A special equipment safety detection data processing method and system

Abstract

The application discloses a kind of special equipment safety detection data processing method and system, and the initial special equipment multi-source data is obtained by sensor;Initial special equipment multi-source data is converted into structured feature vector, and is aligned with sensor data in time dimension;Build a hidden danger prediction network, update node features and edge weights in real time based on time sliding window and GAT graph attention mechanism, capture the timing changes of equipment operating state;The aligned special equipment multi-source data is input into GNN-GAT hidden danger prediction network model for identification;Based on DDPG deep deterministic policy gradient algorithm, determine the graded response strategy according to the level of equipment risk probability.It improves the efficiency of equipment operation, reduces the production loss caused by failure.

Description

Technical Field

[0001] This invention relates to the field of equipment safety testing technology, and in particular to a method and system for processing safety testing data of special equipment. Background Technology

[0002] During the operation of special equipment, collecting multi-source data such as vibration, temperature, pressure, and current through sensors is a crucial means of achieving equipment condition monitoring and fault early warning. However, traditional data processing methods for special equipment safety inspections suffer from numerous problems. On the one hand, the collected multi-source data often contains significant amounts of duplicate, missing, and noise interference. Sensor malfunctions, environmental electromagnetic interference, and other factors can lead to data distortion. Directly using raw data for analysis may result in erroneous conclusions, affecting equipment condition assessment. On the other hand, traditional data processing methods struggle to effectively uncover the complex relationships between multi-source data and cannot accurately capture the temporal changes in equipment operating status. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method for processing safety inspection data of special equipment.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned method for processing safety inspection data of special equipment:

[0005] Vibration, temperature, pressure and current data in special equipment are acquired by sensors to obtain multi-source data of special equipment. The multi-source data of special equipment is then processed by the DBSCAN clustering algorithm to obtain initial multi-source data of special equipment.

[0006] The initial multi-source data of special equipment is converted into structured feature vectors and aligned with sensor data in the time dimension to obtain aligned multi-source data of special equipment.

[0007] A hazard prediction network is constructed based on DynamicGNN, and node features and edge weights are updated in real time based on time sliding window and GAT graph attention mechanism to capture the temporal changes of equipment operating status, thus obtaining the GNN-GAT hazard prediction network model.

[0008] The multi-source data of the aligned special equipment is input into the GNN-GAT hazard prediction network model for identification, and the equipment risk probability is obtained.

[0009] Based on the DDPG deep deterministic policy gradient algorithm, a graded response strategy is determined according to the level of the device risk probability.

[0010] Furthermore, in the above-mentioned method for processing safety inspection data of special equipment, the step of processing the multi-source data of special equipment using the DBSCAN clustering algorithm to obtain initial multi-source data of special equipment includes:

[0011] Based on the multi-source data of the special equipment, determine the neighborhood radius ε and the minimum sample parameter MinPts of the DBSCAN clustering algorithm, and mark the multi-source data of the special equipment as unaccessed.

[0012] Select an unvisited data point P from the multi-source data of special equipment, define a neighborhood range with ε as the radius, and count the number of data points contained in the neighborhood.

[0013] If the number of data points in the neighborhood is greater than or equal to MinPts, then data point P is determined to be the core point, and a new cluster C is created for the core point P.

[0014] For the identified core point P, all data points in its neighborhood that have not yet been assigned to any cluster are assigned to cluster C and marked as visited, and the initial special equipment multi-source data is obtained iteratively.

[0015] Furthermore, in the aforementioned method for processing safety inspection data of special equipment, the step of converting the initial multi-source data of special equipment into structured feature vectors and aligning them with sensor data in the time dimension to obtain aligned multi-source data of special equipment includes:

[0016] Vibration feature data are extracted from the initial multi-source data of special equipment, including at least time-domain features and frequency-domain features;

[0017] The average, maximum, minimum, and rate of change of temperature are extracted from the initial multi-source data of the special equipment to obtain temperature feature data.

[0018] The pressure average, maximum, minimum and fluctuation amplitude features are extracted from the initial multi-source data of special equipment to obtain pressure feature data;

[0019] The average current value, effective value, peak value and frequency characteristics of the initial special equipment multi-source data are extracted to obtain current characteristic data;

[0020] The vibration feature data, pressure feature data, and current feature data are combined in a unified format to obtain a structured feature vector.

[0021] Furthermore, in the above-mentioned method for processing safety inspection data of special equipment, the step of constructing a hazard prediction network based on a Dynamic GNN (Graphical Neural Network) includes:

[0022] A hazard prediction network is constructed based on DynamicGNN, which includes at least a set of nodes, a set of edges, and a dynamic update mechanism.

[0023] Let the set of nodes represent the equipment components of special equipment, including at least bearings, gears and motors, and let the node characteristics include real-time monitoring data, historical operating data and design parameters of the equipment components;

[0024] The edge set represents the connection relationships and mutual influences between equipment components, and the edge weight represents the degree of association between equipment components;

[0025] The dynamic update mechanism uses a time sliding window and GAT graph attention mechanism to update node features and edge weights in real time to capture temporal changes in device operating status.

[0026] Furthermore, in the aforementioned method for processing special equipment safety inspection data, the step of updating node features and edge weights in real time based on a time sliding window and a GAT graph attention mechanism to capture the temporal changes in equipment operating status and obtain a GNN-GAT hazard prediction network model further includes:

[0027] Based on the time scale of equipment failure development and the time resolution of data, by analyzing historical failure data, and based on the average time interval from the appearance of early potential problems to failure, the sliding window size is set to be greater than this time interval to determine the time sliding window size;

[0028] In the GAT graph attention layer, the attention weights between each node and its neighboring nodes are calculated; the importance of different components is learned through the attention weights, and the GNN-GAT hazard prediction network model is obtained.

[0029] Furthermore, in the above-mentioned method for processing safety inspection data of special equipment, the step of inputting the aligned multi-source data of special equipment into the GNN-GAT hazard prediction network model for identification to obtain the equipment risk probability includes:

[0030] The multi-source data of the alignment special equipment is normalized to convert the data with different features to the range of [0,1], and a normalized feature vector is obtained.

[0031] The normalized feature vector is input into the trained GNN-GAT hazard prediction network model, and the features and edge weights of each node are calculated through forward propagation.

[0032] It captures the temporal changes in the operating status of the equipment and the interactions between components, and outputs the risk probability of the equipment.

[0033] Furthermore, in the above-mentioned special equipment safety inspection data processing method, the step of determining a graded response strategy based on the DDPG deep deterministic strategy gradient algorithm according to the level of equipment risk probability includes:

[0034] The operating status of special equipment is regarded as the environment, which includes the risk probability of the equipment, the real-time monitoring data of each sensor, and the historical operating data of the equipment.

[0035] Different graded response strategies are defined as actions, and the corresponding graded response strategies are output using the DDPG algorithm based on the real-time calculated device risk probability.

[0036] Furthermore, in a system for processing safety inspection data of special equipment, the system includes the following modules:

[0037] The data acquisition and clustering module is used to acquire vibration, temperature, pressure and current data in special equipment through sensors to obtain multi-source data of special equipment. The DBSCAN clustering algorithm is used to process the multi-source data of special equipment to obtain initial multi-source data of special equipment.

[0038] The feature extraction and alignment module is used to convert the initial multi-source data of special equipment into structured feature vectors and align them with the sensor data in the time dimension to obtain aligned multi-source data of special equipment.

[0039] The prediction model building module is used to construct a hazard prediction network based on DynamicGNN dynamic graph neural network. It updates node features and edge weights in real time based on time sliding window and GAT graph attention mechanism to capture the temporal changes of equipment operating status and obtain GNN-GAT hazard prediction network model.

[0040] The risk probability identification module is used to input the multi-source data of the aligned special equipment into the GNN-GAT hidden danger prediction network model for identification, and to obtain the equipment risk probability.

[0041] The response strategy specification module is used to determine a graded response strategy based on the risk probability level of the device using the DDPG deep deterministic policy gradient algorithm.

[0042] Furthermore, in the system for implementing the above-mentioned special equipment safety inspection data processing method, the risk probability identification module includes the following sub-modules:

[0043] The normalization submodule is used to normalize the multi-source data of the alignment special equipment, transforming data with different features into the range of [0,1] to obtain a normalized feature vector;

[0044] The calculation submodule is used to input the normalized feature vector into the trained GNN-GAT hazard prediction network model and calculate the features and edge weights of each node through forward propagation.

[0045] The capture submodule is used to capture the temporal changes in the device's operating status and the interactions between components, and output the device's risk probability.

[0046] Furthermore, in the system implementing the above-mentioned special equipment safety inspection data processing method, the response strategy specification module includes the following sub-modules:

[0047] The setting submodule is used to treat the operating status of special equipment as the environment. The status includes the risk probability of the equipment, the real-time monitoring data of each sensor, and the historical operating data of the equipment.

[0048] Define a submodule to define different graded response strategies as actions. Based on the real-time calculated device risk probability, use the DDPG algorithm to output the corresponding graded response strategy.

[0049] Its beneficial effects are as follows: Vibration, temperature, pressure, and current data from special equipment are acquired through sensors to obtain multi-source data for the special equipment. This multi-source data is then processed using the DBSCAN clustering algorithm to obtain initial multi-source data. This initial multi-source data is converted into structured feature vectors and aligned with the sensor data in the time dimension to obtain aligned multi-source data. A hazard prediction network is constructed based on the DynamicGNN dynamic graph neural network. Node features and edge weights are updated in real time using a time sliding window and GAT graph attention mechanism to capture the temporal changes in equipment operating status, resulting in a GNN-GAT hazard prediction network model. The aligned multi-source data is input into the GNN-GAT hazard prediction network model for identification to obtain the equipment risk probability. Based on the DDPG deep deterministic policy gradient algorithm, a graded response strategy is determined according to the level of the equipment risk probability. 1. Early identification of equipment hazards is more sensitive and accurate, allowing for the early detection of potential failure risks, saving more time for equipment maintenance, and reducing equipment failure rates and repair costs. 2. Intelligent and precise response strategies are achieved. This tiered response mechanism can minimize unnecessary downtime, improve equipment operating efficiency, and reduce production losses caused by malfunctions while ensuring equipment safety. It also enhances the safety and reliability of special equipment operation, providing strong support for industrial production and social life. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0051] Figure 1 This is a schematic diagram of the first embodiment of a special equipment safety inspection data processing method according to the present invention;

[0052] Figure 2 This is a schematic diagram of a second embodiment of a special equipment safety inspection data processing method according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the first embodiment of a special equipment safety inspection data processing system according to the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0056] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1 As shown, a method for processing safety inspection data of special equipment includes the following steps:

[0057] Step 101: Obtain vibration, temperature, pressure and current data in special equipment through sensors to obtain multi-source data of special equipment. Process the multi-source data of special equipment through DBSCAN clustering algorithm to obtain initial multi-source data of special equipment.

[0058] Specifically, in this embodiment, the neighborhood radius ε and minimum sample parameter MinPts of the DBSCAN clustering algorithm are determined based on multi-source data of special equipment, and the multi-source data of special equipment is marked as unaccessed.

[0059] Select an unvisited data point P from the multi-source data of special equipment, define a neighborhood range with ε as the radius, and count the number of data points contained in the neighborhood.

[0060] If the number of data points in the neighborhood is greater than or equal to MinPts, then data point P is determined to be the core point, and a new cluster C is created for the core point P.

[0061] For the identified core point P, all data points in its neighborhood that have not yet been assigned to any cluster are assigned to cluster C and marked as visited, and the initial special equipment multi-source data is obtained iteratively.

[0062] Step 102: Convert the initial multi-source data of special equipment into structured feature vectors and align them with the sensor data in the time dimension to obtain aligned multi-source data of special equipment;

[0063] Specifically, in this embodiment, vibration characteristic data is extracted from the initial multi-source data of special equipment, including at least time-domain characteristics and frequency-domain characteristics;

[0064] The average, maximum, minimum, and rate of change of temperature are extracted from the initial multi-source data of special equipment to obtain temperature characteristic data.

[0065] Pressure characteristic data is obtained by extracting the average, maximum, minimum, and fluctuation amplitude features of pressure from the initial multi-source data of special equipment.

[0066] The average current, RMS current, peak current, and frequency characteristics are extracted from the initial multi-source data of special equipment to obtain current characteristic data;

[0067] Vibration feature data, pressure feature data, and current feature data are combined in a unified format to obtain a structured feature vector.

[0068] Step 103: Construct a hazard prediction network based on DynamicGNN dynamic graph neural network, and update node features and edge weights in real time based on time sliding window and GAT graph attention mechanism to capture the temporal changes of equipment operating status and obtain GNN-GAT hazard prediction network model.

[0069] Specifically, in this embodiment, a hazard prediction network is constructed based on DynamicGNN, which includes at least a set of nodes, a set of edges, and a dynamic update mechanism;

[0070] Let the set of nodes represent the equipment components of special equipment, including at least bearings, gears and motors, and let the node characteristics include real-time monitoring data, historical operating data and design parameters of the equipment components;

[0071] The edge set represents the connection relationships and mutual influences between equipment components, and the edge weight represents the degree of association between equipment components;

[0072] The dynamic update mechanism uses a time sliding window and GAT graph attention mechanism to update node features and edge weights in real time to capture temporal changes in device operating status.

[0073] Based on the time scale of equipment failure development and the time resolution of data, by analyzing historical failure data, and based on the average time interval from the appearance of early potential problems to failure, the sliding window size is set to be greater than this time interval to determine the time sliding window size;

[0074] In the GAT graph attention layer, the attention weights between each node and its neighboring nodes are calculated; the importance of different components is learned through the attention weights, and the GNN-GAT hazard prediction network model is obtained.

[0075] Step 104: Input the multi-source data of the special equipment into the GNN-GAT hazard prediction network model for identification to obtain the equipment risk probability;

[0076] Specifically, in this embodiment, the multi-source data of the special equipment will be normalized to convert the data with different features to the range of [0,1], and a normalized feature vector will be obtained.

[0077] The normalized feature vector is input into the trained GNN-GAT hazard prediction network model, and the features and edge weights of each node are calculated through forward propagation.

[0078] It captures the temporal changes in the operating status of the equipment and the interactions between components, and outputs the risk probability of the equipment.

[0079] Step 105: Based on the DDPG deep deterministic policy gradient algorithm, determine the graded response strategy according to the level of equipment risk probability.

[0080] Specifically, in this embodiment, the operating status of special equipment is regarded as the environment, which includes the risk probability of the equipment, the real-time monitoring data of each sensor, and the historical operating data of the equipment;

[0081] Different graded response strategies are defined as actions, and the corresponding graded response strategies are output using the DDPG algorithm based on the real-time calculated device risk probability.

[0082] Its beneficial effects are as follows: 1. It is more sensitive and accurate in identifying early-stage equipment hazards, enabling the early detection of potential failure risks, buying more time for equipment maintenance, and reducing equipment failure rates and repair costs. 2. It achieves intelligent and precise response strategies. This tiered response mechanism can minimize unnecessary downtime, improve equipment operating efficiency, and reduce production losses caused by failures while ensuring equipment safety. It also enhances the safety and reliability of special equipment operation, providing strong support for industrial production and social life.

[0083] Please see Figure 2 In a method for processing safety inspection data of special equipment, constructing a hazard prediction network based on a Dynamic GNN includes the following steps:

[0084] Step 201: Construct a hazard prediction network based on DynamicGNN, which includes at least a set of nodes, a set of edges, and a dynamic update mechanism;

[0085] Step 202: Let the node set represent the equipment components of the special equipment, including at least bearings, gears and motors. The node characteristics include real-time monitoring data, historical operating data and design parameters of the equipment components.

[0086] Step 203: The edge set represents the connection relationship and mutual influence between equipment components, and the edge weight represents the degree of association between equipment components;

[0087] Step 204: The dynamic update mechanism uses a time sliding window and GAT graph attention mechanism to update node features and edge weights in real time to capture the temporal changes in the device's operating status.

[0088] Please see Figure 3 In a special equipment safety inspection data processing system, the system includes the following modules:

[0089] The data acquisition and clustering module is used to acquire vibration, temperature, pressure and current data in special equipment through sensors to obtain multi-source data of special equipment. The DBSCAN clustering algorithm is used to process the multi-source data of special equipment to obtain the initial multi-source data of special equipment.

[0090] The feature extraction and alignment module is used to convert the initial multi-source data of special equipment into structured feature vectors and align them with the sensor data in the time dimension to obtain aligned multi-source data of special equipment.

[0091] The prediction model building module is used to construct a hazard prediction network based on DynamicGNN dynamic graph neural network. It updates node features and edge weights in real time based on time sliding window and GAT graph attention mechanism to capture the temporal changes of equipment operating status and obtain GNN-GAT hazard prediction network model.

[0092] The risk probability identification module is used to input multi-source data of special equipment into the GNN-GAT hidden danger prediction network model for identification and to obtain the risk probability of the equipment.

[0093] The response strategy specification module is used to determine a graded response strategy based on the DDPG deep deterministic policy gradient algorithm and the level of device risk probability.

[0094] Specifically, this embodiment can also be implemented in the following ways:

[0095] I. Data Collection

[0096] (I) Sensor Selection

[0097] Based on the different types and working characteristics of special equipment, select appropriate sensors: Rotating machinery (such as motors, pumps, etc.): focus on monitoring vibration, temperature and current data.

[0098] Pressure vessels (such as boilers and storage tanks): mainly monitor pressure and temperature data.

[0099] For lifting machinery (such as cranes and elevators): it is necessary to monitor vibration, pressure (hydraulic system pressure), and current (motor current) data.

[0100] II. Data Preprocessing

[0101] (a) Clustering Algorithm Selection

[0102] 1. Parameter initialization

[0103] Before applying the DBSCAN algorithm to process multi-source data (vibration, temperature, pressure, and current data) from special equipment, two key parameters need to be determined: the neighborhood radius (ε) and the minimum number of samples (MinPts). The neighborhood radius (ε) defines the range of the data point's neighborhood, and its value needs to be determined in conjunction with the data distribution characteristics and the data volume.

[0104] 2. Determining the core points

[0105] Select an unvisited data point P from the dataset. Define a neighborhood with P as the center and ε as the radius, and count the number of data points within this neighborhood. If the number of data points in the neighborhood is greater than or equal to MinPts, then point P is determined to be a core point, and a new cluster C is created for it. If the number of data points in the neighborhood is less than MinPts, then point P cannot become a core point temporarily and remains in the "unvisited" state. For example, when processing vibration data, if there are enough data points in the neighborhood of a data point that meet the MinPts condition, it is determined to be a core point, and a new cluster is started.

[0106] 3. Cluster growth

[0107] For a identified core point P, all data points within its neighborhood that have not yet been assigned to any cluster are added to cluster C and marked as visited. Next, these newly added data points in cluster C are checked. If a core point exists (i.e., the number of data points in its neighborhood is greater than or equal to MinPts), then the unassigned points in the neighborhood of that core point are also added to cluster C. This process is repeated continuously to achieve the continuous growth and expansion of clusters.

[0108] 4. Boundary point processing

[0109] During cluster growth, some data points may be encountered whose neighborhood contains fewer than MinPts, but these points are located within the neighborhood of a core point. These points are defined as boundary points. Although boundary points do not meet the core point criteria themselves, they are attached to the cluster in which the core point belongs, and are thus assigned to the corresponding cluster, thereby completing the boundaries of the clusters.

[0110] 5. Noise Point Identification

[0111] Once all data points in the dataset have been accessed, those data points that neither belong to any cluster nor are marked as boundary points are identified as noise points. These noise points may be invalid data generated by factors such as sensor malfunctions or environmental interference. Removing them from the dataset yields initial multi-source data for special equipment after clustering processing, effectively improving data quality and providing a reliable data foundation for subsequent data transformation and hazard prediction.

[0112] (ii) Duplicate data removal

[0113] By comparing the timestamps and feature values ​​of the data, it is determined whether the data is duplicated. Data with identical timestamps and feature values ​​within the allowable error range are considered duplicates and are removed to ensure the uniqueness and validity of the data.

[0114] (III) Missing Data Processing

[0115] For missing data, first determine the missing percentage and missing pattern. If the missing percentage is small (e.g., less than 5%), use interpolation to fill the gaps. The appropriate interpolation method should be selected based on the data type: for time-series correlated data such as vibration and current, use linear interpolation or cubic spline interpolation, interpolating data from consecutive time points; for relatively stable data such as temperature and pressure, use mean interpolation, filling the gaps with the average value of the sensor over similar time periods. If the missing percentage is large (e.g., greater than 10%), remove the missing data to avoid affecting the accuracy of subsequent analysis due to excessive data filling.

[0116] III. Data Conversion and Alignment

[0117] (I) Construction of Structured Feature Vectors

[0118] Vibration data feature extraction: Extract time-domain features (such as peak value, mean value, root mean square value, kurtosis, skewness, etc.) to reflect the intensity and waveform characteristics of the vibration signal; extract frequency-domain features (such as frequency, amplitude spectrum, power spectral density, etc.), convert the time-domain signal into a frequency-domain signal through Fast Fourier Transform (FFT), analyze the energy distribution of different frequency components, and identify the abnormal vibration frequency of the equipment.

[0119] Temperature data feature extraction: Extract features such as average temperature, maximum temperature, minimum temperature, and rate of change to determine whether the temperature is within the normal range and the trend of temperature change.

[0120] Pressure data feature extraction: Extract features such as average value, maximum value, minimum value, and fluctuation range of pressure to reflect the stability and changes of pressure.

[0121] Current data feature extraction: Extract features such as average value, effective value, peak value, and frequency of current to determine the load status and normal operation of the motor.

[0122] The extracted features are combined in a unified format to form structured feature vectors, with each feature vector corresponding to the device operating status data at a specific point in time.

[0123] (ii) Time dimension alignment

[0124] Time calibration technology: Time calibration is performed using a combination of hardware and software synchronization. Hardware synchronization provides a unified time reference for each sensor through an external clock source (such as a GPS clock or a BeiDou clock), ensuring that the sensors have accurate timestamps when collecting data. Software synchronization checks and adjusts the timestamps of each sensor's data after data acquisition, and uses time interpolation algorithms (such as linear interpolation and nearest neighbor interpolation) to align the data from different sensors to a unified time grid.

[0125] Time sliding window setting: Based on the operating cycle and data change frequency of the special equipment, set an appropriate time sliding window size. For high-frequency vibration data, the window size can be set to a shorter time interval (e.g., 1 second) to capture rapidly changing signal characteristics; for data with relatively slow changes such as temperature and pressure, the window size can be appropriately increased (e.g., 10 seconds). During data alignment, data from each sensor is processed synchronously in units of the time sliding window to ensure temporal consistency of data within the same window.

[0126] IV. Construction of Hazard Prediction Network

[0127] (I) Principles and Structure of Dynamic GNN

[0128] DynamicGNN consists of a set of nodes, a set of edges, and a dynamic update mechanism. Nodes represent key components of special equipment (such as bearings, gears, and motors), and node features include real-time monitoring data (such as vibration characteristics and temperature characteristics), historical operating data, and design parameters. Edges represent the connections and mutual influences between components, and edge weights reflect the degree of correlation between components. The dynamic update mechanism uses a time sliding window and a GAT graph attention mechanism to update node features and edge weights in real time to capture the temporal changes in the equipment's operating status.

[0129] (ii) Determining the size of the time sliding window

[0130] Based on the timescale of equipment failure development and the temporal resolution of the data, the size of the time sliding window is reasonably determined. By analyzing historical failure data, the average time interval from the appearance of early potential problems to the occurrence of failure is determined. The sliding window size is then set to be slightly larger than this time interval to capture key characteristic changes during the failure development process. Simultaneously, considering the data acquisition frequency, it is ensured that the window contains a sufficient number of valid data points; generally, the number of data points within the window should not be less than 20.

[0131] (III) Application of GAT graph attention mechanism

[0132] In the GAT graph attention layer, for each node, its attention weight with neighboring nodes is calculated. Through the attention weight, the model can automatically learn the importance between different components, focusing on the components that have a greater impact on the operating state of the equipment and their interactions, thereby more accurately capturing the temporal change characteristics of the equipment.

[0133] (iv) Network model training

[0134] Training data preparation: Collect historical monitoring data of special equipment under normal operation and different fault conditions, including multi-source data such as vibration, temperature, pressure, and current, as well as corresponding equipment fault labels (such as normal, bearing wear, gear failure, overload, etc.). Divide the data into training set, validation set, and test set, with a typical ratio of 7:2:1.

[0135] Training parameter settings: Select a suitable optimizer (such as the Adam optimizer), set the learning rate to 0.001-0.01, and set an appropriate number of training epochs based on model complexity and data volume, generally 50-200 epochs. During training, use early stopping to prevent overfitting; stop training when the loss function on the validation set no longer decreases after 10 consecutive training epochs.

[0136] Model evaluation metrics: The model is evaluated using metrics such as accuracy, precision, recall, and F1 score to ensure that the model has high performance in predicting equipment hazard.

[0137] V. Equipment Risk Probability Identification

[0138] (a) Data preprocessing

[0139] Before inputting the aligned multi-source data of special equipment into the GNN-GAT hazard prediction network model, the data is normalized to transform data with different features into the range of [0,1] or [-1,1], avoiding the impact of different feature units on the training and prediction performance of the model. Commonly used normalization methods include min-max normalization and Z-score normalization; the appropriate method should be selected based on the distribution characteristics of the data.

[0140] (II) Model Input and Prediction

[0141] The normalized structured feature vectors are input into the trained GNN-GAT hazard prediction network model. The model calculates the features and edge weights of each node through forward propagation, capturing the temporal changes in equipment operating status and the interactions between components, and finally outputs the risk probability of the equipment. The risk probability ranges from [0,1], with a larger value indicating a higher probability of equipment failure.

[0142] VI. Determining the Tiered Response Strategy

[0143] (a) Risk level classification

[0144] Based on the probability of equipment risk, the equipment risk level is divided into four levels:

[0145] Level 1 (Safe condition, risk probability ≤ 0.2): The equipment is operating normally and requires no special handling. Regular inspections and maintenance should be carried out according to the normal cycle.

[0146] Level 2 (Attention status, 0.2 < Risk probability ≤ 0.5): The equipment may have early-stage hidden dangers. It is necessary to increase the frequency of inspections, closely monitor the equipment's operating status, record abnormal data, and conduct further analysis and diagnosis.

[0147] Level 3 (Warning status, 0.5 < Risk probability ≤ 0.8): The equipment has obvious hidden dangers and may fail at any time. It should be shut down for maintenance immediately, and key components should be thoroughly inspected and tested, and a maintenance plan should be developed.

[0148] Level 4 (Dangerous state, risk probability > 0.8): The equipment is in a high-risk state and may experience serious failure at any time, endangering the safety of personnel and equipment. The equipment must be stopped immediately, emergency shutdown measures must be taken, and professional personnel must be organized to carry out comprehensive inspection and repair. The equipment can only be put back into operation after it has been repaired and passed strict testing.

[0149] (II) Application of DDPG Deep Deterministic Strategy Gradient Algorithm

[0150] The DDPG algorithm is a model-based reinforcement learning algorithm capable of policy optimization in a continuous action space. In this scheme, the state space represents the risk probability and current operating state characteristics of the equipment, while the action space represents different hierarchical response strategies (such as normal inspection, increasing inspection frequency, shutdown for maintenance, emergency shutdown, etc.). The reward function is designed based on the changes in equipment risk and economic benefits after adopting different strategies. By training the DDPG algorithm, the agent can automatically select the optimal hierarchical response strategy based on the equipment's risk probability, achieving effective management and control of equipment risk. The specific steps are as follows:

[0151] Environmental modeling: The operating status of special equipment is regarded as the environment, which includes the risk probability of the equipment, real-time monitoring data of each sensor, and historical operating data of the equipment.

[0152] Action definition: Define different graded response strategies as actions, such as normal inspection actions for level 1 risks and actions to increase inspection frequency for level 2 risks.

[0153] Reward Function Design: The reward function aims to encourage the selection of strategies that reduce equipment risk and improve equipment safety and economy. For example, when a shutdown and maintenance strategy is adopted, the probability of equipment risk decreases, and a positive reward is given; if an inappropriate strategy leads to equipment failure, a negative reward is given.

[0154] Algorithm training: The DDPG algorithm is trained using historical data and real-time feedback data, and the parameters of the policy network and value network are adjusted so that the agent can learn the optimal hierarchical response policy.

[0155] Strategy Execution: Based on the real-time calculated equipment risk probability, the DDPG algorithm outputs a corresponding graded response strategy to guide operators or automatic control systems to handle the equipment accordingly.

[0156] The foregoing has shown and described 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. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for processing safety inspection data of special equipment, characterized in that, The special equipment safety inspection data processing method includes the following steps: Vibration, temperature, pressure and current data in special equipment are acquired by sensors to obtain multi-source data of special equipment. The multi-source data of special equipment is then processed by the DBSCAN clustering algorithm to obtain initial multi-source data of special equipment. The initial multi-source data of special equipment is converted into structured feature vectors and aligned with sensor data in the time dimension to obtain aligned multi-source data of special equipment. A hazard prediction network is constructed based on DynamicGNN, comprising at least a node set, an edge set, and a dynamic update mechanism. The node set represents the equipment components of special equipment, including at least bearings, gears, and motors. Node features include real-time monitoring data, historical operating data, and design parameters of the equipment components. The edge set represents the connections and mutual influences between equipment components, with edge weights indicating the degree of correlation between them. The dynamic update mechanism uses a time-sliding window and GAT graph attention mechanism to update node features and edge weights in real time, capturing the temporal changes in equipment operating status. Based on the time scale of equipment failure development and the temporal resolution of the data, and by analyzing historical failure data, the sliding window size is set to be greater than the average time interval from the appearance of an early hazard to the occurrence of a failure, thus determining the time-sliding window size. In the GAT graph attention layer, the attention weights between each node and its adjacent nodes are calculated. The importance of different components is learned through these attention weights, resulting in the GNN-GAT hazard prediction network model. The multi-source data of the aligned special equipment is input into the GNN-GAT hazard prediction network model for identification, and the equipment risk probability is obtained. Based on the DDPG deep deterministic policy gradient algorithm, a graded response strategy is determined according to the level of the device risk probability.

2. The special equipment safety inspection data processing method as described in claim 1, characterized in that, The process of processing the multi-source data of the special equipment using the DBSCAN clustering algorithm to obtain initial multi-source data of the special equipment includes: Based on the multi-source data of the special equipment, determine the neighborhood radius ε and the minimum sample parameter MinPts of the DBSCAN clustering algorithm, and mark the multi-source data of the special equipment as unaccessed. Select an unvisited data point P from the multi-source data of special equipment, define a neighborhood range with ε as the radius, and count the number of data points contained in the neighborhood. If the number of data points in the neighborhood is greater than or equal to MinPts, then data point P is determined to be a core point, and a new cluster C is created for core point P. For the identified core point P, all data points in its neighborhood that have not yet been assigned to any cluster are assigned to cluster C and marked as visited, and the initial special equipment multi-source data is obtained iteratively.

3. The special equipment safety inspection data processing method as described in claim 1, characterized in that, The step of converting the initial multi-source data of special equipment into structured feature vectors and aligning them with sensor data in the time dimension to obtain aligned multi-source data of special equipment includes: Vibration feature data are extracted from the initial multi-source data of special equipment, including at least time-domain features and frequency-domain features; The average, maximum, minimum, and rate of change of temperature are extracted from the initial multi-source data of the special equipment to obtain temperature feature data. The pressure average, maximum, minimum and fluctuation amplitude features are extracted from the initial multi-source data of special equipment to obtain pressure feature data; The average current value, effective value, peak value and frequency characteristics of the initial special equipment multi-source data are extracted to obtain current characteristic data; The vibration feature data, pressure feature data, and current feature data are combined in a unified format to obtain a structured feature vector.

4. The special equipment safety inspection data processing method as described in claim 1, characterized in that, The step of inputting the multi-source data of the aligned special equipment into the GNN-GAT hazard prediction network model for identification to obtain the equipment risk probability includes: The multi-source data of the alignment special equipment is normalized to convert the data with different features to the range of [0,1], and a normalized feature vector is obtained. The normalized feature vector is input into the trained GNN-GAT hazard prediction network model, and the features and edge weights of each node are calculated through forward propagation. It captures the temporal changes in the operating status of the equipment and the interactions between components, and outputs the risk probability of the equipment.

5. The special equipment safety inspection data processing method as described in claim 1, characterized in that, The DDPG-based deep deterministic policy gradient algorithm determines a graded response strategy based on the risk probability level of the device, including: The operating status of special equipment is regarded as the environment, which includes the risk probability of the equipment, the real-time monitoring data of each sensor, and the historical operating data of the equipment. Different graded response strategies are defined as actions, and the corresponding graded response strategies are output using the DDPG algorithm based on the real-time calculated device risk probability.

6. A special equipment safety inspection data processing system, characterized in that, The special equipment safety inspection data processing system includes the following modules: The data acquisition and clustering module is used to acquire vibration, temperature, pressure and current data in special equipment through sensors to obtain multi-source data of special equipment. The DBSCAN clustering algorithm is used to process the multi-source data of special equipment to obtain initial multi-source data of special equipment. The feature extraction and alignment module is used to convert the initial multi-source data of special equipment into structured feature vectors and align them with the sensor data in the time dimension to obtain aligned multi-source data of special equipment. The prediction model building module is used to construct a hazard prediction network based on DynamicGNN dynamic graph neural network, which includes at least a set of nodes, a set of edges, and a dynamic update mechanism; let the set of nodes represent the equipment components of special equipment, including at least bearings, gears and motors, and the node features include real-time monitoring data, historical operating data and design parameters of the equipment components; The edge set represents the connection relationships and mutual influences between equipment components, and the edge weight represents the degree of association between equipment components; The dynamic update mechanism uses a time-sliding window and a GAT graph attention mechanism to update node features and edge weights in real time to capture the temporal changes in equipment operating status. Based on the time scale of equipment fault development and the temporal resolution of the data, and by analyzing historical fault data, the sliding window size is set to be greater than the average time interval from the appearance of early potential hazards to the occurrence of a fault, thus determining the time-sliding window size. In the GAT graph attention layer, the attention weights between each node and its neighboring nodes are calculated. The importance of different components is learned through these attention weights, resulting in the GNN-GAT hazard prediction network model. The risk probability identification module is used to input the multi-source data of the aligned special equipment into the GNN-GAT hidden danger prediction network model for identification, and to obtain the equipment risk probability. The response strategy specification module is used to determine a graded response strategy based on the risk probability level of the device using the DDPG deep deterministic policy gradient algorithm.

7. The special equipment safety inspection data processing system as described in claim 6, characterized in that, The risk probability identification module includes the following sub-modules: The normalization submodule is used to normalize the multi-source data of the alignment special equipment, transforming data with different features into the range of [0,1] to obtain a normalized feature vector; The calculation submodule is used to input the normalized feature vector into the trained GNN-GAT hazard prediction network model and calculate the features and edge weights of each node through forward propagation. The capture submodule is used to capture the temporal changes in the device's operating status and the interactions between components, and output the device's risk probability.

8. The special equipment safety inspection data processing system as described in claim 6, characterized in that, The response strategy specification module includes the following sub-modules: The setting submodule is used to treat the operating status of special equipment as the environment. The status includes the risk probability of the equipment, the real-time monitoring data of each sensor, and the historical operating data of the equipment. Define a submodule to define different graded response strategies as actions. Based on the real-time calculated device risk probability, use the DDPG algorithm to output the corresponding graded response strategy.

Images