high-frequency time-series feature vectors of acoustic emission signals. .

7. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, The spatiotemporal attention mechanism is used to enhance the spatiotemporal correlation between acoustic emission signals and vibration signals, and the spatiotemporal attention weight matrix it generates... The calculation formula is: ； in, The query vector in the spatiotemporal attention mechanism is obtained by linear mapping from the input feature sequence. This refers to the key vector in the spatiotemporal attention mechanism; This refers to the value vector in the spatiotemporal attention mechanism; Key vector Feature mapping dimension parameters; This indicates a transpose operation on the matrix formed by the key vectors; This is the softmax activation function used for normalization.

8. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 1, characterized in that, The multimodal causal relationship graph model Includes a set of node states With the set of directed edges The node state set It consists of the mapped multimodal feature matrix; The graph convolutional network model in the first... Hidden node state matrix at +1 level The update formula is: ； in, This is the adjacency matrix of the causal graph model, representing the strength of causal connections between multimodal nodes; This is the degree matrix corresponding to the adjacency matrix; For graph convolutional network models in the 1st... The hidden node state matrix of the layer, when =0 represents the initial features of the input multimodal nodes; For graph convolutional network models in the 1st... The learnable network weight parameter matrix of the layer; This is the ReLU activation function used for nonlinear activation of the model.

9. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 8, characterized in that, The risk level continuous evaluation score The calculation formula is: ； in, For internal feature vectors of the device; This represents the device's operational feature vector. For visual feature vectors; These are the weighting coefficients of the device's internal feature vectors in the causal graph; These are the weighting coefficients of the device's operational feature vectors in the causal graph; These are the weight coefficients of the visual feature vectors in the causal graph; This represents the norm operation performed on the eigenvector; This is the output feature matrix of the last layer of the graph convolutional network model; For a fully connected multilayer perceptron network mapping function; The risk level continuous evaluation score The range is normalized to 0 to 100.

10. The intelligent safety early warning system for a production workshop based on the Internet of Things according to claim 9, characterized in that, The cross-domain linkage hierarchical early warning module continuously evaluates the risk level score. Mapped to a four-level risk status: The range [0, 25) represents a low-risk state. The risk level is considered medium within the range of [25, 60). The range [60, 85) is considered a high-risk area. The range [85, 100] represents an emergency risk state; The cross-domain linkage hierarchical early warning module issues control commands based on different risk states, and performs cross-system linkage interventions including equipment control, access control locking, alarms and emergency shutdowns.

Images