Multi-dimensional leakage source positioning method based on 360-degree infrared detection of building

Abstract

The invention relates to the technical field of building leakage detection, and discloses a leakage source multi-dimensional positioning method based on building 360-degree infrared detection. The method comprises the following steps: acquiring 360-degree infrared image data of the periphery of a building through an infrared scanning device, and synchronously acquiring real-time environmental parameters; the method comprises the following steps: performing flat field correction and radiation calibration preprocessing on infrared image data to generate standard infrared image data, and extracting temperature gradient features and thermal anomaly features from the standard infrared image data to form an initial feature set; and completing normalization processing of the initial feature set in combination with real-time environmental parameters, inputting the initial feature set into a convolutional neural network model for feature learning, and outputting a thermal anomaly probability graph. And performing clustering analysis on the thermal anomaly probability graph to identify a thermal anomaly region, and determining the position coordinates of the leakage source through multi-dimensional space analysis combining deep learning and geometric constraint. The method can eliminate detection blind areas and reduce environmental interference, is accurate in positioning and wide in applicability, and is suitable for peripheral leakage detection of various buildings.

Description

Technical Field

[0001] This invention relates to the field of building leakage detection technology, specifically a multi-dimensional method for locating leakage sources based on 360-degree infrared detection of buildings. Background Technology

[0002] Building leakage is a long-standing and prominent problem in the construction engineering field, related to a variety of factors including construction techniques, material performance degradation, and environmental erosion. Leakage not only damages the aesthetics of a building's facade but also seeps into the interior, causing mold growth on walls and peeling of decorative layers, affecting the indoor living and user experience. More seriously, long-term leakage can corrode reinforced concrete and other structural components, leading to problems such as steel corrosion and concrete carbonization, weakening the building's load-bearing capacity and durability, creating safety hazards, and significantly shortening the building's normal service life. Furthermore, the repair process requires substantial manpower, material resources, and time, causing unnecessary losses to building owners and users.

[0003] Currently, various technical methods exist for detecting building leaks. Among traditional methods, visual inspection is the simplest and most direct, but it heavily relies on the experience and judgment of the inspectors, making it highly subjective. Furthermore, it is limited by the inspection perspective, failing to comprehensively observe high points, concealed areas, and the entire 360-degree range of the building facade, easily overlooking hidden leak sources and compromising reliability. Ultrasonic detection utilizes the differences in the propagation of ultrasonic waves in different media to determine leak conditions. However, ultrasonic waves have limited penetration capabilities, making them unsuitable for scenarios with thick building insulation and decorative layers. They are also easily affected by the building surface material and flatness, leading to signal distortion and difficulty in accurately locating leak sources. Radar detection analyzes the internal structure and leak conditions of a building through the principle of electromagnetic wave reflection. However, electromagnetic wave propagation is easily interfered with by environmental parameters such as temperature, humidity, and wind speed. Especially under complex climatic conditions, signal attenuation is significant, resulting in large location errors and an inability to achieve multi-dimensional and accurate identification of leak sources.

[0004] Infrared detection technology has been gradually applied in building leakage detection due to its advantages of being non-contact, fast, efficient, and intuitive. However, there are still many problems to be solved in the existing related technologies. Current infrared detection methods mostly employ local scanning, lacking the ability to capture data across a 360-degree perimeter of a building, resulting in significant blind spots. This makes it difficult to effectively detect leak sources in hard-to-observe areas such as building corners and rear facades, leading to incomplete detection. Furthermore, existing technologies often neglect the influence of real-time environmental parameters when processing infrared image data. Fluctuations in ambient temperature, humidity, and wind speed can cause false signals or distortions in thermal features within infrared images, affecting the accuracy of feature extraction. Current technologies also tend to extract infrared image features in a limited way, focusing only on thermal anomalies and failing to integrate multiple dimensions such as temperature gradients for comprehensive analysis. This results in an initial feature set that lacks comprehensiveness and fails to fully reflect the differences in thermal distribution related to leaks. In the leak source localization stage, current technologies rely heavily on single analytical algorithms, lacking multi-dimensional spatial analysis methods that combine deep learning and geometric constraints. This fails to fully explore the intrinsic relationship between thermal anomaly areas and leak sources, leading to insufficient localization accuracy. This makes it difficult to meet the precise leak source localization requirements in practical engineering projects, impacting the targeted and effective nature of subsequent repair work. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-dimensional method for locating leakage sources based on 360-degree infrared detection of buildings, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a multi-dimensional method for locating leakage sources based on 360-degree infrared detection of buildings, the method comprising: The infrared scanning device collects 360-degree infrared image data of the building's exterior and simultaneously obtains real-time environmental parameters. The infrared image data is preprocessed, including flat field correction and radiometric calibration, to generate standard infrared image data; Temperature gradient features and thermal anomaly features are extracted from the standard infrared image data to form an initial feature set; The initial feature set is normalized by combining the real-time environmental parameters to obtain a normalized feature set; The normalized feature set is input into a convolutional neural network model for feature learning, and a hot anomaly probability map is output. Cluster analysis was performed on the thermal anomaly probability map to identify thermal anomaly regions; Multi-dimensional spatial analysis, including deep learning and geometric constraints, is performed based on the thermal anomaly region to determine the coordinates of the leakage source location.

[0007] Preferably, the method of acquiring 360-degree infrared image data of the building's perimeter using an infrared scanning device, while simultaneously obtaining real-time environmental parameters, includes: The infrared scanning device is controlled to perform a 360-degree rotating scan around the perimeter of the building at a preset scanning speed, acquiring continuous infrared image frames. During the scanning process, ambient temperature, humidity and wind speed data are collected synchronously by environmental sensors as the real-time environmental parameters. The continuous infrared image frames are timestamped to generate a time-synchronized infrared image sequence; The infrared image sequence and the real-time environmental parameters are stored in the database.

[0008] Preferably, the preprocessing operation on the infrared image data includes flat field correction and radiometric calibration to generate standard infrared image data, and the method includes: A flat-field correction algorithm is applied to each image frame in the infrared image sequence to eliminate optical system errors and obtain a corrected image; The corrected image is radiometrically calibrated based on a blackbody radiation reference source, and pixel values ​​are converted into temperature values ​​to generate a temperature image. Spatial filtering is performed on the temperature image to remove high-frequency noise, thereby obtaining the standard infrared image data.

[0009] Preferably, the method for extracting temperature gradient features and thermal anomaly features from the standard infrared image data to form an initial feature set includes: Calculate the spatial temperature gradient matrix of the standard infrared image data, and extract the gradient magnitude and direction features as temperature gradient features; Potential thermal anomaly regions are identified from the standard infrared image data using a threshold segmentation method, and the statistical features of these regions are calculated as thermal anomaly features. The temperature gradient features and thermal anomaly features are combined into a multi-dimensional vector to form the initial feature set.

[0010] Preferably, the method for normalizing the initial feature set by combining the real-time environmental parameters to obtain a normalized feature set includes: Calculate the environmental compensation coefficient based on the ambient temperature and humidity data in the real-time environmental parameters; The environmental compensation coefficient is used to scale and adjust each feature in the initial feature set to achieve environmental compensation; The compensated features are normalized to zero-mean, so that the feature mean is zero and the variance is one, thus generating the normalized feature set.

[0011] Preferably, the method of inputting the normalized feature set into a convolutional neural network model for feature learning and outputting a heat anomaly probability map includes: Construct a convolutional neural network model whose input layer receives the normalized feature set and whose hidden layer includes multiple convolutional layers and pooling layers; The process of constructing the convolutional neural network model includes: Set the input size of the input layer to match the dimension of the normalized feature set: Multiple convolutional layers and pooling layers are stacked sequentially after the input layer, wherein each convolutional layer contains multiple learnable filters for extracting local features, and each pooling layer is used to reduce the spatial size of the feature map; A fully connected layer is added after the stacked convolutional and pooling layers to integrate high-level features; An output layer is connected after the fully connected layer. The number of nodes in the output layer corresponds to the size of the thermal anomaly probability map, and the leakage probability of each pixel is output using the Sigmoid activation function. The convolutional neural network model is trained using historical leakage data as labels to learn feature mappings; The thermal anomaly probability map is generated by calculating the leakage probability of each pixel through forward propagation.

[0012] Preferably, the method for performing cluster analysis on the thermal anomaly probability map to identify thermal anomaly regions includes: A density-based spatial clustering algorithm is applied to cluster high-probability points in the thermal anomaly probability map to form candidate clusters. Calculate the centroid location and spatial extent of each candidate cluster, and select clusters that meet the size threshold as thermal anomaly regions; Record the boundary coordinates and probability values ​​of the thermal anomaly region.

[0013] Preferably, the method for performing multi-dimensional spatial analysis based on the thermal anomaly region, including deep learning and geometric constraints, to determine the coordinates of the leakage source location includes: Multidimensional features, including thermal inertia features and texture features, are extracted from the thermal anomaly region. A pre-trained deep learning model is used to perform regression analysis on the multi-dimensional features to predict the depth information of the leakage source. The prediction results are geometrically constrained and optimized by combining the building geometric model, the spatial position is corrected, and the coordinates of the leakage source are output.

[0014] Preferably, the method for using a pre-trained deep learning model to perform regression analysis on the multi-dimensional features to predict the depth information of the leakage source includes: A fully connected neural network is constructed as a regression model, with the multi-dimensional features as input and depth values ​​as output; The regression model is trained using training data with depth annotations, using the minimum mean squared error loss function. The multidimensional features of the thermal anomaly region are input into the trained regression model to obtain the depth prediction value of each region.

[0015] Preferably, the method of combining the building geometric model to optimize the prediction results by geometric constraints, correcting the spatial position, and outputting the coordinates of the leakage source location includes: Load the 3D geometric model of the building and obtain the surface structure information of the building; The two-dimensional coordinates and depth prediction values ​​of the thermal anomaly region are mapped onto a three-dimensional geometric model to generate a three-dimensional point cloud; The three-dimensional point cloud is registered with the geometric model by an iterative nearest point algorithm to optimize the positional accuracy and finally output the coordinates of the leakage source.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By using an infrared scanning device to acquire 360-degree infrared image data of the building's exterior, this method completely changes the traditional infrared detection model that only scans local areas. It achieves comprehensive, all-around coverage of the building's exterior, effectively eliminating blind spots caused by limited viewing angles and insufficient coverage in traditional detection methods. This ensures that leakage-related thermal anomalies in any area of ​​the building can be accurately captured, guaranteeing comprehensive detection from the source of data acquisition and preventing the omission of leakage sources. Simultaneously, real-time environmental parameters are acquired during the acquisition process, fully considering the potential impact of environmental factors on the infrared detection results. This provides crucial environmental reference data for subsequent feature data processing and lays the foundation for avoiding environmental interference.

[0017] Flat-field correction and radiometric calibration preprocessing of the acquired infrared image data effectively correct image distortion and uneven grayscale caused by differences in the optical characteristics of the equipment itself, sensor performance, and external environmental interference during image acquisition. This optimizes image quality, making the generated standard infrared image data more realistic and objective in reflecting the actual thermal distribution of the building surface, providing high-quality and reliable data support for subsequent feature extraction. By simultaneously extracting temperature gradient features and thermal anomaly features from the standard infrared image data to form an initial feature set, compared to the single feature extraction methods of existing technologies, the feature dimensions are enriched. This allows for a more comprehensive and multi-layered reflection of the differences in thermal distribution on the building surface, accurately capturing thermal feature information related to leakage phenomena, making the initial feature set more representative and discriminative, and better meeting the actual needs of leakage source detection.

[0018] Normalizing the initial feature set using real-time environmental parameters effectively eliminates the impact of fluctuations in parameters such as temperature, humidity, and wind speed under different environmental conditions on thermal feature data. This provides a unified standard for comparison and analysis, avoiding feature distortion and bias caused by environmental factors, ensuring the stability and consistency of feature data, and providing more reliable input data for subsequent model feature learning, thus improving the accuracy of model analysis results. Inputting the normalized feature set into a convolutional neural network model for feature learning fully leverages the advantages of convolutional neural networks in deep feature mining and complex pattern recognition. This allows for in-depth analysis of the inherent patterns and correlations related to leakage sources hidden within the feature data, accurately distinguishing between valid and interfering features. The output thermal anomaly probability map clearly and intuitively reflects the likelihood of thermal anomalies in various areas of the building, providing a strong reference for the accurate identification of subsequent thermal anomaly areas.

[0019] Cluster analysis of thermal anomaly probability maps can quickly aggregate and classify pixels or regions with similar thermal anomaly characteristics, effectively filtering out truly leak-related thermal anomaly areas and eliminating false thermal anomaly signals caused by environmental interference, differences in building surface materials, etc. This improves the accuracy of thermal anomaly area identification, reduces invalid analysis areas, increases overall detection efficiency, and makes detection work more targeted. Based on the identified thermal anomaly areas, multi-dimensional spatial analysis incorporating deep learning and geometric constraints is conducted. Deep learning further explores the spatial distribution patterns, morphological characteristics, and intrinsic connections between thermal anomaly areas and leakage sources. Combined with geometric constraints, key information such as the positional relationship and distance parameters of thermal anomaly areas in the building space is clarified, enabling accurate derivation and locking of the leakage source location from the thermal anomaly area. This avoids positioning errors caused by single analysis methods, making the final determined leakage source location coordinates more consistent with the actual situation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the multi-dimensional location method for leakage sources based on 360-degree infrared detection of buildings as described in this invention. Figure 2 A flowchart for acquiring infrared image data and environmental parameters; Figure 3 Flowchart for infrared image data preprocessing; Figure 4 A diagram showing the temperature gradient and thermal anomaly characteristics; Figure 5 This is a graph showing the probability of thermal anomalies and cluster analysis. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a multi-dimensional method for locating leakage sources based on 360-degree infrared detection of buildings. The method includes: acquiring 360-degree infrared image data of the building's perimeter using an infrared scanning device, and simultaneously acquiring real-time environmental parameters; performing preprocessing operations on the infrared image data, such as flat-field correction and radiometric calibration, to generate standard infrared image data; extracting temperature gradient features and thermal anomaly features from the standard infrared image data to form an initial feature set; normalizing the initial feature set in conjunction with real-time environmental parameters to obtain a normalized feature set; inputting the normalized feature set into a convolutional neural network model for feature learning, and outputting a thermal anomaly probability map; performing cluster analysis on the thermal anomaly probability map to identify thermal anomaly regions; and performing multi-dimensional spatial analysis based on the thermal anomaly regions, including deep learning and geometric constraints, to determine the coordinates of the leakage source location.

[0023] Example 1: See Figure 2 In practical implementation, the infrared scanning device rotates 360 degrees around the building's perimeter at a preset scanning speed, acquiring continuous infrared image frames. The preset scanning speed is set according to the building's size and detection resolution requirements. For example, for large building structures, the preset scanning speed is adjusted to a lower value to obtain high spatial resolution infrared image data, while for conventional buildings, the preset scanning speed can be set to a higher value to improve detection efficiency. The infrared scanning device includes an infrared camera assembly and a rotating platform assembly. The infrared camera assembly uses an uncooled microbolometer or a cooled photon detector, which has high thermal sensitivity and a wide dynamic range, capable of capturing subtle temperature changes on the building surface. The rotating platform assembly is driven by a stepper motor or servo motor to achieve smooth and precise 360-degree rotation. The control unit coordinates the scanning process. The control unit is based on an embedded processor design, runs a real-time operating system, sends control commands to the infrared scanning device, such as starting the scan, adjusting the scan speed, and stopping the scan, and receives infrared image frame data. During the scanning process, the infrared scanning device rotates at a constant angular velocity to ensure that the image frame covers the entire circumference of the building. The acquisition frame rate of continuous infrared image frames is matched with the scanning speed to avoid image overlap or gaps. The acquisition frame rate is set by programming through the control unit, for example, acquiring multiple frames per second to ensure the continuity of the image sequence.

[0024] In practical implementation, environmental sensors simultaneously collect ambient temperature, humidity, and wind speed data as real-time environmental parameters. These sensors are deployed near the infrared scanning device to minimize spatial deviation. The ambient temperature sensor uses platinum resistance or thermocouple elements, the ambient humidity sensor is based on capacitive or resistive principles, and the ambient wind speed sensor uses an ultrasonic or cup-type design. Data acquisition from the environmental sensors and infrared image frame acquisition are synchronized with a hardware mechanism, such as using a unified clock source to trigger acquisition events, ensuring time consistency. Ambient temperature, humidity, and wind speed data are output as digital signals and transmitted to the control unit via a serial communication interface such as RS-485 or I2C. The control unit performs preliminary data processing, such as filtering to remove transient noise and adding timestamps. The synchronous acquisition process is monitored by the control unit; any sensor malfunction or data anomaly triggers an alarm and logs the data for subsequent analysis. The environmental parameter sampling frequency is set according to application requirements, typically higher than the image acquisition frame rate, to capture rapid environmental changes. Sampled data is temporarily stored in a buffer, awaiting alignment with the image frames.

[0025] In practice, continuous infrared image frames are timestamped to generate a time-synchronized infrared image sequence. Timestamp alignment is achieved using a high-precision timer, which employs GPS synchronization or a crystal oscillator as the time reference, with deviations controlled within milliseconds. Each infrared image frame and environmental parameter data packet is marked with a precise timestamp, using either Unix timestamps or a custom time encoding format. The alignment algorithm handles asynchronous data points using interpolation methods. For example, for the deviation between the timestamp of an infrared image frame and the timestamp of environmental parameters, linear interpolation or spline interpolation is used to estimate the missing environmental parameter values, generating an environmental parameter set that precisely matches each image frame. The time-synchronized infrared image sequence is organized as a frame sequence, with each frame containing image data and corresponding environmental parameters. The sequence structure supports random access and streaming processing. The alignment operation is performed in the control unit or on the backend server. Software modules handle timestamp parsing and data matching, ensuring the time consistency of the entire dataset.

[0026] In practical implementation, infrared image sequences and real-time environmental parameters are stored in a database. The database employs a relational database management system or a time-series database system, designed with a multi-table structure, including an image data table, an environmental parameter table, and a time index table. The image data table stores the binary or compressed data of the infrared image frames, the environmental parameter table stores numerical records of ambient temperature, humidity, and wind speed, and the time index table links the image data table and the environmental parameter table through a timestamp field for efficient querying. The storage process includes data verification and compression steps. Data verification checks the integrity of image frames and the rationality of parameters, while compression algorithms such as JPEG or H.264 reduce storage space usage. The database supports transaction processing to ensure data atomicity and consistency and provides a backup mechanism to prevent data loss. The storage interface is implemented through an API or direct database connection, allowing batch import and real-time writes. In some embodiments, the database is deployed on a local server or cloud platform, featuring a scalable architecture to adapt to large-scale data storage needs. Optionally, the database integrates data management tools to support data retrieval, visualization, and export functions.

[0027] In practical implementation, the control logic of the infrared scanning device includes a state machine model, which defines various states during the scanning cycle, such as initialization, scanning, pause, and termination states. The control unit automatically switches states according to a preset scanning protocol, which includes the scanning speed curve, acquisition frame rate, and environmental parameter sampling plan. Calibration of the infrared scanning device is performed periodically, using a standard blackbody source to verify temperature measurement accuracy. Calibration data is stored in a calibration table for subsequent data correction. The placement of environmental sensors considers the influence of micro-environmental factors, such as avoiding measurement deviations caused by direct sunlight or poor ventilation. Sensor positions are adjusted using fixed brackets to maintain consistent relative distances to building surfaces. Timestamp alignment accuracy is verified through simulation tests, such as injecting data packets with known time delays to check if the alignment error is within acceptable limits. Database index optimization uses B-trees or hash indexes to accelerate time range queries, and data partitioning strategies are used to divide data by time or building identifiers to improve query performance. The entire acquisition system is designed with a modular architecture, allowing for component replacement or upgrades, such as replacing the infrared camera with a higher resolution one or adding more types of environmental sensors.

[0028] In practical implementation, the preset scanning speed is set based on the building's geometric features and the target being detected. For example, for high-rise buildings, the preset scanning speed is set to slow to ensure sufficient sampling density in the vertical direction, while for low-rise buildings, the preset scanning speed can be increased. The rotating platform of the infrared scanning device is equipped with encoder feedback to monitor the rotation angle and speed in real time. The encoder data is used for closed-loop control to correct speed fluctuations. The acquisition of continuous infrared image frames is triggered by either hardware or software. Hardware triggering uses photoelectric sensors to detect the rotation position, ensuring that image frames are acquired at fixed angular intervals. Software triggering is based on timer interrupts to achieve uniform time interval acquisition. Synchronous acquisition of environmental sensors uses interrupt service routines to process environmental data with high priority, avoiding the loss of rapidly changing parameters. The timestamp alignment algorithm is optimized on the embedded system to reduce computational latency. Sequence numbers are added to the aligned data packets to ensure correct order. Database storage uses redundant arrays or distributed storage to enhance data reliability, and access control mechanisms restrict unauthorized operations. In some embodiments, infrared image sequences and environmental parameters are streamed to a remote monitoring center in real time, and data transmission integrity is ensured through network protocols such as TCP / IP. Optionally, the data acquisition system integrates self-diagnostic functions to periodically check the equipment status, such as the focal length of the infrared camera or the sensitivity of the sensor, and automatically generate maintenance reports.

[0029] In practical implementation, the installation location of the infrared scanning device is chosen considering field of view coverage and obstacle avoidance, such as deployment at the building's center point or on a mobile platform, stably supported by tripods or fixed bases. A graphical configuration tool is provided for adjusting the preset scanning speed; the operator inputs the building dimensions and desired resolution, and the system automatically calculates the optimal scanning speed. The calibration cycle of the environmental sensors is synchronized with the infrared scanning device, using standard environmental sources such as constant temperature and humidity chambers for periodic calibration, with calibration coefficients stored in non-volatile memory. The timestamp alignment processing pipeline is designed for parallelization, utilizing multi-core processors to accelerate interpolation calculations, and the alignment results are displayed in real-time on the user interface for operator confirmation. The database architecture supports concurrent access by multiple users, transaction isolation levels are set to prevent data conflicts, and the query optimizer analyzes the execution plan to improve retrieval efficiency. The energy management of the acquisition system adopts an energy-saving mode, reducing equipment power consumption during idle periods to extend its lifespan. The entire implementation process is documented, recording configuration parameters and operation logs for easy auditing and reproduction.

[0030] In practical implementation, the scanning path of the infrared scanning device is programmable, supporting custom start and end points to adapt to asymmetrical building structures. The storage format for continuous infrared image frames is selected from standard image formats such as RAW or TIFF, preserving original data accuracy, or compressed formats such as JPEG to balance quality and size. The environmental parameter data structure includes metadata fields, such as sensor identification and acquisition status, enhancing data traceability. The timestamp alignment error handling mechanism includes retry logic; when the alignment deviation exceeds a threshold, the relevant data segment is automatically reacquired. The database backup strategy combines incremental and full backups, periodically synchronizing to off-site storage to ensure disaster recovery. In some embodiments, the acquisition system is integrated with a building information modeling system, automatically associating building identification and scanning data to simplify project management. Optionally, the user interface provides a real-time preview function, displaying the current scanning progress and environmental parameter curves to assist on-site decision-making. It is understood that the system design follows industry standards, such as IP-rated dust and water resistance, adapting to harsh outdoor environments.

[0031] Example 2: See Figure 3 In its implementation, a flat-field correction algorithm is applied to each image frame in the infrared image sequence to eliminate optical system errors. The algorithm calculates the gain and offset coefficients based on a reference image, which is obtained by capturing images in a uniform temperature scene. This uniform temperature scene is achieved using a large blackbody plate or a constant-temperature surface to ensure radiation uniformity. The flat-field correction algorithm process includes acquiring dark-field and flat-field images. The dark-field image is acquired under conditions where the infrared scanning device lens is completely blocked, characterizing the system's inherent noise and dark current. The flat-field image is acquired in a uniform radiation field, characterizing the ideal optical response. The correction formula transforms the pixel values ​​of each image frame in the original infrared image sequence by subtracting the corresponding pixel value from the dark-field image, dividing by the corresponding pixel value of the flat-field image, and multiplying by a normalization coefficient to output the corrected image. The corrected image has a more uniform pixel value distribution, eliminating errors such as lens vignetting and sensor non-uniformity. The corrected image is stored in floating-point format to preserve the dynamic range. The parameters of the flat-field correction algorithm are configurable; for example, the gain and offset coefficients are determined through calibration experiments. These calibration experiments are performed periodically, using a standard radiation source to verify the correction effect.

[0032] In practice, the corrected image is radiometrically calibrated based on a blackbody radiation reference source. Pixel values ​​are converted into temperature values ​​to generate a temperature image. The blackbody radiation reference source is placed within the field of view of the infrared scanning device, and its temperature is precisely controllable, maintained at a stable temperature value by a temperature control unit. The radiometric calibration process applies Planck's radiation law to establish the physical relationship between pixel grayscale values ​​and radiation intensity. By measuring the image pixel values ​​of the blackbody radiation reference source at different temperatures, a calibration curve is fitted. The calibration curve uses a polynomial regression or lookup table to map each pixel value of the corrected image to an absolute temperature value, with the temperature unit selectable as Celsius or Kelvin. The data structure for generating the temperature image includes a temperature matrix and metadata. The metadata records the calibration time, blackbody temperature setting, and environmental parameters. Radiometric calibration accuracy is improved through multi-point calibration, such as acquiring blackbody images at multiple temperature points to minimize fitting errors. The calibration results are stored in a calibration file for subsequent processing.

[0033] In practice, spatial filtering is performed on the temperature image to remove high-frequency noise, resulting in standard infrared image data. The spatial filter chosen is either a Gaussian filter or a median filter. The kernel size and standard deviation of the Gaussian filter are adjusted based on the image resolution and noise characteristics, while the window size of the median filter is determined based on noise statistics. The spatial filtering operation is performed in the spatial domain, convolving the temperature image with the filter kernel to smooth random noise while preserving edge information. The filtered standard infrared image data is stored in matrix form, with pixel values ​​representing temperature values. The data format supports subsequent feature extraction. Spatial filtering parameters are optimized experimentally, for example, by analyzing the noise power spectrum to select the optimal filter type. The filtering effect is evaluated using the signal-to-noise ratio (SNR). The standard infrared image data output is a multi-band image containing temperature information and spatial coordinates, facilitating integration into the processing pipeline.

[0034] In practical implementation, the flat-field correction algorithm utilizes a dedicated image processing library. This library provides flat-field correction functions, taking the original infrared image sequence, dark-field image, and flat-field image as input, and outputting a corrected image sequence. Quality checks on the corrected images include uniformity testing and calculating the image standard deviation to verify the correction effect. If the uniformity fails to meet the standard, a new reference image is acquired. The blackbody radiation reference source for radiometric calibration is integrated into the scanning system. The blackbody radiation reference source has a fixed position to avoid obstructing the building's field of view. The calibration process is automated, with software controlling the temperature control unit and image acquisition. Spatial filtering of the temperature image is accelerated using parallel computing, utilizing a GPU to process large images. The filter kernel size is dynamically adjusted according to image detail requirements. In some embodiments, flat-field correction and radiometric calibration are combined into a single processing step, reducing intermediate data storage and improving efficiency. Optionally, adaptive filtering technology is incorporated into the spatial filtering stage, adjusting the filtering intensity based on local noise levels.

[0035] In practical implementation, the reference image acquisition cycle of the flat-field correction algorithm is synchronized with the maintenance cycle of the infrared scanning device, for example, acquiring new dark-field and flat-field images monthly to adapt to equipment aging. The corrected images are stored using a lossy compression format, such as JPEG2000, balancing accuracy and storage cost. The blackbody radiation reference source temperature range for radiometric calibration covers the expected temperature range of the building surface, for example, from -10°C to 50°C, and the calibration curve is verified by comparison with a third-party temperature measurement device. After spatial filtering of the temperature images, edge enhancement processing is performed to highlight the contours of thermal anomaly areas; edge enhancement uses the Sobel or Laplacian operator. Standard infrared image data is timestamped and geographically labeled, aligned with the building model. In some embodiments, preprocessing operations are distributed across multiple servers to process large-scale image sequences. Optionally, the user interface provides interactive settings for preprocessing parameters, such as filter kernel size and calibration point selection.

[0036] In practical implementation, the acquisition environment for both dark-field and flat-field images in the flat-field correction algorithm is strictly controlled. Dark-field images are acquired in a dark, enclosed space, while flat-field images are captured in a uniformly heated room to avoid environmental interference. The data format of the corrected images is converted to standard temperature units for cross-system compatibility. Planck's radiation law calculation for radiometric calibration uses a numerical approximation method to improve computational speed, and the calibration results are cached in memory to reduce redundant calculations. The spatial filtering stage of the temperature images allows for various filter combinations, such as Gaussian filtering for noise reduction followed by median filtering to remove salt-and-pepper noise. Metadata for standard infrared image data includes preprocessing parameter logs, supporting auditing and reproducibility. The preprocessing module integrates error handling mechanisms, such as automatic re-acquisition when images are corrupted. The entire preprocessing workflow is designed as a pipeline, allowing for real-time or batch processing.

[0037] In practical implementation, the gain and offset coefficients of the flat-field correction algorithm are derived from the reference image through least-squares fitting, using a linear regression model. After correction, the pixel values ​​of the image are normalized to between 0 and 1 to avoid numerical overflow. The blackbody radiation reference source for radiometric calibration is traceable to international temperature standards, ensuring measurement traceability. The spatial filter kernel size for the temperature image is adaptively calculated based on the image resolution; for example, a large kernel is used for smoothing high-resolution images, while a small kernel is used for preserving details in low-resolution images. The standard infrared image data output interface supports multiple formats, such as HDF5 or NetCDF, including temperature values ​​and confidence layers. In some embodiments, the preprocessing system integrates an automated quality assessment module to detect image blur or calibration deviations. Optionally, preprocessing parameters are managed through configuration files, supporting customization for different building types.

[0038] In practical implementation, the flat-field correction algorithm takes into account the differences in infrared camera models; for example, different camera models use different correction coefficient tables. The corrected image is temporarily stored in a cache to accelerate subsequent calibration steps. The multi-point calibration process for radiometric calibration is automated, with software controlling the temperature step of the blackbody radiation reference source and simultaneously acquiring calibration images. The spatial filtering effect of the temperature image is visualized, displaying a comparison before and after filtering via a heatmap. The storage structure of standard infrared image data is hierarchically organized, separating raw data, intermediate data, and result data. The preprocessing module's performance monitoring records processing time and resource usage to optimize load balancing. The preprocessing workflow is seamlessly integrated with subsequent feature extraction modules, with data transfer achieved through shared memory or message queues.

[0039] Example 3: In specific implementation, temperature gradient features and thermal anomaly features are extracted from standard infrared image data. Temperature gradient features are obtained by analyzing the spatial temperature changes in the standard infrared image data, and a spatial temperature gradient matrix is ​​calculated. The spatial temperature gradient matrix contains horizontal and vertical gradient components. The gradient magnitude feature is calculated by taking the square root of the sum of the squares of the horizontal and vertical gradient components, and the gradient direction feature is calculated by taking the arctangent values ​​of the horizontal and vertical gradient components. Thermal anomaly features are used to identify potential thermal anomaly regions from the standard infrared image data using a threshold segmentation method. The threshold segmentation method employs Otsu's algorithm or a statistical segmentation algorithm. A temperature threshold is set to binarize the standard infrared image data, and a connected component analysis algorithm is applied to the binary image to mark continuous regions as potential thermal anomaly regions. For each potential thermal anomaly region, regional statistical features are calculated, including region area, region average temperature, region temperature standard deviation, region maximum temperature, and region minimum temperature. The temperature gradient features and thermal anomaly features are combined into a multi-dimensional vector to form an initial feature set. Each sample in the initial feature set corresponds to an analysis unit in the standard infrared image data. The analysis unit can be the entire image or a sliding window sub-image.

[0040] In practical implementation, the initial feature set is normalized by incorporating real-time environmental parameters, including ambient temperature and humidity data. An environmental compensation coefficient is calculated based on these data. This coefficient is derived from a heat conduction model that considers the influence of ambient temperature on building surface heat radiation and ambient humidity on air heat conduction. The formula for calculating the environmental compensation coefficient is as follows:

[0041] in: Indicates the environmental compensation coefficient. This represents the temperature compensation weighting coefficient. This represents the surface temperature measurement value in standard infrared image data. This represents the ambient temperature data in the real-time environmental parameters. Indicates the reference temperature value. This represents the humidity compensation weighting coefficient. This represents the ambient humidity data in the real-time environmental parameters. This represents the reference humidity value. The temperature and humidity compensation weighting coefficients were determined through experimental data, and the reference temperature and humidity values ​​were set to typical values ​​under standard atmospheric conditions. Environmental compensation coefficients were used to scale and adjust each feature in the initial feature set to achieve environmental compensation.

[0042] In specific implementations, temperature gradient feature extraction is optimized using convolution operations, with convolution kernel sizes of 3x3 or 5x5, enabling parallel computation on GPUs to improve efficiency. Regional statistical feature calculations for thermal anomalies utilize regional attribute functions from an image processing library to quickly extract statistics from multiple regions. The dimensionality of the initial feature set is optimized using feature selection methods, removing redundant features and retaining those with high information content. Environmental compensation coefficients are calculated in real-time, with values ​​dynamically updated as environmental parameters change. Zero-mean normalized statistics μ and σ are periodically recalculated to adapt to data distribution shifts. In some embodiments, feature extraction and normalization processes are pipelined to achieve real-time data stream processing. In some embodiments, the weighting coefficients α and β in the environmental compensation coefficient formula are adjusted based on building material properties; for example, different coefficient values ​​are used for concrete and metal structures. Optionally, a feature validity flag is added to the normalized feature set to indicate data quality.

[0043] In practical implementation, the calculation of the spatial temperature gradient matrix considers image boundary processing, employing a padding method to expand the image boundaries and ensure consistent gradient matrix size. The threshold segmentation method for thermal anomaly features uses an adaptive threshold algorithm, adjusting the threshold size based on local image statistics. The initial feature set is combined using a feature concatenation method, connecting temperature gradient features and thermal anomaly features dimensionally to form a high-dimensional vector. The reference temperature value for the environmental compensation coefficient is used. Set to 293.15K (20°C), with reference humidity value. The relative humidity is set to 50%, and these values ​​can be adjusted according to geographical region and season. A small epsilon value is added to the standard deviation σ of zero-mean normalization to prevent division by zero errors. The output format of the normalized feature set is standardized for compatibility with machine learning frameworks. In some embodiments, multi-scale analysis is incorporated into the feature extraction process to extract features on images at different resolutions. The entire feature engineering module is designed to be configurable, supporting flexible adjustment of feature types and parameters. The gradient direction feature of temperature gradient features is quantized into discrete directional intervals, such as eight directional intervals, reducing feature dimensionality. The area feature of thermal anomalies is normalized to a ratio relative to the image area, eliminating the influence of image size. The initial feature set is stored in an efficient binary format, reducing storage space usage. The calculation of the environmental compensation coefficient ensures dimensional consistency; both temperature and humidity terms in the formula are dimensionless. The mean μ and standard deviation σ of zero-mean normalization use sliding window statistics to adapt to the time-varying characteristics of the data. Data validation of the normalized feature set includes range checks and consistency checks. The feature processing flow is synchronized with data acquisition, achieving end-to-end automation. It is understandable that the feature quality monitoring mechanism detects abnormal feature values ​​and triggers the re-collection of data.

[0044] In practical implementation, the convolution operation of the spatial temperature gradient matrix uses a separable convolution technique, calculating the horizontal gradient first and then the vertical gradient to improve computational efficiency. Connected component analysis of thermal anomaly features employs a two-pass scanning algorithm to efficiently label regions in large images. Dimensionality reduction of the initial feature set uses principal component analysis to retain key feature components. The temperature compensation weight coefficient α and humidity compensation weight coefficient β of the environmental compensation coefficient are learned from historical data through regression analysis. Zero-mean normalized statistics calculation uses an online update algorithm to adapt to the characteristics of streaming data. Metadata of the normalized feature set records feature extraction parameters and environmental conditions. In some embodiments, a time dimension feature is incorporated into the feature extraction process to extract temperature change trends from continuous image sequences. It is understood that the performance analysis tools of the feature processing module monitor computation time and memory usage to optimize resource allocation.

[0045] See Figure 4This chart presents a comprehensive analysis of the infrared temperature distribution and thermal anomaly characteristics of a building surface. The chart clearly displays temperature variations across different areas of the building surface using color gradients, with warm colors representing high-temperature regions and cool colors representing low-temperature regions. Overlaid contour lines show the temperature gradient characteristics; these curves depict the steepness of temperature changes and effectively identify the boundaries of abnormal temperature variations. Simultaneously, blue dashed lines mark the boundaries of thermal anomaly areas detected by a threshold segmentation algorithm; these areas are typically associated with potential leakage problems. The entire chart comprehensively presents multi-layered information including raw temperature data, gradient change characteristics, and anomaly area identification, providing an intuitive visual basis for building health assessment.

[0046] Example 4: In a specific implementation, a convolutional neural network (CNN) model is constructed. The input layer of the CNN model receives a normalized feature set, and the size of the input layer matches the dimension of the normalized feature set. The input layer is designed as a multi-channel structure to accommodate different types of feature maps. The hidden layers of the CNN model include multiple convolutional layers and pooling layers. The convolutional layers contain learnable filters that perform sliding window convolution operations on the input feature maps to extract local features. Each convolutional layer is followed by an activation function to introduce a non-linear transformation. The pooling layers use max pooling or average pooling operations to reduce the spatial size of the feature maps while retaining salient features. The pooling window size is typically set to 2x2. After the stacked convolutional and pooling layers, a fully connected layer is added. The fully connected layer linearly combines the unfolded feature vectors to integrate high-level features. The number of neurons in the fully connected layer is set according to the task complexity. After the fully connected layer, an output layer is connected. The number of nodes in the output layer corresponds to the size of the thermal anomaly probability map. The output layer uses a sigmoid activation function to map the output value of each node to between 0 and 1, representing the leakage probability of each pixel. The convolutional neural network model is trained using historical leakage data as labels. This data includes real-world annotations of known leakage locations. The training process calculates predicted values ​​through forward propagation and adjusts network weights through backpropagation. The optimizer is either stochastic gradient descent or the Adam algorithm, and the loss function is binary cross-entropy. The trained convolutional neural network model is then used for forward propagation to generate a thermal anomaly probability map. This map is the same size as the input image, and each pixel value represents the probability of a leakage source existing at that location.

[0047] In the specific implementation, referring to Table 1, a density-based spatial clustering algorithm is applied to cluster high-probability points in the thermal anomaly probability map. A density-based spatial clustering algorithm such as DBSCAN is used. DBSCAN requires setting a neighborhood radius parameter and a minimum number of points parameter. The neighborhood radius parameter defines the neighborhood range of a sample point, and the minimum number of points parameter defines the minimum number of neighborhood points required for the core object. Thresholding is performed on the thermal anomaly probability map to extract high-probability points with probability values ​​higher than the set threshold. These high-probability points constitute the input dataset for cluster analysis. The DBSCAN algorithm traverses all high-probability points, identifying core points, boundary points, and noise points, merging density-connected core points and their boundary points to form candidate clusters. The centroid position and spatial range of each candidate cluster are calculated. The centroid position is obtained by calculating the average coordinates of all points within the cluster, and the spatial range is obtained by calculating the bounding box or convex hull of the cluster. Candidate clusters are selected based on size thresholds, including a minimum area threshold and a maximum area threshold. Candidate clusters that meet the size threshold conditions are identified as thermal anomaly regions. Record the boundary coordinates and probability values ​​of the thermal anomaly region. The boundary coordinates define the spatial location of the thermal anomaly region, and the probability values ​​are the average or maximum probability of all points within the region.

[0048] Table 1: Training Parameter Configuration for Convolutional Neural Network Models

[0049] In practical implementation, the convolutional neural network model uses a deep learning framework. The network architecture adopts an encoder-decoder structure. The encoder consists of convolutional and pooling layers, progressively compressing the feature map size. The decoder consists of transposed convolutional layers or upsampling layers, restoring the feature map size to the original input size. Learnable filters are initialized using Xavier or He initialization methods to avoid gradient vanishing or exploding. Loss function monitoring during training is performed using a validation set. Early stopping strategies prevent overfitting, and data augmentation techniques such as rotation and flipping increase the diversity of training samples. The generation of the thermal anomaly probability map uses a sliding window approach to handle large images, ensuring full image coverage. Parameter tuning for density-based spatial clustering algorithms is achieved through grid search or Bayesian optimization, and the clustering results are visualized for verification. The boundary coordinates of thermal anomaly regions are stored as a sequence of polygon vertices, with probability values ​​appended with confidence indices. In some embodiments, the convolutional neural network model incorporates an attention mechanism to enhance attention to important features. Optionally, morphological operations, such as opening and closing operations, are added after clustering analysis to optimize the region shape.

[0050] In practical implementation, the normalized feature set received by the input layer of the convolutional neural network model is organized in tensor form, with tensor dimensions including batch size, image height, image width, and number of channels. The convolutional layers use either uniform padding or effective padding to control the output feature map size. The pooling layer stride is set to 2, halving the feature map size. The dropout rate of the fully connected layers is set to prevent overfitting, and the sigmoid activation function of the output layer ensures that the output value is within the probability range. Training data is divided into training, validation, and test sets, typically in proportions of 70%, 15%, and 15%. The distance metric for density-based spatial clustering algorithms is Euclidean distance, and neighborhood queries are accelerated using kd-trees. The filtering conditions for hot anomaly regions are configurable to support the needs of different application scenarios. Essentially, the entire processing flow is automated, requiring no manual intervention from feature input to region output. Optionally, non-maximum suppression is added to the post-processing of the hot anomaly probability map to eliminate overlapping regions.

[0051] In practical implementation, the convolutional layers of the convolutional neural network model use grouped convolutions or depthwise separable convolutions to reduce the number of parameters and improve computational efficiency. The pooling window of the pooling layer can be adjusted to 3x3 to retain more detailed information. The activation function of the fully connected layer is ReLU or LeakyReLU to enhance the network's expressive power. The number of nodes in the output layer is set according to the resolution of the thermal anomaly probability map, such as one-quarter of the original image resolution. The learning rate scheduling during training uses cosine annealing or step descent to improve convergence efficiency. The noise point handling strategy of the density-based spatial clustering algorithm includes secondary clustering or manual review. The boundary coordinates of the thermal anomaly region are transformed to the world coordinate system for easy alignment with the building model. The probability values ​​are stored in floating-point format to preserve computational precision. The deployment of the convolutional neural network model considers embedded device optimization, such as model quantization and pruning. The parallelization of the density-based spatial clustering algorithm is implemented to process large-scale point cloud data. The results of the thermal anomaly region are exported to a standard geographic information format, supporting integration with third-party software.

[0052] See Figure 5 The figure presents the results of deep learning-based analysis of thermal anomaly probability distribution and leakage source location. A probability heatmap is used to illustrate the likelihood of a leakage source at each pixel, with areas having higher probability values ​​appearing brighter.

[0053] Example 5: In a specific implementation, multi-dimensional features are extracted from the thermal anomaly region. These features include thermal inertia and texture features. Thermal inertia is calculated by analyzing the temperature change curves of the thermal anomaly region in a continuous infrared image sequence, reflecting the thermal inertia of the material. Calculating thermal inertia involves measuring the rate of temperature change of the thermal anomaly region during heating and cooling cycles, such as the temperature response speed under varying solar radiation or external heat sources. Texture features are extracted from the infrared images of the thermal anomaly region using a gray-level co-occurrence matrix. Texture features include statistics such as contrast, energy, homogeneity, and correlation, describing the surface texture patterns of the thermal anomaly region. The multi-dimensional feature extraction process is performed for each identified thermal anomaly region, generating a feature vector for each region. This feature vector contains numerical representations of both thermal inertia and texture features. The feature extraction algorithm is implemented using an image processing library.

[0054] In practice, a pre-trained deep learning model is used to perform regression analysis on multi-dimensional features to predict the depth information of leakage sources. The pre-trained deep learning model adopts a fully connected neural network structure. The fully connected neural network is used as the regression model, with the number of input layer nodes matching the dimensions of the multi-dimensional features. The hidden layers contain multiple fully connected layers, each followed by an activation function such as ReLU. The output layer is a single node, using a linear activation function to directly output the predicted depth value of the leakage source. The fully connected neural network regression model is trained using depth-labeled training data from historical leakage detection cases. Each case contains multi-dimensional feature vectors and corresponding true depth values. The training process minimizes the mean squared error loss function, using the Adam algorithm as the optimizer with a learning rate of 0.001. The number of training epochs is determined based on the amount of data. The multi-dimensional features of the thermal anomaly region are input into the trained fully connected neural network regression model, and forward propagation calculates the predicted depth value for each thermal anomaly region. The predicted depth value is expressed in length units, such as centimeters or meters.

[0055] In practice, the predicted results are geometrically constrained and optimized using a building geometry model to correct spatial location and output the coordinates of the leakage source. The building geometry model is loaded from the Building Information Model (BIM) and contains three-dimensional structural information of the building surface, such as wall curvature, pipe layout, and structural hierarchy. The two-dimensional coordinates and predicted depth values ​​of the thermal anomaly area are mapped onto the building geometry model to generate a three-dimensional point cloud. The two-dimensional coordinates are derived from the pixel coordinate system of the infrared image and transformed to the building coordinate system using camera calibration parameters. The three-dimensional point cloud generation process assigns three-dimensional coordinates to each thermal anomaly area, where the X and Y coordinates are derived from the geometric transformation of the two-dimensional coordinates, and the Z coordinate is the predicted depth value. The generated three-dimensional point cloud is registered with the building geometry model using an iterative nearest-neighbor algorithm. This algorithm minimizes the distance error between corresponding points by iteratively calculating the rigid body transformation between the point cloud and the model surface. After optimization, the registered three-dimensional point cloud position is output as the final leakage source coordinates, represented in three-dimensional spatial coordinates, such as (X, Y, Z) triples.

[0056] In practical implementation, the thermal inertia feature is specifically calculated by measuring the temperature fluctuation amplitude and phase delay of the thermal anomaly region within a 24-hour cycle, with temperature data extracted from a continuous infrared image sequence. The gray-level co-occurrence matrix calculation for texture features employs multiple direction and distance parameters, such as 0°, 45°, 90°, and 135° directions, with distance parameters set to 1, 2, and 3 pixels. The input feature vector of the fully connected neural network regression model is standardized to ensure that each feature dimension has a mean of 0 and a variance of 1. Depth annotation of the training data is obtained through borehole detection or ultrasonic thickness gauge measurement to ensure the accuracy of the true depth values. Loading of the architectural geometry model supports multiple file formats, such as IFC or DWG, and surface geometry data is extracted through a parser. During the registration process between the 3D point cloud and the architectural geometry model, the iterative nearest-point algorithm sets a maximum number of iterations and a distance threshold to prevent infinite loops. In some embodiments, time-series features, such as the Fourier coefficients of temperature changes, are added to the multi-dimensional features. Optionally, surface normal vector constraints are added during the geometric constraint optimization stage to improve registration accuracy.

[0057] In practice, multi-dimensional feature extraction of the thermal anomaly region is performed within a specified time window, covering a typical thermal cycle. The number of hidden layers in the fully connected neural network regression model is determined based on feature complexity, typically 3 to 5 layers, with the number of neurons in each layer decreasing progressively. An early stopping strategy is used during training to prevent overfitting; training stops when the validation set loss no longer decreases. The coordinate system of the building geometry model is aligned with the world coordinate system of the infrared scanning device, achieving coordinate unification through ground control points. The 3D point cloud generation considers the camera's perspective projection model, back-projecting the 2D pixel coordinates into 3D space. The distance metric for the iterative nearest point algorithm uses Euclidean distance, and the transformation matrix includes rotation and translation parameters. Registration result evaluation is performed through reprojection error analysis to ensure location accuracy. In some embodiments, depth prediction incorporates uncertainty estimation, outputting a confidence interval. Optionally, the leakage source location coordinates are output in multiple formats, including local and global coordinate systems.

[0058] In practice, thermal inertia feature calculation uses differential temperature sequences to measure the rate of temperature change between adjacent time points. Texture feature extraction employs multi-scale analysis methods to calculate texture statistics at different image resolutions. Training data augmentation for the fully connected neural network regression model is achieved by adding noise and rotation to expand the dataset. Surface structure information of the building geometry model, including material properties and thermophysical parameters, is used to verify the rationality of depth prediction. After the 3D point cloud is generated, outlier removal is performed, eliminating obviously erroneous depth points. Initial registration using the iterative nearest point algorithm employs principal component analysis or quaternion methods to provide a good initial location. The coordinates of the registered and optimized leakage source location are automatically labeled on the building model, generating a detection report. The entire process is automated, requiring no manual intervention from feature extraction to coordinate output. Optionally, location coordinate verification is performed by combining other detection methods, such as hygrometer measurement.

[0059] In practical implementation, the multi-dimensional feature vectors of thermal anomaly regions have high dimensionality. Principal component analysis is used for dimensionality reduction, retaining 95% of the variance contribution rate. Cross-validation is used to optimize the hyperparameters of the fully connected neural network regression model to select the optimal network structure. The quality of the training data annotation is ensured through independent annotation by multiple researchers and consistency checks. The level of detail of the building geometry model is selected based on the required detection accuracy, such as LOD300 or LOD400. A multi-resolution strategy is used for registration between the 3D point cloud and the building geometry model, starting with low-precision registration and gradually increasing the accuracy. The convergence criterion for the iterative nearest point algorithm is set to a change in transformation parameters less than a threshold. The output of the leakage source location coordinates includes a timestamp and region identifier, supporting historical data traceability. It can be understood that the multi-dimensional spatial analysis module is integrated with the front-end visualization system to achieve interactive display of results.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional method for locating leakage sources based on 360-degree infrared detection in buildings, characterized in that, The method includes: The infrared scanning device collects 360-degree infrared image data of the building's exterior and simultaneously obtains real-time environmental parameters. The infrared image data is preprocessed, including flat field correction and radiometric calibration, to generate standard infrared image data; Temperature gradient features and thermal anomaly features are extracted from the standard infrared image data to form an initial feature set; The initial feature set is normalized by combining the real-time environmental parameters to obtain a normalized feature set; The normalized feature set is input into a convolutional neural network model for feature learning, and a hot anomaly probability map is output. Cluster analysis was performed on the thermal anomaly probability map to identify thermal anomaly regions; Multi-dimensional spatial analysis, including deep learning and geometric constraints, is performed based on the thermal anomaly region to determine the coordinates of the leakage source location.

2. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 1, characterized in that, The method of acquiring 360-degree infrared image data of the building's exterior using an infrared scanning device, while simultaneously obtaining real-time environmental parameters, includes: The infrared scanning device is controlled to perform a 360-degree rotating scan around the perimeter of the building at a preset scanning speed, acquiring continuous infrared image frames. During the scanning process, ambient temperature, humidity and wind speed data are collected synchronously by environmental sensors as the real-time environmental parameters. The continuous infrared image frames are timestamped to generate a time-synchronized infrared image sequence; The infrared image sequence and the real-time environmental parameters are stored in the database.

3. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 2, characterized in that, The preprocessing operation on the infrared image data includes flat-field correction and radiometric calibration to generate standard infrared image data. The method includes: A flat-field correction algorithm is applied to each image frame in the infrared image sequence to eliminate optical system errors and obtain a corrected image; The corrected image is radiometrically calibrated based on a blackbody radiation reference source, and pixel values ​​are converted into temperature values ​​to generate a temperature image. Spatial filtering is performed on the temperature image to remove high-frequency noise, thereby obtaining the standard infrared image data.

4. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 3, characterized in that, The method for extracting temperature gradient features and thermal anomaly features from the standard infrared image data to form an initial feature set includes: Calculate the spatial temperature gradient matrix of the standard infrared image data, and extract the gradient magnitude and direction features as temperature gradient features; Potential thermal anomaly regions are identified from the standard infrared image data using a threshold segmentation method, and the statistical features of these regions are calculated as thermal anomaly features. The temperature gradient features and thermal anomaly features are combined into a multi-dimensional vector to form the initial feature set.

5. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 4, characterized in that, The method of normalizing the initial feature set by combining the real-time environmental parameters to obtain a normalized feature set includes: Calculate the environmental compensation coefficient based on the ambient temperature and humidity data in the real-time environmental parameters; The environmental compensation coefficient is used to scale and adjust each feature in the initial feature set to achieve environmental compensation; The compensated features are normalized to zero-mean, so that the feature mean is zero and the variance is one, thus generating the normalized feature set.

6. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 5, characterized in that, The method of inputting the normalized feature set into a convolutional neural network model for feature learning and outputting a heat anomaly probability map includes: Construct a convolutional neural network model whose input layer receives the normalized feature set and whose hidden layer includes multiple convolutional layers and pooling layers; The process of constructing the convolutional neural network model includes: Set the input size of the input layer to match the dimension of the normalized feature set: Multiple convolutional layers and pooling layers are stacked sequentially after the input layer, wherein each convolutional layer contains multiple learnable filters for extracting local features, and each pooling layer is used to reduce the spatial size of the feature map; A fully connected layer is added after the stacked convolutional and pooling layers to integrate high-level features; An output layer is connected after the fully connected layer. The number of nodes in the output layer corresponds to the size of the thermal anomaly probability map, and the leakage probability of each pixel is output using the Sigmoid activation function. The convolutional neural network model is trained using historical leakage data as labels to learn feature mappings; The thermal anomaly probability map is generated by calculating the leakage probability of each pixel through forward propagation.

7. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 6, characterized in that, The method for performing cluster analysis on the thermal anomaly probability map to identify thermal anomaly regions includes: A density-based spatial clustering algorithm is applied to cluster high-probability points in the thermal anomaly probability map to form candidate clusters. Calculate the centroid location and spatial extent of each candidate cluster, and select clusters that meet the size threshold as thermal anomaly regions; Record the boundary coordinates and probability values ​​of the thermal anomaly region.

8. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 7, characterized in that, The method for determining the location coordinates of the leakage source by performing multi-dimensional spatial analysis based on the thermal anomaly region, including deep learning and geometric constraints, includes: Multidimensional features, including thermal inertia features and texture features, are extracted from the thermal anomaly region. A pre-trained deep learning model is used to perform regression analysis on the multi-dimensional features to predict the depth information of the leakage source. The prediction results are geometrically constrained and optimized by combining the building geometric model, the spatial position is corrected, and the coordinates of the leakage source are output.

9. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 8, characterized in that, The method of using a pre-trained deep learning model to perform regression analysis on the multi-dimensional features to predict the depth information of the leakage source includes: A fully connected neural network is constructed as a regression model, with the multi-dimensional features as input and depth values ​​as output; The regression model is trained using training data with depth annotations, using the minimum mean squared error loss function. The multidimensional features of the thermal anomaly region are input into the trained regression model to obtain the depth prediction value of each region.

10. The method for multi-dimensional location of leakage sources based on 360-degree infrared detection of buildings as described in claim 9, characterized in that, The method involves combining a building geometry model to perform geometric constraint optimization on the prediction results, correcting the spatial location, and outputting the coordinates of the leakage source location. The method includes: Load the 3D geometric model of the building and obtain the surface structure information of the building; The two-dimensional coordinates and depth prediction values ​​of the thermal anomaly region are mapped onto a three-dimensional geometric model to generate a three-dimensional point cloud; The three-dimensional point cloud is registered with the geometric model by an iterative nearest point algorithm to optimize the positional accuracy and finally output the coordinates of the leakage source.

Images