Unattended intelligent control system and monitoring method based on multi-source data fusion

Abstract

The application discloses an unattended intelligent control system and a monitoring method based on multi-source data fusion, which is applied to physical equipment monitoring in an evaporation station of alumina production and comprises three layers from bottom to top, namely, a bottom data acquisition layer, a middle data fusion and edge processing layer and a top remote centralized control and interaction layer, wherein the bottom data acquisition layer comprises a PLC control system, a temperature and vibration integrated sensor assembly and a suspended rail inspection robot mobile sensing terminal. The application integrates the PLC, the temperature and vibration sensor and the suspended rail robot to perform data acquisition and inspection, and builds a unified intelligent decision platform, deeply fuses process control, equipment management and safety management and the like, realizes true unattended operation, can replace manual work to complete most of monitoring, adjustment and abnormal treatment work, greatly reduces the dependence on manpower, and can optimize equipment operation conditions, improves equipment stability, efficiency, service life and safety.

Description

Technical Field

[0001] This invention relates to an unattended intelligent control system and monitoring method based on multi-source data fusion, belonging to the field of intelligent monitoring technology for alumina production equipment. Background Technology

[0002] The evaporation process in alumina production is a high-energy-consuming stage, and its stable and efficient operation is crucial to the overall plant efficiency. Currently, most evaporation stations rely on manual, experience-based operation of distributed control systems (DCS) or PLCs, which has significant limitations. The monitoring dimensions are limited: relying on traditional temperature, pressure, and flow meters, there is a lack of online real-time monitoring of the health status (vibration, bearing temperature) of key equipment (such as pumps and motors), resulting in delayed fault warnings; Inspections rely on manpower: Regular manual inspections are required to check for leaks, instrument readings, and equipment appearance, which is inefficient, poses safety risks, and cannot achieve continuous monitoring. Data silos and insufficient linkage: Equipment status data, robot vision data, and process parameter data are isolated from each other, and the control logic is rigid, making it impossible to perform adaptive optimization and collaborative control based on the fusion of multi-dimensional information; High staffing costs: To ensure safe operation, staff need to be on duty 24 hours a day, which results in high labor costs and is prone to misjudgment due to fatigue.

[0003] With the development of the Industrial Internet of Things (IIoT) and artificial intelligence (AI), some technologies have attempted localized improvements, such as deploying wireless vibration sensors in critical units or introducing inspection robots for image acquisition. However, these technologies have the following shortcomings: (1) At the data acquisition level, most existing rail-mounted robots are equipped with only a single type of sensor (such as only having visible light imaging or infrared temperature measurement functions), which makes it difficult to comprehensively collect various physical quantities of the equipment. Data from a single sensor cannot fully reflect the operating status of the equipment, resulting in a significant risk of missed and false alarms in the monitoring results; (2) At the data processing level, a large amount of raw data is directly transmitted back to the cloud or backend server, resulting in high network bandwidth pressure and high transmission latency. Moreover, cloud processing cannot meet real-time requirements. Existing systems lack edge computing capabilities close to the data source, making it impossible to perform real-time preprocessing and preliminary analysis of massive amounts of sensor data. (3) At the system architecture level, most existing rail inspection systems are single-layer or two-layer architectures. The bottom layer data acquisition and the upper layer application are tightly coupled, lacking a clear layered decoupling design, resulting in insufficient system scalability and flexibility. There is a lack of standardized data interfaces and collaboration mechanisms between the various functional modules, making it difficult to achieve effective integration and comprehensive utilization of multi-source data; (4) In terms of unattended operation capability, the existing technologies are mostly independent subsystems, which have failed to achieve deep data integration and intelligent closed-loop control with the core control system (PLC), and have failed to fundamentally realize "unattended operation".

[0004] Therefore, there is an urgent need for a solution that can deeply integrate equipment status, machine vision and process control, and drive PLC and robot collaborative operation through a central intelligent system. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide an unattended intelligent control system and monitoring method based on multi-source data fusion, which can realize a rail-mounted robot monitoring system with comprehensive multi-source data acquisition, real-time edge processing, and remote centralized control, so as to solve the problems of single data acquisition, poor real-time processing, insufficient system scalability, and low level of intelligence in the prior art.

[0006] The objective of this invention is achieved through the following technical solution: An unattended intelligent control system based on multi-source data fusion is applied to the monitoring of physical equipment in an alumina production evaporation station. The system comprises a three-layer architecture from bottom to top: a bottom-level data acquisition layer, a middle-level data fusion and edge processing layer, and a top-level remote centralized control and interaction layer. The bottom data acquisition layer includes a PLC control system, an integrated temperature and vibration sensor assembly, and a rail-mounted inspection robot mobile sensing terminal. The PLC control system is used for process control of the physical equipment in the evaporation station; the integrated temperature and vibration sensor assembly is used for real-time monitoring of key rotating equipment in the evaporation station; and the rail-mounted inspection robot mobile sensing terminal is used for inspection of the entire monitoring area. The middle layer data fusion and edge processing layer includes an edge computing gateway and a robot base station. The edge computing gateway is used to receive real-time data from the PLC and temperature and vibration sensor data, and to aggregate and preprocess the data. The robot base station is used to receive task instructions from the cloud and schedule the robot to execute them, and can also perform local real-time analysis of the high-definition video streams collected by the robot. Top-level remote control and interaction layer: The top-level remote control and interaction layer is a cloud-based intelligent analysis layer, serving as the system's command and visualization hub. It enables full-domain monitoring and the highest level of human-machine interaction, including a cloud-based intelligent analysis platform and a 3D digital twin monitoring platform. The cloud-based intelligent analysis platform receives, stores, and processes historical and real-time data from lower layers, deploys and runs artificial intelligence models to perform data analysis and prediction tasks, and generates control decision commands based on preset rules and model output results. The 3D digital twin monitoring platform is used to construct a high-fidelity 3D model of the evaporation station, drive the model's status to synchronize with the site, map data changes to the dynamic display of 3D graphic elements, and provide user interaction and data transmission between the user and the 3D scene.

[0007] Furthermore, the integrated temperature and vibration sensor assembly consists of several integrated temperature and vibration sensors installed on various rotating equipment in the evaporation station. The integrated temperature and vibration sensor adopts an industrial wireless sensor, which integrates a high-frequency vibration accelerometer, an embedded signal processing unit, a platinum resistance temperature sensor, and a low-power wireless transmission module.

[0008] Furthermore, the mobile sensing terminal of the rail-mounted inspection robot includes the rail-mounted robot body, the track laid above the monitoring area, and a multi-source sensor assembly mounted on the rail-mounted robot body.

[0009] Furthermore, the track is a ring-shaped rigid track laid above the alumina production evaporation station. The track laying path covers all key equipment, and RFID or QR code tags are provided on both sides of the track to provide absolute position reference for the robot.

[0010] Furthermore, the rail-mounted robot body is equipped with a central processing unit and a communication module, and has an intelligent mobile chassis at the bottom. The intelligent mobile chassis integrates a drive system and a power management module that are telecommunicationly connected to the central processing unit.

[0011] Furthermore, the multi-source sensor assembly includes: a visible light camera for acquiring images of the device's appearance and meter readings; an infrared thermal imager for acquiring the device's temperature distribution; an ultrasonic pickup and lidar for obstacle avoidance and positioning; and also includes an acoustic and environmental perception unit, an auxiliary lighting unit, and a positioning unit.

[0012] Furthermore, the mid-layer data fusion and edge processing layer is responsible for local data processing and agile response: The edge computing gateway is an embedded computer deployed in each process area. It is connected to the PLC control system via industrial Ethernet and to the temperature and vibration sensor components via LoRa wireless network. The robot base station functions as a charging dock, communication hub, and edge server. It is an integrated workstation equipped with an edge vision server and communication module, and maintains continuous communication with the rail-mounted inspection robot via high-speed Wi-Fi 6.

[0013] Furthermore, the cloud-based intelligent analysis platform is deployed on an enterprise private cloud or industrial internet platform using a Kubernetes-based microservice architecture, including a data platform service, an AI model service container group, and a rule engine and decision service. The data platform service, AI model service container group, and rule engine and decision service communicate with each other through the microservice architecture and share a unified data bus. The three-dimensional digital twin monitoring platform is deployed on the central control center's large screen and engineer's station via WebGL or a standalone client, and includes model and data-driven, visualization mapping, and interactive interfaces; The industrial internet platform sends the decision results to the interactive interface module of the 3D digital twin monitoring platform through the rule engine and decision service, so as to display the decision effect in the 3D scene.

[0014] An intelligent monitoring method based on the above system includes the following steps: Step S1: Based on the synchronization mechanism, high-precision time synchronization acquisition is performed on three core data sources through three types of interface protocols, including process data acquisition from PLC control system, data acquisition from integrated temperature and vibration sensor, and data acquisition from rail-mounted inspection robot. Step S2: Edge-side preprocessing and fusion Within edge computing gateways and robot base stations deployed on-site, lightweight processing and initial fusion of data from the edge, close to the data source, are performed to improve data quality and reduce cloud load. Specifically, this includes: Data preprocessing: Filtering, rationality verification and engineering unit conversion of raw PLC data; performing time-domain noise reduction and fast Fourier transform on vibration waveform data; and preliminarily calculating spectral characteristics. Edge intelligence and feature extraction: Within the robot base station, a lightweight YOLO or MobileNet model is run to analyze the acquired images in real time and directly output anomaly labels instead of uploading the original video stream; Feature-level fusion and uploading: The processed process time series, equipment vibration features, and structured visual recognition results are packaged according to a unified time frame to generate a standard edge fusion data packet and uploaded to the cloud data platform; Step S3: Perform in-depth mining and collaborative analysis on the fused data. In the cloud-based microservice cluster, three core analytics engines run in parallel to perform in-depth mining and collaborative analysis on the merged data. These three core analytics engines specifically include: Equipment health diagnosis model: Receives vibration spectrum and temperature trend data, analyzes spectrum characteristics through a pre-trained deep convolutional neural network model, automatically identifies fault modes, and calculates a quantitative real-time equipment health index; Process optimization model: Based on real-time process data stream, combined with the equipment performance degradation coefficient provided by the equipment health index, the model predictive control algorithm is used to dynamically constrain the equipment health status using the health index and specific failure mode information as efficiency coefficients. Visual depth analysis model: Perform secondary depth analysis on abnormal images uploaded by the robot or quantify the scale thickness through image segmentation technology to provide more accurate on-site status input for process optimization; Step S4: The cloud-based decision engine automatically makes judgments based on the analysis results of the three analysis engines, triggering different control branches, specifically including: Set comprehensive evaluation conditions, including but not limited to: a) whether the equipment health index is below a preset threshold; b) whether key process parameters deviate from the optimal range; c) whether visual analysis detects any anomalies requiring intervention. Analyze the comprehensive evaluation conditions using the following logic: If any of the above conditions is "yes", the process will proceed to the abnormal collaborative handling branch corresponding to step S5. Otherwise, it will enter the branch that maintains normal autonomous operation and remain in normal autonomous operation mode; Step S5: The abnormal collaborative handling branch performs parallel control, which includes, but is not limited to: Generate process optimization parameters and send instructions to the PLC to adjust equipment operating parameters; Generate targeted tasks and instructions and issue them to the robot for review and special inspection; Step S6: Feed the execution results back to the analysis and decision center.

[0015] Furthermore, during the data acquisition process of the integrated temperature and vibration sensor, the raw acquired data is preprocessed to predict equipment health. Specific processing steps include: (1) Feature extraction - time domain / frequency domain feature values Massive amounts of raw waveform data are compressed into low-dimensional feature vectors containing fault information, including automated multi-dimensional feature extraction of raw vibration signals on cloud or high-performance edge servers. (2) AI health diagnosis Automatically identify fault modes and quantify health status using CNN-LSTM deep learning models, specifically including: Convolutional Neural Network (CNN) branch: It takes the spectrogram as a two-dimensional image input and automatically learns the spatial distribution pattern of faults in the frequency domain through convolutional layers; Long Short-Term Memory (LSTM) branch: Receives time series composed of temporal features and temperature trends, and learns the evolution of device state over time; Feature fusion and fully connected layer: The high-level features extracted by the CNN and LSTM branches are fused, and the diagnostic results are finally output through the fully connected layer and the Softmax classifier; (3) Visual assessment of health status Transforming diagnostic results into decision-making information usable by engineering includes: setting up health status dashboards, fault type prediction alarms, and remaining life (RUL) estimation. (4) Input the results into the process optimization model to achieve adaptive control.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) It achieves true unmanned operation: Through multi-source data fusion and intelligent decision-making, the system can replace human labor to complete most of the monitoring, adjustment and abnormal handling work, greatly reducing the dependence on human labor.

[0017] (2) Improve system stability and efficiency: Adaptive optimization control based on working condition identification can enable the evaporation station to always operate under optimal or near-optimal working conditions, effectively smooth out production fluctuations, improve the qualified rate of discharge concentration, and significantly reduce steam consumption.

[0018] (3) Achieve predictive maintenance: By monitoring the equipment status in real time and analyzing the data, change "post-event maintenance" to "predictive maintenance", discover potential faults in advance, avoid unplanned downtime, and extend the equipment life.

[0019] (4) Enhance inherent safety: Active safety monitoring combined with video analysis can promptly detect risks such as leaks and fires that are difficult for humans to detect, and automatically link with the safety system, greatly improving the safety of the production site.

[0020] (5) Breaking down information silos: A unified intelligent decision-making platform has been built, which deeply integrates subsystems such as process control, equipment management, and safety management, providing a solid data foundation and application examples for the digital and intelligent transformation of the factory.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein: Figure 1 A schematic diagram of the overall architecture of an unattended intelligent control system based on multi-source data fusion provided in an embodiment of the present invention; Figure 2 A schematic diagram of the overall architecture of the rail-mounted inspection robot; Figure 3 The flowchart illustrates the process of an unattended intelligent control method based on multi-source data fusion, as provided in an embodiment of the present invention. Figure 4 This is a logic diagram for predictive maintenance of system health provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0024] like Figures 1-2 As shown, the first embodiment of this invention provides an unattended intelligent control system based on multi-source data fusion, which is applied to an alumina production evaporation station. The physical equipment of the evaporation station includes a multi-effect evaporator, a raw liquid pump, a circulating pump, a condensate pump, various pneumatic / electric regulating valves, a vacuum system, etc. The system comprises a three-layer architecture from bottom to top: a bottom-layer data acquisition layer, a middle-layer data fusion and edge processing layer, and a top-layer remote centralized control and interaction layer. The bottom data acquisition layer includes a PLC control system, an integrated temperature and vibration sensor assembly, and a rail-mounted inspection robot mobile sensing terminal. The PLC control system is used for process control of the physical equipment in the evaporation station; the integrated temperature and vibration sensor assembly is used for real-time monitoring of key rotating equipment in the evaporation station; and the rail-mounted inspection robot mobile sensing terminal is used for inspection of the entire monitoring area. The middle layer data fusion and edge processing layer includes an edge computing gateway and a robot base station. The edge computing gateway is used to receive real-time data from the PLC and temperature and vibration sensor data, and to aggregate and preprocess the data. The robot base station is used to receive task instructions from the cloud and schedule the robot to execute them, and can also perform local real-time analysis of the high-definition video streams collected by the robot. Top-level remote centralized control and interaction layer: This serves as the command and visualization hub of the system, enabling full-domain monitoring and the highest level of human-machine interaction. It includes a cloud-based intelligent analysis platform and a 3D digital twin monitoring platform. The cloud-based intelligent analysis platform adopts a Kubernetes-based microservice architecture and is deployed on an enterprise private cloud or industrial internet platform. The 3D digital twin monitoring platform is developed based on Unity3D or Unreal Engine and is deployed on the central control center's large screen and engineer workstations via WebGL or a standalone client.

[0025] The PLC control system serves as the core control terminal, employing redundant rack-mount PLCs, such as the Siemens S7-1500H series. It directly connects to the actuators (e.g., valves, pump frequency converters) and process sensors (e.g., temperature sensors, pressure transmitters) of all physical equipment in the evaporation station. Its core functions are: first, to execute basic logic and loop control; and second, to act as a "high-speed executor" of intelligent instructions. In addition to traditional control programs, the PLC control system includes an optimized instruction receiving and processing function block (FB). This function block receives optimized setpoints from the cloud via industrial Ethernet and smoothly applies them to the existing control loop using a ramp function or proportional method, achieving seamless switching and ensuring process stability.

[0026] The integrated temperature and vibration sensor assembly consists of several integrated temperature and vibration sensors installed on various rotating equipment in the evaporation station. It is preferably installed on critical rotating mechanical equipment with high reliability requirements and significant impact from failures, such as the drive motors and pump bodies of raw material pumps, various stages of circulating pumps, forced-efficiency circulating pumps, and condensate pumps, measuring axial and radial vibrations (such as acceleration and velocity values), bearing temperatures, etc. The integrated temperature and vibration sensor uses industrial wireless sensors, such as intrinsically safe wireless sensors with built-in edge computing capabilities. These industrial wireless sensors can perform FFT transformations directly at the node while measuring vibration acceleration and temperature, uploading spectral feature values ​​instead of the original waveform, with a battery life of over 3 years. Data is relayed to the edge computing gateway via a self-organizing network. Specifically, this sensor integrates: A high-frequency vibration accelerometer with a sampling frequency of not less than 12.8 kHz is used to capture high-frequency transient signals caused by impact failures (such as bearing spalling). The embedded signal processing unit performs analog-to-digital conversion and preliminary filtering of the vibration signal locally at the sensor node; The platinum resistance temperature sensor (PT100) is installed close to the outer surface of the bearing housing to monitor the temperature change trend of the bearing. Low-power wireless transmission module: Using LoRa or industrial Wi-Fi protocols, encrypted data packets are periodically (e.g., every 5 minutes) or triggered (e.g., when vibration exceeds the limit) to the edge gateway.

[0027] The mobile sensing terminal of the rail-mounted inspection robot autonomously plans its inspection path and task sequence by receiving scheduling instructions from the robot's base station. It includes the rail-mounted robot body, the track laid above the monitoring area, and a multi-source sensor assembly mounted on the rail-mounted robot body.

[0028] The track is a circular rigid track laid above the alumina production evaporation station. The circular rigid track uses high-strength aluminum alloy or steel enclosed I-beams as its main body and is securely installed to the roof truss or side walls of the evaporation workshop via supports. The track laying path requires precise mechanical calculations to ensure coverage of all critical equipment, such as pump sets, valve sets, tank sight glasses, instrument panels, and other visible inspection points, forming a closed loop. In addition, the track can also incorporate high-reliability conductive rails or employ wireless inductive power supply technology to provide continuous power to the robot.

[0029] The rail-mounted robot houses a central processing unit (CPU) and a communication module, and is equipped with an intelligent mobile chassis. The intelligent mobile chassis features a cast aluminum shell with an explosion-proof (Ex d IIC T4 Gb) design, and integrates a drive system and power management module that are telecommunicationly connected to the CPU. The drive system consists of multiple servo motors and high-friction rubber rollers, ensuring smooth movement, precise positioning, and reliable braking even on inclined and slippery tracks, with a positioning accuracy of ±5 mm. The chassis has an IP67 protection rating to withstand the high temperature, high humidity, and alkaline mist environment of the evaporation station. The CPU is a high-performance industrial-grade system-on-a-chip (SoC) based on the ARM Cortex-A series, running a real-time Linux operating system. It is responsible for integrating sensor data, executing navigation algorithms (such as global path planning and real-time obstacle avoidance based on the track map), and controlling the gimbal and payload. The communication module integrates a dual-band Wi-Fi 6 (802.11ax) module and a 5G industrial module as primary and backup communication links, ensuring low-latency and highly reliable transmission of high-definition video streams and command data between the evaporation station physical equipment and the robot base station, supporting breakpoint resume and automatic link switching. The power management module can be equipped with an automatically rechargeable battery. RFID or QR code tags can be pre-embedded on both sides of the track to provide the robot with absolute position references.

[0030] The multi-source sensor assembly includes at least: a visible light camera for acquiring images of the equipment's appearance and meter readings, an infrared thermal imager for acquiring the equipment's temperature distribution, an ultrasonic pickup for obstacle avoidance and positioning, and a lidar.

[0031] The visible light camera is a high-definition optical zoom camera, such as a high-definition starlight-level camera with 30x optical zoom and 4K resolution, mounted on the rail-mounted robot body with at least two degrees of freedom (pitch and horizontal). Combined with a built-in AI acceleration chip, such as Huawei Atlas or Intel Movidius, it can run deep learning models in real time to complete the following core functions: Automatic Instrument Recognition and Reading (OCR): Automatically locates and crops the dial image, recognizes pointer angles or digits, and converts them into engineering values ​​with an accuracy rate >99%. Equipment status visual inspection: Based on trained YOLO and other models, automatically detects appearance abnormalities such as "running, leaking, dripping, and seeping" (such as pipeline leaks and liquid accumulation on the ground), insulation layer damage, and loose bolts.

[0032] The infrared thermal imager is an uncooled infrared thermal imager with a resolution of at least 640x480 and a thermal sensitivity (NETD) of <40mK. It is aligned with the optical axis of a visible light camera to achieve "dual-light fusion." Main applications include: Temperature anomaly monitoring: Automatically scans motor bearing housings, pump casings, valve flanges, electrical cabinets, and other parts to generate thermal images of temperature distribution. The system can be set with temperature alarm thresholds and automatically identifies local hot spots. Fault-assisted diagnosis: By analyzing thermal image features (such as a valve with a significantly higher temperature on one side than the other, indicating internal leakage), key visual evidence is provided for equipment health diagnosis.

[0033] The multi-source sensor assembly also includes an acoustic and environmental sensing unit, an auxiliary lighting unit, and a positioning unit.

[0034] The acoustic and environmental sensing unit includes a highly directional digital microphone array and a multi-gas detection module, primarily used for: Acoustic analysis: Collect equipment operating noise, perform acoustic pattern analysis through edge computing, and extract characteristic frequencies. This can be used to help determine pump cavitation phenomena, bearing dry friction noise, etc. Environmental monitoring: Real-time detection of ammonia (NH3) concentration, temperature and humidity in the environment around the robot, providing environmental context information for safety warnings and process analysis.

[0035] The auxiliary lighting unit includes an integrated LED fill light that supports adaptive brightness adjustment to ensure clear visible light images are still available in low-light conditions at night or in areas at the bottom of the device.

[0036] The positioning unit comprises a combined structure of an inertial measurement unit (IMU), a wheel encoder, and a visual odometry (VO). With the assistance of track tags, it enables continuous and high-precision pose estimation of the robot on long-distance tracks, providing accurate spatial tags for data acquisition.

[0037] The middle-layer data fusion and edge processing layer serves as a bridge connecting the physical and digital worlds, responsible for local data processing and agile response.

[0038] Edge computing gateways are deployed in each process area, aggregating real-time PLC data from up to 500 points within the area via industrial Ethernet and receiving temperature and vibration sensor data via LoRa wireless network. Their core functions include millisecond-level data timestamp alignment, caching, and lightweight edge computing (such as vibration RMS calculation and simple threshold alarms), and uploading the packaged standardized data stream to the cloud. The edge computing gateways utilize industrial-grade ARM or x86 architecture embedded computers, deployed within the evaporation station's field control cabinet. These embedded computers integrate industrial Ethernet (Profinet, EtherNet / IP), Modbus TCP / RTU, OPC UA, and other protocols to access data from the PLC control system's process control core, and are equipped with LoRa or Wi-Fi communication modules to aggregate wireless data packets from the integrated temperature and vibration sensors. The embedded computers incorporate a rule engine and a lightweight AI inference framework (such as TensorFlow Lite) to perform data cleaning, preliminary anomaly threshold judgment, data compression and encryption, and upload the processed standardized data to the cloud via VPN or 5G network.

[0039] In the mid-layer data fusion and edge processing layer, the robot base station functions as a charging dock, communication hub, and edge server. It receives task instructions from the cloud and schedules the robot to execute them. Simultaneously, it can perform real-time local analysis of the high-definition video streams captured by the robot, such as instantly identifying meter readings or detecting obvious leaks, and uploading the results (rather than the raw video stream), significantly reducing network bandwidth pressure. The robot base station connects to the track; when the robot's battery level drops below a threshold, it automatically navigates to the base station and quickly replenishes its power through physical contact or wireless sensing. Furthermore, it serves as a data exchange hub, connecting to the factory network via wired Ethernet and linking to the robots.

[0040] The robot base station is an integrated workstation equipped with an edge vision server and a communication module. It maintains continuous communication with the rail-mounted inspection robot via high-speed Wi-Fi 6, and is responsible for issuing task instructions, monitoring robot status, and managing paths. The edge vision server has a built-in edge AI chip, which can analyze the video stream transmitted back by the robot in real time, run lightweight visual recognition algorithms, realize real-time parsing of instrument readings and preliminary screening of abnormal images, and only upload the result data, which greatly saves bandwidth.

[0041] The role of the robot base station is as follows: The robot continuously transmits raw video streams, thermal image streams, audio streams, and structured data (such as GPS coordinates and device IDs) collected by multi-source sensor components to the robot base station via a high-speed wireless link. Within the edge vision server, the data stream employs efficient compression encoding and encapsulation protocols to balance quality and bandwidth. Specific processing includes: Video stream analytics: Runs lightweight AI models in real time to perform instrument OCR readings and detect emergency anomalies (such as open flames, large leaks); Data simplification and annotation: The processing results (such as "P-203 pump front bearing temperature: 72℃" and "No. 3 sight glass liquid level is normal") are bound with the corresponding key frame images or thermal image thumbnails to generate a lightweight "structured inspection report". By using robotic base stations, the amount of data that needs to be uploaded to the cloud is greatly reduced, realizing the transformation from "transmitting raw data" to "transmitting results," which is the key to cloud-edge collaboration.

[0042] The remote centralized control and interaction layer serves as the system's command and visualization hub, enabling comprehensive monitoring and the highest level of human-machine interaction. This layer is a cloud-based intelligent analysis layer, physically integrated into the AI ​​models and optimization algorithms of the cloud-based intelligent analysis platform, embodying its core algorithm capabilities. It does not correspond to independent hardware but represents a crucial data processing flow: receiving data from lower layers, performing analysis, and then directing instructions to optimize control commands and perform inspection-review tasks.

[0043] The cloud-based intelligent analysis platform adopts a Kubernetes-based microservice architecture and is deployed on an enterprise private cloud or industrial internet platform. Its core services include: Data platform service: responsible for receiving, storing and managing historical and real-time data from the lower layer, using a hybrid storage scheme of "time series database + relational database"; AI model service container group: Encapsulates and runs equipment predictive maintenance models, process optimization algorithm models and machine vision analysis models independently. The models are called through RESTful API or gRPC interface and have elastic scaling capabilities to cope with computing load fluctuations. Rules engine and decision service: Built-in business processes and security rules are responsible for coordinating the output of different AI services and generating final control and task instructions.

[0044] The 3D digital twin monitoring platform is deployed on the central control center's large screen and engineer workstation via WebGL or a standalone client. Its core services include: Model and data driven: A 1:1 high-fidelity 3D model of the evaporation station is built on the platform and connected to a real-time database via the WebSocket protocol to drive the model status (such as valve opening and closing animation, pump rotation speed) to be synchronized with the site; Visualization mapping: The health status (HI) of the equipment is mapped to model colors (green / yellow / red), key process data is displayed as dynamic labels next to the equipment, and the real-time position of the robot and the video stream are integrated into the 3D scene to provide an immersive inspection experience; Interactive Interface: Provides a "one-click intervention" virtual control. When the operator clicks on a device in the twin model, detailed real-time data, historical curves, and manual operation panel of the device will pop up. Example 2

[0045] like Figure 3 As shown, the unattended intelligent monitoring method based on the control system described in Embodiment 1 specifically includes the following steps: Step S1: High-precision time synchronization data is collected from three core data sources using three types of interface protocols. PLC control system process data acquisition: Through industrial Ethernet (such as PROFINET), the operating data of various physical equipment in the evaporation station are collected synchronously at a period of no less than 100ms, such as the temperature and pressure of each effect of the evaporator, the flow rate of materials and steam, the status and opening degree of pumps and valves, and other continuous process variables. Integrated temperature and vibration sensor data acquisition: Through an industrial wireless network (such as WirelessHART), the vibration spectrum (acceleration, velocity) and temperature data of the bearing housing of key rotating equipment (such as feed pumps and forced circulation pumps) are collected synchronously in a cycle of 1-5 seconds. Data acquisition by the rail-mounted inspection robot: It receives structured data packets uploaded by the robot through a high-speed Wi-Fi network triggered by robot events, including but not limited to field instrument readings based on machine vision recognition, temperature anomalies marked on infrared thermal images, audio spectrum characteristics, and equipment appearance status labels.

[0046] Synchronization mechanism: When all data is accessed at the edge layer, the edge computing gateway embeds a unified high-precision timestamp based on IEEE 1588 (PTP) to establish a precise time alignment foundation for subsequent multi-source data fusion; Step S2: Edge-side preprocessing and fusion Within edge computing gateways and robot base stations deployed on-site, lightweight processing and initial fusion of data from the edge, close to the data source, are performed to improve data quality and reduce cloud load. Specifically, this includes: Data preprocessing: Filtering (noise removal), rationality verification (exceeding limits) and conversion of engineering units are performed on the raw PLC data; time-domain noise reduction and fast Fourier transform (FFT) are performed on the vibration waveform data; and preliminary calculation of spectral characteristics is performed. Edge intelligence and feature extraction: Within the robot base station, a lightweight YOLO or MobileNet model is run to analyze the acquired images in real time and directly output instrument readings and abnormal labels such as "running, leaking, and dripping", instead of uploading the original video stream; Feature-level fusion and uploading: The processed process time series, equipment vibration characteristics, and structured visual recognition results are packaged according to a unified time frame to generate a standard edge fusion data message. This message is uploaded to the cloud data platform via the MQTT protocol, either in seconds or through event triggering, effectively reducing uplink data traffic by more than 70%.

[0047] Step S3: Perform in-depth mining and collaborative analysis on the fused data. In the cloud-based microservice cluster, three core analytics engines run in parallel. These three core analytics engines specifically include: Equipment health diagnosis model: Receives vibration spectrum and temperature trend data, analyzes spectrum characteristics through a pre-trained deep convolutional neural network (CNN) model, automatically identifies fault modes such as bearing wear and blade imbalance, and calculates a quantitative real-time equipment health index; Process optimization model: Based on real-time process data stream, combined with the equipment performance degradation coefficient provided by the equipment health index, and using the model predictive control (MPC) algorithm, a rolling optimization problem is dynamically constructed with the goal of "lowest steam consumption" or "most stable output" and with equipment capacity and process safety as constraints. Visual depth analysis model: Performs secondary depth analysis on abnormal images uploaded by the robot. For example, it can accurately calculate the area and temperature gradient of overheated areas in infrared thermal images, or quantify the thickness of scale through image segmentation technology, providing more accurate on-site status input for process optimization.

[0048] Step S4: The cloud-based decision engine automatically makes judgments based on the analysis results of the three analysis engines, triggering different control branches, specifically including: Set comprehensive evaluation conditions, including but not limited to: a) whether the equipment health index is below a preset threshold; b) whether key process parameters (such as soft measurement values ​​of discharge concentration) deviate from the optimal range; c) whether visual analysis reveals any anomalies requiring intervention (such as severe leakage). Analyze the comprehensive evaluation conditions using the following logic: If any of the above conditions is "yes", the process will proceed to the abnormal collaborative handling branch corresponding to step S5. Otherwise, it enters the normal autonomous operation branch. In the normal autonomous operation mode, the system only performs data monitoring, recording and trend visualization, maintains the existing control parameters, and the PLC and robot work according to the established routine mode.

[0049] Step S5: The abnormal collaborative handling branch performs parallel control, which includes, but is not limited to: Generate process optimization parameters and send instructions to the PLC to adjust equipment operating parameters, such as valve and pump frequencies; Generate targeted tasks and instructions and issue them to the robot for review and special inspection.

[0050] Specifically, the generation method of the process optimization parameters includes: when optimization is required, the process optimization model solves the MPC optimization problem online based on the latest global state (including equipment health degradation information), generating the optimal control sequence for the next time domain (e.g., the next 15 minutes). This sequence mainly includes: the adjustment amount of the opening of each effect heating steam valve, the set value of the feed pump frequency, and the vacuum system adjustment command. Before the control command is issued, it must be verified by the Safety Constraint Controller (SCC). The SCC has built-in process safety interlock logic and equipment operation boundaries to ensure that any optimization command is within the safe operating area. Then, through the secure OPC UA communication protocol, the verified command set is sent to the field PLC control system. The advanced process control module in the PLC smoothly transitions the relevant control loops to the new optimization setpoint in a "bumper-free switching" manner, achieving precise and stable process regulation.

[0051] The generation of targeted tasks includes: the decision engine dynamically generates high-priority tasks based on the anomaly type. For example, when vibration analysis indicates an anomaly in a pump, the task is generated: "Navigate to the P-205A pump, perform infrared thermometry (front and side) and high-definition video recording (for 30 seconds)." The task instruction is then issued via the robot's base station, and the robot scheduling system immediately performs dynamic path planning, driving the robot to the target point. Upon arrival, the robot automatically invokes the corresponding sensors to complete the detection and performs preliminary analysis at the edge (such as identifying the highest temperature), feeding back a structured task execution report (including key data, image thumbnails, and preliminary conclusions) to the cloud in real time.

[0052] Step S6: Feedback the execution results to the analysis and decision-making center, specifically including: The system continuously monitors the actual response curves of key process parameters after the PLC executes the optimization instructions, and compares them with the curves predicted by the MPC model to generate control effect evaluation indicators (such as setpoint tracking error and energy saving effect). On-site verification and feedback: The cloud receives the verification report transmitted back by the robot and compares it with the initial diagnostic information that triggered the anomaly. For example, it confirms whether there is a corresponding high-temperature point in the pump body with abnormal vibration, thus completing the closed loop from "remote diagnosis" to "on-site verification". Data closed loop and learning: The complete "data-analysis-decision-execution-result" sequence is stored as a labeled sample in a historical case library. This case library is used to periodically perform incremental training and iterative optimization of the AI ​​models in the three major analysis engines, enabling the system's diagnostic accuracy and control optimization capabilities to continuously improve themselves over time, ultimately achieving complete autonomous adaptation and learning evolution. Example 3

[0053] like Figure 4 As shown, based on Embodiments 1 and 2, the unattended intelligent monitoring method preprocesses the raw collected data and performs equipment health prediction during the data acquisition process of the integrated temperature and vibration sensor.

[0054] This step converts the raw analog signals acquired by the integrated temperature and vibration sensor into time-series digital signals using a 24-bit high-precision ADC. This initial conversion from physical signals to digital information forms the basis for subsequent advanced analysis. The raw analog signals include vibration and temperature signals. The vibration signals contain raw acceleration waveform data (time-domain signals), typically acquired for one second each time, encompassing several complete equipment rotation cycles to ensure the capture of speed-related fault characteristics. The temperature signals are linearized resistance values ​​converted into digital temperature values ​​(unit: °C). These raw signal data packets, along with equipment codes, acquisition timestamps, sensor location, and other information, are packaged together and uploaded via a wireless network. This provides the cloud-based diagnostic model with unprocessed, "first-hand" data, preserving fault information to the maximum extent possible.

[0055] The specific processing steps include: (1) Feature extraction - time domain / frequency domain feature values Its aim is to compress massive amounts of raw waveform data into low-dimensional feature vectors containing fault information. This includes automated multi-dimensional feature extraction of raw vibration signals on the cloud or high-performance edge servers. Time-domain feature extraction: Calculate the statistical characteristics of the waveform, including: RMS (Reflecting the total energy of vibration), peak value (Reflecting instantaneous impact), kurtosis (Sensitive to impact faults), and waveform parameters (Shape Factor). These features are sensitive to faults such as early pitting of bearings and broken gear teeth. Frequency domain feature extraction: After applying the Hanning window function to the original waveform, a Fast Fourier Transform (FFT) is performed to convert the time-domain signal into a spectrum, from which the following features are accurately extracted: Fault characteristic frequency amplitude: Based on the equipment model and rotational speed, calculate and track the amplitude corresponding to the bearing inner ring failure frequency (BPFI), outer ring failure frequency (BPFO), rolling element failure frequency (BSF), and gear meshing frequency. Increases in these frequency amplitudes are clear indicators of wear on the corresponding components. Band energy: Divide the frequency bands (e.g., 0-1kHz, 1k-5kHz) and calculate the energy value of each band to identify different types of fault modes; Temperature trend characteristics: Calculate the recent slope of the temperature, the temperature difference with similar equipment or historical data from the same period, etc.

[0056] (2) AI health diagnosis The core of intelligent diagnostics is to automatically identify fault modes and quantify health status using deep learning models. The method for intelligent diagnostics involves deploying a CNN-LSTM hybrid deep learning model, the structure of which includes: Convolutional Neural Network (CNN) branch: Taking the spectrogram as a two-dimensional image input, it automatically learns the spatial distribution pattern of faults in the frequency domain through convolutional layers, for example, identifying the "sideband" features at specific fault frequencies in the spectrum; Long Short-Term Memory (LSTM) branch: Receives time series composed of temporal features and temperature trends, and learns the evolution of device status over time, such as the slow upward trend of kurtosis value over a week; Feature fusion and fully connected layer: High-level features extracted by CNN and LSTM branches are fused, and the diagnostic results are finally output through a fully connected layer and a Softmax classifier.

[0057] The diagnostic results include the probability that the equipment falls under various fault categories (such as bearing inner ring damage, rotor imbalance, misalignment, etc.). The overall health of the equipment is comprehensively reflected through a Health Index (HI). The Health Index (HI) is output as a continuous value from 0 to 100, where 100 represents brand new and perfect condition, and 0 represents complete failure. This index is calculated based on the degree to which characteristics deviate from the normal baseline model.

[0058] (3) Visual assessment of health status Transforming diagnostic results into decision-making information usable in engineering specifically includes: Set up a health status dashboard: In the digital twin interface of the remote control center, the color of the device's 3D model changes dynamically according to the Health Index (HI) (e.g., green > 80, yellow 80-60, red < 60). At the same time, the historical changes of HI are displayed intuitively in the form of a dashboard or trend curve, making it easy for maintenance personnel to understand.

[0059] Fault type prediction alarm: When the model's confidence level for a certain fault mode exceeds a threshold (e.g., 85%), the system generates a clear predictive alarm, such as: "Warning: Suspected inner ring damage in the drive bearing of P-203A pump (confidence level 92%)". The alarm is accompanied by a key spectrum diagram as evidence for the judgment.

[0060] Remaining useful life (RUL) estimation: Based on the current health index of the equipment and its degradation rate (learned from historical trends through an LSTM network), combined with failure data of similar equipment, particle filtering or deep learning regression models are used to estimate the average time required for the equipment to develop from its current state to functional failure. The output format is: "Predicted remaining useful life: 45±10 days", providing a critical time window for planned maintenance.

[0061] (4) Input the results into the process optimization model to achieve adaptive control. The quantitative results of equipment health diagnosis, especially the Health Index (HI) and specific failure mode information, are used as key input variables and fed into the process optimization model in real time. Specifically, the Health Index (HI) and specific failure mode information serve as efficiency coefficients to dynamically constrain the equipment health status. For example, when the Health Index of a circulating pump drops from 95 to 80, the optimization model will correspondingly lower its hydraulic efficiency coefficient in its internal calculations. Based on this correction, when solving for the optimal operating point, the model may automatically generate an optimization instruction to "appropriately reduce the load of this pump and increase the load of the parallel standby pump." Thus, even when the equipment is in a sub-healthy state, it can both ensure production tasks, prevent accelerated equipment deterioration, and maintain optimal overall system energy efficiency.

[0062] This transforms the entire control strategy from the traditional "ignore equipment status" or "shutdown after failure" mode to an intelligent and resilient control mode that "proactively adapts to changes in equipment status," truly achieving closed-loop linkage between equipment management and process operation. Example 4

[0063] Based on Embodiments 1, 2, and 3, the specific implementation of the multi-source data fusion described in this application is as follows: (1) Establish a unified time standard Synchronization Mechanism: When all data is accessed at the edge layer, a unified high-precision timestamp based on IEEE 1588 (PTP) is embedded by the edge computing gateway, establishing a precise time alignment foundation for subsequent multi-source data fusion. Specifically, a PTP Grand Master clock (compliant with the IEEE 1588 standard) is deployed at the edge node, connecting the PLC control system, the integrated temperature and vibration sensor assembly, and the mobile sensing terminal of the rail-mounted inspection robot via industrial Ethernet. Clock synchronization is achieved through Precision Time Protocol (PTP) or Network Time Protocol (NTP) at the PLC control system's process data acquisition terminal, the integrated temperature and vibration sensor data acquisition terminal, and the rail-mounted inspection robot data acquisition terminal, ensuring that each data acquisition terminal has a unified time reference; the synchronization accuracy of the high-precision clock source is better than or equal to 1 millisecond.

[0064] (2) Perform step S1 in Example 2 to collect multi-source data synchronously, including the PLC control system to collect raw data of process time series, the temperature and vibration integrated sensor to collect raw data of vibration signal, and the rail inspection robot to collect raw data of visual image.

[0065] (3) Perform data preprocessing and feature extraction at edge nodes, specifically: S3-1: Preprocessing of raw process time series data includes: outlier detection and filtering of raw sampled data, removing data points exceeding preset tolerance ranges; resampling of the filtered data according to the duration of the target time frame, aligning the data to the sampling points of the target time frame using the mean method or interpolation method; and normalizing the resampled data to generate standardized process time series data blocks. For example, when receiving raw sampled data of temperature, pressure, and rotational speed through edge nodes, outlier detection is first performed, marking and removing data points exceeding preset tolerance ranges (temperature exceeding [-10℃, 150℃], pressure exceeding [0MPa, 5MPa], rotational speed exceeding [0rpm, 3000rpm]). Then, resampling is performed according to the target time frame duration of 2 seconds. The mean value of the data within each 2-second window is calculated using the mean method as the representative value of that window, and Z-score normalization is performed on the data of each channel.

[0066] S3-2: Feature extraction of the raw vibration signal data, including: performing a short-time Fourier transform or wavelet packet decomposition on the raw vibration signal data at the edge nodes to convert the time-domain vibration signal into a time-frequency domain representation; extracting time-domain and frequency-domain features from the transform results, wherein the time-domain features include one or more of peak value, mean, variance, kurtosis, and crest factor, and the frequency-domain features include one or more of dominant frequency, dominant frequency amplitude, spectral energy, and energy distribution of each frequency band; aggregating the time-domain and frequency-domain features according to a window of the target time frame to generate a vibration feature data block. For example, the edge node receives the raw vibration waveform data from the temperature and vibration sensor, performs a short-time Fourier transform (STFT) on the signal for each axis, uses a Hanning window as the window function, and adaptively determines the window length to 200 sampling points based on the 2kHz sampling frequency and the required frequency resolution (e.g., 10Hz). Extracting time-domain features (peak value, RMS, kurtosis, crest factor) and frequency-domain features (dominant frequency, dominant frequency amplitude, 1kHz frequency band energy, 2kHz frequency band energy) from the transform results. The above feature values ​​are aggregated in a 2-second time window to generate a vibration feature vector for each window.

[0067] S3-3: Perform structured recognition on raw visual image data, including: executing a target detection algorithm on the raw visual image data at edge nodes to identify target objects in the image; extracting structured metadata of the recognition results, which includes one or more of the target object's category label, location coordinates, and confidence score; binding the structured metadata with the timestamp of the image acquisition time to generate a structured visual recognition result data block. For example, if an edge node receives image data acquired by a visible light camera, it calls a lightweight target detection model (such as YOLOv5s) deployed on the edge node to perform real-time inference on the image to identify whether there are surface defects such as cracks and scratches in the image. For each detected defect target, its category label (such as "crack"), bounding box coordinates (x, y, width, height) and confidence score are extracted and bound with the hardware trigger timestamp of the image to generate a structured visual recognition result.

[0068] (4) Data alignment and fusion according to a unified time frame Set a fixed frame duration T, use the frame duration T as the window length, and use the unified time base as a reference to define multiple consecutive time frames. Each time frame is uniquely determined by its start time and end time. For the process time series data block generated in step S3-1, it is assigned to the corresponding time frame according to the timestamp of each data point; For the vibration feature data block generated in step S3-2, it is assigned to the corresponding time frame according to the time window corresponding to each feature value. For the structured visual recognition result data block generated in step S3-3, the nearest neighbor matching method is used to classify it into the time frame to which the timestamp belongs, based on the timestamp bound to each recognition result; if multiple visual recognition results are contained in the same time frame, all recognition results are classified into the time frame in the form of a list. Therefore, each time frame contains time-aligned process time series data, vibration feature data, and structured visual recognition result data, forming a multi-source fusion data unit for that time frame.

[0069] For example, setting the frame duration T=2 seconds, and using the PTP unified time base as a reference, define continuous time frames: Frame #1 is [00:00:00.000, 00:00:02.000), Frame #2 is [00:00:02.000, 00:00:04.000), and so on. Each data point in the process time series data block generated in step (3) is assigned to its corresponding time frame according to its timestamp; the feature vectors of each 2-second window in the vibration feature data block are assigned to their corresponding time frames; the visual recognition results are assigned to the time frame to which the timestamp belongs based on the image acquisition timestamp they are bound to—if the same time frame contains recognition results for multiple images, all recognition results are assigned to that frame in a list format. Thus, each time frame contains time-aligned process data, vibration features, and visual recognition results within that 2-second time period.

[0070] (5) Generate standardized edge data fusion messages According to a preset message format, the multi-source fusion data unit of each time frame is encapsulated into a standardized edge data fusion message, which includes at least the following fields: Message header fields include: protocol version number, unique message identifier, source edge node identifier, target device identifier, and message generation timestamp; Time frame metadata fields include frame start time, frame end time, frame duration, frame sequence number, clock source type, and synchronization precision. Process timing data block fields include: number of sampling points, sampling rate, name and unit of each channel, numerical sequence of each channel, and statistical characteristic value; Vibration characteristic data block fields: include number of sampling points, sampling rate, axial identification, time domain characteristic value, frequency domain characteristic value, and spectrum data; Visual recognition data block fields include: number of visual frames, list of target objects (each target object includes category label, location coordinates, confidence score, and acquisition timestamp), and image metadata; The message tail field contains the data integrity check value, digital signature, message length, and compression identifier.

[0071] According to the preset message format, the multi-source fused data unit of each time frame is encapsulated into a standardized JSON message. The specific structure of the message is as follows: Message Header: Includes protocol version number "1.0", unique message identifier (e.g., "EDGE_MSG_20260708_143022_001"), source edge node identifier (e.g., "EDGE-GW-03"), target device identifier (e.g., "MACHINE-07"), and message generation timestamp (ISO 8601 format).

[0072] Frame Metadata: Includes frame start time, frame end time, frame duration (2000ms), frame sequence number (e.g., 2847), clock source type ("PTP"), and synchronization precision (100ns).

[0073] Process Data Block: Contains the number of sampling points (200), sampling rate (100Hz), information for each channel (channel name "temperature" / "pressure" / "speed", unit "℃" / "MPa" / "rpm", corresponding numerical sequence), and statistical characteristic values ​​of each channel (mean, maximum, minimum).

[0074] Vibration Feature Block: Includes the number of sampling points (4000), sampling rate (2kHz), axial identifiers ("X", "Y", "Z"), time-domain feature values ​​(peak value, RMS value, kurtosis, crest factor for each axis), frequency-domain feature values ​​(dominant frequency, dominant frequency amplitude, total spectral energy, energy of each frequency band), and sparsified spectral data (only the first 20 frequency points with amplitudes exceeding the preset energy threshold are retained).

[0075] Vision Data Block: Contains the number of visual frames within the time frame, the category label for each target object, bounding box coordinates, confidence score, and acquisition timestamp, as well as image metadata (resolution, frame rate, exposure time, etc.).

[0076] Footer: Includes the data integrity verification method ("SHA256"), checksum, digital signature, total message length (in bytes), and compression method ("gzip").

[0077] (6) Message transmission and verification The edge data fusion message generated in step (5) is sent to the remote control and interaction layer via the Message Queuing Telemetry Transport (MQTT) protocol or the OPC UA publish / subscribe model. Before sending, the integrity check value of the message is calculated and a digital signature is attached. The receiving end confirms the integrity and source credibility of the message by verifying the integrity check value and the digital signature.

[0078] Specifically, the encapsulated JSON message is published to a topic on the cloud data platform (e.g., " / edge / fusion / data") via the MQTT protocol at QoS=1. Before sending, the SHA-256 hash value of the message is calculated as an integrity check value, and the hash value is signed using the edge node's private key to generate a digital signature. After receiving the message, the receiving end verifies the digital signature using the edge node's public key and recalculates the hash value, comparing it with the check value carried in the message to confirm that the message has not been tampered with during transmission.

[0079] This embodiment achieves precise time alignment of multi-source heterogeneous data. By deploying a high-precision clock source and employing the PTP protocol to synchronize the clocks of each acquisition terminal, coupled with a hardware-triggered visual acquisition timestamp recording mechanism and a data merging strategy based on a unified time frame, it effectively solves the time misalignment problem caused by different sampling frequencies and independent clocks in process timing data, vibration data, and visual data. The time alignment accuracy can reach millisecond or even microsecond levels. Furthermore, it reduces the data transmission bandwidth requirements from the edge to the cloud. It completes frequency domain feature extraction of vibration signals and structured recognition of visual images at the edge nodes, transmitting only feature data and structured metadata instead of the original waveforms and images. Combined with compression strategies such as dead-zone filtering and spectral sparsity reduction, it significantly reduces data transmission volume, alleviating network bandwidth and cloud storage pressure. Simultaneously, it completes all computationally intensive tasks such as data preprocessing, feature extraction, time alignment, and message encapsulation at the edge nodes, keeping the end-to-end processing latency within 200 milliseconds, fully meeting the needs of industrial real-time monitoring scenarios.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments without departing from the technical solution of the present invention and based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. An unattended intelligent control system based on multi-source data fusion, applied to the monitoring of physical equipment in an alumina production evaporation station, characterized in that, The architecture consists of three layers from bottom to top: a bottom-level data acquisition layer, a middle-level data fusion and edge processing layer, and a top-level remote centralized control and interaction layer. The bottom data acquisition layer includes a PLC control system, an integrated temperature and vibration sensor assembly, and a rail-mounted inspection robot mobile sensing terminal. The PLC control system is used for process control of the physical equipment in the evaporation station; the integrated temperature and vibration sensor assembly is used for real-time monitoring of key rotating equipment in the evaporation station; and the rail-mounted inspection robot mobile sensing terminal is used for inspection of the entire monitoring area. The middle layer data fusion and edge processing layer includes an edge computing gateway and a robot base station. The edge computing gateway is used to receive real-time data from the PLC and temperature and vibration sensor data, and to aggregate and preprocess the data. The robot base station is used to receive task instructions from the cloud and schedule the robot to execute them, and can also perform local real-time analysis of the high-definition video streams collected by the robot. Top-level remote centralized control and interaction layer: This is the cloud-based intelligent analysis layer, including a cloud-based intelligent analysis platform and a 3D digital twin monitoring platform. The cloud-based intelligent analysis platform is used to receive, store, and manage historical and real-time data from the lower layers, deploy and run artificial intelligence models to perform data analysis and prediction tasks, and generate control decision commands based on preset rules and model output results. The 3D digital twin monitoring platform is used to construct a high-fidelity 3D model of the evaporation station, drive the model status to synchronize with the site, map data changes to the dynamic display of 3D graphic elements, and provide user interaction and data transmission between the user and the 3D scene.

2. The system according to claim 1, characterized in that, The integrated temperature and vibration sensor assembly consists of several integrated temperature and vibration sensors installed on various rotating equipment in the evaporation station. The integrated temperature and vibration sensor adopts an industrial wireless sensor, which integrates a high-frequency vibration accelerometer, an embedded signal processing unit, a platinum resistance temperature sensor, and a low-power wireless transmission module.

3. The system according to claim 1, characterized in that, The mobile sensing terminal of the rail-mounted inspection robot includes the rail-mounted robot body, the track laid above the monitoring area, and a multi-source sensor assembly mounted on the rail-mounted robot body.

4. The system according to claim 3, characterized in that, The track is a circular rigid track laid above the alumina production evaporation station. The track laying path covers all key equipment, and RFID or QR code tags are installed on both sides of the track to provide absolute position reference for the robot.

5. The system according to claim 3, characterized in that, The rail-mounted robot body is equipped with a central processing unit and a communication module, and has an intelligent mobile chassis at the bottom. The intelligent mobile chassis integrates a drive system and a power management module that are telecommunicationly connected to the central processing unit.

6. The system according to claim 3, characterized in that, The multi-source sensor assembly includes: a visible light camera for acquiring images of the device's appearance and meter readings; an infrared thermal imager for acquiring the device's temperature distribution; an ultrasonic pickup and lidar for obstacle avoidance and positioning; and also includes an acoustic and environmental perception unit, an auxiliary lighting unit, and a positioning unit.

7. The system according to claim 1, characterized in that, The mid-layer data fusion and edge processing layer is responsible for local data processing and agile response: The edge computing gateway is an embedded computer deployed in each process area. It is connected to the PLC control system via industrial Ethernet and to the temperature and vibration sensor components via LoRa wireless network. The robot base station functions as a charging dock, communication hub, and edge server. It is an integrated workstation equipped with an edge vision server and communication module, and maintains continuous communication with the rail-mounted inspection robot via high-speed Wi-Fi 6.

8. The system according to claim 1, characterized in that, The cloud-based intelligent analysis platform is deployed on an enterprise private cloud or industrial internet platform using a Kubernetes-based microservice architecture. It includes a data platform service, an AI model service container group, and a rule engine and decision service. The data platform service, AI model service container group, and rule engine and decision service communicate with each other through the microservice architecture and share a unified data bus. The three-dimensional digital twin monitoring platform is deployed on the central control center's large screen and engineer's station via WebGL or a standalone client, and includes model and data-driven, visualization mapping, and interactive interfaces; The industrial internet platform sends the decision results to the interactive interface module of the 3D digital twin monitoring platform through the rule engine and decision service, so as to display the decision effect in the 3D scene.

9. An intelligent monitoring method based on the system described in any one of claims 1-8, characterized in that, Includes the following steps: Step S1: Based on the synchronization mechanism, high-precision time synchronization acquisition is performed on three core data sources through three types of interface protocols, including process data acquisition from PLC control system, data acquisition from integrated temperature and vibration sensor, and data acquisition from rail-mounted inspection robot. Step S2: Edge-side preprocessing and fusion Within edge computing gateways and robot base stations deployed on-site, lightweight processing and initial fusion of data from the edge, close to the data source, are performed to improve data quality and reduce cloud load. Specifically, this includes: Data preprocessing: Filtering, rationality verification and engineering unit conversion of raw PLC data; performing time-domain noise reduction and fast Fourier transform on vibration waveform data; and preliminarily calculating spectral characteristics. Edge intelligence and feature extraction: Within the robot base station, a lightweight YOLO or MobileNet model is run to analyze the acquired images in real time and directly output anomaly labels instead of uploading the original video stream; Feature-level fusion and uploading: The processed process time series, equipment vibration features, and structured visual recognition results are packaged according to a unified time frame to generate a standard edge fusion data packet and uploaded to the cloud data platform; Step S3: Perform in-depth mining and collaborative analysis on the fused data. In the cloud-based microservice cluster, three core analytics engines run in parallel to perform in-depth mining and collaborative analysis on the merged data. These three core analytics engines specifically include: Equipment health diagnosis model: Receives vibration spectrum and temperature trend data, analyzes spectrum characteristics through a pre-trained deep convolutional neural network model, automatically identifies fault modes, and calculates a quantitative real-time equipment health index; Process optimization model: Based on real-time process data stream, combined with the equipment performance degradation coefficient provided by the equipment health index, the model predictive control algorithm is used to dynamically constrain the equipment health status using the health index and specific failure mode information as efficiency coefficients. Visual depth analysis model: Perform secondary depth analysis on abnormal images uploaded by the robot or quantify the scale thickness through image segmentation technology to provide more accurate on-site status input for process optimization; Step S4: The cloud-based decision engine automatically makes judgments based on the analysis results of the three analysis engines, triggering different control branches, specifically including: Set comprehensive evaluation conditions, including but not limited to: a) whether the equipment health index is below a preset threshold; b) whether key process parameters deviate from the optimal range; c) whether visual analysis detects any anomalies requiring intervention. Analyze the comprehensive evaluation conditions using the following logic: If any of the above conditions is "yes", the process will proceed to the abnormal collaborative handling branch corresponding to step S5. Otherwise, it will enter the normal autonomous operation branch and maintain the normal autonomous operation mode; Step S5: The abnormal collaborative handling branch performs parallel control, which includes, but is not limited to: Generate process optimization parameters and send instructions to the PLC to adjust equipment operating parameters; Generate targeted tasks and instructions and issue them to the robot for review and special inspection; Step S6: Feed the execution results back to the analysis and decision center.

10. The intelligent monitoring method according to claim 9, characterized in that, During the data acquisition process of the integrated temperature and vibration sensor, the raw acquired data is preprocessed to predict equipment health. The specific processing steps include: (1) Feature extraction - time domain / frequency domain feature values Massive amounts of raw waveform data are compressed into low-dimensional feature vectors containing fault information, including automated multi-dimensional feature extraction of raw vibration signals on cloud or high-performance edge servers. (2) AI health diagnosis The CNN-LSTM deep learning model is used to automatically identify fault modes and quantify health status, specifically including: Convolutional Neural Network (CNN) branch: It takes the spectrogram as input as a two-dimensional image and automatically learns the spatial distribution pattern of faults in the frequency domain through convolutional layers; Long Short-Term Memory (LSTM) branch: Receives time series composed of temporal features and temperature trends, and learns the evolution of device state over time; Feature fusion and fully connected layer: The high-level features extracted by the CNN and LSTM branches are fused, and the diagnostic results are finally output through the fully connected layer and the Softmax classifier; (3) Visual assessment of health status Transforming diagnostic results into decision-making information usable by engineering includes: setting up health status dashboards, fault type prediction alarms, and remaining life (RUL) estimation. (4) Input the results into the process optimization model to achieve adaptive control.

Images