Underground powerhouse construction information dynamic management method and system based on deep learning

Abstract

The invention belongs to the technical field of underground powerhouse construction management, and particularly relates to an underground powerhouse construction information dynamic management method and system based on deep learning, and the method comprises the steps: firstly obtaining historical data including construction image data, environment safety data and structure state data, so as to construct a construction progress analysis model and a construction safety analysis model; dynamic prediction is conducted on the actual construction progress and the construction safety risk through the construction progress analysis model and the construction safety analysis model; and then based on the actual construction progress and the construction safety risk which are dynamically predicted in real time, an optimal construction resource allocation scheme and an optimal construction progress adjustment strategy are generated through a multi-objective optimization algorithm. The real-time dynamic monitoring of the construction progress and the construction safety risk is realized, and the resource allocation rationality and the decision real-time performance are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of underground powerhouse construction management, and specifically relates to a method and system for dynamic management of underground powerhouse construction information based on deep learning. Background Art

[0002] Underground powerhouse construction refers to the engineering activities involved in constructing infrastructure such as large-scale industrial facilities, hydropower plant buildings, and underground transportation hubs in underground spaces. This requires complex excavation, support, and structural construction within a limited and enclosed underground space, involving multiple technologies such as surrounding rock stability control, support structure design, groundwater infiltration prevention, ventilation, and environmental control. Due to the complexity and uncertainty of geological conditions, the underground powerhouse construction process requires continuous monitoring of surrounding rock deformation, the stress state of the support structure, and changes in environmental parameters to ensure construction safety and project quality. The construction principle follows the basic process of "detection first, excavation second, support third, and lining last." Each stage is closely linked and requires precise information management and control.

[0003] Existing technologies use hierarchical models to intelligently evaluate the design of underground powerhouse ancillary caverns for pumped-storage power stations, managing multiple conflicting design objectives. However, these technologies focus solely on design-phase evaluation, rather than managing the entire construction process. Furthermore, they fail to implement real-time dynamic construction progress monitoring and risk warnings, resulting in delayed decision-making for underground powerhouse construction and irrational resource allocation and progress adjustments. Summary of the Invention

[0004] The purpose of the present invention is to address the above-mentioned problems existing in the prior art and to provide a method, system, equipment and medium for dynamic management of underground plant construction information based on deep learning, which realizes real-time dynamic monitoring of construction progress and construction safety risks, thereby improving the rationality of resource allocation and the real-time nature of decision-making.

[0005] To achieve the above objectives, the technical solutions of the present invention are as follows:

[0006] In a first aspect, the present invention provides a method for dynamic management of underground powerhouse construction information based on deep learning, the method comprising:

[0007] Acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, wherein the historical data includes construction image data, environmental safety data, and structural status data;

[0008] Building a training data set based on historical data, and training a construction progress analysis model and a construction safety analysis model based on the training data set. The construction progress analysis model and the construction safety analysis model are used to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site, respectively;

[0009] A constrained multi-objective optimization model for the underground powerhouse construction site is constructed. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress, and the construction safety risk score. A heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site.

[0010] The expression of the comprehensive objective function of the multi-objective optimization model is:

[0011] Γ=α3·f prog +β3·f res +γ3·f risk ;

[0012]

[0013] In the above formula, Γ represents the comprehensive objective function of the multi-objective optimization model; f prog 、f res 、f risk They represent the construction progress difference objective function, construction resource consumption objective function, and construction safety risk objective function respectively; α3, β3, and γ3 represent the relationship between f prog 、f res 、f risk The corresponding weight coefficient is α3+β3+γ3=1; Respectively Actual construction progress and planned construction progress at all times; is the total adjustment period; δ mat , δ lab Respectively The corresponding weight coefficient; Respectively The amount of construction materials consumed and the number of construction workers at each moment; express The construction safety risk score at the moment; Ω(·) is the safety risk mapping quantization function; η2, η3, and η4 are the third, fourth, and fifth anti-zero constants, respectively.

[0014] The construction progress analysis model performs multi-scale feature extraction and scene segmentation on construction image data under a certain construction resource allocation plan and construction progress adjustment strategy to obtain actual hierarchical scene features, calculates the difference between the actual hierarchical scene features and the hierarchical scene features under a baseline state, and weights the difference to calculate the actual construction progress.

[0015] The calculation formula for the actual construction progress is:

[0016]

[0017] In the above formula, P current is the actual construction progress; M2 is the scene feature dimension of the construction image data; γ2 is the progress magnification coefficient; is the m-th dimension scene feature in the actual hierarchical scene feature; is the m-th dimension scene feature in the hierarchical scene feature under the baseline state; ε is the first anti-zero constant.

[0018] The training objective function of the construction progress analysis model is:

[0019]

[0020] F i =ReLU(Conv i (X)+b i );

[0021] In the above formula, L total represents the training objective function of the scene analysis model; α1 and β1 are the classification weight and regression weight respectively; γ1 is the attention weighting coefficient; N1 is the total number of samples; M1 is the number of scales used; Predict the value y for the kth sample k and the true value Cross entropy classification loss; RMSE (F i ,F fused ) is the i-th scale feature F i and fusion feature F fused The root mean square error between loss is the temporal consistency loss based on the image frame sequence; A i is the attention weight of the i-th scale feature; W i is the weight matrix of the i-th scale feature; exp(·) is the exponential function; X is the input construction image data matrix; b i is the bias of the i-th scale feature; Conv i is the convolution operation of the convolution kernel of the i-th scale feature.

[0022] The construction safety analysis model first uses a two-stream convolutional network to extract actual environmental safety features and actual structural status features from environmental safety data and structural status data under a certain construction resource allocation plan and construction progress adjustment strategy. It then uses a temporal attention mechanism to calculate a fused weighted feature of the actual environmental safety features and the actual structural status features. Anomaly detection scores are calculated based on the fused weighted features.

[0023] Then, a feature extraction algorithm is used to extract the actual environmental safety feature vector and the actual structural state feature vector from the environmental safety data and structural state data under a certain construction resource allocation plan and construction progress adjustment strategy. The differences between the actual environmental safety feature vector and the environmental safety feature vector under the baseline state, as well as the differences between the actual structural state feature vector and the structural state characteristics under the baseline state are calculated. By weighting the differences and combining them with the anomaly detection score, the construction safety risk score is calculated.

[0024] The expression of the construction safety risk score is:

[0025]

[0026] In the above formula, R risk Score the construction safety risk; E current 、E baseline are the actual environmental safety feature vector and the environmental safety feature vector under the reference state respectively; S current 、S baseline are the actual structural state eigenvector and the structural state eigenvector under the reference state, respectively; ||·||2 is the Euclidean norm; α2 and β2 are the weighting coefficients of environmental safety data and structural state data, respectively; η1 is the second zero-proof constant; κ is the normalization coefficient of construction safety risk score; S anomaly Scoring anomaly detection;

[0027] The expression of the anomaly detection score is:

[0028]

[0029] F env =ReLU(Conv env (X env )+b env );

[0030]

[0031] F str =ReLU(Conv str (X str )+b str );

[0032] In the above formula, ζ is the anomaly The corresponding weight coefficient; N2 is the feature dimension after fusion weighting; represents the fusion weighted features; is the lth element in the fusion weighted feature; μ l , σ lare the mean and standard deviation of the lth element in the fusion weighted feature respectively; δ is the time decay factor; Δt is the difference between the current time and the reference time; F env 、F str are the actual environmental safety characteristics and the actual structural state characteristics respectively; F′ env , F′ str F env 、F str The features obtained after time series weighting; λ1 is F′ env and F′ str Balanced fusion weight; W env 、W str F env 、F str The temporal attention weight matrix of F env (t), F str (t) are the actual environmental safety characteristics and actual structural state characteristics of the t-th time frame; T is the number of time frames; X env 、X str are the input environmental safety data matrix and structural status data matrix respectively; b str 、b env They are the bias items of environmental safety data and structural status data respectively.

[0033] In a second aspect, the present invention provides a deep learning-based dynamic management system for underground powerhouse construction information, which includes a data acquisition module, a data prediction module, and a multi-objective optimization module;

[0034] The data acquisition module is used to acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, wherein the historical data includes construction image data, environmental safety data and structural status data;

[0035] The data prediction module is used to construct a training data set based on historical data, train a construction progress analysis model and a construction safety analysis model based on the training data set, and use the trained construction progress analysis model and construction safety analysis model to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site respectively;

[0036] The multi-objective optimization module is used to construct a constrained multi-objective optimization model for the underground powerhouse construction site. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress and the construction safety risk score. A heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site.

[0037] The expression of the comprehensive objective function of the multi-objective optimization model is:

[0038] Γ=α3·f prog +β3·f res +γ3·f risk ;

[0039]

[0040] In the above formula, Γ represents the comprehensive objective function of the multi-objective optimization model; f prog 、f res 、f risk They represent the construction progress difference objective function, construction resource consumption objective function, and construction safety risk objective function respectively; α3, β3, and γ3 represent the relationship between f prog 、f res 、f risk The corresponding weight coefficient is α3+β3+γ3=1; Respectively Actual construction progress and planned construction progress at all times; is the total adjustment period; δ mat , δ lab Respectively The corresponding weight coefficient; Respectively The amount of construction materials consumed and the number of construction workers at each moment; express

[0041] The construction safety risk score at the moment; Ω(·) is the safety risk mapping quantization function; η2, η3, and η4 are the third, fourth, and fifth anti-zero constants, respectively.

[0042] The construction progress analysis model is used to perform multi-scale feature extraction and scene segmentation on construction image data under a certain construction resource allocation scheme and construction progress adjustment strategy, obtain actual hierarchical scene features, calculate the difference between the actual hierarchical scene features and the hierarchical scene features under the baseline state, and weight the difference to calculate the actual construction progress.

[0043] The calculation formula for the actual construction progress is:

[0044]

[0045] In the above formula, P current is the actual construction progress; M2 is the scene feature dimension of the construction image data; γ2 is the progress magnification coefficient; is the m-th dimension scene feature in the actual hierarchical scene feature; is the m-th dimension scene feature in the hierarchical scene feature under the baseline state; ε is the first anti-zero constant.

[0046] The training objective function of the construction progress analysis model is:

[0047]

[0048] F i =ReLU(Conv i (X)+b i );

[0049] In the above formula, L total represents the training objective function of the scene analysis model; α1 and β1 are the classification weight and regression weight respectively; γ1 is the attention weighting coefficient; N1 is the total number of samples; M1 is the number of scales used; Predict the value y for the kth sample k and the true value Cross entropy classification loss; RMSE (F i ,F fused ) is the i-th scale feature F i and fusion feature F fused The root mean square error between loss is the temporal consistency loss based on the image frame sequence; A i is the attention weight of the i-th scale feature; W i is the weight matrix of the i-th scale feature; exp(·) is the exponential function; X is the input construction image data matrix; b i is the bias of the i-th scale feature; Conv i is the convolution operation of the convolution kernel of the i-th scale feature.

[0050] The construction safety analysis model is used to first use a two-stream convolutional network to extract actual environmental safety features and actual structural state features from environmental safety data and structural state data under a certain construction resource allocation plan and construction progress adjustment strategy, respectively. The model then uses a temporal attention mechanism to calculate a fusion weighted feature of the actual environmental safety features and the actual structural state features, and calculates an anomaly detection score based on the fusion weighted feature.

[0051] Then, a feature extraction algorithm is used to extract the actual environmental safety feature vector and the actual structural state feature vector from the environmental safety data and structural state data under a certain construction resource allocation plan and construction progress adjustment strategy. The differences between the actual environmental safety feature vector and the environmental safety feature vector under the baseline state, as well as the differences between the actual structural state feature vector and the structural state characteristics under the baseline state are calculated. By weighting the differences and combining them with the anomaly detection score, the construction safety risk score is calculated.

[0052] The expression of the construction safety risk score is:

[0053]

[0054] In the above formula, R risk Score the construction safety risk; E current 、E baseline are the actual environmental safety feature vector and the environmental safety feature vector under the reference state respectively; S current 、S baseline are the actual structural state eigenvector and the structural state eigenvector under the reference state, respectively; ||·||2 is the Euclidean norm; α2 and β2 are the weighting coefficients of environmental safety data and structural state data, respectively; η1 is the second zero-proof constant; κ is the normalization coefficient of construction safety risk score; S anomaly Scoring anomaly detection;

[0055] The expression of the anomaly detection score is:

[0056]

[0057] F env =ReLU(Conv env (X env )+b env );

[0058]

[0059] F str =ReLU(Conv str (X str )+b str );

[0060] In the above formula, ζ is the anomaly The corresponding weight coefficient; N2 is the feature dimension after fusion weighting; represents the fusion weighted features; is the lth element in the fusion weighted feature; μ l , σ l are the mean and standard deviation of the lth element in the fusion weighted feature respectively; δ is the time decay factor; Δt is the difference between the current time and the reference time; F env 、F str are the actual environmental safety characteristics and the actual structural state characteristics respectively; F′ env , F′ str F env 、F str The features obtained after time series weighting; λ1 is F′ env and F′ str Balanced fusion weight; Wenv 、W str F env 、F str The temporal attention weight matrix of F env (t), F str (t) are the actual environmental safety characteristics and actual structural state characteristics of the t-th time frame; T is the number of time frames; X env 、X str are the input environmental safety data matrix and structural status data matrix respectively; b str 、b env They are the bias items of environmental safety data and structural status data respectively.

[0061] In the third aspect, the present invention provides a dynamic management device for underground plant construction information based on deep learning, the dynamic management device for underground plant construction information includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the aforementioned dynamic management method for underground plant construction information according to the instructions in the computer program code.

[0062] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the aforementioned method for dynamic management of underground plant construction information is implemented.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The deep learning-based dynamic management method for underground powerhouse construction information described in the present invention, on the one hand, obtains historical data including construction image data, environmental safety data and structural status data to construct a construction progress analysis model and a construction safety analysis model, uses the construction progress analysis model to accurately identify the actual construction progress, and uses the construction safety analysis model to evaluate the construction safety risk, thereby realizing the dynamic prediction of the actual construction progress and safety risk; on the other hand, based on the difference between the actual construction progress and the planned construction progress predicted in real time, and the construction safety risk predicted in real time, the optimal construction resource allocation plan and construction progress adjustment strategy are generated through a multi-objective optimization algorithm, thereby improving the rationality of resource allocation and the real-time nature of decision-making.

[0065] 2. The deep learning-based dynamic management method for underground powerhouse construction information described in the present invention constructs a comprehensive objective function of multi-task fusion in the multi-objective optimization model, so that the model can take into account both classification and regression accuracy, improve the recognition accuracy of important targets on the construction site, and introduce temporal consistency loss to ensure the stability and continuity of the model output, reduce the model's excessive sensitivity to accidental noise or outliers, and improve the prediction accuracy of time-varying factors such as construction progress and safety risks. It also introduces an adaptive weight adjustment mechanism, which can dynamically adjust the weight coefficients of each objective function according to the real-time status of the construction site, effectively dealing with the complex coupling relationship among construction progress, resource allocation and safety risks, thereby improving the utilization efficiency of underground powerhouse construction resources.

[0066] 3. The deep learning-based dynamic management method for underground powerhouse construction information described in the present invention, in the construction progress analysis model, objectively quantifies the current construction progress by extracting multi-scale features and tracking temporal changes in the complex construction scenes of the underground powerhouse. In the construction safety analysis model, it integrates environmental parameter data and structural status data to establish a two-dimensional safety risk assessment system, integrates and weights the environmental parameter data and structural status data, and compares them with historical benchmark values, and finally outputs a construction safety risk score, thereby improving the safety management and control level of underground powerhouse construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Flowchart of the method for dynamic management of underground powerhouse construction information according to the present invention

[0068] Figure 2 This is a structural block diagram of the underground plant construction information dynamic management system described in the present invention.

[0069] Figure 3 This is a structural block diagram of the underground plant construction information dynamic management equipment described in the present invention. DETAILED DESCRIPTION

[0070] The present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings.

[0071] Example 1:

[0072] See also Figure 1 , a dynamic management method for underground powerhouse construction information based on deep learning, is carried out in the following steps:

[0073] Step 1: Acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, wherein the historical data includes construction image data, environmental safety data, and structural status data;

[0074] Specifically, multi-source sensing devices capable of periodic sampling are deployed in the construction-intensive areas of the underground powerhouse construction site to form an underground powerhouse sensor network. The multi-source sensing devices include image acquisition equipment, environmental monitoring equipment, and structural monitoring equipment. The image acquisition equipment is used to collect construction image data at a frequency of 5-30 frames per second, the environmental monitoring equipment is used to collect environmental safety data including temperature, humidity, oxygen concentration, and harmful gas concentration at a frequency of 1-5 minutes, and the structural monitoring equipment is used to collect structural status data including stress and deformation at a frequency of 10-60 minutes. The collected data is preliminarily screened and compressed by the edge computing node and then encrypted and transmitted to the data processing center.

[0075] In a specific embodiment, the image acquisition equipment is deployed at construction sites, channel intersections, and material storage areas; the environmental monitoring equipment is deployed at key nodes of the ventilation system and worker work areas; and the structural monitoring equipment is deployed at support structures, surrounding rock, and groundwater infiltration-sensitive areas. The aforementioned sensing devices form a distributed sensing network through a combination of wired and wireless methods to achieve all-round monitoring of the construction site.

[0076] The present invention sets optimized acquisition frequencies based on the changing characteristics of different sensor data, including 5-30 frames / second for construction image data, 1-5 minutes / time for environmental safety data, and 10-60 minutes / time for structural status data. This ensures data timeliness while avoiding redundant acquisition. A distributed network architecture combining wired and wireless networks addresses the issue of limited underground communication. The introduction of edge computing nodes enables preliminary data screening and compression, reducing transmission bandwidth requirements and central processing burdens.

[0077] Step 2: Construct a training dataset based on historical data, and train a construction progress analysis model and a construction safety analysis model based on the training dataset. The construction progress analysis model and the construction safety analysis model are used to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site, respectively;

[0078] Specifically, the steps for constructing the training data set are as follows: first, geometric correction, noise reduction, contrast enhancement, and resolution unification are performed on the construction image data to obtain standard construction image data; missing value interpolation, numerical smoothing, and scale normalization are performed on the environmental parameter data to obtain standard environmental parameter data; band-pass filtering, trend removal, and statistical normalization are performed on the structural state data to obtain standard structural state data; then, multi-source data alignment is performed on the standard construction image data, standard environmental parameter data, and standard structural state data based on timestamps and spatial identifiers to construct a training data set for multi-source heterogeneous data association mapping;

[0079] In a specific embodiment, a hierarchical spatiotemporal feature fusion algorithm is used to align multi-source data. The specific steps are as follows: first, a temporal index is established based on a global timestamp; then, a spatial reference coordinate system is established to map standard construction image data, standard environmental parameter data, and standard structural status data to a unified spatial model; finally, the spatiotemporal correlation weights among the standard construction image data, standard environmental parameter data, and standard structural status data are calculated to generate a training dataset that is spatiotemporally aligned and has a correlation structure;

[0080] The present invention adopts different preprocessing methods for different types of data. Among them, geometric correction and scene segmentation are used for construction image data, missing value interpolation and smoothing are used for environmental parameter data, and bandpass filtering and trend removal are used for structural status data, thereby improving the quality and consistency of the original data. By introducing a layered spatiotemporal feature fusion algorithm, high-precision spatiotemporal alignment of three types of heterogeneous data, namely image, environment and structure, in the underground powerhouse construction scene is achieved. Based on the global timestamp index and unified spatial coordinate mapping, the correlation structure between data is established, and the spatiotemporal correlation weight is calculated. This not only overcomes the data island problem in traditional construction management, but also can analyze the construction status from different dimensions, improving the model's ability to understand complex scenes on the construction site.

[0081] Specifically, the construction progress analysis model performs multi-scale feature extraction and scene segmentation on construction image data under a certain construction resource allocation plan and construction progress adjustment strategy to identify different construction components and equipment, obtain actual hierarchical scene features, calculate the difference between the actual hierarchical scene features and the hierarchical scene features under the baseline state, and weight the difference to calculate the actual construction progress;

[0082] The calculation formula for the actual construction progress is:

[0083]

[0084] In the above formula, P current is the actual construction progress; M2 is the scene feature dimension of the construction image data; γ2 is the progress magnification coefficient; is the m-th dimension scene feature in the actual hierarchical scene feature; is the m-th dimension scene feature in the hierarchical scene feature under the baseline state; ε is the first anti-zero constant;

[0085] The training objective function of the construction progress analysis model is:

[0086]

[0087]

[0088] F i =ReLU(Convi (X)+b i );

[0089] In the above formula, L total represents the training objective function of the scene analysis model; α1 and β1 are the classification weight and regression weight respectively; γ1 is the attention weighting coefficient; N1 is the total number of samples; M1 is the number of scales used; Predict the value y for the kth sample k and the true value Cross entropy classification loss; RMSE (F i ,F fused ) is the i-th scale feature F i and fusion feature F fused The root mean square error between loss is the temporal consistency loss based on the image frame sequence; A i is the attention weight of the i-th scale feature; W i is the weight matrix of the i-th scale feature; exp(·) is the exponential function; X is the input construction image data matrix; b i is the bias of the i-th scale feature; Conv i is the convolution operation of the convolution kernel of the i-th scale feature;

[0090] On the one hand, the present invention combines classification loss with regression loss, taking into account the accurate characterization of continuous numerical values such as construction component and equipment identification, construction progress, and structural status in the construction scene, and simultaneously optimizes the performance of different tasks in a comprehensive objective function; on the other hand, the mean square error term between multi-scale features and fusion features is introduced to ensure that the features extracted by the model under different receptive fields can effectively participate in training and be paid attention to. Compared with the traditional method of extracting features only at a single scale, it can better capture the heterogeneous information of the underground plant construction site. Multi-scale feature extraction and attention weighting enable the model to remain sensitive to key information in the diverse and dynamic underground plant construction environment, and is not prone to missing local details, while taking into account the overall trend; in addition, temporal consistency loss is introduced to ensure the stability and coherence of the model output, reduce the model's excessive sensitivity to accidental noise or outliers, and improve the accuracy of predicting changes in construction progress over time. In summary, the objective function of the multi-task fusion of the present invention enables the model to take into account both classification and regression accuracy, improve the recognition accuracy of important targets in the construction site (such as equipment, components, personnel distribution, etc.), and improve the prediction accuracy of time series data (such as environment and structural status).

[0091] Specifically, the construction safety analysis model first uses a two-stream convolutional network to extract actual environmental safety features and actual structural status features from environmental safety data and structural status data under a certain construction resource allocation plan and construction progress adjustment strategy. It then uses a temporal attention mechanism to calculate a fused weighted feature of the actual environmental safety features and the actual structural status features. Anomaly detection scores are then calculated based on the fused weighted features.

[0092] Then, a feature extraction algorithm is used to extract the actual environmental safety feature vector and the actual structural state feature vector from the environmental safety data and structural state data under a certain construction resource allocation plan and construction schedule adjustment strategy. The differences between the actual environmental safety feature vector and the environmental safety feature vector under the baseline state, and between the actual structural state feature vector and the structural state characteristics under the baseline state are calculated. By weighting the differences and combining them with the anomaly detection score, a construction safety risk score is calculated.

[0093] In a specific embodiment, the anomaly detection score is compared with its benchmark value. If it is determined that the anomaly detection score does not exceed its benchmark value, the anomaly detection score is retained; otherwise, an alarm is issued. By establishing a comparison mechanism between the anomaly detection score and the benchmark value, the present invention realizes automatic early warning and constructs an evaluation system that comprehensively perceives the construction status from two dimensions: environmental parameters and structural status, thereby improving the early warning capability of underground powerhouse construction safety management.

[0094] Specifically, the expression of the construction safety risk score is:

[0095]

[0096] In the above formula, R risk Score the construction safety risk; E current 、E baseline are the actual environmental safety feature vector and the environmental safety feature vector under the reference state respectively; S current 、S baseline are the actual structural state eigenvector and the structural state eigenvector under the reference state, respectively; ||·||2 is the Euclidean norm; α2 and β2 are the weighting coefficients of environmental safety data and structural state data, respectively; η1 is the second zero-proof constant; κ is the normalization coefficient of construction safety risk score; S anomaly Scoring anomaly detection;

[0097] The expression of the anomaly detection score is:

[0098]

[0099] F env =ReLU(Conv env (Xenv )+b env );

[0100]

[0101] F str =ReLU(Conv str (X str )+b str );

[0102] In the above formula, ζ is the anomaly The corresponding weight coefficient; N2 is the feature dimension after fusion weighting; represents the fusion weighted features; is the lth element in the fusion weighted feature; μ l , σ l are the mean and standard deviation of the lth element in the fusion weighted feature respectively; δ is the time decay factor; Δt is the difference between the current time and the reference time; F env 、F str are the actual environmental safety characteristics and the actual structural state characteristics respectively; F′ env , F′ str F env 、F str The features obtained after time series weighting; λ1 is F′ env and F′ str Balanced fusion weight; W env 、W str F env 、F str The temporal attention weight matrix of F env (t), F str (t) are the actual environmental safety characteristics and actual structural state characteristics of the t-th time frame; T is the number of time frames; X env 、X str are the input environmental safety data matrix and structural status data matrix respectively; b str 、b env They are the bias items of environmental safety data and structural status data respectively;

[0103] Traditional risk assessments often calculate environmental monitoring results and structural safety status separately, making it difficult to timely reflect the coupling between the two. This method, however, extracts environmental safety eigenvectors and structural status eigenvectors separately and integrates them with corresponding weighting coefficients. This method, on the one hand, more comprehensively captures potential hazards during underground powerhouse construction, while not ignoring the impact of environmental factors on structural safety and promptly identifying stress or deformation anomalies within the structure itself. This allows for comprehensive quantification of construction safety risks and completes collaborative "environment-structure" risk monitoring. Furthermore, it highlights more important or risk-prone dimensions, allowing for flexible adjustment of the assessment focus during different construction phases or working conditions. The Euclidean norm is used to measure the magnitude of differences. Compared to simple subtraction or ratio calculations, the Euclidean norm better reflects the overall magnitude of differences in multi-dimensional eigenvectors, reducing the bias in the final results caused by extreme dimensions. Comparison with a baseline state can immediately reflect the degree of deviation from the normal state of the construction site. Furthermore, the addition of a zero-proof constant prevents division by zero anomalies during normalization or difference calculations, thereby enhancing robustness.

[0104] When calculating the construction safety risk score, the score is normalized using a normalization coefficient so that the risk scores obtained at different stages and under different monitoring conditions can be uniformly compared. The mechanism of temporal attention and time attenuation factor is also introduced. Combined with the time span between the current moment and the historical benchmark, it can more effectively distinguish between short-term sharp fluctuations and long-term slow-changing trends, thereby maintaining sensitivity to both sudden risks and slowly cumulative risks. The Euclidean norm is used to measure the difference between the current actual value of the environment / structure and the historical benchmark value. Assisted by temporal attention and weighting mechanisms, it is possible to more finely characterize the accumulation and mutation process of risks over time in real-time data streams, thereby reducing interference caused by data discreteness or accidental fluctuations.

[0105] The calculated construction safety risk score has temporal consistency and can be used as an input parameter in the multi-objective optimization model, thereby improving the comprehensive decision-making effect of construction progress management, safety monitoring, and resource scheduling, and realizing refined control of the construction process. According to the actual needs of the project, the construction cycle can be divided into multiple time steps by day or hour, or divided into stages according to key construction nodes. As construction continues, the corresponding current construction progress and construction safety risk score can be obtained at each time step. The current construction progress and construction safety risk score of each time step in the entire construction cycle can be expressed as:

[0106]

[0107]

[0108] Step 3: Construct a constrained multi-objective optimization model for the underground powerhouse construction site. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress and the construction safety risk score. A heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site.

[0109] Specifically, the constraints of the multi-objective optimization model include the priority requirements of the construction period, the allocation quotas of various resources, and the construction safety risk score threshold. The expression of the comprehensive objective function of the multi-objective optimization model is:

[0110] Γ=α3·f prog +β3·f res +γ3·f risk ;

[0111]

[0112] In the above formula, Γ represents the comprehensive objective function of the multi-objective optimization model; f prog 、f res 、f risk They represent the construction progress difference objective function, construction resource consumption objective function, and construction safety risk objective function respectively; α3, β3, and γ3 represent the relationship between f prog 、f res 、f risk The corresponding weight coefficient is α3+β3+γ3=1; Respectively Actual construction progress and planned construction progress at all times; is the total adjustment period; δ mat , δ lab Respectively The corresponding weight coefficient; Respectively The amount of construction materials consumed and the number of construction workers at each moment; express

[0113] The construction safety risk score at the moment; Ω(·) is the safety risk mapping quantization function; η2, η3, η4 are the third, fourth, and fifth anti-zero constants respectively;

[0114] In a specific embodiment, a heuristic search algorithm combining a genetic algorithm and a particle swarm algorithm is used to iteratively solve a multi-objective optimization model. In each iteration, the comprehensive objective function value of each candidate solution is calculated, and the optimal solution and several candidate solutions are retained for the next generation until the comprehensive objective function of the multi-objective optimization model converges. During the iteration process, the weight coefficients and search directions in the comprehensive objective function of the multi-objective optimization model are adjusted according to the convergence results to guide the model to continue iterative optimization and better cope with sudden changes or local optimal dilemmas. When the convergence conditions are met, the optimal construction resource allocation plan and construction progress adjustment strategy are output, and the results are pushed to the management personnel in real time, realizing dynamic monitoring of construction progress and safety risks, improving the rationality of resource allocation and the efficiency and real-time decision-making of underground powerhouse construction. The convergence condition is: the fluctuation of the comprehensive objective function value is less than a preset threshold in multiple consecutive iterations, which indicates that the construction progress difference objective function, the construction resource consumption objective function, and the construction safety risk objective function are all tending to be stable, and there is no situation where a certain dimension continues to deteriorate.

[0115] In a specific embodiment, the weight coefficients in the comprehensive objective function are dynamically adjusted according to the construction stage; for example, in the excavation stage, by increasing the weight coefficient of construction safety risk, more attention is paid to construction safety risk, while in the support construction stage, by increasing the weight coefficient of construction resource consumption, more attention is paid to construction resource consumption, thereby achieving flexible optimization in stages, effectively avoiding the extreme optimization of a single goal, the situation of blindly shortening the construction period while ignoring safety risks or excessive consumption of resources, and being conducive to obtaining an optimal solution that is more in line with actual construction needs and more executable.

[0116] Performance Verification:

[0117] The following is an example of the excavation, support, and secondary lining construction of an underground powerhouse at a large hydropower station. The construction of this underground powerhouse has the following requirements:

[0118] 1. Tight construction schedule: Excavation and initial support of the main caverns needed to be completed within the scheduled 18 months;

[0119] 2. Limited resources: The amount of materials (rebar, concrete, anchors, shotcrete, etc.) and equipment (rock drills, concrete pumps, ventilators, etc.) is limited; labor also needs to be reasonably distributed among different types of workers (blasters, support workers, welders, etc.);

[0120] 3. Safety risk control: Areas with unstable surrounding rock and areas sensitive to groundwater infiltration need to avoid risks of collapse and water inrush;

[0121] (1) Construction data:

[0122] Construction data includes historical and current construction data. Historical construction data includes historical construction progress and historical construction safety risk scores. Current construction data includes current construction progress indicators and current construction safety risk scores. The current project is 40% complete, which is approximately 10% behind schedule. The construction safety risk score is 0.68, with higher values in the [0, 1] range indicating higher risk.

[0123] (2) Model constraints:

[0124] Construction period priority: Complete the remaining 60% of the main excavation and initial support work within 6 months;

[0125] Construction progress constraints: The overall progress gap is reduced to within 5% within 6 months;

[0126] Resource constraints: Equipment constraints: 3 rock drills and 2 concrete pump trucks can be used daily, and a maximum of 50 workers can be employed each working day (divided into two shifts, morning and evening); Material constraints: 50 tons of steel bars and 200m3 of concrete 3 , 3,000 anchor rods;

[0127] Construction safety risk score threshold constraint: The construction safety risk score must not exceed the threshold of 0.75 for two consecutive months;

[0128] (3) Model parameter setting

[0129] During the peak excavation phase, more emphasis should be placed on safety and efficiency, so α3, β3, and γ3 are set to 0.45, 0.2, and 0.35, respectively. During the support and secondary lining phase, resource consumption needs to be optimized while ensuring safety, so α3, β3, and γ3 are set to 0.55, 0.3, and 0.15, respectively.

[0130] (4) Algorithm settings:

[0131] A heuristic search strategy combining genetic algorithm and particle swarm optimization was used, with a population size of 100, a maximum number of iterations of 50, a crossover probability of 0.7, and a mutation probability of 0.1. The algorithm's initial solution was as follows: Construction resource allocation plan: 3 rock drills, 2 concrete pump trucks, and 3 ventilators per month, 50 workers, and 20 tons of support materials per month; Construction schedule adjustment strategy: The first three months are the peak excavation period, and the last three months are the support and secondary lining period.

[0132] If the construction safety risk score is too high (over 0.75) in a certain iteration, the weight coefficient of the construction safety risk objective function will be increased. If equipment or material usage exceeds the limit, the weight coefficient of the construction resource consumption objective function will be reduced. If the progress is still behind schedule, the weight coefficient of the construction progress difference objective function will be increased, and a moderate increase in equipment and labor input will be allowed.

[0133] (5) Output the optimal solution;

[0134] A stable solution was found at the 40th iteration (the comprehensive objective function value fluctuated by less than 2% within 5 consecutive iterations). The solution obtained in this iteration was output as the optimal solution, and the optimal construction resource allocation plan and the optimal construction progress adjustment strategy were obtained.

[0135] The optimal construction resource allocation scheme and optimal construction schedule adjustment strategy for the underground powerhouse construction site are obtained as follows:

[0136] Construction resource allocation plan:

[0137] (1) Mechanical equipment configuration:

[0138] Rock drill configuration:

[0139] In the first month, two rock drills were put into operation, focusing on preliminary excavation in Area A, with a daily operation time of 12 hours. In the second month, two rock drills were maintained, and the operation was expanded to areas A and B in parallel, with the daily operation time extended to 18 hours. In the third month, all three rock drills were put into operation, achieving simultaneous excavation in areas A, B, and C, with 24-hour shifts and uninterrupted operation. In the fourth month, all three rock drills were kept in operation, but the focus of work shifted from excavation to auxiliary support hole construction. In the fifth month, the number of operating units was reduced to two, mainly used for finishing excavation and support hole drilling in Area C. In the sixth month, only one rock drill was kept in operation, focusing on local finishing and special geological treatment.

[0140] Concrete pump truck configuration: One concrete pump truck will be deployed in the first month, mainly for experimental support and initial roadway reinforcement; in the second and third months, the number will be increased to two, one fixed in area A for post-excavation support, and the other for mobile deployment between areas B and C; in the fourth to sixth months, two concrete pump trucks will be deployed, one dedicated to support spraying, and the other to start the initial secondary lining operation;

[0141] Ventilation equipment configuration: In the first month, two ventilators were deployed, primarily to ensure air quality in the main passages and initial excavation face. In the second and third months, the number was increased to three, covering all working faces, with a focus on improving smoke exhaust efficiency after blasting. In the fourth month, the ventilators were repositioned to prioritize air quality in the support operation area. In the fifth month, two high-brightness lighting devices were added for fine work in the secondary lining construction area. In the sixth month, an automatic spraying robot was introduced to improve the accuracy and efficiency of secondary lining construction.

[0142] Equipment dynamic adjustment strategy: Equipment rotation system, each rock drill is scheduled for one day of maintenance per week and replaced by spare equipment; the working intensity and drilling layout of the rock drill are adjusted in time according to changes in geological conditions of the working face; monthly equipment efficiency evaluation is enabled, and equipment with efficiency below 80% is repaired or replaced.

[0143] (2) Labor force allocation:

[0144] In the first month, each shift has 20 people (two shifts), with the job ratio being 3 blasters, 3 welders, 6 support workers, 3 mechanics, 2 inspectors, and 3 assistants;

[0145] In the second month, the number of people per shift was 22, which was adjusted to 4 blasters, 4 welders, 7 support workers, 3 mechanics, 2 inspectors, and 2 assistants;

[0146] In the third month, the number of people per shift increased to 25, including 5 blasters, 5 welders, 8 support workers, 4 mechanics, 2 inspectors, and 1 assistant;

[0147] During the first three months, a "high-efficiency shift rotation" system was implemented, with a 30-minute handover period to ensure information flow. At the end of the third month, blasters received a week-long training on the basics of support and concrete construction.

[0148] In the fourth month, the number of workers per shift remained at 25, but the types of workers were adjusted to 3 blasters, 5 welders, 9 support workers, 3 mechanics, 2 inspectors, 2 concrete workers, and 1 auxiliary worker.

[0149] In the fifth month, the number of people per shift was maintained at 25, and was adjusted to 2 blasters, 6 welders, 9 support workers, 2 mechanics, 2 inspectors, and 4 concrete workers;

[0150] In the 6th month, each shift has 25 people, and the full-time blaster is eliminated. There are 8 support workers, 6 welders, 1 mechanic, 2 inspectors, 6 concrete workers, and 2 secondary lining professional workers.

[0151] Starting from the fourth month, some blasters will receive a two-week training on concrete construction skills, as follows:

[0152] Month 4: Maintain 25 workers per shift, but adjust the types of work to 3 blasters, 5 welders, 9 support workers, 3 mechanics, 2 inspectors, 2 concrete workers, and 1 auxiliary worker;

[0153] Month 5: Maintaining 25 people per shift, adjusted to 2 blasters, 6 welders, 9 support workers, 2 mechanics, 2 inspectors, and 4 concrete workers.

[0154] The 6th month: 25 people per shift, the full-time blaster was eliminated, 8 support workers, 6 welders, 1 mechanic, 2 inspectors, 6 concrete workers, and 2 secondary lining professional workers were employed.

[0155] (3)Material supply:

[0156] Rebar allocation: In the first month, the monthly consumption is 25 tons, 65% of which is allocated to the main channel support and 35% is used for the excavation section of Area A; in the second month, it increases to 35 tons, with the allocation ratio being 50% for Area A, 30% for Area B, and 20% for the main channel; in the third month, it reaches 45 tons, with the allocation ratio being 40% for Area A, 35% for Area B, and 15% for Area C, with 10% reserved for emergency use; in the fourth month, it remains at 45 tons, but the use is adjusted to 60% for support steel bars and 40% for secondary lining steel mesh; in the fifth month, it remains at 45 tons, with the ratio adjusted to 40% for support steel bars and 60% for secondary lining steel mesh; in the sixth month, it increases to 48 tons, with the support steel bars reduced to 20%, the secondary lining steel mesh increased to 80%, and the configuration density in the seepage section increased by 15%;

[0157] Concrete supply and distribution: Month 1, monthly consumption 120m 3 , shotcrete accounted for 85%, structural concrete accounted for 15%; in the second month, increased to 150m 3 , shotcrete accounts for 75%, distributed in the ratio of 3:2 between A and B areas; in the third month, it reaches 180m 3 The allocation ratio of A, B and C is 5:3:2; in the fourth month, it increases to 190m 3 , the ratio of shotcrete to structural concrete was adjusted to 6:4; in the fifth month, it reached 200m 3 , the proportion of shotcrete dropped to 40%, and the proportion of structural concrete increased to 60%; in the 6th month, it was maintained at 200m 3 , but further adjusted the proportions to 25% shotcrete and 75% structural concrete;

[0158] Distribution of anchor rods and other special materials: In the first month, 2,000 anchor rods were used, mainly concentrated in the roof and high-stress areas; in the second month, the number of anchor rods increased to 2,500, and the density began to be differentiated according to the surrounding rock grade; in the third month, the number of anchor rods reached 2,800, and the density of anchor rods in the special surrounding rock fracture zone was increased by 20%; in the fourth month, the number of anchor rods remained at 2,800, and the density was increased by 15m 3 Special grouting materials were used for water seepage control; in the fifth month, the number of anchor bolts was reduced to 2,500 and the amount of grouting materials increased to 20m 3 , used for reinforcement of water seepage areas; in the sixth month, the number of anchor rods was further reduced to 2,200, and the grouting material was maintained at 20m 3 , and strengthen treatment of key areas.

[0159] Construction progress adjustment strategy:

[0160] (1) First to third month (peak excavation period):

[0161] Adjustment of process priority: parallel excavation of working faces, with priority given to ensuring that working faces A and B proceed simultaneously, and follow-up excavation of working face C is arranged according to geological conditions;

[0162] Adjustments to shifts and work content: Night shifts for important processes will be strengthened, with the main types of work concentrated on blasting, ventilation, and slag removal to narrow the gap with the planned progress as quickly as possible;

[0163] Increase the frequency of safety inspections: Increase the number of inspections per shift from 2 to 3 times, with a focus on monitoring surrounding rock cracks and groundwater infiltration points.

[0164] (2) The fourth to sixth month (transition period between support and secondary lining):

[0165] Logical arrangement of process: simultaneous support and local closure. After the excavation of A and B working faces is basically completed, rapid support will be started immediately. For local areas that have been excavated but not yet closed, secondary lining operations will be carried out first.

[0166] Phased pouring plan: Schedule concrete pouring windows every week to ensure smooth connection between shotcrete and steel mesh installation;

[0167] Adjustment of material supply gradient: Place material orders with suppliers one week in advance to ensure that the support progress is not affected and the budget is not exceeded.

[0168] (3) Safety linkage and temporary emergency response:

[0169] Real-time risk assessment threshold: If the construction safety risk score is not less than 0.75 in any week, the safety weight coefficient will be immediately increased and the number of inspections will be dynamically increased; if necessary, some processes will be suspended and the construction period will be extended to ensure safety;

[0170] Surrounding rock monitoring and grouting measures: Once surrounding rock displacement or abnormal increase in seepage pressure is detected, timely grouting reinforcement or additional anchor rods are carried out, and support resources are invested first to avoid accidents.

[0171] After applying the underground powerhouse construction information dynamic management method described in the present invention, by focusing on the peak excavation intensity in the first three months, it is expected that the progress difference from the original plan of 10% can be reduced to the range of 3-5%. In addition, a comprehensive balance is taken between personnel scheduling, machinery and equipment, and material supply to avoid waste of resources or idle manpower and machines caused by disorderly scheduling. Dynamic inspections and real-time construction safety risk score responses are set up, so that the safety weight can be timely increased in high-risk stages to reduce the probability of collapse and water inrush accidents.

[0172] Example 2:

[0173] See also Figure 2A deep learning-based dynamic management system for underground powerhouse construction information includes a data acquisition module, a data prediction module, and a multi-objective optimization module. The data acquisition module is used to acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, and the historical data includes construction image data, environmental safety data, and structural status data. The data prediction module is used to construct a training data set based on the historical data, train a construction progress analysis model and a construction safety analysis model based on the training data set, and use the trained construction progress analysis model and construction safety analysis model to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site respectively. The construction progress analysis model is used to perform multi-scale feature extraction and scene segmentation on the construction image data under a certain construction resource allocation scheme and construction progress adjustment strategy to obtain actual hierarchical scene features, calculate the difference between the actual hierarchical scene features and the hierarchical scene features under the baseline state, and weight the difference to calculate the actual construction progress. The calculation formula for the actual construction progress is:

[0174]

[0175] In the above formula, P current is the actual construction progress; M2 is the scene feature dimension of the construction image data; γ2 is the progress magnification coefficient; is the m-th dimension scene feature in the actual hierarchical scene feature; is the m-th dimension scene feature in the hierarchical scene feature under the baseline state; ε is the first anti-zero constant;

[0176] The training objective function of the construction progress analysis model is:

[0177]

[0178] F i =ReLU(Conv i (X)+b i );

[0179] In the above formula, L total represents the training objective function of the scene analysis model; α1 and β1 are the classification weight and regression weight respectively; γ1 is the attention weighting coefficient; N1 is the total number of samples; M1 is the number of scales used; Predict the value y for the kth sample k and the true value Cross entropy classification loss; RMSE (F i ,F fused ) is the i-th scale feature F i and fusion feature F fused The root mean square error between lossA is the temporal consistency loss based on the image frame sequence; i is the attention weight of the i-th scale feature; W i is the weight matrix of the i-th scale feature; exp(·) is the exponential function; X is the input construction image data matrix; b i is the bias of the i-th scale feature; Conv i is the convolution operation of the convolution kernel of the i-th scale feature.

[0180] The construction safety analysis model is used to first use a two-stream convolutional network to extract actual environmental safety features and actual structural state features from environmental safety data and structural state data under a certain construction resource allocation scheme and construction progress adjustment strategy, and calculate a fusion weighted feature of the actual environmental safety features and the actual structural state features in combination with a temporal attention mechanism. An anomaly detection score is calculated based on the fusion weighted feature; then, a feature extraction algorithm is used to extract actual environmental safety feature vectors and actual structural state feature vectors from environmental safety data and structural state data under a certain construction resource allocation scheme and construction progress adjustment strategy, and the difference between the actual environmental safety feature vector and the environmental safety feature vector under a reference state, and between the actual structural state feature vector and the structural state feature under a reference state is calculated. The construction safety risk score is calculated by weighting the differences and combining them with the anomaly detection score. The expression of the construction safety risk score is:

[0181]

[0182] In the above formula, R risk Score the construction safety risk; E current 、E baseline are the actual environmental safety feature vector and the environmental safety feature vector under the reference state respectively; S current 、S baseline are the actual structural state eigenvector and the structural state eigenvector under the reference state, respectively; ||·||2 is the Euclidean norm; α2 and β2 are the weighting coefficients of environmental safety data and structural state data, respectively; η1 is the second zero-proof constant; κ is the normalization coefficient of construction safety risk score; S anomaly Scoring anomaly detection;

[0183] The expression of the anomaly detection score is:

[0184]

[0185] F env =ReLU(Conv env (X env )+b env );

[0186]

[0187] F str =ReLU(Conv str (X str )+b str );

[0188] In the above formula, ζ is the anomaly The corresponding weight coefficient; N2 is the feature dimension after fusion weighting; represents the fusion weighted features; is the lth element in the fusion weighted feature; μ l , σ l are the mean and standard deviation of the lth element in the fusion weighted feature respectively; δ is the time decay factor; Δt is the difference between the current time and the reference time; F env 、F str are the actual environmental safety characteristics and the actual structural state characteristics respectively; F′ env , F′ str F env 、F str The features obtained after time series weighting; λ1 is F′ env and F′ str Balanced fusion weight; W env 、W str F env 、F str The temporal attention weight matrix of F env (t), F str (t) are the actual environmental safety characteristics and actual structural state characteristics of the t-th time frame; T is the number of time frames; X env 、X str are the input environmental safety data matrix and structural status data matrix respectively; b str 、b env They are the bias items of environmental safety data and structural status data respectively;

[0189] The multi-objective optimization module is used to construct a constrained multi-objective optimization model for the underground powerhouse construction site. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress and the construction safety risk score. The heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site. The expression of the comprehensive objective function of the multi-objective optimization model is:

[0190] Γ=α3·f prog +β3·f res +γ3·f risk ;

[0191]

[0192] In the above formula, Γ represents the comprehensive objective function of the multi-objective optimization model; f prog 、f res 、f risk They represent the construction progress difference objective function, construction resource consumption objective function, and construction safety risk objective function respectively; α3, β3, and γ3 represent the relationship between f prog 、f res 、f risk The corresponding weight coefficient is α3+β3+γ3=1; Respectively Actual construction progress and planned construction progress at all times; is the total adjustment period; δ mat , δ lab Respectively The corresponding weight coefficient; Respectively The amount of construction materials consumed and the number of construction workers at each moment; express The construction safety risk score at the moment; Ω(·) is the safety risk mapping quantization function; η2, η3, and η4 are the third, fourth, and fifth anti-zero constants, respectively.

[0193] Example 3:

[0194] See also Figure 3 , a deep learning-based dynamic management device for underground plant construction information, comprising a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the underground plant construction information dynamic management method as described in Example 1 according to the instructions in the computer program code.

[0195] Example 4:

[0196] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for dynamic management of underground plant construction information as described in Example 1.

[0197] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or a combination of software and hardware embodiments. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0198] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0199] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0200] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based dynamic management method for underground powerhouse construction information, characterized by: The underground powerhouse construction information dynamic management method includes: Acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, wherein the historical data includes construction image data, environmental safety data, and structural status data; Building a training data set based on historical data, and training a construction progress analysis model and a construction safety analysis model based on the training data set. The construction progress analysis model and the construction safety analysis model are used to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site, respectively; A constrained multi-objective optimization model for the underground powerhouse construction site is constructed. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress, and the construction safety risk score. A heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site.

2. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 1 is characterized by: The expression of the comprehensive objective function of the multi-objective optimization model is: C=α3·f prog +β3·f res +γ3·f risk ; In the above formula, Γ represents the comprehensive objective function of the multi-objective optimization model; fprog, f res 、f risk They represent the construction progress difference objective function, construction resource consumption objective function, and construction safety risk objective function respectively; α3, β3, and γ3 represent the relationship between fprog and f res 、f risk The corresponding weight coefficient is α3+β3+γ3=1; Respectively Actual construction progress and planned construction progress at all times; is the total adjustment period; δ mat , δ lab Respectively The corresponding weight coefficient; Respectively The amount of construction materials consumed and the number of construction workers at each moment; express The construction safety risk score at the moment; Ω(·) is the safety risk mapping quantization function; η2, η3, and η4 are the third, fourth, and fifth anti-zero constants, respectively.

3. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 1 is characterized in that: The construction progress analysis model performs multi-scale feature extraction and scene segmentation on construction image data under a certain construction resource allocation plan and construction progress adjustment strategy to obtain actual hierarchical scene features, calculates the difference between the actual hierarchical scene features and the hierarchical scene features under a baseline state, and weights the difference to calculate the actual construction progress.

4. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 3 is characterized by: The calculation formula for the actual construction progress is: In the above formula, P current is the actual construction progress; M2 is the scene feature dimension of the construction image data; γ2 is the progress magnification coefficient; is the m-th dimension scene feature in the actual hierarchical scene feature; is the m-th dimension scene feature in the hierarchical scene feature under the baseline state; ε is the first anti-zero constant.

5. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 1 is characterized in that: The training objective function of the construction progress analysis model is: F i =ReLU(Conv i (X)+b i ); In the above formula, L total represents the training objective function of the scene analysis model; α1 and β1 are the classification weight and regression weight respectively; γ1 is the attention weighting coefficient; N1 is the total number of samples; M1 is the number of scales used; Predict the value y for the kth sample k and the true value Cross entropy classification loss; RMSE (F i ,F fused ) is the i-th scale feature F i and fusion feature F fused The root mean square error between loss A is the temporal consistency loss based on the image frame sequence; i is the attention weight of the i-th scale feature; W i is the weight matrix of the i-th scale feature; exp(·) is the exponential function; X is the input construction image data matrix; b i is the bias of the i-th scale feature; Conv i is the convolution operation of the convolution kernel of the i-th scale feature.

6. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 1, characterized in that: The construction safety analysis model first uses a two-stream convolutional network to extract actual environmental safety features and actual structural status features from environmental safety data and structural status data under a certain construction resource allocation plan and construction progress adjustment strategy. It then uses a temporal attention mechanism to calculate a fused weighted feature of the actual environmental safety features and the actual structural status features. Anomaly detection scores are calculated based on the fused weighted features. Then, a feature extraction algorithm is used to extract the actual environmental safety feature vector and the actual structural state feature vector from the environmental safety data and structural state data under a certain construction resource allocation plan and construction progress adjustment strategy. The differences between the actual environmental safety feature vector and the environmental safety feature vector under the baseline state, as well as the differences between the actual structural state feature vector and the structural state characteristics under the baseline state are calculated. By weighting the differences and combining them with the anomaly detection score, the construction safety risk score is calculated.

7. The method for dynamic management of underground powerhouse construction information based on deep learning according to claim 1, characterized in that: The expression of the construction safety risk score is: In the above formula, R risk Score the construction safety risk; E current 、E baseline are the actual environmental safety feature vector and the environmental safety feature vector under the reference state respectively; S current 、S baseline are the actual structural state eigenvector and the structural state eigenvector under the reference state, respectively; ||·||2 is the Euclidean norm; α2 and β2 are the weighting coefficients of environmental safety data and structural state data, respectively; η1 is the second zero-proof constant; κ is the normalization coefficient of construction safety risk score; S anomaly Scoring anomaly detection; The expression of the anomaly detection score is: F env =ReLU(Conv env (X env )+b env ); F str =ReLU(Conv str (X str )+b str ); In the above formula, ζ is the anomaly The corresponding weight coefficient; N2 is the feature dimension after fusion weighting; represents the fusion weighted features; is the lth element in the fusion weighted feature; μ l , σ l are the mean and standard deviation of the lth element in the fusion weighted feature respectively; δ is the time decay factor; Δt is the difference between the current time and the reference time; F env 、F str They are the actual environmental safety characteristics and the actual structural status characteristics; F' env 、F' str F env 、F str The features obtained after time series weighting; λ1 is F' env With F' str Balanced fusion weight; W env 、W str F env 、F str The temporal attention weight matrix of F env (t), F str (t) are the actual environmental safety characteristics and actual structural state characteristics of the t-th time frame respectively; T is the number of time frames; X env 、X str are the input environmental safety data matrix and structural status data matrix respectively; b str 、b env They are the bias items of environmental safety data and structural status data respectively.

8. A dynamic management system for underground powerhouse construction information based on deep learning, characterized by: The underground powerhouse construction information dynamic management system includes a data acquisition module, a data prediction module, and a multi-objective optimization module; The data acquisition module is used to acquire historical data of the underground powerhouse construction site under different construction resource allocation schemes and different construction progress adjustment strategies, wherein the historical data includes construction image data, environmental safety data and structural status data; The data prediction module is used to construct a training data set based on historical data, train a construction progress analysis model and a construction safety analysis model based on the training data set, and use the trained construction progress analysis model and construction safety analysis model to predict the actual construction progress and construction safety risk score of the underground powerhouse construction site respectively; The multi-objective optimization module is used to construct a constrained multi-objective optimization model for the underground powerhouse construction site. The comprehensive objective function of the multi-objective optimization model is established based on the difference between the actual construction progress and the planned construction progress and the construction safety risk score. A heuristic search strategy is used to iteratively solve the multi-objective optimization model to obtain the optimal construction resource allocation plan and the optimal construction progress adjustment strategy for the underground powerhouse construction site.

9. Deep learning-based dynamic management equipment for underground powerhouse construction information, characterized by: The underground plant construction information dynamic management device includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the underground plant construction information dynamic management method as described in claims 1-7 according to the instructions in the computer program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the method for dynamically managing underground powerhouse construction information as described in claims 1-7 is implemented.

Images