Method for detecting the tightness of sgs system of photothermal power station based on 1d-cnn-bi-lstm model

Abstract

This invention, based on a 1D-CNN-Bi-LSTM model, is a leak detection method for SGS systems in concentrated solar power (CSP) plants. It aims to address the problems of existing non-invasive monitoring methods failing to distinguish leaking devices and accurately locate leaks when there are medium leaks on the tube sides of the four components. Traditional disassembly and inspection methods also result in long downtime, high maintenance costs, and secondary damage. This invention deploys multiple types of sensors on the shell sides of the four components to simultaneously collect multi-dimensional time-series data on temperature, pressure, flow rate, sound waves, and hydrogen concentration. After preprocessing, a standard sample set is constructed. A 1D-CNN model is used to extract local leak features, combined with Bi-LSTM to mine bidirectional time-series correlation features of the signal. The leak probability is then output through probability normalization using a fully connected layer to identify the leaking device. This invention achieves non-destructive and accurate location, effectively reducing false alarm and false negative rates, shortening maintenance downtime, saving maintenance costs, avoiding secondary damage to equipment, and meeting the requirements of intelligent online leak detection for SGS systems.

Description

Technical Field

[0001] This invention relates to the field of solar thermal power generation technology, and in particular to a method for rigorous testing of SGS systems in solar thermal power plants based on a 1D-CNN-Bi-LSTM model. Background Technology

[0002] Concentrated solar power (CSP) is a crucial direction for solar thermal utilization. The Storage and Heat Exchange System (SGS) is a core subsystem of a CSP plant, primarily composed of key heat exchange equipment such as a preheater, evaporator, superheater, and reheater (hereinafter referred to as the "four units"). Each of the four units consists of multiple tube bundles with one inlet and one outlet, structurally divided into tube-side and shell-side components. Specifically, the preheater's tube-side contains water, while the shell-side contains molten salt (potassium nitrate and sodium nitrate); the evaporator's tube-side contains a steam-water mixture, while the shell-side contains molten salt; and the superheater and reheater's tube-side contains steam, while the shell-side contains molten salt. During long-term operation of the SGS system, due to the effects of high temperature, high pressure, and molten salt corrosion, leakage of the tube bundles from the tube side to the shell side may occur. This leakage can lead to decreased equipment efficiency, molten salt corrosion, increased heat loss, and even safety accidents. Therefore, rigorous leak testing of the four units is essential for ensuring the safe and stable operation of a CSP plant.

[0003] However, because the SGS system's four components (instrument, valve, and diaphragm) form a coupled flow field through series or parallel pipe connections, current leak detection technology for these components faces a key bottleneck: a lack of precise location capabilities for leaking devices. Existing methods cannot distinguish specific leaking devices through non-invasive monitoring, relying solely on destructive disassembly methods for investigation. This involves cutting and dismantling the shell-side connecting pipes of the four components, verifying the leak source by isolating individual devices or pressurizing them one by one. This process not only leads to lengthy system downtime and high maintenance costs but may also cause secondary damage to the sealing of non-leaking devices due to repeated pipe cutting and welding.

[0004] Existing conventional detection methods, such as visual inspection, conventional ultrasonic testing, and single-parameter threshold discrimination, can only identify obvious macroscopic defects or faults with large parameter limits. They cannot capture the weak temporal characteristics and multi-parameter coupling correlation patterns in the early stages of leakage. Furthermore, they are difficult to adapt to the complex operating conditions of leakage signal propagation and mutual interference under the coupled flow field of four devices. Specifically, CN120212442A discloses a method for identifying leaks in heating pipelines based on secondary decomposition and BiLSTM, using negative pressure waves as a single monitoring source. In the SGS system of a solar thermal power plant with multiple coupled devices, the negative pressure wave signal is severely affected by multiple fluid interferences, and the unidirectional modeling of BiLSTM cannot simultaneously capture the forward accumulation and backward propagation characteristics of the leakage signal in time, resulting in insufficient ability to distinguish leakage sources. CN120593991A discloses an online diagnostic system for condenser heat exchanger tube leaks, which relies on a preset fixed diagnostic priority queue and lacks the ability to adaptively extract local abrupt changes in leakage signal characteristics. Moreover, the localization and correction process relies on historical leakage deviation backtracking, which leads to significant temperature drift effects under the molten salt conditions of solar thermal power plants, easily generating cumulative localization errors. CN121558264A discloses a rapid diagnosis method for heat exchanger leakage faults, which relies on flow velocity and direction correction based on ultrasonic propagation speed in the localization stage. However, in the high-temperature molten salt environment of a concentrated solar power (CSP) plant, the ultrasonic propagation characteristics are complex, limiting the correction model. Furthermore, this method focuses on physical localization at the heat exchanger fin level and cannot directly output leakage classification results at the heat exchanger level. CN121145013A discloses a device cavitation fault detection method based on a hybrid architecture of CNN and BiLSTM, which targets the cavitation impact characteristics in vibration signals. It does not consider the multi-dimensional coupling relationship of five types of heterogeneous sensor data (pressure, temperature, flow rate, acoustic waves, and medium composition) in the SGS system of a CSP plant, and the model input is a single vibration time-series signal, which cannot adapt to the detection scenario of multi-parameter joint monitoring on the shell side.

[0005] In summary, existing technologies for leak detection suffer from insufficient specificity in signal noise reduction, feature extraction, temporal modeling, multi-source fusion, and location correction. In particular, they lack a non-invasive and precise leak location method that is suitable for the coupled flow field of the four heat exchangers in the SGS system of concentrated solar power plants and adapts to multi-dimensional monitoring data on the shell side. There is an urgent need to build an intelligent detection system that integrates local feature extraction and bidirectional temporal dependency modeling to achieve rapid and accurate identification of leaking heat exchangers. This would enable an efficient operation mode of "first accurately locating the leaking heat exchanger, then directionally opening the manhole door, internal inspection, and leak repair welding," solving the technical pain points of long downtime, high maintenance costs, and secondary damage caused by traditional blind disassembly. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a tightness detection method for the SGS system of a solar thermal power plant based on a 1D-CNN-Bi-LSTM model. This method addresses the technical problem that when the tube-side medium of the four components (preheater, evaporator, superheater, and reheater) in the SGS system of a solar thermal power plant leaks to the shell side, it is impossible to distinguish the specific leaking device through non-invasive monitoring and lacks the ability to accurately locate the leaking device.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: This invention provides a method for leak detection of SGS system in solar thermal power plants based on a 1D-CNN-Bi-LSTM model, which includes five steps: real-time data acquisition, data preprocessing and sample construction, model feature extraction and fusion, leakage probability classification and determination, and automatic output of detection results.

[0008] First, differential pressure sensors, vortex flow meters, thermocouples, acoustic emission sensors, and infrared gas analyzers are installed at the inlet and outlet of each heat exchanger on the shell side of the four heat exchangers in the SGS system. The sampling frequency is uniformly set to 100Hz, and independent monitoring channels are configured to synchronously collect five types of monitoring parameters: shell side temperature, pressure, flow rate, sound wave, and hydrogen concentration of each heat exchanger. Secondly, abnormal data are removed from the original monitoring data using the 3σ rule, and the differences in the dimensions and magnitudes of each parameter are eliminated by range normalization. Time series data are extracted with a fixed time window to construct a three-dimensional standardized time series sample set that is adapted to the model input. Next, a 1D-CNN-Bi-LSTM fusion model was constructed, consisting of an input layer, a 1D-CNN module, a Bi-LSTM module, a fully connected layer, and a probability output layer. The 1D-CNN module extracts local leakage features such as pressure abrupt changes, flow fluctuations, and acoustic anomalies in the time-series signal through a sliding convolution kernel. The Bi-LSTM module, relying on its internal forget gate, input gate, cell state unit, and output gate structure, mines the dynamic evolution correlation of leakage features from both forward and reverse bidirectional time-series dimensions and completes bidirectional feature splicing and fusion. Then, the fused features are input into the fully connected layer for feature mapping and dimension adaptation. After probability normalization, the leakage probability distribution of the four heat exchangers is output. The heat exchanger with the maximum probability is identified as the leakage source device. Finally, the system automatically generates a leak location report, which includes the leaking device number, fault confidence level, and key parameter change curves, guiding on-site directional opening of manhole doors, internal inspection, and leak repair welding.

[0009] During the model training phase, cross-entropy loss is used as the evaluation criterion. The Adam optimizer is selected with a fixed learning rate, and iterative training is performed using batch gradient descent. An early stopping mechanism is introduced to prevent model overfitting. During the model validation phase, a four-device simulation platform is built, and gradient leakage flow ranges are set to simulate multiple leakage scenarios. The training set and validation set are divided proportionally to complete the verification of model accuracy, false alarm rate, and false negative rate, as well as parameter tuning.

[0010] The present invention provides a method for airtightness testing of SGS systems in solar thermal power plants based on a 1D-CNN-Bi-LSTM model, which has the following beneficial effects: 1. This invention adopts a deep learning model architecture based on 1D-CNN-Bi-LSTM, which integrates local feature extraction and bidirectional temporal correlation mining capabilities. It solves the technical bottleneck that leakage devices of the four heat exchangers in the SGS system of solar thermal power plants cannot be accurately distinguished by non-invasive means under coupled flow fields, and realizes non-destructive and accurate positioning of leakage heat exchangers.

[0011] 2. This invention designs a multi-dimensional monitoring system with five types of sensors (pressure, flow, temperature, sound wave, and medium composition) on the shell side of the "four devices". It introduces unconventional monitoring methods such as acoustic emission sensors and infrared gas analyzers, which improves the information dimension and sensitivity of leak detection and solves the problem of high false alarm rate and false negative rate caused by traditional methods that rely on only a single parameter of pressure or flow.

[0012] 3. This invention constructs a 1D-CNN module (2 convolutional layers with 32 and 64 kernels respectively, and a kernel length of 3 for each) to extract local leakage features such as pressure changes and acoustic peaks in shell-side monitoring signals. This replaces the traditional manual feature engineering method and improves the efficiency and accuracy of automatic extraction of local abnormal features.

[0013] 4. This invention introduces a Bi-LSTM module (1-layer bidirectional LSTM with 128 hidden units) to simultaneously mine the forward and backward dependencies of leakage features in the time dimension, adapting to the dynamic evolution process of leakage signals from occurrence to spread, and solving the problem that unidirectional time series models cannot capture the complete evolution law of leakage signals.

[0014] 5. This invention adopts a data preprocessing scheme based on Min-Max normalization and the 3σ criterion, which solves the model input quality problem caused by dimensional differences and outliers in multi-source sensor data, and improves the consistency and reliability of model training data.

[0015] 6. This invention sets a time window. The sample construction strategy, based on dimension The sample set input model solves the engineering processing problems of time series data segmentation and multi-parameter alignment, and realizes the standardization and engineering deployability of model input data.

[0016] 7. This invention adopts a training configuration of cross-entropy loss function and Adam optimizer (learning rate 0.001), and introduces an early stopping mechanism (if the accuracy of the validation set does not improve for 5 consecutive rounds, training is stopped), which solves the problems of model overfitting and low training efficiency, and improves the model's generalization ability and training convergence speed.

[0017] 8. This invention constructs a classification output structure with two fully connected layers (64 and 4 neurons respectively) and a Softmax activation function, and maps the fused features to the leakage probability distribution of four heat exchangers, realizing the normalized output of leakage probability of multiple devices and the automatic determination of the device with the highest probability.

[0018] 9. This invention realizes an efficient operation method of "first accurately locating the leaking heat exchanger, then directionally opening the manhole door, internal inspection, and leak repair welding". It solves the problem of long system downtime caused by the traditional method of cutting and disassembling the shell-side connecting pipes of the "four heat exchangers" and pressure testing them one by one, and greatly reduces the operation and maintenance costs.

[0019] 10. This invention avoids secondary damage to the sealing of non-leaking equipment caused by repeated cutting and welding of pipelines, solves the technical pain point of traditional destructive inspection methods that damage the structural integrity of intact equipment, and ensures the operational safety of SGS system equipment.

[0020] 11. This invention verifies that features such as a decrease in shell-side outlet temperature, a sudden increase in pressure, and abnormal fluctuations in local flow on the shell side can be accurately captured by the model through actual detection of evaporator tube-side leakage cases. This solves the problem of difficulty in distinguishing leakage features from normal fluctuations and verifies the reliability and feasibility of the model.

[0021] 12. This invention, through actual detection of a preheater tube-side leakage case, confirmed by on-site disassembly of the manhole door that three tube bundles and one blind flange of the preheater had leakage faults, verifying the high consistency between the model prediction results and the actual situation on site, and solving the problem of doubts about the accuracy of the model in actual engineering scenarios.

[0022] 13. This invention integrates a dual-module architecture that combines 1D-CNN local features with Bi-LSTM temporal correlation, and combines the complementarity of multi-sensor data to realize the leakage difference identification of the "four devices" under coupled flow field. It solves the problem that the traditional single-parameter threshold judgment method cannot distinguish specific leakage devices, and significantly reduces the false alarm rate and false negative rate of tightness detection.

[0023] 14. This invention sets up a batch gradient descent strategy (batch size 32, iterations 100 times), which solves the problem of unstable model training under small sample conditions. Through sufficient training and validation with 10,000 sets of samples (70% training, 30% validation), the generalization detection capability of the model under different leakage levels (0.1~1.0m³ / h) is improved.

[0024] 15. This invention achieves full automation of the entire process, including automatic data acquisition, automatic preprocessing, intelligent feature extraction, automatic leak location, and automatic generation of test reports. It reduces reliance on manual experience and improves the standardization and intelligence level of SGS system tightness testing for solar thermal power plants.

[0025] 16. This invention is adapted to the tube-to-shell leakage characteristics under four different media conditions: preheater (water on the tube side), evaporator (steam-water mixture on the tube side), superheater, and reheater (steam on the tube side). It solves the problem that a single model cannot cover leakage detection under multiple operating conditions and improves the versatility and engineering applicability of the model in the SGS system of solar thermal power plants.

[0026] 17. The present invention adopts a 100Hz sampling frequency and an independent monitoring channel for the inlet and outlet of each heat exchanger, which solves the problem of feature loss caused by insufficient sampling rate of leakage signals, improves the model's ability to capture weak signals in the initial stage of leakage, and ensures the timeliness and accuracy of leakage detection.

[0027] 18. This invention replaces traditional destructive disassembly and inspection with a full-process non-invasive testing method, which effectively shortens the system downtime maintenance cycle, reduces manpower input, disassembly and assembly consumables and power generation losses, significantly reduces the operation and maintenance cost of solar thermal power plants, and avoids the risk of secondary damage to non-leaking equipment caused by blind disassembly. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the detection method of the present invention. Figure 2 This is a flowchart of the diagnostic process for the 1D-CNN-Bi-LSTM model of this invention; Figure 3 This describes the flow direction of the medium on the shell side of the "four evaporators" in the single evaporator mode of Embodiment 2 of the present invention. Figure 4 This describes the flow direction of the shell-side medium in the "four-evaporator" mode under the two-evaporator mode of Embodiment 3 of the present invention; Figure 5 These are schematic diagrams showing the evaporator leakage detection results in Embodiments 2 and 3 of the present invention; Figure 6 This is a schematic diagram of the preheater leakage detection results in Embodiment 4 of the present invention. Detailed Implementation

[0029] The technical solutions of the present invention will be further described below with reference to the embodiments and accompanying drawings: Example 1 like Figure 1 As shown, this embodiment provides a method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model, such as... Figure 1 As shown, the overall process includes five stages: data acquisition, data preprocessing, model building and feature extraction, leakage detection, and result output. The specific steps are as follows: Step 1: Real-time collection of multi-dimensional monitoring data At the shell-side inlet and outlet positions of the four heat exchangers (preheater, evaporator, superheater, and reheater) in the SGS (Storage and Heat Exchange System) of the solar thermal power plant, five types of detection devices, namely differential pressure sensors, vortex flow meters, thermocouples, acoustic emission sensors, and infrared gas analyzers, are respectively deployed to form a complete monitoring array.

[0030] All sensors are uniformly set to a sampling frequency of 100Hz. Each heat exchanger inlet and outlet is independently configured with a monitoring channel to synchronously and without interference collect five types of operating parameters: shell-side temperature, pressure, flow rate, acoustic signal, and medium hydrogen concentration, forming a raw multi-source time-series monitoring data stream.

[0031] Step 2: Data preprocessing and construction of standardized time series sample sets The raw data collected in step 1 undergoes three levels of preprocessing: Outlier removal: Using the 3σ criterion, a threshold is set based on the fluctuation range of the equipment's normal operating data to automatically identify and remove invalid data that deviates from the normal range due to sensor malfunctions or interference from the field environment; Normalization: The range normalization method is used to uniformly process multiple parameters such as temperature, pressure, and flow rate, eliminating the differences in the dimensions and numerical magnitudes of different physical parameters, and mapping all data to a unified numerical range. Sample extraction: Time series data is continuously extracted with a fixed time window length to construct a three-dimensional structured sample set containing the number of samples, time step, and monitoring parameter categories, which is fully adapted to the input requirements of time series neural network models.

[0032] Step 3: Construction and Feature Extraction of the 1D-CNN-Bi-LSTM Fusion Model Construct a fusion model consisting of an input layer, a 1D-CNN (One-Dimensional Convolutional Neural Network) module, a Bi-LSTM (Bidirectional Long Short-Term Memory) module, a fully connected layer, and a probability output layer: 1. Feature extraction of 1D-CNN module: A two-layer convolutional structure is used in conjunction with max pooling. The convolutional kernel slides along the one-dimensional time-series signal for sampling. Local features are weighted, fused, and biased. Nonlinear activation is used to retain effective leakage features such as pressure change, flow fluctuation, and acoustic anomaly, while suppressing steady-state operating noise. Multiple convolutional kernels are used to extract exclusive leakage features for different monitoring parameters in parallel to form a multi-channel feature map. 2. Bi-LSTM module temporal feature mining: A single-layer bidirectional long short-term memory structure is adopted, which works in concert through forget gate, input gate, cell state unit, and output gate. The forget gate discards irrelevant historical information, the input gate incorporates the current effective leakage features, the cell state preserves the temporal information of the entire leakage process, and the output gate outputs the effective temporal features. By splicing and fusing forward temporal deduction and backward temporal backtracking features, the dynamic evolution and correlation of leakage features are fully captured. 3. Model Training and Optimization: A cross-entropy loss evaluation mechanism is adopted to compare the model's predicted probabilities with the true labels and iteratively optimize the model parameters by minimizing the bias; the Adam (Adaptive Moment Estimation) optimizer is used with a fixed learning rate and batch gradient descent is set to train the model, limiting the number of training batches and the maximum number of iterations; an early stopping mechanism is set to automatically terminate training when the accuracy on the validation set does not improve for several consecutive rounds, thus avoiding overfitting. 4. Model Validation: Build an SGS system four-instrument simulation platform, set gradient leakage flow range to simulate multiple leakage conditions, collect normal and leakage samples, divide the training set and validation set according to a fixed ratio, evaluate the model accuracy, false alarm rate, and false negative rate, and complete the model structure and hyperparameter optimization tuning.

[0033] Step 4: Identification and Location of Leaking Equipment The fused features output by Bi-LSTM are fed into a fully connected layer of a two-level neuron structure to perform linear mapping and bias compensation on the features. After probability normalization, the discrimination scores of each heat exchanger are converted into normalized probability distributions, and the probability values ​​directly represent the leakage confidence level of the corresponding heat exchanger. By comparing the leakage probability distribution of the four heat exchangers, the heat exchanger with the highest probability was identified as the leaking device.

[0034] When the medium on the tube side of the heat exchanger leaks to the shell side, the shell side will exhibit a series of interconnected changes, including a decrease in temperature, an abnormal increase in pressure, and disordered fluctuations in local flow. The model relies on the differences in abnormal characteristics of multi-parameter coupling to achieve precise location of the leakage source under the coupling structure of multiple devices, completely replacing the traditional destructive detection method of cutting and dismantling pipelines and pressurizing each unit for inspection.

[0035] Step 5: Automatic generation and output of test reports The system automatically generates a leak detection and leakage location report, which clearly outputs the leaking device number, fault confidence level, and key monitoring parameter change curves. This enables non-destructive leak detection and precise location of leaking devices in the SGS system of solar thermal power plants, providing direct evidence for on-site targeted maintenance.

[0036] Example 2 In another preferred embodiment, based on Embodiment 1, this embodiment provides a method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model, employing a single evaporator operation mode, such as... Figure 3 As shown, the molten salt flows sequentially from hot salt through the heat exchanger, reheater, single evaporator, and preheater, and then returns through cold salt to form a closed loop. Parameter monitoring points are set up at the inlet and outlet of each heat exchanger in SGS to simultaneously collect time-series data on shell-side temperature, pressure, flow rate, sound waves, and hydrogen concentration.

[0037] In the operation scenario of the evaporator of the solar thermal power plant's heat storage and exchange system, under the condition of normal operation after the equipment is put into operation, the equipment exhibits tube-side tube bundle leakage. The tube-side tube bundle leakage fault was confirmed by disassembling the manhole door for inspection.

[0038] The sensor sampling frequency is set to 100Hz, outlier data is removed using the 3σ criterion, and then processed by Min-Max normalization. (1); In the formula, These are the original monitoring parameter values ​​before preprocessing; These are the normalized values; , These are the minimum and maximum values ​​among samples of the same type of monitoring parameters, eliminating the dimensional differences between different physical parameters.

[0039] With time window Extract time-series data and construct dimensions as follows: The sample set, The total number of samples, For a single sample time step, 5 corresponds to five types of monitoring parameters: temperature, pressure, flow rate, sound wave, and hydrogen concentration.

[0040] The calculation expression for the convolutional layer of the 1D-CNN model is: (2); In the formula: This represents the current number of convolutional network layers. For the next layer of network; For the first Layer Each feature map at location The output value; The ReLU (Rectified Linear Unit) activation function takes values ​​of... ; The kernel length is [length]. Input values ​​for the temporal positions corresponding to the feature maps of the previous layer; These are the convolution kernel weight parameters; This is the feature map bias term.

[0041] The core calculation formula for the Bi-LSTM module is as follows: Forgotten Gate: (3); Input Gate: (4); Candidate cell status: (5); Cell status update: (6); Output gate: (7); One-way hidden state: (8); Bidirectional feature fusion: (9); In the formula, For time series data, the time step; The state was hidden in the previous moment; Input data for multidimensional monitoring at the current moment; These are the output values ​​of the forget gate, input gate, and output gate, respectively. These represent the cell state and the candidate cell state, respectively. These are the weight matrices corresponding to the forget gate, input gate, output gate, and cell state, respectively. These are the bias terms corresponding to the forget gate, input gate, output gate, and cell state, respectively. These are the forward and backward LSTM hidden states, respectively. This is a hidden state for bidirectional fusion.

[0042] Formula for probability output of fully connected layer: (10); In the formula, Let be the leakage probability distribution vector for the four heat exchangers; The hidden state is the result of bidirectional fusion at the last time step; The weights and biases corresponding to the output gates of the fully connected layer; The Softmax Function (normalization exponential function) normalizes probabilities, ensuring that the sum of the probabilities of each class is 1.

[0043] The model training uses the cross-entropy loss function: (11); In the formula, The value of the loss function; For the first The actual label of the heat exchanger shows a leak mark as 1 and a no-leak mark as 0. For the model to predict the first Probability of leakage in a heat exchanger.

[0044] Model structure configuration: 1D-CNN is set with 2 layers of convolution + max pooling, with 32 and 64 convolution kernels respectively, kernel length of 3, stride of 1, and pooling window length of 2; Bi-LSTM has 128 hidden units; the fully connected layer has 64 and 4 neurons; training uses Adam optimizer, learning rate of 0.001, batch size of 32, 100 iterations, and introduces an early stopping mechanism.

[0045] In the operating scenario of the evaporator in a solar thermal power plant, using the pressure, temperature, and flow parameters under normal system conditions, the 1D-CNN-Bi-LSTM model was used to monitor the state of shell-side outlet temperature reduction, pressure surge, and local abnormal flow fluctuations, accurately determining the evaporator tube-side leakage fault, verifying the reliability of the model, and avoiding the damage to the equipment caused by traditional pressure testing.

[0046] like Figure 5 As shown, this embodiment simulates an evaporator leakage condition, with the leakage amount controlled within... Based on the characteristics of shell-side temperature drop, pressure surge, and abnormal flow fluctuation, the model accurately identifies leaks from the evaporator tube side to the shell side, and can locate the leak point without cutting and disassembling the pipes.

[0047] Example 3 In another preferred embodiment, based on embodiments 1 and 2, this embodiment provides a method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model, employing a parallel operation mode of two evaporators, such as... Figure 4As shown, molten salt and hot salt enter the superheater and reheater in sequence, then flow into the two evaporators and converge to the preheater, and finally flow out as cold salt. Each heat exchanger inlet and outlet is independently equipped with parameter monitoring channels to collect five types of time-series parameters.

[0048] In the operation scenario of the evaporator of the solar thermal power plant's heat storage and exchange system, under the condition of normal operation after the equipment is put into operation, the equipment exhibits tube-side tube bundle leakage. The tube-side tube bundle leakage fault was confirmed by disassembling the manhole door for inspection.

[0049] Data preprocessing, sample construction, 1D-CNN, Bi-LSTM model structure, calculation formulas and training parameters are all consistent with those in Example 2; In the operating scenario of the evaporator in a solar thermal power plant, using the pressure, temperature, and flow parameters under normal system conditions, the 1D-CNN-Bi-LSTM model was used to monitor the state of shell-side outlet temperature reduction, pressure surge, and local abnormal flow fluctuations, accurately determining the evaporator tube-side leakage fault, verifying the reliability of the model, and avoiding the damage to the equipment caused by traditional pressure testing.

[0050] like Figure 5 As shown, under the condition of coupled flow field of dual evaporators, the model of the present invention can effectively suppress the flow field coupling interference between devices, accurately identify the evaporator device that has leaked, and adapt to the SGS tightness test scenario of multiple devices operating in parallel and coupled.

[0051] Example 4 In another preferred embodiment, based on embodiments 1 to 3, this embodiment provides a leak detection method for the SGS system of a concentrated solar power plant based on a 1D-CNN-Bi-LSTM model. This embodiment performs leak detection on the preheater of the SGS system of the concentrated solar power plant, with the flow path layout being the same. Figure 3 A complete set of sensors is installed at the inlet and outlet of the preheater shell to collect full-dimensional time-series monitoring data.

[0052] In the operation scenario of the evaporator of the solar thermal power plant's heat storage and exchange system, under the condition of normal operation after the equipment is put into operation, the equipment exhibits tube-side tube bundle leakage. The tube-side tube bundle leakage fault was confirmed by disassembling the manhole door for inspection.

[0053] The data processing flow, 1D-CNN, Bi-LSTM model structure, calculation formula and training parameters are the same as in Example 2.

[0054] In the operation scenario of a solar thermal power plant preheater, based on the historical parameters of pressure, temperature, and flow rate during normal system operation, a 1D-CNN-Bi-LSTM model was used to monitor the shell-side outlet temperature decrease, pressure increase, and local abnormal flow fluctuations, which determined that there was a leak on the tube side of the preheater. On-site inspection confirmed that there was a leak in 3 tube bundles and 1 blind flange, verifying the accuracy and reliability of the model.

[0055] like Figure 6 As shown, when the preheater tube bundle and blind plate leak from the tube side to the shell side, the model captures the correlation features of shell-side outlet temperature decrease, pressure increase, and local abnormal flow fluctuation, and outputs the leakage probability via Softmax to accurately determine the preheater leakage. The on-site verification by opening the manhole door shows that the results are completely consistent with the model's diagnosis, verifying the positioning accuracy and reliability of the method of this invention in SGS's actual engineering scenarios.

[0056] In the preferred embodiment, in step 1, five types of detection devices are respectively installed at the inlet and outlet of each heat exchanger on the shell side: differential pressure sensor, vortex flow meter, thermocouple, acoustic emission sensor, and infrared gas analyzer. These devices simultaneously collect five monitoring parameters: shell side temperature, pressure, flow rate, acoustic wave, and hydrogen concentration of the medium. This deployment method achieves comprehensive perception of the medium state on the shell side of each heat exchanger. The five parameters complement each other; acoustic wave and hydrogen concentration are highly sensitive to leaks, pressure and temperature reflect macroscopic thermodynamic changes, flow rate reflects abnormal medium transport, and the simultaneous acquisition of multi-source heterogeneous data provides a complete information foundation for accurate subsequent model judgment, effectively avoiding the risk of misjudgment caused by single-parameter monitoring.

[0057] In the preferred embodiment, the sampling frequency of each sensor is set to 100Hz, and each heat exchanger inlet and outlet is independently configured with a monitoring channel to collect five types of monitoring parameters corresponding to the inlet and outlet of the superheater, reheater, evaporator, and preheater shell side. The high sampling frequency of 100Hz can capture the transient characteristic changes in the early stage of leakage, and the independent monitoring channel ensures that the data of each heat exchanger are not mixed, which facilitates the model to independently distinguish each device, thereby achieving accurate differentiation of the leakage source under the condition of four-electrode coupled flow field.

[0058] In the preferred embodiment, the outlier handling in step 2 employs the 3σ criterion, which, based on the fluctuation range threshold of normal operating data, eliminates abnormal monitoring data that deviates from the normal range due to sensor malfunctions or external environmental interference. This processing method automatically identifies and filters out outlier data using statistical methods, ensuring the reliability of the data quality input to the model, avoiding false leakage alarms caused by occasional sensor malfunctions, and improving the stability and reliability of the model in actual operating environments.

[0059] In the preferred embodiment, the normalization process in step 2 employs range normalization to uniformly process the monitoring parameters, eliminating differences in dimensions and numerical magnitudes between different parameters and ensuring that multi-source monitoring data fall within the same numerical range. After eliminating dimensional differences, each feature dimension of the model participates in learning with equal weight, avoiding training bias caused by significant differences in the numerical magnitudes of parameters such as pressure and temperature, accelerating model convergence, and improving the balance of feature fusion.

[0060] In a preferred embodiment, step 2 involves constructing a standardized time-series sample set. Specifically, this involves extracting time-series data with a fixed time window length and constructing a sample set with a three-dimensional structure including sample quantity, time step size, and parameter categories to meet the input requirements of the time-series neural network model. Building for time windows The dimensional samples not only preserve the complete temporal evolution of the leakage signal from its occurrence to its spread, but also ensure that the sample format is strictly matched with the input layer of the 1D-CNN-Bi-LSTM model, so as to ensure that the model can fully learn the temporal dependence of the leakage features.

[0061] In the preferred embodiment, the 1D-CNN-Bi-LSTM fusion model in step 3 sequentially includes an input layer, a 1D-CNN module, a Bi-LSTM module, a fully connected layer, and a probability output layer. The 1D-CNN module uses a two-layer convolutional structure with max pooling to extract local leakage features such as pressure surges, flow fluctuations, and acoustic anomalies. The Bi-LSTM module employs a single-layer bidirectional long short-term memory structure to capture the temporal dynamic evolution and correlation patterns of leakage features. The fully connected layer uses a two-level neuron structure, with the output dimension matching the classification and recognition requirements of four heat exchangers. This model architecture organically combines local feature extraction with bidirectional temporal modeling. The front-end convolution quickly locates abnormal segments, while the back-end bidirectional LSTM tracks the dynamic propagation process of the leakage signal. Finally, the fully connected layer achieves accurate four-class classification output, resulting in significantly better overall recognition accuracy and robustness than a single model.

[0062] In the preferred embodiment, the model training of the 1D-CNN-Bi-LSTM fusion model in step 3 employs a cross-entropy loss evaluation mechanism. This mechanism compares the model's predicted leakage probability distribution with the actual leakage label of the equipment, quantifies the deviation between the prediction results and the actual operating conditions, and iteratively optimizes the model's internal parameters with the goal of minimizing the deviation. The cross-entropy loss function has strong discriminative power in penalizing the probability distribution, effectively driving the model to converge quickly to the optimal classification boundary, ensuring a high degree of consistency between the predicted probability and the actual leakage state, and guaranteeing the reliability of the output results.

[0063] In the preferred embodiment, the model training employs the Adam optimizer with a fixed learning rate, using batch gradient descent for iterative training, limiting the batch size and maximum number of iterations. An early stopping mechanism is implemented, terminating training when the validation set accuracy shows no improvement for several consecutive rounds, thus avoiding overfitting. The Adam optimizer adaptively adjusts the learning rate, combining the stable update characteristics of batch gradient descent with overfitting protection from the early stopping mechanism, enabling the model to achieve optimal generalization performance within a limited number of training rounds and preventing a decrease in detection accuracy due to overtraining.

[0064] In the preferred embodiment, during the model validation stage of the 1D-CNN-Bi-LSTM fusion model described in step 3, an SGS system four-device simulation platform is built. Gradient leakage flow ranges are set to simulate multiple leakage conditions, and sufficient normal and leakage condition samples are collected. The training set and validation set are divided into a fixed ratio, and the validation samples are used to assess the model's recognition accuracy, false alarm rate, and false negative rate, thus completing the model structure and hyperparameter optimization tuning. Multi-condition simulation verification of gradient leakage: The model was tested on 10,000 samples, which fully demonstrated its detection capability under different leakage levels. The false alarm rate and false negative rate were effectively controlled.

[0065] In the preferred embodiment, the 1D-CNN module in step 3 employs a sliding sampling method along the one-dimensional temporal signal using convolutional kernels to perform weighted fusion and bias correction on local features. It retains effective anomalous features and suppresses steady-state noise through nonlinear activation filtering. Multiple convolutional kernels are used in parallel to extract specific leakage features for different monitoring parameters, forming a multi-channel feature map. 32 and 64 convolutional kernels respectively extract low-order and high-order local features, and the ReLU activation function filters steady-state background noise, enabling the model to focus on the abrupt changes and fluctuations caused by leakage, significantly improving the signal-to-noise ratio and discriminative power of feature extraction.

[0066] In the preferred embodiment, the Bi-LSTM module described in step 3 is configured with a forget gate, an input gate, a cell state unit, and an output gate working collaboratively. The forget gate discards irrelevant historical temporal information, the input gate incorporates the current valid leakage features and updates the memory state, the cell state unit continuously stores the temporal evolution information of the entire leakage process, and the output gate outputs valid temporal features according to the memory state. By splicing and fusing forward temporal deduction and backward temporal backtracking features, the temporal dependency relationship of the leakage signal is fully represented. The bidirectional structure simultaneously captures the forward trend of temperature decrease and pressure increase after the leakage occurs and the backtracking of the precursor information before the leakage, adapting to the entire dynamic evolution process of the leakage signal from occurrence to diffusion, and significantly enhancing the model's ability to capture temporal correlation features.

[0067] In the preferred embodiment, the fully connected layer in step 4 performs linear mapping and bias compensation on the fused features of the Bi-LSTM output. Normalization converts the discrimination scores of each heat exchanger into a normalized probability distribution, using the probability values ​​to characterize the confidence level of leakage in the corresponding heat exchanger. Softmax normalization ensures that the sum of the leakage probabilities of the four heat exchangers is 1, with the maximum probability directly corresponding to the leaking device. The output results are intuitive and clear, facilitating rapid decision-making by maintenance personnel and providing a quantitative confidence basis for subsequent targeted maintenance.

[0068] In the preferred embodiment, when the tube-side medium of the heat exchanger leaks to the shell side in step 4, it will cause a correlation of changes, including a decrease in shell-side medium temperature, an abnormal increase in pressure, and disordered fluctuations in local flow. The model relies on the differences in anomaly characteristics of multi-parameter coupling to achieve precise location of leakage sources in multiple heat exchangers under a coupled structure, replacing the destructive detection method of traditional pipe cutting and disassembly, and pressure testing of each unit. This location method allows maintenance personnel to directly open the manhole of the leaking heat exchanger for targeted repair, avoiding secondary damage caused by repeated cutting and welding of non-leaking equipment, significantly shortening downtime and reducing maintenance costs. In actual cases, leaks in the evaporator and preheater have been successfully detected and verified through on-site disassembly.

[0069] In summary, this invention proposes a leak detection method for the SGS system of a solar thermal power plant based on a 1D-CNN-Bi-LSTM model. This method effectively solves the technical problems of existing leak detection methods, which cannot accurately locate leaking equipment under the coupled flow field of the preheater, evaporator, superheater, and reheater in the SGS system of a solar thermal power plant. These methods rely solely on pipe cutting and dismantling, and pressure testing of each unit, resulting in long downtime, high maintenance costs, and easy secondary seal damage to the equipment.

[0070] This invention applies a deep fusion model of 1D-CNN and Bi-LSTM to the SGS "four-element" leak detection field, utilizing non-invasive multi-dimensional monitoring data from the shell side to accurately locate leaking heat exchangers, breaking through the traditional technical route that relies on destructive disassembly for investigation. Addressing the technical bottleneck of being unable to distinguish leak sources under the coupled flow field of the "four elements," a fusion architecture is constructed to simultaneously extract local leak features and bidirectional temporal dependencies, enabling the classification and identification of the four heat exchangers and the output of leak probability.

[0071] At the data acquisition level, five types of sensors—pressure, flow, temperature, sound wave, and medium composition—are simultaneously deployed on the shell side of the four components and collect data at a frequency of 100Hz to construct a... The three-dimensional time series sample set forms a standardized data system that adapts to the model input.

[0072] The method of this invention designs a dual-module collaborative architecture that uses 1D-CNN to extract local leakage features such as pressure mutations and acoustic peaks, and Bi-LSTM to capture the forward and backward temporal evolution of leakage signals. This not only solves the problem that a single CNN cannot model the dynamic evolution of time series, but also overcomes the problem that a single LSTM is not sensitive enough to local mutation features, thus significantly improving the leakage identification accuracy.

[0073] Simultaneously, by utilizing the abnormal characteristics of multiple parameters such as temperature drop, pressure surge, and abnormal flow fluctuation caused by the leakage of medium from the pipe side to the shell side, and combining the complementarity of multi-sensor data, the leakage difference identification of the "four devices" is realized. The detection mode is changed from "disassembly first and then location" to "location first and then maintenance", which greatly reduces the false alarm rate and the missed alarm rate.

[0074] In addition, this method introduces the 3σ criterion for outlier removal, Min-Max normalization to eliminate dimensional differences, and a time window. Preprocessing strategies such as sample construction, combined with the training paradigm of cross-entropy loss function and Adam optimizer, and early stopping mechanism to prevent overfitting, form a complete and rigorous testing technology system from data acquisition to model deployment. It has been verified by 10,000 sets of samples and confirmed by on-site disassembly cases of evaporators and preheaters, and has practical engineering application value.

Claims

1. A method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model, characterized in that, Includes the following steps: Step 1: Real-time acquisition of multi-dimensional monitoring data on the shell side of the four components of the SGS system of the solar thermal power plant, namely the preheater, evaporator, superheater and reheater. Step 2: The collected raw monitoring data are sequentially subjected to outlier removal, normalization, and fixed time window sample extraction to construct a standardized time series sample set; Step 3: Construct a 1D-CNN-Bi-LSTM fusion model. Input the standardized time series sample set into the model, extract local leakage features of the monitoring signal through the 1D-CNN module, and use the Bi-LSTM module to mine the bidirectional temporal correlation of features and complete feature fusion. Step 4: Input the fused features into the fully connected layer, and output the leakage probability distribution of each heat exchanger after probability normalization. The heat exchanger with the maximum probability is identified as the leaking device. Step 5: Automatically generate a leak location report, outputting the leaking device number, fault confidence level, and key parameter change curves, to achieve non-destructive tightness testing and precise location of leaking devices using the SGS system.

2. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: In step 1, five types of detection devices are installed at the inlet and outlet of each heat exchanger on the shell side of the four devices: differential pressure sensor, vortex flow meter, thermocouple, acoustic emission sensor, and infrared gas analyzer. These devices simultaneously collect five types of monitoring parameters: shell side temperature, pressure, flow rate, sound wave, and hydrogen concentration of the medium.

3. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 2, characterized in that: Each heat exchanger is equipped with an independent monitoring channel at its inlet and outlet, which collects five types of monitoring parameters corresponding to the inlet and outlet of the superheater, reheater, evaporator, and preheater shell side.

4. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: The outlier handling described in step 2 adopts the 3σ criterion, which eliminates abnormal monitoring data that deviates from the normal range due to sensor malfunctions or external environmental interference, based on the fluctuation range threshold of the normal operating data.

5. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: The normalization process described in step 2 uses range normalization to uniformly process the monitoring parameters, eliminating differences in dimensions and numerical magnitudes between different parameters, so that the multi-source monitoring data are in the same numerical range.

6. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: In step 2, a standardized time series sample set is constructed. Specifically, time series data is extracted with a fixed time window length, and a sample set containing a three-dimensional structure of sample quantity, time step size, and parameter category is constructed to adapt to the input requirements of the time series neural network model.

7. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: The 1D-CNN-Bi-LSTM fusion model described in step 3 includes an input layer, a 1D-CNN module, a Bi-LSTM module, a fully connected layer, and a probability output layer. The 1D-CNN module uses a two-layer convolutional structure with max pooling to extract local leakage features such as pressure surges, flow fluctuations, and acoustic anomalies. The Bi-LSTM module uses a single-layer bidirectional long short-term memory structure to capture the temporal dynamic evolution and correlation of leakage features. The fully connected layer has a two-level neuron structure, and the output dimension matches the classification and recognition requirements of four heat exchangers.

8. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: The 1D-CNN-Bi-LSTM fusion model described in step 3 uses a cross-entropy loss evaluation mechanism for model training. It compares the model's predicted leakage probability distribution with the actual leakage label of the equipment, quantifies the deviation between the prediction results and the actual working conditions, and iteratively optimizes the model's internal parameters with the goal of minimizing the deviation.

9. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 8, characterized in that: The model training uses the Adam optimizer with a fixed learning rate and employs batch gradient descent for iterative training, limiting the training batch size and the maximum number of iterations. An early stopping mechanism is set up to terminate training when the accuracy of the validation set does not improve for several consecutive rounds, thus avoiding the problem of model overfitting.

10. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: In step 3, the model validation stage of the 1D-CNN-Bi-LSTM fusion model involves building an SGS system four-device simulation platform, setting gradient leakage flow ranges to simulate multiple leakage conditions, and collecting sufficient normal and leakage condition samples. The training set and validation set are divided into a fixed ratio, and the model recognition accuracy, false alarm rate, and false negative rate are evaluated using the validation samples. The model structure and hyperparameters are then optimized and tuned.

11. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: The 1D-CNN module described in step 3 uses a sliding sampling method along the one-dimensional time-series signal to perform weighted fusion and bias correction on local features. It retains effective abnormal features and suppresses steady-state noise through nonlinear activation screening. It extracts exclusive leakage features of different monitoring parameters in parallel through multiple convolutional kernels to form a multi-channel feature map.

12. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: In step 3, the Bi-LSTM module is configured with a forget gate, an input gate, a cell state unit, and an output gate working together. The forget gate discards irrelevant historical time-series information, the input gate incorporates the current effective leakage features and updates the memory state, the cell state unit continuously stores the time-series evolution information of the entire leakage process, and the output gate outputs the effective time-series features according to the memory state. By splicing and fusing forward time-series deduction and backward time-series backtracking features, the time-series dependency of the leakage signal is fully represented.

13. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: In step 4, the fully connected layer performs linear mapping and bias compensation on the fused features of the Bi-LSTM output. Through normalization processing, the discrimination scores of each heat exchanger are converted into normalized probability distributions, and the probability values ​​are used to characterize the confidence level of the corresponding heat exchanger leakage.

14. The method for airtightness testing of a solar thermal power plant SGS system based on a 1D-CNN-Bi-LSTM model according to claim 1, characterized in that: When the tube-side medium of the heat exchanger leaks to the shell side as described in step 4, it will cause a correlation change in the shell-side medium temperature, abnormal pressure increase, and disordered local flow fluctuation. The model relies on the differences in abnormal characteristics of multi-parameter coupling to achieve accurate location of leakage sources of multiple heat exchangers under the coupled structure, replacing the destructive detection method of traditional pipe cutting and disassembly and pressure testing of each unit.

Images