Water conservancy and hydropower construction safety question and answer processing optimization method based on time sequence deduction

Abstract

The application discloses a water conservancy and hydropower construction safety question and answer processing optimization method based on time sequence deduction, relates to the technical field of water conservancy projects, and acquires historical working condition matrixes and historical operation matrixes through preprocessing by collecting construction site basic working condition data and historical control logs; receives natural language questions and analyzes intentions, constructs intention operation vectors, iteratively deduces based on a time sequence deduction model combined with historical matrixes, and acquires a predicted state matrix; a constraint bipartite graph is constructed based on the predicted state matrix to detect violations, trace disaster-causing parameters, and generate a prohibited mask matrix; a search tree is constructed based on the prohibited mask matrix, a heuristic pruning search is used to replace a scheme, an optimal alternative construction sequence and compliance degree and safety degree are acquired, and finally, comprehensive utility evaluation is carried out and a decision reply is generated. The application realizes prospective deduction of operation intentions and violation tracing, actively searches for alternative schemes considering safety progress, and thus improves the decision assistance capability of complex construction scenes.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, and in particular to an optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series simulation. Background Technology

[0002] The water conservancy and hydropower engineering construction safety question-and-answer system is an important technical tool to assist in on-site construction safety management. Its core value lies in integrating scattered construction specifications, historical experience, and real-time monitoring data into an interactive knowledge service, helping on-site managers quickly obtain safety decision support information in complex construction environments. As water conservancy and hydropower projects develop towards larger scale and greater complexity, the number of safety specifications involved in construction sites is increasing, and the coupling and constraint relationships between operational parameters are becoming increasingly tight. Traditional methods that rely on manual review of specifications or personal experience for safety judgment are no longer sufficient to meet the dual requirements of modern construction management for response speed and decision accuracy. Chinese patent application CN119396971A discloses a water conservancy engineering knowledge question-and-answer construction method and system based on a large model. It loads water conservancy engineering data from multiple formats and data sources and performs text segmentation processing. It uses an open-source embedding model to convert the text into vector form and stores it in the Milvus vector library to realize the storage and management of document vectorized data. In the question-and-answer application stage, after the user's query question is converted into vector form, semantic matching is performed in the vector database to retrieve relevant answers, and the large model is used to summarize the retrieval results to generate the final answer. The underlying logic of the knowledge question-answering system in the aforementioned existing technical solutions is still based on the retrieval and reorganization of existing static documents. Once the content in the knowledge base is entered, it becomes fixed information and cannot be used to extrapolate and analyze construction scenarios that have not yet occurred or are beyond the scope of the documents. In other words, its answers are essentially a reproduction and summary of historical knowledge or standard texts, and it does not have the ability to predict the safety of construction operation plans in the future time dimension. The retrieval and recall mechanism of the existing technology relies on the similarity matching of semantic vectors. For complex construction intentions proposed by users involving the coordinated adjustment of multiple operation parameters, it is difficult to capture the coupling effect between various operation variables and their combined impact on the construction safety status. When the operation plan of the user's intention has the risk of violating the regulations, the solution can only provide relevant clauses or cases already existing in the documents, but cannot locate the key operation parameters that cause the violation, let alone automatically search for and provide alternative operation plans that meet the regulatory constraints. This greatly limits its practical value in assisting on-site construction decision-making. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies are difficult to meet the actual needs of water conservancy and hydropower engineering construction scenarios for forward-looking safety simulation of construction operation plans, accurate tracing of the root causes of violations, and automatic generation of compliant alternative solutions.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: an optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series deduction, comprising the following steps: Step S1: Collect basic working condition data and historical control logs from the construction site, perform normalization preprocessing, and obtain the historical working condition matrix and historical operation matrix. Step S2: Receive natural language question text and perform intent parsing, extract operation parameters to construct intent operation vector, and perform iterative deduction based on the time-series inference model combined with the historical working condition matrix, historical operation matrix and operation vector to obtain the predicted state matrix; Step S3: Construct a constraint bipartite graph based on the predicted state matrix to perform normative violation detection. When a violation is detected, trace the source of the violation, obtain the operation parameters that caused the violation, and generate a prohibition mask matrix based on the operation parameters. Step S4: Construct a search tree based on the forbidden mask matrix, and search for compliant alternatives through heuristic pruning to obtain the optimal alternative construction sequence and its corresponding compliance and safety. Step S5: Perform a comprehensive utility assessment based on the compliance and security levels, and generate a response based on the comparison between the comprehensive utility assessment results and the decision threshold.

[0005] As an optimization, step S1 includes the following sub-steps: Step S101: Collect temperature data, flow rate data and wind speed data as basic operating condition data, and extract discrete data points within a preset time window and align them according to timestamps to form the original operating condition matrix. Step S102: Extract the historical control log within the preset time window, parse the actual execution value of each operation parameter and form a historical operation column vector, and concatenate the historical operation column vectors of each time step in chronological order to form the original operation matrix. Step S103: Based on the preset operation parameter boundary table, normalization preprocessing is performed on each element in the original operating condition matrix and the historical operating matrix to obtain the historical operating condition matrix and the historical operating matrix respectively.

[0006] As an optimization, step S2 includes the following sub-steps: Step S201: Input the question text into the pre-trained named entity recognition network, output entity labels, and parse the structured operation parameters from the text based on the entity labels; Step S202: After normalizing the parsed operation parameters, fill them into the corresponding dimension positions in the operation vector of the preset length to obtain the intention operation vector.

[0007] As an optimization, step S2 further includes the following sub-steps: Step S203: Vertically concatenate each column vector in the historical working condition matrix with the column vector of the corresponding time step in the historical operation matrix to form multiple historical composite state vectors, and input them sequentially into the time series inference model to activate network memory and obtain the initial memory state. The time-series extrapolation model includes an input layer, a stacked long short-term memory network layer, and a fully connected output layer.

[0008] As an optimization, step S2 also includes the following sub-steps: Step S204: Extract the last column vector of the historical working condition matrix and concatenate it with the operation vector to form a first composite state vector; Step S205: Using the time-series deduction model combined with the initial memory state, and taking the first composite state vector and the initial memory state as input, iteratively calculate the state prediction vectors for multiple future time steps, and horizontally concatenate the state prediction vectors of each time step to form the prediction state matrix.

[0009] As an optimization, step S3 includes the following sub-steps: Step S301: Construct a constrained bipartite graph, which includes a left node set, a right node set, and edges; The set of left-side nodes is constructed based on the physical state variables corresponding to each row of the predicted state matrix; The set of nodes on the right is constructed based on the normative constraint limits in the preset normative knowledge base; The process of constructing the edges specifically includes: If the physical state variable is subject to the specification constraint limit, then an edge is constructed between the corresponding left and right nodes, and the maximum out-of-bounds magnitude of the physical state variable exceeding the specification threshold is used as the weight of the edge. Step S302: Traverse all edges and perform violation determination. The logic of violation determination includes: if the weight is greater than zero, then a violation is determined to have occurred. Step S303: Based on a differentiable automatic differentiation framework, calculate the gradient of the edge weights that are determined to have violated the rules with respect to the intentional operation vector; Step S304: Extract the index of the operation variable with the largest absolute gradient value and the corresponding gradient value.

[0010] As an optimization, step S3 further includes the following sub-steps: Step S305: Construct a prohibition mask matrix, wherein the rows of the prohibition mask matrix represent operation parameters and the columns represent perturbation directions; Based on the sign of the gradient value, the elements in the prohibition mask matrix corresponding to the index of the operation variable and the perturbation direction are set as prohibition markers, and the remaining elements are set as allow markers.

[0011] As an optimization, step S4 includes the following sub-steps: Step S401: Construct a search tree with operation combinations as nodes and perturbation actions on operation parameters as branches; Step S402: Perform the first pruning, the second pruning, and the third pruning in sequence; The first pruning process includes: Traverse the preset set of discrete actions, read the prohibition mask matrix, and if the element corresponding to the current action is a prohibition flag, then prune and discard the branch; if it is a permission flag, then execute the action to generate the second operation vector. The second pruning process includes: A second composite state vector is constructed based on the second operation vector. The corresponding predicted state vector is obtained through a time series deduction model. The constraint bipartite graph is used to detect violations. If a violation is detected, the node is treated as a violation dead node and pruned. The third pruning process includes: Calculate the comprehensive score of the candidate nodes that have passed the first and second pruning, sort them in descending order of comprehensive score, and retain the preset number of nodes with the highest scores to enter the next level of expansion; Step S403: Obtain the compliance and security of the candidate node, wherein the compliance is the total number of consecutive time steps from the root node to the current node without triggering a violation of the specification; The security level is the sum of the number of deduction steps and the compliance level from the current node, keeping the current operation vector constant and extrapolating to the future until the first violation of the specification is triggered. The process of obtaining the comprehensive score specifically includes: The security level is input into a preset gating function to obtain a security discount factor. The security discount factor is then multiplied by the compliance level to obtain the comprehensive score.

[0012] As an optimization, step S5 includes the following sub-steps: Step S501: Based on the decision hesitation degree, the progress benefits and accountability losses are weighted and integrated to obtain the comprehensive utility; The schedule benefit is calculated based on the ratio of the compliance of the optimal alternative construction sequence to the expected total number of operation time steps; The accountability loss is calculated based on the safety of the optimal alternative construction sequence; The decision hesitation level is calculated based on the environmental mutation index; Step S502: Compare the overall utility with preset automatic release threshold and forced interception threshold; When the overall utility is greater than or equal to the automatic release threshold, a compliant release decision response is generated. When the overall utility is less than or equal to the mandatory interception threshold, a mandatory interception decision response is generated. When the overall utility falls between the automatic release threshold and the mandatory interception threshold, a manual decision response is generated.

[0013] As an optimization, the process of calculating the environmental mutation index specifically includes: Select a preset set of key sensors, and calculate the first-order difference sequence of the normalized sequence of each sensor in the set within the preset time window; The first-order difference sequence is subjected to moving average filtering, and the variance of the filtered sequence is calculated. The environmental mutation index is obtained by taking the mean of the variances of the filtered sequences of each sensor in the set.

[0014] The beneficial effects of this invention are as follows: The invention constructs a constrained bipartite graph and uses a differentiable automatic differentiation framework for violation judgment and source tracing. It can not only quantify and predict the extent of the deviation from the norm threshold, but also track and locate the core operational parameters that cause the violation and their deterioration direction by calculating gradients in reverse. This invention uses the generated prohibition mask matrix combined with a heuristic pruning search tree to lock and block dangerous deterioration paths, while automatically and efficiently finding the optimal alternative construction scheme that balances compliance and safety in the operational variable space. Attached Figure Description

[0015] Figure 1 The flowchart illustrates the steps of an optimization method for handling safety questions in water conservancy and hydropower construction based on time-series deduction, as provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example, refer to Figure 1 This paper provides an optimized method for handling safety questions in water conservancy and hydropower construction based on time-series analysis, including the following steps: Step S1: Collect basic working condition data and historical control logs from the construction site, perform normalization preprocessing, and obtain the historical working condition matrix and historical operation matrix. Step S2: Receive natural language question text and perform intent parsing, extract operation parameters to construct intent operation vector, and perform iterative deduction based on time series inference model combined with historical working condition matrix, historical operation matrix and operation vector to obtain prediction state matrix; Step S3: Construct a constraint bipartite graph based on the predicted state matrix to perform normative violation detection. When a violation is detected, trace the source of the violation, obtain the operation parameters that caused the violation, and generate a prohibition mask matrix based on the operation parameters. Step S4: Construct a search tree based on the prohibition mask matrix, and search for compliant alternatives through heuristic pruning to obtain the optimal alternative construction sequence and its corresponding compliance and safety. Step S5: Conduct a comprehensive utility assessment based on compliance and security, and generate a response based on the comparison between the comprehensive utility assessment results and the decision threshold.

[0018] This invention introduces a time-series inference model and deeply integrates historical real-world working conditions and operation logs to achieve forward-looking physical state prediction of users' natural language construction operation intentions. When a predicted violation risk is detected, it can not only proactively trace the disaster-causing parameters, but also automatically optimize alternative construction sequences that balance safety and progress in a complex space of operational variables based on search trees and heuristic pruning algorithms. Finally, it combines comprehensive utility evaluation to generate refined multi-level decision responses, upgrading the question-and-answer system from a passive knowledge retrieval tool to a proactive dynamic expert system with intelligent inference and decision support capabilities.

[0019] Step S1 includes the following sub-steps: Step S101: Collect temperature data, flow rate data and wind speed data as basic operating condition data, and extract discrete data points within a preset time window and align them according to timestamps to form the original operating condition matrix. Step S102: Extract the historical control log within the preset time window, parse the actual execution value of each operation parameter and form a historical operation column vector, and concatenate the historical operation column vectors of each time step in chronological order to form the original operation matrix. Step S103: Based on the preset operation parameter boundary table, normalization preprocessing is performed on each element in the original operating condition matrix and the historical operation matrix to obtain the historical operating condition matrix and the historical operation matrix respectively.

[0020] In one specific embodiment of the present invention, temperature data is collected through multiple sets of distributed optical fiber temperature measuring cables, and the multiple sets of distributed optical fiber temperature measuring cables are arranged in an S-shape along the vertical direction of the concrete silo and the key horizontal layer. Flow data is collected by electromagnetic flow meters, which are fixedly installed on the main cooling water supply pipelines that enter each concrete slab to collect real-time flow data of cooling water flowing through the pipelines. The wind speed data is the high-altitude wind speed collected by an ultrasonic anemometer, which is transformed into the equivalent convective heat transfer wind speed on the concrete surface by the vertical wind speed profile equation based on the power law distribution of the atmospheric boundary layer. An ultrasonic anemometer is installed at the highest point of the tower crane or other unobstructed high point in the on-site collaborative operation to collect the high-altitude wind speed in the construction site environment. Since the high-altitude wind speed cannot be directly equivalent to the convective heat transfer wind speed on the concrete surface, this embodiment converts the high-altitude wind speed data into the equivalent convective heat transfer wind speed on the concrete surface based on the vertical wind speed profile equation of the power law distribution of the atmospheric boundary layer. The mathematical expression for the equivalent convective heat transfer velocity is: ; in, The equivalent wind speed over the concrete surface. The high-altitude wind speed collected at the top of the tower crane. The relative height of the concrete pouring slab from the bedrock reference surface. This represents the relative height of the anemometer from the bedrock reference surface. The bedrock reference surface, also known as the dam foundation elevation, is a known constant in the engineering design. The preset ground roughness index is 0.15 in this embodiment; The data acquisition frequency of all the above sensors was uniformly set to once every 10 minutes; When a user enters a natural language question text, triggering the security question-and-answer process of this invention, the system extracts all continuous basic operating condition data from the past 24 hours, with the question time as the endpoint. Within the 24-hour time window, each sensor will collect 144 discrete data points arranged in chronological order. The discrete data acquired by all sensor channels at the same time step are used to construct an original operating condition column vector. All the original operating condition column vectors are arranged in column vector order to form an original operating condition matrix. The number of rows in the original operating condition matrix is ​​the total number of sensors involved in the monitoring. The column number represents the number of time steps. Each element in the original operating condition matrix represents the data collected by each sensor at each time step; Based on the data in the pre-stored working condition data boundary table, the data in the original working condition matrix is ​​normalized and preprocessed using max-min normalization to obtain the historical working condition matrix. The working condition data boundary table sets the absolute lower and absolute upper limits for each type of working condition data under the current construction scenario. The absolute upper and lower limits of each physical state variable in this table are determined based on relevant design specifications for water conservancy and hydropower projects, statistical extreme values ​​of historical construction monitoring data, and the physical capabilities of the equipment. For example, the preset concrete temperature range is... The internal surface temperature difference range is preset to The cooling water flow rate range is preset to The wind speed range is preset to ; Synchronously retrieve the historical control logs of the field programmable logic controller for the corresponding past 24 hours. The historical control logs record the actual execution setpoints of each adjustable operating parameter in each control cycle in the form of a time series. Based on the preset operation parameter mapping table, the descriptive field names of each operation parameter in the log are parsed into the corresponding dimension indexes. ,in For each time step ,in Extract the actual execution values ​​of each operation parameter corresponding to that time step to form a... Initialize the column vector of the dimension operation; The specific operating parameters include the set value of the water flow rate of each cooling water pipe, the speed multiplier of the cable crane, and the speed multiplier of the tower crane. If there are multiple records in the same time step, the value with the closest timestamp is taken; if there is no record, the value of the previous moment is used to ensure the continuity of the time series. The actual execution value of each dimension in the initial operation column vector is normalized according to the preset upper and lower boundaries in the operation parameter boundary table corresponding to the operation parameter, and mapped to the [0,1] interval to obtain the historical operation column vector. All The normalized historical operation column vectors of each time step are concatenated horizontally in chronological order to obtain the historical operation matrix. The number of rows in the historical operation matrix is ​​equal to the total number of operation parameters. The column number represents the number of time steps. , its first Line 1 The elements of the column represent the first The operation parameter in the first The normalized value of the actual execution setpoint for each historical time step; The operation parameter boundary table contains the preset absolute lower and upper bounds for each operation parameter. For example, the boundary of the continuous operation plan duration is preset to [0 hours, 24 hours], the boundary of the cooling water flow rate is preset to [0 m³ / h, 5.0 m³ / h], and the boundary of the cable car running speed ratio is preset to [0.1, 1.0].

[0021] Step S2 includes the following sub-steps: Step S201: Input the question text into the pre-trained named entity recognition network, output entity labels, and parse the structured operation parameters from the text based on the entity labels; Step S202: After normalizing the parsed operation parameters, fill them into the corresponding dimension positions in the operation vector of the preset length to obtain the intention operation vector.

[0022] In one specific embodiment of the present invention, a pre-trained named entity recognition network is used to parse the natural language security query input by the user into an operation vector, and predict the future state based on historical data. The Named Entity Recognition Network receives a sequence of Chinese characters from a user's query text. After character vector mapping, bidirectional long short-term memory network encoding, and conditional random field decoding, it outputs the entity label corresponding to each Chinese character. Based on the label sequence, it parses the structured operation parameter information from the text. For example, when receiving a natural language question text from a construction user, "We plan to carry out a large-volume foundation slab pouring for 12 consecutive hours tonight on section 3 of the dam. Please assess the risk of setting the cooling water flow rate to 2.0 cubic meters per hour," the system ultimately parses out that the planned operation duration is 12 hours and the cooling water flow rate is set to 2.0 cubic meters per hour. Then, the system normalizes the parsed operation parameters, namely 12 hours and 2.0 cubic meters per hour. The normalized discrete operation parameters are integrated and mapped into a preset length. operation vector In the process, the intent operation vector is obtained. Each dimension of the intent operation vector uniquely corresponds to a specific, executable operation parameter, such as cooling water flow rate or tower crane operating speed ratio. It is the total number of all remotely controllable operations in the project; operation vector The filling rules for each dimension are as follows: The operation parameters mentioned in the user intent parsing are assigned normalized discrete parameter values, such as the normalized values ​​of parameters 12 hours and 2.0 cubic meters per hour. For other operational parameters not included in the intent resolution, the normalized actual operating settings of the corresponding device in the field programmable logic controller are automatically read to ensure the continuity between the simulation baseline and the actual field conditions, and the dimension of the operational vector. The number of rows in the forbidden mask matrix should be consistent with the number of rows in subsequent steps. The Named Entity Recognition Network consists of a character vector mapping layer, a bidirectional long short-term memory network layer, and a conditional random field layer connected in series. The pre-training process of the character vector mapping layer is independent of the training of the named entity recognition network. It employs the continuous bag-of-words architecture of the Word2Vec model for unsupervised learning on a large-scale unstructured construction text corpus of water conservancy and hydropower projects. Specifically, 10GB of plain text documents, including construction logs, daily supervision reports, technical briefing records, and equipment operation reports from water conservancy and hydropower projects, are pre-collected to form an unsupervised pre-training corpus. The Word2Vec model scans each sentence in the corpus using a sliding window approach, taking the context characters of the central Chinese character as input to predict the central character itself. It iterative optimization using a stochastic gradient descent algorithm maximizes the probability of correctly predicting the central character, ultimately mapping each Chinese character to a 300-dimensional dense real-number vector. After pre-training, the vector representations of all Chinese characters are solidified, forming a character vector mapping matrix. Its dimensions are ,in This represents the total number of unique Chinese characters in the corpus. When the query text is received, the word vector mapping layer performs character segmentation on the query text, breaking it down into characters. A character sequence consisting of Chinese characters, among which Given the total number of Chinese characters in the text, for each Chinese character in the character sequence, query the character vector mapping matrix. This maps Chinese characters to their corresponding 300-dimensional character vectors. Will The character vectors of each Chinese character are arranged in order to form a character vector sequence, which serves as the input to the bidirectional long short-term memory network layer in the named entity recognition network. Training a named entity recognition network requires labeled supervisory data. From the aforementioned construction corpus, 50,000 sentences containing scheduling instructions were selected. Each Chinese character was manually labeled with its entity tag, using the BIO tagging system: B-OP represents the starting character of an operation parameter type, I-OP represents the inner or ending character of an operation parameter type; B-VAL represents the starting character of a parameter value, I-VAL represents the inner or ending character of a parameter value; O represents a non-entity character. During training, the word vector sequence is input into a bidirectional long short-term memory network layer, which consists of two parts: a forward LSTM and a backward LSTM. The forward LSTM scans from the beginning to the end of the sentence, while the backward LSTM scans from the end to the beginning of the sentence. For the first... The word vector at each position will be used to feed the hidden output of the LSTM at that position. Hidden output of inverse LSTM By concatenating the vectors, we obtain the contextual semantic feature vector at that location. ; All The sequence input to the Conditional Random Field (CRF) layer is composed of contextual semantic feature vectors at each position. The CRF layer jointly predicts the label sequence of the entire sentence by modeling the global transition probability of the label sequence and outputs the optimal label sequence. The training aims to maximize the log-likelihood function of the correct label sequence. The stochastic gradient descent algorithm is used for backpropagation to update the parameters of the bidirectional long short-term memory network layer and the CRF layer. When the F1 score on the independent validation set reaches 0.90 or above, training stops and all parameters are fixed.

[0023] This invention utilizes a pre-trained named entity recognition network to extract core operational parameters from query text and perform standardized normalization mapping. At the same time, it automatically inherits the current actual operating settings of other devices not explicitly mentioned by the user. This not only accurately captures the user's local control intentions but also fully preserves the complex global construction baseline conditions on site, avoiding inference distortion caused by missing intent information.

[0024] Step S2 also includes the following sub-steps: Step S203: Vertically concatenate each column vector in the historical operating condition matrix with the column vector of the corresponding time step in the historical operation matrix to form multiple historical composite state vectors, and input them sequentially into the time series extrapolation model to activate the network memory and obtain the initial memory state. The time-series extrapolation model includes an input layer, a stacked long short-term memory network layer, and a fully connected output layer; In one specific embodiment of the present invention, a time-series deduction model based on a long short-term memory network is constructed to predict the working state after the user's intention is performed. Each column of the historical operating condition matrix is ​​vertically concatenated with the historical operation column vector of the corresponding time step in the historical operation matrix to form a 144-dimensional historical composite state vector. These historical composite state vectors are then sequentially input into the pre-trained time series inference model. For each input historical composite state vector, the hidden states in the time series extrapolation model are... and cell state After updating once, once all historical composite state vectors have been input, the time series simulation model acquires and records the evolution of the operating conditions over the past 24 hours, extracts and retains the hidden state vector at this point. and cell state vector Hidden state vector and cell state vector This constitutes the initial memory state; The input of the time series extrapolation model is a composite state vector, and the output is the predicted value of the working condition data at the future time step. The structure of the time series extrapolation model includes an input layer, an LSTM stacked layer, and a fully connected output layer. The input layer is used to receive The sum of the column vectors of working conditions in dimension Dimensional operations on column vectors, The sum of the column vectors of working conditions in dimension Dimensional operations concatenate column vectors into A composite state vector of dimension ; The LSTM stacked layer is composed of a first LSTM layer and a second LSTM layer stacked vertically. The first LSTM layer receives the signals from the input layer. A composite state vector is mapped to a 128-dimensional hidden state sequence. The first LSTM layer is configured to return the complete sequence, and its output 128-dimensional hidden state sequence serves as the input to the second LSTM layer. The second LSTM layer contains 128 hidden units and is configured to return only the hidden state at the last time step. After extracting temporal features from the input sequence, it outputs the hidden state at the final time step. LSTM layers are used to capture the temporal dependencies of the physical field evolution and pass the hidden state between time steps. and cell state ; The fully connected output layer is used to store the hidden state of the second LSTM layer at the final time step. Mapped to The predicted state vector is of dimension 1. The fully connected output layer does not use a bias term. An adaptive activation function is used, and its mathematical expression is: ; in, The output value of the activation function. The linear output value of the fully connected layer. The overflow permeability coefficient, To smooth out control parameters, this embodiment takes... , Adaptive activation function on input Approximately linear, in Controlled overflow is allowed at certain times, which enables model predictions to break through normalized historical extremes; The pre-training process of the time series extrapolation model is as follows: Training data was extracted from the historical multimodal time series database accumulated from similar dam pouring operations in the early stages of this water conservancy project. The historical multimodal time series database recorded the actual execution values ​​of all sensor values ​​and all operational variables during each pouring process at 10-minute intervals. With a sliding step of 10 minutes, continuous time windows were extracted from each historical pouring record to construct training sample pairs. Each training sample pair contains an input and a ground truth label. The input is a composite state vector at a certain time step, derived from the state at that time step. Normalized values ​​of each sensor and The truth label is formed by concatenating the normalized values ​​of the operands; the truth label is the next time step, i.e., 10 minutes later. Normalized values ​​from each sensor.

[0025] The training uses mean squared error as the loss function. For each training sample, the input composite state vector is fed into the network for forward propagation to obtain the predicted physical state vector for the next time step. The average of the squared differences between each element of the predicted vector and the ground truth label vector is calculated, which yields the loss value for that sample. The loss value for a batch of training samples is the arithmetic mean of the loss values ​​of all samples. The Adam optimizer is used for gradient backpropagation. The partial derivative of the loss value with respect to the weights of each layer of the network is calculated, and the weights of each layer are iteratively updated along the gradient descent direction to gradually reduce the loss value. During training, after each training round, the validation loss is calculated on an independent validation dataset. When the validation loss no longer decreases in several consecutive rounds, the network is considered to have converged and training is stopped.

[0026] This invention concatenates the historical working condition matrix and the historical operation matrix step by step and inputs them sequentially into the stacked layers of the Long Short-Term Memory Network. This forces the model to fully reproduce and preheat the physical field evolution process over a period of time before receiving the current operation instruction. As a result, the initial memory state extracted by the model not only includes the static slice of the current moment, but also contains complex dynamic historical inertial characteristics such as the release of heat of hydration of concrete and heat exchange with the environment. This improves the physical temporal coherence and accuracy of subsequent predictions of future working conditions.

[0027] Step S2 also includes the following sub-steps: Step S204: Extract the last column vector of the historical working condition matrix and concatenate it with the operation vector to form the first composite state vector; Step S205: Using the time-series extrapolation model combined with the initial memory state, the first composite state vector and the initial memory state are used as inputs to iteratively calculate the state prediction vectors for multiple future time steps, and the state prediction vectors of each time step are horizontally concatenated to form a prediction state matrix.

[0028] In one specific embodiment of the present invention, the last column vector of the historical working condition matrix is ​​extracted, and this vector represents the entirety of the current real-time data. Normalized values ​​from individual sensors; The last column vector of the historical operating condition matrix is ​​vertically concatenated with the intended operation vector to obtain the first composite state vector. ; Using the first composite state vector and the initial memory state as input, the predicted state vector for the next time step, i.e., the first future time step, is obtained through a time-series extrapolation model. and update the initial memory state. , obtain ; The predicted state vector is an N-dimensional column vector, where N is the total number of sensors. Each element in the predicted state vector corresponds to the predicted value of a physical state variable monitored by a sensor at a certain time step in the future. The predicted state vector and the intended operation vector at the next time step are concatenated, and then processed through a time-series inference model and... Obtain the predicted state vector at the second time step; This embodiment sets up the projection of the next 12 hours, corresponding to 72 predicted state vectors with a step size of 10 minutes. For the _th Next simulation, Input the composite state column vector from the previous time step into the model, and the model combines it with the memory state from the previous time step stored internally. Perform forward propagation to obtain the first... The predicted state vector at each time step and time step memory state ; The predicted state vectors at all time steps are concatenated to obtain the predicted state matrix, which has the following dimensions: Each row of the predicted state matrix corresponds to a physical state variable monitored by a sensor, and each column corresponds to a future time step.

[0029] This invention extracts the latest historical working condition vector and the intention operation vector to form the first inference base point, and relies on the fully activated initial memory state to carry out rolling forward propagation. Without relying on the intervention of real future data, it can continuously infer and predict the physical state evolution trajectory of multiple future time steps by relying on the cell state transmission within the model. This allows for forward-looking trial and error in the digital space, providing complete and fine-grained panoramic predictive data support for subsequent long-term violation risk assessment.

[0030] Step S3 includes the following sub-steps: Step S301: Construct a constrained bipartite graph, which includes a left set of nodes, a right set of nodes, and edges; The left-hand node set is constructed based on the physical state variables corresponding to each row of the predicted state matrix; The set of nodes on the right is constructed based on the normative constraint limits in the preset normative knowledge base; The process of constructing edges specifically includes: If the physical state variable is subject to the specification limit, then an edge is constructed between the corresponding left and right nodes, and the maximum out-of-bounds magnitude of the physical state variable exceeding the specification threshold is used as the weight of the edge. Step S302: Traverse all edges and perform violation judgment. The logic for violation judgment includes: if the weight is greater than zero, then a violation is judged to have occurred. Step S303: Based on a differentiable automatic differentiation framework, calculate the gradient of the edge weights that are determined to have violated the rules with respect to the intention operation vector. Step S304: Extract the index of the operation variable with the largest absolute gradient value and the corresponding gradient value.

[0031] In one specific embodiment of the present invention, the derived predicted state matrix is ​​compared with the construction specifications. Specifically, the left-hand node set represents the real-world physical states parsed from the predicted state matrix. Nodes are constructed using the category of physical state variables as the dimension; that is, each row of the predicted state matrix is ​​transformed into a physical state variable node in the left-hand node set, with a total number of nodes. For the left node Extract the predicted state matrix. The row vector of a row is used as its initial feature vector; The set of nodes on the right represents the quantified constraint limits in the construction specifications, and the construction process specifically includes: A pre-defined structured specification knowledge base is provided, which includes multiple quantified construction specification requirements. Each requirement includes the following fields: Construction specification name and clause number, name of constrained physical state variable, constraint type, constraint value, and physical unit; The names of the constrained physical state variables include internal surface temperature difference, internal concrete temperature, and surface temperature. Constraint types include upper bound constraints and lower bound constraints; When constructing the right-side node set, based on the current construction type, such as large-volume concrete pouring, all relevant constraint records are retrieved from the specification knowledge base. For each retrieved constraint record, its constraint value is normalized according to the working condition data boundary table corresponding to the physical state variable. Each constraint record is transformed into a right node in the set of right nodes. This node It contains two attributes: the upper bound of the normalization constraint. and normalization constraint lower bound ; If the constraint type is an upper limit constraint, then Let the above normalization constraint value be set. Set to 0; if the constraint type is a lower bound constraint, then Set as the normalization constraint value. Set to 1; If node The physical state variable represented, and If the names of the constrained physical state variables are consistent with the requirements of the corresponding construction specifications, then it is determined that... by Constraints, in and Connecting edges; The weight of an edge represents the maximum extent to which a physical state variable exceeds the specification threshold. The mathematical expression for the edge weight is: ; in, To connect the first The physical state variable node and the first The edge weights of a standard constraint node. To calculate the number of steps, For the first The physical state variable at the th ... The predicted value at each time step, and The first Normalized upper and lower bounds of a normative constraint; If the boundary is not exceeded throughout the entire process, the edge weight is 0; Traverse all edges, if If a violation is determined, the violation will be traced back to its source. Specifically, The calculation starts with the first composite state vector. After 72 iterations of the time-series deduction model, a prediction state matrix is ​​generated. The elements of the prediction state matrix are compared with the upper and lower bounds and obtained by the max function. The computation process constitutes a computation graph, which is complete and differentiable. In this embodiment, the automatic differentiation framework is used to call its backpropagation function to calculate the gradient of the weight of the violation edge relative to the intended operation vector. The automatic differentiation framework automatically records the input-output dependencies of each step of the computation during forward propagation and constructs a computation graph; during back propagation, it traverses the computation graph in reverse and automatically calculates the partial derivatives of each intermediate variable with respect to the final result based on the chain rule. For non-differentiable operations such as the max function in the computation graph, the automatic differentiation framework has a built-in subgradient handling rule. When the function is at a nondifferentiable point, the subgradient is automatically used to replace the regular gradient to ensure the continuity of the computation process. Through the aforementioned automatic differentiation and backpropagation, the long-range violation risk is traced back to the operational parameter that has the most significant impact on the critical moment, and the index of the operational parameter with the largest absolute gradient value is extracted. It also generates a text traceability record containing the specific time point, the value of the exceedance, and the violated clause.

[0032] This invention constructs a weighted constrained bipartite graph connecting the predicted physical state nodes and the specification constraint nodes, quantifies the maximum out-of-bounds magnitude of the predicted state with respect to multiple specification thresholds, and utilizes the backpropagation and chain rule of the differentiable automatic differentiation framework. When a violation is detected, by calculating the gradient of the violation edge weight with respect to the input intention operation vector, it is possible to backtrack and quantify the core disaster-causing operation parameters that lead to the violation from complex coupled variables.

[0033] Step S3 also includes the following sub-steps: Step S305: Construct a prohibition mask matrix. The rows of the prohibition mask matrix represent operation parameters, and the columns represent perturbation directions. Based on the sign of the gradient value, the elements in the forbidden mask matrix corresponding to the index of the operation variable and the perturbation direction are set as forbidden markers, and the remaining elements are set as allowed markers.

[0034] In a specific embodiment of the present invention, a dimension is constructed. A matrix of all zeros, used as a forbidden mask matrix. Each row of the prohibition mask matrix corresponds to an operation parameter. Each operation parameter represents its taboo state through the elements in the prohibition mask matrix. The two columns of the prohibition mask matrix are used to record the taboo state of each operation parameter in two different directions. Forbidden column indexes of the mask matrix Represents the direction of decrease of the operation parameter, column index. Indicates the direction of increase of the operating parameter; No. Is the decreasing direction of column records prohibited? Is increasing the column record direction prohibited? Index of operation parameter for extracting the maximum absolute gradient and the specific values ​​of their gradients ; like This indicates that under the current operating conditions, adjusting any single operating parameter to alleviate the local gradient violation of the specification is not significant. To prevent invalid perturbations, all elements of the prohibition mask matrix are set to 1, which means that all operating variables are prohibited from being adjusted in any direction. This setting will cause the subsequent search tree to fail to generate any legitimate branches, thus naturally terminating the search and ultimately providing the user with feedback that the intended operation is unsafe and that no alternative solution can be automatically generated. like This indicates that increasing this operand will lead to a higher violation rate, thus prohibiting the first element in the mask matrix. Row and column indexes are Setting an element to 1 only prevents the parameter from being increased. like This indicates that reducing this operand will exacerbate the violation, thus prohibiting the first element in the mask matrix. Setting the element in the first column of the row to 1 only prevents the parameter from being decreased. This targeted prohibition operation not only precisely locks down the path that leads to disaster worsening, but also fully preserves the operational parameters for reverse regulation and self-rescue. Since the second pruning of the subsequent search tree will detect and remove all operational branches that will lead to future violations in real time, prohibiting only the direction of the main cause of violation is enough to block the most dangerous direct path at present. Other potentially dangerous directions will be naturally eliminated in subsequent searches.

[0035] Based on the local gradient numerical symbols obtained from violation tracing, this invention selectively locks down specific operational parameters and their single perturbation direction that exacerbate the violation trend within the mask matrix, rather than blindly freezing all operational parameters. This asymmetric targeted prohibition strategy blocks the shortest disastrous path to a high-risk violation state while retaining the flexibility to adjust in the opposite direction for self-rescue and the degree of freedom to adjust other non-disastrous variables, thus avoiding excessive restriction of the safe search space in the subsequent optimization process.

[0036] Step S4 includes the following sub-steps: Step S401: Construct a search tree with operation combinations as nodes and perturbation actions on operation parameters as branches; Step S402: Perform the first pruning, the second pruning, and the third pruning in sequence; The first pruning process includes: Traverse the preset set of discrete actions, read the prohibition mask matrix, and if the element corresponding to the current action is a prohibition flag, then prune and discard the branch; if it is a permission flag, then execute the action to generate the second operation vector. The second pruning process includes: A second composite state vector is constructed based on the second operation vector. The corresponding predicted state vector is obtained through the time series deduction model. The constraint bipartite graph is used to detect violations. If a violation is detected, the node is treated as a violation dead node and pruned. The third pruning process includes: Calculate the comprehensive score of the candidate nodes that have passed the first and second pruning, sort them in descending order of comprehensive score, and retain the preset number of nodes with the highest scores to enter the next level of expansion; Step S403: Obtain the compliance and security of the candidate nodes. The compliance is the total number of consecutive time steps from the root node to the current node where no violation of the specification has been triggered. The safety level is the sum of the number of calculation steps and the compliance level, starting from the current node, keeping the current operation vector constant and extrapolating to the future until the first violation of the norm is triggered or the maximum number of calculation steps is reached. The process of obtaining the overall score includes: The security level is input into a preset gating function to obtain a security discount factor. The security discount factor is then multiplied by the compliance level to obtain a comprehensive score.

[0037] In one specific embodiment of the present invention, the purpose of this step is to automatically search for compliant alternative construction solutions when a violation is detected, and to answer the question of how to adjust the construction plan to ensure safe operation. In order to automatically search for compliant alternative construction solutions, this embodiment constructs a search tree based on a prohibition mask matrix and uses heuristic search to prune the search tree; Each level of the search tree corresponds to a future time step, i.e., 10 minutes. Each time the search tree expands downward, it represents advancing time by one step. The search tree includes nodes and branches. A node represents the state of the working condition after performing an action at a specific future time step. This state includes the operation vector value at that time step, the predicted state value, and the hidden memory state of the time series extrapolation model. A branch represents an action performed from the parent node to the child node. The action includes an index of operation parameters. Disturbance direction and perturbation step size ,in ; Disturbance direction for When, it indicates an increase, and the direction of the disturbance. for When, it indicates a decrease; the operation parameter index indicates the type of operation parameter; and the disturbance step size indicates the magnitude of the adjustment. Operating parameters include continuous and discrete variables. Continuous variables, such as flow velocity, can take any value, depending on the direction of the disturbance. and perturbation step size Adjust the corresponding operation parameter index The magnitude of the corresponding operating parameters, and for continuous variables within the operating parameters, the perturbation step size. The default value is 0.05; Discrete variables, such as gear positions, take discrete values. Their action is defined as switching to adjacent discrete values, such as adjacent gear positions. Specifically, the gear position values ​​of discrete variables are pre-normalized and mapped to unique normalized values ​​in the interval [0,1], forming an ordered set of discrete points. When a perturbation is applied to a discrete variable, the current normalized value is searched in the ordered discrete point set according to the perturbation direction δ. If δ=+1, then the smallest discrete point larger than the current normalized value is selected as the target discrete point; if δ=-1, then the largest discrete point smaller than the current normalized value is selected as the target discrete point. If the current normalized value is already at the upper boundary of the ordered discrete point set and δ=+1, or is already at the lower boundary and δ=-1, then there are no reachable neighboring discrete points in that direction, this action is considered invalid, and the branch is pruned directly with a perturbation step size. The current normalized value The absolute value of the difference between the normalized value and the target discrete point; The heuristic pruning process includes first pruning, second pruning, and third pruning in sequence. For the root node of the search tree, the physical state of the root node is the last column vector of the historical working condition matrix, and the operation vector is the user's intention operation vector. For the root node, the preset discrete action set is directly traversed. Perform the first, second, and third pruning. The first pruning is a directed prohibition pruning, with a preset set of discrete actions for the current node. The Cartesian product of all adjustable operating parameters and their two perturbation directions, i.e. ,in For the first The perturbation step size corresponding to each operating parameter This represents the total number of operation parameters. Traverse the action set Each action in the process is performed using a formula. Disturbance direction Mapped to column indices of the forbidden mask matrix ,Right now correspond , correspond ; Read the prohibition mask matrix; if the prohibition mask matrix... OK Column elements If the value is 0, it indicates that the action is in a direction that exacerbates the disaster, then the branch is directly pruned and discarded, and will not expand further downwards; if the value is 0, then the action is executed. After executing the action, based on the perturbation direction and perturbation step size, the updated operating parameters are obtained, i.e., the perturbed operating parameters. For the continuous variables in the updated operating parameters, an action is applied. Boundary clipping generates new operation parameters, and the new operation parameters are: ; For discrete variables, the updated operation parameter is directly taken as the normalized value of the target discrete point. Since the target discrete point itself is a valid normalized value of the operation parameter, the updated value does not need to be clipped by the boundary. A new operation vector is generated based on the updated operation parameters and used as the second operation vector. The second operation vector is then concatenated with the predicted state vector corresponding to the time step represented by the current level to generate a new composite state vector, which is used as the second composite state vector. The second operation vector is forcibly inherited to subsequent child nodes until it is covered by a new perturbation. The second pruning is the pruning of illegal dead nodes. Based on the second composite state vector, the predicted state vector of the next time step corresponding to the second composite state vector is obtained through the time series inference model. Based on the constrained bipartite graph, the illegality judgment of the predicted state vector of the next time step corresponding to the second composite state vector is performed. If any side is found to be illegal, that is, the edge weight is greater than 0, the node is judged as an illegal dead node and is pruned directly without further expansion. The third pruning is bundle search pruning. If, at each level, a comprehensive score is calculated for all candidate nodes that have passed the first and second pruning, and the comprehensive score values ​​are sorted in descending order, only the top 5 nodes with the largest comprehensive score values ​​are retained as bundle widths to enter the next level for expansion, and the remaining nodes are pruned, the tree width of the search tree is the maximum number of candidate nodes retained at each time step level, which is 5 in this embodiment. Table 1 shows a comparison of the computational scale of candidate child nodes of a single node in each layer under different bundle width values; For the root node of the search tree, the physical state of the root node is the last column vector of the historical condition matrix, the operation vector is the user's intent operation vector, and the memory state is... For the root node, traverse the preset discrete action set. Perform the first, second, and third pruning. In this embodiment, during the process from the root node at level 0 to level 1, a preset set of discrete actions is traversed. Perform the first pruning, and for an action that passes the prohibition mask check, generate the second operation vector; The second operation vector obtained from the first pruning in the first layer is concatenated with the last column vector of the historical working condition matrix to generate the second composite state vector of the first layer. ,Will and memory state The time-series extrapolation model outputs the predicted state vector for the next time step. and the updated memory state ; Continue with the second pruning step, using the constrained bipartite graph to... Conduct testing and continue with the third pruning step; In the process of moving from the first-layer nodes to the second-layer nodes, for each node in the first layer, its predicted state vector is: The operation vector is the second operation vector of the first layer, and the memory state is... ,based on The second operation vector of the first layer and Perform the second and third pruning; For those located at the The nodes of the layer have a known predicted state vector as follows: The operation vector is The memory state is ; Traverse the action set Each action in the process, if it meets the condition If so, the action is executed: obtain the updated operation parameters based on the perturbation direction and perturbation step size, and generate a new operation vector. ; Will The predicted state vector of the current node is Vertical concatenation is performed to form a composite state column vector. ,Will Together with the accumulated memory state of the current node The input is fed into the time series extrapolation model, and a forward propagation is performed to obtain the predicted state vector at step k+1. Meanwhile, the time-series extrapolation model outputs the updated memory state. ; The calculation process for the overall score includes: Obtain the compliance and security of candidate nodes; The compliance level is the depth of the current level, representing the total number of consecutive time steps from the root node to the current node in the deduction that have not triggered any violations. At the root node of the search tree, the initial value of the compliance level is 0. The safety level is calculated as follows: Assume that, starting from the current node, the new operation vector of the current node is maintained at all future time points. The predicted state vector corresponding to the current node remains constant. and Starting from the current node, the time-series deduction model is used to perform deduction until the first violation is triggered. The number of deduction steps from the current node to the first violation is recorded. The safety of the current node is the sum of the compliance and the number of deduction steps. The number of deduction steps is the difference between 72 and the current node level k. Based on the above compliance and security levels, a comprehensive scoring function O reflecting security priorities is constructed, and its mathematical expression is as follows: ; in, For comprehensive scoring, For safety reasons, The step is the minimum acceptable absolute safety depth. This is the steepness parameter; Sigmoid gate function Will Mapped to a safety discount factor in the interval (0,1), when far below hour, The score rapidly decays to near zero, suppressing progress; when When it is large, Approaching 1, this scoring function fully rewards progress, ensuring that only paths with sufficient historical compliance steps and long-term safety depth receive high scores, thus guiding searches toward operational sequences that are both safe and make substantial progress. After the search terminates, if there are surviving leaf nodes, select the path with the highest comprehensive score as the optimal alternative construction sequence, and extract its compliance and safety. To determine the minimum acceptable absolute safety depth, the typical single-compartment operation time for continuous pouring of large-volume concrete is usually 6 to 12 hours. Taking the conservative midpoint of 4 hours, corresponding to 24 10-minute increments, we take this as the minimum acceptable threshold for safety depth. Schemes below this value are considered to have insufficient long-term safety. The steepness parameter of the Sigmoid function controls the score. Near the transition rate, when At that time, the Sigmoid function is The value is 0.5, and the width of the transition interval is moderate; Table 2 shows the different Values The calculation results; As can be seen from the table above, This enables the safety discount factor to be in A smooth transition nearby avoids creating a sharp cutoff at the threshold while significantly penalizing solutions with insufficient security depth.

[0038] This invention progressively removes invalid perturbations, deterministic disaster-causing paths, and low-scoring edge branches by sequentially chaining a first pruning based on a prohibition mask matrix, a second pruning based on dead node detection, and a third pruning based on bundle search. This effectively overcomes the computational dimensionality explosion caused by the combination of multiple operation parameters. The comprehensive scoring function incorporates a safety discount factor from the gating function, forcibly guiding the model to prioritize exploring high-quality paths with sufficient safety depth, ensuring that the final output alternative has sufficient safety margin in actual engineering implementation.

[0039] Step S5 includes the following sub-steps: Step S501: Based on the degree of decision hesitation, the progress benefits and accountability losses are weighted and integrated to obtain the comprehensive utility; Schedule benefits are calculated based on the ratio of compliance of the optimal alternative construction sequence to the expected total number of operation time steps; The accountability loss is calculated based on the safety factor of the optimal alternative construction sequence; Decision hesitation is calculated based on the environmental abruptness index; Step S502: Compare the overall utility with the preset automatic release threshold and forced interception threshold; When the overall utility is greater than or equal to the automatic release threshold, a compliant release decision response is generated. When the overall utility is less than or equal to the mandatory interception threshold, a mandatory interception decision response is generated. When the overall utility falls between the automatic release threshold and the mandatory interception threshold, a manual decision response is generated.

[0040] In one specific embodiment of the present invention, the purpose of this step is to conduct a comprehensive utility evaluation of the alternative construction scheme in order to determine the final action and form a question-and-answer conclusion; The mathematical expression for comprehensive utility assessment is: ; in, For overall utility, For progress benefits, To hold those responsible for the losses accountable, For decision-making hesitation; The mathematical expression for schedule benefits is: ; in, To establish a standard value for the reference dimension, this embodiment uses a value of 15. The value is determined based on the absence of environmental fluctuations. The value is 0, and the solution is fully compliant. and Under equal ideal working conditions, accountability losses When it approaches 0, the overall utility is... To ensure that the efficiency under this ideal operating condition significantly exceeds the automatic release threshold And with sufficient decision margin, the value is set to 15; This constant is set based on the comparison between the expected maximum compliance progress benefit and the release threshold, so that the utility value of the fully compliant solution is 15, which is significantly higher than the release threshold of 8.0, ensuring sufficient decision margin. This is an empirical parameter in prospect theory that characterizes diminishing returns sensitivity. To ensure compliance with the optimal alternative construction sequence, The expected total number of time steps for the task; ratio This refers to the percentage of time spent working safely. The mathematical expression for accountability loss is: ; in, The base unit is set to a constant, sharing the same base with the schedule benefits, so that the benefit items and loss items have equivalent magnitudes; To mitigate the amplification factor for losses, The empirical parameters for loss sensitivity are all derived from the classic settings of foreground theory; The mathematical expression for decision hesitation is: ; in, The scaling factor is a preset value; in this embodiment, it is set to 5.0. This is an environmental mutation index; when hour, The overall utility is determined by both schedule gains and accountability losses. Increase Approaching 1, progress gains The weights are compressed to 0, at which point... Decision-making is entirely based on accountability for losses This leads to a decision-making logic where safety is prioritized in harsher environments. If the search tree has no surviving child nodes due to the prohibition mask matrix being set to all 1s or all search paths being eliminated by the second pruning, it is determined that there is no compliant alternative under the current working condition. The compliance and security scores are forcibly assigned to 0. In the comprehensive utility evaluation, when the compliance and security scores are forcibly assigned to 0, the calculated comprehensive utility evaluation is far lower than the forced interception threshold. This will automatically trigger a forced interception decision response, providing the user with a clear conclusion that the current construction plan has irreconcilable regulatory violations and that no alternative solution can be generated by adjusting the operating parameters.

[0041] Automatic release threshold and forced interception threshold The process of obtaining the value includes: The specific process of building an expert decision-making tag library includes: Based on a pre-trained and parameterized time-series model, historical management logs from past projects are retrieved. Each log entry contains environmental monitoring time-series data, actual operating setpoints of each controllable device, and the final processing result from the project's chief engineer, including release, review, and interception. The actual operating setpoints recorded in the logs are used as the expected operation vector for that operating condition. Combined with the operating condition data at that time, the algorithm calculates the expected operation vector for each historical operating condition under the current algorithm architecture. And the chief engineer's actual handling results will be used as labels; The dataset was divided chronologically into a training set (first 70%) and a test set (last 30%). For each sample in the training set, a Bayesian optimization algorithm was used in a predefined search space. Bayesian optimization seeks the optimal combination of thresholds within each iteration, selecting a set of candidate thresholds. Calculate its overall utility, if Then it is classified as release, if If it is classified as interception, it is classified as review; This yields the confusion matrix for the three classifications. The macro-average F1 score is then calculated. Bayesian optimization aims to maximize this macro-average F1 score, iteratively updating the candidate thresholds until convergence. Ultimately, the algorithm converges on the training set to... ; Based on overall utility and automatic release threshold and forced interception threshold Based on the comparison results, generate the corresponding response content; when When responding, the response includes: the user's intended operation is found to be irregular and infeasible based on the analysis; a safe and feasible alternative construction plan has been found; the complete operation parameters of the generated optimal alternative construction sequence are attached, along with the corresponding safety, compliance and overall effectiveness. when The response included: the user's intended operation was found to be illegal and infeasible based on the analysis, and the optimal alternative construction sequence could not be found. The generated disaster tracing record was attached, clearly pointing out the specific information of the violation. when The response included: the user's intended operation was found to be in violation of regulations, and an alternative construction plan was found. However, the plan was at the critical point between the benefits of rushing the work and the liability for accidents, and the safety margin was insufficient. Therefore, its feasibility could not be automatically determined, and manual review was recommended. The above-mentioned responses are then transformed into natural language responses that the user can understand using a large language model, thus completing the question-and-answer process.

[0042] This invention not only assesses safety levels but also introduces prospect theory to quantify and offset the expected progress benefits of alternative solutions with potential safety accountability losses. By comparing the weighted and fused comprehensive utility with preset dual decision thresholds, it can adaptively output three-level decision instructions: automatic compliance release, mandatory interception, or transfer to manual review. This maximizes the efficiency space for compliant and expedited work while adhering to the absolute safety bottom line of the project, providing clear decision support for on-site management personnel.

[0043] The process of calculating the environmental mutation index specifically includes: Select a preset set of key sensors and calculate the first-order difference sequence of the normalized sequence of each sensor in the set within a preset time window; Apply a moving average filter to the first-order difference sequence and calculate the variance of the filtered sequence. The environmental mutation index is obtained by taking the mean of the variances of the filtered sequences of each sensor in the set.

[0044] In one specific embodiment of the present invention, the environmental mutation index is a comprehensive quantification of the intensity of fluctuations in the normalized values ​​of the key sensor set over the past 24 hours. The larger the value, the more unstable the recent environmental physical field and the more frequent the sudden disturbances. The mathematical expression for the environmental mutation index is: ; in, This is an environmental mutation index. This is a key sensor set, composed of sensors most sensitive to disturbances in the construction environment. In this embodiment, a concrete surface temperature sensor and an ambient wind speed sensor are selected to form the set. , Represents a set The number of sensors in the middle; For the first The first-order difference variance of the normalized sequence of the key sensors over the past 24 hours is calculated as follows: For the first... The normalized sequence of the sensor is used to calculate its first-order difference sequence. The obtained first-order difference sequence is then filtered by a moving average with a window length of 3 to suppress high-frequency noise introduced by the difference operation. During the filtering process, the edges at both ends are filled with duplicate data from the beginning and end. Then the variance of the filtered difference sequence is calculated.

[0045] This invention quantifies the recent volatility of the physical field at the construction site by sequentially performing first-order difference extraction, moving average filtering for noise reduction, and variance calculation on recent key environmental sensor sequences. This environmental mutation index is deeply embedded in the calculation of decision hesitation, enabling dynamic adjustment of the decision logic. When the external environment is stable, the progress benefits are calculated normally, while when there are drastic fluctuations in the external climate or working conditions, the hesitation is automatically increased and the progress weight is suppressed, tilting the decision towards conservative safety and risk avoidance, thus improving the robustness of decision-making under harsh construction environment conditions.

[0046] This invention breaks through the limitations of traditional water conservancy engineering question-and-answer systems, which can only perform static text retrieval and knowledge reorganization. By introducing a time-series deduction model and combining historical working conditions and control logs, it achieves forward-looking physical state deduction of the user's natural language construction operation intentions, upgrading safety question-and-answer from passive norm query to dynamic risk prediction. This invention constructs a constrained bipartite graph and uses a differentiable automatic differentiation framework for violation judgment and source tracing. It can not only quantify the extent of deviation between the predicted state and norm threshold, but also trace and locate the core operational parameters that lead to violations and their deterioration direction by calculating gradients. This invention uses a generated prohibition mask matrix combined with a heuristic pruning search tree to lock and block dangerous deterioration paths while automatically and efficiently finding the optimal alternative construction scheme that balances compliance and safety in the operational variable space. This invention also introduces decision hesitation based on environmental mutation index to perform a weighted comprehensive utility evaluation of construction progress benefits and accountability safety losses, and constructs a three-level dynamic diversion decision mechanism that can adapt to fluctuations in the on-site environment, thus providing intelligent decision-making assistance that balances safety and construction efficiency for complex and ever-changing water conservancy and hydropower construction sites.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, users of the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. An optimized method for handling safety questions in water conservancy and hydropower construction based on time-series extrapolation, characterized in that: Includes the following steps: Step S1: Collect basic working condition data and historical control logs from the construction site, perform normalization preprocessing, and obtain the historical working condition matrix and historical operation matrix. Step S2: Receive natural language question text and perform intent parsing, extract operation parameters to construct intent operation vector, and perform iterative deduction based on the time-series inference model combined with the historical working condition matrix, historical operation matrix and operation vector to obtain the predicted state matrix; Step S3: Construct a constrained bipartite graph based on the predicted state matrix to perform normative violation detection. When a violation is detected, perform gradient backtracking on the weights of the violation edges based on the differentiable automatic differentiation framework to locate the core operational parameters that cause the violation and their deterioration direction. Generate a prohibition mask matrix for directional prohibition of dangerous perturbation directions based on the backtracking results. Step S4: Construct a search tree based on the forbidden mask matrix, and search for compliant alternatives through heuristic pruning to obtain the optimal alternative construction sequence and its corresponding compliance and safety. Step S5: Perform a comprehensive utility assessment based on the compliance and security levels, and generate a response based on the comparison between the comprehensive utility assessment results and the decision threshold.

2. The optimization method for handling safety questions in water conservancy and hydropower construction based on time-series deduction as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Collect temperature data, flow rate data and wind speed data as the basic operating condition data, and extract discrete data points within a preset time window and align them according to timestamps to form the original operating condition matrix. Step S102: Extract the historical control log within the preset time window, parse the actual execution value of each operation parameter and form a historical operation column vector, and concatenate the historical operation column vectors of each time step in chronological order to form the original operation matrix. Step S103: Based on the preset operation parameter boundary table, normalization preprocessing is performed on each element in the original operating condition matrix and the historical operating matrix to obtain the historical operating condition matrix and the historical operating matrix respectively.

3. The optimization method for handling safety questions in water conservancy and hydropower construction based on time-series deduction as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S201: Input the question text into the pre-trained named entity recognition network, output entity labels, and parse the structured operation parameters from the text based on the entity labels; Step S202: After normalizing the parsed operation parameters, fill them into the corresponding dimension of the operation vector of the preset length to obtain the intention operation vector.

4. The optimization method for handling safety questions in water conservancy and hydropower construction based on time-series deduction as described in claim 3, characterized in that, Step S2 further includes the following sub-steps: Step S203: Vertically concatenate each column vector in the historical working condition matrix with the column vector of the corresponding time step in the historical operation matrix to form multiple historical composite state vectors, and input them sequentially into the time series inference model to activate network memory and obtain the initial memory state. The time-series extrapolation model includes an input layer, a stacked long short-term memory network layer, and a fully connected output layer.

5. The optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series deduction as described in claim 4, characterized in that, Step S2 also includes the following sub-steps: Step S204: Extract the last column vector of the historical working condition matrix and concatenate it with the operation vector to form a first composite state vector; Step S205: Using the time-series deduction model combined with the initial memory state, and taking the first composite state vector and the initial memory state as input, iteratively calculate the state prediction vectors for multiple future time steps, and horizontally concatenate the state prediction vectors of each time step to form the prediction state matrix.

6. The optimization method for handling safety questions in water conservancy and hydropower construction based on time-series deduction as described in claim 5, characterized in that, Step S3 includes the following sub-steps: Step S301: Construct the constrained bipartite graph, which includes a left node set, a right node set, and edges; The set of left-side nodes is constructed based on the physical state variables corresponding to each row of the predicted state matrix; The set of nodes on the right is constructed based on the normative constraint limits in the preset normative knowledge base; The process of constructing the edges specifically includes: If the physical state variable is subject to the specification constraint limit, then an edge is constructed between the corresponding left and right nodes, and the maximum out-of-bounds magnitude of the physical state variable exceeding the specification threshold is used as the weight of the edge. Step S302: Traverse all edges and perform violation determination. The logic of violation determination includes: if the weight is greater than zero, then a violation is determined to have occurred. Step S303: Based on a differentiable automatic differentiation framework, calculate the gradient of the edge weights that are determined to have violated the rules with respect to the intentional operation vector; Step S304: Extract the index of the operation variable with the largest absolute gradient value and the corresponding gradient value.

7. The optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series extrapolation as described in claim 6, characterized in that, Step S3 further includes the following sub-steps: Step S305: Construct a prohibition mask matrix, wherein the rows of the prohibition mask matrix represent operation parameters and the columns represent perturbation directions; Based on the sign of the gradient value, the elements in the prohibition mask matrix corresponding to the index of the operation variable and the perturbation direction are set as prohibition markers, and the remaining elements are set as allow markers.

8. The optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series deduction as described in claim 7, characterized in that, Step S4 includes the following sub-steps: Step S401: Construct a search tree with operation combinations as nodes and perturbation actions on operation parameters as branches; Step S402: Perform the first pruning, the second pruning, and the third pruning in sequence; The first pruning process includes: Traverse the preset set of discrete actions, read the prohibition mask matrix, and if the element corresponding to the current action is a prohibition flag, then prune and discard the branch; if it is a permission flag, then execute the action to generate the second operation vector. The second pruning process includes: A second composite state vector is constructed based on the second operation vector. The corresponding predicted state vector is obtained through a time series deduction model. The constraint bipartite graph is used to detect violations. If a violation is detected, the node is treated as a violation dead node and pruned. The third pruning process includes: Calculate the comprehensive score of the candidate nodes that have passed the first and second pruning, sort them in descending order of comprehensive score, and retain the preset number of nodes with the highest scores to enter the next level of expansion; Step S403: Obtain the compliance and security of the candidate node, wherein the compliance is the total number of consecutive time steps from the root node to the current node without triggering a violation of the specification; The security level is the sum of the number of deduction steps and the compliance level from the current node, keeping the current operation vector constant and extrapolating to the future until the first violation of the specification is triggered. The process of obtaining the comprehensive score specifically includes: The security level is input into a preset gating function to obtain a security discount factor. The security discount factor is then multiplied by the compliance level to obtain the comprehensive score.

9. The optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series deduction as described in claim 8, characterized in that, Step S5 includes the following sub-steps: Step S501: Based on the decision hesitation degree, the progress benefits and accountability losses are weighted and integrated to obtain the comprehensive utility; The schedule benefit is calculated based on the ratio of the compliance of the optimal alternative construction sequence to the expected total number of operation time steps; The accountability loss is calculated based on the safety of the optimal alternative construction sequence; The decision hesitation level is calculated based on the environmental mutation index; Step S502: Compare the overall utility with preset automatic release threshold and forced interception threshold; When the overall utility is greater than or equal to the automatic release threshold, a compliant release decision response is generated. When the overall utility is less than or equal to the mandatory interception threshold, a mandatory interception decision response is generated. When the overall utility falls between the automatic release threshold and the mandatory interception threshold, a manual decision response is generated.

10. The optimization method for handling safety questions and answers in water conservancy and hydropower construction based on time-series deduction as described in claim 9, characterized in that, The process of calculating the environmental mutation index specifically includes: Select a preset set of key sensors, and calculate the first-order difference sequence of the normalized sequence of each sensor in the set within the preset time window; The first-order difference sequence is subjected to moving average filtering, and the variance of the filtered sequence is calculated. The environmental mutation index is obtained by taking the mean of the variances of the filtered sequences of each sensor in the set.

Images