Task-based regional intelligent inspection system

Abstract

The invention discloses a task-based regional intelligent inspection system, and the system comprises a multi-source data collection module which is used for collecting various data of a task region; the task generation module is used for generating a dynamic task list according to the data of the multi-source data acquisition module and the feedback optimization module; the intelligent task assignment module is used for generating an inspection task according to the data in the task list and in combination with the man-machine efficiency data; the path planning module is used for generating a differentiated inspection path based on a regional three-dimensional map and the inspection task; the intelligent analysis module is used for establishing a new LSTM model and predicting the current and future states of the equipment; and the feedback optimization module is used for automatically adjusting parameters and task generation strategies of the LSTM new model. According to the method, automatic path planning and feedback optimization are carried out through multi-data fusion, leap-type upgrading of chemical inspection from civil defense to technical defense is realized, and remarkable benefits are generated in the aspects of efficiency, safety, cost and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital factory management, and particularly relates to a task-based regional intelligent inspection system. BACKGROUND

[0002] Some areas, such as chemical plant production, have characteristics such as high temperature and high pressure, flammable and explosive, toxic and harmful, etc. The direct economic loss caused by chemical accidents due to equipment failure or human error is over ten billion US dollars per year. The traditional manual inspection mode has significant defects: efficiency bottleneck: relying on paper records and manual judgment, a single comprehensive inspection takes 6-8 hours, and the missed detection rate is as high as 23% (China Anke Institute 2024 data). Safety risk: manual intervention in high-risk areas (such as reaction kettle area, storage tank area) leads to an annual accident rate of up to 41%. Data island: vibration monitoring, gas detection and other subsystems operate independently, lacking cross-dimensional data correlation analysis.

[0003] With the development of technology, intelligent inspection systems gradually replace traditional modes. The first generation system (2010-2015) uses a wired sensor network to realize basic parameter monitoring; the second generation system (2016-2020) introduces unmanned aerial vehicles and robots to expand the detection dimension; the third generation system (2021 to now) begins to integrate AI algorithms, but there are problems such as weak model generalization ability and poor dynamic adaptability. The current regional inspection, especially the chemical plant equipment inspection, has obvious limitations, and all have significant defects: ① manual inspection mode, the inspection personnel hold the detection instrument and patrol according to the fixed route, which has the following problems: the detection accuracy is greatly affected by the experience of personnel, the paper record data is difficult to digitalize and trace, and the manual approach detection in high-risk areas has safety hazards. ② Fixed monitoring system, vibration sensors, temperature transmitters, etc. are installed on key equipment, but there are the following problems: the installation position is limited (such as unable to monitor the stress concentration position of the pipeline), the sensor wiring cost is high (a single set of device needs to increase

[0004] 8-12 million yuan of transformation cost), the data acquisition frequency is fixed (usually 5-10 minutes / time), and transient abnormalities (such as pump body sudden surge) cannot be captured. Typical defects of existing intelligent inspection technology: rigid path planning: static topological map is used to plan the path, without considering dynamic obstacles such as maintenance occupation and personnel flow, resulting in the need for temporary adjustment of the inspection route. Multi-source data fragmentation: video analysis, vibration detection, process parameters are scattered in different systems, lacking a timestamp synchronization mechanism, affecting fault correlation analysis. Model update lag: the diagnosis model relies on historical data offline training, and cannot absorb new discovered fault modes (such as new type of seal failure characteristics) in real time. SUMMARY

[0005] The present application aims at the problems in the prior art and discloses a task-based regional intelligent inspection system, which realizes the leapfrog upgrade of chemical inspection from "people defense" to "technical defense" through multi-data fusion, automatic path planning and feedback optimization, and produces significant benefits in efficiency, safety, cost and other dimensions.

[0006] The present application is realized through the following technical solutions:

[0007] The present application first provides a task-based regional intelligent inspection system, which comprises

[0008] A multi-source data acquisition module is configured to acquire various data of a task region.

[0009] A task generation module is configured to generate a dynamic task list according to the data of the multi-source data acquisition module and the feedback optimization module.

[0010] An intelligent task assignment module is configured to generate an inspection task according to the data in the task list and in combination with human-machine efficiency data. A path planning module is configured to generate a differentiated inspection path based on a regional three-dimensional map and the inspection task.

[0011] An intelligent analysis module is configured to establish an LSTM new model to predict the current and future states of equipment, wherein the formula of an output gate in the LSTM new model is o t =σ(W o ·[h t-1 ,x t ]+b o ), wherein o t is the output of the output gate, sigma is an activation function, h t-1 is a hidden state, and then is transmitted to the next time step, W o is a weight matrix, x t is the input of the current time step, and b o is a bias term. A feedback optimization module is configured to automatically adjust the parameters of the LSTM new model and the task generation strategy according to the deviation value between the prediction result of the LSTM new model and the actual equipment state.

[0012] As a further solution, the task generation module comprises a data acquisition subunit, a data analysis subunit and a priority sorting subunit,

[0013] The data acquisition subunit is responsible for collecting equipment operation parameters, historical fault data and environmental monitoring data in real time and transmitting them to the data analysis subunit in real time.

[0014] The data analysis subunit processes and analyzes the data of the data acquisition subunit, extracts key features and trends and transmits them to the priority sorting subunit.

[0015] The priority sorting subunit prioritizes inspection tasks based on the analysis results of the data analysis subunit, using a preset algorithm to ensure that high-priority tasks are executed first.

[0016] As a further solution, the intelligent task assignment module includes:

[0017] Task attribute parsing unit: used to extract the task type, security risk level, and required detection equipment type from the dynamic task list;

[0018] Resource status monitoring unit: Real-time acquisition of personnel location, robot battery level, and sensor availability data through the database of the multi-source data acquisition module;

[0019] Human-machine performance evaluation unit: Constructs a task completion quality-efficiency matrix based on historical data;

[0020] Dynamic dispatch engine unit: Based on the data from the task attribute parsing unit, resource status monitoring unit, and human-machine efficiency evaluation unit, it uses the AHP-TOPSIS hybrid decision model to output the optimal inspection task.

[0021] As a further solution, the path planning module includes

[0022] The map processing unit determines the location of obstacles such as equipment, pipelines, and buildings on the 3D map, as well as road information, and then divides the 3D map into a grid.

[0023] The path optimization unit calculates the optimal path from the starting point to the ending point, using a grid as the unit, based on the current inspection task; the optimal path includes both the personnel path and the robot path.

[0024] The dynamic obstacle avoidance unit optimizes the obstacle path on the optimal path by combining real-time road condition information.

[0025] As a further solution, the multi-source data acquisition module includes a mobile inspection terminal and an intelligent robot cluster.

[0026] As a further solution, the intelligent analysis module includes a data preprocessing unit, a model training unit, a risk management unit, and an online identification unit.

[0027] The data preprocessing unit cleans and extracts features from the data retrieved from the multi-source data acquisition module and stores it for subsequent analysis.

[0028] The model training unit is used to build and train new LSTM models.

[0029] The risk management unit is used to determine the risk level after marking equipment failure, human error, and environmental risk factors;

[0030] The online identification unit is used to input real-time collected risk data into a pre-trained LSTM model to predict the current and future states of the device.

[0031] As a further approach, the method for constructing and training a new LSTM neural network model is as follows:

[0032] S1. Normalize the feature data required for training obtained from the feature database.

[0033] S2. Constructing the LSTM neural network structure

[0034] S3, Model Training

[0035] Collect a large amount of historical monitoring data and label it; select a loss function and optimizer to train the model;

[0036] S4. Online Monitoring and Forecasting

[0037] The real-time collected time-series monitoring data is input into the trained LSTM neural network model;

[0038] S5, State Prediction:

[0039] The LSTM neural network predicts the current and future state of a device based on real-time input data and previously learned patterns and rules.

[0040] S6. Results Analysis and Feedback:

[0041] Analyze the prediction results, and if a potential equipment malfunction or other abnormal situation is predicted, issue an alarm in a timely manner and take corresponding measures.

[0042] As a further solution, the feedback optimization module includes: a data collection and processing unit: collecting and processing equipment failures, abnormal event records, predicted equipment status, and failure probabilities found during actual inspections;

[0043] Deviation Calculation Unit: Compares the actual inspection results with the predicted results from the data collection and processing unit to calculate the deviation value;

[0044] Deviation Analysis Unit: Performs statistical analysis on the deviation values ​​of the deviation calculation unit, calculates statistical indicators such as average deviation and standard deviation, and evaluates the overall accuracy of the prediction results.

[0045] As a further embodiment, the system also includes a digital twin modeling module, which includes...

[0046] The data acquisition and preprocessing unit includes a sensor data receiver, a data format converter, and a data buffer. The sensor data receiver is responsible for collecting real-time operating data from various devices in the factory. The data format converter converts heterogeneous data output from different types of sensors into a standardized data format. The data buffer temporarily stores the acquired data.

[0047] The 3D modeling unit includes a modeling engine, a model library, and a texture mapping subunit. The modeling engine creates the 3D geometric model of the chemical plant; the model library stores prefabricated 3D models of various equipment, parts, and buildings; and the texture mapping subunit maps realistic texture maps onto the surface of the 3D model.

[0048] The model fusion and update unit includes a data fusion engine, a model update controller, and a semantic parser;

[0049] The visualization and interaction unit includes a graphics rendering engine, a user interface, and scene navigation tools.

[0050] As a further embodiment, the system also includes a temporary task processing module, which comprises an emergency task receiving unit, a temporary task planning unit, a task coordination and execution unit, and a result feedback and integration unit.

[0051] The features and beneficial effects of this invention are as follows:

[0052] (1) This invention achieves a leapfrog upgrade of chemical inspection from "human defense" to "technical defense" by integrating multiple data, performing automatic path planning and feedback optimization, and generating significant benefits in terms of efficiency, safety and cost.

[0053] (2) This invention achieves closed-loop management of risk warning and emergency response through the linkage of real-time environmental monitoring and digital twin virtual model.

[0054] (3) The unmanned inspection terminal set up in this invention replaces more than 60% of manual inspections in high-risk areas, reducing the labor cost of a single inspection by 50% while ensuring the safety of personnel. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of a task-based regional intelligent inspection system according to an embodiment of the present invention. Detailed Implementation

[0057] To facilitate understanding of the present invention, a more comprehensive description of the present invention will be given below, and embodiments of the present invention will be provided, but this does not limit the scope of the present invention.

[0058] A task-based regional intelligent inspection system, such as Figure 1 As shown, it includes: a multi-source data acquisition module, used to collect equipment and environmental data in the task area and generate a database;

[0059] Task generation module: Dynamically adjusts the task execution order and generates a dynamic task list based on data from the multi-source data acquisition module and the feedback optimization module;

[0060] Intelligent task assignment module: Generates inspection tasks based on the data in the task list and combined with human-machine efficiency data; Path planning module: Generates differentiated inspection paths based on the regional 3D map and the inspection tasks;

[0061] The intelligent analysis module builds a new LSTM model and inputs real-time data into the model to predict the current and future states of the device.

[0062] The feedback optimization module automatically adjusts the parameters of the new LSTM model and the task generation strategy based on the deviation between the prediction results of the new LSTM model and the actual device status.

[0063] This invention can automatically optimize inspection routes and dynamically adjust task priorities. For example, by integrating equipment maintenance records with real-time road conditions, the system can increase the inspection frequency of critical equipment (such as reactors and storage tanks) from once per hour to real-time monitoring, while reducing the inspection frequency of non-critical areas by 40%, thus achieving optimal resource allocation.

[0064] A multi-source data fusion diagnostic system was constructed. For example, by cross-modal correlation between vibration spectrum and thermal imaging data, early wear of a compressor bearing was successfully predicted, avoiding unplanned downtime losses.

[0065] The task generation module includes a priority evaluation unit, which generates a dynamically weighted task priority score based on real-time equipment operation data, historical failure probability matrix, and environmental risk map.

[0066] The priority evaluation unit includes a data acquisition subunit, a data analysis subunit, and a priority ranking subunit.

[0067] The data acquisition subunit is responsible for collecting equipment operating parameters, historical fault data, and environmental monitoring data in real time, and transmitting them to the data analysis subunit in real time.

[0068] The data analysis subunit processes and analyzes the data from the data acquisition subunit, extracts key features and trends, and passes them to the priority ranking subunit.

[0069] The priority sorting subunit prioritizes inspection tasks based on the analysis results of the data analysis subunit, using a preset algorithm to ensure that high-priority tasks are executed first.

[0070] In one or more embodiments, the data acquisition subunit comprises a temperature sensor for testing the reactor temperature and a gas sensor for testing the surrounding environment. The data is transmitted to the data analysis subunit, which analyzes historical fault data of the reactor to indicate that similar faults have occurred frequently. This information is then sent to the priority ranking subunit, which immediately prioritizes the reactor inspection task and re-plans the inspection sequence. The task generation module updates the inspection plan promptly based on the task list adjusted by the priority evaluation unit, ensuring that inspection personnel or robots prioritize detailed inspections of the reactor, promptly identifying potential safety hazards and taking appropriate measures. This process fully demonstrates the task generation module's ability to dynamically adjust the task execution sequence based on various data sources, effectively improving the timeliness and targeting of inspections and enhancing the safety of the chemical plant.

[0071] The intelligent task assignment module includes:

[0072] Task attribute parsing unit: used to extract features such as task type, security risk level, and required detection equipment type;

[0073] Resource status monitoring unit: used to acquire real-time data on personnel location, robot battery level, and sensor availability;

[0074] Human-machine performance evaluation unit: Constructs a task completion quality-efficiency matrix based on historical data;

[0075] Dynamic assignment engine: The AHP-TOPSIS hybrid decision model is used to output the optimal assignment scheme.

[0076] The intelligent task assignment module acts like an "intelligent dispatch commander," breaking through the traditional fixed human-machine division of labor model. By quantitatively analyzing the matching relationship between task characteristics and the capabilities of the executing entity, it achieves the following:

[0077] (1) Dynamic human-machine collaboration: Based on the urgency of the task (such as gas leaks requiring a second-level response), the task is automatically assigned to an aerial drone, routine inspections are handled by ground robots, and complex fault diagnosis tasks are assigned to professional personnel.

[0078] (2) Multidimensional constraint optimization: Simultaneously consider hard constraints such as personnel qualification certification (e.g., confined space operation permit), robot sensor configuration (e.g., explosion-proof vs. ordinary), and task geographical accessibility (e.g., high-altitude pipe gallery requires climbing robots);

[0079] (3) Flexible load balancing: When multiple devices in a certain area alarm at the same time, the optimal allocation combination is automatically calculated to avoid overload of a single device.

[0080] The path planning module includes...

[0081] The map processing unit determines the location of obstacles such as equipment, pipelines, and buildings on the 3D map, as well as road information, and then divides the 3D map into a grid.

[0082] The path optimization unit calculates the optimal path from the starting point to the ending point in grid units, based on the current inspection task.

[0083] The dynamic obstacle avoidance unit optimizes the obstacle avoidance path by combining real-time road condition information, selecting the path with the lowest cost to bypass obstacles and continue moving towards the target point.

[0084] The path planning module plans paths for both personnel and robots. Personnel paths include safety passages and emergency evacuation point markers, while robot paths include charging station docking strategies and dynamic obstacle avoidance rules.

[0085] ① The path planning module uses DDPG (Deep Deterministic Policy Gradient), an algorithm based on deep reinforcement learning. Deep reinforcement learning is a technique that combines reinforcement learning and deep neural networks, enabling the learning of optimal behavioral policies through interaction with the environment. By generating actions through a policy network and evaluating the value of actions through a value network, dynamic path planning based on a 3D mesh map is achieved in complex chemical environments.

[0086] The path planning method of the path planning module is as follows:

[0087] A1. The system loads a detailed 3D map of the chemical plant area, which includes the location information of various equipment, pipelines, buildings, and other obstacles, as well as the road topology. The map is preprocessed and divided into a grid to facilitate pathfinding by the algorithm.

[0088] A2. Based on the current inspection task, determine the starting point (current location) and ending point (next inspection target point) of the inspection robot.

[0089] A3. Using a grid as the unit, starting from the starting point, the optimal path from the starting point to the ending point is calculated using the algorithm's heuristic search strategy. During the search process, dynamic factors such as equipment maintenance records, personnel location data, and real-time traffic information are incorporated into the path planning model.

[0090] A4. During the inspection robot's movement along the planned path, real-time monitoring of dynamic information such as equipment maintenance records, personnel flow, and road conditions is conducted. Upon discovering new obstacles (such as temporary maintenance equipment or personnel activity areas) or changes in road conditions (such as a section of road becoming impassable due to material leakage), the map is immediately updated, and the route is replanned. Traditional AI algorithms typically use Manhattan distance or Euclidean distance to estimate the cost from the current node to the target node. In the improved AI algorithm, the heuristic function is optimized by considering the actual layout of the chemical plant area and inspection requirements. In addition to distance factors, weighting coefficients for equipment importance, inspection priority, and road accessibility are also incorporated.

[0091] The new heuristic function expression is: h(n) = α × d(n,goal) + β × w(equipment) + γ × w(road), where d(n,goal) represents the Euclidean distance from the current node n to the target node; w(equipment) is the equipment weight coefficient, such as: critical equipment (reactor = 1.2, compressor = 1.1, storage tank = 1.0); w(road) is the road access difficulty coefficient, defined according to the 3D map grid attributes: maintenance passage = 0.8, pipe gallery area = 1.2, explosion-proof area = 1.5; α, β, and γ are weight adjustment factors, and the optimal combination is determined by orthogonal experimentation: α = 0.55 ± 0.05, β...

[0092] =0.35±0.03, γ=0.10±0.02, the sum of the three is 1.

[0093] Example: In an inspection scenario of an ethylene plant, α = 0.6, β = 0.3, and γ = 0.1 can be set. At this time, the path planning of the compressor area can shorten the journey by 20% compared with the traditional algorithm, while ensuring that the coverage of key equipment reaches 100%.

[0094] Furthermore, the path planning module also includes a dynamic obstacle avoidance mechanism. When an obstacle is detected in front of the inspection robot, the obstacle avoidance path is calculated in real time based on the size, shape, and position of the obstacle. At the same time, combined with real-time road condition information, the obstacle avoidance path is optimized to select the path with the least cost to bypass the obstacle and continue moving towards the target point.

[0095] The multi-source data acquisition module includes a mobile inspection terminal and an intelligent robot cluster. The mobile inspection terminal includes a head-mounted AR device, an integrated multispectral environmental perception sensor, and an augmented reality navigation interface, supporting voice command interaction and 3D anomaly localization annotation. The intelligent robot cluster includes: an explosion-proof ground robot equipped with a gas concentration sensor array and a robotic arm; a tethered drone integrating an infrared thermal imager and high-gain communication; and a track-mounted wall-climbing robot equipped with a magnetic flux leakage detection probe and a high-definition video device. The unmanned inspection terminal (explosion-proof robot, AR glasses) replaces more than 60% of manual inspections in high-risk areas, reducing labor costs per inspection by 50%.

[0096] The intelligent analysis module includes a data preprocessing unit, a model training unit, an online identification unit, and a risk management unit. The data preprocessing unit includes a data cleaning subunit, a feature extraction subunit, and a feature database.

[0097] The data cleaning subunit removes noisy data and missing values ​​from the data output from the multi-source data acquisition module to ensure the quality of the input data. It uses interpolation to fill in missing values ​​and employs smoothing filtering algorithms to reduce noise interference.

[0098] Feature extraction subunit: Extracts key features characterizing the equipment's operating status from the data cleaning subunit, such as frequency components in vibration signals and the rate of change of temperature signals. Based on the equipment's characteristics and fault modes, wavelet transform is used to extract time-frequency domain features from the signals.

[0099] Feature database: Used to store data after cleaning and feature extraction, for subsequent analysis modules to use.

[0100] The model training unit includes:

[0101] Model building subunit: This unit acquires the feature data required for training from the feature database to build the deep learning model architecture, using a hybrid model combining Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN). LSTM is used to process temporal data and capture temporal dependencies; CNN is used to extract spatial features from the data.

[0102] Training Subunit: Using historical monitoring data and corresponding fault labels, the model in the model building subunit is trained. The Adam optimization algorithm is used, with the cross-entropy loss function as the optimization objective, to iteratively update the model parameters until the model converges. After training, the updated model parameters are stored in the model repository.

[0103] The model construction subunit includes a risk assessment model for risk assessment and an equipment status prediction model for predicting equipment conditions. The risk assessment model involves acquiring equipment failure information, inspection records, and environmental monitoring data from a feature database, and using a fuzzy comprehensive evaluation model to determine equipment failure, human error, and environmental factors as input indicators. An evaluation indicator system is constructed, and expert experience is used to perform fuzzy quantification of the indicators. Fuzzy computation is then used to derive the fuzzy membership degree of the risk level, thereby determining the risk level.

[0104] After the model runs, the risk assessment results are output to the risk management unit of the intelligent analysis module and connected to the production scheduling system to provide decision support for production scheduling.

[0105] The risk level classification standard is divided into four levels. Based on a comprehensive assessment of the likelihood of the risk occurring and the severity of its consequences, the risk levels are divided into Level 1 (highest risk) to Level 5 (lowest risk).

[0106] The method for establishing the equipment condition prediction model is as follows:

[0107] S1. Normalize the feature data required for training obtained from the feature database, scaling the data of different dimensions to the same range (such as between 0 and 1) to facilitate processing by the LSTM neural network.

[0108] S2. Constructing the LSTM neural network structure: Input layer: The input layer receives the preprocessed time-series monitoring data from S1. Each time step inputs a fixed-length vector representing the device status at that time point.

[0109] LSTM layer: An LSTM layer consists of multiple LSTM units, each of which contains a cell state, a forget gate, an input gate, and an output gate.

[0110] Cell state: For a given time step, there are actually old cell states, candidate cell states, and new cell states. The new cell state is calculated by weighting the old cell states and candidate cell states, as detailed in the formula in S24. The candidate cell state is temporary information generated in the current time step using input data and hidden states, representing potential new content that needs to be updated.

[0111] (1) Forget Gate: Determines how much old information should be discarded, using the formula:

[0112] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0113] Among them, f t The output of the forget gate is represented by σ, which is the activation function (constraining the output value between 0 and 1), and W is the output of the forget gate. f It is the weight matrix, b f It is a bias term, h t-1 It is the hidden state of the previous time step, x t This is the input for the current time step.

[0114] (2) Input gate: Determines how much new information should be stored in the cell state, using the following formula:

[0115] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0116] Among them, i t The output of the input gate is a vector that determines how much new information will be written into the cell state C at the current time step. t Middle. W i It is the weight matrix, b i It is a bias term, h t-1 It is the hidden state of the previous time step, x t This is the input for the current time step.

[0117] (3) Candidate cell states: New candidate information is generated to update the cell state, using the following formula:

[0118] C t ′=tanh(W C ·[h t-1 ,x t ]+b C ), where C t ′ represents the candidate cell state, and the tanh activation function limits the output value to between -1 and 1. W C It is the weight matrix of the candidate cell states, h t-1 It is the hidden state of the previous time step, x t It is the input for the current time step, b C It is the bias of the candidate cell state.

[0119] (4) New cell state: The cell state is updated by combining the outputs of the forget gate and the input gate, using the following formula:

[0120] C t =f t ·C t-1 +i t ·C t

[0121] Among them, C t-1 It represents the cell state at the previous time step.

[0122] (5) Output gate: Determines how much information should be output. The formula is:

[0123] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0124] Among them, o t The output of the output gate is a vector that determines how much cell state information at the current time step will be output to the hidden state h. t This is then passed on to the next time step. o It is the weight matrix, b o It is a bias term.

[0125] (6) Hidden State: The hidden state at the current time step is calculated based on the output of the output gate and the new cell state. The hidden state is carried over to the next time step for subsequent calculations. The formula is: h t =o t ·tanh(C t ).

[0126] (7) Construct a deeper LSTM network by stacking multiple LSTM layers together to form a deep LSTM network. For example, you can first create a network structure containing multiple LSTM layers, where the output of each layer serves as the input of the next layer.

[0127] S3, Model Training

[0128] S31. Training data preparation:

[0129] Collect a large amount of historical monitoring data and label it (such as whether a fault occurred, the type of fault, etc.).

[0130] S32, Model Building:

[0131] Construct an LSTM neural network model and define hyperparameters such as network structure, number of layers, and number of units.

[0132] S33. Loss Function and Optimizer:

[0133] Choose an appropriate loss function (such as mean squared error, cross-entropy loss, etc.) and optimizer (such as Adam, RMSprop, etc.) to train the model.

[0134] S34, Training Process:

[0135] The training data is input into the LSTM neural network, the predicted value is calculated through forward propagation, then the loss between the predicted value and the true value is calculated, and finally the weight parameters of the network are updated through back propagation until the model converges.

[0136] Specifically, the methods include the following:

[0137] (1) Forward propagation phase:

[0138] The input is standardized time series data with dimensions [batch size, time step, feature dimension] = [64, 24, 15].

[0139] Calculation using gated cyclic units:

[0140] Forgotten Gate: f t =σ(W f ·[h t-1 ,x t ]+b f +λ·C t-1 2 )

[0141] The newly added quadratic term λ·C t-1 2 (λ=0.02) is used to enhance long-term memory retention.

[0142] (2) Loss Calculation:

[0143] Construct a hybrid loss function L = 0.7L MAE +0.3L SSIM , where: L MAE The mean absolute error L SSIM =1-SSIM(x,y) is the structural similarity loss;

[0144] (3) Backpropagation optimization:

[0145] The Adam optimizer, accelerated by Nesterov, is used with an initial learning rate η = 0.001.

[0146] Weight update formula:

[0147]

[0148] Where β1 = 0.9, β2 = 0.999, ε = 1e-7; gradient clipping is performed every 500 steps, with the threshold set to 10.0.

[0149] (4) Early termination mechanism:

[0150] Training is terminated when the validation set loss decreases by less than 0.5% for three consecutive epochs.

[0151] Ultimately, in the compressor vibration prediction task, the test set MAE was 0.023 and the F1 score was 0.914.

[0152] Example: Compressor vibration early warning

[0153] Forward propagation phase:

[0154] Input vibration data from the past 24 hours (1 time per second, totaling 86,400 data points).

[0155] The model found that whenever the vibration frequency is maintained at 5-7 Hz for 10 minutes, a sudden temperature rise will occur.

[0156] Test with new data: Given the data from the first 12 hours, let the model predict the vibration value for the next hour.

[0157] The first test score was 60 (MAE = 0.1), and it was found that the predicted curve shape differed significantly from the actual data.

[0158] Backpropagation optimization:

[0159] Adjust the learning focus, increase the attention weight of frequency features by 30%, and increase the number of daily practice sessions from 10 to 20.

[0160] Early shutdown mechanism:

[0161] Stop mechanical practice if you find that your performance improves by less than 0.5% in three consecutive attempts (to prevent overtraining).

[0162] Practical results:

[0163] The average error in predicting vibrations in the next hour is only 0.023 (equivalent to detecting anomalies 40 minutes in advance), and the fault identification accuracy rate is 91.4% (91 out of 100 warnings actually resulted in a fault).

[0164] S4. Online Monitoring and Forecasting

[0165] Real-time data input:

[0166] The real-time collected time-series monitoring data is input into the trained LSTM neural network model.

[0167] S5, State Prediction:

[0168] LSTM neural networks predict the current and future states of a device based on real-time input data and previously learned patterns and rules.

[0169] S6. Results Analysis and Feedback:

[0170] Analyze the prediction results, and if a potential equipment malfunction or other abnormal situation is predicted, issue an alarm in a timely manner and take corresponding measures.

[0171] The feedback optimization module automatically adjusts the parameters of the LSTM neural network and the task generation strategy based on the deviation between the prediction results and the actual equipment status, in order to improve the accuracy and adaptability of the model.

[0172] The online recognition unit includes: a feature matching subunit, a recognition decision subunit, and a recognition result database.

[0173] Feature matching subunit: Matches device feature data collected in real time from the feature database with known abnormal pattern features in the model. Calculates the similarity between real-time features and abnormal pattern features.

[0174] Identification Decision Subunit: Based on the feature matching results from the feature matching subunit, it determines whether the device's operating status is abnormal. If the similarity exceeds a set threshold, an anomaly alarm is triggered. Simultaneously, the category of the abnormal pattern is determined, and the anomaly identification result is pushed to the alarm center of the intelligent analysis module in real time. The identification result is also stored in the identification result database for subsequent analysis and traceability.

[0175] The equipment status prediction model in the intelligent analysis module uses an LSTM neural network to process time-series monitoring data, enabling accurate prediction of equipment operating status. The anomaly pattern recognition engine and risk rating unit can identify abnormal patterns in the equipment and assess the risk level. This gives the system powerful fault prediction and risk warning capabilities.

[0176] The feedback optimization module includes a data collection and processing unit, a deviation calculation unit, and a deviation analysis unit.

[0177] Data Collection and Processing Unit: Collects execution result data of inspection tasks from the historical records of the path planning module and the multi-source data acquisition module, including equipment failures and abnormal event records found during actual inspections. It also collects the prediction result data from the corresponding intelligent analysis module, i.e., the equipment status, failure probability, and other information predicted by the intelligent analysis module based on the then-input monitoring data before the inspection task is executed.

[0178] Deviation Calculation Unit: This unit compares the actual inspection results from the data collection and processing unit with the predicted results to calculate the deviation value. For numerical data (such as equipment temperature and pressure), the difference between the actual and predicted values ​​can be calculated directly. For categorical data (such as whether equipment is faulty), a 0-1 loss function can be used, where the deviation is 0 for correct prediction and 1 for incorrect prediction.

[0179] Deviation Analysis Unit: This unit performs statistical analysis on the deviation values ​​from the deviation calculation unit, calculating statistical indicators such as mean deviation and standard deviation to assess the overall accuracy of the prediction results. It analyzes the distribution of deviations to identify specific equipment, specific fault types, or specific time periods with large deviations, allowing for targeted model optimization.

[0180] Feedback optimization module parameter adjustment process:

[0181] (1) Learning rate adjustment: If the bias analysis unit finds that the prediction error of the equipment status prediction model of the intelligent analysis module is large and the convergence speed is slow, the learning rate of the model can be increased appropriately to speed up the parameter update speed; conversely, if the model becomes unstable in the later stage of training, the learning rate can be reduced appropriately to improve the stability of the model.

[0182] (2) Weight Update: Based on the magnitude and direction of the deviation, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and then the model's weight parameters are updated in the direction of gradient descent. For specific features or neurons with large deviations, the update magnitude of their weights can be appropriately increased to enhance the model's learning ability for that feature or neuron.

[0183] (3) Adjustment of regularization parameter: In order to prevent the equipment condition prediction model from overfitting or underfitting, the regularization parameter can be adjusted according to the results of the bias analysis. If the model performs well on the training data, but has a large bias on the actual inspection data (overfitting), the regularization parameter can be increased to enhance the generalization ability of the model; if the model performs poorly on both the training data and the actual data (underfitting), the regularization parameter can be decreased to enable the model to fit the data better.

[0184] In one or more embodiments, a task-driven AI-powered intelligent inspection system for chemical plants was deployed in the reactor area of ​​a chemical plant. The feedback optimization module played a crucial role in this system.

[0185] (1) Comparison between actual inspection and prediction results: In one inspection task, the AI ​​model predicted that the temperature of a certain reactor would fluctuate within the normal range, but the actual inspection results showed that the temperature of the reactor had risen abnormally.

[0186] (2) Deviation calculation and analysis: The feedback optimization module calculates the deviation between the predicted temperature value and the actual value and finds that the deviation exceeds the preset threshold.

[0187] (3) Adjustment of Equipment Status Prediction Model Parameters: Based on the deviation analysis results, the feedback optimization module adjusted the relevant parameters in the equipment status prediction model. Specifically, this included: appropriately increasing the weight coefficient of the reactor temperature monitoring data to improve the model's sensitivity to this parameter; and optimizing the connection weights of neurons related to reactor temperature prediction in the model.

[0188] (4) Task generation strategy adjustment: In subsequent task generation, the feedback optimization module increased the inspection priority of the reactor, temporarily adjusted the inspection frequency from once per hour to once every 30 minutes, and replanned the inspection route to ensure that the inspection robot can prioritize the detailed inspection of the reactor.

[0189] The system also includes a digital twin modeling module. This module comprises a data acquisition and preprocessing unit, a 3D modeling unit, a model fusion and update unit, and a visualization and interaction unit.

[0190] The data acquisition and preprocessing unit includes a sensor data receiver, a data format converter, and a data buffer. The sensor data receiver is responsible for collecting real-time operational data from various equipment within the factory, such as temperature, pressure, and vibration signals. The data format converter transforms heterogeneous data output from different types of sensors into a standardized data format. The data buffer temporarily stores the acquired data to ensure its continuity and integrity. This unit provides the digital twin modeling module with an accurate real-time data source, ensuring that the model reflects the latest state of the physical factory.

[0191] The 3D modeling unit comprises a modeling engine, a model library, and a texture mapping subunit. The modeling engine uses professional 3D modeling software (such as 3ds Max, Maya, etc.) or industrial modeling tools (such as AutoCAD, SolidWorks, etc.) to create 3D geometric models of the chemical plant. The model library stores prefabricated 3D models of various equipment, components, and buildings. The texture mapping subunit maps realistic texture maps onto the surface of the 3D model, enhancing its realism. The 3D modeling unit constructs a realistic 3D virtual model of the chemical plant, providing a foundational framework for subsequent visualization and simulation analysis.

[0192] The model fusion and update unit comprises a data fusion engine, a model update controller, and a semantic parser. The data fusion engine deeply integrates real-time collected equipment operation data with the 3D model, associating equipment status information and process parameters with corresponding model components. The model update controller triggers model update operations periodically or in real-time based on data change frequency and model accuracy requirements. The semantic parser parses the semantic relationships between equipment data and model components, achieving semantic-level fusion of data and model. This ensures dynamic synchronization between the digital twin modeling module and the physical factory, enabling the model to be updated in real-time and accurately reflect dynamic changes in equipment and the environment.

[0193] The visualization and interaction unit includes a graphics rendering engine, a user interface, and scene navigation tools. The graphics rendering engine employs advanced real-time rendering technologies (such as physically based rendering, PBR) to render and display 3D models and related data on the screen with high fidelity. The user interface supports input devices such as mouse, keyboard, and touchscreen, allowing users to easily adjust the viewing angle by rotating, scaling, and panning the model. The scene navigation tools provide functions such as path planning, bookmark positioning, and automatic navigation, assisting users in quickly locating and browsing target areas in complex 3D scenes. This provides users with a more intuitive and convenient visualization and interactive experience, enabling them to comprehensively observe and analyze the chemical plant's operational status.

[0194] The connection relationships between the digital twin modeling module and the external environment are as follows:

[0195] (1) Connection with the multi-source data acquisition module:

[0196] Data Acquisition Interface: The digital twin modeling module connects to the multi-source data acquisition module via a high-speed data acquisition bus (such as Industrial Ethernet, PROFIBUS, etc.) to acquire real-time equipment operation data from various sensors and monitoring devices. The multi-source data acquisition module preprocesses the acquired infrared thermal imaging data, gas concentration data, vibration spectrum data, and high-definition video data, and then sends them to the data acquisition and preprocessing module of the digital twin modeling module according to the agreed data format and transmission protocol.

[0197] Data synchronization mechanism: Establish a data timestamp synchronization mechanism to ensure consistency of data from different sources across time. Implement temporary data storage and sequential retrieval through message queues or data buffers to guarantee data integrity and accuracy.

[0198] (2) Connection with the intelligent analysis module:

[0199] Model data output interface: The model fusion and update module of the digital twin modeling module sends real-time updated digital twin model data, including equipment status, process parameters, and environmental information, to the intelligent analysis module via the data bus. The intelligent analysis module then performs analytical tasks such as equipment status prediction, anomaly pattern recognition, and risk assessment based on this data.

[0200] Analysis Result Feedback Interface: The intelligent analysis module feeds back the analyzed equipment fault prediction results, abnormal mode information, and risk ratings to the visualization and interaction module of the digital twin modeling module. The digital twin modeling module presents these analysis results on the 3D model in an intuitive visualization manner (such as highlighting model components and pop-up warning boxes), providing users with clear decision support information.

[0201] (3) Connection with the task generation module:

[0202] Task planning data interface: The digital twin modeling module provides the task generation module with a detailed 3D model of the chemical plant and equipment layout information through a data interface. Based on this information, the task generation module plans the optimal path and task allocation strategy for the inspection tasks.

[0203] Task execution feedback interface: The task generation module sends the generated inspection task information (such as task route, inspection points, task priority, etc.) to the digital twin modeling module. The digital twin modeling module displays the task execution status in real time in the 3D model, such as the current position of the inspection robot, completed inspection points, and remaining inspection paths, providing visual monitoring support for task execution.

[0204] (4) Connection with the factory management system:

[0205] Enterprise-level data interface: The digital twin modeling module integrates with factory production management systems (such as MES and ERP) and equipment maintenance systems via industrial communication protocols such as OPC-UA and Modbus TCP / IP. It acquires data such as production plans, equipment maintenance records, and material information from these systems, while simultaneously sending relevant information from the digital twin model (such as equipment operating status, inspection results, and fault warnings) to the factory management system, achieving data sharing and collaborative management. The system supports over 20 industrial protocols, including OPC-UA and Modbus, and can seamlessly integrate with existing DCS / PLC control systems. In cross-plant applications, data interoperability with MES and ERP systems is achieved through API interfaces, forming a closed loop of "perception-analysis-decision-execution".

[0206] By linking real-time environmental monitoring (combustible gases, toxic gases, temperature and humidity) with a digital twin virtual model, the system achieves closed-loop management of risk warning and emergency response. For example, when abnormal fluctuations in pipeline pressure are detected, the system automatically triggers an AR interface to mark the leak point and pushes a response plan to the smart terminal of maintenance personnel. Production safety risks are reduced by more than 40%.

[0207] Decision Support Interface: This provides a decision support interface for factory management, presenting key indicators and analysis results extracted from the digital twin model to managers through web services or data reports. Managers then use this information to optimize production scheduling, develop equipment maintenance plans, and make safety management decisions.

[0208] The system also includes a temporary task processing module, which comprises an emergency task receiving unit, a temporary task planning unit, a task coordination and execution unit, and a result feedback and integration unit.

[0209] The emergency task receiving unit includes a high-sensitivity signal receiving antenna, a signal decoder, and a task parser. The signal receiving antenna is responsible for capturing emergency detection request signals sent from human input devices (such as handheld terminals, control center computers, etc.). The signal decoder decodes the received signals, restoring them into recognizable instructions and data formats. The task parser performs semantic analysis and structured processing on the decoded task information, extracting key task elements such as task type (equipment detection, anomaly investigation, etc.), target equipment or area, and task priority. It receives and accurately parses manually input emergency detection requests in real time, transforming them into task information that the system can process.

[0210] The temporary task planning unit includes a task planner, a path generator, and a resource allocator. The task planner formulates temporary inspection plans based on the type, priority, and target location of the emergency task, combined with the overall operational status of the current inspection system. The path generator plans the optimal inspection path for inspection robots or personnel performing the temporary task based on a 3D map of the plant area and real-time road condition information. The resource allocator evaluates currently available inspection resources (such as the number of robots, sensor status, etc.) and allocates resources rationally to ensure the efficient execution of the temporary task. It quickly generates temporary inspection plans for urgent inspection needs, including task steps, path planning, and resource allocation.

[0211] The task coordination and execution unit includes a task scheduler, an instruction transmitter, and an execution status tracker. The task scheduler is responsible for assigning temporary tasks to specific inspection entities (robots or personnel) and coordinating the time sequence and resource usage conflicts between temporary tasks and existing inspection tasks. The instruction transmitter sends detailed inspection instructions to the inspection robot or personnel, including task content, path guidance, and operational requirements. The execution status tracker monitors the execution progress and status of temporary tasks in real time, such as task initiation, execution, completion, or termination. This ensures that temporary tasks can be executed smoothly according to the planned scheme and provides real-time monitoring of task execution.

[0212] The results feedback and integration unit comprises a data collector, a results evaluator, and a data integrator. The data collector gathers inspection result data from inspection robots or personnel performing temporary tasks, including detected equipment status, anomaly information, and on-site environmental data. The results evaluator performs a preliminary assessment of the inspection results, determining whether the task achieved its intended objectives, such as whether faults were detected and their severity. The data integrator integrates the inspection results of temporary tasks into the system's overall data architecture, including updating the equipment status database and historical inspection records. It collects and evaluates the execution results of temporary tasks and integrates them into the system's data architecture, providing a basis for subsequent data analysis and decision-making.

[0213] Connection relationship between temporary task processing module and external systems

[0214] (1) Connection with the task generation module

[0215] Task Information Input Interface: The temporary task processing module connects to the task generation module via a dedicated data interface, receiving routine inspection task information and task priority adjustment instructions from the task generation module. After generating an inspection task list according to preset inspection procedures, the task generation module sends basic task information (such as task ID, inspection equipment list, inspection route, etc.) to the temporary task processing module. Based on this information, the temporary task processing module understands the current routine inspection task schedule and can rationally integrate and adjust tasks when receiving urgent tasks.

[0216] Task Integration and Feedback Interface: After the temporary task processing module generates a temporary inspection plan, it uses this interface to feed back the temporary task information to the task generation module, including the temporary task's priority, estimated execution time, and required resources. Based on this information, the task generation module re-organizes and adjusts the overall inspection task plan to ensure that regular and temporary tasks can be carried out in a coordinated manner.

[0217] (2) Connection with the path planning module

[0218] Path planning request interface: When the temporary task processing module needs to plan a path for a temporary task, it sends a path planning request to the path planning module through this interface. The request information includes the task start point (the current location of the inspection robot or personnel), the task end point (the location of the target equipment or area), the task priority, and real-time traffic information. Upon receiving the request, the path planning module generates the optimal temporary inspection path based on its built-in 3D map of the factory area and an improved AI algorithm.

[0219] Path Update Feedback Interface: The path planning module sends the generated temporary inspection path back to the temporary task processing module, which then sends detailed path guidance instructions to the inspection robot or personnel. During the inspection, if real-time road conditions change (such as the appearance of new obstacles or road closures), the inspection robot or personnel use this interface to feed back the path update information to the path planning module. The path planning module then replans the path and updates the inspection route in a timely manner.

[0220] (3) Connection with multi-source data acquisition module

[0221] Data Request Interface: The temporary task processing module sends specific data acquisition instructions to the multi-source data acquisition module through the data request interface, specifying the data type to be acquired (such as infrared thermal imaging data, gas concentration data, etc.), acquisition frequency, and acquisition duration. During the execution of the temporary task, the data acquisition strategy is adjusted in real time according to the task requirements, such as increasing the acquisition frequency of vibration spectrum data in response to abnormal conditions of a certain device.

[0222] Data Receiving and Processing Interface: After the multi-source data acquisition module collects the corresponding data according to the instructions, it transmits the data to the temporary task processing module in real time through this interface. Upon receiving the data, the temporary task processing module performs preliminary data processing and analysis, extracting key information related to the temporary task to provide data support for subsequent task execution and result evaluation.

[0223] (4) Connection with inspection robots or personnel

[0224] Command Issuance Interface: The temporary task processing module connects to the control terminal of the inspection robot or personnel via a wireless communication network (such as Wi-Fi, 4G / 5G, etc.) or a wired communication link (such as industrial Ethernet) to issue temporary task commands. The command content includes detailed information such as task type, target location, inspection route, and data collection requirements. The inspection robot or personnel execute the corresponding inspection operations according to the received commands.

[0225] Status Monitoring and Feedback Interface: During the execution of temporary tasks, inspection robots or personnel can use this interface to provide real-time feedback on the task execution status to the temporary task processing module, such as current location, completed inspection steps, and any encountered anomalies. The temporary task processing module adjusts its task execution strategy promptly based on the feedback information; for example, when the inspection robot encounters an obstacle, it replans the path and sends new path guidance instructions.

[0226] (5) Connection with intelligent analysis module

[0227] Data Output Interface: The temporary task processing module transmits the collected temporary task inspection results data (including equipment status data, on-site environmental data, etc.) to the intelligent analysis module via the data bus. The intelligent analysis module uses its built-in equipment status prediction model and anomaly pattern recognition engine to perform in-depth analysis of the data, extracting equipment health status information, potential fault modes, etc.

[0228] Analysis Result Input Interface: The intelligent analysis module feeds back the analysis results (such as equipment failure probability, abnormal mode description, risk rating, etc.) to the temporary task processing module. Based on this, the temporary task processing module evaluates the execution effect of the temporary tasks and integrates the analysis results into the overall system data architecture, providing a basis for subsequent task decisions and optimization.

[0229] (6) Connection with the user interface of the inspection system (such as the control center and maintenance personnel terminal)

[0230] Task Receiving and Display Interface: The temporary task processing module connects to the user interface of the inspection system via network communication protocols (such as HTTP, WebSocket, etc.) to receive urgent inspection requests input by the user through the interface. The user interface provides an intuitive operation interface, facilitating user input of task information, such as selecting target equipment and setting task priorities. The temporary task processing module displays the received task information on the interface in real time, allowing users to understand the current task receiving status.

[0231] Task execution status and result display interface: The temporary task processing module provides real-time feedback on the execution status of temporary tasks (such as task in progress, completed, encountering problems, etc.) and the final inspection results (such as equipment normal, fault found and fault details, etc.) to the user interface through this interface. The user interface displays the task execution status to the user in an intuitive graphical way (such as progress bars, status indicator lights, pop-up windows, etc.), making it convenient for the user to understand the task progress and results in a timely manner.

[0232] This invention constructs an Equipment Health Index (EHI) model and a digital twin visualization platform to achieve global monitoring of production status. Managers can view equipment status in real time (such as pipeline pressure fluctuations and reactor temperature gradients) through a 3D visualization interface and predict the remaining lifespan of the equipment based on historical data (error <8%). It supports voice command interaction and 3D annotation guidance. A self-learning mechanism: through online incremental learning, the model's prediction drift suppression rate is improved by 72%, and the deployment cycle for new production lines is shortened to 3 days.

[0233] This invention, through technological innovation and system integration, has achieved a leapfrog upgrade of chemical inspection from "human-based" to "technology-based," generating significant benefits in terms of efficiency, safety, and cost, and possessing significant market competitiveness and industry promotion value.

[0234] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A task-based regional intelligent inspection system, characterized in that: include Multi-source data acquisition module, used to collect various data in the task area; Task generation module: Generates a dynamic task list based on data from the multi-source data acquisition module and the feedback optimization module; Intelligent task assignment module: Generates inspection tasks based on the data in the task list and combined with human-machine efficiency data; Path planning module: Generates differentiated inspection paths based on the regional 3D map and the inspection task; The intelligent analysis module establishes a new LSTM model and inputs real-time data of differentiated inspection paths into the new LSTM model to predict the current and future status of the equipment. The formula for the output gate in the new LSTM model is: the t =σ(W o ·[h t-1 ,x t ]+b o ), Among them, o t It is the output of the output gate, σ is the activation function, and h is the output of the output gate. t-1 It is a hidden state, which is then passed to the next time step, W. o It is the weight matrix, x t It is the input for the current time step, b o It is a bias term; The feedback optimization module automatically adjusts the parameters of the new LSTM model and the task generation strategy based on the deviation between the prediction results of the new LSTM model and the actual device status.

2. The task-based regional intelligent inspection system according to claim 1, characterized in that: The task generation module includes a data acquisition subunit, a data analysis subunit, and a priority ranking subunit. The data acquisition subunit is responsible for collecting equipment operating parameters, historical fault data, and environmental monitoring data in real time, and transmitting them to the data analysis subunit in real time. The data analysis subunit processes and analyzes the data from the data acquisition subunit, extracts key features and trends, and passes them to the priority ranking subunit. The priority sorting subunit prioritizes inspection tasks based on the analysis results of the data analysis subunit, using a preset algorithm to ensure that high-priority tasks are executed first.

3. The task-based regional intelligent inspection system according to claim 1, characterized in that: The intelligent task assignment module includes: Task attribute parsing unit: used to extract the task type, security risk level, and required detection equipment type from the dynamic task list; Resource status monitoring unit: Real-time acquisition of personnel location, robot battery level, and sensor availability data through the database of the multi-source data acquisition module; Human-machine performance evaluation unit: Constructs a task completion quality-efficiency matrix based on historical data; Dynamic dispatch engine unit: Based on the data from the task attribute parsing unit, resource status monitoring unit, and human-machine efficiency evaluation unit, it uses the AHP-TOPSIS hybrid decision model to output the optimal inspection task.

4. The task-based regional intelligent inspection system according to claim 1, characterized in that: The route planning module includes: The map processing unit determines the location of obstacles such as equipment, pipelines, and buildings on the 3D map, as well as road information, and then divides the 3D map into a grid. The path optimization unit calculates the optimal path from the starting point to the ending point, using a grid as the unit, based on the current inspection task; the optimal path includes both the personnel path and the robot path. The dynamic obstacle avoidance unit optimizes the obstacle path on the optimal path by combining real-time road condition information.

5. A task-based regional intelligent inspection system according to claim 1, characterized in that: The multi-source data acquisition module includes a mobile inspection terminal and an intelligent robot cluster.

6. The task-based regional intelligent inspection system according to claim 1, characterized in that: The intelligent analysis module includes a data preprocessing unit, a model training unit, a risk management unit, and an online identification unit. The data preprocessing unit cleans and extracts features from the data retrieved from the multi-source data acquisition module and stores it for subsequent analysis. The model training unit is used to build and train new LSTM models. The risk management unit is used to determine the risk level after marking equipment failure, human error, and environmental risk factors; The online identification unit is used to input real-time collected risk data into a trained LSTM model to predict the current and future states of the device.

7. A task-based regional intelligent inspection system according to claim 6, characterized in that: The method for constructing and training a new LSTM model is as follows: S1. Normalize the feature data required for training obtained from the feature database. S2. Construct the LSTM neural network structure; S3, Model Training Collect and label historical monitoring data; select loss function and optimizer to train the model; S4. Online Monitoring and Forecasting The real-time collected time-series monitoring data is input into the trained LSTM model; S5, State Prediction: The LSTM neural network predicts the current and future state of a device based on real-time input data and previously learned patterns and rules. S6. Results Analysis and Feedback: Analyze the prediction results, and if a potential equipment malfunction or other abnormal situation is predicted, issue an alarm in a timely manner and take corresponding measures.

8. A task-based regional intelligent inspection system according to claim 1, characterized in that: The feedback optimization module includes: Data collection and processing unit: collects and processes equipment failures, abnormal event records, predicted equipment status, and failure probabilities discovered during actual inspections; Deviation Calculation Unit: Compares the actual inspection results with the predicted results from the data collection and processing unit to calculate the deviation value; Deviation Analysis Unit: Performs statistical analysis on the deviation values ​​of the deviation calculation unit, calculates statistical indicators such as average deviation and standard deviation, and evaluates the overall accuracy of the prediction results.

9. A task-based regional intelligent inspection system according to claim 1, characterized in that: It also includes a digital twin modeling module, which interacts with the multi-source data acquisition module to ensure the consistency of data from different sources in the time dimension; The digital twin modeling module interacts with the intelligent analysis module, and presents the analysis results from the intelligent analysis module. The digital twin modeling module interacts with the task generation module and displays the task execution status in real time.

10. A task-based regional intelligent inspection system according to claim 1, characterized in that: It also includes a temporary task processing module, which comprises an emergency task receiving unit, a temporary task planning unit, a task coordination and execution unit, and a result feedback and integration unit.

Images