Intelligent monitoring system for civil construction equipment of thermal power plant

Abstract

The invention relates to the technical field of electric digital data processing, in particular to an intelligent monitoring system for civil construction equipment of a thermal power plant, which comprises a digital acquisition module, a digital processing module, a digital analysis module, an anomaly identification module and an optimization module. The digital acquisition module deploys a sensor at a key position of construction equipment, executes end-side noise filtering and feature extraction, and outputs a standardized feature vector. The digital processing module constructs a digital twinborn body, calibrates sensor coordinate deviation and timestamp deviation through multi-body dynamics simulation, and generates a multi-modal synchronous data stream. And the optimization module generates an adversarial sample by using the false alarm record, updates model parameters through federated learning and deploys the model parameters to an edge end. According to the method, through edge preprocessing, twin calibration, dynamic fusion and closed-loop optimization, the problem of real-time fusion of heterogeneous sensor data is solved, and the real-time performance and accuracy of abnormal state recognition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric digital data processing, in particular to an intelligent monitoring system for civil construction equipment in a thermal power plant. Background Art

[0002] Civil construction equipment for thermal power plants is used to complete the engineering tasks of the plant foundation, main structure, and ancillary buildings, covering the earthwork, basic construction, and installation phases. The project begins with earthmoving machinery such as bulldozers and excavators clearing the surface and excavating the foundation pit, forming a stable working surface to support subsequent construction. Concrete equipment includes mixing stations and pump trucks to prepare and transport the mix, and cooperates with vibrators to fill the formwork and complete the foundation pouring and curing process. Lifting equipment such as tower cranes and crawler cranes lift pre-assembled boiler steel structures, chimney segments, and plant trusses to improve installation accuracy and construction efficiency. The entire process is managed through a quality control system and safety regulations to reduce the risk of project delays.

[0003] When the intelligent monitoring system for civil construction equipment in thermal power plants is used to analyze data from earth-moving machinery, concrete equipment, and lifting devices in real time, there is a technical pain point that makes data fusion difficult due to the heterogeneity of sensor information. That is, when multi-source sensor data (such as vibration, temperature, and position signals) is input, the system cannot efficiently normalize and process data in different formats and time series, affecting the accuracy of fault prediction. For example, during crane hoisting operations, the accelerometer and visual sensor data are asynchronous or have conflicting formats, which prevents the monitoring model from identifying structural stress anomalies and generating early warnings, leading to potential equipment damage and construction delays. This is because the computing architecture is difficult to adapt to the data structure changes in the dynamic construction environment, and requires optimization through advanced algorithms to improve decision-making reliability. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an intelligent monitoring system for civil construction equipment in a thermal power plant to solve the problem of delayed abnormal state recognition caused by the failure of real-time fusion of heterogeneous sensor data of construction equipment in a dynamic operating environment.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: The intelligent monitoring system for civil construction equipment of a thermal power plant provided by the present invention comprises: The digital acquisition module deploys sensors on earthmoving machinery, concrete equipment, and lifting equipment to collect raw vibration, posture, and temperature data in real time. It then performs on-device noise filtering and feature extraction, and outputs standardized feature vectors to the digital processing module. a digital processing module that receives the standardized feature vectors, constructs a digital twin of the construction equipment, calibrates the sensor coordinate offset and timestamp deviation using multi-body dynamics simulation, generates a multimodal synchronous data stream, and transmits it to the digital analysis module; A digital analysis module parses the multimodal synchronous data stream, uses a pre-trained working condition classifier to identify load and environmental parameters, dynamically assigns sensor feature weight coefficients, generates a fused feature tensor, and provides it to the anomaly recognition module; The anomaly recognition module receives the fused feature tensor, mines the fault sequence chain through the causal graph, outputs an early warning signal to the monitoring terminal when an abnormal causal pattern is detected, and simultaneously sends a false alarm record to the optimization module; The optimization module receives the warning signal and false alarm record, generates an adversarial sample set to update the recognition model parameters, and deploys the optimized model parameters to the digital analysis module and the anomaly recognition module.

[0006] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital acquisition module includes: The vibration sensor embeds a wavelet packet decomposition chip to separate the effective spectrum from the impact noise, and outputs the spectrum data to the lightweight model on the device side. The LiDAR unit is configured with SLAM positioning firmware to generate a point cloud of the device's pose and provide the point cloud data to the lightweight model on the device side. Infrared thermal imager with built-in temperature field reconstruction algorithm outputs heat distribution matrix and transmits the heat distribution matrix to the lightweight model on the terminal side; The lightweight model on the end receives the spectrum data, point cloud data, and heat distribution matrix, performs millisecond-level anomaly detection and feature fusion, and outputs a standardized feature vector.

[0007] Furthermore, in the intelligent monitoring system for civil construction equipment in a thermal power plant according to the present invention, the terminal-side lightweight model is configured as follows: continuously monitoring the rate of change of bearing temperature in the heat distribution matrix; When the bearing temperature change rate exceeds the preset threshold, the spectrum data and time domain vibration waveform within the period are captured synchronously; The MobileNetV3 architecture with cropped convolution kernels is used to fuse multi-source feature fragments and generate standardized feature vectors.

[0008] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital processing module includes: Parametric modeling unit, importing equipment CAD drawings to drive multi-body dynamics solutions and build a digital twin framework for construction equipment; A spatiotemporal graph convolutional network unit receives the point cloud data and vibration signal in the standardized feature vector, converts the coordinate system based on the joint angle of the twin body, and outputs the spatiotemporal alignment data to the virtual calibration unit; The virtual calibration unit corrects the sensor temperature drift error in the spatiotemporal alignment data according to the motor current simulation theoretical vibration waveform, and generates a multi-modal synchronous data stream.

[0009] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the virtual calibration unit is configured as follows: extracting a hoist vibration spectrum from the spatiotemporal aligned data; Collect real-time current data from the crane hoist and input it into the digital twin physics engine to calculate the theoretical cable tension value; Receive the actual value measured by the tension sensor and compare the deviation between the theoretical value and the actual value; The spectrum compensation coefficient is generated based on the deviation value, and the vibration spectrum data is corrected and output to the multimodal synchronous data stream.

[0010] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital analysis module includes: MAML pre-trained network, using simulation data from hundreds of working conditions to initialize the dynamic weight allocator; A working condition classifier receives the load rate and wind speed parameters in the multimodal synchronous data stream, and outputs a priority instruction to the dynamic weight allocator when the wind speed is greater than level 6 or the load rate is greater than 85%; a dynamic weight allocator, receiving the priority instruction to allocate sensor feature weight coefficients and generate a weighted feature vector; The attention weighting unit performs multimodal feature fusion on the weighted feature vector and outputs the fused feature tensor to the anomaly recognition module.

[0011] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the dynamic weight allocator is configured to perform the following operation on a concrete pump truck blocked pipe condition: Receive pressure fluctuation characteristics and motor vibration characteristics data of pipe blockage conditions; A weight coefficient of 0.6 is assigned to the pressure fluctuation feature; Assign a weight coefficient of 0.1 to the motor vibration characteristics; Output the low-dimensional weighted feature vector of the enhanced pressure feature to the attention weighted unit.

[0012] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the abnormality identification module includes: The causal graph construction unit applies the FCI algorithm to generate the Bayesian network of the fault causal chain and transfers the causal graph parameters to the temporal convolutional network; A temporal convolutional network receives the fused feature tensor and causal graph parameters, detects the temporal correlation of abnormal events, and outputs a warning signal to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the fault propagation path, verifies the validity of the early warning signal, and then outputs it to the monitoring terminal and optimization module.

[0013] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the temporal convolutional network is configured to perform crane hoisting monitoring operations: receiving the crane deformation data and the cable tension data in the fused feature tensor; Analyze the causal probability that deformation data becomes abnormal earlier than cable tension data; When the causal probability exceeds 85%, a structural instability warning signal is generated and transmitted to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the cable slipping process, verifies the effectiveness of the warning, and then outputs it to the monitoring terminal.

[0014] Furthermore, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the optimization module is configured as follows: Receiving the false alarm record sent by the anomaly recognition module; The edge node extracts false alarm cases, fine-tunes the recognition model parameters, and uploads them to the cloud in an encrypted manner; The cloud aggregator uses the FedProx algorithm to fuse multi-node model parameters to generate a lightweight model; The adversarial sample generator constructs extreme working condition data and injects it into the digital twin to expand the training samples; Deploy the updated lightweight model to the lightweight model on the edge of the digital acquisition module.

[0015] Beneficial effects of the present invention: This invention systematically solves the problem of data fusion lag in heterogeneous sensors of construction equipment through the three-level collaboration of edge multi-source data collaborative processing, digital twin-driven physical simulation calibration, dynamic feature fusion, and federated closed-loop optimization: Motion sensors, LiDAR units, and infrared thermal imagers perform signal decomposition and feature extraction on the device side. The lightweight model uses a temperature change rate trigger mechanism to synchronously intercept multi-source data fragments, generate standardized feature vectors, reduce the amount of original data transmission, and provide low-latency, high-quality input for subsequent fusion.

[0016] The digital twin establishes a unified spatiotemporal benchmark based on the dynamic parameters of the equipment, the spatiotemporal graph convolutional network aligns the coordinate systems of multiple source signals, and the virtual calibration unit corrects the sensor temperature drift error through the theoretical value of motor current inversion, eliminating the spatiotemporal misalignment between vibration signals and posture data, making the multimodal synchronous data stream physically interpretable.

[0017] A meta-learning weight allocator dynamically adjusts feature weights based on operating risk, and an attention mechanism integrates these weights to generate low-dimensional feature tensors. The anomaly recognition module combines causal graphs with twin verification to achieve high-precision early warnings. Federated learning aggregates edge node parameters to update lightweight models, and adversarial examples are injected into digital twins to expand the training sample library, forming a closed loop that accelerates anomaly recognition response speeds to milliseconds. This three-tiered technology addresses bottlenecks in data transmission, spatiotemporal calibration, and model generalization, significantly reducing the rates of missed and false alarms caused by delayed heterogeneous data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0019] Figure 1 This is a system architecture diagram of the intelligent monitoring system for civil construction equipment in a thermal power plant provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0021] See also Figure 1 The present invention provides an intelligent monitoring system for civil construction equipment in a thermal power plant, comprising: The digital acquisition module deploys sensors on earthmoving machinery, concrete equipment, and lifting equipment to collect raw vibration, posture, and temperature data in real time. It then performs on-device noise filtering and feature extraction, and outputs standardized feature vectors to the digital processing module. a digital processing module that receives the standardized feature vectors, constructs a digital twin of the construction equipment, calibrates the sensor coordinate offset and timestamp deviation using multi-body dynamics simulation, generates a multimodal synchronous data stream, and transmits it to the digital analysis module; A digital analysis module parses the multimodal synchronous data stream, uses a pre-trained working condition classifier to identify load and environmental parameters, dynamically assigns sensor feature weight coefficients, generates a fused feature tensor, and provides it to the anomaly recognition module; The anomaly recognition module receives the fused feature tensor, mines the fault sequence chain through the causal graph, outputs an early warning signal to the monitoring terminal when an abnormal causal pattern is detected, and simultaneously sends a false alarm record to the optimization module; The optimization module receives the warning signal and false alarm record, generates an adversarial sample set to update the recognition model parameters, and deploys the optimized model parameters to the digital analysis module and the anomaly recognition module.

[0022] Multi-source sensors are deployed at key locations, such as the traveling mechanisms of earth-moving machinery and crane articulation points. Vibration sensors are embedded with wavelet packet decomposition chips, which separate the raw vibration signal into effective spectral components reflecting the wear status of the equipment and mechanical impact noise, outputting a spectral feature vector. The LiDAR unit generates a 3D point cloud of the equipment's position and posture using SLAM positioning firmware, addressing visual positioning drift in dynamic operating scenarios. The infrared thermal imager performs non-uniformity correction and outputs a thermal distribution matrix representing the surface temperature distribution of the equipment. The lightweight model on the device (MobileNetV3 cropped architecture) receives three types of data. When the bearing temperature change rate exceeds a preset threshold, it simultaneously captures the spectrum and vibration waveform of the associated time period and fuses them to generate a 128-dimensional normalized feature vector. This module reduces transmission latency through edge computing.

[0023] The parametric modeling unit imports equipment CAD drawings, drives multibody dynamics solutions, and constructs the digital twin framework. The spatiotemporal graph convolutional network unit extracts point clouds and vibration signals from standardized feature vectors, transforms the coordinate system based on the twin's joint angles, and eliminates coordinate drift between the visual and vibration sensor data. The virtual calibration unit infers the theoretical vibration waveform based on the motor current (for example, calculating the theoretical cable tension from the winch current). This is compared with the actual measured value from the tension sensor to generate a compensation coefficient. After correcting for temperature drift errors, the unit outputs a spatiotemporally synchronized multimodal data stream. This module addresses the issue of data asynchrony from multiple sources through physical simulation and calibration.

[0024] The MAML pre-trained network initializes a dynamic weight allocator based on simulated data from hundreds of working conditions. The working condition classifier analyzes the load factor and wind speed parameters in the multimodal synchronized data stream. When the load factor exceeds 85% or the wind speed exceeds level 6, a high-risk working condition priority strategy is triggered. The dynamic weight allocator adjusts feature weights based on priority instructions (for example, in the case of a concrete pump truck pipe blockage, the pressure feature is weighted 0.6, and the motor vibration feature is weighted 0.1). The attention weighting unit performs cross-modal fusion on the weighted feature vectors to generate a low-dimensional fused feature tensor. This module adapts to the dynamic construction environment through meta-learning.

[0025] The causal graph construction unit applies the FCI algorithm to mine historical fault chains, such as hydraulic leaks, pressure drops, and pump shutdowns, and generates Bayesian network parameters. A temporal convolutional network, combined with Bayesian network parameter analysis, fuses feature tensors. When the crane deformation data is detected to be abnormal before the cable tension data with a probability exceeding 85%, a structural instability warning signal is generated. The digital twin verification unit uses the twin to simulate the cable slip process, matching the fault propagation path to verify the warning validity, and then outputs a signal to the monitoring terminal. Simultaneously, false alarm records are sent to the optimization module. This module integrates data-driven and physical verification mechanisms.

[0026] Edge nodes extract false alarms, fine-tune local recognition model parameters, and upload encrypted gradient data. The cloud aggregator uses the FedProx algorithm to fuse cross-device model parameters and generate a lightweight version. The adversarial example generator constructs extreme operating condition data (such as light and shadow interference deformation data) and injects it into the digital twin to expand the training sample library. The updated lightweight model is deployed to the lightweight model on the device side of the digital acquisition module.

[0027] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital acquisition module includes: The vibration sensor embeds a wavelet packet decomposition chip to separate the effective spectrum from the impact noise, and outputs the spectrum data to the lightweight model on the device side. The LiDAR unit is configured with SLAM positioning firmware to generate a point cloud of the device's pose and provide the point cloud data to the lightweight model on the device side. Infrared thermal imager with built-in temperature field reconstruction algorithm outputs heat distribution matrix and transmits the heat distribution matrix to the lightweight model on the terminal side; The lightweight model on the end receives the spectrum data, point cloud data, and heat distribution matrix, performs millisecond-level anomaly detection and feature fusion, and outputs a standardized feature vector.

[0028] Vibration sensors are deployed at key vibration sources, such as equipment transmission bearings and hydraulic motors. An embedded wavelet packet decomposition chip performs frequency band decomposition on the raw vibration signal. The chip uses a preset frequency band energy threshold (for example, high-frequency bands above 500 Hz) to separate the effective spectral components reflecting equipment wear from mechanical impact noise. The separated effective spectral components are converted into fixed-dimensional spectral data and transmitted to the lightweight device-side model with millisecond latency. This process eliminates high-amplitude random interference while preserving low-frequency features that represent equipment health.

[0029] LiDAR units are installed at the end of moving mechanisms such as crane turntables and pump truck booms. SLAM positioning firmware calculates the equipment's spatial pose in real time. The firmware uses an iterative closest point algorithm to match the real-time scanned point cloud frames with the equipment's CAD model, generating a pose point cloud consisting of 3D spatial coordinates and Euler angles. The point cloud data is downsampled and compressed using a voxel grid, forming a structured array format and delivered to the lightweight model on the device via a timestamp synchronization mechanism. This technology addresses the problem of cumulative visual positioning errors in dynamic construction scenarios.

[0030] The infrared thermal imager is aimed at the surfaces of heat-generating components such as motor windings and gearbox housings. A built-in temperature field reconstruction algorithm generates a temperature distribution matrix based on the raw voltage signals from the thermopile array. The algorithm performs nonuniformity correction and radiation compensation calculations, mapping pixel-level temperature values ​​into a 32×32 resolution thermal distribution matrix. The matrix elements correspond to physical quantities in the device surface temperature field. A hardware clock synchronizer maintains temporal alignment with the vibration spectrum data and pose point cloud, preventing misalignment in the timing of multi-source data acquisition.

[0031] The lightweight model on the device runs on an embedded edge computing unit and simultaneously receives three types of input data: spectrum data obtained from the vibration sensor, the pose point cloud array provided by the LiDAR unit, and the heat distribution matrix transmitted by the infrared thermal imager. The model performs three operations: Real-time monitoring of the temperature change rate of the bearing area in the heat distribution matrix; When the temperature change rate exceeds a preset threshold (e.g., 5°C / s), the interception mechanism is triggered to synchronously acquire the spectrum data and time-domain vibration waveform within that period; The MobileNetV3 architecture with cropped convolution kernels is used to fuse multi-source features, reducing the dimensionality through a 1×1 convolutional layer to output a 128-dimensional normalized feature vector. This fusion process preserves cross-modal spatiotemporal correlations and eliminates data scale differences caused by device model differences.

[0032] Specifically, in the intelligent monitoring system for civil construction equipment in a thermal power plant according to the present invention, the terminal-side lightweight model is configured as follows: continuously monitoring the rate of change of bearing temperature in the heat distribution matrix; When the bearing temperature change rate exceeds the preset threshold, the spectrum data and time domain vibration waveform within the period are captured synchronously; The MobileNetV3 architecture with cropped convolution kernels is used to fuse multi-source feature fragments and generate standardized feature vectors.

[0033] The lightweight model on the device side analyzes the coordinates of the bearings in the thermal distribution matrix in real time (based on the preset position indexes in the device CAD model) and calculates the temperature differential values ​​for consecutive time frames. The differential algorithm calculates the temperature gradient within a fixed time window (for example, 1 second). When the gradient exceeds a preset threshold (for example, 5°C / second), subsequent data capture is triggered. This process is executed by a hardware accelerator, with monitoring latency controlled within 10 milliseconds.

[0034] When the bearing temperature change rate exceeds the threshold, the model starts the synchronous interception mechanism: lock the current timestamp T, and extract the multi-source data fragments within the time window [T-δt, T+δt] from the cache queue. Specifically, it includes: The spectrum data and time-domain vibration waveform collected by the vibration sensor during the period [T-δt, T+δt]; The pose point cloud keyframes generated by the LiDAR unit in the same period; The thermal distribution matrix frame capture operation corresponding to the timestamp T of the infrared thermal imager is achieved through precise timestamp matching, and the time deviation is controlled within ±5 milliseconds to ensure the time domain alignment of multimodal data.

[0035] Feature fusion is performed using the MobileNetV3 architecture with cropped convolution kernels: Architecture modification: Remove the last three residual blocks of the original model and halve the number of 3×3 depth-separable convolution kernels to reduce the amount of computation. Fusion process: The time-domain vibration waveform is input into the 1D convolution branch to extract the time-frequency features; The spectrum data and pose point cloud are input into the 2D convolution branch to extract spatial features; The heat distribution matrix input channel attention layer strengthens the key temperature area; Output processing: The multi-branch features are compressed into a 128-dimensional standardized feature vector through a fully connected layer, and the vector elements are normalized by Z-score.

[0036] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital processing module includes: Parametric modeling unit, importing equipment CAD drawings to drive multi-body dynamics solutions and build a digital twin framework for construction equipment; A spatiotemporal graph convolutional network unit receives the point cloud data and vibration signal in the standardized feature vector, converts the coordinate system based on the joint angle of the twin body, and outputs the spatiotemporal alignment data to the virtual calibration unit; The virtual calibration unit corrects the sensor temperature drift error in the spatiotemporal alignment data according to the motor current simulation theoretical vibration waveform, and generates a multi-modal synchronous data stream.

[0037] The parametric modeling unit imports the CAD drawing library of construction equipment and extracts the geometric parameters and mass properties of key moving components. Based on the principles of Lagrangian mechanics, multibody dynamics equations are established to solve the constraints of the equipment's joint degrees of freedom. A physics engine simulation is used to construct a digital twin framework that includes joint displacement, velocity, and acceleration parameters. This twin framework maps the kinematic state of the physical equipment in real time, providing a reference coordinate system for spatiotemporal alignment.

[0038] The spatiotemporal graph convolutional network unit receives the point cloud data and vibration signal in the normalized feature vector and performs three-level processing: Spatial alignment: Using the twin joint angles as the reference system, point cloud data is converted from the world coordinate system to the device local coordinate system; Time synchronization: Align the vibration signal and the point cloud frame timestamp through interpolation algorithm to generate sequence data with matching time axis; Feature association: Establish the spatiotemporal correlation matrix between the spectral features of the vibration signal and the posture change; Output the spatiotemporal alignment data to the virtual calibration unit to eliminate the measurement deviation caused by the difference in sensor installation posture.

[0039] The virtual calibration unit infers the theoretical vibration waveform based on the motor current signal: Extracting the hoist vibration spectrum from the spatiotemporal aligned data; Collect the real-time current data of the motor and input it into the digital twin, and calculate the theoretical cable tension value through the electromagnetic torque equation; Receive the actual value measured by the tension sensor and calculate the frequency spectrum characteristic deviation between the theoretical value and the actual value; Generate frequency response compensation coefficient based on the deviation to correct the amplitude-frequency characteristics of the vibration spectrum; The corrected vibration spectrum is fused with the pose point cloud and temperature data to generate a multimodal synchronized data stream.

[0040] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the virtual calibration unit is configured as follows: extracting a hoist vibration spectrum from the spatiotemporal aligned data; Collect real-time current data from the crane hoist and input it into the digital twin physics engine to calculate the theoretical cable tension value; Receive the actual value measured by the tension sensor and compare the deviation between the theoretical value and the actual value; The spectrum compensation coefficient is generated based on the deviation value, and the vibration spectrum data is corrected and output to the multimodal synchronous data stream.

[0041] The virtual calibration unit extracts the vibration spectrum of the winch's specific frequency band (0-500Hz) from the spatiotemporal alignment data. A Fourier transform is used to convert the time-domain vibration signal into a complex spectrum vector, preserving both amplitude and phase information. The extraction process locates the corresponding data channel for the winch based on the device ID index, preventing interference from vibration signals from other devices.

[0042] Real-time three-phase current data from the crane's hoist motor is collected and fed into the digital twin's physics engine. The engine calculates the theoretical output torque based on the electromagnetic torque equation Te = Kt * Iq (Kt is the motor torque constant). This is combined with the pulley radius parameter r to solve for the theoretical cable tension Ftheory = Te / r. The calculated results are mapped to the frequency domain to generate a theoretical vibration spectrum, reflecting the vibration characteristics under ideal operating conditions.

[0043] The corrected vibration spectrum is re-fused with the original pose point cloud and temperature data. This fusion process maintains the timestamp synchronization of the spatiotemporally aligned data, generating a multimodal, synchronized data stream containing the calibrated vibration signature. This data stream is transmitted to downstream modules to ensure that the vibration signature is consistent with the physical state of the device.

[0044] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the digital analysis module includes: MAML pre-trained network, using simulation data from hundreds of working conditions to initialize the dynamic weight allocator; A working condition classifier receives the load rate and wind speed parameters in the multimodal synchronous data stream, and outputs a priority instruction to the dynamic weight allocator when the wind speed is greater than level 6 or the load rate is greater than 85%; a dynamic weight allocator, receiving the priority instruction to allocate sensor feature weight coefficients and generate a weighted feature vector; The attention weighting unit performs multimodal feature fusion on the weighted feature vector and outputs the fused feature tensor to the anomaly recognition module.

[0045] The MAML pre-trained network runs on a cloud server, loaded with a simulation dataset covering hundreds of operating conditions, including earthmoving machinery collapse and pump truck pipe blockage. The network employs a model-agnostic meta-learning framework, employing a dual-loop optimization mechanism: an inner loop fine-tunes the weight allocator parameters on a subset of simulation conditions, while an outer loop updates the meta-initialization parameters. This process enables the dynamic weight allocator to obtain initial weights that quickly adapt to new operating conditions, eliminating the need for repetitive training for traditional models.

[0046] The operating condition classifier parses structured parameters from the multimodal synchronized data stream in real time: it calculates the load factor from the pose point cloud data and reads wind speed values ​​from environmental sensors. The classifier uses a gradient boosting decision tree architecture. When wind speed exceeds level 6 (13.8 m / s threshold) or the load factor exceeds 85%, it generates a high-risk operating condition priority instruction. This instruction is encoded as a four-dimensional vector (wind speed level, load level, equipment type, operating condition code) and transmitted to the dynamic weight allocator to activate the safety policy.

[0047] The dynamic weight allocator performs feature enhancement operations after receiving priority instructions: Analyze the operating condition risk level in the instruction vector; Increase the weight coefficient of key sensors according to preset rules (for example, increase the weight of posture features by 0.3 when the wind speed is greater than level 6); The non-critical features are downgraded (for example, the vibration feature weight is reduced by 0.2 under low-risk working conditions) to generate a weighted feature vector, whose feature dimension is consistent with the original input but the numerical distribution is reconstructed.

[0048] The attention weighted unit performs cross-modal interaction on the weighted feature vector: Calculate the correlation matrix of vibration-pose-temperature features through a multi-head attention mechanism; Perform weighted averaging on the eigenvectors according to the correlation matrix; A 1×1 convolutional layer is used to compress the feature dimension to 30% of the original data volume; Output the low-dimensional fused feature tensor to the anomaly recognition module to retain cross-modal key correlation information.

[0049] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the dynamic weight allocator is configured to perform the following operations in the concrete pump truck pipe blocking condition: Receive pressure fluctuation characteristics and motor vibration characteristics data of pipe blockage conditions; A weight coefficient of 0.6 is assigned to the pressure fluctuation feature; Assign a weight coefficient of 0.1 to the motor vibration characteristics; Output the low-dimensional weighted feature vector of the enhanced pressure feature to the attention weighted unit.

[0050] The dynamic weight allocator receives specific inputs from a multimodal synchronized data stream: pressure fluctuation features refer to the time-domain pulsation feature vector of the pump truck's delivery pipe pressure sensor, and motor vibration features refer to the high-frequency vibration spectrum of the drive motor. The data is flagged by the operating condition classifier as a blockage risk state (pressure fluctuation frequency > 2 Hz and amplitude exceeding a threshold), triggering the weight allocator to execute a specific action.

[0051] Assign a weight coefficient of 0.6 to the pressure fluctuation feature, and the execution process is as follows: Extract time-frequency domain indicators of pressure fluctuation characteristics (main frequency amplitude, waveform kurtosis); Multiply the original eigenvector by the gain factor 0.6; Keep the vector dimension unchanged but strengthen the pressure-related dimension value; This operation highlights the core characterization value of pressure anomalies under pipe plugging conditions.

[0052] Assign a weight coefficient of 0.1 to the motor vibration characteristics. The execution process is: Identify high-frequency noise in the vibration spectrum that is unrelated to pressure fluctuations, multiply the original eigenvector by an attenuation coefficient of 0.1, suppress the interference of non-critical vibration components on the fusion results, and perform weight reduction to prevent secondary features from diluting the core fault signal.

[0053] The operation of generating low-dimensional weighted feature vectors is as follows: the pressure feature vector after gain is concatenated with the vibration feature vector after attenuation, the concatenated vector is reduced to 50% of the original dimension through the PCA algorithm, and the vector elements are normalized to conform to the Z-score distribution. The output data retains the main features of pressure fluctuations and is transmitted to the attention weighted unit for cross-modal fusion.

[0054] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the abnormality identification module includes: The causal graph construction unit applies the FCI algorithm to generate the Bayesian network of the fault causal chain and transfers the causal graph parameters to the temporal convolutional network; A temporal convolutional network receives the fused feature tensor and causal graph parameters, detects the temporal correlation of abnormal events, and outputs a warning signal to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the fault propagation path, verifies the validity of the early warning signal, and then outputs it to the monitoring terminal and optimization module.

[0055] The causal graph construction unit loads a historical fault database (including thousands of cases such as hydraulic leaks and bearing failures) and applies the Fast Causal Inference (FCI) algorithm to discover causal dependencies between fault events. The algorithm performs conditional independence tests, identifies direct causal chains between variables, and generates a Bayesian network with a directed acyclic graph structure. Network parameters, including a node conditional probability table, are transmitted via Ethernet to a time-series convolutional network to initialize the fault detection model.

[0056] The temporal convolutional network receives two inputs: the fused feature tensor output by the digital analysis module and the Bayesian network parameters transmitted by the causal graph construction unit. The network uses a dilated convolution architecture and performs three-order temporal analysis: Predefine the abnormal event detection priority based on the cause-effect diagram parameters (e.g., oil temperature abnormality detection priority is higher than vibration abnormality); Detecting abnormal event sequences on the time axis of feature tensors; Calculate the matching degree between the event sequence and the causal chain (e.g., the temporal relationship of deformation anomaly occurring earlier than cable tension anomaly); When the event sequence conforms to the fault causal chain and the matching degree exceeds 85%, a structured early warning signal (including device ID, fault type, and risk level) is generated.

[0057] After receiving the warning signal, the digital twin verification unit calls the digital twin physical engine to perform fault propagation simulation: Injecting the initial fault parameters corresponding to the warning signal into the twin (e.g., a 30% reduction in the friction coefficient of the crane cable); Simulate the fault propagation process (e.g., cable slippage causing boom deformation growth); Compare simulation results with actual monitoring data in the feature tensor; When the deformation growth trend matches over 90%, the effectiveness of the early warning is verified. The verification results are output to the monitoring terminal, and false alarm cases are simultaneously sent to the optimization module.

[0058] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the temporal convolutional network is configured to perform crane hoisting monitoring operations: receiving the crane deformation data and the cable tension data in the fused feature tensor; Analyze the causal probability that deformation data becomes abnormal earlier than cable tension data; When the causal probability exceeds 85%, a structural instability warning signal is generated and transmitted to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the cable slipping process, verifies the effectiveness of the warning, and then outputs it to the monitoring terminal.

[0059] The temporal convolutional network extracts channel-specific data from the fused feature tensor: crane deformation data refers to the sequence of microstrain values ​​collected by the strain sensor installed on the boom, and cable tension data refers to the time-domain waveform output by the strain sensor. Data is then matched to a crane-specific analysis channel based on the device ID to avoid confusion between data from multiple devices.

[0060] Analyze the causal probability that deformation data is earlier than tension data. The execution process is as follows: The time difference Δt between abnormal deformation points (such as strain value mutation exceeding threshold value) and abnormal tension points is detected by mutual information calculation; When the causal probability exceeds the threshold, a structural instability warning signal is generated: The coded warning type is "wire rope slip risk" Mark the abnormal time window [t_start, t_end]; Additional deformation and tension correlation parameters; The signal is transmitted to the digital twin verification unit via a low-latency message queue.

[0061] The digital twin verification unit calls the twin to perform physical simulation: injecting the friction coefficient attenuation parameter (μ is reduced by 40%) into the twin cable model, simulating the dynamic process of cable slippage, and recording the boom deformation growth curve.

[0062] Compare the correlation coefficient between the simulation curve and the actual deformation data: when the correlation coefficient is greater than 0.9, the early warning is verified to be effective and the signal is output to the monitoring terminal.

[0063] Specifically, in the intelligent monitoring system for civil construction equipment of a thermal power plant according to the present invention, the optimization module is configured as follows: Receiving the false alarm record sent by the anomaly recognition module; The edge node extracts false alarm cases, fine-tunes the recognition model parameters, and uploads them to the cloud in an encrypted manner; The cloud aggregator uses the FedProx algorithm to fuse multi-node model parameters to generate a lightweight model; The adversarial sample generator constructs extreme working condition data and injects it into the digital twin to expand the training samples; Deploy the updated lightweight model to the lightweight model on the edge of the digital acquisition module.

[0064] The optimization module receives false alarm record data packets from the anomaly identification module via a message queue. These packets contain three elements: a warning timestamp, a fault type code, and a fragment of raw sensor data at the monitoring point. During the receiving process, the module performs a data integrity check, filters out format errors, and retains valid false alarm cases for model iteration.

[0065] The edge node extracts valid cases from the false alarm records, loads the local recognition model copy and performs fine-tuning operations: the raw sensor data in the case is input into the model for forward propagation, the cross entropy loss between the model output and the true label (false alarm label) is calculated, and the parameters of the fully connected layer of the model are updated through backpropagation.

[0066] The fine-tuned parameter gradients are uploaded to the cloud after being encrypted with AES-256, and the original data is retained at the edge node to ensure privacy and security.

[0067] The cloud aggregator receives encrypted gradients from multiple edge nodes and performs federated aggregation using the FedProx algorithm: Decrypt the gradient data of each node, add L2 regularization terms to constrain the parameter offset amplitude, and use weighted average fusion gradients to update the global model parameters.

[0068] Generate a lightweight model version (with the number of parameters compressed to 30% of the original model) to solve the problem of insufficient model generalization caused by scattered data islands.

[0069] The adversarial sample generator constructs extreme working condition data based on false alarm cases: the physical parameter deviation corresponding to the false alarm is injected into the digital twin (such as a 50% reduction in the friction coefficient of the steel cable), and the extreme working condition sensor data (such as unstable vibration waveform) is simulated and generated.

[0070] Add adversarial perturbations (FGSM algorithm generates pixel-level noise) to generate data to expand the training sample library and enhance the model's robustness to edge cases.

[0071] The updated lightweight model is converted to the TensorRT engine format and deployed to the edge computing unit of the digital acquisition module through a secure channel. The deployment process performs version compatibility verification, replaces the lightweight model parameters on the device without changing the feature fusion architecture, and ensures continuous service availability of the system.

[0072] The algorithm and model of the present invention are explained as follows: On-device lightweight model (MobileNetV3 pruned architecture): This model is a streamlined convolutional neural network architecture deployed on edge computing devices. By removing some residual blocks (such as the last three residual layers) from the standard MobileNetV3 and halving the number of convolution kernels, computational complexity is reduced. Its function is to simultaneously process multi-source heterogeneous data (vibration spectrum, pose point cloud, thermal distribution matrix), using a channel-by-channel processing mechanism: The 1D convolution branch extracts the time-frequency features of the time-domain vibration waveform; The 2D convolution branch parses spatial pose data; The channel attention layer strengthens the key regions in the temperature matrix; The lightweight model on the edge (MobileNetV3 pruned architecture) compresses the original data into a 128-dimensional standardized vector while retaining multimodal correlation features, achieving millisecond-level response and resolving the real-time processing bottleneck in scenarios where edge device computing power is limited.

[0073] Spatiotemporal Graph Convolutional Network (ST-GCN): Based on the graph neural network structure, the topological relationship of device joints is constructed as a spatiotemporal graph. Spatial graph convolution: Using the joint angles of the digital twin as the reference coordinate system, the pose point cloud is converted from the world coordinate system to the device local coordinate system; Temporal convolution: aligning the timestamp sequence of the vibration signal and the point cloud through cubic spline interpolation; The spatiotemporal graph convolutional network (ST-GCN) eliminates the differences in sensor installation positions and sampling time deviations, generates spatiotemporally aligned data streams, and provides a unified benchmark for multimodal fusion.

[0074] Virtual calibration unit (physical drive compensation): uses the motor current signal to invert the physical state of the device (electromagnetic torque equation ), calculate the theoretical vibration waveform. By comparing it with the actual value measured by the tension sensor, the spectrum compensation coefficient (such as the amplitude correction coefficient α) is generated.

[0075] Technical Effect: Corrects sensor measurement deviations caused by temperature drift and aging, making the vibration spectrum physically interpretable and improving the reliability of subsequent anomaly detection.

[0076] Meta-learning initialization (MAML framework): Pre-training dynamic weight allocator using model-independent meta-learning algorithm: Inner loop: fine-tuning weight parameters on a subset of simulation conditions (e.g., pipe blockage, overload); Outer loop: Update the initialization parameters to make the model quickly adapt to the new working conditions; Meta-learning initialization (MAML framework) gives the weight allocator the ability to generalize across working conditions, avoiding the defect that existing models need to be repeatedly trained in new scenarios.

[0077] Causal graph construction (FCI algorithm): Analyze the historical fault database based on the Fast Causal Inference algorithm: Perform conditional independence tests to identify direct causal relationships between variables; Construct a Bayesian network with a directed acyclic graph structure (including node conditional probability table); Technical Effect: A time chain model of fault propagation (such as "abnormal deformation → abnormal cable tension") is established to provide prior knowledge for time series analysis.

[0078] Temporal Convolutional Network (Dilated Convolutional Architecture): Combining Causal Graph Prior Knowledge with Fusion Feature Tensors: Setting the priority of abnormal event detection based on Bayesian network; Expand the temporal receptive field through dilated convolution; Calculate the probability of matching the event sequence with the causal chain; The temporal convolutional network (dilated convolutional architecture) identifies physically plausible failure modes (e.g., deformation occurs before tension anomalies), reducing false positive rates.

[0079] Federated Learning Optimization (FedProx algorithm): The closed-loop update mechanism includes: Edge fine-tuning: using false alarm cases to locally update the fully connected layer parameters of the model; Parameter aggregation: FedProx algorithm is used in the cloud to add L2 regularization term , constrain node parameter offset amplitude Technical effect: Integrate multi-node knowledge while protecting data privacy and improve model generalization ability.

[0080] Adversarial example generation (physical simulation injection): Digital twin-based physics engine: Inject extreme working condition parameters (such as a 40% decrease in the friction coefficient of the steel cable); Simulate and generate adversarial sensor data (e.g., unstable vibration waveforms); Technical effect: Expands the diversity of training samples and enhances the robustness of the model to edge cases.

[0081] The optimization modules of all embodiments perform closed-loop updates: false alarm cases trigger edge nodes to fine-tune model parameters and encrypt and upload gradient data; the cloud uses a federated learning algorithm to fuse multi-node parameters; the adversarial sample generator constructs extreme working condition data and injects it into the digital twin to expand the sample library; the updated lightweight model is deployed to the end-side device to continuously improve the accuracy and timeliness of anomaly recognition.

[0082] The specific implementation of this invention is based on the actual needs of civil construction scenarios in thermal power plants. Addressing the technical bottleneck of lagging multi-source heterogeneous sensor data fusion during dynamic operations of earthmoving machinery, concrete equipment, and lifting equipment, this invention achieves real-time abnormal state identification through a three-level architecture combining edge computing, digital twin calibration, and adaptive model optimization. Multi-source sensing units are deployed at the earthmoving machinery's travel mechanisms, crane articulation points, and key bearings in concrete pump trucks. A vibration sensor embedded with a wavelet packet decomposition chip performs frequency band segmentation on the raw vibration signal, separating the effective spectral components representing equipment wear from mechanical impact noise, and outputting spectral feature vectors to edge computing nodes. The LiDAR unit is equipped with SLAM positioning firmware, which uses an iterative closest point algorithm to match real-time scanned point cloud frames with the equipment's CAD model, generating a pose point cloud containing 3D spatial coordinates and pose Euler angles. This pose point cloud is then compressed through voxel downsampling to form a structured array. The infrared thermal imager performs non-uniformity correction and radiation compensation calculations, mapping pixel-level temperature values ​​into a 32x32 resolution heat distribution matrix. A hardware clock synchronizer maintains time-domain alignment with the vibration spectrum data. The lightweight model on the device runs on an embedded edge computing unit, continuously monitoring the temperature change rate of the bearing area in the heat distribution matrix. When the temperature gradient within a 1-second time window exceeds the preset threshold of 5°C / second, the synchronous interception mechanism is triggered to lock the current timestamp, extract the spectrum data, time-domain vibration waveform, and pose point cloud keyframes within the ±0.5-second window, and use the MobileNetV3 architecture with cropped convolution kernels to fuse multi-source feature fragments. After dimensionality reduction through 11 convolutional layers, the output is a 128-dimensional Z-score normalized normalized feature vector.

[0083] The parametric modeling unit imports a library of construction equipment CAD drawings and establishes multibody dynamics equations based on Lagrangian mechanics. The unit solves the constraints on the joint degrees of freedom of the equipment and constructs a digital twin framework through physics engine simulation, mapping the displacement, velocity, and acceleration parameters of the joints in real time. The spatiotemporal graph convolutional network unit receives the point cloud and vibration signals in the standardized feature vector and performs third-order processing using the twin's joint angles as the reference coordinate system. The pose point cloud is converted from the world coordinate system to the equipment's local coordinate system to eliminate installation deviations. A cubic spline interpolation algorithm is used to align the vibration signal and point cloud timestamps to generate a timeline matching sequence. A spatiotemporal correlation matrix is ​​then established between the vibration spectrum characteristics and pose changes. The virtual calibration unit corrects sensor errors based on physical mechanisms. The unit extracts the winch's unique 0-500Hz frequency spectrum from the spatiotemporal alignment data. The unit collects three-phase current data from the crane winch and inputs it into the digital twin. The theoretical cable tension is calculated based on the electromagnetic torque equation. The unit receives measured values ​​from the tension sensor and calculates the spectral feature deviation. Frequency response compensation coefficients are generated to correct the amplitude-frequency characteristics of the vibration spectrum, outputting a spatiotemporally synchronized multimodal data stream.

[0084] The digital analysis module's MAML pre-trained network loads a dataset of hundreds of working condition simulations on a cloud server. It initializes a dynamic weight allocator through an inner loop fine-tuning and outer loop updating of meta-parameters. The working condition classifier analyzes the load factor and wind speed parameters in the multimodal data stream in real time, employing a gradient boosting decision tree architecture for working condition risk assessment. A four-dimensional priority instruction vector is generated when the wind speed exceeds 13.8 m / s or the load factor exceeds 85%. The dynamic weight allocator responds to instructions and performs feature enhancement: Under the concrete pump truck pipe blockage condition, a weight of 0.6 is assigned to pressure fluctuation features to enhance the main frequency amplitude and waveform kurtosis, while a weight of 0.1 is assigned to motor vibration features to suppress high-frequency noise interference. The output is a weighted feature vector reduced to 50% of its original dimension using the PCA algorithm. The attention weighting unit uses a multi-head attention mechanism to calculate the correlation matrix of vibration, posture, and temperature features. Based on the cross-modal interaction weights, a weighted average is performed to generate a low-dimensional fused feature tensor.

[0085] The causal graph construction unit of the anomaly identification module applies the FCI algorithm to mine the historical fault database and performs conditional independence tests to generate a Bayesian network causal graph. A temporal convolutional network, combined with causal graph parameter analysis, fuses feature tensors. This involves extracting boom deformation data and cable tension data from crane hoisting monitoring and detecting the time difference between abnormal deformation and abnormal tension. A structural instability warning signal is generated when the deformation anomaly occurs earlier than the tension anomaly and the causal probability calculated by mutual information exceeds 85%. The digital twin verification unit uses the twin to inject a 40% attenuation parameter for the cable friction coefficient, simulating the dynamic process of cable slippage to generate a boom deformation growth curve. The warning's effectiveness is verified when the correlation coefficient between the simulated curve and the actual monitoring data exceeds 0.9. The optimization module receives false alarm record data packets, the edge node extracts cases to fine-tune the fully connected layer parameters of the recognition model, and the encrypted gradient is uploaded to the cloud via AES-256; the cloud aggregator uses the FedProx algorithm to add L2 regularization terms to constrain parameter offsets, and weighted average fusion of multi-node gradients updates the global model; the adversarial sample generator injects false alarms into the digital twin corresponding to physical parameter deviations to generate extreme working condition data, expands the training samples, converts them into the TensorRT engine format, and deploys them to the lightweight model at the edge to complete the closed-loop optimization.

[0086] This invention addresses the issue of real-time fusion and anomaly identification lags in heterogeneous sensor data through a three-level technical architecture. The first level of edge preprocessing eliminates noise in raw data: the digital acquisition module deploys intelligent sensors on the device side, the vibration sensor separates the effective spectrum from impact noise, the LiDAR unit generates a pose point cloud, and the infrared thermal imager outputs a heat distribution matrix. A lightweight model on the device side triggers the simultaneous capture of multi-source data based on the temperature change rate, which is then fused to generate a standardized feature vector. This process reduces the amount of transmitted data while preserving key features, addressing processing delays caused by data heterogeneity at the edge layer.

[0087] The second level of virtual simulation calibration achieves spatiotemporal unification: the digital processing module constructs the digital twin framework, and the spatiotemporal graph convolutional network transforms the coordinate system based on joint angles, eliminating deviations between visual and vibration sensor data. The virtual calibration unit infers the theoretical tension value based on the motor current, compares it with the measured value, generates spectrum compensation coefficients to correct for temperature drift errors, and outputs a multimodal synchronized data stream. This level of technology resolves the spatiotemporal misalignment of multi-source signals through physical simulation calibration, ensuring the basic consistency of data fusion.

[0088] The third-level dynamic fusion and closed-loop optimization improve the real-time decision-making: the digital analysis module uses meta-learning to initialize the dynamic weight allocator, the working condition classifier identifies high-risk scenarios, activates the priority strategy, and dynamically allocates sensor feature weight coefficients; the attention weighted unit fusion generates a low-dimensional feature tensor; the anomaly recognition module detects fault chains through causal graphs and temporal convolutional networks, and the digital twin verification unit simulates and verifies the effectiveness of the warning; the optimization module generates adversarial samples based on false alarm records, and federated learning updates the model parameters and deploys them to the edge.

Claims

1. Intelligent monitoring system for civil construction equipment in thermal power plants, characterized by: include: The digital acquisition module deploys sensors on earthmoving machinery, concrete equipment, and lifting equipment to collect raw vibration, posture, and temperature data in real time. It then performs on-device noise filtering and feature extraction, and outputs standardized feature vectors to the digital processing module. a digital processing module that receives the standardized feature vectors, constructs a digital twin of the construction equipment, calibrates the sensor coordinate offset and timestamp deviation using multi-body dynamics simulation, generates a multimodal synchronous data stream, and transmits it to the digital analysis module; A digital analysis module parses the multimodal synchronous data stream, uses a pre-trained working condition classifier to identify load and environmental parameters, dynamically assigns sensor feature weight coefficients, generates a fused feature tensor, and provides it to the anomaly recognition module; The anomaly recognition module receives the fused feature tensor, mines the fault sequence chain through the causal graph, outputs an early warning signal to the monitoring terminal when an abnormal causal pattern is detected, and simultaneously sends a false alarm record to the optimization module; The optimization module receives the warning signal and false alarm record, generates an adversarial sample set to update the recognition model parameters, and deploys the optimized model parameters to the digital analysis module and the anomaly recognition module.

2. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 1 is characterized in that: The digital acquisition module includes: The vibration sensor embeds a wavelet packet decomposition chip to separate the effective spectrum from the impact noise, and outputs the spectrum data to the lightweight model on the device side. The LiDAR unit is configured with SLAM positioning firmware to generate a point cloud of the device's pose and provide the point cloud data to the lightweight model on the device side. Infrared thermal imager with built-in temperature field reconstruction algorithm outputs heat distribution matrix and transmits the heat distribution matrix to the lightweight model on the terminal side; The lightweight model on the end receives the spectrum data, point cloud data, and heat distribution matrix, performs millisecond-level anomaly detection and feature fusion, and outputs a standardized feature vector.

3. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 2 is characterized in that: The end-side lightweight model is configured as follows: continuously monitoring the rate of change of bearing temperature in the heat distribution matrix; When the bearing temperature change rate exceeds the preset threshold, the spectrum data and time domain vibration waveform within the period are captured synchronously; The MobileNetV3 architecture with cropped convolution kernels is used to fuse multi-source feature fragments and generate standardized feature vectors.

4. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 3 is characterized in that: The digital processing module includes: Parametric modeling unit, importing equipment CAD drawings to drive multi-body dynamics solutions and build a digital twin framework for construction equipment; A spatiotemporal graph convolutional network unit receives the point cloud data and vibration signal in the standardized feature vector, converts the coordinate system based on the joint angle of the twin body, and outputs the spatiotemporal alignment data to the virtual calibration unit; The virtual calibration unit corrects the sensor temperature drift error in the spatiotemporal alignment data according to the motor current simulation theoretical vibration waveform, and generates a multi-modal synchronous data stream.

5. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 4 is characterized in that: The virtual calibration unit is configured to: extracting a hoist vibration spectrum from the spatiotemporal aligned data; Collect real-time current data from the crane hoist and input it into the digital twin physics engine to calculate the theoretical cable tension value; Receive the actual value measured by the tension sensor and compare the deviation between the theoretical value and the actual value; The spectrum compensation coefficient is generated based on the deviation value, and the vibration spectrum data is corrected and output to the multimodal synchronous data stream.

6. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 5, characterized in that: The digital analysis module includes: MAML pre-trained network, using simulation data from hundreds of working conditions to initialize the dynamic weight allocator; A working condition classifier receives the load rate and wind speed parameters in the multimodal synchronous data stream, and outputs a priority instruction to the dynamic weight allocator when the wind speed is greater than level 6 or the load rate is greater than 85%; a dynamic weight allocator, receiving the priority instruction to allocate sensor feature weight coefficients and generate a weighted feature vector; The attention weighting unit performs multimodal feature fusion on the weighted feature vector and outputs the fused feature tensor to the anomaly recognition module.

7. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 6, characterized in that: The dynamic weight allocator is configured to: Receive pressure fluctuation characteristics and motor vibration characteristics data of pipe blockage conditions; A weight coefficient of 0.6 is assigned to the pressure fluctuation feature; Assign a weight coefficient of 0.1 to the motor vibration characteristics; Output the low-dimensional weighted feature vector of the enhanced pressure feature to the attention weighted unit.

8. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 7, characterized in that: The abnormality identification module includes: The causal graph construction unit applies the FCI algorithm to generate the Bayesian network of the fault causal chain and transfers the causal graph parameters to the temporal convolutional network; A temporal convolutional network receives the fused feature tensor and causal graph parameters, detects the temporal correlation of abnormal events, and outputs a warning signal to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the fault propagation path, verifies the validity of the early warning signal, and then outputs it to the monitoring terminal and optimization module.

9. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 8, characterized in that: The temporal convolutional network is configured to perform crane hoisting monitoring operations: receiving the crane deformation data and the cable tension data in the fused feature tensor; Analyze the causal probability that deformation data becomes abnormal earlier than cable tension data; When the causal probability exceeds 85%, a structural instability warning signal is generated and transmitted to the digital twin verification unit; The digital twin verification unit calls the twin to simulate the cable slipping process, verifies the effectiveness of the warning, and then outputs it to the monitoring terminal.

10. The intelligent monitoring system for civil construction equipment of a thermal power plant according to claim 9, characterized in that: The optimization module is configured to: Receiving the false alarm record sent by the anomaly recognition module; The edge node extracts false alarm cases, fine-tunes the recognition model parameters, and uploads them to the cloud in an encrypted manner; The cloud aggregator uses the FedProx algorithm to fuse multi-node model parameters to generate a lightweight model; The adversarial sample generator constructs extreme working condition data and injects it into the digital twin to expand the training samples; Deploy the updated lightweight model to the lightweight model on the edge of the digital acquisition module.

Images