Ai analysis method for life line inspection of construction project

Abstract

The application relates to the technical field of computer vision, and discloses an AI analysis method for construction engineering lifeline inspection, which comprises the following steps: performing space-time registration on multi-period visual and thermal imaging data based on semantic key points, and constructing a 3D space-time difference tensor containing a bimodal evolution channel; extracting space-time features by using a disease precursor analysis model, and calculating the cross-correlation of thermal and visual features through a multi-modal fusion gate unit, so that the feature weight is dynamically modulated to suppress the thermal deformation response; and the application solves the common problem that benign respiratory interference and weak disease evolution are difficult to distinguish in a complex dynamic scene by establishing physical causal constraints and forcibly decoupling environmental thermal noise and structural intrinsic damage in the feature dimension.

Description

TECHNICAL FIELD

[0001] The application relates to an AI analysis method for construction engineering life line inspection and belongs to the technical field of computer vision. BACKGROUND

[0002] In the current maintenance of construction engineering life lines such as subway tunnels, comprehensive pipe galleries and deep foundation pits, it is a mainstream practice to use unmanned aerial vehicles or inspection robots to carry high-resolution visual acquisition equipment to monitor the structural health. The existing technology usually uses a convolutional neural network to extract features and classify single-frame images, identify morphological features such as cracks, leaks or peeling, and evaluate the structural state. Based on the single-period static image observation mode, there are principle limitations in the face of complex dynamic engineering environments. Water stains, oil stains, moss or light shadows in the engineering scene are highly similar to real cracks in terms of gray scale distribution and edge gradient features, forming interference of the same spectrum of foreign matter. Conventional algorithms are difficult to distinguish between static stains and expanding structural cracks in a single frame dimension, and false positives are easily produced in a high-noise environment. Large building structures are affected by day and night temperature differences, causing periodic thermal expansion and contraction, resulting in small displacements. The existing technology lacks the ability to decouple the cause and effect of temperature and deformation, and is prone to misjudgment of normal structure breathing as a structural displacement risk.

[0003] In addition to the shortcomings of visual recognition, the existing supervision methodology has limitations in responding to complex scene emergency response mechanisms. For example, the authorized publication number CN120409967B Chinese invention patent discloses a smart city life line engineering emergency supervision method and a thing networking large model system. Although the scheme dynamically adjusts the robot inspection frequency and sampling strategy based on gas environmental characteristics, soil data and spatial connectivity information, realizes risk-driven service scheduling, and still relies on environmental sensor threshold triggering mechanism and statistical prediction, the method focuses on solving the problem of when to perform the inspection task and how to schedule it, and does not solve the problem of how to accurately diagnose when the robot arrives at the scene in the face of non-gas leakage diseases. In the face of structure intrinsic damage such as micro-crack evolution without gas parameter anomaly, simple frequency adjustment cannot make the system distinguish between structure intrinsic damage and environmental thermal deformation and identify non-rigid attachment interference in a strong noise background. Ultimately, disease confirmation needs to rely on manual review, which cannot achieve intelligent autonomous closed loop. Although some technologies try to introduce infrared thermal imaging data or increase the size of training samples, they mostly stay at the level of simple channel splicing or pixel-level superposition of visual and thermal images, and do not establish logical constraints between visual changes and thermal changes. It is difficult to distinguish between physiological disturbance caused by environmental temperature driving and pathological evolution caused by structure failure. For the small dynamic artifacts of spider webs or attachments caused by micro air flow, the visual model lacks physical property verification and is difficult to remove.

[0004] Therefore, how to construct a technology that integrates time series evolution information and physical causal logic to accurately lock the structural disease evolution characteristics from complex environmental interference has become a technical problem to be solved by the present application. SUMMARY

[0005] To solve the problems raised in the background art, the technical scheme of the present application is as follows: an AI analysis method for life line inspection of construction projects, comprising the following steps:

[0006] Obtain multi-period visual image data and synchronous environmental thermal imaging data for the same inspection area, and use a semantic segmentation network to extract semantic key points in the visual image data that have spatial position invariance;

[0007] Based on the semantic key points, a unified spatial coordinate mapping relationship is constructed, and spatio-temporal registration is performed on the visual image data and the environmental thermal imaging data to generate aligned time series data sets;

[0008] Calculate the feature difference value in the adjacent time interval in the aligned time series data set, stack the feature difference value along the time axis, and construct a multi-modal 3D spatio-temporal difference tensor containing a visual evolution feature channel and a thermal evolution feature channel;

[0009] Input the multi-modal 3D spatio-temporal difference tensor into a disease precursor analysis model, use the 3D convolutional neural network unit in the disease precursor analysis model to perform convolution operation on the multi-modal 3D spatio-temporal difference tensor in the spatio-temporal dimension, and extract the spatio-temporal feature vector representing the dynamic evolution of the structure;

[0010] Use the multi-modal fusion gate unit in the disease precursor analysis model to calculate the temperature change trend of the thermal evolution feature channel data and the time series cross-correlation coefficient of the visual evolution feature channel data;

[0011] Generate a dynamic weight mask based on the time series cross-correlation coefficient, perform point multiplication modulation on the spatio-temporal feature vector using the dynamic weight mask, reduce the feature response with a time series cross-correlation coefficient higher than a preset correlation threshold, and retain the feature response showing a monotonic cumulative change;

[0012] Based on the modulated spatio-temporal feature vector, use a fully connected layer classifier to calculate and output the evolution trend prediction result of the structure disease.

[0013] Preferably, before the step of constructing a multi-modal 3D spatio-temporal difference tensor containing a visual evolution feature channel and a thermal evolution feature channel, a non-rigid target filtering step based on micro-motion field entropy value is further included, a dense optical flow field between adjacent frames in the aligned time series data set is calculated, and a residual optical flow field after removing the camera self-motion component is calculated based on the spatial coordinate mapping relationship; for each pixel position in the residual optical flow field, the optical flow vector direction dispersion in a preset time series window is calculated, and the optical flow vector direction dispersion is mapped to a micro-motion entropy value; the area with a micro-motion entropy value exceeding a preset rigidity threshold is defined as a non-rigid interference area and a non-rigid mask is generated, and the corresponding area data participating in the construction of the multi-modal 3D spatio-temporal difference tensor is subjected to zero processing using the non-rigid mask.

[0014] Preferably, in the step of mapping the optical flow vector direction dispersion to the micro-motion entropy value, the micro-motion entropy value is calculated using the following formula. Where N represents the number of frames within the preset timing window, and θ i p(θ) represents the direction angle of the residual optical flow vector at the pixel location in the i-th frame. i The ) represents the normalized probability distribution frequency of the orientation angle within the time window; the micro-motion entropy value H motion Used to characterize the degree of motion disorder of the physical entity corresponding to the pixel position, when the micro-motion entropy value H motion If the value exceeds a preset rigidity threshold, the physical entity corresponding to the pixel position is determined to be a non-rigid object.

[0015] Preferably, the step of generating aligned temporal data sets further includes a bidirectional spatiotemporal cyclic consistency verification step. For adjacent video frames in the aligned temporal data sets, the forward optical flow vector field from the current frame to the next frame and the backward optical flow vector field from the next frame to the current frame are calculated respectively. Based on the forward and backward optical flow vector fields, pixel-level cyclic projection is performed to calculate the cyclic regression coordinate error of each pixel and generate a cyclic consistency error map. The cyclic consistency error map is mapped to a registration confidence weight matrix using an inverse proportional function, and the registration confidence weight matrix is ​​used as a spatial attention operator to perform weighted processing on the multimodal 3D spatiotemporal difference tensor to reduce the feature values ​​in the region where the cyclic regression coordinate error is higher than a preset error threshold.

[0016] Preferably, the step of performing dot product modulation on the spatiotemporal feature vector using a dynamic weight mask includes: obtaining the corresponding temporal cross-correlation coefficient for each spatial location in the multimodal 3D spatiotemporal difference tensor; mapping the temporal cross-correlation coefficient to a suppression weight coefficient with a value between zero and one using the Sigmoid activation function; calculating the reciprocal of the suppression weight coefficient as the weight value at the corresponding location in the dynamic weight mask, so that the weight value is negatively correlated with the temporal cross-correlation coefficient; and performing element-wise dot product operation between the dynamic weight mask and the spatiotemporal feature vector to suppress the feature intensity of the thermally deformed region.

[0017] Preferably, before constructing the multimodal 3D spatiotemporal difference tensor containing visual evolution feature channels and thermal evolution feature channels, an illumination artifact decoupling step is also included: based on the visual images in the aligned temporal data set, the brightness to saturation ratio of each pixel is calculated to generate a static specular probability distribution map; based on the optical flow field corresponding to the aligned temporal data set, the divergence of the local optical flow vector is calculated to generate a dynamic glide probability distribution map; the intersection region of the static specular probability distribution map and the dynamic glide probability distribution map is calculated to generate a dynamic specular suppression mask; the aligned temporal data set is weighted using the dynamic specular suppression mask, and the feature numerical intensity of the intersection region in the multimodal 3D spatiotemporal difference tensor is set to zero.

[0018] Preferably, the method further comprises an information entropy based stream weight adaptive adjustment step, calculating image information entropy of each local region in the aligned time series data set as a texture richness indicator; based on the texture richness indicator, calculating confidence weight of the 3D convolutional neural network unit output result using a preset nonlinear mapping function; the nonlinear mapping function is set to output a reduced confidence weight value when the texture richness indicator is lower than a first preset threshold or higher than a second preset threshold; in the step of outputting the evolution trend prediction result of the structural disease, multiplying the confidence weight by the evolution trend prediction result to obtain the weighted final prediction result.

[0019] Preferably, the disease precursor analysis model further comprises a parallel normal element identification branch, and the method further comprises: inputting a single frame image in the aligned time series data set into the normal element identification branch, and extracting static semantic features using a 2D convolutional neural network; identifying non-structural normal elements in the inspection area based on the static semantic features, the non-structural normal elements including construction bare soil, engineering vehicles and operating personnel; logically associating the identification result of the non-structural normal elements with the evolution trend prediction result of the structural disease, and generating an interference exclusion signal when the spatial position corresponding to the evolution trend prediction result of the structural disease coincides with the spatial position of the non-structural normal elements.

[0020] Preferably, the semantic key points are corner points or edge feature points of building components; the step of constructing a unified spatial coordinate mapping relationship comprises: identifying a building component region in the image using a semantic segmentation network, and extracting corner points or edge feature points in the building component region; based on the position coordinates of the extracted corner points or edge feature points in different period images, calculating a homography transformation matrix using a least squares method; and resampling all period visual image data and environmental thermal imaging data to the same reference coordinate system using the homography transformation matrix.

[0021] Preferably, the step of outputting the evolution trend prediction result of the structural disease comprises: inputting the modulated spatiotemporal feature vector into a fully connected layer classifier to generate a disease class probability distribution vector; determining the disease type based on the disease class probability distribution vector, and calculating the disease evolution rate based on the pixel value change of the corresponding position in the visual evolution feature channel; comparing the disease evolution rate with a preset safety threshold, and outputting early warning information containing a high risk level mark and corresponding spatial coordinates when the disease evolution rate is greater than the preset safety threshold.

[0022] Compared with the prior art, the present application has the following advantages:

[0023] 1. In the AI analysis of construction engineering lifeline inspection, a space-time difference tensor containing visual evolution and thermal evolution information is constructed, the difference values of adjacent time step features are stacked along the time axis to generate a high-dimensional tensor, the evolution data representing the dynamic change process of the structure is directly processed, the time series static or quasi-static background interference such as water stains, oil stains, moss or light shadows is filtered out in the feature extraction stage by using the time series difference mathematical characteristics, the structural change information with space-time evolution trend is retained, the real disease with extension trend is locked by using the physical constraint of data structure, false positives caused by the same spectrum of foreign objects in complex engineering environment are avoided, and a large number of negative sample training is not needed.

[0024] 2. A physical causal logic feature weighting mechanism based on a multi-modal fusion gate unit is established, thermal imaging data collected synchronously is used as a logic modulation signal, the time series cross-correlation between thermal features and visual features is analyzed, respiratory deformation synchronized with temperature periodic change is identified, dynamic weights are generated according to the cross-correlation, and neural network feature channels are inhibited or enhanced in real time, the periodic disturbance caused by environmental temperature and the cumulative mutation caused by structural failure are decoupled in the feature space, the physical causal constraint is introduced to improve the algorithm's ability to distinguish between structural physiological activity and pathological evolution logic, and the reliability of long-period monitoring of weak structural precursors is ensured.

[0025] 3. A residual optical flow field micro-motion entropy value rigidity discrimination logic is introduced, a physical attribute verification dimension is added outside the visual recognition dimension, non-structural attachments that cause dynamic artifacts are removed, the direction dispersion and modulus fluctuation of the residual optical flow vector in the time series window after removing the camera self-motion component are calculated, which are mapped to the micro-motion entropy value, based on the low entropy of residual motion of rigid structure cracks and the high entropy of flexible attachments, an entropy value threshold is set to generate a reverse mask, the high-entropy area response is shielded at the feature input layer, the rigidity attribute of the object is inversed by using statistical characteristics, specific interference object modeling is not needed, and non-rigid dynamic interference is filtered out by zero-sample generalization through the essential difference in physical attributes. BRIEF DESCRIPTION OF DRAWINGS

[0026] Fig. 1 The construction engineering lifeline inspection AI analysis flowchart based on the space-time difference tensor of the present application;

[0027] Fig. 2 The dynamic weight modulation and thermal inhibition characteristic curve of the multi-modal fusion gate unit of the present application;

[0028] Fig. 3 The full-link data flow conversion architecture diagram of the inspection system integrated with edge desensitization function of the present application. DETAILED DESCRIPTION

[0029] The present application will be described in detail below with specific embodiments, and it should be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application.

[0030] This invention discloses an AI analysis method for lifeline inspection of construction projects. Logically, this method consists of four tightly coupled processing stages: high-precision spatiotemporal benchmark construction, physical attribute removal of multi-source interference, tensor quantization representation of high-dimensional spatiotemporal features, and feature gating analysis based on causal logic. These stages are interconnected through standardized data interfaces, ensuring that the structural evolution information contained in the original visual data is preserved and enhanced after a series of transformations, while environmental noise is progressively removed. Addressing the interference from foreign objects and dynamic artifacts caused by the movement of acquisition equipment, changes in illumination, and environmental thermal noise in construction project inspection scenarios, this solution constructs a unified spatial coordinate system based on semantic invariance. Before performing spatiotemporal registration, the system calls a pre-built semantic segmentation network, such as a lightweight network improved based on the DeepLabV3+ architecture, to process the input... Semantic parsing is performed on multi-phase visual image data to extract semantic key points with spatial invariance. These key points are preferably corner points of concrete segments, intersections of construction joints, or edge feature points of embedded parts, rather than textured corner points easily affected by dirt. The system uses the least squares method to calculate the homography transformation matrix based on semantic key points between adjacent image phases. All phases of visual image data and synchronously acquired environmental thermal imaging data are resampled to the same reference coordinate system to generate aligned temporal data sets. To address the registration deviation problem caused by feature scarcity in textured areas such as large areas of smooth cement walls in underground engineering projects, the system executes a bidirectional spatiotemporal cyclic consistency check procedure in parallel while generating aligned data. This procedure calls optical flow estimation algorithms, such as RAFT or Farneback, to calculate the optical flow from the current frame I. t Point to the next frame I t+1 Forward optical flow vector field F t→t+1 And by the next frame I t+1 Pointing to the current frame I t Backward optical flow vector field F t+1→t The system uses the forward optical flow vector field to determine the position of each pixel P in the current frame. t Project it to the next frame, and then project it back to the current frame coordinate system based on the backward optical flow vector field to obtain the regression point. Calculate the original point P t Regression point The Euclidean distance between them is used as the cyclic regression coordinate error, and the system utilizes an inverse proportional function. The cyclic regression coordinate error is mapped to a normalized registration confidence weight matrix, where The preset sensitivity coefficient ranges from 0.5 to 1.0. This registration confidence weight matrix serves as a spatial attention operator, which performs weighted suppression on unreliable regions in the subsequent feature extraction stage, thereby blocking edge artifact interference caused by registration failure.

[0031] In view of the dynamic highlight interference generated by wet pipe wall or water accumulation area under the moving light source, the system performs a light artifact decoupling procedure based on the chrominance-motion double constraint before constructing the feature tensor, calculates the brightness-to-saturation ratio of each pixel point in the aligned time series data set, and when the ratio is greater than a preset highlight threshold, such as 3.5, it is determined that the pixel is in a static highlight probable interval. At the same time, the system calculates the divergence of the local optical flow vector, and when the divergence of the local region is greater than a preset non-rigid sliding threshold, such as 0.8, it is determined that the region has a non-physical light spot sliding phenomenon. The system defines the intersection of the static highlight probable interval and the non-rigid sliding interval as the dynamic highlight region, generates a dynamic highlight suppression mask, and applies the mask to the input data, setting the pixel values of the corresponding region to zero. In view of the micro-motion interference generated by non-structural attachments such as spider webs and plant roots under micro air flow, the system performs a non-rigid target filtering procedure based on the micro-motion field entropy value. Based on the residual optical flow field after removing the camera self-motion component, the system calculates the probability distribution p(θ i ) of the direction angle of the residual optical flow vector at each pixel position within a preset time window, such as 16 frames. The system calculates the micro-motion entropy value of the position using the Shannon entropy formula . Since the residual motion direction of rigid structure cracks is highly consistent, it exhibits low entropy, while the motion direction of flexible attachments is randomly distributed, exhibiting high entropy. The system marks the region with a calculated micro-motion entropy value greater than a preset rigidity threshold, such as 1.5 bits, as a non-rigid interference region, generates a non-rigid mask, and excludes the feature response of the region in subsequent processing, thereby achieving zero-sample generalization filtering of pseudo-diseases based on physical properties.

[0032] After completing the above multiple physical and logical denoising, the system constructs a multi-modal 3D spatio-temporal difference tensor containing visual evolution feature channels and thermal evolution feature channels. The system calculates the visual feature difference value and the thermal feature difference value of adjacent time steps in the aligned time series data set and stacks them along the time axis to form a 4D tensor with dimensions (C, T, H, W), where C is the number of feature channels containing visual and thermal information, and T is the time series length. The tensor is input into a disease precursor analysis model. The backbone network of the model uses a 3D convolutional neural network unit, such as I3D or SlowFast architecture, to perform convolution operations in the space-time dimension and extract spatio-temporal feature vectors representing the dynamic evolution of the structure. To distinguish between benign respiratory deformation driven by environmental temperature and pathological evolution caused by structural failure, a multi-modal fusion gating unit is embedded in the model to calculate the time series cross-correlation coefficient between the change gradient of the thermal evolution feature channel data and the change gradient of the visual evolution feature channel data in real time. The system sets a dynamic weight mask generation function based on physical causal constraints wherein, sigma is a Sigmoid activation function, beta is an adjustment factor, when the thermal change and the visual change present a high positive correlation, i.e. rho tends to 1, the generated mask value tends to 0, thereby suppressing the feature response of the region; when the visual change is independent of the thermal change, i.e. rho tends to 0, the mask value remains high response, the system performs element-wise point multiplication modulation on the dynamic weight mask and the spatio-temporal feature vector, forcing the decoupling of the thermal deformation component in the feature space, in addition, to adapt to different texture richness of engineering scenes, the system also performs a flow weight self-adaptive adjustment procedure based on information entropy in parallel, calculates the local information entropy of the input image as a texture richness indicator, when the indicator is lower than a first preset threshold (representing extremely smooth) or higher than a second preset threshold (representing extremely chaotic), the confidence weight of the 3D convolutional neural network unit output is reduced through a nonlinear mapping function, and the weight of the parallel running 2D normal element recognition branch is correspondingly increased, finally, the spatio-temporal feature vector after modulation and weighting is input into the full connection classifier, and the evolution trend prediction result of the structure disease is output, when the predicted disease evolution rate is greater than a preset safety threshold, the system outputs the early warning information containing the high-risk level label and the corresponding spatial coordinates.

[0033] Example 1: In the scene of a subway tunnel section located in a soft soil area and affected by high groundwater level for more than ten years, the inner wall of the segment presents typical composite disease characteristics, including large area irregularly distributed water stains, moss growing due to dampness and spider webs attached to the vault, affected by the piston wind generated by train operation, the above-mentioned attachments present non-periodic micro-swing, in view of the problem that the conventional single-frame visual detection has a high false positive rate due to the interference of foreign matter with the same spectrum and dynamic artifacts in the scene, the AI analysis method for construction engineering lifeline inspection of the application is applied to the continuous monitoring of the section, the system obtains the inspection video stream and synchronous thermal imaging data of the region for 30 consecutive days, and calls the semantic segmentation network to extract the bolt hole and the intersection point of the joint as the semantic key point, constructs a unified spatial coordinate mapping relationship, and realizes the spatio-temporal registration of multi-period data at the sub-pixel level.

[0034] After constructing the multi-modal 3D spatio-temporal differential tensor, the system captures a small area covered by water stains, which shows a monotonic extension trend independent of the background water stains in continuous time series, which is different from the characteristics of quasi-static or disordered diffusion of water stain edges over time. The spatio-temporal feature vector extracted by the disease precursor analysis model shows obvious structural evolution signals. At the same time, for another area where the pipe joint produces a 0.5mm periodic opening and closing due to the diurnal temperature difference, although the traditional displacement monitoring system may issue a structural risk alert, the multi-modal fusion gating unit in the embodiment calculates the time series cross-correlation coefficient p between the displacement change curve and the ambient temperature change curve to be 0.85. Based on the physical causal constraint, the system generates a low-value dynamic weight mask to suppress the feature response of this area, and determines it as a benign thermal breathing rather than a pathological evolution. In addition, for the spider web area that produces optical flow motion under the action of wind, the non-rigid target filtering procedure based on micro-motion field entropy calculates the micro-motion entropy H motion up to 2.1 bits, which is higher than the preset rigid threshold of 1.35 bits. Accordingly, the system marks this area as a non-rigid interference area and filters it out to prevent false positives caused by non-structural micro-motion. Through the above analysis based on spatio-temporal evolution and physical causal decoupling, the system successfully identifies and warns the above-mentioned small crack expansion trend covered by water stains without triggering false alarms.

[0035] Embodiment 2: This embodiment quantitatively verifies the detection accuracy and data confidence of the bidirectional spatio-temporal cycle consistency checking mechanism and the illumination artifact decoupling mechanism in the method of the present application under extreme visual conditions by constructing a controlled physical test platform containing typical industrial strong noise and illumination disturbance. The test selects a simulated area of about 50 meters of underground comprehensive pipe gallery as the physical test environment. The inner wall of this area is a newly built smooth concrete surface, lacking texture features, and the bottom is provided with a water accumulation area to simulate the common humid environment. In the test, the high-frequency point light source carried by the inspection robot forms a strong highlight reflection on the moving track of the water accumulation surface and the wet pipe wall. During data acquisition, a Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed to simulate the thermal noise of the on-site sensor, and a 50Hz illumination frequency flash disturbance is introduced.

[0036] The experiment included two control groups and the prototype of this invention. Control group A used the basic optical flow method without any verification and decoupling modules; control group B enabled the unidirectional optical flow verification module; the prototype of this invention fully enabled the bidirectional spatiotemporal cyclic consistency verification and illumination artifact decoupling mechanism of this invention. For registration stability in texture degradation areas, the experimental system continuously acquired and processed a 10-minute video stream. Data showed that control group A, due to a lack of feature points, experienced registration drift, producing numerous edge artifacts caused by misalignment on smooth walls, with a false crack detection rate as high as 1.2 per meter. Although control group B showed some improvement, it still exhibited 0.5 false cracks per meter in areas of drastic illumination change. In contrast, the prototype of this invention, through parallel calculation of the forward optical flow F... t→t+1 With backward optical flow F t+1→t Utilizing cyclic projection error A dynamic confidence weight matrix was generated, successfully identifying high-error features in texture-poor areas and suppressing feature weights in unreliable areas by more than 85%, reducing the false detection rate to 0.1 per meter. Regarding specular interference, control group A misidentified light spots moving with the camera as dynamic cracks, with the false alarm rate increasing linearly with the increase in water accumulation area. In contrast, the present invention's sample group accurately identified over 92% of dynamic specular areas and implemented masking suppression by calculating the brightness / saturation ratio (threshold set to 3.0) and optical flow divergence (threshold set to 1.0). The unprocessed raw data contained a large amount of noise features caused by specular highlights and texture loss, resulting in a signal-to-noise ratio (SNR) of 15.2 dB. After enabling the dual mechanism, the present invention's sample group effectively filtered out unstructured artifacts, increasing the SNR to 27.5 dB. Furthermore, in the detection of minor defects (width < 0.5 mm), the recall rate increased from 72.1% in control group A to 91.5%. Table 1 shows the comparison data of detection performance under different processing mechanisms.

[0037] Table 1: Comparison of Detection Performance under Different Verification and Decoupling Mechanisms

[0038]

[0039] The above data objectively confirms that by introducing a two-way verification based on physical reversibility and a decoupling mechanism based on optical properties, this invention effectively avoids the shortcomings of existing technologies in adaptability under texture degradation and strong light interference. Experimental results show that in complex engineering environments, this method can stably suppress environmental noise and ensure that the spatiotemporal evolution characteristics of the input originate from the real physical changes of the structural entity.

[0040] Example 3: This example combines Figs. 1 to 3 An explanation of the AI ​​analysis method for lifeline inspection of construction projects, such as... Fig. 1As shown, the method inputs multi-period visual image stream and synchronous environmental thermal data, extracts semantic key points with position invariance using a semantic segmentation network, maps unified spatial coordinates based on homography transformation matrix, generates aligned time series data set, in the process, performs light artifact decoupling step based on chrominance and motion double constraints to generate dynamic highlight suppression mask, and non-rigid target filtering step based on micro-motion field entropy to calculate residual optical flow direction dispersion and remove non-structural dynamic interference, and then stacks feature difference values along the time axis to build a multi-modal 3D spatio-temporal difference tensor containing visual evolution channel and thermal evolution channel. The step simultaneously accepts registration confidence weight matrix generated by forward and backward optical flow regression error from bidirectional spatio-temporal cycle consistency verification module. Finally, the tensor is input into a disease precursor analysis model containing a 3D convolutional neural network unit, the mutual correlation coefficient of thermal and visual features is calculated through a multi-modal fusion gating unit, and a dynamic weight modulation is generated to decouple thermal-induced deformation, and the output includes disease category, evolution rate and warning of structural disease evolution trend prediction results.

[0041] As shown in Fig. 2 , the horizontal coordinate represents the time series cross-correlation coefficient p, with a value range of 0.0 to 1.0, the left vertical coordinate represents the dynamic weight mask value M, and the right vertical coordinate represents the percentage. The dynamic weight mask value M represented by the solid line in the figure shows a monotonically decreasing trend with the increase of the time series cross-correlation coefficient rho, which embodies the negative correlation inhibition logic. The dotted line represents the thermal-induced deformation suppression rate increasing with the increase of , and reaches a maximum value when p approaches 1.0, while the dot-dash line represents the structural disease retention rate maintaining a high level in the whole interval; as shown in Fig. 3 , the system architecture includes a front-end inspection platform and an AI analysis server cluster. The front-end inspection platform is composed of an inspection robot or a UAV carrying a collection device, connected to the AI analysis server cluster through a 5G or private network transmission channel. The edge privacy desensitization unit is used to blur the human face and license plate before the data enters the server. The data space alignment engine and the disease precursor analysis engine are deployed inside the server cluster, and the multi-modal fusion gating unit is integrated to perform thermal-visual causal decoupling. The processed data is transmitted to the monitoring and early warning terminal through a gigabit local area network. The terminal finally outputs the disease evolution trend prediction and high-risk coordinate thermal map, realizing the full-link closed loop from end-side collection, edge desensitization, cloud decoupling analysis to terminal early warning.

[0042] Example 4: In this example, the determination of the preset rigidity threshold in the calculation of micro-motion entropy value adopts the following standardized calibration procedure: A video sample set containing typical rigid structures such as concrete walls and typical flexible interference objects such as swinging cables and spider webs is selected. The sample duration is not less than 10 seconds, and the frame rate is 30fps. For each pixel in the sample set, the probability distribution of the residual optical flow vector direction angle and the corresponding Shannon entropy value within a 16-frame sliding window are calculated. Statistical results show that the micro-motion entropy value of the rigid structure surface is mainly distributed in the range of 0.1 to 0.8 bits, following a normal distribution N(0.4, 0.15). 2 The micro-motion entropy values ​​of flexible interference objects are mainly distributed in the range of 1.8 to 3.5 bits. Based on the principle of minimizing Bayesian classification error, the intersection point of the two distributions is set as a rigid threshold.

[0043] In the calibration environment of this embodiment, the threshold is determined to be 1.35 bits. Regarding the calibration of the adjustment factor β in the multimodal fusion gating unit, a hybrid dataset containing known thermal expansion and contraction displacements (i.e., true values ​​known) and structural crack propagation displacements is constructed. The normalized cross-correlation coefficient ρ is substituted into the suppression function M=σ(1-β⋅ρ) to maximize the suppression rate of benign thermally induced displacements (target greater than 95%) and minimize the false suppression rate of malignant crack displacements (target less than 5%). As the optimization objective, the optimization is performed in the interval [5.0, 15.0] using a grid search method. Experiments show that when... When the value is 10.0, the system achieves the best balance in suppressing highly correlated thermally induced deformation. Therefore, β is set to 10.0. The above calibration procedure ensures that the algorithm parameters are not subjectively set, but are deterministic values ​​objectively derived based on physical statistical laws and optimization objectives.

[0044] Example 5: This example provides a defense mechanism against extreme data loss scenarios that may occur when the system operates unattended for extended periods. In the actual operation of lifeline inspection in construction projects, the inspection robot may experience discontinuous loss of visual or thermal imaging data in some periods due to instantaneous fluctuations in the communication link or occasional sensor failures, resulting in an interruption of the spatiotemporal difference tensor in the time dimension. To address this boundary condition, the system has a built-in time-series self-repair and completion module. When a certain time step T is detected... i When data is missing, the system does not discard the sample directly, but instead bases its response on the adjacent valid time steps T. i-1 With T i+1 The data is used to generate visually complete frames using an optical flow-based frame interpolation algorithm, and thermally complete frames are generated using linear interpolation.

[0045] Meanwhile, the system generates a completion confidence label channel and incorporates it into the multi-modal 3D spatio-temporal difference tensor, which is set to a low confidence value such as 0.5 at the time position corresponding to the completion frame and a high confidence value 1.0 at the original frame position. When the disease precursor analysis model is processed, the label channel is read through the attention mechanism layer, and the weight contribution of the completion frame feature in the final decision is automatically reduced. In addition, to prevent privacy-sensitive information such as the occasional appearance of non-staff faces or license plates in the inspection screen from being uploaded to the cloud analysis server, the system integrates a privacy desensitization unit based on a lightweight face and license plate detection algorithm in the front-end acquisition device. Before the original image is input into the semantic segmentation network, Gaussian blur processing is performed on the detected sensitive areas with a kernel size of 15x15 and a standard deviation of 5.0, ensuring that only de-identified structural ontology data is input into the subsequent analysis process.

[0046] Embodiment 6: This embodiment provides a set of standardized parameter calibration and model training procedures. When constructing an analysis system that can adapt to various texture environments such as subway tunnels and utility tunnels, it is necessary to determine the effective working interval of the texture richness index, i.e., the image information entropy. A gradient verification data set is constructed, which includes an extremely smooth newly poured concrete wall surface to an extremely rough pipe wall surface with a large amount of attached silt. For each frame of image in the data set, the local gray level information entropy is calculated using the formula , where p(i) is the normalized probability of the gray level . At the same time, the average end point error EPE of inter-frame registration is calculated using the optical flow algorithm. The experimental data shows that when H img is lower than 1.2 bits, the EPE of the optical flow algorithm increases exponentially due to the lack of sufficient corner features, resulting in the failure of the construction of the spatio-temporal difference tensor. When H img is higher than 7.8 bits, the EPE increases again and the feature signal-to-noise ratio decreases due to the interference of high-frequency noise and chaotic texture. Based on this physical experiment result, the first preset threshold is set to 1.2 bits, and the second preset threshold is set to 7.8 bits. A non-linear mapping function is constructed for calculating the confidence weight, where μ is set to 4.5 bits, which is the mean value of the optimal texture interval, and σ is set to 2.0 bits, ensuring that the system gives the highest weight to the 3D convolutional neural network unit in the moderate texture area, and automatically reduces the weight and smoothly switches to the 2D normal element identification branch in the texture-poor or overloaded area.

[0047] After determining the above preprocessing parameters, for the training of the disease precursor analysis model, a physically constrained mixed loss function optimization strategy is adopted to force the model to learn the decoupling logic of structural evolution and environmental thermal noise. A training set containing three types of samples, i.e., real structure displacement, thermal-induced breathing deformation, and pure background noise, is constructed. During the training process, the total loss function L totalTo classify the loss L cls The weighted sum of the physical consistency regularization term L phy , that is, , where L cls Adopt the binary cross-entropy loss function to optimize the disease identification accuracy, L phy is designed as a penalty term to constrain the behavior of the multi-modal fusion gating unit, which is mathematically expressed as , where N is the number of feature points, M j is the weight mask value output by the gating unit, and p j is the time series cross-correlation coefficient of thermal and visual features. The physical meaning of this regularization term is that when the cross-correlation coefficient p j is high, indicating thermal-induced deformation, if the model output retention weight M j is also high, a huge penalty loss will be generated. By setting the hyperparameter to 0.5 and using the Adam optimizer for end-to-end training, the model is forced to automatically suppress those features that are highly related to temperature changes in the feature extraction stage, thereby learning the ability to distinguish between physiological breathing and pathological evolution without human intervention.

[0048] Embodiment 7: This embodiment describes the standardization engineering calibration procedure of the key discriminant threshold in the non-rigid target filtering step based on the micro-motion field entropy value. In actual deployment, the selection of the preset rigidity threshold directly determines the rejection accuracy of non-rigid interference and the retention rate of structural diseases, and the optimal working point needs to be determined through an offline calibration process. The calibration process selects a representative scene containing typical rigid structures, such as concrete segments, and typical non-rigid interference objects, such as hanging cables and vegetation, in the target monitoring area, and collects video sample data with a duration of not less than 1 hour. The system performs optical flow calculation and self-motion compensation on the sample data to obtain the residual optical flow vector sequence of each pixel point. For each pixel position, the optical flow direction angle probability distribution within a preset time window is calculated, and the micro-motion entropy value is calculated according to the formula . The entropy value distribution characteristics of the rigid structure region and the non-rigid interference region are statistically analyzed, and the entropy value frequency histogram of the two types of targets is constructed. Based on the Bayesian minimum error rate criterion, the intersection of the distribution curves of the two types of targets is found as the initial threshold value. A weighted cost function is introduced, where t is the to-be-determined threshold value, P fn is the missed detection rate of the rigid structure being misjudged as non-rigid, P fp is the false alarm rate of the non-rigid interference being misjudged as rigid, and l1 and l2 are the penalty weights of missed detection and false alarm, respectively. By performing grid search in the neighborhood of the initial threshold value, the entropy value that minimizes the cost function J(t) is found as the final preset rigidity threshold value, ensuring that the setting of the discriminant parameter is based on objective statistical laws and quantitative optimization objectives rather than empirical guesses.

[0049] Embodiment 8: This embodiment describes in detail the initial calibration and debugging process of the unified spatial coordinate mapping relationship and bidirectional space-time cyclic consistency verification procedure based on semantic key points in the multi-source heterogeneous data fusion scene. After the system first accesses a new inspection area or replaces the collection device, a standardized reference calibration procedure needs to be performed. In a period without rain and fog and with uniform illumination, the collection device is controlled to perform full coverage scanning on the calibration area along the preset trajectory, to obtain a reference video stream containing rich texture features and known geometric size reference objects, such as standard targets or pipe joint seams with known sizes. A motion structure recovery algorithm (SfM) is used to reconstruct a three-dimensional sparse point cloud of the area from the reference video stream, and an absolute scale factor is calculated based on the physical size of the reference object, so as to construct a reference coordinate system with true physical measurement.

[0050] For each type of heterogeneous sensor, such as a visible light camera and a thermal imager, extrinsic parameter calibration and time synchronization calibration are performed. The system, in the reference coordinate system, identifies the projection positions of the same physical feature points in different modal images, uses a PnP algorithm to solve the initial pose matrix of each sensor, and calculates the timestamp deviation between different sensors. After completing the initial calibration, the system enters an online adaptive debugging phase. In this phase, the system continuously monitors the cyclic regression coordinate error calculated by the bidirectional optical flow When the average error of consecutive N frames, for example N = 30, exceeds a preset safety threshold, for example 2.0 pixels, it is determined that the current environmental parameters have drifted or the initial calibration has failed. At this time, the system automatically triggers the recalibration mechanism, and uses the current high-confidence feature points to perform local fine-tuning optimization on the pose matrix until the average error falls below the safety threshold, ensuring that the system can always maintain a high-precision space-time registration state in long-term operation and complex working conditions.

[0051] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.

Claims

1. An AI analysis method for construction project life line inspection, characterized in that, The method comprises the following steps: acquiring multi-period visual image data and synchronous environmental thermal imaging data of the same inspection area, and extracting semantic key points in the visual image data with spatial position invariance by using a semantic segmentation network; constructing a unified spatial coordinate mapping relationship based on the semantic key points, performing space-time registration on the visual image data and the environmental thermal imaging data, and generating aligned time series data sets; calculating feature difference values in adjacent time intervals in the aligned time series data sets, stacking the feature difference values along the time axis, and constructing a multi-modal 3D space-time difference tensor containing visual evolution feature channels and thermal evolution feature channels; inputting the multi-modal 3D space-time difference tensor into a disease precursor analysis model, performing convolution operation on the multi-modal 3D space-time difference tensor in the time-space dimension by using a 3D convolutional neural network unit in the disease precursor analysis model, and extracting a space-time feature vector representing the dynamic evolution of the structure; calculating the temperature change trend of the thermal evolution feature channel data and the time series cross-correlation coefficient of the visual evolution feature channel data by using a multi-modal fusion gating unit in the disease precursor analysis model; generating a dynamic weight mask based on the time series cross-correlation coefficient, performing point multiplication modulation on the space-time feature vector by using the dynamic weight mask, reducing the feature response of the time series cross-correlation coefficient higher than a preset correlation threshold, and retaining the feature response showing a monotonic cumulative change; based on the modulated space-time feature vector, calculating and outputting the evolution trend prediction result of the structure disease by using a fully connected layer classifier.

2. The AI analysis method for construction engineering lifeline inspection according to claim 1, characterized in that, Before the step of constructing the multi-modal 3D space-time difference tensor containing the visual evolution feature channels and the thermal evolution feature channels, a non-rigid target filtering step based on micro-motion field entropy value is further included, a dense optical flow field between adjacent frames in the aligned time series data sets is calculated, and a residual optical flow field after removing the camera self-motion component is calculated based on the spatial coordinate mapping relationship; for each pixel position in the residual optical flow field, the direction dispersion of the optical flow vector in a preset time window is calculated, the micro-motion entropy value is mapped from the direction dispersion of the optical flow vector; the area with the micro-motion entropy value exceeding a preset rigidity threshold is defined as a non-rigid interference area and a non-rigid mask is generated, and the corresponding area data participating in the construction of the multi-modal 3D space-time difference tensor is processed by zero.

3. The AI analysis method for construction engineering lifeline inspection according to claim 2, characterized in that, In the step of mapping the optical flow vector direction dispersion to the micro-motion entropy value, the micro-motion entropy value is calculated by using the following formula , wherein N represents the frame number in the preset time sequence window, θ i represents the residual optical flow vector direction angle of the pixel position in the i-th frame, p(θ i ) represents the normalized probability distribution frequency of the direction angle in the time sequence window; the micro-motion entropy value H motion is used to represent the motion confusion degree of the physical entity corresponding to the pixel position, and when the micro-motion entropy value H motion is greater than a preset rigidity threshold, it is determined that the physical entity corresponding to the pixel position is a non-rigid object.

4. The AI analysis method for construction project life line inspection according to claim 1, wherein, In the step of generating the aligned time series data sets, a bidirectional space-time cycle consistency verification step is further included, for adjacent video frames in the aligned time series data sets, a forward optical flow vector field pointing from the current frame to the next frame and a backward optical flow vector field pointing from the next frame to the current frame are calculated respectively; pixel-level cyclic projection is performed based on the forward optical flow vector field and the backward optical flow vector field, the cyclic regression coordinate error of each pixel point is calculated, and a cycle consistency error map is generated; the cycle consistency error map is mapped into a registration confidence weight matrix by using an inverse proportional function, and the registration confidence weight matrix is used as a spatial attention operator to perform weighted processing on the multi-modal 3D space-time difference tensor, and the feature values of the area with the cyclic regression coordinate error higher than a preset error threshold are reduced.

5. The AI analysis method for construction project life line inspection according to claim 1, wherein, The step of performing point multiplication modulation on the space-time feature vector by using the dynamic weight mask comprises: obtaining a corresponding time series cross-correlation coefficient for each spatial position in the multi-modal 3D space-time difference tensor; mapping the time series cross-correlation coefficient to a suppression weight coefficient with a value range of zero to one by using a Sigmoid activation function; calculating the reciprocal of the suppression weight coefficient as a weight value of a corresponding position in the dynamic weight mask, so that the weight value is in a negative correlation relationship with the time series cross-correlation coefficient; and performing element-wise point multiplication operation on the dynamic weight mask and the space-time feature vector.

6. The AI analysis method for construction engineering lifeline inspection according to claim 1, characterized in that, Before the step of constructing the multi-modal 3D space-time difference tensor comprising the visual evolution feature channel and the thermal evolution feature channel, there is further included a light artifact decoupling step: based on the visual images in the aligned time series data set, calculating the brightness-to-saturation ratio of each pixel point to generate a static highlight probability distribution map; Based on the optical flow field corresponding to the aligned time series data set, the divergence of the local optical flow vector is calculated to generate a dynamic slip probability distribution map; The intersection region of the static highlight probability distribution map and the dynamic slip probability distribution map is calculated to generate a dynamic highlight suppression mask; and the aligned time series data set is weighted by using the dynamic highlight suppression mask, and the feature value intensity of the intersection region in the multi-modal 3D space-time difference tensor is set to zero.

7. The AI analysis method for construction engineering lifeline inspection according to claim 1, characterized in that, The method further comprises a flow weight adaptive adjustment step based on information entropy, which calculates the image information entropy of each local region in the aligned time series data set as a texture richness indicator; based on the texture richness indicator, a preset nonlinear mapping function is used to calculate the confidence weight of the output result of the 3D convolutional neural network unit; The nonlinear mapping function is set to output a reduced confidence weight value when the texture richness indicator is lower than a first preset threshold or higher than a second preset threshold; In the step of outputting the evolution trend prediction result of the structure disease, the confidence weight is multiplied by the evolution trend prediction result to obtain a weighted final prediction result.

8. The AI analysis method for construction project life line inspection according to claim 1, wherein, The disease precursor analysis model further comprises a parallel normal element identification branch, and the method further comprises: inputting a single frame image in the aligned time series data set into the normal element identification branch, and extracting static semantic features by using a 2D convolutional neural network; identifying non-structural normal elements in the inspection area based on the static semantic features, the non-structural normal elements including construction bare soil, engineering vehicles and operating personnel; and logically associating the identification result of the non-structural normal elements with the evolution trend prediction result of the structure disease, and generating an interference exclusion signal when the spatial position corresponding to the evolution trend prediction result of the structure disease coincides with the spatial position of the non-structural normal elements.

9. The AI analysis method for construction project life line inspection according to claim 1, wherein, The semantic key points are corner points or edge feature points of building components; and the step of constructing a unified spatial coordinate mapping relationship comprises: identifying a building component region in an image by using a semantic segmentation network, and extracting corner points or edge feature points in the building component region; calculating a homography transformation matrix by using a least squares method based on the position coordinates of the extracted corner points or edge feature points in different period images; and resampling all period visual image data and environmental thermal imaging data to the same reference coordinate system by using the homography transformation matrix.

10. The AI analysis method for construction project life line inspection according to claim 1, wherein, The step of outputting the evolution trend prediction result of the structure disease includes: inputting the modulated space-time feature vector into a full connection layer classifier to generate a disease category probability distribution vector; determining a disease type based on the disease category probability distribution vector, and calculating a disease evolution rate based on pixel value changes of corresponding positions in the visual evolution feature channel; comparing the disease evolution rate with a preset safety threshold, and when the disease evolution rate is greater than the preset safety threshold, outputting early warning information containing a high-risk level mark and corresponding spatial coordinates.

Images