Superheater pipe explosion dynamic early warning system and method based on multi-mode fusion

Abstract

The invention discloses a superheater pipe burst dynamic early warning system and method based on multi-modal fusion, and belongs to the technical field of thermal power plant safety monitoring, the system comprises an acquisition module, an edge calculation module, a cloud analysis module and a visual interaction module, the acquisition module obtains multi-source heterogeneous sensing signals in each processing period, and the edge calculation module obtains the multi-source heterogeneous sensing signals in each processing period; the edge calculation module processes collected signals into key feature vectors, the time sequence dependency of the key feature vectors is learned through a preliminary evaluation model to output a preliminary abnormal probability, local alarm is achieved, and the key feature vectors are packaged into data packets to be uploaded at the cloud end. And the physical topological structure of the superheater pipeline and the uploaded data packet are fused for space-time correlation analysis and trend prediction, so that high-real-time analysis and high-precision early warning under edge-cloud collaboration are realized.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for thermal power plants, specifically to a dynamic early warning system and method for superheater tube rupture based on multimodal fusion. Background Technology

[0002] The superheater of a thermal power plant boiler is a core component that ensures the safe and stable operation of the unit. It is necessary to monitor the tube wall temperature field, stress and strain, internal microcracks and other dimensions in real time to detect the hidden danger of tube rupture. Distributed fiber optic temperature measurement, high temperature strain monitoring and acoustic emission sensor capture are key technologies for monitoring the superheater status, respectively undertaking the tasks of detecting the overall temperature field of the tube wall, structural deformation and material micro-damage.

[0003] With the growth in electricity demand, thermal power units face more complex operating conditions. Superheaters experience more severe temperature and pressure fluctuations, and the damage accumulation process of tube wall materials is more complex. This leads to a further increase in the demand for real-time and intelligent monitoring systems. Existing superheater monitoring technologies (such as temperature measurement, strain measurement, and acoustic emission) often analyze data independently or simply overlay it, relying on human experience or preset fixed rules for comprehensive judgment. They cannot make intelligent decisions based on the current operating conditions of the unit and the dynamic correlation of multi-source data. In addition, the alarm thresholds of early warning systems often use fixed values ​​or rely on human experience to set, which not only increases the calibration burden of operation and maintenance personnel, but also causes the system to frequently malfunction under variable loads due to the mismatch between the thresholds and the actual operating conditions, resulting in low reliability of the early warning results.

[0004] In summary, existing technologies have not effectively solved key challenges such as sensor high-temperature tolerance, multi-source data collaborative analysis, and real-time decision-making at the edge, resulting in insufficient accuracy of early warning of pipe bursts and failing to provide maintenance personnel with enough response time to take preventive measures. Therefore, there is an urgent need in this field for an intelligent early warning system that can operate stably in high-temperature and harsh environments, has deep multi-modal data fusion capabilities, and strong real-time performance. Summary of the Invention

[0005] This invention provides a dynamic early warning system and method for superheater tube rupture based on multimodal fusion. It constructs an edge-cloud collaborative computing architecture and deeply integrates the physical topology of the equipment with multimodal sensor data in the cloud. This solves the problems of existing technologies such as single monitoring dimension, poor real-time performance, high false alarm rate and weak adaptability to operating conditions, and achieves the goal of early and accurate early warning of tube rupture faults.

[0006] The technical solution to achieve the objective of this invention is as follows:

[0007] In a first aspect, the present invention provides a dynamic early warning system for superheater tube rupture based on multimodal fusion, comprising:

[0008] The acquisition module acquires the acquisition signals for each cycle, including acoustic emission signals, fiber optic temperature signals, strain signals, and vibration signals.

[0009] The edge computing module processes the acquired signals of each cycle into corresponding key feature vectors, puts them into the rolling update time-series feature queue, uses a preliminary evaluation model to learn the data time-series dependency of the updated time-series feature queue, obtains the preliminary anomaly probability of each cycle, compares the preliminary anomaly probability with the local alarm threshold to decide whether to immediately repair or package the key feature vectors and preliminary anomaly probabilities into a data package and upload it to the cloud analysis module.

[0010] The cloud-based analysis module receives data packets in each cycle and constructs them into a dynamic attribute graph to obtain the real-time operating conditions of the unit. It uses a graph neural network to consider the physical topology of the superheater to perform spatiotemporal correlation analysis on the dynamic attribute graph. Combined with the dynamic adjustment and early warning logic based on the operating conditions, it obtains the risk assessment results for each cycle.

[0011] The visualization and interaction module retrieves and displays the risk assessment results for each period.

[0012] Furthermore, the preliminary evaluation model processes the updated queue data through a long short-term memory network, outputs a hidden state vector, and uses a fully connected layer to map the hidden state vector into a logical value and convert it into a preliminary anomaly probability.

[0013] Furthermore, the graph neural network is a spatiotemporal graph attention network model. The specific method for risk assessment in each cycle is as follows: a spatial attention mechanism is used to calculate and aggregate the spatial association information of each node in the dynamic attribute graph for each cycle, calculate the attention weight of each node in the dynamic attribute graph, and aggregate the features of neighboring nodes in a weighted manner to generate node representations that integrate spatial association information. A temporal attention mechanism is used to sum the weighted representations of historical nodes to update the node state in the current cycle. Based on the updated node state, the future risk probability sequence of each node is output, which is then input into the classifier to determine and output the risk assessment results that include multi-level risk classification and specific pipeline location.

[0014] Furthermore, the edge computing module includes a feature extraction unit, which uses wavelet transform to extract frequency domain features from the acoustic emission signal, uses median filter and bilinear interpolation to reconstruct the one-dimensional temperature data of the fiber optic temperature signal into a two-dimensional temperature field image, and uses the isolated forest algorithm to identify and extract the spatial geometric features of the abnormal hot zone.

[0015] Furthermore, the cloud-based analysis module includes an adaptive threshold adjustment unit, which acquires multi-dimensional operating condition parameters such as unit load, operating status, and coal type changes. It uses an operating condition-feature baseline mapping model to predict the sensor feature baseline range, combines it with the floating tolerance range related to the operating condition, and constructs a dynamic value to determine the alarm threshold of the risk assessment results.

[0016] Furthermore, the edge computing module includes a preliminary diagnostic unit that performs tiered responses, compares the preliminary anomaly probability with the local alarm threshold, and determines that there is a high-confidence risk of pipe bursting if the preliminary anomaly probability is equal to or higher than the threshold, generates an alarm event and prompts immediate repair; if it is lower than the threshold, the data packet is uploaded to the cloud analysis module for further spatiotemporal correlation analysis.

[0017] Furthermore, the acquisition module includes a data synchronization unit that uses a network time protocol to maintain microsecond-level synchronization with the master clock server, acquiring a high-precision timestamp in each system cycle to eliminate timing errors between different sensing systems.

[0018] Furthermore, the cloud analytics module includes a historical data management unit that stores historical sensor data and maintenance records, supports the optimization of graph neural networks and adaptive threshold adjustment units, and provides historical data for fault tracing.

[0019] Furthermore, the risk assessment results are displayed in the form of a heat map to locate high-risk pipe sections, and maintenance priorities are generated and recommended based on maintenance records.

[0020] Secondly, the present invention provides a dynamic early warning method for superheater tube rupture based on multimodal fusion, comprising the following steps:

[0021] Acquire the acquisition signals for each cycle, including acoustic emission signals, fiber optic temperature signals, strain and vibration signals;

[0022] The acquired signal of each cycle is processed into a corresponding key feature vector and placed into a rolling update time series feature queue. The preliminary evaluation model is used to learn the data time series dependency of the updated time series feature queue to obtain the preliminary anomaly probability of each cycle. The preliminary anomaly probability is compared with the local alarm threshold to decide whether to immediately repair or package the key feature vector and preliminary anomaly probability into a data package and upload it to the cloud analysis module.

[0023] Data packets are received and constructed into a dynamic attribute graph to obtain the real-time operating conditions of the unit. A graph neural network is used to perform spatiotemporal correlation analysis on the dynamic attribute graph, taking into account the physical topology of the superheater. Combined with the operating condition dynamic adjustment early warning logic, the risk assessment results for each cycle are obtained.

[0024] Obtain and display the risk assessment results for each period.

[0025] Compared with the prior art, the advantages of this invention are as follows:

[0026] 1. The new edge-cloud collaborative computing architecture enables real-time early warning response. This invention solves the data transmission latency bottleneck caused by the traditional cloud-based centralized processing architecture by deploying edge computing modules at the data source for front-end intelligent processing, thus ensuring the timeliness of fault response and early warning.

[0027] 2. The deep analysis model integrating physical topology improves the accuracy and adaptability of early warning. The graph neural network of this invention incorporates the physical structure of the superheater pipe into the analysis. Combined with the adaptive threshold adjustment unit, it solves the problems of independent analysis of multi-source data, inability to perform deep spatial correlation, and inability of fixed thresholds to adapt to changing operating conditions, thereby improving the accuracy of early warning and reducing the false alarm rate. Attached Figure Description

[0028] Figure 1 This is a structural framework diagram of a superheater tube rupture dynamic early warning system based on multimodal fusion.

[0029] Figure 2 A schematic diagram of the multimodal data acquisition deployment;

[0030] Figure 3 This is a flowchart of the intelligent computing module.

[0031] Meaning of the reference numerals in the attached figures:

[0032] 1. Superheater tube wall; 2. High-temperature resistant piezoelectric ceramic acoustic emission sensor; 3. Distributed fiber optic temperature sensor; 4. High-temperature strain / vibration sensor. Detailed Implementation

[0033] To illustrate the workflow of each module of this invention, a specific application scenario is used as an example. Taking a 660MW boiler superheater in a power plant as the background, the tube bank of the boiler's last stage superheater is subjected to high-temperature steam and high-concentration flue gas at approximately 605°C for a long time. The invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0034] Example 1:

[0035] like Figure 1 As shown, this invention discloses a dynamic early warning system for superheater tube rupture based on multimodal fusion, including a data acquisition module, an edge computing module, a cloud analysis module, and a visualization interaction module;

[0036] During each system processing cycle, the acquisition module obtains multi-dimensional and heterogeneous raw sensor signals reflecting the tube wall health status from the final superheater of the 660MW boiler. Through the built-in data synchronization unit, it performs high-precision timestamp calibration on all signal streams to ensure absolute alignment of data in the time dimension, providing high-quality and synchronized data input for subsequent fusion analysis.

[0037] The edge computing module is deployed in the boiler instrument room to acquire the synchronous data stream output by the acquisition module. In each system processing cycle, it uses the built-in high-performance AI chip to perform real-time and intelligent processing on massive amounts of raw data, including key feature extraction and quantization, signal noise reduction and filtering, generating key feature vectors and placing them into a rolling update time-series feature queue. Based on a lightweight time-series model, it learns the time-series dependency of the time-series feature queue data, makes a preliminary judgment on single-point anomaly trends, and performs hierarchical processing based on the judgment. If the probability exceeds the alarm threshold, it is judged that the tube has burst and is dealt with immediately. If it is below the threshold, the key feature vector and the preliminary anomaly probability are packaged into a data packet and uploaded to the cloud analysis module.

[0038] The cloud-based analysis module runs on the central server cluster in the power plant's main control building. It receives key feature datasets uploaded from the edge computing module, builds and runs a deep learning model that integrates the physical topology of the superheater, and performs global and deep spatiotemporal correlation analysis on multimodal features. The module can also dynamically adjust its early warning logic based on the real-time operating conditions of the boiler transmitted from the distributed control system, and finally outputs the risk level of tube rupture and the risk trend prediction for the next 24 hours, accurate to the specific tube bank and tube number.

[0039] The visualization and interaction module is deployed on the engineer station in the power plant's central control room. It receives the early warning results output by the cloud analysis module. Based on the three-dimensional digital twin model of the superheater equipment, it renders and presents the abstract risk data in real time in the form of a dynamic and intuitive three-dimensional heat map. The module also combines the knowledge of the historical data management unit to provide scientific maintenance priority recommendations for operation and maintenance personnel, and transforms early warning information into operation and maintenance decisions that can directly guide actions.

[0040] Furthermore, the acquisition module includes a sensing unit and a data synchronization unit;

[0041] like Figure 2As shown, the sensing unit uses acoustic emission sensors 2 to capture high-frequency transient elastic waves released when micro-cracks occur in the pipe wall material under stress or when small leaks occur in high-temperature steam. Multiple high-temperature resistant piezoelectric ceramic acoustic emission sensors 2 are spot-welded to both sides of key welds near the superheater tube wall 1, the roof tube bank, and the outlet header via specially designed stainless steel waveguides, forming a monitoring array covering key areas. The piezoelectric crystals built into the sensors have an upper operating temperature limit of up to 700℃ and a frequency response of 50kHz~900kHz. Their weak output charge signals are connected to a data acquisition system deployed in a local instrument room via a low-noise coaxial cable for synchronous high-speed acquisition of multi-channel acoustic emission signals. The distributed fiber optic temperature sensor 3 uses a specially packaged special sensing fiber that can withstand temperatures up to 800℃. The fiber is tightly laid in an S-shaped reciprocating path on the outer wall of the final stage superheater roof tube bank and fixed with special high-temperature resistant clips to ensure... With good heat conduction with the pipe wall, a continuous and seamless temperature measurement network is formed. The DTS host connected to it emits nanosecond-level laser pulses and receives Raman backscattered light caused by temperature effects in the optical fiber. Using the OTDR principle, it demodulates and calculates the precise temperature of each spatial resolution point on the optical fiber at 0.5m intervals, thereby obtaining the complete and continuous temperature field distribution on the pipe surface in real time. The high-temperature strain / vibration sensor 4 adopts a high-temperature resistant strain gauge based on FBG and is welded to both ends of the outlet header of the final stage superheater and its main support rods, which are subjected to thermal stress and gravity loads. It responds to changes in temperature and strain by the drift of the center wavelength of its grid area. In order to eliminate temperature cross-sensitivity, each measuring point uses a pair of strain FBG and temperature compensation FBG. The sensors are connected in series through armored optical cables to monitor in real time the long-term plastic deformation trend of the pipeline caused by high-temperature creep, the alternating stress caused by start-up, shutdown or load fluctuation, and the structural vibration caused by unstable combustion or abnormal steam flow.

[0042] The data synchronization unit provides a unified, high-precision time reference for all heterogeneous and asynchronous data acquisition devices. Physically integrated into the data access port of the edge computing module, it runs a high-precision clock synchronization service in software. As an NTP client, this unit maintains microsecond-level synchronization with the power plant's main clock server to obtain a unified UTC. When any sensor generates a data packet, whether it is an acoustic emission signal with a sampling rate of MHz or a temperature signal with a sampling rate of Hz, the system immediately adds a high-precision UTC timestamp based on this synchronization clock. Through this mechanism, all signals used for subsequent analysis, regardless of their source or original sampling rate, are precisely aligned on the time reference. This effectively eliminates timing errors caused by factors such as clock differences within different sensing systems, different data transmission paths, and hardware processing delays, providing a reliable time guarantee for the effective fusion and analysis of multimodal data.

[0043] Furthermore, the edge computing module includes a feature extraction unit and a preliminary diagnosis unit;

[0044] The feature extraction unit adaptively executes the optimal feature extraction algorithm based on the physical characteristics of different signals. Specifically, it includes four main functions: AE frequency domain analysis, DTS anomaly identification, strain trend assessment, and data denoising and compression. For the acquired multi-channel acoustic emission raw waveform data, the AE frequency domain analysis first filters out the main noise frequency bands using an adaptive bandpass filter. The passband range is set to 100kHz~600kHz, filtering out low-frequency noise generated by boiler combustion and high-frequency electromagnetic interference in the field environment. Then, WPT is applied to perform a 5-level decomposition on the denoised signal. This process decomposes the original signal into 32 sub-bands of equal energy. This unit focuses on calculating the energy of these sub-bands within this range. And energy entropy, which characterizes the complexity and disorder of a signal. The calculation formula is as follows:

[0045] , ,

[0046] in, For the first Discrete signal sequences after wavelet packet decomposition of each sub-band. For the first Signal energy of each frequency band For the first The energy percentage of discrete signals within each sub-band. For the first The signal energy entropy of each sub-band, along with these two frequency domain features, are extremely sensitive to identifying microscopic damage events within the material. Ultimately, each identified acoustic emission event is transformed into a 16-dimensional frequency domain feature vector containing 8 energy values ​​and 8 energy entropy values. For the identified acoustic emission events, the system also utilizes the time difference of arrival of signals received by multiple sensors to calculate the three-dimensional spatial coordinates of the sound source using the Chan-Ho algorithm, ultimately generating a 16-dimensional frequency domain feature vector. An additional three-dimensional coordinate is added, giving it spatial attributes. DTS anomaly detection takes a frame of one-dimensional temperature data containing 1600 temperature measurement points, smooths the original temperature sequence using a sliding window mid-range filter, removes random jump noise from individual temperature measurement points, and then reconstructs the smoothed temperature sequence into a 256x256 pixel two-dimensional temperature field grayscale image using bilinear interpolation based on the actual laying coordinates of the optical fiber on the roof tube bank of the final-stage superheater. Then, the isolated forest algorithm was used to identify anomalies in the temperature map. In the offline phase, numerous isolated trees were trained using temperature maps from a large number of boilers operating under normal conditions. During online inference, Each pixel in the model is treated as a data point and input into a pre-trained model. The model calculates an anomaly score s(p, n) for each pixel. A score close to 1 indicates that the pixel is highly likely to be an anomaly, while a score much less than 0.5 indicates that it is a normal pixel. Finally, pixels with scores higher than a preset threshold are clustered to form anomaly regions, and features such as the total area, highest temperature, average temperature, and centroid location of these regions are calculated to form an 8-dimensional spatial anomaly feature vector. The centroid coordinates of the abnormal region are clearly defined, directly corresponding to the spatial location of the anomaly on the physical unfolded diagram of the superheater tube bank. Strain trend assessment evaluates the long-term creep trend and short-term abnormal fluctuations based on the acquired multi-channel strain data. First, a low-pass filter is applied to separate the quasi-static strain component from the original strain signal, reflecting the pipe's high-temperature creep and large load changes. The slope and residual sum of squares over the past 24 hours are calculated using linear regression. Simultaneously, the quasi-static component is subtracted from the original signal to obtain the dynamic strain component. Its spectrum is calculated using FFT to extract the main vibration frequencies and amplitudes, used to determine the presence of resonance risks caused by unstable combustion or abnormal steam flow. Finally, each strain measurement point is transformed into a 4-dimensional trend feature vector containing strain rate, nonlinearity, main vibration frequency, and amplitude. After all the features have been extracted, the data noise reduction and compression function normalizes all modal feature vectors and scales the feature values ​​to the range to eliminate the influence of dimensions. Then, it performs efficient binary serialization through Google's Protocol Buffers protocol, packing the structured feature data into a compact message format. Compared with the original waveform or full temperature data, the overall data volume is compressed to one-thousandth of the original data.

[0047] Based on feature extraction, the preliminary diagnostic unit deploys a lightweight preliminary evaluation model for each single-modality feature data stream. This model does not simply make predictions, but rather quantifies the degree of anomaly in the current data pattern. To achieve unified processing of data streams with different properties, key feature vectors are placed into a rolling update time-series feature queue. Specifically, this unit maintains an independent rolling update time-series feature queue for each feature vector stream. This queue is implemented as a first-in-first-out (FIFO) cache structure, and its data filling mechanism varies depending on the nature of the data source. For periodic data such as DTS and strain, a new feature vector is added at fixed intervals. For event-driven data such as acoustic emissions, a new feature vector is added to the queue only when a valid event is detected. Regardless of the situation, the queue maintains a constant length by popping the oldest data. After a rolling update, a preliminary evaluation model learns the temporal dependencies of the updated queue data. This involves using the complete data of the queue as input to the preliminary evaluation model. The specific implementation structure of this preliminary evaluation model is as follows: First, an LSTM model with two LSTM layers, each with 32 hidden units, processes the input time-series tensor. Then, the LSTM network outputs the hidden state vector of its last time step. Then the hidden state vector It is fed into a single fully connected layer, and The 32-dimensional linear transformation is converted into a logit. Finally, this logistic value undergoes a non-linear mapping using a Sigmoid activation function, ultimately outputting an initial anomaly probability ranging from 0 to 1. Its calculation process can be expressed as follows:

[0048]

[0049]

[0050]

[0051] in, Let t be the time series tensor input to the LSTM model. This is the 32-dimensional hidden state vector output by the LSTM network at the last time step. The learnable linear transformation weight matrix of the fully connected layer. For the bias term of the fully connected layer, The logical value output by the fully connected layer. Let be the initial anomaly probability at time t, with a value ranging from [0,1]. This probability rigorously characterizes the likelihood that the data at the current time mode exhibits a significant anomaly pattern. Based on this initial anomaly probability, the system implements a hierarchical response mechanism at the edge. The system presets a local alarm threshold. In each calculation cycle, the preliminary diagnosis unit will make the following judgment: If the initial anomaly probability is less than the local alarm threshold, it indicates that there are certain abnormal signs in the current data pattern, but it has not yet reached the severity level requiring immediate on-site intervention, or more dimensions of information are needed for corroboration. At this time, the edge computing module will use the key feature vector at that time along with its corresponding initial anomaly probability as a focus-level data packet. This data packet includes a timestamp, sensor ID, and other information. Information such as key feature vectors and preliminary anomaly probabilities is uploaded to the cloud analysis module for deeper spatiotemporal correlation analysis. If the preliminary anomaly probability is greater than or equal to the local alarm threshold, it indicates that the current data pattern has an extremely significant and high-confidence anomaly, indicating an impending serious failure. The edge computing module determines that there is a high-confidence risk of pipe bursting, triggers a local alarm, and immediately hands it over to the field for handling. It also uploads the data to the cloud analysis module for recording, archiving, and in-depth source tracing analysis. Through this hierarchical response mechanism, the functional positioning of the edge computing module becomes clear and efficient. It can provide the fastest local response to serious alarms and entrust uncertain events that require comprehensive analysis to the cloud brain for processing, truly realizing the value of edge-cloud collaboration.

[0052] Furthermore, the cloud analytics module includes a data fusion unit, an adaptive threshold adjustment unit, and a historical data management unit;

[0053] The core of the data fusion unit is to run a unique fusion analysis model based on a spatiotemporal graph attention network, which is an innovation of this invention. The implementation of this model includes two parts: graph construction and model operation. Graph construction is based on the design drawings and 3D design model of the superheater, abstracting the precise spatial coordinates of 32 sensor nodes, and constructing a weighted undirected graph G=(V, E, W), where the node set V represents the 32 sensor nodes. If the physical distance between two nodes on the pipe is less than 5m, or they are located on the same pipe segment or pipe segments directly connected by a header, then an edge e (e∈E) is established between these two nodes, with the edge weight w... w∈W) can be set as the product of the reciprocal of the physical distance and the thermal conductivity of the pipe at that temperature to characterize the strength of the physical influence between nodes. The constructed graph structure is persistently stored in the Neo4j graph database for easy querying and version management. The model's working mechanism is as follows: ST-GAT consists of four stacked spatiotemporal attention layers and a final output layer. Each layer contains a spatial attention module and a temporal attention module. At each time step, the spatial attention module calculates the attention coefficient of each neighbor node to the central node through a self-attention mechanism, realizing the weighted aggregation of spatial association information. For each node The neighbors are calculated using attention logits, normalization, and a weighted aggregation formula. Attention coefficient And aggregate information:

[0054]

[0055] in, For time steps Next node Its neighboring nodes The attention logits value, For ReLU activation function with leakage correction, The weight vector of a single-layer feedforward network transpose, The weight matrix is ​​a learnable linear transformation. , They are time steps Next node ,node The hidden state vector, This is a vector concatenation operator, representing the concatenation of vectors. and Dimensionally stitched;

[0056]

[0057] in For time steps Next node Its neighboring nodes The normalized attention coefficient, For nodes All neighboring nodes Perform softmax normalization. For nodes The set of all neighboring nodes, It is a natural exponential function. For nodes All neighboring nodes of Summation, For nodes For any neighbor The original attention score;

[0058]

[0059] in, For time steps Next, node The new hidden state vector output after spatial attention weighted aggregation It is the activation function of the exponential linear unit. For nodes All neighboring nodes Summation, In time step Next, node Its neighboring nodes The normalized attention coefficient, The weight matrix is ​​a learnable linear transformation. In time step Next, neighboring nodes The hidden state vector is obtained through a time attention module, which employs a gated recurrent unit structure with an attention mechanism. This module integrates the node representations obtained in the previous step, which incorporate spatial information. As the current input, examine the node's past. The model calculates the importance weight of each historical moment in predicting the current state, based on the feature sequence of each historical time step, to accurately capture the evolutionary pattern over time. After propagation through all spatiotemporal attention layers, each node obtains a final representation vector that deeply integrates spatiotemporal information. This vector is then fed into a prediction head composed of fully connected layers to decode and predict the risk probability prediction sequence of the node in the next 24 time steps. Finally, this prediction sequence is fed into a decision head, which first extracts key indicators such as the highest risk probability, average value, and growth slope of the sequence. These indicators are then input into a Softmax classifier, which determines and provides a comprehensive risk assessment of four levels: normal, attention, alarm, and severe, as well as a precise location result down to the specific pipe segment. The entire model is trained end-to-end using the AdamW optimizer, with a cross-entropy loss function with class weights to address the imbalance problem where there are far more normal samples than faulty samples.

[0060] The adaptive threshold adjustment unit obtains 15 operating condition parameters, including unit load, main steam pressure, main steam temperature, and coal type code, through the power plant's real-time database interface. Internally, this unit pre-builds a gradient boosting decision tree (GBDT) operating condition feature baseline mapping model based on Python's XGBoost library through offline learning. During online execution, this model uses the current 15-dimensional real-time operating condition parameter vector... As input, the normal baseline range of each sensor characteristic under this operating condition is accurately predicted. The final alarm threshold used by the system is a dynamic value consisting of "the predicted normal baseline + a floating tolerance range related to the operating condition," calculated as follows:

[0061]

[0062] in, for The dynamic alarm threshold of the time system for The mean value of the normal baseline of sensor characteristics predicted by the baseline mapping model of operating conditions at any given time. This is the sensitivity coefficient. for The standard deviation of the normal baseline of sensor characteristics predicted by the baseline mapping model of operating conditions at any given time. It constitutes a floating tolerance range related to operating conditions, which is used to adapt to the inherent fluctuation characteristics of the characteristic baseline under different operating conditions;

[0063] The historical data management unit adopts a hybrid database architecture to meet the storage and query needs of different types of data. It uses a distributed time-series database, clustered and deployed on three servers, to store all sensor feature data with high-precision timestamps uploaded from the edge. A graph database is used to store the structural map model of the superheater, the relationships between sensor nodes, and the propagation paths and correlations of historical early warning events, providing strong support for fault tracing analysis. A relational database is used to store manually entered equipment ledgers, material information, and detailed historical maintenance records. This architecture not only supports continuous online optimization of the ST-GAT model and the adaptive threshold model, but also provides maintenance personnel with powerful, multi-dimensional fault tracing analysis support when an early warning occurs.

[0064] Furthermore, the visualization interaction module, based on the Unity3D engine, constructs a high-fidelity 3D digital twin model of the superheater equipment. When a cloud-based early warning result is received, the module matches the ID number of the high-risk pipe section with the 3D object in the twin model, and dynamically renders the color and glowing effects of the pipe section according to the risk level, forming a striking and intuitive 3D heat map. When the user clicks on the highlighted pipe section, an information box pops up on the interface, displaying detailed feature data and risk evolution curves, and automatically retrieves relevant historical maintenance records from the historical data management unit. Finally, by integrating all the information, a scientific maintenance priority recommendation is generated.

[0065] Example 2:

[0066] Reference Figure 3 The present invention also discloses a method for early warning of superheater tube rupture based on the above system. Figure 3 The overall workflow of the intelligent computing module shown includes the following steps:

[0067] S1: During each system cycle, the acquisition module synchronously acquires acoustic emission, temperature, strain, and vibration signals;

[0068] S2: The collected signal is sent to the edge computing module, where the signal is processed into a key feature vector by the feature extraction unit. The vector is then placed into a rolling update queue. By initially evaluating the temporal dependency of the data in the queue after the model learns the update, the initial anomaly probability of the period is obtained, and the key feature vector and the corresponding initial anomaly probability are packaged into a data packet.

[0069] S3: The edge computing module performs a tiered response. If the alarm value is higher than the local alarm threshold, it is determined to be a pipe burst and is handled immediately. If the alarm value is lower than the alarm threshold, the data packet is uploaded to the cloud analysis module for further analysis.

[0070] S4: After receiving the data packet, the cloud analysis module uses a graph neural network to perform deep fusion analysis and prediction, while the adaptive threshold adjustment unit corrects the model's warning logic according to the real-time operating conditions.

[0071] S5: The cloud-based analysis module outputs the final warning level, future risk trends, and precise spatial coordinates of high-risk sections for this period, forming a risk assessment result;

[0072] S6: The risk assessment results are sent to the visualization and interaction module, where they are displayed as a heat map on the three-dimensional digital twin model, along with intelligent maintenance suggestions.

[0073] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic early warning system for superheater tube rupture based on multimodal fusion, characterized in that, include: The acquisition module acquires the acquisition signals for each cycle, including acoustic emission signals, fiber optic temperature signals, strain signals, and vibration signals. The edge computing module processes the acquired signals of each cycle into corresponding key feature vectors, puts them into the rolling update time-series feature queue, uses a preliminary evaluation model to learn the data time-series dependency of the updated time-series feature queue, obtains the preliminary anomaly probability of each cycle, compares the preliminary anomaly probability with the local alarm threshold to decide whether to immediately repair or package the key feature vectors and preliminary anomaly probabilities into a data package and upload it to the cloud analysis module. The cloud-based analysis module receives data packets in each cycle and constructs them into a dynamic attribute graph to obtain the real-time operating conditions of the unit. It uses a graph neural network to consider the physical topology of the superheater to perform spatiotemporal correlation analysis on the dynamic attribute graph. Combined with the dynamic adjustment and early warning logic based on the operating conditions, it obtains the risk assessment results for each cycle. The visualization and interaction module retrieves and displays the risk assessment results for each period.

2. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The initial evaluation model processes the updated queue data through a long short-term memory network, outputs a hidden state vector, and uses a fully connected layer to map the hidden state vector into a logical value and convert it into a preliminary anomaly probability.

3. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The graph neural network is a spatiotemporal graph attention network model. The specific method for risk assessment in each cycle is as follows: a spatial attention mechanism is used to calculate and aggregate the spatial correlation information of each node in the dynamic attribute graph for each cycle, calculate the attention weight of each node in the dynamic attribute graph, and aggregate the features of neighboring nodes in a weighted manner to generate a node representation that integrates spatial correlation information. A temporal attention mechanism is used to sum the weighted representations of historical nodes to update the node state in the current cycle. Based on the updated node state, the future risk probability sequence of each node is output, which is then input into a classifier to determine and output the risk assessment result that includes multi-level risk classification and specific pipeline location.

4. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The edge computing module includes a feature extraction unit, which uses wavelet transform to extract frequency domain features from acoustic emission signals, uses median filter and bilinear interpolation to reconstruct one-dimensional temperature data of fiber optic temperature signals into two-dimensional temperature field images, and uses the isolated forest algorithm to identify and extract the spatial geometric features of anomalous hot areas.

5. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The cloud-based analysis module includes an adaptive threshold adjustment unit, which acquires multi-dimensional operating parameters such as unit load, operating status, and coal type changes. It uses an operating condition-feature baseline mapping model to predict the range of sensor feature baselines, combines the floating tolerance range related to the operating condition, and forms a dynamic value to determine the alarm threshold of the risk assessment results.

6. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The edge computing module includes a preliminary diagnosis unit, which performs a hierarchical response, compares the preliminary anomaly probability with the local alarm threshold, and if the preliminary anomaly probability is equal to or higher than the threshold, it determines that there is a high-confidence risk of pipe bursting, generates an alarm event, and prompts immediate repair. If the data packet is below the threshold, it will be uploaded to the cloud analysis module for further spatiotemporal correlation analysis.

7. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The acquisition module includes a data synchronization unit that uses a network time protocol to maintain microsecond-level synchronization with the master clock server and acquires a high-precision timestamp in each system cycle to eliminate timing errors between different sensing systems.

8. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The cloud analytics module includes a historical data management unit that stores historical sensor data and maintenance records, supports the optimization of graph neural networks and adaptive threshold adjustment units, and provides historical data for fault tracing.

9. The superheater tube rupture dynamic early warning system based on multimodal fusion according to claim 1, characterized in that, The risk assessment results are displayed in the form of a heat map to locate high-risk pipe sections. Combined with maintenance records, maintenance priorities are generated and recommended.

10. A dynamic early warning method for superheater tube rupture based on multimodal fusion, executed by the dynamic early warning system for superheater tube rupture based on multimodal fusion as described in any one of claims 1-9, characterized in that, Includes the following steps: Acquire the acquisition signals for each cycle, including acoustic emission signals, fiber optic temperature signals, strain and vibration signals; The acquired signal of each cycle is processed into a corresponding key feature vector and placed into a rolling update time series feature queue. The preliminary evaluation model is used to learn the data time series dependency of the updated time series feature queue to obtain the preliminary anomaly probability of each cycle. The preliminary anomaly probability is compared with the local alarm threshold to decide whether to immediately repair or package the key feature vector and preliminary anomaly probability into a data package and upload it to the cloud analysis module. Data packets are received and constructed into a dynamic attribute graph to obtain the real-time operating conditions of the unit. A graph neural network is used to perform spatiotemporal correlation analysis on the dynamic attribute graph, taking into account the physical topology of the superheater. Combined with the operating condition dynamic adjustment early warning logic, the risk assessment results for each cycle are obtained. Obtain and display the risk assessment results for each period.

Images