Power transmission project construction progress state detection method and device and electronic equipment

Abstract

The application relates to the technical field of image processing, and discloses a power transmission and transformation engineering construction progress state detection method and device and electronic equipment, the method comprising the following steps: acquiring multi-temporal satellite remote sensing images of a power transmission and transformation engineering construction area; inputting each temporal satellite remote sensing image into a potential diffusion model to extract multi-scale fusion feature maps corresponding to each temporal satellite remote sensing image; inputting the multi-scale fusion feature maps corresponding to each temporal phase into a state classification model to perform construction progress state classification prediction, and obtaining state probability distributions corresponding to each temporal phase; according to a pre-constructed restricted state transition matrix, combining a dynamic programming algorithm to perform state evolution path optimization on the state probability distributions corresponding to the multiple temporal phases respectively, and obtaining a construction progress state sequence; performing process node mapping according to the construction progress state sequence and a pre-acquired construction schedule, calculating a construction period deviation, and triggering corresponding grade early warning according to the construction period deviation. The application can improve the construction progress recognition accuracy of the power transmission and transformation engineering.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method, device, and electronic equipment for detecting the construction progress status of power transmission and transformation projects. Background Technology

[0002] Power transmission and transformation projects are typically characterized by long construction periods and complex construction site environments. Their construction procedures strictly follow a sequential progression, including foundation construction (including excavation, pouring, and curing), tower erection (including material handling and erection), and line stringing (including laying and tensioning). Traditional construction progress monitoring and unplanned operation inspections rely primarily on ground-based monitoring systems and "four no's and two direct" methods. However, in order to meet deadlines, construction teams sometimes deliberately fail to report plans or bring ground-based monitoring systems, creating significant safety hazards. The accuracy of safety monitoring during construction will affect the safe and stable operation of the power system and the safety of construction personnel.

[0003] The existing methods for monitoring the construction progress of power transmission and transformation projects are difficult to effectively obtain clear characteristics of construction targets (such as tower materials and excavators), resulting in decreased recognition accuracy. Furthermore, they often ignore the continuity and irreversibility of construction projects, which can easily lead to misjudgments in construction progress identification. Summary of the Invention

[0004] This invention provides a method, device, and electronic equipment for detecting the construction progress status of power transmission and transformation projects, in order to solve the technical problem of low accuracy and precision in identifying the construction progress of power transmission and transformation projects in related technologies.

[0005] In a first aspect, the present invention provides a method for detecting the construction progress status of a power transmission and transformation project, the method comprising: Acquire multi-temporal satellite remote sensing images of the construction area of ​​the power transmission and transformation project to be inspected; Satellite remote sensing images of each time phase are input into a preset potential diffusion model for forward inference, and multi-level output features of the model network are extracted for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. The multi-scale fusion feature maps corresponding to satellite remote sensing images of each time phase are input into a preset state classification model to predict the construction progress status and obtain the state probability distribution of various candidate construction progress statuses corresponding to satellite remote sensing images of each time phase. Based on the pre-constructed constrained state transition matrix, and combined with the preset dynamic programming algorithm, the state evolution path is optimized for the state probability distribution corresponding to the multi-temporal satellite remote sensing images to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images. Based on the construction progress status sequence and the pre-acquired construction plan, process nodes are mapped, the construction period deviation is calculated, and corresponding level warnings are triggered based on the construction period deviation.

[0006] In some embodiments, the latent diffusion model includes a feature encoder of a variational autoencoder (VAE) and a diffusion denoising network (U-Net); the step of inputting satellite remote sensing images of each time phase into a preset latent diffusion model for forward inference, and extracting multi-layer output features of the model network for multi-scale feature fusion to obtain a multi-scale fused feature map corresponding to each time phase satellite remote sensing image includes: Each temporal satellite remote sensing image is input into the feature encoder for feature compression encoding to obtain the potential feature vector of each temporal satellite remote sensing image; The potential feature vectors are superimposed with Gaussian noise corresponding to a preset optimal number of diffusion steps and then input into the diffusion denoising network for forward inference. During the forward inference process, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and spliced ​​together to obtain the multi-scale fused feature map.

[0007] In some embodiments, the feature encoder includes an initial convolutional layer, a multi-scale downsampling stacking module, an intermediate bottleneck attention module, and a latent spatial mapping module connected in sequence. The step of inputting each temporal satellite remote sensing image into the feature encoder for feature compression encoding to obtain the latent feature vector of each temporal satellite remote sensing image includes: Each temporal satellite remote sensing image is input into the initial convolutional layer for convolutional mapping processing to extract shallow texture feature maps; The shallow texture feature map is subjected to step-by-step downsampling compression and channel amplification processing by the multi-scale downsampling stacking module to output a high-level semantic feature map. The high-level semantic feature map is input into the intermediate bottleneck attention module for residual transformation and global semantic association modeling. The latent feature vector is obtained by performing normalization activation, channel compression, distribution parameter prediction, and reparameterization sampling on the high-level semantic feature map output by the intermediate bottleneck attention module through the latent space mapping module.

[0008] In some embodiments, the diffusion denoising network U-Net includes an encoding module, an intermediate bottleneck layer, and a decoding module. The encoding module contains multiple feature extraction layers, each consisting of two stacked residual blocks and a downsampling convolutional layer connected to the end of the layer. The decoding module is structurally symmetrical to the encoding module, and each layer of the decoding module is configured with an upsampling module and two residual blocks. The output of the upsampling module at each layer is connected to the output of the corresponding layer of the encoding module via a skip connection. The intermediate bottleneck layer is located between the encoding module and the decoding module and includes two cascaded residual blocks. The step of superimposing the latent feature vector with Gaussian noise corresponding to a preset optimal number of diffusion steps and then inputting it into the diffusion denoising network for forward inference includes: The preset optimal diffusion step number is obtained by sinusoidal position encoding mapping to obtain the diffusion step number embedding vector, and then injected into each residual block of the diffusion denoising network through a multilayer perceptron; The latent feature vector is superimposed with Gaussian noise corresponding to the preset optimal diffusion step number and then input into the encoding module. The residual transformation is performed layer by layer based on the injected diffusion step number embedding vector, and the feature dimension and channel number are adjusted by downsampling to obtain the output features of each level. The hierarchical output features at the end of the encoding module are input into the intermediate bottleneck layer, and residual transformation is performed based on the injected diffusion step embedding vector to obtain the high-level bottleneck features. The bottleneck high-level features are input into the decoding module, and the feature dimension and number of channels are adjusted by upsampling layer by layer. The residual transformation is performed based on the injected diffusion step embedding vector. The output features after upsampling at each level are concatenated with the corresponding level output features in the encoding module and then input into the residual block of the current level for residual transformation. Extracting the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network, and splicing and fusing them to obtain a multi-scale fused feature map, includes: extracting the output features of at least one preset intermediate level upsampling module in the decoding module and the hierarchical output features of the corresponding skip connection encoding module level, and splicing and fusing them to obtain the multi-scale fused feature map.

[0009] In some embodiments, the potential diffusion model is trained using the following methods: Construct an anti-interference training sample set, which includes multiple training image samples, including multiple reference image samples and noisy image samples that correspond one-to-one with the reference image samples. In the first stage, the variational autoencoder (VAE) is pre-trained based on each training image sample. After training, the network parameters of the VAE are frozen to obtain the feature encoder of the variational autoencoder (VAE). In the second stage, each training image sample is input into the feature encoder of the frozen variational autoencoder (VAE) to infer the latent feature vector of each training image sample. The diffusion steps are randomly sampled from the preset candidate diffusion step set. According to the potential diffusion forward process, Gaussian noise corresponding to the current diffusion step is superimposed on the potential feature vector of each training image sample to generate a noisy potential feature vector. The current diffusion step number is mapped by sinusoidal position encoding to obtain the diffusion step number embedding vector. The noisy latent feature vector and the current diffusion step number embedding vector are input into the diffusion denoising network to predict the superimposed noise. The sample loss is calculated based on the superimposed Gaussian noise and the predicted noise. The network parameters of the diffusion denoising network are updated by backpropagation based on the sample loss of each training image sample. The VAE keeps the parameters frozen. Iterate through the training process until the network converges to obtain the trained potential diffusion model.

[0010] In some embodiments, the anti-interference training sample set further includes a construction progress status label corresponding to each training image sample, and the status classification model is trained using the following method: Each training image sample is input into the trained feature encoder for feature compression encoding to obtain the latent feature vector of each training image sample. Randomly sample diffusion steps from a preset set of candidate diffusion steps, and superimpose Gaussian noise corresponding to the noise intensity of the current diffusion step onto the latent feature vector of each training image sample to obtain noisy latent features; The noisy latent features of each training image sample and the embedding vector of the current diffusion step number are input into the trained diffusion denoising network for forward inference. During the forward inference process, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and spliced ​​together to obtain a multi-scale fusion feature map of the training image samples. The multi-scale fusion feature map of each training image sample is input into the state classification model to predict the construction progress status, and the predicted construction progress status corresponding to each training image sample is obtained. Based on the construction progress status labels and predicted construction progress status of each training image sample, the classification loss is calculated, and the model parameters of the state classification model are updated by backpropagation based on the classification loss. The state classification model is trained iteratively until the classification loss converges, resulting in a fully trained state classification model.

[0011] In some embodiments, constructing the anti-interference training sample set includes: Collect clear image samples of multiple time phases from the construction areas of each power transmission and transformation project, and obtain the coordinates of the tower locations during the construction of the power transmission and transformation projects; Orthorectification, atmospheric correction, and image registration were performed on each clear image sample in sequence, and the sample was cropped with the tower position coordinates as the center to obtain the reference image sample. According to the construction procedures of power transmission and transformation projects, the construction progress status is divided into seven categories: foundation not started, foundation construction, foundation completed, tower erection construction, tower erection completed, line erection construction, and line erection completed. The construction progress status is marked for each of the cropped reference image samples. For each of the reference image samples, Gaussian noise is used to simulate the low signal-to-noise ratio environment of the sensor, and Perrault noise is used to simulate the cloud and fog occlusion effect, generating noisy image samples that correspond one-to-one with the reference image samples, thus constructing the anti-interference training sample set.

[0012] In some embodiments, the step of optimizing the state evolution path based on the state probability distributions corresponding to the multi-temporal satellite remote sensing images according to a pre-constructed constrained state transition matrix and a preset dynamic programming algorithm to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects includes: Based on the initial state probability distribution of each candidate construction progress state and the state probability distribution of each candidate construction progress state in the first time phase, initialize the cumulative probability corresponding to each candidate construction progress state in the first time phase. According to the time order in the multi-time phase, traverse from the second time phase to the last time phase. For each candidate construction progress state in the current time phase, calculate the transition path probability from each candidate construction progress state in the previous time phase to the current candidate construction progress state in the current time phase based on the cumulative probability and the restricted state transition matrix corresponding to each candidate construction progress state in the previous time phase. Determine the maximum transition path probability corresponding to each candidate construction progress state in the current time phase, and record the source of the transition path corresponding to each candidate construction progress state in the current time phase through a pointer matrix; Based on the maximum transition path probability corresponding to each candidate construction progress state in the current time phase and the state probability distribution of each candidate construction progress state in the current time phase, calculate the cumulative probability corresponding to each candidate construction progress state in the current time phase. When traversing to the last time phase, among the candidate construction progress states in the last time phase, the candidate construction progress state corresponding to the highest cumulative probability is selected as the endpoint of the optimized state evolution path. Using the end point of the state evolution path as the starting point for backtracking, the state evolution path is backtracked in reverse through the pointer matrix. The optimal construction progress state corresponding to each time phase before the last time phase is determined in turn. The optimal construction progress states under each time phase are sorted in chronological order to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects.

[0013] Secondly, the present invention provides a device for detecting the construction progress status of power transmission and transformation projects, the device comprising: The image acquisition module is used to acquire multi-temporal satellite remote sensing images of the construction area of ​​power transmission and transformation projects; The feature extraction module is used to input satellite remote sensing images of each time phase into a preset potential diffusion model for forward inference, and extract multi-layer output features of the model network for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. The state prediction module is used to input the multi-scale fusion feature map corresponding to each phase of satellite remote sensing imagery into a preset state classification model to perform construction progress state classification prediction, and obtain the state probability distribution of various candidate construction progress states corresponding to each phase of satellite remote sensing imagery. The state timing correction module is used to optimize the state evolution path of the state probability distribution corresponding to the multi-temporal satellite remote sensing images according to the pre-constructed constrained state transition matrix and the preset dynamic programming algorithm, so as to obtain the construction progress state sequence that conforms to the construction procedure logic of the power transmission and transformation project. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images. The status early warning module is used to map process nodes according to the construction progress status sequence and the pre-acquired construction plan, calculate the schedule deviation, and trigger the corresponding level of early warning based on the schedule deviation.

[0014] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method for detecting the construction progress status of power transmission and transformation projects as described in the first aspect or any corresponding embodiment.

[0015] The power transmission and transformation project construction progress status detection method according to embodiments of the present invention uses multi-temporal satellite remote sensing images as data input, utilizes a latent diffusion model for forward inference to extract multi-layer output features and performs multi-scale feature fusion, extracting construction detail features of different granularities from the remote sensing images, improving the ability to extract construction features in complex scenarios and enhancing status recognition accuracy; it uses a state classification model to predict the state probability distribution of candidate construction progress states for each temporal phase, avoiding early recognition errors caused by hard classification labels; it introduces a constrained state transition matrix based on the process logic of power transmission and transformation projects, and combines a dynamic programming algorithm to globally optimize the temporal state evolution path, using the inherent prior knowledge of the project's processes to constrain state jumps, suppressing state jumps, temporal logic contradictions, and single-temporal misjudgments caused by single-frame remote sensing image classification prediction, outputting a globally optimal construction progress status sequence that conforms to the actual construction process logic, ensuring the monotonicity and accuracy of the construction progress status detection results on the time axis; finally, it maps the time-optimized construction progress status sequence to the construction schedule table for process node mapping, quantitatively calculates the schedule deviation and implements hierarchical early warning, realizing automated management and control from multi-source remote sensing data acquisition to closed-loop early warning of construction progress. This invention eliminates the need for extensive manual on-site surveys, overcoming the high cost and low efficiency of manual on-site surveys. It automatically identifies and verifies the construction progress status and timing of power transmission and transformation projects using satellite remote sensing imagery. This effectively improves upon the problems of single-phase remote sensing independent classification being susceptible to noise interference and producing chaotic timing logic in the output construction progress status. It achieves automatic detection of construction progress, timing logic correction, and graded early warning of construction period anomalies for power transmission and transformation projects based on multi-phase satellite remote sensing imagery. This enables objective and automated monitoring of the actual construction status of power transmission and transformation projects, providing reliable data support for construction management and control. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram illustrating an application scenario of an embodiment of the present invention; Figure 2 A flowchart illustrating a method for detecting the construction progress status of a power transmission and transformation project, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-scale feature fusion extraction process in an embodiment of the present invention; Figure 4 This is a schematic diagram of the model architecture of the preset potential diffusion model in the embodiments of the present invention; Figure 5 A structural block diagram of a power transmission and transformation project construction progress status detection device provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] Figure 1 This is a schematic diagram of one application scenario of the present invention. As an optional application scenario of the present invention, such as... Figure 1 As shown, the power transmission and transformation project construction progress status monitoring system may include at least one terminal device and at least one server. Figure 1 The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0021] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0022] In related technologies, there are several methods for detecting construction status: threshold-based change detection methods, static semantic segmentation methods based on convolutional neural networks (CNN) or Transformers, and image enhancement methods based on GANs and their improvements. Among these, threshold-based change detection methods (such as the difference method and the ratio method) primarily work by comparing pixel differences between two consecutive images to determine whether a change has occurred in the monitored area, thereby inferring whether construction has taken place. However, this method has limitations: it struggles to distinguish specific construction stages (e.g., it cannot effectively differentiate between changes caused by "foundation construction" and "tower erection"), and it is highly susceptible to interference from seasonal vegetation growth or shedding, leading to a high false positive rate.

[0023] Static semantic segmentation methods based on convolutional neural networks (CNNs) or Transformers (such as DeepLab and Swin Transformer) work by using deep learning models to extract and segment features from single-period satellite imagery, thereby identifying and extracting the construction area of ​​the tower site. However, this method has a drawback: it only utilizes the spatial feature information of the image at the current moment, ignoring the temporal logical information contained in the historical image sequence during the construction process. Furthermore, these models exhibit poor robustness when dealing with blurred or noisy images, making it difficult to guarantee recognition accuracy.

[0024] To address common interference issues in time-series images, such as cloud cover, lighting variations, and complex terrain (e.g., mountain shadows), Generative Adversarial Networks (GANs) are used for image dehazing or enhancement preprocessing. The principle behind this approach is to remove image noise or cloud cover before recognition through generative adversarial methods, thereby improving image quality. However, the training process of GAN networks is often unstable, and artifacts are easily generated during generation. Especially in power transmission engineering scenarios, these artifacts can distort the features of critical construction details such as small tower components, severely impacting the accuracy of subsequent construction status recognition.

[0025] In summary, construction status detection methods based on convolutional neural networks or ordinary visual models struggle to extract clear features of construction targets (such as tower materials and excavators) when faced with cloud cover, lighting changes, and complex terrain (such as mountain shadows) in images, leading to a significant decrease in recognition accuracy. Furthermore, they are usually based on independent recognition of single images, ignoring the continuity and irreversibility of construction projects (e.g., irreversible process logic). This often results in logical paradoxes in the monitoring results (e.g., identifying "tower assembly construction" at one time point and "foundation construction" at another), causing misjudgments in construction progress identification.

[0026] Therefore, this invention proposes a method, device, and electronic equipment for detecting the construction progress status of power transmission and transformation projects, aiming to effectively improve the accuracy and precision of identifying the construction progress of power transmission and transformation projects in related technologies.

[0027] According to an embodiment of the present invention, a method for detecting the construction progress status of a power transmission and transformation project is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] This invention provides a method for detecting the construction progress status of power transmission and transformation projects, which can be used in the aforementioned terminal equipment or server. Figure 2 This is a flowchart illustrating a method for detecting the construction progress status of power transmission and transformation projects according to an embodiment of the present invention, as shown below. Figure 2 As shown, the method for monitoring the construction progress of this power transmission and transformation project includes the following steps: Step S201: Obtain multi-temporal satellite remote sensing images of the construction area of ​​the power transmission and transformation project to be inspected.

[0029] In this embodiment of the invention, multi-temporal phase generally refers to reflecting the characteristics of a set of remote sensing images over a time series. Broadly speaking, a set of images, maps, or geographic data of the same region acquired at different times can all be regarded as "multi-temporal" data.

[0030] In some embodiments, high-resolution satellite remote sensing images of the construction area (tower monitoring point) of the power transmission and transformation project to be monitored along the power transmission and transformation project can be acquired periodically through satellite ground receiving stations or commercial data interfaces. The image acquisition source includes, but is not limited to, artificial satellites. The spatial resolution is required to meet the preset requirements to ensure that the foundation pits and tower material textures of the power transmission and transformation towers can be clearly identified.

[0031] In some embodiments, before detecting the construction progress status of the power transmission and transformation project construction area, the method further includes: obtaining the tower location coordinates and construction schedule of the power transmission and transformation project construction area to be detected. For example, by connecting to the project construction management platform, the tower location coordinates (such as latitude and longitude information) of the power transmission and transformation project construction area in the engineering design drawings, as well as the WBS (Work Breakdown Structure) construction schedule in P6 or Project format, are obtained, and a related database of "geographic location - tower number - planned construction period" is established.

[0032] In some embodiments, after step S201, the multi-temporal satellite remote sensing images are preprocessed. For example, the acquired raw satellite remote sensing images are subjected to orthorectification, atmospheric correction, and image registration to eliminate geometric and radiation distortions caused by terrain undulations and atmospheric scattering. Based on the tower location coordinates of the power transmission and transformation project construction area, the multi-temporal satellite remote sensing images are cropped at a fixed size (e.g., 512×512 pixels) to obtain multi-temporal local image slices, forming a satellite remote sensing image set ordered by time sequence centered on the tower location.

[0033] Step S202: Input the satellite remote sensing images of each time phase into the preset potential diffusion model for forward inference, and extract the multi-layer output features of the model network for multi-scale feature fusion to obtain the multi-scale fused feature map corresponding to the satellite remote sensing images of each time phase.

[0034] Latent Diffusion Models (LDMs) are deep network models that perform diffusion processes in a low-dimensional latent space. In this embodiment of the invention, the latent diffusion model is trained in advance using remote sensing image samples. The latent diffusion model is then used to perform denoising learning on the remote sensing image samples in a low-dimensional latent space. Robust high-dimensional semantic features are reconstructed from remote sensing image samples affected by clouds, fog, and shadows, thereby training the preset latent diffusion model.

[0035] In step S202, each frame of satellite remote sensing image corresponding to each time is input into a preset latent diffusion model. Forward inference is performed through the latent diffusion model. By utilizing the powerful generative prior ability of the latent diffusion model, robust high-dimensional semantic features are reconstructed from satellite remote sensing images affected by clouds, fog, and shadows. During the forward inference process, multi-layer output features of the model network are extracted and multi-scale feature fusion is performed to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. This is beneficial to improving the accuracy of construction progress status recognition under complex backgrounds.

[0036] In this embodiment of the invention, the latent diffusion model does not perform a complete image generation (denoising) process. Instead, it uses the latent diffusion model as a feature extractor to extract multi-layer output features of the model network during the model's forward inference process and perform multi-scale feature fusion to obtain a multi-scale fused feature map for construction progress status classification.

[0037] Step S203: Input the multi-scale fusion feature map corresponding to each phase of satellite remote sensing image into the preset state classification model to perform construction progress state classification prediction, and obtain the state probability distribution of various candidate construction progress states corresponding to each phase of satellite remote sensing image.

[0038] The multi-scale fusion feature maps corresponding to satellite remote sensing images of each time phase are input into a pre-trained state classification model. The state classification model performs construction progress state classification prediction based on the multi-scale fusion feature map of each time phase, and obtains the predicted state probability corresponding to each candidate construction progress state under that time phase. After normalization processing, the state probability distribution under that time phase is formed. After traversing each time phase, the state probability distribution corresponding to each time phase is obtained. The state probability distribution corresponding to each time phase includes the predicted state probability corresponding to each candidate construction progress state under that time phase.

[0039] For example, each candidate construction progress state includes seven types: foundation not started, foundation under construction, foundation completed, tower erection under construction, tower erection completed, line erection under construction, and line erection completed. The probability distribution of each state in each time phase includes the predicted state probability corresponding to foundation not started, foundation under construction, foundation completed, tower erection under construction, tower erection completed, line erection under construction, and line erection completed.

[0040] Step S204: Based on the pre-constructed constrained state transition matrix, and combined with the preset dynamic programming algorithm, optimize the state evolution path of the state probability distribution corresponding to the multi-temporal satellite remote sensing images to obtain a construction progress state sequence that conforms to the construction procedure logic of the power transmission and transformation project. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images.

[0041] In this embodiment of the invention, a constrained state transition matrix is ​​pre-constructed based on the fixed construction procedure logic of the power transmission and transformation project to constrain the state transition relationship of the construction progress. This constrains the forward transition of the normal state, the reverse rollback of the prohibited state, and the abnormal transition of the cross-sequence state. Based on a preset dynamic programming algorithm, the state probability distribution obtained by independent prediction of each time phase is used as the observation basis. Combined with the state transition constraints, the global temporal state evolution path is optimized to correct the misjudgment of single time phase identification and the problem of disordered state transition. Finally, a construction progress state sequence that meets the actual construction process is output. This construction progress state sequence includes the optimal construction progress state corresponding to each of the multi-time phase satellite remote sensing images.

[0042] In this embodiment of the invention, the constrained state transition matrix is ​​a state transition probability matrix constructed based on the actual construction process logic of power transmission and transformation projects. The matrix dimension is consistent with the total number of candidate construction progress states. Each matrix element represents the state transition relationship between different candidate construction progress states. In the matrix, the state transition relationship corresponding to each matrix element is quantified into the corresponding state transition probability according to the state transition constraints.

[0043] In a constrained state transition matrix, matrix rows can represent the construction progress state at the previous time step (i.e., the previous time phase), and matrix columns can represent the construction progress state at the next time step (i.e., the current time phase), and vice versa. Taking the example of matrix rows representing the construction progress state at the previous time step and matrix columns representing the construction progress state at the next time step, the matrix element T in the i-th row and j-th column... ij This represents the state transition probability from the i-th construction progress state to the j-th construction progress state. It should be noted that the preceding and following moments are two consecutive moments (i.e., phases).

[0044] For reverse jumps and cross-process jumps that violate the construction sequence, the state transition probability of the corresponding matrix element is set to 0 or a minimum value, so that such jump paths are filtered out during dynamic programming optimization. For forward jumps and state maintenance that conform to the process logic, a higher state transition probability value is assigned. This achieves business logic constraints on the state evolution path and avoids the problems of random state jumps and temporal contradictions caused by single-phase classification prediction.

[0045] The term "restricted" refers to the fact that the state transition is not completely free. The transition between two adjacent moments (i.e., the previous moment and the next moment) is constrained according to the business rules governing construction procedures. These constraint rules may include: (1) Allow normal forward process jumps: for example, the previous moment is the basic non-started state, and the next moment is the basic construction state; the previous moment is the basic construction state, and the next moment is the basic completed state, etc. (2) Reverse rollback is prohibited: For example, if the tower is in the tower construction state at the previous moment, it is prohibited to rollback to the foundation construction state at the next moment. (3) Abnormal jumps across process states are prohibited: For example, if the foundation is not started at the previous moment, it is prohibited to jump directly to the overhead line construction state at the next moment. (4) Allowing adjacent moments to be in the same construction progress state: For example, the previous moment may be in the foundation construction state, and since the process may continue for multiple cycles, the next moment may also be in the foundation construction state.

[0046] For prohibited jumps, the corresponding matrix element is set to 0 or a minimum value, so that the path will not be selected as the optimal path in subsequent path optimization processes.

[0047] The embodiments of the present invention do not impose special restrictions on dynamic programming algorithms; for example, the Viterbi algorithm can be used.

[0048] Step S205: Map the process nodes according to the construction progress status sequence and the pre-acquired construction plan, calculate the schedule deviation, and trigger the corresponding level of early warning based on the schedule deviation.

[0049] After obtaining the construction progress status sequence that conforms to the construction procedure logic of power transmission and transformation projects, the construction progress status identified in each time phase is mapped and matched with the planned procedure nodes in the pre-built construction plan table to determine the procedure nodes corresponding to the construction progress status of each time phase.

[0050] By comparing the monitoring time corresponding to the construction progress status of each time phase with the planned end time of the corresponding process node, the construction period deviation is quantitatively calculated, and the deviation level is further determined according to the preset multi-level deviation threshold. Based on different deviation levels, corresponding level warnings are triggered to realize the monitoring and graded warning of construction progress deviations in power transmission and transformation projects.

[0051] The power transmission and transformation project construction progress status detection method according to embodiments of the present invention uses multi-temporal satellite remote sensing images as data input, utilizes a latent diffusion model for forward inference to extract multi-layer output features, and performs multi-scale feature fusion to extract construction detail features of different granularities from the remote sensing images, thereby improving the feature extraction capability of power transmission and transformation projects in complex scenarios and enhancing status recognition accuracy. It also uses a status classification model to predict the status probability distribution of candidate construction progress states for each temporal phase, avoiding early recognition errors caused by hard classification labels. Furthermore, it introduces a constrained state transition matrix based on the logic of power transmission and transformation project procedures, combined with dynamic rules... The algorithm performs global optimization of the temporal state evolution path, uses the inherent prior knowledge of engineering procedures to constrain state transitions, and suppresses the problems of state jumps, temporal logic contradictions, and single-phase misjudgments caused by classification and prediction of single-frame remote sensing images. It outputs a globally optimal construction progress state sequence that conforms to the logic of actual construction procedures, ensuring the monotonicity and accuracy of construction progress state detection results on the time axis. Finally, the time-optimized construction progress state sequence is mapped to the construction schedule table for process nodes, quantitatively calculates the schedule deviation, and realizes hierarchical early warning, achieving full-process automated management and control from multi-source remote sensing data acquisition to closed-loop early warning of construction progress.

[0052] This invention eliminates the need for extensive manual on-site surveys, overcoming the high cost and low efficiency of manual on-site surveys. It automatically identifies and verifies the construction progress status and timing of power transmission and transformation projects using satellite remote sensing imagery. This effectively improves upon the problems of single-phase remote sensing independent classification being susceptible to noise interference and producing chaotic timing logic in the output construction progress status. It achieves automatic detection of construction progress, timing logic correction, and graded early warning of construction period anomalies for power transmission and transformation projects based on multi-phase satellite remote sensing imagery. This enables objective and automated monitoring of the actual construction status of power transmission and transformation projects, providing reliable data support for construction management and control.

[0053] In some embodiments, the preset potential diffusion model includes a feature encoder of a variational autoencoder (VAE) and a diffusion denoising network U-Net.

[0054] Figure 3 This is a schematic diagram of the multi-scale feature fusion extraction process in an embodiment of the present invention. In some embodiments, such as... Figure 3 As shown, in step S202 above, satellite remote sensing images of each time phase are input into a preset potential diffusion model for forward inference, and multi-layer output features of the model network are extracted for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase, including steps S301 to S303.

[0055] Step S301: Input each temporal satellite remote sensing image into the feature encoder for feature compression encoding to obtain the potential feature vector of each temporal satellite remote sensing image.

[0056] To significantly reduce computational complexity while preserving high-frequency texture details (such as tower edges and surface cracks) in satellite remote sensing imagery, a fully convolutional variational autoencoder based on KL regularization is employed for the perceptual compression module of the latent diffusion model. The feature encoder in the variational autoencoder performs convolution and downsampling on the input satellite remote sensing imagery to compress the image from a high-dimensional pixel space to a low-dimensional latent space, resulting in a low-dimensional latent feature vector. This latent feature vector preserves the semantic and spatial structure information of ground features in the power transmission and transformation construction area of ​​the satellite remote sensing imagery, reducing the redundancy of the original pixel data and lowering the computational burden for subsequent diffusion denoising.

[0057] Figure 4 This is a schematic diagram of the model architecture of a preset potential diffusion model in an embodiment of the present invention. In some embodiments, such as... Figure 4 As shown, the feature encoder of the Variational Autoencoder (VAE) includes an initial convolutional layer, a multi-scale downsampling stacking module, an intermediate bottleneck attention module, and a latent spatial mapping module connected in sequence. In step S301, each temporal satellite remote sensing image is input into the feature encoder for feature compression encoding to obtain the latent feature vector of each temporal satellite remote sensing image. This includes: inputting each temporal satellite remote sensing image into the initial convolutional layer for convolutional mapping processing to extract shallow texture feature maps; performing stepwise downsampling compression and channel augmentation processing on the shallow texture feature maps through the multi-scale downsampling stacking module to output high-level semantic feature maps; inputting the high-level semantic feature maps into the intermediate bottleneck attention module for residual transformation and global semantic association modeling processing; and performing normalization activation, channel compression, distribution parameter prediction, and reparameterization sampling processing on the high-level semantic feature maps output by the intermediate bottleneck attention module through the latent spatial mapping module to obtain the latent feature vector.

[0058] In some embodiments, combined with Figure 4 As shown, the input satellite remote sensing image is a preprocessed satellite remote sensing image. The satellite remote sensing image obtained after preprocessing... , Representing the real number field, H×W×C represents the size of the image, where H represents the image height, W represents the image width, and C represents the number of channels. For example, H is 512 pixels, W is 512 pixels, and C is 3.

[0059] In some embodiments, combined with Figure 4 As shown, the initial convolutional layer uses a convolutional layer with a preset kernel size to perform convolution processing on the input satellite remote sensing image, extracting shallow texture information such as tower material edges, foundation pit outlines, and surface textures. For example, the initial convolutional layer can be a convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, mapping the input satellite remote sensing image from 3 channels to a 128-channel shallow texture feature map. .

[0060] In some embodiments, the feature encoder sets the downsampling factor. This involves performing an 8-fold spatial downsampling compression on the input satellite remote sensing imagery to map it into a latent feature space. Combined with... Figure 4 As shown, in order to achieve the compression target of the downsampling factor, the multi-scale downsampling stacking module in the feature encoder adopts three downsampling stacking modules stacked sequentially. Each downsampling stacking module can be composed of two residual blocks (ResNet Block) and one downsampling convolutional layer. The input feature map is downsampled and compressed and the channels are amplified step by step through the three downsampling stacking modules to gradually increase the number of feature channels of the feature map and reduce the spatial resolution, extract deep construction semantic information layer by layer, and output high-level semantic feature map.

[0061] For example, the first-level downsampling stacking module contains two residual blocks with 128 channels each, and is downsampled through a convolutional layer with a stride of 2. The output feature dimension of the first-level downsampling stacking module becomes 256×256×128. The second-level downsampling stacking module contains two residual blocks with 256 channels each, and is downsampled through a convolutional layer with a stride of 2. The output feature dimension of the second-level downsampling stacking module becomes 128×128×256. The third-level downsampling stacking module contains two residual blocks with 512 channels each, and is downsampled through a convolutional layer with a stride of 2. The output feature dimension of the third-level downsampling stacking module becomes 64×64×512.

[0062] In some embodiments, combined with Figure 4As shown, in order to capture the global semantic context of the image (e.g., understanding the connection between the tower base and the tower body), after downsampling, residual transformation and global semantic association modeling are performed through an intermediate bottleneck attention module. The intermediate bottleneck attention module includes a residual block, a self-attention block, and a residual block connected in sequence. The two sets of residual blocks are used to perform nonlinear feature enhancement on the high-level semantic features to alleviate the gradient degradation of the deep network. Based on the self-attention mechanism of the self-attention module, the features are flattened into a sequence, and the correlation matrix between pixels is calculated, so that the model can focus on regions that are far away but semantically related, establish the spatial relationship between distant construction targets (such as tower base, tower materials, construction pits, etc.), filter irrelevant background interference such as vegetation and mountain shadows, and finally the output feature dimension is consistent with the input feature dimension, for example, the feature dimension is kept at 64×64×512.

[0063] In some embodiments, combined with Figure 4 As shown, the latent space mapping module includes a normalization unit, an activation unit, and a convolutional unit. The high-level semantic feature map, after being processed by the intermediate bottleneck attention module, is sequentially processed by the group normalization of the normalization unit, the activation of the activation unit (e.g., SiLU activation), and the channel compression of the convolutional unit. Based on the output feature map of the convolutional unit, the distribution parameters are predicted to obtain the latent distribution mean and logarithmic variance. Finally, a differentiable low-dimensional latent feature vector is generated through the reparameterization trick (RT).

[0064] For example, a convolutional unit using a 3×3 kernel compresses the number of input feature map channels to twice the dimension of the latent space, resulting in an output feature dimension of 64×64×8. The first four channel features output by this convolutional unit are defined as the latent distribution mean. The last four channel features are defined as log-variance. To make the network differentiable, instead of random sampling, standard Gaussian noise is introduced. By reparameterizing the sampling formula: This generates the final latent feature vector z0. This latent feature vector retains the key visual information after compression and also has the characteristics of a standard normal distribution through KL regularization, serving as the baseline input for the subsequent diffusion denoising network.

[0065] Step S302: The potential feature vector is superimposed with Gaussian noise corresponding to the preset optimal diffusion step number and then input into the diffusion denoising network for forward inference.

[0066] In some embodiments, the diffusion denoising network adopts an improved U-Net architecture based on the diffusion step condition. This network receives the noisy latent features obtained by superimposing the latent feature vector with Gaussian noise corresponding to the preset optimal diffusion step number, and predicts the noise distribution through a deep convolutional network, thereby achieving feature denoising and reconstruction. In order to reduce the consumption of computing resources and adapt to the high-frequency texture characteristics of remote sensing images of power transmission and transformation projects, the diffusion denoising network of this embodiment removes the complex spatial attention module and adopts a fully convolutional residual structure.

[0067] In some embodiments, combined with Figure 4 As shown, the diffusion denoising network U-Net includes an encoder, a middle block, and a decoder. The network input consists of noisy latent features. , from the compressed latent feature vector This is obtained by superimposing Gaussian noise. During the training phase, the network output is the predicted noise. The dimension remains consistent with the network input. The predicted noise of the output represents the "noise component" that the network believes exists in the current image. By subtracting this noise component from the input, the feature can be denoised and enhanced. In the inference stage, the network no longer performs the complete image generation (denoising) process. Instead, it uses the trained diffusion denoising network U-Net as a feature extractor to extract its deep semantic features for construction progress status classification.

[0068] The encoding module contains multiple feature extraction layers, each consisting of two stacked residual blocks and a downsampling convolutional layer connected at the end of the layer.

[0069] For example, such as Figure 4 As shown, the encoding module contains four feature extraction layers: Encoding Module Layer 1, Encoding Module Layer 2, Encoding Module Layer 3, and Encoding Module Layer 4. Each layer consists of two stacked ResNet residual blocks, responsible for progressively extracting the texture and semantic features of the image. At the end of each layer, a convolutional layer with a stride of 2 and a kernel size of 3×3 is connected for downsampling to compress the spatial dimension and increase the number of channels. The ResNet residual block is the basic unit of the network, implemented using the following formula: ; in, Given the input feature map, This indicates that the input is subjected to group normalization. This represents the embedding vector of the current diffusion step. This indicates activation function processing. This represents the convolution operation. This represents the final output of the residual block.

[0070] The decoding module and the encoding module have symmetrical structures. Each level of the decoding module is configured with an up-sampling block and two residual blocks. The output of the up-sampling block at each level is connected to the output of the corresponding level of the encoding module through skip connections.

[0071] In the decoding module, each layer first performs upsampling through nearest neighbor interpolation and convolutional layers to restore spatial dimensions. At each layer, the feature map output from the corresponding layer in the encoding path is concatenated with the feature map output after upsampling in the current layer through skip connections. This design can directly pass high-frequency details lost in the encoding module (such as tower material edges) to the decoding module to assist in the reconstruction of deep semantic features. The concatenated feature map is then fused through two ResNet residual blocks in the same layer.

[0072] For example, in combination Figure 4 As shown, the encoding module includes encoding module level 1, encoding module level 2, encoding module level 3, and encoding module level 4 arranged sequentially from input to output. The decoding module includes four feature decoding levels, namely decoding module level 1, decoding module level 2, decoding module level 3, and decoding module level 4 arranged sequentially from input to output. The output of encoding module level 1 is transmitted to the output of the upsampling module of decoding module level 4 via a skip connection, and the features are concatenated and fused with the output features of the upsampling module of decoding module level 4. The output of encoding module level 2 is transmitted to the output of the upsampling module of decoding module level 3 via a skip connection, and the features are concatenated and fused with the output features of the upsampling module of decoding module level 3. The output of encoding module level 3 is transmitted to the output of the upsampling module of decoding module level 2 via a skip connection, and the features are concatenated and fused with the output features of the upsampling module of decoding module level 2. The output of encoding module level 4 is transmitted to the output of the upsampling module of decoding module level 1 via a skip connection, and the features are concatenated and fused with the output features of the upsampling module of decoding module level 1.

[0073] Through the aforementioned skip connection method, each level of the decoding module utilizes deep semantics to recover the main semantics and introduces high-frequency details from each level of the encoding module to assist each level of the decoding module in reconstructing deep semantic features, suppressing interference from cloud and fog noise and light noise, reducing the interference of remote sensing imaging noise on features, making the final extracted intermediate features more focused on the construction target of power transmission and transformation projects, and improving the accuracy of subsequent construction progress status classification.

[0074] The intermediate bottleneck layer is located between the encoding and decoding modules and consists of two cascaded residual blocks. Specifically, combined with... Figure 4 As shown, the intermediate bottleneck layer connects the output of the encoding module level 4 at the end of the encoding module and the input of the decoding module level 1. The intermediate bottleneck layer consists of two consecutive ResNet residual blocks and is responsible for processing the highest-dimensional semantic features (dimension 8×8×512) compressed by the encoding module level 4. Without changing the feature map size, the non-linear expressive power of the features is improved by increasing the network depth, thus capturing the macroscopic construction scene structure.

[0075] In some embodiments, in step S302, the potential feature vector is superimposed with Gaussian noise corresponding to a preset optimal diffusion step number and then input into the diffusion denoising network for forward inference, including the following steps 1 to 4.

[0076] Step 1: The preset optimal diffusion step number is mapped by sinusoidal position encoding to obtain the diffusion step number embedding vector, and then injected into each residual block in the diffusion denoising network through a multilayer perceptron.

[0077] The diffusion step number, also known as the diffusion time step, represents the corresponding noise intensity level. Different diffusion step numbers correspond to different noise intensities.

[0078] The current preset optimal diffusion step number is mapped using a sinusoidal position encoding method to obtain the diffusion step number embedding vector. And embed the diffusion step count into the vector. By injecting a multilayer perceptron (MLP) into each residual module of the diffusion denoising network, the network can perceive the current noise level and adaptively adjust the weights according to different noise intensities, effectively balancing denoising degree and detail preservation. The optimal number of diffusion steps is preset. .

[0079] In some embodiments, the preset optimal diffusion step number is the optimal diffusion time step parameter obtained in advance through a validation sample set after the potential diffusion model and state classification model have been trained. The method for obtaining the preset optimal diffusion step number includes: after the state classification model has been trained, fixing the model parameters of the VAE feature encoder, the diffusion denoising network, and the state classification model, constructing a validation sample set with labels of remote sensing image samples and ground truth values ​​of construction progress status; encoding each remote sensing image sample in the validation sample set with the trained VAE feature encoder to obtain a potential feature vector; traversing the preset candidate diffusion step number set, for each candidate diffusion step number, superimposing Gaussian noise with the noise intensity corresponding to the current diffusion step number onto the potential feature vector of each remote sensing image sample to obtain noisy potential features; inputting the noisy potential features of each remote sensing image sample and the diffusion step number embedded vector corresponding to the current diffusion step number into the diffusion denoising network for forward inference, extracting multi-scale fusion feature maps and inputting them into the state classification model to obtain the construction progress status prediction results of the validation sample set; calculating indicators such as classification accuracy and F1 value based on the prediction results, and selecting the diffusion step number corresponding to the optimal comprehensive evaluation index as the current preset optimal diffusion step number. During the inference phase, this optimal number of diffusion steps is used to extract features from multi-temporal satellite remote sensing images of the business scenario.

[0080] Step 2: The latent feature vector is superimposed with Gaussian noise corresponding to the preset optimal diffusion step number and then input into the encoding module. The residual transformation is performed layer by layer based on the injected diffusion step number embedding vector, and the feature dimension and number of channels are adjusted by downsampling to obtain the output features of each level.

[0081] Combination Figure 4 As shown, the latent feature vector is superimposed with Gaussian noise corresponding to the preset optimal diffusion step number to obtain noisy latent features. The noisy latent features are then input into the encoding module. At each level of the encoding module, the features are first transformed through two sets of residual blocks to increase the number of feature channels. Then, the spatial size is compressed through downsampling convolution with a stride of 2, thereby extracting multi-scale semantic features step by step. The output features of each level of the encoding module are cached for subsequent skip connections in the decoding module.

[0082] Step 3: Input the hierarchical output features at the end of the encoding module into the intermediate bottleneck layer, and perform residual transformation processing based on the injected diffusion step embedding vector to obtain the high-level features of the bottleneck.

[0083] Combination Figure 4 As shown, the hierarchical output features at the end of the encoding module are input into the intermediate bottleneck layer. The two sets of residual blocks in the intermediate bottleneck layer perform deep semantic transformation based on the injected diffusion step embedding vector, aggregate global information of the power transmission and transformation construction scenario, and output the high-level features of the bottleneck.

[0084] Step 4: Input the high-level bottleneck features into the decoding module, adjust the feature dimension and number of channels by upsampling layer by layer, and perform residual transformation based on the injected diffusion step embedding vector. In this step, the output features after upsampling at each level are concatenated with the output features of the corresponding level in the encoding module and then input into the residual block of the current level for residual transformation.

[0085] Combination Figure 4 As shown, the hierarchical structure of the decoding module and the encoding module is symmetrically set. The high-level bottleneck features output from the intermediate bottleneck layer are input into the decoding module. Each level of the decoding module first recovers the spatial resolution of the feature map through upsampling, and then splices the layer output features cached by the corresponding level of the encoding module through skip connections. The input high-level global semantic features are fused with the shallow features output by the corresponding level of the encoding module, and then input into the residual block of this level for further feature transformation.

[0086] Step S303: During the forward inference process, extract the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network, and concatenate and fuse them to obtain a multi-scale fused feature map.

[0087] In some embodiments, extracting the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network, and splicing and fusing them to obtain a multi-scale fused feature map, includes: extracting the output features of at least one preset intermediate level upsampling module in the decoding module and the hierarchical output features of the corresponding skip connection encoding module level, and splicing and fusing them to obtain a multi-scale fused feature map.

[0088] For example, the output features of the upsampling module at the i-th level (e.g., i=3) in the decoding module are extracted, and the extracted output features of the upsampling module at the decoding module level are set as follows. (For example (32×32×512). To fuse multi-scale information, output features from corresponding coding module levels (e.g., coding module level 3) are fused via skip connections. The final multi-scale fused feature map is obtained. .

[0089] In some embodiments, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and spliced ​​together to obtain a multi-scale fused feature map. This includes: extracting the output features of at least one preset intermediate level upsampling module in the decoding module and the hierarchical output features of the corresponding skip connection encoding module level, the diffusion step embedding vector, and the output features of the final level of the decoding module, and splicing them together to obtain a multi-scale fused feature map.

[0090] Combination Figure 4As shown, the output features of the preset intermediate layers (i.e., encoding module level 2 and encoding module level 3) in the encoding module are extracted, the output features of the upsampling modules of the preset intermediate layers (i.e., decoding module level 2 and decoding module level 3) in the decoding module are extracted, and the output features of decoding module level 4 are also extracted. The resolution and channel dimension alignment of the above features are processed. The aligned features are then merged into the diffusion step embedding vector corresponding to the preset optimal diffusion step number, and channel dimension splicing and fusion are performed to generate a multi-scale fused feature map.

[0091] In some embodiments, the potential diffusion model is trained by the following steps: steps a1 to g1.

[0092] Step a1: Construct an anti-interference training sample set. The anti-interference training sample set includes multiple training image samples, which include multiple reference image samples and noisy image samples that correspond one-to-one with the reference image samples.

[0093] In some embodiments, constructing an anti-interference training sample set includes: collecting clear image samples of multiple time phases of each power transmission and transformation project construction area, and obtaining the tower location coordinates of the power transmission and transformation project construction; performing orthorectification, atmospheric correction, and image registration on each clear image sample in sequence, and cropping it with the tower location coordinates as the center to obtain a reference image sample; classifying the construction progress status into seven categories according to the power transmission and transformation project construction sequence: foundation not started, foundation construction, foundation completed, tower erection, tower erection completed, line erection, and line erection completed, and labeling the construction progress status of each cropped reference image sample; for each reference image sample, using Gaussian noise to simulate the low signal-to-noise ratio environment of the sensor and Perrault noise to simulate the cloud and fog obstruction effect, generating noisy image samples that correspond one-to-one with the reference image samples, and constructing an anti-interference training sample set.

[0094] Specifically, high-resolution, clear image samples of the construction areas of each power transmission and transformation project are periodically acquired through satellite ground receiving stations or commercial data interfaces. Image sources include, but are not limited to, artificial satellites. The spatial resolution is required to meet preset requirements to ensure that the foundation pits and tower material textures of the power transmission and transformation towers can be clearly identified. By connecting to the project construction management platform, the coordinates (latitude and longitude) of the tower locations in the engineering design drawings and the WBS (Work Breakdown Structure) construction plan in P6 or Project format are obtained to establish a related database of "geographic location - tower number - planned construction period".

[0095] Image preprocessing is performed on clear images from multiple time phases, such as orthorectification, atmospheric correction, and image registration, to eliminate geometric and radiation distortions caused by topographic undulations and atmospheric scattering. Based on the tower location coordinates in the power transmission and transformation project construction area, each clear image sample is cropped to a fixed size (e.g., 512×512 pixels) to obtain local image slices from multiple time phases, forming a reference image sample set ordered by time sequence centered on the tower location.

[0096] Based on the construction procedures for power transmission and transformation projects as stipulated in the "Code for Acceptance of Construction Quality of Power Transmission and Transformation Projects", the construction progress status is refined into seven categories: "foundation not started, foundation construction (including excavation / pouring), foundation completed, tower erection, tower erection completed, line stringing, and line stringing completed". The construction progress status of each cropped benchmark image sample is manually marked.

[0097] At the sample augmentation level, in order to give the model robustness in dealing with complex weather conditions, a "noise injection" strategy is adopted to construct an anti-interference training sample set. Gaussian noise is used to simulate a low signal-to-noise ratio imaging environment, and Perlin noise is used to simulate the cloud and fog obstruction effect. Pairs of baseline image samples and noisy image samples are generated, which forces the model to learn the ability to recover construction features from noise in subsequent training.

[0098] Specifically, the process of generating noisy image samples is as follows: for any pixel in the reference image sample ( , Pixel values ​​in a clear, cloud-free state. ( , ), calculate the corresponding damaged observation image ( , This refers to noisy image samples. The calculation process is as follows: First, calculate the medium transmittance map. ( , ): ( , )=1 ( , );in, ( , The region ∈ [0,1] represents the generated Perlin noise texture. ∈[0.4,0.9] represents a randomly selected cloud and fog concentration coefficient; Then, the pixel values ​​of the damaged observation images, simulating sensor imaging noise and cloud / fog processes, are calculated: ( , )= ( , ) ( , )+ [1 ( , )]+ ( , ); in, This represents the global atmospheric light value, typically a high brightness value (e.g., 220-255). ( , Let be a random variable that follows a zero-mean Gaussian distribution. (0, 2 When generating damaged observation images, randomly select... ∈[0.01,0.15] to simulate low signal-to-noise ratio imaging quality under different gains.

[0099] Step b1: In the first stage, the variational autoencoder (VAE) is pre-trained based on each training image sample. After training, the network parameters of the VAE are frozen to obtain the feature encoder of the variational autoencoder (VAE).

[0100] The pre-training process includes: inputting each training image sample into the feature encoder of the Variational Autoencoder (VAE), which generates the distribution of latent feature vectors based on the input training image samples and predicts the output distribution parameters: mean μ and variance σ²; sampling latent feature vector z from the distribution of latent feature vectors using reparameterization techniques; inputting the sampled latent feature vector z into the feature decoder of the VAE, which decodes the latent feature vector z into a reconstructed image; calculating the loss function, which consists of reconstruction loss and KL divergence loss, respectively calculating the difference between the reconstructed image and the original training image samples, and the difference between the latent variable distribution and the standard normal distribution; backpropagation: calculating the gradient of the loss function and updating the network parameters of the VAE using optimization algorithms (such as SGD, Adam, etc.); iterative training until the network converges or reaches the preset maximum number of iterations, resulting in a trained Variational Autoencoder (VAE).

[0101] Step c1: In the second stage, each training image sample is input into the feature encoder of the frozen variational autoencoder (VAE) to infer the latent feature vector of each training image sample.

[0102] Step d1: Randomly sample or rotate the diffusion steps from the preset candidate diffusion step set, and add Gaussian noise corresponding to the current diffusion step to the potential feature vector of each training image sample according to the potential diffusion forward process to generate a noisy potential feature vector.

[0103] Step e1: Obtain the diffusion step embedding vector by sinusoidal position encoding mapping of the current diffusion step number. Input the noisy latent features and the current diffusion step embedding vector into the diffusion denoising network to predict the superimposed noise.

[0104] Step f1: Calculate the sample loss based on the superimposed Gaussian noise and the predicted noise, and update the network parameters of the diffusion denoising network by backpropagation based on the sample loss of each training image sample. The VAE keeps the parameters frozen.

[0105] Step g1: Iterate the training until the network converges to obtain the trained potential diffusion model.

[0106] In some embodiments, the anti-interference training sample set further includes a construction progress status label corresponding to each training image sample. The status classification model is trained using the following method: Each training image sample is input into a trained VAE feature encoder for feature compression encoding to obtain a latent feature vector for each training image sample; diffusion steps are randomly sampled or rotated from a preset candidate diffusion step set; Gaussian noise corresponding to the noise intensity of the current diffusion step is superimposed on the latent feature vector of each training image sample to obtain noisy latent features; the noisy latent features of each training image sample and the diffusion step embedding vector corresponding to the current diffusion step are input into a trained diffusion denoising network for further processing. Forward inference is performed; during the forward inference process, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and concatenated to obtain a multi-scale fused feature map of the training image samples; the multi-scale fused feature map of each training image sample is input into the state classification model to predict the construction progress status, and the predicted construction progress status corresponding to each training image sample is obtained; the classification loss is calculated based on the construction progress status label and the predicted construction progress status of each training image sample, and the model parameters of the state classification model are updated by backpropagation based on the classification loss; the state classification model is iteratively trained until the classification loss converges, and the trained state classification model is obtained.

[0107] By training the state classification model under different diffusion step conditions, the state classification model can be adapted to multi-scale fusion features under different noise intensities.

[0108] In some embodiments, in step S203, the state classification model uses a lightweight classification head to predict the state probability of satellite remote sensing images at each time phase under each construction progress state.

[0109] Combination Figure 4 As shown, the state classification model includes a global average pooling layer, a fully connected layer, and an activation layer. The classification loss of the state classification model can be the cross-entropy loss.

[0110] Among them, the Global Average Pooling (GAP) layer compresses the input three-dimensional feature map into a one-dimensional feature vector, eliminating spatial location sensitivity and focusing on global semantics.

[0111] Fully connected (FC) layer: Maps feature vectors to construction progress state category space. Construction progress state categories include foundation not under construction state s1, foundation under construction state s2, foundation completed state s3, tower erection state s4, tower erection completed state s5, line erection state s6, and line erection completed state s7.

[0112] Activation layer: Employing the Softmax activation function, the output layer shows the probability distribution of the current phase of satellite remote sensing imagery belonging to various construction progress states.t The calculation formula is as follows:

[0113] in, and For the weights and biases of the fully connected layer, F final This represents a multi-scale fused feature map, where N is the number of categories for construction progress status, for example, N=7, s i Let I represent the i-th construction progress state, where i = 1, 2, 3, ..., N. t P(s) represents the satellite remote sensing image at the current time t. i |I t The following indicates that the satellite remote sensing image at time t belongs to the construction progress state s. i The state probability.

[0114] In some embodiments, in step S204 above, a preset constrained state transition matrix is ​​constructed in advance in the following manner: Define state space This corresponds to the aforementioned 7 construction progress states (N=7). To describe the evolution of these states over time, a... State transition matrix T:

[0115] In matrix T, the rows represent the construction progress state i at the previous time phase t, and the columns represent the construction progress state j at the next time phase t+1, where i, j = 1, 2, 3, ..., N; each matrix element... q represents the construction progress status from the previous time phase t. t =s i The construction progress status q at time t+1 after the transfer t+1 =s j The state transition probability.

[0116] Initialize all elements of matrix T to 1. To ensure that the derivation results conform to the construction process specifications for power transmission and transformation projects, construct a physically constrained matrix. Mask matrix The initialized matrix T is corrected element-wise by the Hadamard product, i.e. Mask matrix The specific constraint rules are as follows: (1) Irreversible constraint, that is, normal forward process jump is allowed, but reverse rollback jump is prohibited: Define a Mask matrix ,matrix The rows of the matrix correspond to the construction progress state i in the previous time phase t, and the columns correspond to the construction progress state j in the next time phase t+1. Construction progress has a physical unidirectional nature; therefore, the mask matrix... The lower triangular region (excluding the main diagonal) is all zeros, that is:

[0117] In this way, it is possible to prevent a previous phase from regressing from a higher-order state to a lower-order state in a subsequent phase; for example, it is possible to prevent a regression from a higher-order state to a lower-order state in a subsequent phase. "Tower assembly completed" revert to "Basic construction".

[0118] (2) Constraints on the state evolution of adjacent time phases: Considering that the satellite revisit cycle (usually 1-3 days) is shorter than the construction cycle (weeks to months), the construction progress status usually remains unchanged or only advances to the next process state in adjacent time phases. Therefore, for the upper triangular region (including the main diagonal) of the mask matrix M, only the matrix elements on the main diagonal (self-maintaining) and the matrix elements on the first superdiagonal (one step forward) are retained as having high state transition probabilities, while the remaining positions are set to minimum values. (e.g. 10) -6 ).

[0119] The state transition probabilities corresponding to the matrix elements on the main diagonal and the matrix elements on the first superdiagonal can be set according to actual needs. This embodiment of the invention does not impose any special restrictions on this. For example, the state transition probability of the matrix elements on the main diagonal is set to P1=0.9, and the state transition probability of the matrix elements on the first superdiagonal is set to P2=1-P1=0.1.

[0120] For example, a mask matrix M can be constructed as follows:

[0121] Through physical constraints Mask matrix The initialized matrix T is corrected element-wise using the Hadamard product to construct the final restricted state transition matrix. , , Each matrix element T ij It can be represented as: ; If the construction progress state i of the previous time phase is the same as the construction progress state j of the next time phase, then the corresponding state transition probability is set to P1, for example, P1 = 0.9; if the construction progress state j of the next time phase is the next construction progress state after the construction progress state i of the previous time phase, then the corresponding state transition probability is set to P2, P2 = 1 - P1, for example, P2 = 0.1; if the construction progress state j of the next time phase is a construction progress state before the construction progress state i of the previous time phase, then the corresponding state transition probability is set to 0; for other cases, the corresponding state transition probability is set to a minimum value. ,For example .

[0122] Constructing a sparse, constrained state transition matrix in the above manner not only conforms to the logic of actual construction procedures in engineering, but also significantly reduces the computational complexity of subsequent calculations.

[0123] In some embodiments, in step S204 above, based on the pre-constructed constrained state transition matrix, and combined with a preset dynamic programming algorithm, the state evolution path is optimized for the state probability distribution corresponding to the multi-temporal satellite remote sensing images to obtain a construction progress state sequence that conforms to the construction procedure logic of the power transmission and transformation project, including steps a2 to f2.

[0124] Step a2: Based on the initial state probability distribution of each candidate construction progress state and the state probability distribution of each candidate construction progress state in the first time phase, initialize the cumulative probability corresponding to each candidate construction progress state in the first time phase.

[0125] Assuming there are T time phases, the observation sequence of the state probability distribution of the construction area (tower monitoring points) of the power transmission and transformation project to be monitored, ordered sequentially within the monitoring time window T, is predicted using the above state classification model. ,in, This represents the state probability distribution of each candidate construction progress state at the t-th time phase output by the state classification model. t=1,2,3,…,T, where T is the maximum number of time phases acquired, I t This represents the satellite remote sensing image acquired at time t. This indicates that the t-th time-phase satellite remote sensing image is in the candidate construction progress state s. i The state probability.

[0126] In dynamic programming algorithms, to prevent numerical underflow caused by long sequence multiplication, this embodiment of the invention converts probability multiplication into logarithmic addition, logarithmically converts the state probabilities of each candidate construction progress state in the state probability distribution corresponding to each time into logarithmic emission probabilities, and logarithmically converts the constrained state transition matrix into a logarithmic state transition matrix.

[0127] Define logarithmic emission probability , O represents the state probability distribution corresponding to the t-th time. t The state probability of the j-th candidate construction progress state, B tj for The logarithmic emission probability obtained after logarithmic transformation To prevent the appearance of a small constant in ln(0), similarly, a logarithmic state transition matrix is ​​defined. A is the constrained state transition matrix constructed above.

[0128] definition For the t-th time phase, the construction progress state s is reached. i The cumulative probability represents the probability of reaching the construction progress state s at time t. i The higher the cumulative probability of a state path, the lower the state path cost, and vice versa.

[0129] In the initialization phase (t=1), based on the initial state probability distribution O0 of each candidate construction progress state and the state probability distribution of each candidate construction progress state after logarithmic transformation in the t=1 phase, the cumulative probability corresponding to each candidate construction progress state in the t=1 phase is calculated. The cumulative probability corresponding to the i-th candidate construction progress state in the t=1 phase is expressed as: ; Where N represents the number of categories of candidate construction progress status. To initialize the initial state probability of the i-th candidate construction progress state in the initial state probability distribution O0, initially, the initial state probability distribution O0 is set to the state s where the foundation has not started. i The initial state probability of B is 1, and the initial state probability of the other candidate construction progress states is 0; 1i Let represent the logarithmic emission probability corresponding to the i-th candidate construction progress state in the state probability distribution of the t=1-th time phase.

[0130] Step b2: Traverse the second to the last time phase in the time sequence of the multi-time phase. For each candidate construction progress state in the current time phase, calculate the transition path probability from each candidate construction progress state in the previous time phase to the current candidate construction progress state in the current time phase based on the cumulative probability and the restricted state transition matrix corresponding to each candidate construction progress state in the previous time phase.

[0131] Specifically, during the recursive phase (2≤t≤T), the second to the Tth time phases are traversed in chronological order across multiple time phases. For the j-th candidate construction progress state in the current time phase t, the following transition path probability calculation is performed: Iterate through all candidate construction progress states in the previous time phase (t-1). For the i-th candidate construction progress state in the previous time phase (t-1), based on the cumulative probability corresponding to the i-th candidate construction progress state in the previous time phase (t-1) and the state transition probability of the i-th candidate construction progress state to the j-th candidate construction progress state in the restricted state transition matrix after logarithmic transformation, calculate the transition path probability of the i-th candidate construction progress state in the previous time phase (t-1) to the j-th candidate construction progress state in the current time phase t. The transition path probability = , Let be the cumulative probability corresponding to the i-th candidate construction progress state in the previous time phase (t-1). Let be the state transition probability of the i-th candidate construction progress state to the j-th candidate construction progress state in the constrained state transition matrix.

[0132] Step c2: Determine the maximum transition path probability corresponding to each candidate construction progress state in the current time phase, and record the source of the transition path corresponding to each candidate construction progress state in the current time phase through a pointer matrix.

[0133] For the j-th candidate construction progress state at the current time t, the maximum transition path probability corresponding to the j-th candidate construction progress state is: The maximum transition path probability corresponding to the j-th candidate construction progress state at the current time t is the maximum value among the transition path probabilities corresponding to the j-th candidate construction progress state at the current time t; simultaneously, using the pointer matrix... Record the origin of the transfer path for future backtracking.

[0134] In the pointer matrix In the pointer matrix Each row in the matrix represents a time phase, and each column represents a candidate construction progress status, or vice versa. Matrix elements As the corresponding source of the transition path, the source of the transition path corresponding to the j-th candidate construction progress state in the current time phase t can be represented as: ; The source of the transition path corresponding to the j-th candidate construction progress state in the current time phase t represents the candidate construction progress state with the highest probability of generating a transition path with the j-th candidate construction progress state in the previous time phase (t-1). It can be represented by the category number of the candidate construction progress state (such as 1, 2, 3, ..., N).

[0135] Step d2: Based on the maximum transition path probability corresponding to each candidate construction progress state in the current time phase and the state probability distribution of each candidate construction progress state in the current time phase, calculate the cumulative probability corresponding to each candidate construction progress state in the current time phase.

[0136] For the j-th candidate construction progress state at the current time phase t, the cumulative probability corresponding to the j-th candidate construction progress state for: ; Among them, B tj Let represent the logarithmic emission probability of the j-th candidate construction progress state in the t-th time phase of the current time phase t state probability distribution.

[0137] Step e2: When traversing to the last time phase, select the candidate construction progress state with the highest cumulative probability from among the candidate construction progress states in the last time phase as the endpoint of the optimized state evolution path.

[0138] When the traversal reaches the last time phase T, through steps b2 to d2 above, the cumulative probability corresponding to each candidate construction progress state in the T-th time phase is calculated. Among the candidate construction progress states in the T-th time phase, the candidate construction progress state corresponding to the maximum cumulative probability is selected as the endpoint of the optimal state evolution path obtained through optimization, that is, the best construction progress state in the T-th time phase. It can be represented as: , The candidate construction progress status can be represented by its category number.

[0139] Step f2: Using the end point of the state evolution path as the backtracking starting point, the state evolution path is backtracked in reverse through the pointer matrix. The optimal construction progress state corresponding to each time phase before the last time phase is determined in turn. The optimal construction progress states under each time phase are sorted in chronological order to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects.

[0140] In the pointer matrix, the candidate construction progress state of the (T-1)th time phase recorded in the source of the transition path corresponding to the optimal construction progress state in the (T-1)th time phase is obtained, and this is taken as the optimal construction progress state of the (T-1)th time phase; the candidate construction progress state of the (T-2)th time phase recorded in the source of the transition path corresponding to the optimal construction progress state in the (T-1)th time phase is obtained, and this is taken as the optimal construction progress state of the (T-2)th time phase; and so on, until the optimal construction progress state of the first time phase is obtained by backtracking through the pointer matrix. The optimal construction progress state of the first time phase is the starting point of the state evolution path obtained through optimization. The optimal construction progress state of the kth time phase (k < T) can be expressed as: ;in, This represents the optimal construction progress state at the k-th time phase. This represents the optimal construction progress state at the (k+1)th time phase in the pointer matrix. The corresponding transfer path origin.

[0141] The final construction progress state sequence is a construction progress state chain after full-time logic correction. This sequence has mathematical monotony and non-decreasing property, which effectively eliminates noise in single-frame recognition and ensures the absolute correctness of the construction progress monitoring results in terms of time logic.

[0142] In some application scenarios, in step S205, the objective construction progress status retrieved by remote sensing is mapped to the planning dimension of project management, and a digital twin model is established to realize quantitative comparison and automatic closed-loop control of construction progress.

[0143] First, an automatic mapping mechanism is pre-built to connect the construction progress status identified by remote sensing to the work process nodes in the Work Breakdown Structure (WBS) construction plan. The system reads the pre-set project WBS construction plan and establishes a dynamic association rule base to map the remotely sensed construction progress status to the work process nodes in the construction plan. Specifically, the system maps the identified "foundation construction" status to the "foundation engineering" node in the WBS, the "tower erection construction" status to the "tower engineering" work process node, and the "line stringing construction" status to the "line stringing" work process node, extracting the "planned end time" for each node as benchmark reference data.

[0144] Secondly, the system automatically calculates the schedule deviation based on the time-series correction status. For each tower monitoring point along the power transmission and transformation project, the system obtains the current construction progress status after time-series logic correction. If the tower is in an incomplete intermediate state (e.g., during foundation pouring or tower assembly), the system automatically queries the planned end time of the corresponding process node in the rule base and compares the current monitoring time with the planned end time, calculating the difference to determine the specific schedule deviation. This schedule deviation can be expressed in days or hours. If the current monitoring time is less than or equal to the planned end time, the construction progress is determined to be normal or ahead of schedule; if the current monitoring time has exceeded the planned end time but the process is still not completed, the construction progress is determined to be substantially delayed.

[0145] Finally, based on the schedule deviation of each tower monitoring point along the power transmission and transformation project, a graded early warning and visualization closed-loop release are implemented. The system triggers the corresponding grade early warning and performs visualization rendering on the GIS digital twin platform based on the calculated schedule deviation.

[0146] For example, when the construction period deviation of a tower monitoring point is less than or equal to the first deviation threshold (e.g., 0 days), the construction progress is determined to be at the first deviation level, indicating normal or ahead-of-time progress. In this case, the tower location element corresponding to the tower monitoring point on the GIS digital twin platform is visualized in green. When the construction period deviation is greater than the first deviation threshold but less than or equal to the second deviation threshold (e.g., 7 days), the construction progress is determined to be at the second deviation level, indicating general delays. In this case, the tower location element corresponding to the tower monitoring point on the GIS digital twin platform is visualized in flashing yellow, and a prompt message is generated on the system interface. When the construction period deviation exceeds the second deviation threshold, the construction progress is determined to be at the third deviation level, indicating severe delays. In this case, the tower location element corresponding to the tower monitoring point on the GIS digital twin platform is visualized in highlighted red, and the system will automatically generate supervision information and push it to the project manager and supervisor's terminals in real time via SMS or email interface, thereby realizing closed-loop management from "satellite detection of delays" to "on-site supervision and rectification".

[0147] In actual testing, the power transmission and transformation engineering construction progress status detection method of this invention was used to conduct unplanned operation inspections on 78 towers of a 750 kV line project in a certain area. It can achieve automatic early warning. After evaluation, the average cross-interaction ratio (mIoU) for judging the tower operation status reached 0.89, which is 34.8% higher than the detection method using the SAM model and 21.5% higher than the detection method based on a single time-series discrimination model of U-Net.

[0148] This invention addresses the problems of long construction cycles, complex construction environments (such as cloud cover and mountain shadows) in power transmission and transformation projects, and the high false alarm rate caused by the lack of temporal logic constraints in traditional single-frame recognition methods. Through the aforementioned power transmission and transformation project construction progress status detection method, it achieves fully automated management and control from multi-source remote sensing data acquisition to closed-loop early warning of construction progress.

[0149] On the one hand, in order to solve the problem of scarce training image samples under severe weather conditions, this invention acquires high-resolution satellite images with a resolution better than 0.5 meters and engineering WBS construction plan tables, and adopts a "physical level noise injection" strategy. By introducing Gaussian noise to simulate the low signal-to-noise ratio environment of the sensor, and using Perlin noise to simulate the atmospheric scattering and occlusion effects of clouds and fog, pairs of reference image samples and noisy image samples are generated. Through data augmentation, the subsequent model is forced to be able to identify various extreme environments during the training phase, thereby possessing robustness in recovering the characteristics of construction targets from noise and blur.

[0150] On the other hand, to address the problem that small construction targets (such as tower materials and pits) are easily obscured, this embodiment of the invention constructs a feature encoder of a latent diffusion model and a diffusion denoising network based on the U-Net architecture. During the training phase, the model learns to restore noisy latent features to clear features. During the inference phase, the powerful generative prior ability of the latent diffusion model is utilized to "fill in" the obscured texture details. The upsampled output features of the decoding module of the diffusion denoising network and the output features of the corresponding skip connection encoding module are extracted and fused to extract high-dimensional semantic features that are not sensitive to clouds and shadows, thereby significantly improving the accuracy of construction progress status recognition in complex backgrounds.

[0151] Furthermore, to address common logical paradoxes in single-frame recognition, such as "state regression" (e.g., the tower reverts to a foundation pit after construction) or "abnormal jumps," engineering process constraints are introduced. By constructing a constrained state transition matrix, applying a physical mask of "irreversibility" and "continuous evolution" using the Hadamard product, and combining it with the Viterbi algorithm to perform dynamic programming optimization within the entire time window, the independent state recognition results are concatenated into a time series that conforms to physical laws, eliminating noise that violates the logic of construction procedures, and ensuring the monotonicity and accuracy of the monitoring results on the time axis.

[0152] Ultimately, by establishing an automatic mapping mechanism from construction progress status to the process nodes of the construction plan, the monitoring time of the corrected objective construction progress status is compared with the planned end time of the corresponding process node in real time. The system automatically calculates the schedule deviation. Based on the schedule deviation, the system triggers a three-level warning system of "red, yellow, and green" and displays it on the GIS digital twin platform. It also automatically generates supervision information and pushes it to the management terminal. This breaks down the barriers between technology and management, transforms remote sensing data into direct decision support information, and realizes a digital closed-loop management from "satellite discovery lag" to "on-site supervision and rectification".

[0153] This invention also provides a device for detecting the construction progress status of power transmission and transformation projects. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0154] Figure 5 This is a structural block diagram of a power transmission and transformation project construction progress status detection device provided in an embodiment of the present invention. The present invention provides a power transmission and transformation project construction progress status detection device, such as... Figure 5 As shown, the power transmission and transformation project construction progress status monitoring device includes: Image acquisition module 501 is used to acquire multi-temporal satellite remote sensing images of the construction area of ​​power transmission and transformation projects; The feature extraction module 502 is used to input satellite remote sensing images of each time phase into a preset potential diffusion model for forward inference, and extract multi-layer output features of the model network for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. The state prediction module 503 is used to input the multi-scale fusion feature map corresponding to each phase of satellite remote sensing image into the preset state classification model to perform construction progress state classification prediction, and obtain the state probability distribution of various candidate construction progress states corresponding to each phase of satellite remote sensing image. The state timing correction module 504 is used to optimize the state evolution path of the state probability distribution corresponding to the multi-temporal satellite remote sensing images according to the pre-constructed constrained state transition matrix and the preset dynamic programming algorithm, so as to obtain the construction progress state sequence that conforms to the construction procedure logic of the power transmission and transformation project. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images. The status early warning module 505 is used to map process nodes according to the construction progress status sequence and the pre-acquired construction plan, calculate the schedule deviation, and trigger the corresponding level of early warning based on the schedule deviation.

[0155] The power transmission and transformation project construction progress status detection device provided in this embodiment of the invention can execute the power transmission and transformation project construction progress status detection method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0156] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0157] The following is a detailed reference. Figure 6 This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0158] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0159] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the transmission and transformation engineering construction progress status detection method of the embodiments of the present invention.

[0160] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0161] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the power transmission and transformation engineering construction progress status detection method shown in the above embodiments is implemented.

[0162] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0163] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for detecting the construction progress status of power transmission and transformation projects, characterized in that, The method includes: Acquire multi-temporal satellite remote sensing images of the construction area of ​​the power transmission and transformation project to be inspected; Satellite remote sensing images of each time phase are input into a preset potential diffusion model for forward inference, and multi-level output features of the model network are extracted for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. The multi-scale fusion feature maps corresponding to satellite remote sensing images of each time phase are input into a preset state classification model to predict the construction progress status and obtain the state probability distribution of various candidate construction progress statuses corresponding to satellite remote sensing images of each time phase. Based on the pre-constructed constrained state transition matrix, and combined with the preset dynamic programming algorithm, the state evolution path is optimized for the state probability distribution corresponding to the multi-temporal satellite remote sensing images to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images. Based on the construction progress status sequence and the pre-acquired construction plan, process nodes are mapped, the construction period deviation is calculated, and corresponding level warnings are triggered based on the construction period deviation.

2. The method according to claim 1, characterized in that, The latent diffusion model includes a feature encoder of a variational autoencoder (VAE) and a diffusion denoising network (U-Net). The process involves inputting satellite remote sensing images from each time phase into the preset latent diffusion model for forward inference, extracting multi-layer output features from the model network, and performing multi-scale feature fusion to obtain a multi-scale fused feature map corresponding to each time phase satellite remote sensing image, including: Each temporal satellite remote sensing image is input into the feature encoder for feature compression encoding to obtain the potential feature vector of each temporal satellite remote sensing image; The potential feature vectors are superimposed with Gaussian noise corresponding to a preset optimal number of diffusion steps and then input into the diffusion denoising network for forward inference. During the forward inference process, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and spliced ​​together to obtain the multi-scale fused feature map.

3. The method according to claim 2, characterized in that, The feature encoder includes an initial convolutional layer, a multi-scale downsampling stacking module, an intermediate bottleneck attention module, and a latent spatial mapping module connected in sequence. The step of inputting each temporal satellite remote sensing image into the feature encoder for feature compression encoding to obtain the latent feature vector of each temporal satellite remote sensing image includes: Each temporal satellite remote sensing image is input into the initial convolutional layer for convolutional mapping processing to extract shallow texture feature maps; The shallow texture feature map is subjected to step-by-step downsampling compression and channel amplification processing by the multi-scale downsampling stacking module to output a high-level semantic feature map. The high-level semantic feature map is input into the intermediate bottleneck attention module for residual transformation and global semantic association modeling. The latent feature vector is obtained by performing normalization activation, channel compression, distribution parameter prediction, and reparameterization sampling on the high-level semantic feature map output by the intermediate bottleneck attention module through the latent space mapping module.

4. The method according to claim 2, characterized in that, The diffusion denoising network U-Net includes an encoding module, an intermediate bottleneck layer, and a decoding module. The encoding module contains multiple feature extraction layers, each consisting of two stacked residual blocks and a downsampling convolutional layer connected to the end of the layer. The decoding module is structurally symmetrical to the encoding module. Each layer of the decoding module is configured with an upsampling module and two residual blocks. The output of the upsampling module at each layer is connected to the output of the corresponding layer of the encoding module via a skip connection. The intermediate bottleneck layer is located between the encoding module and the decoding module and includes two cascaded residual blocks. The step of superimposing the latent feature vector with Gaussian noise corresponding to a preset optimal number of diffusion steps and then inputting it into the diffusion denoising network for forward inference includes: The preset optimal diffusion step number is obtained by sinusoidal position encoding mapping to obtain the diffusion step number embedding vector, and then injected into each residual block of the diffusion denoising network through a multilayer perceptron; The latent feature vector is superimposed with Gaussian noise corresponding to the preset optimal diffusion step number and then input into the encoding module. The residual transformation is performed layer by layer based on the injected diffusion step number embedding vector, and the feature dimension and channel number are adjusted by downsampling to obtain the output features of each level. The hierarchical output features at the end of the encoding module are input into the intermediate bottleneck layer, and residual transformation is performed based on the injected diffusion step embedding vector to obtain the high-level bottleneck features. The bottleneck high-level features are input into the decoding module, and the feature dimension and number of channels are adjusted by upsampling layer by layer. The residual transformation is performed based on the injected diffusion step embedding vector. The output features after upsampling at each level are concatenated with the corresponding level output features in the encoding module and then input into the residual block of the current level for residual transformation. Extracting the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network, and splicing and fusing them to obtain a multi-scale fused feature map, includes: extracting the output features of at least one preset intermediate level upsampling module in the decoding module and the hierarchical output features of the corresponding skip connection encoding module level, and splicing and fusing them to obtain the multi-scale fused feature map.

5. The method according to claim 2, characterized in that, The potential diffusion model was trained using the following method: Construct an anti-interference training sample set, which includes multiple training image samples, including multiple reference image samples and noisy image samples that correspond one-to-one with the reference image samples. In the first stage, the variational autoencoder (VAE) is pre-trained based on each training image sample. After training, the network parameters of the VAE are frozen to obtain the feature encoder of the variational autoencoder (VAE). In the second stage, each training image sample is input into the feature encoder of the frozen variational autoencoder (VAE) to infer the latent feature vector of each training image sample. The diffusion steps are randomly sampled from the preset candidate diffusion step set. According to the potential diffusion forward process, Gaussian noise corresponding to the current diffusion step is superimposed on the potential feature vector of each training image sample to generate a noisy potential feature vector. The current diffusion step number is mapped by sinusoidal position encoding to obtain the diffusion step number embedding vector. The noisy latent feature vector and the current diffusion step number embedding vector are input into the diffusion denoising network to predict the superimposed noise. The sample loss is calculated based on the superimposed Gaussian noise and the predicted noise. The network parameters of the diffusion denoising network are updated by backpropagation based on the sample loss of each training image sample. The VAE keeps the parameters frozen. Iterate through the training process until the network converges to obtain the trained potential diffusion model.

6. The method according to claim 5, characterized in that, The anti-interference training sample set also includes a construction progress status label corresponding to each training image sample. The status classification model is trained using the following method: Each training image sample is input into the trained feature encoder for feature compression encoding to obtain the latent feature vector of each training image sample. Randomly sample diffusion steps from a preset set of candidate diffusion steps, and superimpose Gaussian noise corresponding to the noise intensity of the current diffusion step onto the latent feature vector of each training image sample to obtain noisy latent features; The noisy latent features of each training image sample and the embedding vector of the current diffusion step number are input into the trained diffusion denoising network for forward inference. During the forward inference process, the mid-level output features of the encoding module and the mid-level output features of the decoding module of the diffusion denoising network are extracted and spliced ​​together to obtain a multi-scale fusion feature map of the training image samples. The multi-scale fusion feature map of each training image sample is input into the state classification model to predict the construction progress status, and the predicted construction progress status corresponding to each training image sample is obtained. Based on the construction progress status labels and predicted construction progress status of each training image sample, the classification loss is calculated, and the model parameters of the state classification model are updated by backpropagation based on the classification loss. The state classification model is trained iteratively until the classification loss converges, resulting in a fully trained state classification model.

7. The method according to claim 5, characterized in that, The construction of the anti-interference training sample set includes: Collect clear image samples of multiple time phases from the construction areas of each power transmission and transformation project, and obtain the coordinates of the tower locations during the construction of the power transmission and transformation projects; Orthorectification, atmospheric correction, and image registration were performed on each clear image sample in sequence, and the sample was cropped with the tower position coordinates as the center to obtain the reference image sample. According to the construction procedures of power transmission and transformation projects, the construction progress status is divided into seven categories: foundation not started, foundation construction, foundation completed, tower erection construction, tower erection completed, line erection construction, and line erection completed. The construction progress status is marked for each of the cropped reference image samples. For each of the reference image samples, Gaussian noise is used to simulate the low signal-to-noise ratio environment of the sensor, and Perrault noise is used to simulate the cloud and fog occlusion effect, generating noisy image samples that correspond one-to-one with the reference image samples, thus constructing the anti-interference training sample set.

8. The method according to claim 1, characterized in that, The process involves optimizing the state evolution path of the state probability distributions corresponding to the multi-temporal satellite remote sensing images based on a pre-constructed constrained state transition matrix and a pre-defined dynamic programming algorithm, to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects, including: Based on the initial state probability distribution of each candidate construction progress state and the state probability distribution of each candidate construction progress state in the first time phase, initialize the cumulative probability corresponding to each candidate construction progress state in the first time phase. According to the time order in the multi-time phase, traverse from the second time phase to the last time phase. For each candidate construction progress state in the current time phase, calculate the transition path probability from each candidate construction progress state in the previous time phase to the current candidate construction progress state in the current time phase based on the cumulative probability and the restricted state transition matrix corresponding to each candidate construction progress state in the previous time phase. Determine the maximum transition path probability corresponding to each candidate construction progress state in the current time phase, and record the source of the transition path corresponding to each candidate construction progress state in the current time phase through a pointer matrix; Based on the maximum transition path probability corresponding to each candidate construction progress state in the current time phase and the state probability distribution of each candidate construction progress state in the current time phase, calculate the cumulative probability corresponding to each candidate construction progress state in the current time phase. When traversing to the last time phase, among the candidate construction progress states in the last time phase, the candidate construction progress state corresponding to the highest cumulative probability is selected as the endpoint of the optimized state evolution path. Using the end point of the state evolution path as the starting point for backtracking, the state evolution path is backtracked in reverse through the pointer matrix. The optimal construction progress state corresponding to each time phase before the last time phase is determined in turn. The optimal construction progress states under each time phase are sorted in chronological order to obtain a construction progress state sequence that conforms to the construction procedure logic of power transmission and transformation projects.

9. A device for detecting the construction progress status of power transmission and transformation projects, characterized in that, The device includes: The image acquisition module is used to acquire multi-temporal satellite remote sensing images of the construction area of ​​power transmission and transformation projects; The feature extraction module is used to input satellite remote sensing images of each time phase into a preset potential diffusion model for forward inference, and extract multi-layer output features of the model network for multi-scale feature fusion to obtain multi-scale fused feature maps corresponding to satellite remote sensing images of each time phase. The state prediction module is used to input the multi-scale fusion feature map corresponding to each phase of satellite remote sensing imagery into a preset state classification model to perform construction progress state classification prediction, and obtain the state probability distribution of various candidate construction progress states corresponding to each phase of satellite remote sensing imagery. The state timing correction module is used to optimize the state evolution path of the state probability distribution corresponding to the multi-temporal satellite remote sensing images according to the pre-constructed constrained state transition matrix and the preset dynamic programming algorithm, so as to obtain the construction progress state sequence that conforms to the construction procedure logic of the power transmission and transformation project. The construction progress state sequence includes the optimal construction progress state corresponding to the multi-temporal satellite remote sensing images. The status early warning module is used to map process nodes according to the construction progress status sequence and the pre-acquired construction plan, calculate the schedule deviation, and trigger the corresponding level of early warning based on the schedule deviation.

10. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the power transmission and transformation project construction progress status detection method according to any one of claims 1 to 8.

Images