Equipment digital inspection evaluation method and system based on big data

Abstract

The invention discloses an equipment digital inspection evaluation method and system based on big data, and belongs to the field of artificial intelligence, and the method comprises the steps: S1, collecting multi-source heterogeneous data; s2, unifying timestamps by adopting a time synchronization technology, and realizing accurate alignment of multi-source data; s3, cleaning and structuring data by using a distributed stream processing framework, eliminating invalid redundancy, and completing coding standardization; s4, constructing a graph neural network model based on equipment topology, and extracting a state feature vector; s5, combining with historical fault labels, applying a reinforcement learning training evaluation strategy, and improving prediction accuracy and inspection efficiency; s6, dynamically sorting the inspection tasks according to a strategy, and determining a priority and a path in combination with the risk and the dependency relationship; s7, generating a digital inspection evaluation report fusing real-time analysis and a model result; and S8, performing digital signature and timestamp registration through an intelligent contract uplink evidence storage, and ensuring that a result is verifiable, traceable and non-tampered. The beneficial effects are that inspection efficiency is improved and data safety and reliability are ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more specifically, to a method and system for digital inspection and evaluation of equipment based on big data. Background Art

[0002] As the scale and complexity of industrial equipment continue to increase, traditional equipment inspection methods face challenges such as low efficiency, insufficient accuracy, and high labor costs. In recent years, with the advancement of big data technology, the Internet of Things, and artificial intelligence, digital equipment inspection and assessment methods have rapidly developed. By leveraging sensors to collect real-time equipment status data and combining it with big data analysis and machine learning models, accurate predictions of equipment status and fault warnings are achieved, significantly enhancing the intelligence and automation of inspections. Furthermore, the introduction of graphical models and path planning algorithms has further optimized inspection routes, improving operational efficiency and safety.

[0003] Despite significant progress in existing technologies, several shortcomings remain. First, the fixed edge weights in traditional graph models make it difficult to adapt to dynamic changes in equipment status, resulting in insufficient prediction accuracy and model adaptability. Second, path planning algorithms are prone to computing resource bottlenecks in complex equipment networks and lack pruning strategies that incorporate historical equipment fault information, affecting planning efficiency and the rationality of priority sorting. In addition, existing methods fail to adequately consider the actual location of inspection personnel and dynamic environmental constraints, limiting the practicality and flexibility of the solutions. Finally, the lack of effective coordination between model training and path planning makes it difficult to form a closed-loop optimization, limiting the overall performance improvement of digital inspection systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for digital inspection and evaluation of equipment based on big data, so as to solve the problems raised in the above-mentioned background technology: First, the edge weights in the traditional graph model are fixed, which makes it difficult to adapt to the dynamic changes in the equipment status, resulting in insufficient prediction accuracy and model adaptability. Secondly, the path planning algorithm is prone to computing resource bottlenecks in complex equipment networks, and lacks a pruning strategy that combines historical equipment fault information, which affects planning efficiency and the rationality of priority sorting. In addition, the existing methods do not give sufficient consideration to the actual location of inspection personnel and dynamic environmental constraints, which limits the practicality and flexibility of the solution. Finally, there is a lack of effective coordination between model training and path planning, making it difficult to form a closed-loop optimization, which limits the improvement of the overall performance of the digital inspection system.

[0005] Technical solution: The big data-based equipment digital inspection and evaluation method includes the following steps:

[0006] S1. Collecting multi-source heterogeneous data on equipment operation, including real-time sensor data, historical maintenance records, and environmental status information;

[0007] S2. Perform unified timestamp correction on the multi-source heterogeneous data based on time synchronization technology, and construct a high-precision timeline index structure to accurately align various types of data;

[0008] S3. Use a distributed stream processing framework to clean and structure the time-corrected data. The cleaning process removes invalid and redundant items and completes data encoding standardization to generate a standardized data stream.

[0009] S4. Build a graph neural network model of the device operating status, wherein the graph neural network model establishes a node graph structure based on the device topology relationship, uses a standardized data stream as input, and extracts a state feature vector;

[0010] S5. Use a reinforcement learning algorithm to train the evaluation strategy for the state feature vector, combining historical inspection data and equipment fault labels to optimize strategy parameters with the goal of maximizing state prediction accuracy and inspection coverage efficiency;

[0011] S6. Perform dynamic inspection task prioritization based on the trained strategy results, combining device risk values ​​and operational dependencies to determine the inspection path and priority level for each device;

[0012] S7. Generate a digital inspection assessment report that integrates real-time data analysis results, model evaluation scores, and a recommended task list;

[0013] S8. The contents of the assessment report shall be stored on the blockchain through a smart contract mechanism, and digitally signed and timestamped to ensure that the assessment results are verifiable, traceable and cannot be tampered with.

[0014] Preferably, when the S1 acquisition device runs multi-source heterogeneous data, the real-time sensor data is pre-processed by the edge computing node, and after the abnormal data is removed, it is transmitted to the central server, including the following steps:

[0015] S1-1. Real-time sampling data for batch outlier detection;

[0016] S1-2. Eliminate noise interference data based on the sliding window algorithm;

[0017] S1-3. Convert the format of the removed data to adapt to subsequent data stream processing requirements.

[0018] Preferably, the S3 uses a distributed stream processing framework to clean and structure the corrected data to form a standardized data stream.

[0019] The steps also include customizing a multi-level cache mechanism to achieve efficient multi-threaded concurrent cleaning of data streams, and the cleaning rules are driven by a dynamic rule engine.

[0020] The distributed stream processing framework implements a hybrid scheduling strategy based on micro-batch processing and event-driven processing to reduce latency and improve throughput.

[0021] The dynamic rule engine supports online rule version management and automatic switching to achieve seamless upgrade of data cleaning rules.

[0022] Preferably, when the S4 constructs the device operation status model based on the graph neural network, the device operation status model constructs a multi-layer graph convolutional network based on the device topology structure, adopts node feature normalization and edge weight adaptive adjustment strategy, and includes the following steps:

[0023] S4-1. Calculate the historical data correlation matrix between device nodes;

[0024] S4-2. Dynamically adjust the weights of adjacent edges based on the prediction error minimization criterion;

[0025] S4-3. Regularly update edge weight adjustment parameters to ensure the timeliness of graph convolutional network training.

[0026] Preferably, the reinforcement learning algorithm is used to train the evaluation strategy for the state feature vector in combination with historical inspection results and equipment fault labels.

[0027] The reinforcement learning algorithm adopts a deep deterministic policy gradient framework. The training process uses a multi-step reward mechanism and combines temporal difference error for dynamic policy optimization.

[0028] Reinforcement learning training introduces an environment state simulation module and enhances sample diversity through generative adversarial networks.

[0029] Preferably, the S6 performs dynamic inspection task priority sorting, adopts a Monte Carlo tree search-based path planning algorithm, and implements joint optimization of paths and tasks in combination with equipment risk scores, including the following steps:

[0030] S6-1. Establish pruning rules to limit the search tree expansion depth based on the historical failure frequency of the device;

[0031] S6-2. Integrate real-time device status weights and dynamically adjust node selection probability;

[0032] S6-3. Improving path planning efficiency using parallel Monte Carlo simulation.

[0033] Preferably, the digital inspection assessment report is generated by using multimodal data fusion technology to integrate text, images and sensor data, store them in a standardized JSON format, and complete the automatic signing and timestamp registration of the inspection results through the blockchain smart contract module;

[0034] Digital inspection assessment report generation combined with distributed storage system.

[0035] Preferably, the edge weight adaptive adjustment strategy dynamically adjusts the edge weights in the graph by minimizing the device state prediction error, and the update cycle is after each model iteration is completed; the dynamic adjustment further includes the following steps:

[0036] S4-1-1. Monitor the error distribution trend during model training;

[0037] S4-1-2. Combine gradient descent information to automatically adjust the edge weight update step size;

[0038] S4-1-3. Set a threshold to filter edge weight updates to prevent overfitting.

[0039] Preferably, the Monte Carlo tree search introduces pruning rules based on the historical failure frequency of equipment in path planning, limits the expansion depth of the search tree, and ensures that the algorithm completes the task priority sorting within a limited time, including the following steps:

[0040] S6-1-1. Dynamically evaluate the equipment failure probability threshold;

[0041] S6-1-2. Adjust pruning sensitivity in real time;

[0042] S6-1-3. Optimize the search space based on the inspection personnel’s location constraints.

[0043] Preferably, the equipment digital inspection and evaluation system based on big data is designed based on the equipment digital inspection and evaluation method based on big data and provides hardware support for it.

[0044] Compared with the prior art, the advantages of the present invention are:

[0045] (1) Automatically adjust edge weights based on device state prediction errors to improve model prediction accuracy and adaptability; intelligently adjust edge weight update step size based on error distribution trends and gradient information to enhance training stability; set thresholds to filter edge weight updates to avoid noise interference and overfitting.

[0046] (2) Dynamically limit the search tree expansion depth based on the historical failure probability of the equipment to improve the efficiency of path planning; adjust the pruning sensitivity in real time, taking into account the task priority and computing resources; integrate the location constraints of the inspection personnel, optimize the search space, and improve the practical feasibility of the planning.

[0047] (3) Achieve closed-loop optimization of model training and path planning, and improve the overall intelligence level of digital inspection and evaluation.

[0048] (4) Through adaptive adjustment and pruning, the computational complexity is significantly reduced to meet the needs of real-time online inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of the overall process of the device digital inspection and evaluation method for big data of the present invention; DETAILED DESCRIPTION

[0050] For examples, see Figure 1 ,The equipment digital inspection and evaluation method based on big data includes the following steps:

[0051] S1. Collect multi-source heterogeneous data on equipment operation, including real-time sensor data, historical maintenance records, and environmental status information;

[0052] S2. Use time synchronization technology to uniformly timestamp multi-source heterogeneous data and build a high-precision timeline index structure to accurately align various data types.

[0053] S3. Use a distributed stream processing framework to clean and structure the time-corrected data. The cleaning process removes invalid and redundant items and completes data encoding standardization to generate a standardized data stream.

[0054] S4. Build a graph neural network model of device operating status. The graph neural network model establishes a node graph structure based on the device topology, uses standardized data streams as input, and extracts state feature vectors.

[0055] S5. Use a reinforcement learning algorithm to train an evaluation strategy for the state feature vector. Combined with historical inspection data and equipment fault labels, optimize strategy parameters to maximize state prediction accuracy and inspection coverage efficiency.

[0056] S6. Perform dynamic inspection task prioritization based on the trained strategy results, combining device risk values ​​and operational dependencies to determine the inspection path and priority level for each device;

[0057] S7. Generate a digital inspection assessment report that integrates real-time data analysis results, model evaluation scores, and a list of recommended tasks;

[0058] S8. The contents of the assessment report will be stored on the chain through the smart contract mechanism, and digitally signed and timestamped to ensure that the assessment results are verifiable, traceable and cannot be tampered with.

[0059] Specifically, multi-source heterogeneous data collection: sensor data such as temperature, current, voltage, vibration, and images, combined with maintenance logs (JSON / XML) and meteorological data, enables comprehensive equipment modeling.

[0060] Time correction: Using NTP or GPS clock sources, a timestamp cache queue mechanism is introduced at the edge nodes to synchronize data alignment with an accuracy of ±1 second.

[0061] Data cleaning and structuring: Leveraging the Apache Flink / Storm framework, stream processing and state management are combined to achieve high-throughput, real-time cleaning.

[0062] Graph neural network modeling: Using models such as GCN / GAT, topology nodes represent devices, edges represent dependencies, and dynamic edge weight iteration;

[0063] Reinforcement learning strategy optimization: Build a simulation environment to train inspection strategies based on historical labels and state changes;

[0064] Path planning and task sequencing: Based on MCTS (Monte Carlo Tree Search), it integrates real-time risk levels to dynamically generate the optimal inspection path;

[0065] Assessment report generation and chain storage: Introduce a multimodal processor to output text + graphic reports, and achieve tamper-proof evidence storage through chain structures such as Ethereum or Fabric.

[0066] When the S1 acquisition device runs multi-source heterogeneous data, the real-time sensor data is pre-processed by the edge computing node, and after removing abnormal data, it is transmitted to the central server. The following steps are included:

[0067] S1-1. Real-time sampling data for batch outlier detection;

[0068] S1-2. Eliminate noise interference data based on the sliding window algorithm;

[0069] S1-3. Convert the format of the removed data to adapt to subsequent data stream processing requirements.

[0070] Specifically, during the S1 collection process, the edge node performs the following preprocessing process:

[0071] Batch anomaly detection: For data in a sliding window batch every 5 seconds, the mean and standard deviation are calculated in real time, and points outside 3σ are considered anomalies.

[0072] Verification of value range and logic rules: For example, "temperature < -50℃ or > 200℃", "voltage sudden change > ±20%" are considered invalid points and eliminated;

[0073] Format conversion: Raw data is collected in various formats, such as Modbus and CAN, and is uniformly converted into structured KV formats (such as JSON and Parquet);

[0074] Edge caching mechanism: Data is temporarily stored in the L2 / SSD buffer, compressed in batches, and sent to the central node to improve transmission efficiency and reliability.

[0075] S3 uses a distributed stream processing framework to clean and structure the corrected data to form a standardized data stream;

[0076] It also includes a custom multi-level caching mechanism to achieve efficient multi-threaded concurrent cleaning of data streams, and the cleaning rules are driven by a dynamic rule engine;

[0077] The distributed stream processing framework implements a hybrid scheduling strategy based on micro-batch processing and event-driven;

[0078] The dynamic rule engine supports online rule version management and automatic switching to achieve seamless upgrade of data cleaning rules.

[0079] Specifically, the cleaning and structuring process includes the following mechanisms: Distributed stream processing architecture: Based on Apache Flink, configured with micro-batch mechanism (batch size = 50ms) and event-driven scheduling;

[0080] Custom dynamic rule engine:

[0081] Example rule: R1: If the null value rate of a field in the data exceeds 20%, the entire row will be removed;

[0082] R2: The temperature value should satisfy the local time monotonicity. If there is a sudden rise or fall, it should be marked as a false value.

[0083] Configuration method: Use JSON or YAML to describe the rule set, including the rule version number and timestamp, and support hot loading;

[0084] Multi-level cache optimization: Introducing the LRU+time-driven cache cleaning mechanism, combined with Flink's state backend (RocksDB) to support high-concurrency writes;

[0085] Standardized output: The cleaned data is stored in a unified schema data stream (Parquet / Avro format) for consumption by the graph model and the policy model.

[0086] When S4 builds a device operation status model based on a graph neural network, the device operation status model builds a multi-layer graph convolutional network based on the device topology structure, adopts node feature normalization and edge weight adaptive adjustment strategy, including the following steps:

[0087] S4-1. Calculate the historical data correlation matrix between device nodes;

[0088] S4-2. Dynamically adjust the weights of adjacent edges based on the prediction error minimization criterion;

[0089] S4-3. Regularly update edge weight adjustment parameters to ensure the timeliness of graph convolutional network training.

[0090] Specific, concrete, topology modeling principles:

[0091] Node: Each device is a node, and its attributes include temperature, voltage, current, historical alarm count, etc.

[0092] Edges: Physical connections (such as cables), process sequences, and logical dependencies all constitute edges;

[0093] Feature matrix construction: construct node feature vector matrix Edge weight matrix

[0094] Training process: implemented using PyTorchGeometric or DGL;

[0095] Introducing DropEdge+LayerNorm to stabilize training;

[0096] After each training round, calculate the prediction error E i ,in accordance with

[0097]

[0098] in, Indicates the weight value of the edge between device node i and node j at the tth iteration;

[0099] Indicates update to the t+1th iteration value;

[0100] α: learning rate (step factor), used to control the amplitude of edge weight updates, usually a positive decimal;

[0101] E: The prediction error of the current model (loss function value), which is used to measure the difference between the prediction result of the graph neural network and the true value;

[0102] The loss function E is related to the edge weight w ij The partial derivative of , which indicates the sensitivity of the error to the edge weight.

[0103] S5 uses a reinforcement learning algorithm, combined with historical inspection results and equipment fault labels, to train evaluation strategies for state feature vectors;

[0104] The reinforcement learning algorithm adopts a deep deterministic policy gradient framework. The reinforcement training process uses a multi-step reward mechanism and combines the temporal difference error for dynamic policy optimization.

[0105] Reinforcement learning training introduces an environment state simulation module and enhances sample diversity through generative adversarial networks.

[0106] Specifically, the reinforcement learning training environment is constructed as follows:

[0107] State space: characteristic vectors of device nodes (such as current fluctuation amplitude, historical abnormality counts, and device type codes);

[0108] Action space: whether to inspect, adjust the inspection order, skip nodes;

[0109] Reward function design:

[0110] +1: hit device abnormality prediction;

[0111] -0.5: Inspect low-risk equipment or path redundancy;

[0112] -1: Missed detection or causing cascading failures;

[0113] Training model: DDPG (Deep Deterministic Policy Gradient) is used in conjunction with GAN to enhance sample diversity;

[0114] Optimization method: Use TD-error to guide strategic network updates;

[0115] After each round of training, the mini-batch is sampled and replayed through the replaybuffer;

[0116] Add softupdate to keep the target network stable.

[0117] S6 performs dynamic inspection task prioritization and uses a Monte Carlo tree search-based path planning algorithm combined with device risk scoring to achieve joint optimization of paths and tasks. The algorithm includes the following steps:

[0118] S6-1. Establish pruning rules to limit the search tree expansion depth based on the historical failure frequency of the device;

[0119] S6-2. Integrate real-time device status weights and dynamically adjust node selection probability;

[0120] S6-3. Improving path planning efficiency using parallel Monte Carlo simulation.

[0121] Specifically, the scheduling is performed based on MCTS (Monte Carlo Tree Search):

[0122] Node definition: Each inspection task is considered a node, including device ID, risk level, and location;

[0123] UCT formula: used to decide the next task selection:

[0124]

[0125] where w i is the node score, n iNumber of visits, N is the number of visits to the parent node;

[0126] Pruning strategy: If the device risk score is less than the threshold (e.g., 0.3), the child node will not be expanded;

[0127] Parallel simulation: Simulate multiple paths concurrently, with priority based on the total risk coverage of the paths;

[0128] Example application: Prioritize inspections of areas with frequent status changes, such as device groups with fluctuations > 10% / hour.

[0129] S7 generates digital inspection and assessment reports using multimodal data fusion technology, integrating text, images, and sensor data, storing them in a standardized JSON format, and automatically signing and timestamping the inspection results through a blockchain smart contract module.

[0130] Digital inspection assessment report generation combined with distributed storage system.

[0131] Specifically, the report generation and evidence storage process is as follows:

[0132] Evaluation report generation: Summary: inspection task completion status, model score (prediction accuracy, recall rate), risk heat map;

[0133] Format: Structured JSON+PDF report with embedded charts and trend analysis;

[0134] Multimodal fusion: comprehensive display of images (such as camera evidence), text (maintenance logs), and model output;

[0135] On-chain mechanism: Generate report hash summary using Ethereum or Hyperledger Fabric platform;

[0136] Call the smart contract and write it on the chain;

[0137] Contents include: report hash, generation timestamp, node signature;

[0138] Encrypted signature: Use ECC or RSA to sign the report to ensure uniqueness;

[0139] Result verification interface: Achieve a posteriori traceability through hash comparison and timestamp review.

[0140] The edge weight adaptive adjustment strategy dynamically adjusts the edge weights in the graph by minimizing the device state prediction error. The update cycle is after each model iteration. The dynamic adjustment also includes the following steps:

[0141] S4-1-1. Monitor the error distribution trend during model training;

[0142] S4-1-2. Combine gradient descent information to automatically adjust the edge weight update step size;

[0143] S4-1-3. Set a threshold to filter edge weight updates.

[0144] Specifically, the equipment operation status data and historical inspection records are collected, the equipment status graph is constructed, and the weights of each edge are initialized to empirical values ​​or initial values ​​calculated based on historical data.

[0145] Set the error threshold and step size range, and initialize the training parameters.

[0146] The device state is trained using a graph neural network (e.g., GCN), with mean squared error (MSE) used as the loss function, and the training error distribution is calculated periodically.

[0147] Error distribution trend monitoring (step S4-1-1): After each round of training, the error change trend corresponding to each edge is counted to identify the edge weights with increasing or decreasing errors.

[0148] Combine the error gradient with the training gradient information (step S4-1-2) and automatically adjust the update step size of the edge weight according to the error change rate.

[0149] When the error decreases slowly, increase the step size appropriately to speed up convergence; when the error oscillates, reduce the step size to ensure stability.

[0150] Set the error update threshold (step S4-1-3), and gradients smaller than the threshold will not participate in the edge weight update to avoid overfitting and noise interference.

[0151] After each round of training is completed, the edge weights are updated based on the above strategy and fed back to the next round of model training.

[0152] In path planning, Monte Carlo tree search introduces pruning rules based on the historical failure frequency of equipment to limit the expansion depth of the search tree, allowing the algorithm to complete task priority sorting within a limited time. The algorithm includes the following steps:

[0153] S6-1-1. Dynamically evaluate the equipment failure probability threshold;

[0154] S6-1-2. Adjust pruning sensitivity in real time;

[0155] S6-1-3. Optimize the search space based on the inspection personnel’s location constraints.

[0156] Specifically, historical equipment failure records are collected, the failure frequency of each equipment is calculated, and mapped to failure probability.

[0157] Set the fault probability threshold as the judgment criterion for pruning rules.

[0158] According to the current inspection task, initialize the search tree root node and set the search depth and time limit.

[0159] During the search process, for the device corresponding to the extended node, it is dynamically determined whether its failure probability is lower than a threshold (step S6-1-1).

[0160] For low-risk devices, limit the expansion depth, prune branches in advance, and reduce invalid searches.

[0161] Prioritize searches for high-risk equipment to ensure that critical equipment is inspected first.

[0162] According to factors such as inspection urgency, equipment environment status, and inspection personnel load (step S6-1-2), the pruning threshold and sensitivity are dynamically adjusted to ensure that pruning does not miss important nodes and effectively reduces the search space.

[0163] The real-time geographic location data of patrol personnel is introduced (step S6-1-3), which limits the path planning to expand only within a reasonable range, avoids unrealistic long-distance jumps, and improves planning efficiency.

[0164] After Monte Carlo tree search and pruning optimization, the priority ranking of equipment inspection tasks is obtained and output to the inspection system to guide actual operations.

[0165] The equipment digital inspection and evaluation system based on big data includes any one of the equipment digital inspection and evaluation methods based on big data according to claims 1-9.

[0166] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed in the present invention is defined by the attached claims and their equivalents.

Claims

1. The equipment digital inspection and evaluation method based on big data is characterized by: The device digital inspection and evaluation method based on big data includes the following steps: S1. Collect multi-source heterogeneous data on equipment operation, including real-time sensor data, historical maintenance records, and environmental status information; S2. Perform unified timestamp correction on the multi-source heterogeneous data based on time synchronization technology, and construct a high-precision timeline index structure to accurately align various types of data; S3. Use a distributed stream processing framework to clean and structure the time-corrected data. The cleaning process removes invalid and redundant items and completes data encoding standardization to generate a standardized data stream. S4. Build a graph neural network model of the device operating status, wherein the graph neural network model establishes a node graph structure based on the device topology relationship, uses a standardized data stream as input, and extracts a state feature vector; S5. Use a reinforcement learning algorithm to train the evaluation strategy for the state feature vector, combining historical inspection data and equipment fault labels to optimize strategy parameters with the goal of maximizing state prediction accuracy and inspection coverage efficiency; S6. Perform dynamic inspection task prioritization based on the trained strategy results, combining device risk values ​​and operational dependencies to determine the inspection path and priority level for each device; S7. Generate a digital inspection assessment report that integrates real-time data analysis results, model evaluation scores, and a recommended task list; S8. The contents of the assessment report shall be stored on the blockchain through a smart contract mechanism, and digitally signed and timestamped to ensure that the assessment results are verifiable, traceable and cannot be tampered with.

2. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: During the multi-source heterogeneous data collection process of the S1 acquisition device, the real-time sensor data is pre-processed by the edge computing node, and after removing abnormal data, it is transmitted to the central server, including the following steps: S1-1. Real-time sampling data for batch outlier detection; S1-2. Eliminate noise interference data based on the sliding window algorithm; S1-3. Convert the format of the removed data to adapt to subsequent data stream processing requirements.

3. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: The S3 uses a distributed stream processing framework to clean and structure the corrected data to form a standardized data stream; It also includes a custom multi-level caching mechanism to achieve efficient multi-threaded concurrent cleaning of data streams, and the cleaning rules are driven by a dynamic rule engine; The distributed stream processing framework implements a hybrid scheduling strategy based on micro-batch processing and event-driven; The dynamic rule engine supports online rule version management and automatic switching to achieve seamless upgrade of data cleaning rules.

4. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: When S4 constructs a device operation status model based on a graph neural network, the device operation status model constructs a multi-layer graph convolutional network based on the device topology structure, adopts node feature normalization and edge weight adaptive adjustment strategy, and includes the following steps: S4-1. Calculate the historical data correlation matrix between device nodes; S4-2. Dynamically adjust the weights of adjacent edges based on the prediction error minimization criterion; S4-3. Regularly update edge weight adjustment parameters to ensure the timeliness of graph convolutional network training.

5. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: The S5 uses a reinforcement learning algorithm to combine historical inspection results and equipment fault labels to train an evaluation strategy for the state feature vector; The reinforcement learning algorithm adopts a deep deterministic policy gradient framework, and the reinforcement training process uses a multi-step reward mechanism and combines the temporal difference error for dynamic policy optimization; The reinforcement learning training introduces an environment state simulation module and enhances sample diversity by generating an adversarial network.

6. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: The S6 performs dynamic inspection task priority sorting, adopts a Monte Carlo tree search-based path planning algorithm, and combines equipment risk scoring to achieve joint optimization of paths and tasks, including the following steps: S6-1. Establish pruning rules to limit the search tree expansion depth based on the historical failure frequency of the device; S6-2. Integrate real-time device status weights and dynamically adjust node selection probability; S6-3. Improving path planning efficiency using parallel Monte Carlo simulation.

7. The device digital inspection and evaluation method based on big data according to claim 1 is characterized in that: The S7 generates digital inspection and assessment reports using multimodal data fusion technology, integrating text, images, and sensor data, storing them in a standardized JSON format, and completing automatic signing and timestamp registration of inspection results through a blockchain smart contract module; The digital inspection assessment report is generated in conjunction with a distributed storage system.

8. The device digital inspection and evaluation method based on big data according to claim 4 is characterized in that: The edge weight adaptive adjustment strategy dynamically adjusts the edge weights in the graph by minimizing the device state prediction error, with an update cycle after each model iteration. The dynamic adjustment also includes the following steps: S4-1-1. Monitor the error distribution trend during model training; S4-1-2. Combine gradient descent information to automatically adjust the edge weight update step size; S4-1-3. Set a threshold to filter edge weight updates.

9. The device digital inspection and evaluation method based on big data according to claim 6 is characterized in that: The Monte Carlo tree search introduces pruning rules based on the historical failure frequency of equipment in path planning, limits the expansion depth of the search tree, and enables the algorithm to complete task priority sorting within a limited time. The algorithm includes the following steps: S6-1-1. Dynamically evaluate the equipment failure probability threshold; S6-1-2. Adjust pruning sensitivity in real time; S6-1-3. Optimize the search space based on the inspection personnel’s location constraints.

10. The equipment digital inspection and evaluation system based on big data is characterized by: The big data-based equipment digital inspection and evaluation system is designed based on any one of the big data-based equipment digital inspection and evaluation methods described in claims 1-9 and provides hardware support therefor.

Images