A Steam Leakage Data Analysis Method and System Based on Collaborative Manifold Learning

Abstract

This application provides a steam leak data analysis method and system based on collaborative manifold learning, applied in the field of industrial monitoring system technology. After acquiring an image to be identified, the method uses a target detection model to identify candidate regions from the image. Based on the candidate region information, dynamic evolution features of the candidate regions in continuous images to be identified are extracted. Then, collaborative manifold learning is used to embed the dynamic evolution features into a low-dimensional space to obtain low-dimensional embedding features, thereby generating analysis results based on the low-dimensional embedding features and preset steam leak physical behavior features. This method can compress data dimensionality while maintaining physical semantic consistency, effectively extract features related to the steam leak evolution process, and improve the accuracy of steam leak data analysis results.

Description

Technical Field

[0001] This application relates to the field of industrial monitoring system technology, and in particular to a steam leakage data analysis method and system based on collaborative manifold learning. Background Technology

[0002] Steam leakage occurs during industrial production processes when steam escapes from pipelines due to damage or seal failure in components such as pipes, valves, and equipment. Because steam has a high temperature and may release toxic substances, steam leaks pose hazards such as personal injury, equipment corrosion, production interruption, and environmental pollution. Therefore, to mitigate the hazards of steam leaks, real-time monitoring is necessary to promptly identify leak points and implement appropriate countermeasures.

[0003] In industrial settings, real-time monitoring of steam leaks can be achieved through image acquisition. This involves using infrared image or video acquisition devices to capture images of the pipeline system, obtaining infrared image or video data. Analysis of this data then determines whether a steam leak exists. Because the infrared images and video data generated during a steam leak contain complex spatial structures, temporal dynamics, and temperature gradient information, they exhibit characteristics such as high dimensionality, strong spatiotemporal coupling, nonlinearity, and multi-scale. Therefore, effectively extracting key features to support subsequent analyses such as principal component analysis is fundamental to achieving intelligent sensing and behavioral modeling of steam leaks.

[0004] However, data analysis methods such as principal component analysis suffer from poor adaptability when dealing with high-dimensional nonlinear data, making it difficult to accurately represent the evolution patterns and physical characteristics of steam leaks, resulting in insufficient feature representation capabilities. Furthermore, the lack of physical consistency modeling capabilities for industrial leak data prevents the full integration of important semantic features such as temperature field diffusion and dynamic trajectories, thus reducing the accuracy of steam leak data analysis results and limiting the practicality of steam leak analysis in industrial scenarios. Summary of the Invention

[0005] In view of this, embodiments of this application provide a steam leakage data analysis method and system based on collaborative manifold learning to solve the problem of low accuracy of steam leakage data analysis results.

[0006] According to a first aspect of this application, a method for analyzing steam leakage data based on collaborative manifold learning is provided, the method comprising:

[0007] Acquire an image to be identified, wherein the image to be identified is an infrared image obtained by an infrared video surveillance device performing image acquisition on the on-site operating scene;

[0008] Candidate regions are identified from the image to be identified using an object detection model to obtain candidate region information. The object detection model is a deep learning model trained using a training dataset. The candidate region information includes region prediction boxes, region categories, and confidence scores.

[0009] Based on the candidate region information, the dynamic evolution features of the candidate region in the continuous images to be identified are extracted. The dynamic evolution features include temperature distribution features, spatial diffusion morphology features, and time series change features.

[0010] The dynamic evolution features are embedded into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features, which are generated by extracting nonlinear structural relationships within the data from the dynamic evolution features.

[0011] Analysis results are generated based on the low-dimensional embedding features and the preset physical behavior features of steam leakage.

[0012] According to a second aspect of this application, a steam leak data analysis system based on collaborative manifold learning is provided. The system includes: an infrared video monitoring device, a server, a communication device, and a terminal device; wherein the infrared video monitoring device and the terminal device establish a communication connection with the server through the communication device; the infrared video monitoring device is configured to perform image acquisition of the on-site operating scene; the server includes:

[0013] The image acquisition module is used to acquire the image to be identified, which is an infrared image obtained by the infrared video monitoring device from the on-site operating scene;

[0014] The candidate region identification module is used to identify candidate regions from the image to be identified using an object detection model to obtain candidate region information. The object detection model is a deep learning model trained using a training dataset. The candidate region information includes a region prediction box, a region category, and a confidence score.

[0015] The feature extraction module is used to extract the dynamic evolution features of the candidate region in the continuous image to be identified based on the candidate region information. The dynamic evolution features include temperature distribution features, spatial diffusion morphology features and time series change features.

[0016] The dimensionality reduction module is used to embed the dynamic evolution features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features, which are generated by extracting nonlinear structural relationships within the data from the dynamic evolution features.

[0017] The analysis module is used to generate analysis results information based on the low-dimensional embedded features and preset steam leakage physical behavior features.

[0018] According to a third aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described steam leakage data analysis method based on cooperative manifold learning.

[0019] According to a fourth aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described steam leakage data analysis method based on cooperative manifold learning.

[0020] By employing the above technical solutions, embodiments of this application provide a steam leakage data analysis method and system based on collaborative manifold learning. After acquiring an image to be identified, the method uses a target detection model to identify candidate regions from the image. Then, based on the candidate region information, it extracts the dynamic evolution features of the candidate regions in continuous images to be identified. Next, through collaborative manifold learning, the dynamic evolution features are embedded into a low-dimensional space to obtain low-dimensional embedded features. Analysis results are then generated based on these low-dimensional embedded features and preset steam leakage physical behavior features. This method can fuse physical features such as spatial structure, temperature information, and temporal dynamics to construct a physically consistent high-dimensional feature representation, and achieve nonlinear dimensionality reduction through manifold embedding. This method can compress data dimensions while maintaining physical semantic consistency, effectively extracting features related to the steam leakage evolution process, and improving the accuracy of steam leakage data analysis results.

[0021] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 A schematic diagram of the steam leakage data analysis method based on collaborative manifold learning provided in the embodiments of this application;

[0024] Figure 2 This is a schematic diagram of the system deployment structure provided in the embodiments of this application;

[0025] Figure 3 This is a schematic diagram of the candidate region extraction process provided in an embodiment of this application;

[0026] Figure 4 This is a schematic diagram illustrating the steam leak detection effect provided in an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of the collaborative manifold learning dimensionality reduction process provided in the embodiments of this application;

[0028] Figure 6 This is a schematic diagram of the collaborative manifold learning physical feature dimensionality reduction process provided in the embodiments of this application;

[0029] Figure 7 This is a schematic diagram of the process for generating analysis result information provided in the embodiments of this application;

[0030] Figure 8 A schematic diagram of the structure of a steam leakage data analysis system based on collaborative manifold learning provided in this application embodiment. Detailed Implementation

[0031] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0032] In this embodiment, steam leakage is an abnormal production phenomenon in industrial production processes. Steam leakage is caused by damage or seal failure of components such as pipes, valves, and equipment, resulting in steam escaping from the pipes.

[0033] Because steam has a high temperature and may release toxic substances, steam leaks pose hazards such as personal injury, equipment corrosion, production interruption, and environmental pollution. Therefore, to reduce the hazards of steam leaks, it is necessary to monitor steam leaks in real time, promptly identify leak points, and take appropriate countermeasures.

[0034] To achieve real-time monitoring of steam leaks, some embodiments involve acquiring infrared images or video data of areas prone to steam leakage at the production site, including pipes, valves, and equipment. The presence of a steam leak is then determined through analysis of the infrared images or video data.

[0035] In the analysis of infrared images or video data, the large amount of infrared images and video data generated during steam leakage processes contains complex spatial structures, temporal dynamics, and temperature gradient information, exhibiting characteristics such as high dimensionality, strong spatiotemporal coupling, nonlinearity, and multi-scale. Therefore, effectively extracting key features from these data to support subsequent analysis is fundamental to achieving intelligent sensing and behavioral modeling of steam leakage.

[0036] With the continuous improvement of industrial process automation, the demand for steam leak monitoring is increasing, and related image and sensor data exhibit significant characteristics such as high dimensionality, dynamics, and nonlinearity. Especially when using infrared imaging equipment to collect steam leak data, not only is there obvious temperature field diffusion behavior, but it is also accompanied by spatiotemporal dynamic changes and complex environmental interference. Therefore, the processing and analysis of this type of data, especially feature extraction and dimensionality reduction modeling, has become a fundamental step in achieving intelligent identification and early warning.

[0037] In some embodiments, linear methods such as Principal Component Analysis (PCA) can be used to transform a set of potentially correlated variables into a set of linearly uncorrelated variables through orthogonal transformation, reducing the dimensionality of the data while preserving as much of the original data's variability as possible. However, infrared images and video data generated during steam leaks exhibit typical spatiotemporal coupling characteristics, containing multidimensional information such as temperature diffusion, dynamic changes, and local perturbations. In actual steam leak images, temperature diffusion paths, local anomalous hotspots, and dynamic morphological changes often do not conform to the linear assumption, making it difficult for the aforementioned linear dimensionality reduction methods to retain key information, effectively capture the nonlinear structure in the data, and result in unrepresentative dimensionality reduction results, affecting subsequent identification and analysis.

[0038] In some embodiments, nonlinear manifold algorithms such as t-distributed stochastic neighbor embedding (t-SNE), isometric mapping (Isomap), and locally linear embedding (LLE) can be used to represent high-dimensional data in a low-dimensional space while preserving the data's inherent structure as much as possible. While manifold learning algorithms can be used for dimensionality reduction of nonlinear data, in practical applications they do not consider the physical relationships between temperature evolution, spatial structure, and time series in steam leak data. This results in a lack of physical consistency in the dimensionality reduction results, making it difficult to serve downstream leak behavior analysis tasks. Consequently, these nonlinear manifold algorithms lack the ability to model physical semantics and cannot fully integrate important semantic features such as temperature field diffusion and dynamic trajectories, making them unsuitable for industrial leak data compression, thus limiting their practicality in industrial scenarios. For example, while dimensionality reduction algorithms can reveal the manifold structure of data, they do not consider the unique temporal, spatial diffusion, and thermophysical characteristics of steam leaks, resulting in a lack of engineering interpretability in the dimensionality reduction results.

[0039] Furthermore, steam leak data typically consists of multiple image sequences, and the steam distribution and temperature field vary significantly within each frame. However, the dimensionality reduction methods described in the above embodiments only process single-frame images or static samples. Performing dimensionality reduction on each frame individually easily overlooks the temporal continuity and diffusion evolution trends, causing the embedded features after dimensionality reduction to lose their temporal structure and making it difficult to characterize complete behavioral patterns. This makes it difficult to fuse multi-source heterogeneous data using the dimensionality reduction methods described in the above embodiments, resulting in a lack of synergy in data dimensionality reduction, a lack of collaborative compression capabilities for spatiotemporal dynamic features, and an inability to model the continuous evolution trajectory and behavioral patterns during the leak process, thus limiting its applicability in subsequent diagnostic tasks.

[0040] To address the issue of low accuracy in steam leak data analysis, this application provides a steam leak data analysis method based on collaborative manifold learning in some embodiments. This method can perform dimensionality reduction on steam leak data based on collaborative manifold learning. It combines multi-channel image physical features and behavioral dynamics, and achieves high-dimensional feature compression and semantic preservation through nonlinear embedding, providing an interpretable low-dimensional representation for subsequent intelligent recognition and behavioral modeling. By fusing three types of physical features—spatial structure, temperature information, and temporal dynamics—a physically consistent high-dimensional feature representation space is constructed, and nonlinear dimensionality reduction is achieved through manifold embedding. While maintaining physical and semantic consistency, the data dimensionality is compressed, effectively extracting potential features closely related to the steam leak evolution process, providing low-dimensional representation support for subsequent recognition, judgment, and modeling.

[0041] The steam leakage data analysis method based on collaborative manifold learning can be applied to electronic devices with data processing and communication capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control units. For ease of description, this embodiment uses a server as the execution subject of the method. It should be understood that the method can also be applied to other types of execution subjects, which are not illustrated in this embodiment. Figure 1 As shown, the method includes:

[0042] S101. Obtain the image to be recognized.

[0043] To monitor the on-site operation, the server first acquires an image to be identified. This image is an infrared image obtained by infrared video monitoring equipment capturing images of the on-site operation. The infrared video monitoring equipment can establish a communication connection with the server and work with the user's terminal equipment to build an intelligent steam leak monitoring system for real-time monitoring of the on-site operation.

[0044] In some embodiments, such as Figure 2As shown, the intelligent steam leak monitoring system adopts a layered architecture design, including a front-end sensing layer, a data transmission layer, and a back-end processing layer. The front-end sensing layer is used to collect monitoring-related data in the on-site operating environment. Therefore, the front-end sensing layer can include infrared video monitoring equipment deployed in the on-site operating environment.

[0045] For example, the front-end sensing layer deploys high-precision infrared thermal imaging cameras within a safe area of ​​2-3 meters around the steam pipeline using a universal adjustable bracket. The spherical node supports ±45° omnidirectional adjustment, ensuring the equipment covers the critical area of ​​the pipeline with an optimal downward viewing angle of 15-20°. The bracket body is made of 304 stainless steel and reliably connected to building facades such as walls, columns, and ceilings using M12 anti-corrosion expansion bolts, ensuring both equipment stability and adaptability to thermal expansion displacement.

[0046] The data transmission layer is used to realize data transmission between the front-end sensing layer, the back-end processing layer, and the client layer, and can include communication devices such as network cables and switches. For example, an infrared thermal imaging camera is directly connected to an industrial-grade PoE switch via a Category 6 double-shielded Ethernet cable, and after passing through a redundant ring network to the core switch, the thermal map data is uploaded to the back-end processing layer in real time via a fiber optic link.

[0047] The back-end processing layer is used to process the monitoring process data. It can be built by installing applications related to the steam leak intelligent monitoring system on the server. For example, the back-end processing layer includes a monitoring server, which synchronously interfaces with the enterprise's distributed control system (DCS) to achieve data fusion and processing.

[0048] The backend processing layer can also include user-accessible terminal devices for displaying the identification and processing results. For example, these devices could be mobile phones, tablets, or other mobile terminals. The monitoring server can run AI algorithms. When a steam leak is detected by the AI ​​algorithm, a three-tiered alarm mechanism is automatically triggered. This includes immediate alerts from on-site audible and visual alarms, a pop-up 3D location alarm window on the management platform, and a push of a work order to maintenance personnel's mobile terminals via SMS gateway. Simultaneously, an electronic contingency plan is generated, including the leak rate, affected area, and recommended actions, providing multi-dimensional data support for emergency decision-making.

[0049] In some embodiments, to acquire the image to be identified, the server, during real-time monitoring, can receive data analysis instructions input by the user and, in response, send a data acquisition request to the infrared video surveillance device. Upon receiving the data acquisition request, the infrared video surveillance device can activate its image acquisition function to capture infrared images or record infrared video of the scene, forming an image data stream. This image data stream is then sent to the server, allowing the server to acquire the image to be identified from it.

[0050] To meet the requirements of subsequent steam leak data analysis, in some embodiments, the server can obtain an image that meets the input requirements through frame segmentation and normalization during the acquisition of the image to be identified. Therefore, the server can first receive the video stream data sent by the infrared video monitoring device, and then perform frame segmentation on the video stream data to obtain a continuous image sequence including multiple consecutive image frames. Then, based on a preset normalization term, the image frames in the continuous image sequence are normalized to obtain the image to be identified.

[0051] The preset normalization terms include size normalization, format normalization, and pixel normalization to adapt to the input requirements of the target detection model. For example, an infrared camera has high-resolution thermal imaging capabilities and can record the heat distribution information of the equipment surface and its surrounding environment in real time. Infrared video surveillance equipment deployed around industrial equipment collects video stream data during on-site operation and performs frame-by-frame processing to convert the video stream data into a continuous infrared image sequence. Then, the framed images undergo preprocessing operations such as size standardization, format conversion, and pixel normalization to ensure that the preprocessed images can adapt to the input requirements of the deep learning target detection model, providing a high-quality data foundation for subsequent candidate region extraction and physical feature modeling.

[0052] S102. Use an object detection model to identify candidate regions from the image to be identified in order to obtain candidate region information.

[0053] After acquiring the image to be identified, the server can invoke the object detection model. This object detection model is a deep learning model trained using a training dataset. In some embodiments, the server can pre-train the model to obtain the object detection model. When training the model, the server can first acquire the training dataset.

[0054] The training dataset may include a certain amount of sample images, and each sample image may be labeled with a region label. That is, the training dataset includes sample images and region labels, and the region labels are used to annotate the steam leak areas in the sample images. For example, such as... Figure 3 As shown, the server can annotate steam leak areas in infrared images using image annotation tools to construct a standardized training dataset. Annotation tools such as labelimg can be used to annotate steam leak areas in infrared image sequences, generating annotation files and constructing a standardized training dataset. Furthermore, during annotation, bounding boxes need to be drawn for the steam leak targets in each image, and a category needs to be assigned to each bounding box.

[0055] Given the acquired training dataset, the sample images from the dataset can be sequentially input into the trained deep learning model to obtain the training results output by the deep learning model. The training loss is then calculated by comparing the training results with the region labels, and iterative training is performed on the deep learning model based on this loss to obtain the object detection model.

[0056] After obtaining a standardized training dataset through annotation, the server can input this dataset into a deep learning model for training. The deep learning model can be configured to extract spatial and semantic features from sample images and perform classification calculations based on these features to obtain training results. Specifically, it determines whether a steam leak target is present in the sample image, and identifies the predicted bounding box, region category, and confidence score of the steam leak target within the sample image. The training loss is then calculated by comparing the training results with the region labels. This training loss optimizes the deep learning model's ability to detect steam leak targets. Therefore, the training loss is the sum of the category classification loss, bounding box regression loss, and distributed regression loss.

[0057] For example, the deep learning model being trained can be a model like YOLO v8. YOLO v8 employs an anchor-free design, allowing direct regression of target bounding boxes and categories via the detection head, thus simplifying the training process. When the server inputs a standardized training dataset into the YOLO v8 deep learning model for training, the deep learning model can extract spatial and semantic features from infrared images to optimize the model's ability to detect steam leak targets.

[0058] To enhance the model's generalization ability, YOLO v8's built-in Automatic Mixed Precision (AMP), optical flow transformation, and adaptive sample selection are employed to enrich the variability and diversity of the training samples. Data augmentation effectively alleviates the detection difficulty caused by the complexity of steam leak scenarios. During training, model parameters are optimized by minimizing the overall loss function. The overall loss function can be expressed as:

[0059] L total = L cls + L box + L dfl ;

[0060] in, L cls It is a category classification loss; L box It is the bounding box regression loss, which calculates the overlap between the predicted box and the ground truth box, the center distance, and the aspect ratio difference. Ldfl It is a distributed regression loss used to improve the accuracy of bounding box regression.

[0061] After calculating the training loss, the server can compare it with a pre-set loss threshold. When the training loss is greater than the threshold, it indicates that the deep learning model has not yet converged. Therefore, the model parameters can be adjusted based on the training loss, and iterative training can continue until the training loss is less than or equal to the loss threshold. At this point, the corresponding model parameters are output to obtain the target detection model. It is evident that through continuous iterative training, the deep learning model's ability to detect steam leak targets can be gradually optimized, enabling the target detection model to identify candidate regions from the image to be identified, providing accurate candidate region information.

[0062] After obtaining the object detection model through iterative training, the server can input the real-time acquired image to be identified into the trained object detection model for object detection, and then output candidate region information through the object detection model. The candidate region information includes a region prediction box, a region category, and a confidence score. That is, the object detection model can output the prediction box, category, and confidence score of the steam leak candidate region.

[0063] To obtain candidate region information, in some embodiments, when the server uses an object detection model to identify candidate regions from the image to be identified, it can first input the image to be identified into the object detection model to extract spatial and semantic features of the image. Then, based on the spatial and semantic features, the identification probability of the steam leak region in the image to be identified is calculated. Next, the region category is determined based on the label of the region with the highest identification probability, and a confidence score is calculated based on the identification probability. Thus, a region prediction box is delineated in the image to be identified according to the identification probability.

[0064] For example, after inputting real-time acquired infrared images into a standardized trained YOLO v8 object detection model, the model can output predicted bounding boxes, categories, and confidence scores for candidate regions of steam leaks. Then, a Non-Maximum Suppression (NMS) algorithm is used to filter the predicted bounding boxes, eliminating highly overlapping boxes and retaining only the optimal candidate regions. The final output is the coordinates, category label, and corresponding confidence score p for each candidate region; that is, the candidate region information can be represented as:

[0065] ;

[0066] in, Indicates candidate region information; x , y, w , h () represents the coordinates and dimensions of the bounding box. c As a category, p Let I be the confidence score and I be the input image. Through the above object detection process, candidate region image patches for physical feature modeling can be obtained, and the detection results of the candidate regions are shown as follows: Figure 4 As shown.

[0067] It should be noted that the object detection model described in the above embodiments, which identifies candidate regions from the image to be identified to obtain candidate region information, can serve as the first stage of the steam leakage data analysis method based on collaborative manifold learning. In the first stage, the candidate region information obtained by the object detection model can provide high-confidence candidate region locations, providing a spatially defined range for physical feature modeling, that is, providing candidate region coordinates and image basis for feature extraction and dimensionality reduction modeling in the subsequent stage (second stage).

[0068] S103. Based on candidate region information, extract the dynamic evolution features of the candidate regions in continuous images to be identified.

[0069] After identifying candidate regions from the image to be identified using a target detection model to obtain candidate region information, the server can execute the second stage of the steam leak data analysis based on collaborative manifold learning. In the second stage, the server can perform feature extraction and dimensionality reduction modeling based on the candidate region information, that is, extract the dynamic evolution features of the candidate regions in the continuous images to be identified based on the candidate region information. The dynamic evolution features include temperature distribution features, spatial diffusion morphology features, and time series change features.

[0070] like Figure 5 As shown, in order to extract dynamic evolution features, in some embodiments, when the server extracts the dynamic evolution features of the candidate region in consecutive images to be identified based on the candidate region information, it may first obtain associated video data based on the images to be identified. The associated video data includes multiple consecutive image frames, and the multiple consecutive image frames include the image to be identified.

[0071] To acquire associated video data, the server can extract a certain number or duration of image frames from a continuous image sequence based on the position of the current image to be identified within the video data stream. These extracted image frames, together with the current image to be identified, constitute a video of a certain length, i.e., associated video data. For example, for a single frame of the image to be identified in the video data stream, the server needs to acquire associated video data of at least 10 seconds. Therefore, the server needs to extract multiple consecutive image frames within 10 seconds of the current image to be identified from the video data stream. That is, when the infrared video surveillance device is a high-speed thermal imaging infrared camera with a frame rate of 60, the server needs to extract 600 consecutive infrared images from the video data stream to form associated video data.

[0072] After acquiring the associated video data, the server then marks the steam leakage areas in multiple consecutive image frames within the associated video data based on candidate regions, and extracts physical features from the associated video data according to the steam leakage areas. These physical features include temperature distribution data, spatial diffusion patterns, and time-series variation characteristics. Furthermore, by statistically analyzing the physical features in multiple consecutive image frames, a dynamic evolution feature is constructed.

[0073] In other words, the server can extract temperature distribution data, spatial diffusion patterns, and temporal variation characteristics of the candidate steam leakage regions detected in the first stage from continuous infrared video. These characteristics reflect the dynamic evolution of steam under actual operating conditions. The extracted physical information is then integrated into a high-dimensional physical feature set, comprehensively considering its spatial structure, heat distribution patterns, and temporal evolution trends. By constructing a complete, multi-dimensional candidate region feature space, a foundation is provided for subsequent dimensionality reduction.

[0074] For example, the server can extract features from the candidate steam leak regions obtained in the first stage of target detection to extract their physical features in continuous infrared video. These physical features include, but are not limited to, temperature distribution data T, spatial diffusion morphology S, and time-series variation features f. By performing statistical analysis on continuous frame data of the candidate regions, a multidimensional physical feature set, F={T, S, f}, is constructed and used as input to the collaborative manifold learning model for dimensionality reduction modeling.

[0075] S104. By using collaborative manifold learning, dynamic evolutionary features are embedded into a low-dimensional space to obtain low-dimensional embedded features.

[0076] After extracting dynamic evolution features from multiple consecutive frames associated with the image to be identified through feature extraction, the server can perform the second stage of the method, dimensionality reduction modeling. This involves embedding the dynamic evolution features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features. These low-dimensional embedded features are generated by extracting nonlinear structural relationships within the data from the dynamic evolution features.

[0077] In some embodiments, when the server performs the embedding of dynamically evolving features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features, it may first invoke the collaborative manifold learning module. This collaborative manifold learning module is a learning module constructed by combining manifold learning and collaborative learning algorithm frameworks. The collaborative manifold learning module can be configured to mine the intrinsic manifold structure of features in dynamically evolving features and the collaborative relationships between different features.

[0078] Based on the invoked collaborative manifold learning module, the server can input dynamically evolving features into the module to obtain a fused feature representation. Then, based on this fused feature representation, neighborhood relationships between different modalities are constructed, and adaptive nearest neighbor selection is performed according to these relationships to obtain the features to be embedded. Finally, a nonlinear manifold dimensionality reduction algorithm is used to compute the low-dimensional embedding representation of the features to be embedded, thus obtaining the low-dimensional embedding features. The fused feature representation is obtained by performing a weighted concatenation of multimodal features on the dynamically evolving features. The nonlinear manifold dimensionality reduction algorithm includes at least one of local linear embedding, t-distributed random nearest neighbor embedding, and isometric mapping embedding.

[0079] The server can input high-dimensional physical features into the collaborative manifold learning module. The collaborative manifold learning module can combine spatial, temporal, and temperature information to construct a high-dimensional physical feature space and use nonlinear manifold dimensionality reduction algorithms such as Local Linear Embedding (LLE), t-distributed random nearest neighbor embedding (t-SNE), or isomap to obtain a low-dimensional embedding representation to express the potential physical patterns of candidate targets.

[0080] For example, to more intuitively demonstrate the collaborative manifold learning method, when modeling the physical features of candidate regions, the server can be based on, for example... Figure 6 The core processing step shown performs dimensionality reduction modeling. As can be seen, the dimensionality reduction modeling process can include key steps such as high-dimensional feature construction, nonlinear dimensionality reduction, and low-dimensional embedding generation. During dimensionality reduction modeling, the server can first input multimodal features, which can include three different types of feature inputs: spatial features representing the spatial characteristics of the candidate region in the image to be identified; temporal features representing the characteristics of the candidate region data changing over time; and temperature image features representing the temperature data of the candidate region in the image to be identified, i.e., the temperature distribution in the thermal image.

[0081] Based on the multimodal feature input, the server can perform multimodal weighted concatenation at the feature fusion layer. This involves setting multimodal weighting values ​​to weight and fuse the three types of features together, forming a comprehensive feature representation. During the weighted concatenation process, different modal features can be weighted to highlight the importance of certain features.

[0082] After forming a fused feature representation through weighted concatenation, the server can perform adaptive nearest neighbor selection to construct cross-modal neighborhoods. In order to find similar or related data points between different modalities for better understanding and analysis of the data, the server can construct neighborhood relationships between different modalities and perform adaptive nearest neighbor selection to obtain the features to be embedded.

[0083] After obtaining the features to be embedded, the server can also use manifold optimization algorithms to preserve the local structure of the data. Manifold learning is a non-linear dimensionality reduction technique that aims to retain important structural information of the data while reducing dimensionality, thus preserving the data's intrinsic geometric structure in a lower-dimensional space.

[0084] Ultimately, the server can express physical patterns through a low-dimensional embedding space, mapping the processed data into this space to represent physical patterns. In this low-dimensional space, the data representation reflects its patterns or structures in the physical world, allowing for the extraction and expression of physical patterns within the data.

[0085] In some embodiments, when the server performs collaborative manifold learning to embed dynamic evolution features into a low-dimensional space to obtain low-dimensional embedded features, it can first extract feature trajectories of candidate regions in consecutive image frames based on the low-dimensional embedded features. Then, a neighborhood consistency analysis algorithm is applied to calculate the consistency metric parameters of the candidate regions on the feature trajectories. Furthermore, a trajectory clustering algorithm is used to identify abnormal trajectory behavior based on the spatiotemporal distribution and changing trends of trajectory points in the feature trajectories, thereby obtaining anomaly identification results. Based on the consistency metric parameters and the anomaly identification results, the diffusion pattern and thermal dynamic characteristics of the feature trajectories are evaluated.

[0086] For example, the server can use collaborative manifold learning techniques to process the aforementioned high-dimensional physical features, extract the nonlinear structural relationships within the data, and embed them into a low-dimensional space. This embedding significantly reduces the data dimensionality while preserving as much of the original physical behavior features as possible, thereby improving the efficiency and accuracy of subsequent analysis. Its dimensionality reduction objective can be formally expressed as:

[0087]

[0088] in, Z Represents the set of features to be reduced in dimensionality; NThis represents the total number of data points; F i and F j They represent the first i The and the first j Features of each data point; N ( i ) represents a sample i The neighborhood, ω ij To reconstruct the weights.

[0089] The server can then model and analyze the feature trajectories of candidate regions across consecutive image frames within the embedded low-dimensional manifold space. In the obtained low-dimensional manifold space Z, neighborhood consistency analysis and dynamic trajectory clustering algorithms are applied to model and verify the evolution trajectory of the physical behavior of candidate regions. The trajectory consistency metric of candidate targets in the embedding space is calculated. D ( z i , z j This assesses whether it conforms to the typical diffusion pattern and thermal dynamics characteristics of steam leakage. The neighborhood consistency measure can be formalized as:

[0090]

[0091] in, D ( z i , z j This represents a measure of trajectory consistency. z i and z j These represent the i-th and j-th state vectors, respectively; ∆ t Indicates a time interval.

[0092] Meanwhile, trajectory clustering algorithms such as DBSCAN and OPTICS are introduced to identify abnormal trajectory behavior by analyzing the spatiotemporal distribution and changing trends of trajectory points, and to determine whether candidate targets conform to the typical diffusion pattern and thermal dynamic characteristics of steam leakage.

[0093] S105. Generate analysis results information based on low-dimensional embedding features and preset steam leakage physical behavior features.

[0094] After obtaining the low-dimensional embedding features, the server can determine whether the image to be identified conforms to the preset steam leakage physical behavior features based on the low-dimensional embedding features and preset steam leakage physical behavior features, and generate analysis result information based on the judgment result. That is, when the low-dimensional embedding features corresponding to the image to be identified conform to the preset steam leakage physical behavior features, an analysis result can be obtained to characterize that the image to be identified contains a steam leakage target, indicating that there is a steam leakage phenomenon in the current on-site operation scenario. When the low-dimensional embedding features corresponding to the image to be identified do not conform to the preset steam leakage physical behavior features, an analysis result can be obtained to characterize that the image to be identified does not contain a steam leakage target. In this case, the server can continue to monitor the on-site operation scenario according to the analysis method provided in the above embodiments.

[0095] For example, in the reduced-dimensional space, the server can further analyze the evolution trajectory and diffusion pattern of candidate regions to determine whether they conform to the physical behavior characteristics of typical steam leaks. For regions that meet the characteristics, their embedding representation is output as the effective low-dimensional feature result. Abnormal samples are removed or marked to provide reliable input for subsequent identification, early warning, and other applications.

[0096] Therefore, in some embodiments, when the server generates analysis results based on the low-dimensional embedding features and preset steam leakage physical behavior features, it can first extract behavioral features from the low-dimensional embedding features and obtain typical steam leakage diffusion patterns. The behavioral features include spatial structure features and temporal evolution trajectories. The typical steam leakage diffusion patterns include preset steam leakage physical behavior features.

[0097] By comparing the behavioral features with preset physical behavioral features of steam leakage, if the behavioral features are consistent with the preset physical behavioral features of steam leakage, the low-dimensional embedding features corresponding to the behavioral features are retained. If the behavioral features are inconsistent with the preset physical behavioral features of steam leakage, the low-dimensional embedding features corresponding to the behavioral features are removed or labeled.

[0098] For example, the server can construct behavioral features by combining the spatial structure and temporal evolution trajectory of low-dimensional embedded features. If the behavioral features are highly consistent with typical steam leakage diffusion patterns, the low-dimensional embedded features are retained, and a low-dimensional feature set is output. The output low-dimensional feature set will serve as the input data basis for subsequent identification, clustering, or prediction models for subsequent analysis and modeling. If the behavioral features do not meet the consistency condition with typical steam leakage diffusion patterns, they are discarded to improve the effectiveness and expressive power of the final dimensionality reduction result.

[0099] As can be seen, by combining multidimensional physical feature modeling with manifold embedding methods, the server can effectively reduce data dimensionality, retain key information, and provide strong feature support for subsequent steam leak identification, trend judgment, and early warning. Furthermore, to adapt to the computational resource and real-time requirements of different application scenarios, the dimensionality reduction algorithm and candidate region extraction method can be flexibly adjusted according to actual application needs.

[0100] By applying the technical solutions of the above embodiments, the steam leakage data analysis method based on collaborative manifold learning provided in the above embodiments can fuse temperature, spatial, and temporal features to construct a high-dimensional physical feature set and perform dimensionality reduction. Furthermore, a nonlinear embedding algorithm is used to achieve manifold dimensionality reduction of candidate region features, improving the physical consistency and interpretability of the dimensionality reduction results. Trajectory clustering and neighborhood consistency analysis are also used to screen out low-dimensional features that conform to the evolution law of steam leakage, improving the effectiveness and expressive power of the dimensionality-reduced data. Therefore, the method not only effectively compresses data dimensionality and preserves key evolution patterns in the steam leakage process, but also possesses good physical consistency and semantic interpretability, providing high-quality low-dimensional embedded features for subsequent anomaly detection, behavior modeling, and intelligent analysis.

[0101] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a steam leakage data analysis method based on collaborative manifold learning. Based on the above embodiments, this method can also generate analysis result information based on the retained low-dimensional embedding features, such as... Figure 7 As shown, the method includes:

[0102] S201. Generate an effective low-dimensional feature set based on the retained low-dimensional embedding features;

[0103] S202. Calculate the classification probability of the image to be identified for the steam leak classification label based on the effective low-dimensional feature set;

[0104] S203. If the classification probability is greater than or equal to the probability threshold, generate the first analysis result information.

[0105] S204. If the classification probability is less than the probability threshold, generate the second analysis result information.

[0106] The server can filter low-dimensional embedding features based on the comparison results between behavioral characteristics and typical steam leakage diffusion patterns. The retained low-dimensional embedding features can then be collected and organized to generate an effective low-dimensional feature set. This effective low-dimensional feature set is then input into the analysis model, enabling the model to calculate the classification probability of the image to be identified for the steam leakage classification label. The analysis model can be a neural network model trained using sample images containing low-dimensional embedding features. The training process of the analysis model can be similar to the target recognition model training method described in the above embodiments, with differences only in the classification labels and loss functions, which will not be elaborated further here.

[0107] After inputting the effective low-dimensional feature set into the trained analysis model, the model outputs the classification probability of the effective low-dimensional feature set for the analysis result label. The server then compares the classification probability with a preset probability threshold to determine different analysis results. Specifically, if the classification probability is greater than or equal to the probability threshold, a first analysis result is generated. This first analysis result indicates that the image to be identified contains a steam leak target. If the classification probability is less than the probability threshold, a second analysis result is generated, indicating that the image to be identified does not contain a steam leak target.

[0108] By applying the technical solutions of the above embodiments, the steam leakage data analysis method based on collaborative manifold learning provided in the above embodiments can start from the physical properties of steam leakage image data, integrate multi-dimensional information for modeling, and improve the problems of insufficient feature modeling and weak dimensionality reduction expression ability. Furthermore, using the collaborative manifold learning method for nonlinear dimensionality reduction can better preserve the spatial structure and dynamic change characteristics of the leakage area, improving the structural preservation and semantic expression ability of the data dimensionality reduction results. In addition, by introducing trajectory consistency analysis and clustering mechanisms, low-dimensional features are effectively screened, enhancing the stability and applicability of the dimensionality reduction results and providing high-quality feature input for subsequent tasks such as identification, clustering, and prediction.

[0109] As can be seen, the proposed method can focus on infrared image data of steam leaks, which is characterized by high dimensionality, nonlinearity, strong spatiotemporal coupling, and significant dynamic features. A data dimensionality reduction technique based on collaborative manifold learning is proposed, which achieves a structured representation of steam leak data in a low-dimensional space through candidate region extraction, multi-dimensional physical feature construction, and nonlinear manifold embedding. Unlike shallow processing methods based solely on edge, brightness, or temperature thresholds, this method can integrate spatial diffusion patterns, temperature evolution trends, and temporal dynamic information. While maintaining the consistency of the original physical features, it significantly reduces the feature dimensionality, facilitating subsequent tasks such as clustering, recognition, and trend analysis.

[0110] Therefore, the method described above is applicable to the efficient compression and feature extraction of infrared image data under complex working conditions. It can be widely used in fields such as steam system anomaly analysis, industrial intelligent monitoring, behavior modeling and data-driven fault early warning, and has good engineering practicality and industrial promotion prospects.

[0111] In some embodiments, as a specific implementation of the steam leakage data analysis method based on collaborative manifold learning described in the above embodiments, some embodiments of this application also provide a steam leakage data analysis system based on collaborative manifold learning, such as... Figure 8 As shown, the system includes: infrared video surveillance equipment, server, communication equipment, and terminal equipment.

[0112] The infrared video monitoring device and the terminal device establish a communication connection with the server through the communication device; the infrared video monitoring device is configured to perform image acquisition of the on-site operating scene.

[0113] For example, to support the implementation of the steam leakage data analysis method based on collaborative manifold learning described in the above embodiments, the system may consist of three parts: an infrared camera monitoring device, a server, and the monitored object. The infrared camera monitoring device, as an infrared video monitoring device for image acquisition, can be installed around critical pipelines or equipment in the industrial site to continuously acquire infrared image sequences and obtain thermal distribution information of the equipment surface and surrounding area.

[0114] The server, acting as the server-side component, deploys a deep learning object detection model and a collaborative manifold learning algorithm module. This module receives infrared image data and performs candidate region detection, physical feature extraction, and dimensionality reduction. The monitored objects refer to critical thermal equipment or pipeline areas in industrial sites that pose a potential risk of steam leakage.

[0115] By linking the image acquisition terminal with the backend server, standardized acquisition and dimensionality reduction analysis of high-dimensional image data can be achieved, providing high-quality, structurally consistent low-dimensional feature representations for subsequent steam leakage behavior modeling and intelligent analysis. The server includes:

[0116] The image acquisition module is used to acquire the image to be identified, which is an infrared image obtained by the infrared video monitoring device from the on-site operating scene;

[0117] The candidate region identification module is used to identify candidate regions from the image to be identified using an object detection model to obtain candidate region information. The object detection model is a deep learning model trained using a training dataset. The candidate region information includes a region prediction box, a region category, and a confidence score.

[0118] The feature extraction module is used to extract the dynamic evolution features of the candidate region in the continuous image to be identified based on the candidate region information. The dynamic evolution features include temperature distribution features, spatial diffusion morphology features and time series change features.

[0119] The dimensionality reduction module is used to embed the dynamic evolution features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features, which are generated by extracting nonlinear structural relationships within the data from the dynamic evolution features.

[0120] The analysis module is used to generate analysis results information based on the low-dimensional embedded features and preset steam leakage physical behavior features.

[0121] By applying the technical solutions of the above embodiments, the steam leakage data analysis system based on collaborative manifold learning provided in the above embodiments can, after the image acquisition module acquires the image to be identified, the candidate region identification module uses a target detection model to identify candidate regions from the image to be identified. Then, the feature extraction module extracts the dynamic evolution features of the candidate regions in continuous images to be identified based on the candidate region information. Then, the dimensionality reduction module embeds the dynamic evolution features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features. Finally, the analysis module can generate analysis result information based on the low-dimensional embedded features and preset steam leakage physical behavior features. The system can integrate physical features such as spatial structure, temperature information, and temporal dynamics to construct a physically consistent high-dimensional feature expression, and achieve nonlinear dimensionality reduction through manifold embedding. The system can compress data dimensions while maintaining physical semantic consistency, effectively extract features related to the steam leakage evolution process, and improve the accuracy of steam leakage data analysis results.

[0122] It should be noted that other corresponding descriptions of the functional units involved in the steam leakage data analysis system based on collaborative manifold learning provided in the embodiments of this application can be found in the corresponding descriptions in the steam leakage data analysis method based on collaborative manifold learning provided in the above embodiments, and will not be repeated here.

[0123] This application also provides a computer device, specifically a personal computer, server, network device, etc. The computer device includes a bus, processor, memory, and communication interface, and may also include input / output interfaces and a display device. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores location information. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0124] Those skilled in the art will understand that the structure of the computer device described above is only a partial structure related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. A specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0125] In one embodiment, a computer-readable storage medium is also provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0126] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0129] Any references to memory, database, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc.

[0130] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take many forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0131] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for analyzing steam leakage data based on collaborative manifold learning, characterized in that, The method includes: Acquire an image to be identified, wherein the image to be identified is an infrared image obtained by an infrared video surveillance device performing image acquisition on the on-site operating scene; Candidate regions are identified from the image to be identified using an object detection model to obtain candidate region information. The object detection model is a deep learning model trained using a training dataset. The candidate region information includes region prediction boxes, region categories, and confidence scores. Based on the candidate region information, dynamic evolution features of the candidate regions in consecutive images to be identified are extracted. These dynamic evolution features include temperature distribution features, spatial diffusion morphology features, and temporal series variation features. Extracting these dynamic evolution features based on the candidate region information in consecutive images to be identified includes: acquiring associated video data based on the image to be identified, the associated video data including multiple consecutive image frames, each including the image to be identified; marking steam leakage areas in the multiple consecutive image frames of the associated video data according to the candidate regions; extracting physical features of the associated video data according to the steam leakage areas, the physical features including temperature distribution data, spatial diffusion morphology, and temporal series variation features; and constructing the dynamic evolution features by statistically analyzing the physical features in the multiple consecutive image frames. The dynamic evolutionary features are embedded into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features. These low-dimensional embedded features are generated by extracting nonlinear structural relationships within the data from the dynamic evolutionary features. The process of embedding the dynamic evolutionary features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features includes: calling a collaborative manifold learning module, which is a learning module constructed by combining manifold learning and collaborative learning algorithm frameworks; inputting the dynamic evolutionary features into the collaborative manifold learning module to obtain a fused feature representation, which is a concatenation result obtained by performing multimodal feature weighted concatenation on the dynamic evolutionary features; constructing neighborhood relationships between different modalities based on the fused feature representation; and constructing neighborhood relationships according to the neighborhood relationships. Adaptive nearest neighbor selection is used to obtain the features to be embedded; a nonlinear manifold dimensionality reduction algorithm is used to calculate the low-dimensional embedding representation of the features to be embedded, thereby obtaining the low-dimensional embedding features; the nonlinear manifold dimensionality reduction algorithm includes at least one of local linear embedding, t-distribution random nearest neighbor embedding, and isometric mapping embedding; based on the low-dimensional embedding features, feature trajectories of the candidate regions in consecutive image frames are extracted; a neighborhood consistency analysis algorithm is applied to calculate the consistency metric parameter of the candidate regions on the feature trajectories; according to the trajectory clustering algorithm, abnormal trajectory behavior is identified by the spatiotemporal distribution and change trend of trajectory points in the feature trajectories, thereby obtaining anomaly identification results; based on the consistency metric parameter and the anomaly identification results, the diffusion pattern and thermal dynamic characteristics of the feature trajectories are evaluated; The analysis results are generated based on the low-dimensional embedding features and the preset steam leakage physical behavior features, including: extracting behavioral features from the low-dimensional embedding features, the behavioral features including spatial structure features and temporal evolution trajectories; obtaining typical steam leakage diffusion patterns, the typical steam leakage diffusion patterns including the preset steam leakage physical behavior features; if the behavioral features are consistent with the preset steam leakage physical behavior features, retaining the low-dimensional embedding features corresponding to the behavioral features; if the behavioral features are inconsistent with the preset steam leakage physical behavior features, removing or marking the low-dimensional embedding features corresponding to the behavioral features.

2. The method according to claim 1, characterized in that, Obtain the image to be recognized, including: Receive video stream data sent by the infrared video monitoring device; The video stream data is subjected to frame segmentation to obtain a continuous image sequence, which includes multiple consecutive image frames; Based on a preset normalization term, the image frames in the continuous image sequence are normalized to obtain the image to be identified. The preset normalization term includes size normalization, format normalization, and pixel normalization to adapt to the input requirements of the target detection model.

3. The method according to claim 1, characterized in that, The method further includes: Obtain a training dataset, which includes sample images and region labels, wherein the region labels are used to label the steam leak areas in the sample images; The sample image is input into the trained deep learning model to obtain the training result output by the deep learning model. The training loss is calculated by comparing the training results with the region labels. The training loss is the sum of the category classification loss, the bounding box regression loss, and the distributed regression loss. The deep learning model is iteratively trained based on the training loss to obtain the object detection model.

4. The method according to claim 1, characterized in that, Using an object detection model to identify candidate regions from the image to be identified, candidate region information is obtained, including: The image to be identified is input into the target detection model to extract the spatial and semantic features of the image to be identified. Based on the spatial features and the semantic features, the probability of identifying the steam leak area in the image to be identified is calculated; The region category is determined based on the region label with the highest recognition probability, and the confidence score is calculated based on the recognition probability. Based on the recognition probability, a region prediction box is delineated in the image to be recognized.

5. The method according to claim 1, characterized in that, Analysis results are generated based on the low-dimensional embedding features and preset steam leakage physical behavior features, including: A valid set of low-dimensional features is generated based on the retained low-dimensional embedding features; The classification probability of the image to be identified for the steam leak classification label is calculated based on the effective low-dimensional feature set. If the classification probability is greater than or equal to the probability threshold, first analysis result information is generated, which is used to characterize that the image to be identified contains a steam leak target. If the classification probability is less than the probability threshold, a second analysis result is generated. The second analysis result is used to characterize that the image to be identified does not contain a steam leak target.

6. A steam leakage data analysis system based on collaborative manifold learning, characterized in that, The system includes: an infrared video monitoring device, a server, a communication device, and a terminal device; wherein, the infrared video monitoring device and the terminal device establish a communication connection with the server through the communication device; the infrared video monitoring device is configured to perform image acquisition of the on-site operating scene; the server includes: The image acquisition module is used to acquire an image to be identified, which is an infrared image obtained by the infrared video surveillance equipment from the on-site operating scene. The candidate region identification module is used to identify candidate regions from the image to be identified using an object detection model to obtain candidate region information. The object detection model is a deep learning model trained using a training dataset. The candidate region information includes a region prediction box, a region category, and a confidence score. The feature extraction module is used to extract dynamic evolution features of the candidate regions in consecutive images to be identified based on the candidate region information. The dynamic evolution features include temperature distribution features, spatial diffusion morphology features, and time-series change features. Extracting dynamic evolution features of the candidate regions in consecutive images to be identified based on the candidate region information includes: acquiring associated video data based on the image to be identified, the associated video data including multiple consecutive image frames, the multiple consecutive image frames including the image to be identified; marking steam leakage areas in the multiple consecutive image frames of the associated video data according to the candidate regions; extracting physical features of the associated video data according to the steam leakage areas, the physical features including temperature distribution data, spatial diffusion morphology, and time-series change features; and constructing the dynamic evolution features by statistically analyzing the physical features in the multiple consecutive image frames. The dimensionality reduction module is used to embed the dynamic evolution features into a low-dimensional space through collaborative manifold learning to obtain low-dimensional embedded features. These low-dimensional embedded features are generated by extracting nonlinear structural relationships within the data from the dynamic evolution features. The process of embedding the dynamic evolution features into a low-dimensional space through collaborative manifold learning includes: calling a collaborative manifold learning module, which is a learning module constructed by combining manifold learning and collaborative learning algorithm frameworks; inputting the dynamic evolution features into the collaborative manifold learning module to obtain a fused feature representation, which is a concatenation result obtained by performing multimodal feature weighted concatenation on the dynamic evolution features; constructing neighborhood relationships between different modalities based on the fused feature representation; and following the... The neighborhood relationship adaptive nearest neighbor selection is used to obtain the feature to be embedded; a nonlinear manifold dimensionality reduction algorithm is used to calculate the low-dimensional embedding representation of the feature to be embedded, thereby obtaining the low-dimensional embedding feature; the nonlinear manifold dimensionality reduction algorithm includes at least one of local linear embedding, t-distribution random nearest neighbor embedding, and isometric mapping embedding; based on the low-dimensional embedding feature, the feature trajectory of the candidate region under consecutive image frames is extracted; a neighborhood consistency analysis algorithm is applied to calculate the consistency metric parameter of the candidate region on the feature trajectory; according to the trajectory clustering algorithm, abnormal trajectory behavior is identified by the spatiotemporal distribution and change trend of trajectory points in the feature trajectory, thereby obtaining anomaly identification results; based on the consistency metric parameter and the anomaly identification results, the diffusion mode and thermal dynamic characteristics of the feature trajectory are evaluated; The analysis module is used to generate analysis result information based on the low-dimensional embedding features and preset steam leakage physical behavior features, including: extracting behavioral features from the low-dimensional embedding features, the behavioral features including spatial structure features and temporal evolution trajectory; obtaining typical steam leakage diffusion patterns, the typical steam leakage diffusion patterns including the preset steam leakage physical behavior features; if the behavioral features are consistent with the preset steam leakage physical behavior features, retaining the low-dimensional embedding features corresponding to the behavioral features; if the behavioral features are inconsistent with the preset steam leakage physical behavior features, removing or marking the low-dimensional embedding features corresponding to the behavioral features.

Images