Defect prediction method, device and equipment for heat transfer tube of condenser

Abstract

The invention discloses a condenser heat transfer tube defect prediction method, device and equipment, and relates to the technical field of defect prediction, and the method comprises the steps: obtaining temperature field data of the surface of a condenser heat transfer tube, refrigerant flow and pressure difference of a refrigerant inlet and a refrigerant outlet, and a leakage detection signal of the refrigerant; performing temperature field decomposition according to the temperature field data to obtain a target space mode and a time evolution curve, and performing simulation mapping inversion according to the refrigerant flow and the pressure difference to obtain a stock state in a pipe; constructing a high-dimensional feature vector, and performing graph neural network processing on the high-dimensional feature vector to obtain a spatial enhancement feature; identifying an abnormal mode in the target space mode based on the space enhancement feature, and generating a residual temperature field according to the abnormal mode; and performing defect classification and defect evolution prediction according to the residual temperature field, the leakage detection signal and the modal change rate of the target space modal. According to the invention, high-precision prediction of the defect type and the defect evolution trend of the heat transfer tube can be realized.

Description

Technical Field

[0001] This application relates to the field of defect prediction technology, specifically to a method, apparatus, and equipment for predicting defects in condenser heat transfer tubes. Background Technology

[0002] As a core component of a thermodynamic cycle system, the condenser's heat transfer efficiency directly impacts the system's energy efficiency and safe operation. In critical industrial sectors such as nuclear power plants and ships, condenser heat transfer tubes are exposed to high-temperature, high-pressure, and corrosive media environments for extended periods, making them prone to defects such as corrosion thinning, scaling and blockage, micro-crack propagation, and refrigerant leakage. If these defects are not detected and addressed promptly, they can lead to a sharp drop in heat transfer efficiency and even trigger unplanned shutdowns of nuclear power plants or loss of power in ships, causing significant economic losses and safety hazards.

[0003] Currently, commonly used detection methods mainly rely on periodic shutdowns for maintenance or localized inspections, which suffer from problems such as large blind spots and poor timeliness. The efficiency and accuracy of defect detection are insufficient to meet the demands for online, real-time, and high-precision early warning of defects. Among related technologies, non-destructive testing technology based on infrared thermal imaging is gradually gaining attention. However, infrared thermal imaging technology is mostly limited to static temperature field analysis, making it difficult to capture dynamic evolution characteristics. Especially under complex operating conditions, its defect identification capability is insufficient, still failing to meet the needs of early defect warning. Summary of the Invention

[0004] This application provides a method, apparatus, and equipment for predicting defects in condenser heat transfer tubes, which can solve the technical problem that the existing technology cannot meet the requirements for early warning of defects.

[0005] In a first aspect, this application provides a method for predicting defects in condenser heat transfer tubes, the method comprising: Acquire temperature field data on the surface of the condenser heat transfer tubes, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signals; Based on the above temperature field data, temperature field decomposition is performed to obtain the target spatial mode and time evolution curve. Based on the above refrigerant flow rate and pressure difference, simulation mapping inversion is performed to obtain the internal quantity state of the pipe. Based on the aforementioned time evolution curve, pipe internal quantity status, refrigerant flow rate and pressure difference, and leakage detection signal, a high-dimensional feature vector is constructed, and the aforementioned high-dimensional feature vector is processed by a graph neural network to obtain spatially enhanced features. Based on the aforementioned spatial enhancement features, anomalous modes in the aforementioned target spatial modes are identified, and residual temperature fields are generated based on the aforementioned anomalous modes; Based on the residual temperature field, the leakage detection signal, and the modal change rate of the target space mode, defect classification and defect evolution prediction are performed to obtain the defect type and defect evolution trend.

[0006] In conjunction with the first aspect, in one implementation, temperature field decomposition is performed based on the aforementioned temperature field data to obtain the target spatial modes and time evolution curves, specifically including: The temperature field data is reconstructed into a snapshot matrix, and a sliding time window is constructed. The sliding time window is used to slide and truncate on the snapshot matrix to obtain multiple sub-snapshot matrices. Singular value decomposition is performed on each of the above sub-snapshot matrices to obtain the spatial modes and temporal coefficients corresponding to different time windows; Modal registration is performed on the spatial modes of adjacent time windows to obtain the registered spatial modes; Determine the energy weights of each registered spatial mode, and sort the registered spatial modes from high to low according to the energy weights to obtain the registered spatial mode sequence; The first N spatial modes in the registered spatial mode sequence are selected as target spatial modes, and time evolution curves are generated based on the target spatial modes and their corresponding time coefficients, where N is a positive integer.

[0007] In conjunction with the first aspect, in one implementation, the above-mentioned modal registration of spatial modes in adjacent time windows to obtain registered spatial modes specifically includes: The spatial modes of two adjacent time windows are represented as two vector sequences respectively; Calculate the Euclidean distance between each pair of corresponding vectors in the two vector sequences to obtain the initial distance matrix, and optimize the distance metric to obtain the target distance matrix; Based on dynamic programming, the optimal matching path is obtained by progressively searching from the starting point of the target distance matrix to the endpoint. Based on the above optimal matching path, the alignment relationship of corresponding vectors in adjacent time window spatial modes is determined, so as to perform alignment processing on the spatial modes of adjacent time windows and obtain the registered spatial modes.

[0008] In conjunction with the first aspect, in one implementation, the above-mentioned modal registration of spatial modes in adjacent time windows to obtain registered spatial modes specifically includes: Expand the spatial modes of adjacent time windows into vector form to obtain the spatial mode vector of adjacent time windows; Calculate the correlation coefficients between all pairs of spatial mode vectors in adjacent time windows, and construct a correlation matrix; Based on the above correlation matrix, the maximum correlation matching algorithm is adopted. Starting from the first time window, the algorithm searches for the spatial mode vector with the highest correlation in the subsequent time window for each spatial mode vector and matches it until the spatial mode vector of the last time window is matched, thus obtaining the matched spatial mode vector pair. Based on the aforementioned spatial mode vector pairs, the correspondence between spatial modes in adjacent time windows is determined, and the spatial modes in adjacent time windows are aligned and adjusted to obtain the registered spatial modes.

[0009] In conjunction with the first aspect, in one implementation, the internal quantity state of the refrigerant is obtained by simulation mapping inversion based on the aforementioned refrigerant flow rate and pressure difference, specifically including: Construct a thermal-fluid coupling model for the condenser heat transfer tubes; Based on the refrigerant flow rate and pressure difference, as well as the tube length and inner diameter of the condenser heat transfer tube, the above heat-fluid coupling model is solved by numerical simulation to obtain the flow velocity distribution and pressure distribution of the refrigerant inside the tube. Based on the above flow velocity and pressure distribution, the internal quantity status of the pipe can be calculated by inversion.

[0010] In conjunction with the first aspect, in one implementation, the aforementioned high-dimensional feature vectors are processed using a graph neural network to obtain spatially enhanced features, specifically including: Map the above high-dimensional feature vectors into a graph structure; The above graph structure is subjected to feature extraction by graph convolutional layers to obtain spatially enhanced features.

[0011] In conjunction with the first aspect, in one implementation, identifying anomalous modes in the target spatial modes based on the aforementioned spatial enhancement features, and generating a residual temperature field based on the anomalous modes, specifically includes: The aforementioned spatial enhancement features are fused with the target spatial modality to obtain a fused feature matrix. The above-mentioned fused feature matrix is ​​scanned by a preset anomaly detection threshold range to identify abnormal modes that exceed the threshold range. Calculate the difference between the above-mentioned anomalous mode and the above-mentioned target space mode, and map the above-mentioned difference into a residual temperature field.

[0012] In conjunction with the first aspect, in one implementation, based on the aforementioned residual temperature field, the aforementioned leakage detection signal, and the modal change rate of the target spatial mode, defect classification and defect evolution prediction are performed to obtain the defect type and defect evolution trend, specifically including: Anomalies are extracted from the residual temperature field, and the leakage detection signal is time-series aligned with the anomalies to construct a joint feature vector. The anomalies include extreme values ​​of temperature deviation, area of ​​temperature anomaly region, and spatial distribution characteristics. Calculate the rate of change of the above target spatial modes in the time dimension to obtain the modal change rate; The joint feature vector and the modal change rate are input into the pre-trained defect classification model to identify defects and determine the defect type. The aforementioned joint feature vector, modal change rate, and historical defect evolution data are input into the defect evolution prediction model to generate a defect evolution trend. The defect evolution trend is quantitatively characterized by the growth rate of extreme temperature deviations, the expansion rate of abnormal area, and the migration path of spatial distribution.

[0013] Secondly, this application provides a condenser heat transfer tube defect prediction device, the device comprising: The preprocessing module is used to acquire temperature field data on the surface of the condenser heat transfer tubes, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signals. The decomposition module is used to decompose the temperature field based on the above temperature field data to obtain the target spatial mode and time evolution curve; The inversion module is used to perform simulation mapping inversion based on the above refrigerant flow rate and pressure difference to obtain the internal quantity status of the pipe. The processing module is used to construct a high-dimensional feature vector based on the above time evolution curve, the internal quantity status of the pipe, the refrigerant flow rate and pressure difference, and the leakage detection signal, and to perform graph neural network processing on the above high-dimensional feature vector to obtain spatially enhanced features. The identification module is used to identify anomalous modes in the target spatial modes based on the aforementioned spatial enhancement features, and to generate a residual temperature field based on the aforementioned anomalous modes. The prediction module is used to classify defects and predict their evolution based on the residual temperature field, the leak detection signal, and the modal change rate of the target spatial mode, thereby obtaining the defect type and the defect evolution trend.

[0014] Thirdly, this application provides a condenser heat transfer tube defect prediction device, which includes a processor, a memory, and a condenser heat transfer tube defect prediction program stored in the memory and executable by the processor. When the condenser heat transfer tube defect prediction program is executed by the processor, it implements the steps of the condenser heat transfer tube defect prediction method described above.

[0015] The beneficial effects of the technical solutions provided in this application include: This application integrates temperature field data, refrigerant flow rate and pressure difference at refrigerant inlet and outlet, and leakage detection signals to decompose the temperature field and extract the target spatial mode. Combined with the inversion of the tube state, it achieves high-precision prediction of the type of heat transfer tube defects and the trend of defect evolution, including location and development rate. This effectively avoids performance degradation and safety accidents caused by heat transfer tube defects, improves the operational reliability and service life of the condenser, and solves the technical problem of difficulty in meeting the early warning requirements of defects in related technologies. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the condenser heat transfer tube defect prediction method of this application. Figure 2 This is a schematic diagram of the functional modules of an embodiment of the condenser heat transfer tube defect prediction device of this application. Figure 3 This is a schematic diagram of the hardware structure of the condenser heat transfer tube defect prediction device involved in the embodiments of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0018] In a first aspect, embodiments of this application provide a method for predicting defects in condenser heat transfer tubes.

[0019] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the condenser heat transfer tube defect prediction method of this application. The above-mentioned condenser heat transfer tube defect prediction method includes: S1. Acquire temperature field data of the condenser heat transfer tube surface, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signal; S2. Based on the above temperature field data, perform temperature field decomposition to obtain the target spatial mode and time evolution curve, and perform simulation mapping inversion based on the above refrigerant flow rate and pressure difference to obtain the internal quantity state of the pipe; S3. Based on the above time evolution curve, pipe internal quantity status, refrigerant flow rate and pressure difference, and leakage detection signal, a high-dimensional feature vector is constructed, and the above high-dimensional feature vector is processed by a graph neural network to obtain spatial enhancement features. S4. Based on the above spatial enhancement features, identify the anomalous modes in the above target spatial modes, and generate a residual temperature field based on the above anomalous modes; S5. Based on the above residual temperature field, the above leakage detection signal, and the modal change rate of the target space mode, perform defect classification and defect evolution prediction to obtain the defect type and defect evolution trend.

[0020] In this embodiment, temperature field data of the condenser heat transfer tube surface, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signal are acquired. Temperature field decomposition is performed based on the temperature field data to obtain the target spatial mode and time evolution curve. Simulation mapping inversion is then performed based on the refrigerant flow rate and pressure difference to obtain the internal quantity state within the tube. A high-dimensional feature vector is constructed based on the time evolution curve, internal quantity state, refrigerant flow rate and pressure difference, and leakage detection signal. This high-dimensional feature vector is then processed using a graph neural network to obtain spatial enhancement features. Abnormal modes in the target spatial mode are identified based on these spatial enhancement features, and a residual temperature field is generated based on these abnormal modes. Defect classification and defect evolution prediction are performed based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode to obtain the defect type and defect evolution trend.

[0021] By integrating temperature field data, refrigerant flow rate and pressure difference at refrigerant inlet and outlet, and leakage detection signals, the target spatial mode is extracted through temperature field decomposition. Combined with in-tube state inversion, high-precision prediction of heat transfer tube defect types, defect evolution trends including location and development rate is achieved. This effectively avoids performance degradation and safety accidents caused by heat transfer tube defects, improves the operational reliability and service life of the condenser, and solves the technical problem of difficulty in meeting the early warning requirements of defects in related technologies.

[0022] In this embodiment, the condenser heat transfer tube is a key component in the operation of the condenser, and its surface temperature distribution and internal refrigerant state directly reflect the health status of the heat transfer tube. Due to long-term operation and environmental factors, the heat transfer tube may develop defects such as tube wall corrosion, refrigerant leakage, and decreased heat transfer efficiency. If these defects are not detected and addressed in a timely manner, they will seriously affect the performance and safety of the condenser.

[0023] Based on the above embodiments, in one embodiment, step S1 further includes preprocessing the acquired temperature field data, refrigerant flow rate and pressure difference, and leakage detection signal, so as to enable more accurate defect prediction in the subsequent process using the preprocessed temperature field data, preprocessed refrigerant flow rate and pressure difference, and preprocessed leakage detection signal.

[0024] Preferably, the temperature field data is an infrared thermogram sequence; the leakage detection signal is a leakage intensity signal.

[0025] Understandably, an infrared thermal image sequence is a continuous image of the temperature distribution on the surface of a condenser heat transfer tube, acquired using infrared thermal imaging technology. It can visually reflect the temperature changes on the heat transfer tube surface. During the acquisition process, it is necessary to ensure the accuracy and stability of the infrared thermal imager to obtain an accurate and reliable infrared thermal image sequence.

[0026] In this embodiment, the refrigerant flow rate, differential pressure, and refrigerant leakage intensity signals at the refrigerant inlet and outlet are also important monitoring parameters, reflecting changes in the refrigerant's state inside the heat transfer tube. Accurate measurement of refrigerant flow rate and differential pressure relies on high-precision flow meters and differential pressure sensors, which require regular calibration to ensure data reliability. The refrigerant leakage intensity signal can be acquired using a specialized leak detection device, which must possess high sensitivity and anti-interference capabilities to accurately capture leakage signals in complex environments.

[0027] In this embodiment, by preprocessing the infrared thermal image sequence, refrigerant flow rate, pressure difference, and leakage intensity signals, noise interference can be effectively eliminated, data quality can be improved, and a synchronous data stream can be formed.

[0028] Specifically, the above preprocessing process includes: First, an infrared thermal image sequence of the surface of the condenser heat transfer tube is continuously acquired using an infrared thermal imager at a preset sampling frequency; the refrigerant flow rate at the refrigerant inlet and outlet is measured in real time using a flow sensor; the pressure difference across the condenser heat transfer tube is measured in real time using a differential pressure sensor; and the refrigerant leakage intensity signal is monitored in real time using a leakage detection device.

[0029] Then, the above infrared thermal image sequence is subjected to non-uniformity correction, bad pixel repair and spatial registration to obtain the registered infrared thermal image sequence; the above refrigerant flow rate, above pressure difference and above leakage intensity signals are denoised and normalized to obtain normalized refrigerant flow rate, normalized pressure difference and normalized leakage intensity signals.

[0030] Finally, the above-registered infrared thermal image sequence, the above-normalized refrigerant flow rate, the above-normalized pressure difference, and the above-normalized leakage intensity signal are aligned according to a preset timestamp to obtain the pre-processed infrared thermal image sequence, the pre-processed refrigerant flow rate, the pre-processed pressure difference, and the pre-processed leakage intensity signal.

[0031] In this embodiment, a high-resolution infrared thermal imager is deployed to continuously acquire a sequence of temperature field images (i.e., an infrared thermal image sequence) of the condenser's outer wall surface at a preset sampling frequency, where the preset sampling frequency can be 1Hz. High-precision mass flow meters and differential pressure sensors are installed at the condenser inlet and outlet to measure the refrigerant flow rate and the pressure difference across the heat transfer tubes in real time, respectively, and the refrigerant flow rate is recorded. With pressure difference The refrigerant leakage intensity signal is monitored in real time by a high-sensitivity leak detection device arranged around the condenser. This device can capture even minute refrigerant leaks. For example, a laser absorption spectroscopy sensor can accurately quantify the leakage intensity by detecting the unique absorption spectral characteristics of refrigerant molecules.

[0032] In the data preprocessing stage, a scene-based correction algorithm is used to correct the non-uniformity of the infrared thermal image sequence. This algorithm selects multiple uniformly distributed reference points on the heat transfer tube surface and utilizes the radiation characteristics of these reference points to correct the non-uniformity of the entire thermal image, thereby eliminating image inhomogeneity caused by factors inherent to the thermal imager itself. For bad pixel repair, a neighborhood interpolation method is used to estimate and replace the bad pixel value based on the grayscale values ​​of normal pixels surrounding the bad pixel, restoring the integrity of the thermal image. During spatial registration, image feature point matching technology is used to align the acquired infrared thermal image with the heat transfer tube array, ensuring spatial consistency of thermal images acquired at different times.

[0033] For denoising signals of refrigerant flow rate, differential pressure, and leakage intensity, wavelet threshold denoising is employed. By setting an appropriate threshold, high-frequency noise components in the signal are filtered out while retaining the main characteristics of the signal, thereby improving the signal-to-noise ratio. Normalization processes map signal data of different dimensions to the same dimension range, such as the [0,1] interval, to eliminate the influence of dimensional differences on subsequent analysis.

[0034] Timestamp alignment processing precisely aligns the infrared thermal image sequence, refrigerant flow rate, pressure difference, and leakage intensity signals according to the time of signal acquisition, ensuring that multi-source data are strictly synchronized in the time dimension and forming a synchronized data stream.

[0035] Based on the above embodiments, in one embodiment, step S2 involves decomposing the temperature field based on the temperature field data to obtain the target spatial mode and time evolution curve, specifically including: Step S201. Reconstruct the above temperature field data into a snapshot matrix, and construct a sliding time window. Use the above sliding time window to slide and truncate on the above snapshot matrix to obtain multiple sub-snapshot matrices. Step S202. Perform singular value decomposition on each of the above sub-snapshot matrices to obtain the spatial modes and temporal coefficients corresponding to different time windows; Step S203. Perform mode registration on the spatial modes of adjacent time windows to obtain the registered spatial modes; Step S204. Determine the energy weights of each registered spatial mode, and sort the registered spatial modes from high to low according to the energy weights to obtain the registered spatial mode sequence; Step S205. Select the first N spatial modes in the above-registered spatial mode sequence as target spatial modes, and generate time evolution curves based on the above target spatial modes and the corresponding time coefficients, where N is a positive integer.

[0036] Specifically, in this embodiment, the surface temperature field of the heat transfer tube is decomposed based on dynamic intrinsic orthogonal decomposition (VIOD) using a preprocessed infrared thermogram sequence to obtain the target spatial modes and time evolution curves. VOD is an effective method for data dimensionality reduction and feature extraction, capable of extracting the main changing patterns of the heat transfer tube surface temperature field, i.e., the target spatial modes, from complex infrared thermogram sequences. These modes reflect the key characteristics of the heat transfer tube surface temperature distribution at different times. Simultaneously, VOD also yields the time evolution curves, which describe the changes of the target spatial modes over time.

[0037] In this embodiment, the preprocessed infrared thermal image sequence is first constructed into a data matrix. Then, dynamic intrinsic orthogonal decomposition is performed on this data matrix to obtain a series of orthogonal basis functions. These orthogonal basis functions can preserve the variance information in the data to the greatest extent, i.e., the main variation characteristics of the temperature field on the heat transfer tube surface. Next, the original infrared thermal image sequence is reconstructed according to the orthogonal basis functions to obtain target spatial modes. These modes represent the typical state of the temperature distribution on the heat transfer tube surface at different times. At the same time, by calculating the weight coefficient of each target spatial mode at different times, a time evolution curve can be obtained. This curve intuitively shows the trend of the temperature field on the heat transfer tube surface over time.

[0038] Preferably, in this embodiment, the preprocessed infrared thermal image sequence is reconstructed into a snapshot matrix to facilitate subsequent dynamic intrinsic orthogonal decomposition. The snapshot matrix is ​​a matrix composed of a series of infrared thermal image data arranged in chronological order, where each column represents the infrared thermal image data at a given moment, and the rows correspond to individual pixels or feature regions in the thermal image. This reconstruction method transforms the time-varying infrared thermal image sequence into a two-dimensional matrix, making it easier to process and analyze using the dynamic intrinsic orthogonal decomposition method.

[0039] Understandably, constructing a snapshot matrix requires following certain rules, arranging the infrared thermogram data from each moment sequentially into a matrix form. Each row or column can represent the temperature information at different locations on the heat transfer pipe surface at a specific moment. During construction, it's crucial to ensure the accuracy and completeness of the data, avoiding data loss or incorrect arrangement, which could directly impact subsequent temperature field decomposition results. For example, a frame-by-frame scanning method can be used, sequentially filling the corresponding positions of the snapshot matrix with the data from each frame of the infrared thermogram, forming a matrix structure containing temperature information from multiple moments. This lays the foundation for subsequently extracting the spatiotemporal dominant modes of the temperature field using dynamic intrinsic orthogonal decomposition.

[0040] Constructing a sliding time window and performing sliding cuts on the snapshot matrix is ​​to extract time-local subsequences from the original infrared thermogram sequence, enabling a more detailed analysis of the dynamic changes in the surface temperature field of the heat transfer tube. The size of the sliding time window and the sliding step size can be flexibly set according to actual needs and analytical precision. The window size determines the amount of data extracted each time; windows that are too large or too small may affect the accuracy of the analysis results. The sliding step size controls the speed at which the window slides; a step size that is too large will reduce the temporal resolution, while a step size that is too small will increase the computational load.

[0041] In this embodiment, the sliding time window setting needs to comprehensively consider the operating characteristics and computational efficiency of the condenser heat transfer tubes. Optionally, a window length covering 3-5 consecutive sampling times can be selected, which can capture local temperature fluctuations while avoiding data redundancy. The sliding step size is usually set to 1 / 3 to 1 / 2 of the window length to achieve a balance between time resolution and computational burden. The sliding time window slides column by column on the snapshot matrix according to the set step size. Each slide extracts a sub-snapshot matrix. The sub-snapshot matrix not only contains the spatiotemporal information of the temperature field on the heat transfer tube surface, but also reflects the dynamic change characteristics of the temperature field at different time scales. Using this data, the dominant spatiotemporal modes of the temperature field can be extracted more accurately, thereby achieving high-precision prediction of heat transfer tube defects.

[0042] In the specific implementation, let the infrared thermal image sequence be T. IR (x,y,t), sampling frequency f s The total duration is T total There are a total of M frames, with a sliding window length W and a step size Δt. w The i-th window contains the time interval [t] i ,t i+W All image frames within the range form a subset of the dataset, also known as a sub-snapshot matrix. Where n = x × y, is the number of spatial points, or the total number of pixels. This represents the number of frames within the window.

[0043] The above snapshot matrix for each window i Perform singular value decomposition as follows:

[0044] in, For the first N spatial modes of window i, It is a singular value diagonal matrix, representing the energy of each mode. This is the time coefficient matrix corresponding to the first N spatial modes.

[0045] In this embodiment, modal registration of spatial modes in adjacent time windows is performed to eliminate differences such as translation, rotation, or scale changes that may exist between spatial modes in different time windows, thereby obtaining a more accurate and consistent spatial modal representation. The modal registration process can be implemented using a dynamic time warping algorithm or a maximum correlation matching algorithm.

[0046] Dynamic time warping (VTW) is an algorithm used to measure the similarity between two time series. By adjusting the alignment along the time axis, it aims to achieve an optimal match between the two time series in a certain sense. This algorithm can handle time series of varying lengths and exhibits good robustness to scaling and distortion along the time axis. In modal registration, VTW can find the optimal alignment path between spatial modes by comparing the time series features of adjacent time windows, thereby eliminating modal differences caused by time variations and obtaining the registered spatial modes.

[0047] The maximum correlation matching algorithm is based on the principle of signal correlation. It calculates the correlation coefficient between spatial modes in adjacent time windows and finds the matching method with the highest correlation. This method can directly reflect the similarity of spatial modes in different time windows. By maximizing the correlation, it achieves mode registration and ensures that the registered spatial modes can accurately reflect the real changes in the temperature field on the surface of the heat transfer tube.

[0048] Modal registration yields registered spatial modes, which exhibit better temporal consistency and spatial accuracy, providing a reliable foundation for subsequent temperature field decomposition and defect prediction. The registered spatial modes can more accurately describe the dynamic changes in the temperature field on the heat transfer tube surface, contributing to improved prediction accuracy for heat transfer tube defects.

[0049] Furthermore, in this embodiment, step S203 above, which involves modal registration of spatial modes in adjacent time windows to obtain registered spatial modes, specifically includes: First, the spatial modes of two adjacent time windows are represented as two vector sequences respectively; In this approach, the spatial modes of adjacent time windows are considered as two time series, and represented as two vector sequences respectively. Secondly, the Euclidean distance between each pair of corresponding vectors in the two vector sequences is calculated to obtain the initial distance matrix, and the distance metric is optimized to obtain the target distance matrix; that is, the initial distance matrix is ​​dynamically adjusted to obtain the target distance matrix. Then, based on dynamic programming, the search proceeds step by step from the starting point of the target distance matrix to the endpoint to obtain the optimal matching path; where the optimal matching path is the best alignment between spatial modes of adjacent time windows; Finally, the alignment relationship of corresponding vectors in adjacent time window spatial modes is determined based on the above optimal matching path, so as to align the spatial modes of adjacent time windows and obtain the registered spatial modes.

[0050] In this embodiment, the dynamic time warping algorithm is used for modal registration. First, the spatial modal data of adjacent time windows are serialized to ensure that the two modalities to be registered have the same dimensional structure. Then, the Euclidean distance between each pair of corresponding vectors in the two serialized modal vectors is calculated to construct an initial distance matrix, which reflects the degree of difference between the two modes at different locations.

[0051] By dynamically adjusting the initial distance matrix, for example by using local weighting or global constraints, the distance metric is optimized to more accurately reflect the true differences between modes, ultimately yielding the target distance matrix. Subsequently, a dynamic programming algorithm is used to search from the top-left corner (the starting point) of the target distance matrix to the bottom-right corner (the ending point), selecting the path that minimizes the cumulative distance at each step, ultimately obtaining the optimal matching path that traverses the entire matrix. This optimal matching path represents the best alignment between spatial modes in adjacent time windows, describing how adjusting the correspondence on the time axis achieves optimal matching between two modes in the time dimension.

[0052] The alignment relationship of corresponding vectors in adjacent time window spatial modes is determined based on the optimal matching path, and the spatial modes are aligned according to this alignment relationship, including operations such as translation, rotation or scaling transformation, so that the registered spatial modes maintain consistency and continuity in the time dimension, thereby obtaining the registered spatial modes.

[0053] In other embodiments, step S203 above, which involves modal registration of spatial modes in adjacent time windows to obtain registered spatial modes, specifically includes: First, the spatial modes of adjacent time windows are expanded into vector form to obtain the spatial mode vectors of adjacent time windows; Secondly, calculate the correlation coefficients between all pairs of spatial mode vectors in adjacent time windows and construct a correlation matrix; Then, based on the above correlation matrix, the maximum correlation matching algorithm is used. Starting from the first time window, the algorithm searches for the spatial mode vector with the highest correlation in the subsequent time window for each spatial mode vector and matches it until the spatial mode vector of the last time window is matched, thus obtaining the matched spatial mode vector pair. Finally, based on the aforementioned spatial mode vector pairs, the correspondence between spatial modes in adjacent time windows is determined, and the spatial modes in adjacent time windows are aligned and adjusted to obtain the registered spatial modes.

[0054] In this embodiment, when using the maximum correlation matching algorithm for modal registration, the spatial modes of adjacent time windows are first expanded into vector form to convert the modal data into one-dimensional vectors, facilitating subsequent correlation calculation. Then, the correlation coefficients between all pairwise spatial mode vectors of adjacent time windows are calculated to reflect the similarity of spatial modes under different time windows, and a correlation matrix is ​​constructed to store these coefficients. For each mode in window i+1... Calculate its relationship with all modes of window i. The correlation is as follows:

[0055] in, For modality and The correlation, For the k-th mode in window i+1, Let j be the j-th modality in window i.

[0056] The algorithm described above starts from the first time window and sequentially searches for the most relevant spatial mode vector in subsequent time windows for matching for each spatial mode vector. This process is iterated until the spatial mode vectors of the last time window are matched, resulting in matched spatial mode vector pairs. Based on these pairs, the correspondence between spatial modes in adjacent time windows can be determined, describing how spatial modes match and correlate in different time windows. Aligning and adjusting the spatial modes of adjacent time windows according to this correspondence yields the registered spatial modes, providing a more accurate and reliable foundation for subsequent temperature field decomposition and defect prediction.

[0057] Furthermore, the energy weights of the aforementioned spatial modes reflect the proportion of that mode in the overall temperature field change and are a key indicator for measuring the importance of a mode. The process of determining the energy weights is essentially a process of assessing the importance of each registered spatial mode.

[0058] Specifically, the energy weights of each spatial mode are determined based on the singular value diagonal matrix obtained from singular value decomposition. Each element in the singular value diagonal matrix represents the energy level of the corresponding mode. Therefore, the energy weights of each spatial mode can be obtained by calculating the proportion of each singular value in the total energy.

[0059] After obtaining the energy weights, the registered spatial modes are sorted in descending order of weight to form a registered spatial mode sequence, which reflects the consistency and continuity of the spatial modes in the time dimension and reflects the relative importance of each mode in the overall temperature field change.

[0060] Preferably, N is the number of modes determined based on the cumulative energy contribution rate threshold.

[0061] In this embodiment, the cumulative energy contribution rate threshold is an important reference indicator when selecting the target spatial mode. This threshold is usually set according to actual needs and application scenarios, aiming to select spatial modes that can represent the main characteristics of the overall temperature field change. The cumulative energy contribution rate can be obtained by calculating the ratio of the sum of the energy weights of the first N spatial modes to the total energy. When the cumulative energy contribution rate reaches or exceeds the preset threshold, the value of N can be determined, and the first N spatial modes are selected as the target spatial modes. In this embodiment, taking a cumulative energy contribution rate threshold of 95% as an example, the smallest N is selected so that the cumulative energy proportion exceeds 95%, as shown in the following formula:

[0062] in, Let n be the k-th singular value of the i-th window, and n be the number of spatial points. The number of frames within window i.

[0063] The selection method described above ensures that the target spatial mode dominates the overall temperature field change, thus more accurately reflecting the dynamic changes in the temperature field on the heat transfer tube surface. Furthermore, by appropriately setting the cumulative energy contribution rate threshold, the computational and data storage requirements can be reduced while maintaining prediction accuracy, thereby improving the efficiency and practicality of the prediction method.

[0064] After determining the value of N, the top N spatial modes from the registered spatial mode sequence can be selected as the target spatial modes, i.e., the dominant spatial modes. By focusing on the target spatial modes, interference from irrelevant or secondary modes on the analysis results can be effectively avoided, thereby improving the ability to identify potential defects in heat transfer tubes and the accuracy of prediction. At the same time, this mode selection method based on energy contribution rate has strong adaptability and flexibility. The cumulative energy contribution rate threshold can be dynamically adjusted according to the operating characteristics of different condensers or actual testing needs to balance the relationship between prediction accuracy and computational efficiency.

[0065] It should be noted that the target spatial modes reflect the main changing characteristics of the temperature field on the heat transfer tube surface at different times, while the corresponding time coefficients reflect the amplitude and trend of these modes changing over time. By combining the target spatial modes with the corresponding time coefficients, the dynamic evolution of the temperature field on the heat transfer tube surface over time can be visually demonstrated.

[0066] Specifically, the time coefficient corresponding to each target spatial mode is obtained. With the target spatial mode as the horizontal axis and the time coefficient as the vertical axis, a time evolution curve for these target spatial modes is plotted. This curve clearly reflects the changes in the temperature field on the heat transfer tube surface at different time points, including temperature rises, falls, and fluctuations. By observing the time evolution curve, it is possible to visually see which time periods show more drastic temperature changes and which periods are relatively stable, thus providing important temporal dimension information for subsequent defect prediction. Simultaneously, the time evolution curve helps to locate anomalies or abrupt changes in the temperature field, which are often closely related to potential defects in the heat transfer tube.

[0067] In this embodiment, spatial modes are extracted and registered using dynamic intrinsic orthogonal decomposition technology, and the dominant spatial modes are selected by combining energy weight and cumulative energy contribution rate threshold, thereby generating time evolution curves. This makes the extraction of spatial modes more accurate and can capture subtle changes in the surface temperature field of the heat transfer tube at different times, thereby further improving the ability to identify and predict potential defects in the heat transfer tube.

[0068] Furthermore, in one embodiment, step S2 above, which involves performing a simulation mapping inversion based on the refrigerant flow rate and pressure difference to obtain the internal quantity state in the pipe, specifically includes: Step S205. Construct a heat-fluid coupling model of the condenser heat transfer tubes; Step S206. Based on the above refrigerant flow rate and pressure difference, as well as the tube length and inner diameter of the above condenser heat transfer tube, the above heat-fluid coupling model is solved by numerical simulation to obtain the flow velocity distribution and pressure distribution of the refrigerant inside the tube. Step S207. Based on the above flow velocity distribution and pressure distribution, the internal quantity state in the pipe is calculated by inversion.

[0069] The refrigerant state within the tube refers to the distribution of refrigerant reserves inside the heat transfer tube, including the total refrigerant quantity and flow parameters within the tube. Since directly measuring the refrigerant state within the tube is quite difficult, this embodiment uses simulation mapping inversion to achieve an accurate estimate.

[0070] Understandably, simulation mapping inversion is a method based on physical models and numerical calculations. It uses known refrigerant flow and pressure difference data to construct a fluid dynamics model inside the heat transfer tube to simulate the flow and distribution of refrigerant within the tube.

[0071] Specifically, a mathematical model of the fluid flow inside the heat transfer tube is first established based on the tube's geometry, refrigerant properties, and boundary conditions. Then, pre-processed refrigerant flow rate and pressure difference data are used as input, and the model is solved numerically to obtain the refrigerant's stock distribution and flow state parameters, such as velocity and turbulence intensity. These parameters collectively constitute the internal stock state within the tube. Through simulation mapping inversion, the internal stock state can be accurately obtained without damaging the heat transfer tube structure.

[0072] Preferably, obtaining the state of the amount of memory in the pipe specifically includes: First, a heat-fluid coupling model of the condenser heat transfer tubes is constructed, and the tube length and inner diameter of the condenser heat transfer tubes are obtained. Then, based on the pre-processed refrigerant flow rate, pre-processed pressure difference, tube length, and inner diameter, the heat-fluid coupling model is solved using numerical simulation to obtain the refrigerant velocity distribution and pressure distribution inside the tubes. Finally, based on the velocity and pressure distributions and combined with thermodynamic principles, the refrigerant storage state inside the tubes is calculated by inversion. The refrigerant storage state includes the amount of refrigerant stored inside the tubes and the refrigerant flow state parameters inside the tubes.

[0073] It should be noted that the heat-fluid coupling model is built upon the principles of heat transfer and fluid mechanics, and is used to describe the thermodynamic and fluid properties of the refrigerant within the heat transfer tubes of a condenser. This model considers the flow state of the refrigerant within the tubes, the heat transfer process, and its interaction with the tube walls, and can accurately simulate the dynamic behavior of the refrigerant within the tubes. The formula for the heat-fluid coupling model is:

[0074] in, For effective thermal conductivity, For internal heat generation rate, For fluid density, For isobaric specific heat capacity, For quality flow, T For temperature, t Let z be the time, and z be the axial coordinate.

[0075] Solving this model using numerical simulation allows us to obtain the refrigerant velocity and pressure distribution within the tubes. Specifically, this involves determining the initial and boundary conditions for the numerical simulation based on known pre-processed parameters such as refrigerant flow rate, pressure difference, tube length, and inner diameter. These conditions form the basis for solving the heat-fluid coupling model, reflecting the physical state of the condenser heat transfer tubes during actual operation. For example, the initial conditions can be set as the initial velocity and pressure distribution of the refrigerant within the tubes at a certain moment, while the boundary conditions can include parameters such as the inlet and outlet pressures and temperatures at both ends of the tube. Next, a suitable numerical simulation algorithm is used to discretize the heat-fluid coupling model. Numerical simulation algorithms can include the finite difference method, the finite element method, and the finite volume method, among others. These algorithms transform continuous partial differential equations into a discrete system of algebraic equations, facilitating their solution.

[0076] During discretization, the condenser heat transfer tubes need to be divided into multiple small control volumes or mesh elements. The equations of the heat-fluid coupling model are then applied to each element, and the discretized algebraic equations are solved iteratively. During the iteration, the refrigerant velocity and pressure distributions within the tubes need to be continuously updated until a convergence condition is met. The convergence condition can be set by ensuring that the error between two consecutive iterations is less than a given threshold. After solving the heat-fluid coupling model using the above numerical simulation method, the refrigerant velocity and pressure distributions within the tubes can be obtained.

[0077] The above distribution data can intuitively reflect the flow of refrigerant in the pipe, such as the magnitude and direction of flow velocity, and pressure variations. Based on these velocity and pressure distribution data, and combined with thermodynamic principles, the refrigerant state within the pipe can be calculated, including the amount of refrigerant and its flow parameters, such as flow velocity and turbulence intensity. Specifically, this includes: Based on the laws of conservation of mass and energy in thermodynamics, the amount of refrigerant inside the tube can be estimated. The law of conservation of mass states that in a closed system, the mass of a substance does not change with time. Therefore, by analyzing the refrigerant velocity distribution and the tube cross-sectional area, the mass of refrigerant flowing through a certain cross-section per unit time can be determined, and then the total amount of refrigerant inside the tube can be obtained by integration. Simultaneously, by combining this with the law of conservation of energy and considering the enthalpy change of the refrigerant inside the tube and its heat exchange with the tube wall, a more accurate inverse calculation of the refrigerant's flow state parameters can be performed. For example, by analyzing the trends in velocity and pressure distribution, it is possible to infer whether complex flow phenomena such as backflow and vortices exist within the tube, and the impact of these phenomena on the performance of the heat transfer tube.

[0078] Furthermore, in one embodiment, in step S3 above, the high-dimensional feature vector is processed by a graph neural network to obtain spatially enhanced features, specifically including: First, the high-dimensional feature vectors mentioned above are mapped to a graph structure; Then, feature extraction is performed on the above graph structure through graph convolutional layers to obtain spatially enhanced features.

[0079] In this embodiment, the high-dimensional feature vector is a feature representation obtained by fusing multi-source information such as time evolution curves, pipe internal quantity status, refrigerant flow rate, pressure difference, and leakage intensity signals. It can comprehensively reflect the operating status and health condition of the heat transfer tube. When constructing the high-dimensional feature vector, different types of data need to be standardized to eliminate the influence of dimensions and numerical ranges. Then, the standardized data are spliced ​​or combined according to certain rules to form the high-dimensional feature vector.

[0080] Understandably, a graph neural network is a deep learning model specifically designed for processing graph-structured data, capable of capturing spatial relationships and dependencies between nodes. In this embodiment, high-dimensional feature vectors are treated as node features in the graph-structured data. By constructing a heat transfer pipe diagram structure, a graph neural network is used to enhance the spatial features of the high-dimensional feature vectors.

[0081] Specifically, after constructing the heat transfer pipe diagram structure, it is input into a graph neural network. The graph neural network first performs a linear transformation on the features of each node, mapping the high-dimensional feature vector to a new feature space. Then, based on the connectivity relationships between nodes described by the adjacency matrix, it performs an aggregation operation on the node features. For each node, the graph neural network collects the feature information of its neighboring nodes and fuses this information with its own features through weighted summation and other methods. This aggregation operation allows nodes to retain their own features while incorporating the feature information of their neighboring nodes, thereby capturing the spatial dependencies between heat transfer pipes.

[0082] In each layer of the graph neural network, feature transformation and aggregation operations are repeatedly performed. As the network depth increases, node features are continuously fused and updated, gradually extracting more representative spatially enhanced features. These spatially enhanced features not only contain information about the operating status of the heat transfer tube itself, but also incorporate the influence of surrounding heat transfer tubes, enabling a more comprehensive and accurate reflection of the operating status and potential defect information of the heat transfer tubes in the entire condenser system. Through the processing of the graph neural network, the final feature representation after spatial feature enhancement is obtained, providing richer and more valuable input information for subsequent defect prediction.

[0083] Specifically, step S3 above includes: First, the time evolution curve, the internal volume state of the pipe, the refrigerant flow rate after pretreatment, the pressure difference after pretreatment, and the leakage intensity signal after pretreatment are combined according to preset rules to construct a high-dimensional feature vector. Secondly, an initial graph structure for the condenser heat transfer tubes is constructed, where each node in the initial graph structure represents a heat transfer tube, and the edges between nodes represent the proximity relationships between heat transfer tubes. The high-dimensional feature vectors are then assigned to the corresponding nodes in the initial graph structure as the initial feature representations of each node, thus obtaining the target graph structure. Then, the above target graph structure is input into the graph neural network, and the initial feature representation of each node is aggregated through multi-layer graph convolution operation. In each layer of graph convolution, the feature interaction between the current node and its neighboring nodes is calculated using the adjacency matrix to generate intermediate features containing local neighborhood information. The intermediate features generated by each layer are subjected to nonlinear transformation and fusion, and the original feature information is preserved by combining the residual connection mechanism to obtain the aggregated features of each node after update. Finally, the normalized exponential function is used to activate the updated aggregated features of each node to obtain spatially enhanced features.

[0084] It should be noted that a high-dimensional feature vector is constructed by combining the time evolution curve, the internal volume state of the pipe, the pre-processed refrigerant flow rate, pressure difference, and leakage intensity signals with the refrigerant inlet and outlet temperature difference. The preset rules can be formulated according to actual needs and data characteristics. For example, the various signal data can be concatenated in chronological order, or different weights can be assigned based on the importance and correlation of the signal data before combination. In this embodiment, the above features are concatenated in a fixed order to form a high-dimensional feature vector. As shown in the following formula:

[0085] in, Let r be the time coefficient corresponding to the r-th target spatial mode. This refers to the refrigerant flow rate. For pressure difference, To manage memory status, This is a leakage intensity signal. This refers to the temperature difference between the refrigerant inlet and outlet.

[0086] Constructing the initial diagram structure of the heat transfer tube Where V is the node set, with each heat transfer tube corresponding to a node vi, and a total of N nodes, such as N=10000. For an edge set, if two heat transfer tubes are physically adjacent, such as if the Manhattan distance is ≤1, then an undirected edge is established.

[0087] The high-dimensional feature vectors are assigned to the corresponding nodes in the initial graph structure, serving as the initial feature representation for each node, i.e., the initial feature vector for each node vi. From the basic feature vector The observations at the corresponding positions, for example, Features such as these are shared globally and broadcast to all nodes. However, parameters specific to each heat transfer tube, such as the tube's internal volume and the refrigerant flow rate, are precisely assigned to the corresponding nodes to form an initial feature representation with inter-tube differences, thereby obtaining the target graph structure.

[0088] After obtaining the target graph structure, it is input into a graph neural network for processing. The graph neural network performs deep aggregation of the initial feature representations of each node through multiple layers of graph convolution operations. During each layer of graph convolution, the graph neural network uses the adjacency matrix to accurately calculate the feature interactions between the current node and all its neighboring nodes. This interaction calculation can generate intermediate features containing rich local neighborhood information. These intermediate features not only reflect the characteristics of the node itself but also incorporate information from its surrounding nodes, as shown in the following equation:

[0089] in, Let i be the set of neighbors of node i. It is an aggregate function. For trainable weight matrix, For activation function, l This is a layer index.

[0090] After L-layer propagation, each node acquires context-aware features containing two- or even multi-hop neighborhood information. These intermediate features undergo nonlinear transformations and fusion operations to further extract and integrate useful information. Simultaneously, the residual connection mechanism effectively preserves the original feature information, preventing information loss or gradient vanishing during multi-layer graph convolution. Finally, by activating the updated aggregated features of each node using a normalized exponential function, spatially enhanced features are obtained. These spatially enhanced features not only have higher discriminative power but also better reflect the spatial distribution characteristics of condenser heat transfer tubes and the needs for defect prediction, providing strong support for subsequent defect identification and classification.

[0091] Furthermore, in one embodiment, step S4 above, identifying anomalous modes in the target spatial modes based on the spatial enhancement features, and generating a residual temperature field based on the anomalous modes, specifically includes: First, the spatial enhancement features mentioned above are fused with the target spatial modality to obtain a fused feature matrix; Then, the above-mentioned fused feature matrix is ​​scanned by a preset anomaly detection threshold range to identify abnormal modes that exceed the threshold range; Finally, the difference between the above-mentioned anomalous mode and the above-mentioned target space mode is calculated, and the difference is mapped into a residual temperature field.

[0092] Specifically, calculating the difference between the aforementioned anomalous mode and the aforementioned target spatial mode, and mapping the aforementioned difference into a residual temperature field, includes: First, based on the location information of the abnormal modes in the target space modes, the temperature data of the corresponding regions are extracted from the original infrared thermal image sequence. Then, the extracted temperature data is compared with the temperature distribution model under normal conditions, and the temperature deviation value of each abnormal region is calculated. Finally, a residual temperature field is constructed based on the temperature deviation value, where each pixel in the residual temperature field represents the abnormal temperature change at that location relative to the normal state.

[0093] In this embodiment, the spatial enhancement features contain the operating status and spatial correlation information of the condenser heat transfer tubes. By conducting in-depth analysis and mining of these features, abnormal modes in the target spatial modes can be effectively identified.

[0094] Specifically, a series of judgment criteria and thresholds based on normal operating conditions can be set, and the spatial enhancement features obtained through graph neural network processing can be compared with the above criteria and thresholds. When certain features exceed the normal range, the corresponding spatial mode can be determined as an abnormal mode. For example, if the temperature change, pressure fluctuation, and other parameters shown by the spatial enhancement features of a heat transfer tube in a specific region deviate significantly from the normal situation, then the spatial mode corresponding to that region can be identified as an abnormal mode.

[0095] The residual temperature field reflects the difference between the actual temperature field and the normal or expected temperature field. To generate the residual temperature field, the temperature distribution data of the condenser heat transfer tubes under normal conditions must first be obtained. This can be achieved through historical normal operation data, theoretical calculation models, or simulation experiments. Then, the actual monitored temperature field data is subtracted from the normal temperature field data to obtain the temperature difference distribution, which is the residual temperature field. The residual temperature field can visually demonstrate the areas and extent of temperature anomalies caused by abnormal modes.

[0096] Optionally, when fusing spatial enhancement features with the target spatial mode, weighted summation or splicing methods can be used to ensure full interaction of information between the two, resulting in a fused feature matrix that comprehensively reflects the operating status and spatial correlation of the heat transfer tubes. The preset anomaly detection threshold range needs to be determined based on the characteristic distribution range of the condenser heat transfer tubes during normal operation. This can be obtained through statistical analysis of a large amount of historical normal operation data, or it can be set by combining theoretical models and expert experience.

[0097] By scanning the fused feature matrix within a preset anomaly detection threshold range, anomalous modes exceeding the threshold range can be quickly and accurately identified. Based on the location information of the anomalous modes within the target spatial modes, temperature data of the corresponding regions are extracted from the original infrared thermogram sequence. This process must ensure that the extracted temperature data accurately corresponds to the anomalous mode regions to avoid data deviations affecting subsequent analysis.

[0098] Temperature distribution models under normal conditions can be constructed in various ways, such as establishing statistical models using historical normal operation data or establishing physical models based on heat transfer theory. When constructing a residual temperature field based on temperature deviation values, interpolation and other methods can be used to transform discrete temperature deviation values ​​into a continuous residual temperature field, where each pixel represents the amount of abnormal temperature change at that location relative to the normal state, which can intuitively reflect the abnormal area and degree.

[0099] Further, in one embodiment, in step S5 above, defect classification and defect evolution prediction are performed based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode to obtain the defect type and defect evolution trend, specifically including: Step S501. Extract abnormal features from the above residual temperature field, align the above leakage detection signal with the above abnormal features in time sequence, and construct a joint feature vector; the above abnormal features include extreme values ​​of temperature deviation, area of ​​temperature abnormal region, and spatial distribution characteristics; Step S502. Calculate the rate of change of the above target spatial modes in the time dimension to obtain the modal change rate; Step S503. Input the above joint feature vector and the above modal change rate into the pre-trained defect classification model to identify defects and determine the defect type; Step S504. Input the above-mentioned joint feature vector, the above-mentioned modal change rate and historical defect evolution data into the defect evolution prediction model to generate the defect evolution trend; the above-mentioned defect evolution trend is quantitatively characterized by the growth rate of the extreme value of temperature deviation, the expansion rate of the abnormal area and the migration path of the spatial distribution.

[0100] In this embodiment, the residual temperature field visually presents the temperature anomaly area and degree, the leakage intensity signal reflects the possible leakage situation, and the modal change rate reflects the dynamic change characteristics of the spatial modes. Classifying defects based on the above information allows for a more accurate determination of the defect type, such as pipe wall corrosion, refrigerant leakage, or heat transfer efficiency degradation.

[0101] Simultaneously, by combining modal change rates, the evolution trend of defects can be predicted. This prediction can utilize historical data and machine learning algorithms to analyze the changes in the residual temperature field, leakage intensity signal, and modal change rate over time. For example, by establishing a time series model, the changing trend of the residual temperature field, the changes in the leakage intensity signal, and the development direction of the modal change rate can be predicted over a future period. Based on the prediction results, corresponding measures can be taken in advance to prevent further deterioration of defects and ensure the safe and stable operation of the condenser.

[0102] To improve the accuracy of defect classification and defect evolution prediction, the classification and prediction models can be continuously updated and optimized. By incorporating new operational data and real-world cases, the models can be trained and adjusted to better adapt to condenser heat transfer tube defects under different operating conditions. Simultaneously, expert experience and knowledge can be introduced to manually review and correct the classification and prediction results, ensuring their reliability and practicality.

[0103] Optionally, step S5 above specifically includes: First, the extreme values ​​of temperature deviation, the area of ​​temperature anomaly region, and spatial distribution characteristics are extracted from the above residual temperature field. Then, the leakage intensity signal is time-aligned with the above extreme values ​​of temperature deviation, the area of ​​temperature anomaly region, and spatial distribution characteristics to construct a joint feature vector. Secondly, the rate of change of the target spatial modes in the time dimension is calculated to obtain the modal change rate; Then, the joint feature vector and the modal change rate are input into the pre-trained defect classification model to identify defects and determine the defect type. The defect type includes at least one of pipe wall corrosion, refrigerant leakage, and heat transfer efficiency decay. Finally, the aforementioned joint feature vector, modal change rate, and historical defect evolution data are input into the defect evolution prediction model to generate the defect evolution trend within the future time window.

[0104] Among them, the extreme values ​​of temperature deviation, the area of ​​temperature anomaly regions, and spatial distribution characteristics extracted from the residual temperature field are important indicators reflecting the degree of abnormality in the current operating state of the condenser heat transfer tubes. The extreme values ​​of temperature deviation directly reflect the magnitude of the temperature difference between the abnormal region and the normal state, while the area of ​​the temperature anomaly region reflects the scope of the abnormal situation, and the spatial distribution characteristics help to understand the specific location and distribution pattern of the anomaly in the heat transfer tubes. By temporally aligning the leakage intensity signal with the extracted features to construct a joint feature vector, a more comprehensive and accurate description of the operating state and potential defects of the condenser heat transfer tubes at a specific moment can be achieved. Temporal alignment ensures the consistency of each feature across the time dimension.

[0105] The rate of change of the target spatial mode in the time dimension is calculated by integrating or differentiating the time coefficient of the target spatial mode. This rate of change can reflect the speed of dynamic evolution of the spatial mode of the condenser heat transfer tube.

[0106] The pre-trained defect classification model is built upon a large amount of historical data and professional knowledge. It can analyze and judge the input features to accurately identify different types of defects, such as pipe wall corrosion, refrigerant leakage, and heat transfer efficiency degradation. This classification method helps to take appropriate maintenance and treatment measures for different types of defects.

[0107] Optionally, the defect evolution prediction model is built on a long short-term memory network. Through deep learning of historical data, it can capture the potential patterns of defect evolution in condenser heat transfer tubes. When inputting data, the joint feature vector provides detailed state information of the heat transfer tube at the current moment, the modal change rate reflects the dynamic changes of the spatial modes, and the historical defect evolution data provides the model with a reference trajectory of past defect development.

[0108] After receiving the input data, the defect evolution prediction model uses its complex internal neural network structure for calculation and analysis. The unique gating mechanism of the Long Short-Term Memory (LSTM) network enables it to effectively handle long-term dependencies in time-series data, thus more accurately predicting the evolution trend of defects within future time windows. The growth rate of extreme temperature deviations provides a direct indication of the rate of aggravation of temperature anomalies in abnormal regions; the expansion rate of the abnormal region reflects the expansion of the defect's impact range; and the spatial migration path helps predict the propagation direction of defects in the heat transfer tubes and the areas that may be affected. These quantitative characteristics provide maintenance personnel with clear indicators to develop maintenance plans in advance and take targeted measures, such as strengthening monitoring, arranging repairs, or replacing damaged components, thereby effectively preventing further deterioration of defects, ensuring the safe and stable operation of the condenser heat transfer tubes, and reducing the risk and losses of production accidents caused by defects.

[0109] The method in this embodiment integrates multi-source information such as infrared thermal images, refrigerant flow rate, pressure difference, and leakage signals. It utilizes an improved dynamic intrinsic orthogonal decomposition to extract the spatiotemporal dominant modes of the temperature field and combines them with inversion of the tube state. This enables accurate prediction of the type and development rate of defects in heat transfer tubes, thus solving the problem that current detection methods rely on periodic shutdowns for maintenance or local spot checks, resulting in poor defect detection efficiency and accuracy.

[0110] Secondly, embodiments of this application also provide a device for predicting defects in condenser heat transfer tubes.

[0111] In one embodiment, reference is made to Figure 2 , Figure 2This is a functional module diagram of an embodiment of the condenser heat transfer tube defect prediction device of this application. The condenser heat transfer tube defect prediction device includes a preprocessing module, a decomposition module, an inversion module, a processing module, an identification module, and a prediction module.

[0112] The aforementioned preprocessing module is used to acquire temperature field data on the surface of the condenser heat transfer tubes, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signals; The above decomposition module is used to decompose the temperature field based on the above temperature field data to obtain the target spatial mode and time evolution curve; The aforementioned inversion module is used to perform simulation mapping inversion based on the refrigerant flow rate and pressure difference to obtain the internal quantity status of the pipe; The above processing module is used to construct a high-dimensional feature vector based on the above time evolution curve, the internal quantity state of the pipe, the refrigerant flow rate and pressure difference, and the leakage detection signal, and to perform graph neural network processing on the above high-dimensional feature vector to obtain spatially enhanced features. The aforementioned identification module is used to identify anomalous modes in the aforementioned target spatial modes based on the aforementioned spatial enhancement features, and to generate a residual temperature field based on the aforementioned anomalous modes; The aforementioned prediction module is used to classify defects and predict their evolution based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode, thereby obtaining the defect type and the defect evolution trend.

[0113] Furthermore, in one embodiment, the preprocessing module is used to collect temperature field data on the surface of the condenser heat transfer tube, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signal, and perform preprocessing.

[0114] Furthermore, in one embodiment, the above-mentioned decomposition module is used for: The temperature field data is reconstructed into a snapshot matrix, and a sliding time window is constructed. The sliding time window is used to slide and truncate on the snapshot matrix to obtain multiple sub-snapshot matrices. Singular value decomposition is performed on each of the above sub-snapshot matrices to obtain the spatial modes and temporal coefficients corresponding to different time windows; Modal registration is performed on the spatial modes of adjacent time windows to obtain the registered spatial modes; Determine the energy weights of each registered spatial mode, and sort the registered spatial modes from high to low according to the energy weights to obtain the registered spatial mode sequence; The first N spatial modes in the registered spatial mode sequence are selected as target spatial modes, and time evolution curves are generated based on the target spatial modes and their corresponding time coefficients, where N is a positive integer.

[0115] Furthermore, in one embodiment, the above-mentioned decomposition module is also used for: The spatial modes of two adjacent time windows are represented as two vector sequences respectively; the Euclidean distance between each pair of corresponding vectors in the two vector sequences is calculated to obtain the initial distance matrix, and the distance metric is optimized to obtain the target distance matrix; based on dynamic programming, the search proceeds step by step from the starting point of the target distance matrix to the endpoint to obtain the optimal matching path; the alignment relationship of corresponding vectors in the spatial modes of adjacent time windows is determined according to the optimal matching path to align the spatial modes of adjacent time windows and obtain the registered spatial modes.

[0116] Furthermore, in one embodiment, the above-mentioned decomposition module is also used for: The spatial modes of adjacent time windows are expanded into vector form to obtain spatial mode vectors of adjacent time windows. The correlation coefficients between all pairs of spatial mode vectors of adjacent time windows are calculated to construct a correlation matrix. Based on the above correlation matrix, the maximum correlation matching algorithm is used to find the spatial mode vector of the subsequent time window with the highest correlation for each spatial mode vector, starting from the first time window, until the spatial mode vector of the last time window is matched, and the matched spatial mode vector pairs are obtained. The correspondence between spatial modes of adjacent time windows is determined according to the above spatial mode vector pairs, and the spatial modes of adjacent time windows are aligned and adjusted to obtain the registered spatial modes.

[0117] Furthermore, in one embodiment, the above-described inversion module is used for: A heat-fluid coupling model of the condenser heat transfer tube is constructed. Based on the refrigerant flow rate and pressure difference, as well as the tube length and inner diameter of the condenser heat transfer tube, the heat-fluid coupling model is solved by numerical simulation to obtain the flow velocity distribution and pressure distribution of the refrigerant inside the tube. Based on the flow velocity distribution and pressure distribution, the internal quantity state of the refrigerant inside the tube is calculated by inversion.

[0118] Furthermore, in one embodiment, the above-described processing module is used to: The high-dimensional feature vectors are mapped to a graph structure; features are extracted from the graph structure through graph convolutional layers to obtain spatially enhanced features.

[0119] Furthermore, in one embodiment, the identification module described above is used for: The aforementioned spatial enhancement features are fused with the target spatial mode to obtain a fused feature matrix; the fused feature matrix is ​​scanned by a preset anomaly detection threshold range to identify anomalous modes that exceed the threshold range; the difference between the anomalous modes and the target spatial mode is calculated and mapped to a residual temperature field.

[0120] Furthermore, in one embodiment, the prediction module described above is used for: Anomalies are extracted from the residual temperature field, and the leakage detection signal is time-series aligned with the anomalies to construct a joint feature vector. The anomalies include extreme values ​​of temperature deviation, area of ​​temperature anomaly region, and spatial distribution characteristics. Calculate the rate of change of the above target spatial modes in the time dimension to obtain the modal change rate; The joint feature vector and the modal change rate are input into the pre-trained defect classification model to identify defects and determine the defect type. The aforementioned joint feature vector, modal change rate, and historical defect evolution data are input into the defect evolution prediction model to generate a defect evolution trend. The defect evolution trend is quantitatively characterized by the growth rate of extreme temperature deviations, the expansion rate of abnormal area, and the migration path of spatial distribution.

[0121] The functions of each module in the above-mentioned condenser heat transfer tube defect prediction device correspond to the steps in the above-mentioned condenser heat transfer tube defect prediction method embodiment, and their functions and implementation processes will not be described in detail here.

[0122] Thirdly, embodiments of this application provide a condenser heat transfer tube defect prediction device, which can be a personal computer (PC), laptop computer, server or other device with data processing capabilities.

[0123] Reference Figure 3 , Figure 3 This is a schematic diagram of the hardware structure of the condenser heat transfer tube defect prediction device involved in the embodiments of this application. In the embodiments of this application, the condenser heat transfer tube defect prediction device may include a processor, a memory, a communication interface, and a communication bus.

[0124] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0125] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces for interconnecting components within the condenser heat transfer tube defect prediction device, as well as interfaces for interconnecting the condenser heat transfer tube defect prediction device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0126] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0127] The processor can be a general-purpose processor, which can call the condenser heat transfer tube defect prediction program stored in the memory and execute the condenser heat transfer tube defect prediction method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the condenser heat transfer tube defect prediction program is called can refer to the various embodiments of the condenser heat transfer tube defect prediction method of this application, and will not be repeated here.

[0128] Those skilled in the art will understand that Figure 3 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0129] It should be noted that the sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not represent a sequential order, nor do they limit "first," "second," and "third" to different types.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0131] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting defects in condenser heat transfer tubes, characterized in that, The method includes: Acquire temperature field data on the surface of the condenser heat transfer tubes, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signals; Temperature field decomposition is performed based on the temperature field data to obtain the target spatial mode and time evolution curve, and simulation mapping inversion is performed based on the refrigerant flow rate and pressure difference to obtain the internal quantity state of the pipe. Based on the time evolution curve, the internal volume status of the pipe, the refrigerant flow rate and pressure difference, and the leak detection signal, a high-dimensional feature vector is constructed, and the high-dimensional feature vector is processed by a graph neural network to obtain spatially enhanced features. Based on the spatial enhancement features, abnormal modes in the target spatial modes are identified, and a residual temperature field is generated based on the abnormal modes; Based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode, defect classification and defect evolution prediction are performed to obtain the defect type and defect evolution trend.

2. The method for predicting defects in condenser heat transfer tubes as described in claim 1, characterized in that, Based on the temperature field data, temperature field decomposition is performed to obtain the target spatial modes and time evolution curves, specifically including: The temperature field data is reconstructed into a snapshot matrix, and a sliding time window is constructed. The sliding time window is used to slide and truncate on the snapshot matrix to obtain multiple sub-snapshot matrices. Singular value decomposition is performed on each of the sub-snapshot matrices to obtain the spatial modes and temporal coefficients corresponding to different time windows; Modal registration is performed on the spatial modes of adjacent time windows to obtain the registered spatial modes; Determine the energy weights of each registered spatial mode, and sort the registered spatial modes from high to low according to the energy weights to obtain the registered spatial mode sequence; The first N spatial modes in the registered spatial mode sequence are selected as target spatial modes, and a time evolution curve is generated based on the target spatial modes and the corresponding time coefficients, where N is a positive integer.

3. The method for predicting defects in condenser heat transfer tubes as described in claim 2, characterized in that, The modal registration of spatial modes in adjacent time windows to obtain the registered spatial modes specifically includes: The spatial modes of two adjacent time windows are represented as two vector sequences respectively; Calculate the Euclidean distance between each pair of corresponding vectors in the two vector sequences to obtain the initial distance matrix, and optimize the distance metric to obtain the target distance matrix; Based on dynamic programming, the optimal matching path is obtained by progressively searching from the starting point of the target distance matrix to the endpoint. The alignment relationship of corresponding vectors in adjacent time window spatial modes is determined based on the optimal matching path, so as to perform alignment processing on the spatial modes of adjacent time windows and obtain the registered spatial modes.

4. The method for predicting defects in condenser heat transfer tubes as described in claim 2, characterized in that, The modal registration of spatial modes in adjacent time windows to obtain the registered spatial modes specifically includes: Expand the spatial modes of adjacent time windows into vector form to obtain the spatial mode vector of adjacent time windows; Calculate the correlation coefficients between all pairs of spatial mode vectors in adjacent time windows, and construct a correlation matrix; Based on the correlation matrix, the maximum correlation matching algorithm is used. Starting from the first time window, the algorithm searches for the spatial modality vector with the highest correlation in the subsequent time window for each spatial modality vector and matches it until the spatial modality vector of the last time window is matched, thus obtaining the matched spatial modality vector pair. The correspondence between spatial modes in adjacent time windows is determined based on the spatial mode vector pairs, and the spatial modes in adjacent time windows are aligned and adjusted to obtain the registered spatial modes.

5. The method for predicting defects in condenser heat transfer tubes as described in claim 1, characterized in that, Based on the refrigerant flow rate and pressure difference, a simulation mapping inversion is performed to obtain the internal refrigerant state in the pipe, specifically including: Construct a thermal-fluid coupling model for the condenser heat transfer tubes; Based on the refrigerant flow rate and pressure difference, as well as the tube length and inner diameter of the condenser heat transfer tube, the heat-fluid coupling model is solved by numerical simulation to obtain the flow velocity distribution and pressure distribution of the refrigerant inside the tube. Based on the flow velocity and pressure distribution, the internal quantity status of the pipe can be calculated by inversion.

6. The method for predicting defects in condenser heat transfer tubes as described in claim 1, characterized in that, The high-dimensional feature vector is processed by a graph neural network to obtain spatially enhanced features, specifically including: Map the high-dimensional feature vectors to a graph structure; The graph structure is used to extract features through graph convolutional layers to obtain spatially enhanced features.

7. The method for predicting defects in condenser heat transfer tubes as described in claim 1, characterized in that, Based on the spatial enhancement features, anomalous modes in the target spatial modes are identified, and a residual temperature field is generated according to the anomalous modes, specifically including: The spatial enhancement features are fused with the target spatial modality to obtain a fused feature matrix; The fused feature matrix is ​​scanned by a preset anomaly detection threshold range to identify abnormal modes that exceed the threshold range; The difference between the anomalous mode and the target spatial mode is calculated, and the difference is mapped to a residual temperature field.

8. The method for predicting defects in condenser heat transfer tubes as described in claim 1, characterized in that, Based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode, defect classification and defect evolution prediction are performed to obtain the defect type and defect evolution trend, specifically including: Anomalies are extracted from the residual temperature field, and the leakage detection signal is time-aligned with the anomalies to construct a joint feature vector. The anomalies include extreme values ​​of temperature deviation, area of ​​temperature anomalies, and spatial distribution characteristics. Calculate the rate of change of the target spatial mode in the time dimension to obtain the modal change rate; The joint feature vector and the modal change rate are input into a pre-trained defect classification model to identify defects and determine the defect type. The joint feature vector, the modal change rate, and historical defect evolution data are input into the defect evolution prediction model to generate a defect evolution trend. The defect evolution trend is quantitatively characterized by the growth rate of extreme temperature deviations, the expansion rate of abnormal area, and the migration path of spatial distribution.

9. A device for predicting defects in condenser heat transfer tubes, characterized in that, The device includes: The preprocessing module is used to acquire temperature field data on the surface of the condenser heat transfer tubes, refrigerant flow rate and pressure difference at the refrigerant inlet and outlet, and refrigerant leakage detection signals. The decomposition module is used to decompose the temperature field based on the temperature field data to obtain the target spatial mode and time evolution curve. The inversion module is used to perform simulation mapping inversion based on the refrigerant flow rate and pressure difference to obtain the internal quantity status of the pipe; The processing module is used to construct a high-dimensional feature vector based on the time evolution curve, the pipe internal quantity status, the refrigerant flow rate and pressure difference, and the leak detection signal, and to perform graph neural network processing on the high-dimensional feature vector to obtain spatially enhanced features. The identification module is used to identify anomalous modes in the target spatial mode based on the spatial enhancement features, and to generate a residual temperature field based on the anomalous modes; The prediction module is used to classify defects and predict their evolution based on the residual temperature field, the leakage detection signal, and the modal change rate of the target spatial mode, so as to obtain the defect type and the defect evolution trend.

10. A device for predicting defects in condenser heat transfer tubes, characterized in that, The condenser heat transfer tube defect prediction device includes a processor, a memory, and a condenser heat transfer tube defect prediction program stored in the memory and executable by the processor, wherein when the condenser heat transfer tube defect prediction program is executed by the processor, it implements the steps of the condenser heat transfer tube defect prediction method as described in any one of claims 1 to 8.

Images