Building engineering wall flatness detection method and system based on AI visual recognition

Abstract

The application discloses a building engineering wall surface flatness detection method and system based on AI visual recognition, relates to the field of AI visual perception technology, and comprises the following steps: an AI visual driving circuit containing a plurality of double MOS switch driving circuits is built; three types of cameras, i.e., depth, industrial and infrared thermal imaging cameras, are driven to detect a building wall surface, and wall surface visual data sets and visual driving monitoring data sets are obtained; three types of wall surface graphs are constructed through driving self-adaptive noise reduction; a sliding window detection is performed on a depth perception graph and an industrial visual graph respectively to obtain two flatness detection sequences; and finally, a wall surface flatness detection report is generated after compensation according to a thermal imaging graph. The application solves the technical problem that traditional wall surface flatness detection cannot comprehensively and accurately obtain wall surface flatness related information under a complex building site environment, achieves comprehensive and accurate detection of building engineering wall surface flatness, and meets the technical effect of accurate evaluation and effective control of wall surface flatness.

Description

Technical Field

[0001] This invention relates to the field of AI visual perception technology, and in particular to a method and system for detecting the flatness of building walls based on AI visual recognition. Background Technology

[0002] Wall flatness is a core indicator for measuring building quality, directly impacting structural safety, finishing effects, and subsequent user experience. Accurate detection is crucial for project acceptance and quality control. Current wall flatness testing methods largely rely on manual measurement and simple optical instruments, playing a role in inspecting regular indoor walls. However, as engineering quality requirements increase, their limitations become apparent when applied to the complex environments of construction sites. Due to variable lighting, wide wall surfaces, and susceptibility to construction dust, traditional testing methods cannot accurately capture flatness differences in different areas, resulting in incomplete and inaccurate data. This makes it difficult to meet the needs for precise assessment and effective control of wall flatness in complex building scenarios. Summary of the Invention

[0003] This application provides a method and system for detecting the flatness of building walls based on AI visual recognition, which solves the technical problem that traditional wall flatness detection methods are unable to comprehensively and accurately obtain relevant information on wall flatness in complex building site environments.

[0004] The first aspect of this application provides a method for detecting the flatness of building walls based on AI visual recognition. The method includes: constructing an AI visual driving circuit, which includes multiple dual-MOS switch driving circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera respectively; performing visual inspection on the building wall using the AI ​​visual driving circuit to obtain a wall visual dataset and a visual driving monitoring dataset; performing adaptive noise reduction processing on the wall visual dataset based on the visual driving monitoring dataset to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map; performing sliding window geometric flatness detection on the wall depth perception map according to the wall design scheme of the building wall to obtain a geometric flatness detection sequence; performing sliding window surface flatness detection on the wall industrial visual map according to the wall design scheme to obtain a surface flatness detection sequence; and performing flatness hazard detection compensation on the geometric flatness detection sequence and the surface flatness detection sequence based on the wall thermal imaging map to generate a wall flatness detection report.

[0005] A second aspect of this application provides a building wall flatness detection system based on AI visual recognition. The system includes: a switch-driven circuit construction module for building an AI visual driving circuit, the AI ​​visual driving circuit including multiple dual-MOS switch-driven circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera respectively; a dataset acquisition module for performing visual detection on the building wall according to the AI ​​visual driving circuit to obtain a wall visual dataset and a visual driving monitoring dataset; and a wall image acquisition module for performing adaptive noise reduction processing on the wall visual dataset based on the visual driving monitoring dataset to construct a wall depth perception system. The system includes: a knowledge map, a wall surface industrial visual image, and a wall surface thermal imaging image; a geometric flatness detection sequence acquisition module, used to perform sliding window geometric flatness detection on the wall surface depth perception image according to the wall surface design scheme of the building project wall, to obtain a geometric flatness detection sequence; a surface flatness detection sequence acquisition module, used to perform sliding window surface flatness detection on the wall surface industrial visual image according to the wall surface design scheme, to obtain a surface flatness detection sequence; and a wall flatness detection report acquisition module, used to perform flatness defect detection compensation on the geometric flatness detection sequence and the surface flatness detection sequence according to the wall surface thermal imaging image, to generate a wall flatness detection report.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application acquires visual and driving monitoring data of the wall surface by building a driving circuit adapted to multiple types of vision cameras. After adaptive noise reduction processing, it constructs wall depth, industrial vision, and thermal imaging images. Combined with the wall design scheme, it detects the geometry and surface flatness in layers. Then, it uses thermal imaging to compensate for hidden dangers in the detection results, thereby comprehensively acquiring information on wall flatness and hidden dangers. This makes the wall flatness detection results of building projects more accurate and comprehensive, meeting the needs of accurate assessment and control of project quality. It achieves the technical effect of comprehensive and accurate detection of wall flatness in building projects, and meets the needs of accurate assessment and effective control of wall flatness. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart illustrating the method for detecting the flatness of building walls based on AI visual recognition provided in this application embodiment.

[0009] Figure 2This is a schematic diagram of the structure of the building wall flatness detection system based on AI visual recognition provided in the embodiments of this application.

[0010] Figure labeling: Switch drive circuit construction module 1, dataset acquisition module 2, wall image acquisition module 3, geometric flatness detection sequence acquisition module 4, surface flatness detection sequence acquisition module 5, wall flatness detection report acquisition module 6. Detailed Implementation

[0011] This application provides a method and system for detecting the flatness of building walls based on AI visual recognition, which solves the technical problem that traditional wall flatness detection methods are unable to comprehensively and accurately obtain relevant information on wall flatness in complex building site environments.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, a method for detecting the flatness of building walls based on AI visual recognition is provided, wherein the method includes: An AI vision driving circuit is constructed, which includes multiple dual MOS switch driving circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera, respectively.

[0015] In this embodiment, the dual MOS switch drive circuit consists of two MOS transistors, often complementary N-channel and P-channel transistors, or similar types used for current amplification / backup, gate drive chips, and controllers. It can achieve efficient switching and current control of loads such as motors, cameras, and power modules, and is widely used in power electronics scenarios such as motor drive, power conversion, and equipment power supply control.

[0016] Specifically, the core requirement of AI vision driving circuits is to provide independent and stable drive control for depth cameras, industrial cameras, and infrared thermal imaging cameras. Based on the design logic of multi-device driving circuits, the operating parameters of the three types of cameras are analyzed first to determine the rated power supply voltage, operating current range, and gate drive signal requirements of each camera, providing parameter basis for subsequent circuit design.

[0017] Next, referring to the mature topology of existing dual MOS switch drive circuits, suitable components are selected for the parameter requirements of each camera: combined with the functional requirement of adaptive backup switching, that is, direct replacement in case of failure, the dual MOS transistors are configured as the same channel type, both of which are N-channel or both of which are P-channel. Based on the gate drive signal amplitude, a gate drive chip that can realize signal amplification is selected. At the same time, an MCU controller with PWM signal output function is selected to ensure that the controller output signal can accurately control the MOS transistors to turn on and off through the gate drive chip.

[0018] Then, based on the requirement that the three cameras work independently and avoid mutual interference, the layout of a dual MOS switch drive circuit for each camera is determined, the connection paths between each dual MOS switch drive circuit and the power supply, external control unit and corresponding camera are clarified, and independent power supply links and signal transmission links are defined.

[0019] Next, the three dual MOS switch driver sub-circuits are integrated into the same circuit framework to build a complete AI vision driver circuit: the MCU controllers of all sub-circuits are connected to a unified signal bus to ensure that the main control module can send control commands to each sub-circuit synchronously, a common stable DC power supply is configured for all sub-circuits, and filter components are connected in series in the power supply line to reduce the interference of voltage fluctuations on the camera drive.

[0020] Finally, the completed AI vision driving circuit was tested: rated voltage was applied to the circuit, and on / off control signals were sent to each sub-circuit through the main control module. The start-up, operation, and stop states of each camera under the action of the driving signal were detected. At the same time, the operating current and voltage stability of the camera were monitored to confirm that each dual MOS switch driving sub-circuit can independently and reliably drive the corresponding camera, meeting the equipment operation requirements of AI vision inspection.

[0021] The AI ​​vision-driven circuit is used to perform visual inspection on the walls of building projects to obtain a wall visual dataset and a vision-driven monitoring dataset.

[0022] Optionally, the completed AI vision driving circuit is first activated. The MCU controller within the circuit sends a start command to each dual-MOS switch driving circuit. Upon receiving the command, the dual-MOS switch driving circuits activate their corresponding power supply circuits, providing stable operating voltage and current to the depth camera, industrial camera, and infrared thermal imaging camera, respectively, enabling the three types of cameras to enter normal operating mode. Next, based on the actual area and structural characteristics of the building's wall surface, those skilled in the art plan the camera's shooting path and area division scheme, determining the shooting coverage, shooting angle, and sampling interval for each type of camera on the wall surface. This ensures that the wall surface areas captured by the three types of cameras are complete and mutually matched, laying the foundation for subsequent integration of complete wall surface data.

[0023] Then, the depth camera is controlled to scan the wall surface according to the preset path, collect the depth information of each point on the wall surface, and generate wall surface depth data; the industrial camera is controlled to simultaneously collect images of the wall surface, capture the texture, color and minor defect information of the wall surface, and generate wall surface image data; the infrared thermal imaging camera is controlled to collect the heat distribution information of the wall surface in parallel, and generate wall surface thermal imaging data. The three types of data are integrated to form a wall surface visual dataset.

[0024] Meanwhile, during the process of the camera acquiring visual data, the current monitoring component in the AI ​​vision driving circuit collects the operating parameters of each dual MOS switch driving circuit in real time, including the circuit output current, voltage fluctuation value and MOS transistor conduction status signal. These data reflecting the operating status of the driving circuit are transmitted to the data storage module in real time to form a vision driving monitoring dataset.

[0025] Next, the collected wall visual dataset and the vision-driven monitoring dataset are correlated and matched according to the acquisition timestamp to ensure that the wall visual data at the same time point corresponds one-to-one with the monitoring data of the corresponding driving circuit. Finally, a preliminary integrity check is performed on the two types of datasets after correlation to check whether the wall visual data covers the entire preset shooting area, whether the data format of each type meets the requirements for subsequent processing, and whether there are any missing or abnormal values ​​in the vision-driven monitoring data. After confirming that the data is complete and valid, the acquisition of the two types of datasets is completed.

[0026] Based on the vision-driven monitoring dataset, the wall visual dataset is subjected to adaptive noise reduction processing to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map.

[0027] In one embodiment of this application, a wall visual dataset is first read to obtain wall depth perception data, wall industrial visual data, and wall thermal imaging data. Simultaneously, a vision-driven monitoring dataset is read to obtain corresponding depth-driven monitoring data, industrial vision-driven monitoring data, and thermal imaging-driven monitoring data. Next, each type of wall visual data is matched with its corresponding drive monitoring data. Adaptive noise reduction processing is then performed on the corresponding wall visual data based on the matched drive monitoring data. Finally, the processed data yields a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map.

[0028] Based on the wall design scheme of the building project wall, a sliding window geometric flatness detection is performed on the wall depth perception map to obtain a geometric flatness detection sequence.

[0029] In this embodiment of the application, the wall design scheme is formulated based on building design specifications and project functional requirements, and clarifies the selection of wall materials, structural layers, geometric dimensional parameters, and performance requirements such as surface flatness and appearance. It is a technical document used to guide the construction and acceptance of the wall.

[0030] Specifically, firstly, the wall depth perception map is divided into multiple window wall depth maps according to a preset sliding window. Then, these window wall depth maps are registered and projected according to the wall design scheme to establish multiple wall depth-design coupling maps. Subsequently, multi-scale perturbation transfer learning is carried out by combining the wall coupling map sample set and the geometric flatness sample set to construct a wall geometric flatness detection channel. Finally, the multiple wall depth-design coupling maps are traversed, the first wall depth-design coupling map is extracted and input into the detection channel to obtain the first wall geometric flatness, which is then added to the geometric flatness detection sequence.

[0031] Based on the wall design scheme, the surface flatness of the wall industrial visual image is detected by sliding window to obtain a surface flatness detection sequence.

[0032] Specifically, this step is similar to obtaining the geometric flatness detection sequence mentioned above, but it focuses on the characteristics of industrial visual data, performing detailed adaptation on visual features such as wall surface texture, minor bumps / dimples, and uneven color. Specifically, it works as follows: First, the wall is divided according to the sliding window. The window size and step size are set based on the texture resolution of the wall's industrial visual image, ensuring that adjacent windows overlap to avoid missed detections. Then, the wall's industrial visual image is traversed according to the set parameters, and visual data from each window area is extracted to obtain multiple window industrial visual images.

[0033] Next, the baseline visual features of the wall surface are extracted from the wall design scheme, including standard texture patterns, color uniformity parameters, and the allowable visual deviation range of surface flatness. Through texture feature point matching, such as extracting texture corner points and edge lines from the window industrial visual image, the window visual data is aligned with the design baseline features. After correcting the coordinate offset, the correspondence between actual visual features and design baseline is established, and industrial visual-design coupling diagrams corresponding to multiple windows are generated.

[0034] Subsequently, based on the industrial vision-design coupling map sample set and the surface flatness sample set, where the surface flatness sample set is labeled with the flatness deviation values ​​corresponding to the visual features in the samples, a surface flatness detection channel is constructed through multi-scale perturbation transfer learning. During perturbation injection, common visual interferences in industrial scenarios, such as changes in illumination and minor lens damage, are simulated. Each industrial vision-design coupling map is sequentially input into this detection channel, and the corresponding window's surface flatness quantification result is output. Finally, the surface flatness results of each window are added to the empty sequence one by one in traversal order, and the coordinates of the wall area corresponding to each result are associated, forming a surface flatness detection sequence covering the entire wall surface and a traceable area.

[0035] Based on the wall thermal imaging image, the geometric flatness detection sequence and the surface flatness detection sequence are used to detect and compensate for potential flatness defects, and a wall flatness detection report is generated.

[0036] Specifically, firstly, the wall thermal imaging images are divided into multiple window thermal imaging images according to a preset sliding window. These window thermal imaging images are then registered and projected according to the wall design scheme to establish multiple wall thermal imaging-design coupling diagrams. Subsequently, the AdamW optimizer is used to perform decoupling weight decay optimization training on the thermal imaging flatness defect detection event set to construct a wall flatness defect detection channel. Multiple wall thermal imaging-design coupling diagrams are input into this detection channel to obtain a flatness defect detection sequence. Finally, the geometric flatness detection sequence, surface flatness detection sequence, and flatness defect detection sequence are visualized and organized to generate a wall flatness inspection report.

[0037] Furthermore, the method provided in this application embodiment includes: The dual MOS switch driving circuit includes an MCU controller, a gate driver chip, a first MOS transistor, and a second MOS transistor. The MCU controller outputs a PWM pulse width modulation signal to the gate driver chip, which controls the on and off states of the first MOS transistor. The second MOS transistor is used for adaptive standby switching of the first MOS transistor.

[0038] In this embodiment, the MCU controller is a miniature control unit that integrates a microprocessor, memory, input / output interfaces, and peripheral control modules. It can receive external signals, execute preset programs, and output control commands. The gate driver chip is a dedicated chip that can amplify the weak signals output by the MCU and other control units into strong drive signals to precisely control the gate on / off state of power devices such as MOSFETs and IGBTs.

[0039] Optionally, firstly, based on the design specifications of the dual MOS switch drive circuit and the power supply control requirements of the camera equipment, the core components of the dual MOS switch drive circuit are selected. By analyzing the rated current, voltage range, and switching response speed requirements during camera operation, an MCU controller with PWM signal output function is selected, matched with a gate driver chip capable of signal amplification and isolation, and the first and second MOS transistors are selected according to their current carrying capacity to ensure that the parameters of each component meet the stable drive requirements of the camera.

[0040] Next, the component connection links are built according to the circuit topology: the PWM signal output pin of the MCU controller is electrically connected to the signal input terminal of the gate driver chip, so that the MCU controller can transmit control signals to the gate driver chip; the signal output terminal of the gate driver chip is connected to the gate of the first MOSFET, and the source and drain of the first MOSFET are connected to the power supply circuit and the camera power supply terminal respectively, to construct the main control circuit of the first MOSFET for power supply to the camera; then the source and drain of the second MOSFET are connected in parallel between the power supply circuit of the first MOSFET and the camera power supply terminal, and its gate is connected to the backup control terminal of the gate driver chip to form the backup switching circuit of the second MOSFET for the first MOSFET. The first MOSFET and the second MOSFET are of the same channel type, both of which are N-channel or both of which are P-channel.

[0041] Then, the startup circuit implements the conventional driving function of the first MOSFET. Through PWM signal control logic, the MCU controller outputs a PWM pulse width modulation signal with a specific duty cycle according to the preset camera startup sequence and power supply requirements. This signal is transmitted to the gate driver chip, where it undergoes internal signal amplification and level conversion to generate a control signal that can effectively drive the first MOSFET. The control signal acts on the gate of the first MOSFET, adjusting the conduction level and on / off state of the first MOSFET to achieve precise control of the camera power supply. At the same time, by gradually adjusting the duty cycle of the PWM signal, it avoids the sudden increase in current during startup from impacting the camera.

[0042] Subsequently, based on the circuit fault monitoring and backup switching mechanism, the adaptive backup function of the second MOSFET is implemented. During circuit operation, the operating status of the first MOSFET is monitored in real time through the state detection module built into the gate driver chip or an external current detection component, including whether the conduction current, voltage drop, and switching response are normal. When a fault is detected in the first MOSFET, such as conduction failure, abnormal current, or response delay, the gate driver chip immediately switches the control signal output path and sends a conduction signal to the gate of the second MOSFET; after receiving the signal, the second MOSFET quickly conducts, replacing the first MOSFET to connect to the camera power supply circuit, maintaining the continuity of camera power supply, and preventing the camera from stopping work due to the failure of the first MOSFET.

[0043] Finally, the operational status of the dual MOS switch drive circuit was tested using existing circuit function verification methods. The rated operating voltage was applied to the circuit, and the MCU controller output PWM signals with different duty cycles. The switching response speed and power supply stability of the first MOS transistor under different signals were detected. A fault state of the first MOS transistor was artificially simulated to verify the response time of the second MOS transistor during standby switching and the reliability of the power supply after switching. After multiple tests, it was confirmed that the circuit could stably implement the camera drive and standby switching functions, until it met the technical requirements of those skilled in the art for camera drive circuits.

[0044] Furthermore, the method provided in this application embodiment includes: The dual MOS switch drive circuit integrates a current monitoring component, which is used to monitor the camera's operating current in real time. When the camera's operating current is abnormal, the dual MOS switch drive circuit is triggered to perform a power-off protection operation.

[0045] Specifically, firstly, based on the rated current carrying range of the dual MOS switch drive circuit, the rated operating current of the camera, and the peak value of the starting current, a shunt or Hall current sensor is selected as the current monitoring component to ensure that the current measurement range of the component covers the current range under normal and abnormal conditions of the camera, and that the electrical parameters are compatible with the dual MOS switch drive circuit, so as to avoid monitoring deviation or circuit failure due to mismatch of component parameters.

[0046] Then, the selected current monitoring component is connected in series to the main power supply circuit of the dual MOS switch drive circuit, specifically between the power output terminal and the common input terminal of the first MOS transistor and the second MOS transistor, so that all the current when the camera is working flows through the current monitoring component; at the same time, the signal output terminal of the current monitoring component is electrically connected to the signal input terminal of the MCU controller of the dual MOS switch drive circuit, thus constructing a transmission path for current data from the monitoring component to the control unit, ensuring that the monitored current signal can be fed back to the control core in real time.

[0047] Next, based on the camera's technical specifications, determine the upper and lower limits of its normal operating current. Combining the threshold setting principles for overload and undercurrent protection in circuit protection standards, write the above current threshold parameters into the control program through the MCU controller's programming interface to establish a benchmark for judging abnormal current. The upper limit must be lower than the maximum safe carrying current of the dual MOS switch drive circuit, and the lower limit must be higher than the camera's minimum normal operating current to prevent false protection or leakage protection caused by improper threshold setting.

[0048] Subsequently, when the dual MOS switch drive circuit is activated and supplies power to the camera, the current monitoring component begins to collect current data in the main power supply circuit in real time. The collected analog current signal is converted into a digital signal through the internal signal conversion module, and then the digital current signal is continuously sent to the MCU controller through the established transmission path, so that the MCU controller can dynamically obtain the real-time changes in the camera's operating current.

[0049] Subsequently, the MCU controller continuously compares the real-time received current data with the preset normal current threshold. When the current data exceeds the set upper limit, such as when the camera short circuit causes a sudden increase in current, or falls below the set lower limit, such as when the camera power supply circuit has poor contact causing a sudden drop in current, the MCU controller determines that the camera operating current is abnormal, immediately generates a power-off protection control command, and transmits the command to the gate driver chip.

[0050] Finally, upon receiving the power-off protection control command, the gate driver chip quickly cuts off the conduction signals output to the gates of the first and second MOSFETs, causing both MOSFETs to turn off simultaneously and disconnecting the power supply circuit between the dual MOSFET switch driver circuit and the camera. Simultaneously, the MCU controller can issue an abnormal warning signal via external indicator elements such as LEDs or alarm modules to inform the operator that the circuit has implemented power-off protection, preventing damage to the camera due to continuous abnormal current or other circuit safety risks.

[0051] Furthermore, the method provided in this application embodiment includes: Read the wall visual dataset to obtain wall depth perception data, wall industrial visual data, and wall thermal imaging data; read the vision-driven monitoring dataset to obtain depth-driven monitoring data, industrial vision-driven monitoring data, and thermal imaging-driven monitoring data; perform adaptive noise reduction processing on the wall depth perception data based on the depth-driven monitoring data to obtain the wall depth perception map; perform adaptive noise reduction processing on the wall industrial visual data based on the industrial vision-driven monitoring data to obtain the wall industrial visual map; perform adaptive noise reduction processing on the wall thermal imaging data based on the thermal imaging-driven monitoring data to obtain the wall thermal imaging map.

[0052] Specifically, the data storage module's read interface is first called to extract data from the wall visual dataset according to the preset data format. The wall depth perception data collected by the depth camera, the wall industrial visual data collected by the industrial camera, and the wall thermal imaging data collected by the infrared thermal imaging camera are filtered out to ensure that the acquisition time and corresponding wall area identification of the three types of data are complete, providing basic visual information for subsequent processing.

[0053] Next, the vision-driven monitoring dataset is accessed through the same data reading interface. Based on the camera device identifier and data acquisition timestamp, the depth-driven monitoring data corresponding to the depth camera, the industrial vision-driven monitoring data corresponding to the industrial camera, and the thermal imaging-driven monitoring data corresponding to the infrared thermal imaging camera are extracted respectively. The depth-driven monitoring data includes the power supply voltage fluctuation value, frame rate parameters, and current change of the dual MOS switch drive circuit when the depth camera is working. The industrial vision-driven monitoring data includes the exposure parameters and lens drive motor status of the industrial camera. The thermal imaging-driven monitoring data includes the temperature of the thermal imaging detector and the signal gain value, forming drive status information that corresponds one-to-one with the three types of wall vision data.

[0054] Then, basic noise reduction is performed on the wall depth sensing data to obtain an initial wall depth map. Anomaly detection is then performed based on the depth-driven monitoring data to obtain depth-driven anomaly detection results. Subsequently, based on these anomaly detection results, depth feature interference recognition is performed on the building wall to generate corresponding results. Finally, combined with the depth feature interference recognition results, enhanced noise reduction processing is applied to the initial wall depth map to ultimately obtain the wall depth sensing map. This step will be explained in detail later.

[0055] Similar to the steps described above for obtaining the wall depth perception map, the process of obtaining the wall industrial visual image and the wall thermal image also follows the core logic of basic noise reduction—anomaly localization—interference recognition—precise optimization, combined with corresponding driver monitoring data to complete adaptive noise reduction, as detailed below: Obtaining the industrial visual image of the wall: First, perform basic noise reduction on the industrial visual data of the wall to obtain an initial industrial visual image. Then, based on industrial vision-driven monitoring data, such as camera exposure parameters and lens drive status, perform anomaly detection to locate the abnormal imaging time periods and parameters. Based on the anomaly results, identify characteristic interferences in the image caused by drive anomalies, such as brightness and darkness interference and speckle clutter. Finally, combine the interference identification results to perform enhanced noise reduction on the initial industrial visual image to obtain the final industrial visual image of the wall.

[0056] Obtaining a wall thermal image: First, perform basic noise reduction on the wall thermal imaging data to obtain an initial thermal image. Then, based on thermal imaging-driven monitoring data, such as infrared sensor temperature and signal gain values, perform anomaly detection to identify sensor malfunctions or parameter drift. Based on the anomaly results, identify characteristic interferences in the thermal imaging data caused by driving issues, such as temperature deviations and false hot spots. Finally, combine the interference identification results to enhance and reduce noise in the initial thermal image, generating the wall thermal image.

[0057] Finally, the generated wall depth perception map, wall industrial visual map, and wall thermal imaging map are checked for data integrity and validity. The wall area coverage of each map is checked to ensure that the key features are clear. It is confirmed that there is no loss of effective information due to cause-driven adaptive noise reduction processing, and that the three types of maps meet the technical requirements for subsequent wall flatness detection.

[0058] Furthermore, the method provided in this application embodiment includes: Basic noise reduction processing is performed on the wall depth sensing data to obtain an initial wall depth map; anomaly detection is performed on the depth-driven monitoring data to obtain a depth-driven anomaly detection result; depth feature interference recognition is performed on the building wall based on the depth-driven anomaly detection result to obtain a depth feature interference recognition result; and enhanced noise reduction processing is performed on the initial wall depth map based on the depth feature interference recognition result to obtain the wall depth sensing map.

[0059] Specifically, firstly, a Gaussian filtering algorithm is used to perform basic noise reduction on the wall depth sensing data. Based on the noise distribution characteristics of the wall depth sensing data, such as the density of random noise, a 3×3 or 5×5 Gaussian convolution kernel is selected. The kernel size is chosen to effectively filter noise without blurring the wall depth contour; a 3×3 kernel is used if there is little random noise, and a 5×5 kernel is used if there is more noise. Then, the Gaussian convolution kernel is slid across each pixel in the wall depth sensing data. By calculating the weighted average of the pixel and its neighboring pixels (the weights are determined by a Gaussian function, with the highest weight for the central pixel and decreasing towards the edges), the depth value of the original pixel is replaced. This filters out random noise caused by environmental interference (such as light fluctuations). After processing, an initial wall depth map is output.

[0060] Next, anomaly detection is performed based on the depth-driven monitoring data. First, key monitoring parameters are extracted from the data, including the output voltage and operating current of the dual-MOS switch drive circuit corresponding to the depth camera, and the camera's frame rate. The normal ranges for these parameters need to be determined with reference to the depth camera's technical specifications and the design standards of the dual-MOS circuit. Then, each time point of the depth-driven monitoring data is iterated, and the actual parameter values ​​at each point are compared with the preset normal ranges. If a parameter value at a certain point exceeds the normal range, that point is marked as an anomaly, and the time of the anomaly, the type of the anomaly parameter, and the parameter deviation value are recorded. Finally, the results are compiled to form the depth-driven anomaly detection results.

[0061] Then, based on the depth-driven anomaly detection results, depth feature interference recognition is performed on the building wall. First, the abnormal time period is extracted from the depth-driven anomaly detection results, locating the segment in the wall depth sensing data corresponding to this abnormal time period, for example, from second 10 to second 15. Next, the depth features of this data segment are extracted, including the wall's edge lines and the contour features of local protrusions or depressions. Simultaneously, the depth features of normal time periods before and after the abnormal time period are extracted as reference samples, for example, from second 5 to second 9 and from second 16 to second 20. By calculating the similarity between the depth features of the abnormal time period and the features of the reference samples, such as edge overlap and contour shape matching, if the similarity is lower than a preset threshold, such as 80%, it is determined that the depth features of that time period have interference, such as edge breakage or contour distortion. The specific coordinates of the interference area and the type of interference are then marked, such as depth value jumps caused by voltage anomalies, forming the depth feature interference recognition result.

[0062] Finally, the initial wall depth map undergoes enhancement and noise reduction processing using a bilateral filtering algorithm that balances noise reduction and feature preservation. Specifically: First, based on the depth feature interference recognition results, the interference regions requiring focused processing are determined. Bilateral filtering parameters are set for these regions: the spatial domain standard deviation is set to 1.5 to control the filtering range, and the gray-level domain standard deviation is set to 20 to control the gray-level similarity weight. Non-interference regions retain the basic Gaussian filtering parameters for noise reduction to reduce computation. Then, the bilateral filtering kernel slides across the initial wall depth map. For each pixel in the interference region, the spatial distance and gray-level difference of its neighboring pixels are comprehensively considered, with closer distances and similar gray levels receiving higher weights. A weighted average depth value is calculated to replace the original pixel value. This smooths interference noise while preserving detailed information about wall depth features such as the actual height of protrusions and the depth of indentations. After processing, a wall depth perception map is output.

[0063] Furthermore, the method provided in this application embodiment includes: The wall depth perception map is divided into multiple window wall depth maps according to a preset sliding window; the multiple window wall depth maps are registered and projected according to the wall design scheme to establish multiple wall depth-design coupling maps; multi-scale perturbation transfer learning is performed based on the wall coupling map sample set and the geometric flatness sample set to construct a wall geometric flatness detection channel; the multiple wall depth-design coupling maps are traversed to extract the first wall depth-design coupling map; the first wall depth-design coupling map is input into the wall geometric flatness detection channel to obtain the first wall geometric flatness, and the first wall geometric flatness is added to the geometric flatness detection sequence.

[0064] In one embodiment, firstly, the size parameters of the wall depth sensing map and the accuracy requirements for geometric flatness detection are determined. The size and step size of the sliding window are set. The sliding window size is set to 256×256 pixels, and the step size is set to 128 pixels to ensure a 50% overlap between adjacent windows, avoiding missed detection of flatness deviations in local areas of the wall. Then, starting from the upper left corner of the wall depth sensing map, the sliding window is controlled to move sequentially in the horizontal and vertical directions according to the set step size. Each time it moves, the depth data of the area covered by the current window is captured, forming an independent window wall depth map, until the window has traversed the entire wall depth sensing map, ultimately obtaining multiple continuous window wall depth maps that cover the entire wall.

[0065] Next, the baseline geometric parameters of the wall are extracted from the wall design scheme, including the spatial coordinates of the design plane, the theoretical depth values ​​of each area of ​​the wall, and the allowable deviation range of flatness. These parameters are then imported into the coordinate transformation module to generate a standard design coordinate system. A registration method based on feature point matching is used to extract corner points, edges, and other feature points from each window wall depth map. Simultaneously, theoretical feature points of the corresponding wall area are extracted from the standard design coordinate system. By calculating the Euclidean distance between the feature points, the coordinate offset between the window wall depth map and the design scheme is determined. Based on the offset, the window wall depth map is coordinate corrected to align the window depth data with the design parameters in the same coordinate system. Then, the corrected window depth data is projected and fused with the theoretical design data of the corresponding area, establishing a correspondence between actual depth and theoretical depth at each pixel point, ultimately generating multiple wall depth-design coupled maps.

[0066] Next, a detection sample space is constructed by aligning the wall coupling diagram with the geometric flatness sample set. After training the initial model, multi-scale perturbations are injected to obtain the perturbation space, and an augmentation model is trained. Iterative transfer learning is carried out with the initial model as the student layer and the augmentation model as the teacher layer. When the iterative transfer loss coefficient is less than the threshold, a wall geometric flatness detection channel is generated. This step will be explained in detail in the following content.

[0067] Then, multiple wall depth-design coupling diagrams are systematically labeled. Based on the generation time of the coupling diagram or the spatial coordinates of the corresponding wall area, a unique index number is assigned to each coupling diagram, establishing a mapping relationship from the index to the coupling diagram data. This ensures accurate location of each coupling diagram during traversal. After starting the traversal program, the program retrieves coupling diagram data starting from the initial value according to the index number, prioritizing the extraction of the coupling diagram with index number 1 as the first wall depth-design coupling diagram, completing the first round of coupling diagram extraction.

[0068] Next, the extracted first wall depth-design coupling diagram undergoes input adaptation processing. Based on the input specifications of the wall geometric flatness detection channel, size normalization is performed, adjusting the coupling diagram resolution to the preset standard size of the detection channel. Simultaneously, the coupling format of the depth data and design data in the diagram is verified to ensure that the data format matches the input interface of the detection channel. After adaptation, the first wall depth-design coupling diagram is input to the constructed wall geometric flatness detection channel via the data transmission interface.

[0069] Subsequently, the wall geometric flatness detection channel initiates the detection process, using trained model parameters to extract features from the input coupled graph. It captures the difference between the actual depth and the designed depth of the wall in the graph through convolution operations, and then performs quantitative analysis of the difference features through a fully connected layer. Combined with the geometric flatness evaluation algorithm, it calculates the corresponding flatness quantification value, which is the first wall geometric flatness. The detection channel feeds back the result to the data processing module through the output interface.

[0070] Finally, the geometric flatness detection sequence is initialized as a dynamic array structure. The first wall surface geometric flatness result is written into this sequence using tail insertion. Simultaneously, the sequence is associated with the corresponding first wall surface depth-design coupling diagram index number and wall area information to ensure a one-to-one correspondence between the detection result and the wall location. After addition, the traversal program automatically jumps to the next index number, repeating the extraction, detection, and addition steps until all wall surface depth-design coupling diagrams have been processed, ultimately forming a complete geometric flatness detection sequence.

[0071] Furthermore, the method provided in this application embodiment includes: Align the wall surface coupling map sample set and the geometric flatness sample set to obtain a geometric flatness detection sample space; train an initial geometric flatness detection model based on the geometric flatness detection sample space; inject multi-scale perturbations into the geometric flatness detection sample space to obtain multiple geometric flatness detection perturbation spaces corresponding to multiple perturbation scales; train multiple geometric flatness detection enhancement models based on the multiple geometric flatness detection perturbation spaces; use the initial geometric flatness detection model as the flatness detection student layer and the multiple geometric flatness detection enhancement models as the flatness detection teacher layer; perform iterative transfer learning on the flatness detection student layer based on the flatness detection teacher layer to obtain iterative transfer loss coefficients; if the iterative transfer loss coefficients are less than the iterative transfer loss threshold, generate the wall surface geometric flatness detection channel.

[0072] Optionally, the wall coupling image sample set and the geometric flatness sample set are aligned first. The wall coupling image sample set contains wall depth-design coupling images for different wall scenarios, while the geometric flatness sample set corresponds to the actual geometric flatness quantification value of each coupling image, such as the flatness error in millimeters. A bidirectional mapping is established by using the unique identifier of the wall area in the sample set and the acquisition timestamp, accurately associating each wall depth-design coupling image with its corresponding geometric flatness quantification value. Subsequently, the 3σ principle is used to remove outlier samples. The mean and standard deviation of the geometric flatness quantification values ​​are calculated, and samples exceeding the mean ± 3 times the standard deviation are judged as outliers and removed. Finally, a geometric flatness detection sample space with a one-to-one correspondence between coupling images and flatness values ​​is formed, ensuring the validity of the sample data.

[0073] Secondly, an initial model for geometric flatness detection was trained based on the geometric flatness detection sample space. A convolutional neural network (CNN) was selected as the basic model architecture, such as a simplified version of LeNet-5, containing two convolutional layers, two pooling layers, and two fully connected layers. The geometric flatness detection sample space was divided into a training set and a validation set in a 7:3 ratio. The wall coupling map was used as input, and the geometric flatness quantification value was used as output. The batch size was set to 16, and the initial learning rate was 0.001. The Adam optimizer was used to minimize the mean squared error (MSE) loss function. The model weights were updated iteratively through backpropagation. The model accuracy was evaluated using the validation set every 10 training epochs. Training was stopped when the loss on the validation set did not decrease significantly for three consecutive epochs, thus obtaining an initial model for geometric flatness detection that could initially output flatness results.

[0074] Next, multi-scale perturbation injection is performed on the geometric flatness detection sample space to obtain multiple perturbation spaces. Specifically, three perturbation scales are set: small-scale perturbation involves scaling the coupled image to 0.9 times its original size and restoring it, then adding Gaussian noise with a standard deviation of 0.01; medium-scale perturbation involves rotating the coupled image ±5° and cropping it, then adjusting the pixel brightness ±10%; large-scale perturbation involves scaling the coupled image to 0.8 times its original size and restoring it, then adding salt-and-pepper noise with a noise density of 0.02. Each coupled image in the detection sample space is subjected to the three perturbations, generating new sample sets corresponding to the three perturbation scales. Each sample set is still associated with the original geometric flatness quantification value, ultimately resulting in three independent geometric flatness detection perturbation spaces.

[0075] Subsequently, the geometric flatness detection enhancement model was trained using the same CNN architecture and training parameters as the initial geometric flatness detection model to ensure fairness in the model performance comparison. The three perturbation spaces were used as independent training data, and the training process for the initial model was followed sequentially, including a 7:3 dataset partitioning, Adam optimizer, and MSE loss function. Each perturbation space corresponded to one enhancement model: enhancement model A for small-scale perturbation spaces, enhancement model B for medium-scale perturbation spaces, and enhancement model C for large-scale perturbation spaces. This resulted in three geometric flatness detection enhancement models adaptable to different perturbation scenarios.

[0076] Next, the hierarchical relationship of the models was clarified. The pre-trained initial geometric flatness detection model was directly set as the flatness detection student layer, and the three geometric flatness detection enhancement models were combined as the flatness detection teacher layer. The core role of the student layer is to learn the robust detection capabilities of the teacher layer, while the teacher layer provides a more reliable basis for flatness judgment by integrating the outputs of multiple enhancement models. This can be achieved through model function localization.

[0077] Next, a knowledge distillation approach is adopted for iterative transfer learning from the teacher layer to the student layer. During transfer learning, the student layer receives the same coupled graph samples as the teacher layer, and outputs student predictions and predictions from the three teacher models. Two losses are calculated: the basic loss (MSE) between the student's predicted value and the true smoothness value, and the distillation loss (average of the three teacher model outputs) between the student's predicted value and the average predicted value from the teacher layer, also using MSE. The total loss is a weighted sum of the basic loss and the distillation loss, with each weight set to 0.5. Using the total loss as the objective, the student layer weights are iteratively updated using the Adam optimizer. The total loss is calculated once per iteration as the iterative transfer loss coefficient until the loss coefficient reaches a stable fluctuation state.

[0078] Finally, the relationship between the iterative transfer loss coefficient and the preset threshold is determined. The iterative transfer loss threshold is determined through previous experiments, and the minimum stable loss coefficient obtained from multiple model training iterations is statistically analyzed. 1.2 times this minimum stable loss coefficient is taken as the iterative transfer loss threshold. If the current iterative transfer loss coefficient is less than this threshold, it indicates that the student layer has fully learned the robustness of the teacher layer, and transfer learning is stopped. The current student layer model is then encapsulated as a wall geometric flatness detection channel. If the threshold is not met, iteration continues until the condition is satisfied.

[0079] Furthermore, the method provided in this application embodiment includes: The wall thermal imaging image is divided into multiple window thermal imaging images according to a preset sliding window. These multiple window thermal imaging images are then registered and projected according to the wall design scheme to establish multiple wall thermal imaging-design coupling images. The AdamW optimizer is used to decouple and adjust the weight decay of the thermal imaging flatness hazard detection event set to construct a wall flatness hazard detection channel. The multiple wall thermal imaging-design coupling images are input into the wall flatness hazard detection channel to obtain a flatness hazard detection sequence. The geometric flatness detection sequence, the surface flatness detection sequence, and the flatness hazard detection sequence are then visualized and organized to obtain the wall flatness detection report.

[0080] In one embodiment, the resolution parameters of the wall thermal imaging image and the accuracy requirements for detecting flatness hazards are first defined. The size and step size of the sliding window are then set. To accurately capture local thermal distribution anomalies, the window size is set to 256×256 pixels, and the step size is set to 128 pixels, ensuring that adjacent windows have a 50% overlap area to avoid missing hazard areas due to excessive window spacing. Using the upper left corner of the thermal imaging image as the origin, the window is controlled to slide sequentially along the horizontal and vertical directions. Each slide captures the thermal imaging data covered by the current window, generating an independent window wall thermal imaging image, until the window has traversed the entire thermal imaging image, obtaining multiple continuous window wall thermal imaging images that cover the entire wall surface.

[0081] Next, spatial coordinate parameters and theoretical heat distribution standard values ​​for each area of ​​the wall are extracted from the wall design scheme. These theoretical heat distribution standard values ​​need to be determined in conjunction with the thermal conductivity of the wall building materials and construction process requirements; for example, the normal heat distribution range for concrete walls is 20℃-25℃. A registration method based on feature point matching is used to extract feature points such as abrupt changes in thermal gradients and hotspot edges from the thermal imaging of each window wall. Simultaneously, theoretical feature points for the corresponding wall areas are extracted from the spatial coordinate system of the wall design scheme. By calculating the coordinate deviation between the actual and theoretical feature points, the window wall thermal imaging is coordinate-corrected to align the thermal imaging data with the design parameters in the same spatial coordinate system. Subsequently, the corrected actual heat distribution data and the corresponding theoretical heat distribution standard values ​​are projected and fused, establishing a correspondence between actual and theoretical heat values ​​at each pixel, thus creating multiple wall thermal imaging-design coupled maps.

[0082] Subsequently, based on the thermal imaging flatness hazard detection event set, the AdamW optimizer was used to carry out decoupling weight decay optimization training to construct a wall flatness hazard detection channel. The specific training process will be detailed in subsequent steps.

[0083] After the channel construction is completed, multiple wall thermal imaging-design coupling images are preprocessed. Following the input specifications of the wall flatness hazard detection channel, the resolution of the coupling images is uniformly adjusted to the channel's preset standard size. Simultaneously, the format and range of the calorific value data are verified to ensure compatibility with the channel's input interface. The preprocessed coupling images are then input into the wall flatness hazard detection channel one by one, following the window traversal order. This channel analyzes the deviation between the actual and theoretical calorific values ​​in the coupling images to identify flatness hazards reflected by abnormal heat distribution, such as lower temperatures due to thermal insulation effects in hollow areas of the wall or sudden temperature changes due to differences in heat conduction in crack areas. The channel outputs the hazard judgment result for each window, including no hazard, minor hazard, serious hazard, and quantified deviation value. These results are then organized according to the input order to obtain a flatness hazard detection sequence, where each wall thermal imaging-design coupling image corresponds to one flatness hazard detection result.

[0084] Finally, existing data visualization tools were used to integrate the geometric flatness inspection sequence, surface flatness inspection sequence, and flatness defect detection sequence. Using the actual spatial coordinates of the wall as a reference, the inspection results corresponding to the same wall area in the three sequences were correlated. Different colors were used to label different types of inspection data: geometric flatness deviations were represented by a blue gradient, surface flatness defects were marked in green, and defects identified by thermal imaging were highlighted in red. Simultaneously, a data statistics table was generated, clearly indicating the location coordinates, deviation values, defect levels, and corresponding inspection criteria for each abnormal area. The labeled wall visualization map was integrated with the statistics table, supplemented with basic information such as inspection time, inspection equipment model, and inspection personnel, to form a complete, intuitive, and clear wall flatness inspection report.

[0085] Furthermore, the method provided in this application embodiment includes: Based on the thermal imaging flatness hazard detection event set, a flatness hazard detection path is traced to establish a flatness hazard detection path space; the initialization configuration sequence of the AdamW optimizer is obtained; based on the initialization configuration sequence, a convolutional network is trained on the flatness hazard detection path space using the AdamW optimizer to generate an initial flatness hazard detection model; the initialization configuration sequence is decoupled and weight decay is optimized based on the hazard detection loss characteristics of the initial flatness hazard detection model to obtain an optimized configuration sequence; the initial flatness hazard detection model is iteratively optimized and trained using the optimized configuration sequence and the AdamW optimizer to generate the wall flatness hazard detection channel.

[0086] In one embodiment, firstly, the thermal imaging flatness hazard detection event set contains a large number of labeled samples. Each sample records thermal imaging data of the wall surface, such as low-temperature zones in hollow areas and abnormal temperature gradients in cracked areas, corresponding hazard types including hollowness, cracks, and protrusions, as well as severity labels. Key thermal features are extracted from each event sample. Temperature distribution ranges are obtained through grayscale histogram analysis, and contour features of thermal anomaly areas are extracted using edge detection algorithms such as the Sobel operator. These thermal features are then mapped to the hazard type and severity labeled in the sample, forming a detection path of thermal features → hazard determination → severity assessment. Subsequently, all detection paths are classified according to hazard type, for example, low-temperature zone + irregular contour → hollowness hazard → moderate, linear temperature gradient abrupt change → crack hazard → mild. Finally, all classified paths are integrated to construct a flatness hazard detection path space covering multiple hazard scenarios.

[0087] Secondly, obtain the initial configuration sequence of the AdamW optimizer. The core configuration parameters of the AdamW optimizer include the learning rate, weight decay coefficient, and first-moment estimation exponent. Second-order moment estimation index Referring to the general initial parameter range for training convolutional networks: the learning rate is set to 0.001 to balance convergence speed and stability, and the weight decay coefficient is set to 0.01 to initially suppress overfitting. Set to 0.9 Set the value to 0.999 to ensure smooth gradient estimation. Simultaneously, set the batch size to 16 to adapt to the performance of ordinary computing devices and the maximum training epochs to 20 to avoid overtraining. These parameters are then adjusted according to the learning rate minus the weight decay coefficient. - Arrange the batch size and maximum round number in order to form the initialization configuration sequence of the AdamW optimizer.

[0088] Next, a lightweight CNN with a simple structure was selected as the basic network architecture, containing two convolutional layers, two max-pooling layers, and two fully connected layers. First, the samples in the flatness hazard detection path space were divided into training and validation sets in a 7:3 ratio. The training set was used for model parameter learning, and the validation set was used for accuracy evaluation. The training set samples were input into the CNN, and the AdamW optimizer was started with the parameters in the initialization configuration sequence. The optimizer updated the network weights according to the learning rate and applied a weight decay coefficient to the weights of the fully connected layers to prevent overfitting. The deviation between the model's predicted results (hazard type and severity) and the true labels of the samples was calculated using the cross-entropy loss function. The model accuracy was evaluated using the validation set after each training epoch. Training was stopped when the validation set loss showed no decrease or fluctuation was less than 0.005 for three consecutive epochs. The network parameters and structure at this point were saved, generating the initial flatness hazard detection model.

[0089] Then, analyze the loss curve of the initial model for detecting flatness hazards: if the training set loss continues to decrease but the validation set loss first decreases and then increases, indicating overfitting, it means the weight decay is insufficient, and the learning rate needs to be maintained. , Without changing the model's performance, the weight decay coefficient is increased from 0.01 to decouple weight decay from gradient update, thus strengthening weight constraints independently. If the model converges slowly, for example, the loss is still higher than 0.1 after 10 training epochs, then while maintaining the weight decay coefficient, , Without changing the initialization parameters, the learning rate is finely adjusted downwards from 0.001 to slow down the learning pace and stabilize convergence. If the model has low accuracy for detecting certain types of defects, such as hollow areas (e.g., below 85%), a local learning rate boost for the hollow area feature extraction layer is added to the initialization configuration sequence, and the learning rate of that layer is finely adjusted upwards, while other parameters remain unchanged. The initialization configuration sequence is adjusted according to these rules to ultimately form an optimized configuration sequence that adapts to the initial model's loss characteristics.

[0090] Finally, using the optimized configuration sequence as parameters, the AdamW optimizer is restarted, and the training and validation sets of the flatness hazard detection path space are input back into the initial model. During iterative training, the hazard detection loss coefficient of the model, i.e., the cross-entropy loss value, is calculated every 5 rounds. If the loss coefficient is less than the preset threshold of 0.01, the iteration stops; if it does not meet the threshold, training continues according to the optimized configuration until the loss coefficient meets the requirements. After training is completed, the final model structure, parameters, and input / output interfaces are encapsulated to form a wall flatness hazard detection channel that can directly receive wall thermal imaging-design coupling diagrams and output flatness hazard detection results.

[0091] In summary, the AI-based visual recognition-based wall flatness detection method for building engineering provided in this application has the following technical effects: This application acquires a visual and driven monitoring dataset of walls using an AI vision-driven circuit, and obtains three types of wall images through adaptive noise reduction processing. A detection channel is constructed by combining sliding window detection and transfer learning, and a detection report is generated by compensating for geometric and surface flatness detection sequences with thermal imaging. This achieves comprehensive and accurate detection of wall flatness in building engineering, meeting the technical requirements for precise assessment and effective control of wall flatness.

[0092] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a building engineering wall flatness detection system based on AI visual recognition, the system comprising: Switch driving circuit construction module 1 is used to build an AI vision driving circuit, which includes multiple dual MOS switch driving circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera respectively.

[0093] Data set acquisition module 2 is used to perform visual inspection on the wall surface of the building project according to the AI ​​vision driving circuit, and obtain the wall visual dataset and the vision driving monitoring dataset.

[0094] The wall image acquisition module 3 is used to perform adaptive noise reduction processing on the wall visual dataset based on the visual driven monitoring dataset to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map.

[0095] The geometric flatness detection sequence acquisition module 4 is used to perform sliding window geometric flatness detection on the wall depth perception map according to the wall design scheme of the building project wall, and obtain a geometric flatness detection sequence.

[0096] The surface flatness detection sequence acquisition module 5 is used to perform sliding window surface flatness detection on the industrial visual image of the wall according to the wall design scheme, and obtain a surface flatness detection sequence.

[0097] The wall flatness test report acquisition module 6 is used to perform flatness defect detection compensation on the geometric flatness test sequence and the surface flatness test sequence based on the wall thermal imaging image, and generate a wall flatness test report.

[0098] Furthermore, the geometric flatness detection sequence acquisition module 4 is used to perform the following steps: The wall depth perception map is divided into multiple window wall depth maps according to a preset sliding window; the multiple window wall depth maps are registered and projected according to the wall design scheme to establish multiple wall depth-design coupling maps; multi-scale perturbation transfer learning is performed based on the wall coupling map sample set and the geometric flatness sample set to construct a wall geometric flatness detection channel; the multiple wall depth-design coupling maps are traversed to extract the first wall depth-design coupling map; the first wall depth-design coupling map is input into the wall geometric flatness detection channel to obtain the first wall geometric flatness, and the first wall geometric flatness is added to the geometric flatness detection sequence.

[0099] Furthermore, the geometric flatness detection sequence acquisition module 4 is used to perform the following steps: Align the wall surface coupling map sample set and the geometric flatness sample set to obtain a geometric flatness detection sample space; train an initial geometric flatness detection model based on the geometric flatness detection sample space; inject multi-scale perturbations into the geometric flatness detection sample space to obtain multiple geometric flatness detection perturbation spaces corresponding to multiple perturbation scales; train multiple geometric flatness detection enhancement models based on the multiple geometric flatness detection perturbation spaces; use the initial geometric flatness detection model as the flatness detection student layer and the multiple geometric flatness detection enhancement models as the flatness detection teacher layer; perform iterative transfer learning on the flatness detection student layer based on the flatness detection teacher layer to obtain iterative transfer loss coefficients; if the iterative transfer loss coefficients are less than the iterative transfer loss threshold, generate the wall surface geometric flatness detection channel.

[0100] Furthermore, the wall flatness test report acquisition module 6 is used to perform the following steps: The wall thermal imaging image is divided into multiple window thermal imaging images according to a preset sliding window. These multiple window thermal imaging images are then registered and projected according to the wall design scheme to establish multiple wall thermal imaging-design coupling images. The AdamW optimizer is used to decouple and adjust the weight decay of the thermal imaging flatness hazard detection event set to construct a wall flatness hazard detection channel. The multiple wall thermal imaging-design coupling images are input into the wall flatness hazard detection channel to obtain a flatness hazard detection sequence. The geometric flatness detection sequence, the surface flatness detection sequence, and the flatness hazard detection sequence are then visualized and organized to obtain the wall flatness detection report.

[0101] Furthermore, the wall flatness test report acquisition module 6 is used to perform the following steps: Based on the thermal imaging flatness hazard detection event set, a flatness hazard detection path is traced to establish a flatness hazard detection path space; the initialization configuration sequence of the AdamW optimizer is obtained; based on the initialization configuration sequence, a convolutional network is trained on the flatness hazard detection path space using the AdamW optimizer to generate an initial flatness hazard detection model; the initialization configuration sequence is decoupled and weight decay is optimized based on the hazard detection loss characteristics of the initial flatness hazard detection model to obtain an optimized configuration sequence; the initial flatness hazard detection model is iteratively optimized and trained using the optimized configuration sequence and the AdamW optimizer to generate the wall flatness hazard detection channel.

[0102] Furthermore, the wall image acquisition module 3 is used to perform the following steps: Read the wall visual dataset to obtain wall depth perception data, wall industrial visual data, and wall thermal imaging data; read the vision-driven monitoring dataset to obtain depth-driven monitoring data, industrial vision-driven monitoring data, and thermal imaging-driven monitoring data; perform adaptive noise reduction processing on the wall depth perception data based on the depth-driven monitoring data to obtain the wall depth perception map; perform adaptive noise reduction processing on the wall industrial visual data based on the industrial vision-driven monitoring data to obtain the wall industrial visual map; perform adaptive noise reduction processing on the wall thermal imaging data based on the thermal imaging-driven monitoring data to obtain the wall thermal imaging map.

[0103] Furthermore, the wall image acquisition module 3 is used to perform the following steps: Basic noise reduction processing is performed on the wall depth sensing data to obtain an initial wall depth map; anomaly detection is performed on the depth-driven monitoring data to obtain a depth-driven anomaly detection result; depth feature interference recognition is performed on the building wall based on the depth-driven anomaly detection result to obtain a depth feature interference recognition result; and enhanced noise reduction processing is performed on the initial wall depth map based on the depth feature interference recognition result to obtain the wall depth sensing map.

[0104] Furthermore, the switch drive circuit construction module 1 is used to perform the following steps: The dual MOS switch driving circuit includes an MCU controller, a gate driver chip, a first MOS transistor, and a second MOS transistor. The MCU controller outputs a PWM pulse width modulation signal to the gate driver chip, which controls the on and off states of the first MOS transistor. The second MOS transistor is used for adaptive standby switching of the first MOS transistor.

[0105] Furthermore, the switch drive circuit construction module 1 is used to perform the following steps: The dual MOS switch drive circuit integrates a current monitoring component, which is used to monitor the camera's operating current in real time. When the camera's operating current is abnormal, the dual MOS switch drive circuit is triggered to perform a power-off protection operation.

[0106] The AI-based visual recognition wall flatness detection system for building engineering provided in this embodiment of the invention can execute the AI-based visual recognition wall flatness detection method for building engineering provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0107] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0108] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for detecting the flatness of building walls based on AI visual recognition, characterized in that, include: An AI vision driving circuit is constructed, which includes multiple dual MOS switch driving circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera respectively. The AI ​​vision driving circuit is used to perform visual inspection on the wall surface of the building project to obtain a wall visual dataset and a vision driving monitoring dataset. Based on the vision-driven monitoring dataset, the wall visual dataset is subjected to adaptive noise reduction processing to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map. Based on the wall design scheme of the building project wall, the geometric flatness of the wall depth perception map is detected by sliding window to obtain a geometric flatness detection sequence; The surface flatness of the wall is detected by sliding window based on the wall design scheme, and a surface flatness detection sequence is obtained. Based on the wall thermal imaging image, the geometric flatness detection sequence and the surface flatness detection sequence are used to perform flatness defect detection compensation, and a wall flatness detection report is generated. Based on the wall thermal imaging image, the geometric flatness detection sequence and the surface flatness detection sequence are used to detect and compensate for potential flatness defects, generating a wall flatness detection report, including: The wall thermal imaging image is divided according to a preset sliding window to obtain multiple window wall thermal imaging images; Based on the wall design scheme, the thermal images of the multiple window walls are registered and projected to establish multiple wall thermal imaging-design coupled images; Based on the AdamW optimizer, a wall flatness hazard detection channel is constructed by decoupling weight decay optimization training on the thermal imaging flatness hazard detection event set. The multiple wall thermal imaging-design coupled images are input into the wall flatness defect detection channel to obtain a flatness defect detection sequence; The geometric flatness detection sequence, the surface flatness detection sequence, and the flatness defect detection sequence are visualized and organized to obtain the wall flatness detection report; Based on the vision-driven monitoring dataset, the wall visual dataset is subjected to driven adaptive noise reduction processing to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map, including: Read the wall visual dataset to obtain wall depth perception data, wall industrial visual data, and wall thermal imaging data; Read the vision-driven monitoring dataset to obtain depth-driven monitoring data, industrial vision-driven monitoring data, and thermal imaging-driven monitoring data; Adaptive noise reduction processing is performed on the wall depth perception data based on the depth-driven monitoring data to obtain the wall depth perception map; Based on the industrial vision-driven monitoring data, adaptive noise reduction processing is performed on the wall surface industrial vision data to obtain the wall surface industrial vision image. The wall thermal imaging data is adaptively denoised based on the thermal imaging-driven monitoring data to obtain the wall thermal imaging image. Adaptive noise reduction processing is performed on the wall depth sensing data based on the depth-driven monitoring data to obtain the wall depth sensing map, including: Based on the wall depth sensing data, basic noise reduction processing is performed to obtain an initial wall depth map; Anomaly detection is performed based on the depth-driven monitoring data to obtain depth-driven anomaly detection results; Based on the depth-driven anomaly detection results, depth feature interference recognition is performed on the wall surface of the building project to obtain the depth feature interference recognition results. The initial wall depth map is enhanced and denoised based on the depth feature interference recognition result to obtain the wall depth perception map; To obtain an industrial visual image of a wall, the process involves performing basic noise reduction on the wall's industrial visual data to obtain an initial industrial visual image. Based on industrial vision-driven monitoring data, camera exposure parameters, and lens drive status, anomaly detection is performed to locate the abnormal imaging time period and parameters. Based on the anomaly results, the brightness and darkness interference and speckle clutter feature interference caused by drive anomalies in the image are identified. Combined with the interference identification results, the initial industrial visual image is enhanced and denoised to obtain the final industrial visual image of the wall. To obtain a wall thermal image, the process involves performing basic noise reduction on the wall thermal imaging data to obtain an initial thermal image. Based on the thermal imaging-driven monitoring data, anomaly detection is performed on the infrared sensor temperature and signal gain values ​​to identify sensor malfunctions or parameter drift. Based on the anomaly results, temperature deviations and pseudo-hotspot interference caused by driving problems in the thermal imaging data are identified. Finally, the initial thermal image is enhanced and noise-reduced by combining the interference identification results to generate a wall thermal image.

2. The method for detecting the flatness of building walls based on AI visual recognition as described in claim 1, characterized in that, Based on the wall design scheme of the building project wall, a sliding window geometric flatness detection is performed on the wall depth perception map to obtain a geometric flatness detection sequence, including: The wall depth perception map is divided according to a preset sliding window to obtain multiple window wall depth maps; Based on the wall design scheme, the wall depth maps of the multiple windows are registered and projected to establish multiple wall depth-design coupling maps; Multi-scale perturbation transfer learning is performed based on the wall coupling map sample set and the geometric flatness sample set to construct a wall geometric flatness detection channel. The wall coupling map sample set contains wall depth-design coupling maps for different wall scenes, and the geometric flatness sample set corresponds to the actual geometric flatness quantification value of each coupled image. Traverse the multiple wall depth-design coupling diagrams and extract the first wall depth-design coupling diagram; The first wall depth-design coupling diagram is input into the wall geometric flatness detection channel to obtain the first wall geometric flatness, and the first wall geometric flatness is added to the geometric flatness detection sequence.

3. The method for detecting the flatness of building walls based on AI visual recognition as described in claim 2, characterized in that, Based on the wall surface coupling map sample set and the geometric flatness sample set, multi-scale perturbation transfer learning is performed to construct a wall surface geometric flatness detection channel, including: Align the wall surface coupling image sample set and the geometric flatness sample set to obtain the geometric flatness detection sample space; Based on the geometric flatness detection sample space, train an initial model for geometric flatness detection; Multi-scale perturbation injection is performed on the geometric flatness detection sample space to obtain multiple geometric flatness detection perturbation spaces corresponding to multiple perturbation scales; Based on the multiple geometric flatness detection perturbation spaces, train multiple geometric flatness detection enhancement models; The initial model for geometric flatness detection is used as the student layer for flatness detection, and the multiple enhanced models for geometric flatness detection are used as the teacher layer for flatness detection. The teacher layer of the flatness detection performs iterative transfer learning on the student layer of the flatness detection to obtain the iterative transfer loss coefficient; If the iterative migration loss coefficient is less than the iterative migration loss threshold, the wall surface geometric flatness detection channel is generated.

4. The method for detecting the flatness of building walls based on AI visual recognition as described in claim 1, characterized in that, Based on the AdamW optimizer, a wall flatness hazard detection channel is constructed by decoupling weight decay and training on the thermal imaging flatness hazard detection event set, including: Based on the thermal imaging flatness hazard detection event set, the flatness hazard detection path is traced to establish a flatness hazard detection path space. Specifically, the thermal imaging flatness hazard detection event set contains multiple labeled samples. Each sample records wall thermal imaging data and the corresponding hazard type. Key thermal features are extracted for each event sample. The temperature distribution range is obtained through grayscale histogram analysis. The contour features of the thermal anomaly area are extracted through edge detection algorithms such as the Sobel operator. The thermal features are mapped to the hazard type and severity of the labeled sample to obtain the detection path. All detection paths are classified according to hazard type. All classified paths are integrated to construct a flatness hazard detection path space covering multiple hazard scenarios. Obtain the initialization configuration sequence of the AdamW optimizer; Based on the initialization configuration sequence, a convolutional network is trained on the path space of the flatness hazard detection using the AdamW optimizer to generate an initial model for flatness hazard detection. Based on the hazard detection loss characteristics of the initial model for flatness hazard detection, the initial configuration sequence is decoupled and weight decay is adjusted to obtain the optimized configuration sequence. The initial model for detecting flatness hazards is iteratively tuned and trained based on the tuning configuration sequence and the AdamW optimizer to generate the wall flatness hazard detection channel.

5. The method for detecting the flatness of building walls based on AI visual recognition as described in claim 1, characterized in that, The dual MOS switch drive circuit includes an MCU controller, a gate drive chip, a first MOS transistor, and a second MOS transistor. The MCU controller outputs a PWM pulse width modulation signal to the gate driver chip, which controls the turn-on and turn-off of the first MOS transistor. The second MOSFET is used to perform adaptive standby switching of the first MOSFET.

6. The method for detecting the flatness of building walls based on AI visual recognition as described in claim 1, characterized in that, The dual MOS switch drive circuit integrates a current monitoring component, which is used to monitor the camera's operating current in real time. When the camera's operating current is abnormal, the dual MOS switch drive circuit is triggered to perform a power-off protection operation.

7. A building engineering wall flatness detection system based on AI visual recognition, characterized in that, The system is used to implement the AI-based visual recognition method for detecting the flatness of building walls as described in any one of claims 1-6, the system comprising: A switch driver circuit construction module is used to build an AI vision driver circuit, which includes multiple dual MOS switch driver circuits that drive a depth camera, an industrial camera, and an infrared thermal imaging camera respectively. The dataset acquisition module is used to perform visual inspection of the building wall based on the AI ​​vision driving circuit to obtain the wall visual dataset and the vision driving monitoring dataset. The wall image acquisition module is used to perform adaptive noise reduction processing on the wall visual dataset based on the visual-driven monitoring dataset to construct a wall depth perception map, a wall industrial visual map, and a wall thermal imaging map. The geometric flatness detection sequence acquisition module is used to perform sliding window geometric flatness detection on the wall depth perception map according to the wall design scheme of the building project wall, and obtain a geometric flatness detection sequence. The surface flatness detection sequence acquisition module is used to perform sliding window surface flatness detection on the industrial visual image of the wall according to the wall design scheme, and obtain a surface flatness detection sequence. The wall flatness test report acquisition module is used to perform flatness defect detection compensation on the geometric flatness test sequence and the surface flatness test sequence based on the wall thermal imaging image, and generate a wall flatness test report.

Images