BIM-based ethylene plant hoisting operation control system and method

Abstract

The application relates to the technical field of pre-filled syringes, in particular to the technical field of hoisting strategy generation, and particularly relates to an ethylene device hoisting operation control system and method based on BIM. A BIM space simulation module performs digital modeling on an operation area, a hoisting target and a hoisting position, generates a BIM virtual operation space; a hoisting path generation module acquires a standardized data set Ds of the BIM virtual operation space, and generates an alternative path set Cr; a hoisting strategy generation module utilizes a neural network to perform hoisting state decomposition and screening on path elements in the alternative path set Cr, and outputs an optimal hoisting path and a hoisting state sequence. The closed-loop hoisting strategy generation is completed, the effective strategy generation of the overall hoisting operation is realized, the hoisting efficiency is improved, and the safety of the operation is strengthened.

Description

Technical Field

[0001] This application relates to the field of hoisting strategy generation technology, specifically to a BIM-based hoisting operation control system and method for ethylene plants. Background Technology

[0002] Ethylene plants are core equipment in the petrochemical industry. Their hoisting operations involve heavy equipment such as large towers, reactors, and heat exchangers, with individual units weighing hundreds of tons and exceeding 100 meters in height. The work environments are characterized by complex spatial constraints, dense obstacles, and stringent process requirements. During hoisting, the attitude control and path planning of the equipment directly determine the safety, efficiency, and economy of the operation, making them key bottlenecks in ethylene plant construction.

[0003] For a long time, the attitude control planning for hoisting operations of ethylene plants has mainly relied on manual experience, supplemented by basic spatial simulation of BIM (Building Information Modeling) for simple path planning. The overall operation planning has a low level of intelligence, the path planning is not effective enough and is disconnected from attitude control, resulting in insufficient overall operation efficiency and safety. Summary of the Invention

[0004] This application provides a BIM-based hoisting operation control system and method for ethylene plants to solve or partially solve the problems mentioned in the background art.

[0005] This application provides a BIM-based hoisting operation control system for ethylene plants, including a BIM spatial simulation module, a hoisting path generation module, and a hoisting strategy generation module. The functions of each module are as follows: The BIM space simulation module digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space. The hoisting path generation module obtains the standardized dataset Ds of the BIM virtual work space and generates a set of alternative paths Cr; The hoisting strategy generation module uses a neural network to decompose and filter the hoisting state of the path elements in the candidate path set Cr, and outputs the optimal hoisting path and hoisting state sequence. The hoisting state refers to the instantaneous spatial pose of the hoisting target at a certain moment during the hoisting process.

[0006] Preferably, the standardized dataset Ds includes collision detection information, which is obtained based on the Separating Axis Theorem (SAT).

[0007] Preferably, the hoisting path generation module uses an improved A* path generation algorithm to receive a standardized dataset Ds and generate a set of alternative paths Cr, including the following steps: Load differentiated preset constraints based on the hoisting target classification; Secondly, an improved cost function is designed, introducing a contour safety distance penalty term and an attitude change penalty term: In the formula, Let n be the total cost of the current node n. Let S be the cumulative three-dimensional path length from the starting point S to the current node n. Let n be the heuristic distance from node n to the endpoint T. This is a penalty term for attitude change, which is related to the cumulative change in attitude angle from the starting point S to the current node n. As a contour safety distance penalty term, it refers to the shortest distance between the lifting target contour surface and the obstacle when the lifting target posture corresponds to the minimum angle change from the previous node to node n. , , The penalty item weighting coefficient is set with different values ​​based on different hoisting types; When executing the search loop, after finding the endpoint T, only the path is saved without terminating the loop. The remaining nodes of the OpenList are processed to obtain all possible paths. A preset number of alternative paths are obtained through differential preset constraints and the principle of minimizing total cost, and a set of alternative paths Cr is generated.

[0008] Preferably, the hoisting strategy generation module is a neural network model based on a CNN-Transformer fusion architecture. It takes the standardized dataset Ds from the BIM space simulation module, the alternative path set Cr from the hoisting path generation module, and the environmental parameters of the work area as input, and directly outputs the optimal hoisting state sequence and the optimal hoisting path end-to-end. It includes: an input layer, a path-node encoding layer, an attention layer, a dual-branch feature fusion layer, and a state sequence generation layer. The functions of each layer are as follows: The input layer standardizes the input feature data and concatenates features according to the path node dimension, outputting the original feature set. (i,j) represents the j-th node of the i-th path; The path-node encoding layer adds a path attribution identifier to each node and outputs node features carrying path attribution information. ; The attention layer extracts spatial and temporal features of nodes within the same path, masks feature interactions between nodes across paths using an attention mask, and outputs a global node feature set. ; The branch feature fusion layer extracts spatial matching features between nodes and hoisting states through CNN branches, and extracts temporal optimization features of the state sequence through Transformer branches. The two branch features are fused to output a fused feature vector. ; The state sequence generation layer maps the fused features into a sequence of pose angles and assigns each state to a node on the same path, outputting the optimal state sequence and the optimal path.

[0009] Preferably, the input layer first removes the dimensions and standardizes each feature parameter. The standardization algorithm uses the Z-Score standardization algorithm, and the standardized features are concatenated according to the path node dimension to obtain the original feature set. .

[0010] Preferably, the path-node encoding layer uses one-hot encoding and feature fusion to add a path attribution identifier to each node, outputting node features carrying path attribution information. The mathematical expression is as follows: In the formula, Let be the one-hot encoded vector of the i-th path, with the dimension being the number of candidate paths. Emb is a learnable fully connected layer used to map the low-dimensional one-hot encoding to a high-dimensional attribution embedding vector. Let be the embedding vector belonging to the i-th path, be the unique identifier of the path, and ⊕ represent the element-wise addition operation. Let j represent the j nodes on the i-th alternative path.

[0011] Preferably, the attention layer includes a path space coding sublayer and a perceptual attention sublayer; The path spatial coding sublayer performs 1D-CNN coding on the node features within a single path, extracting the spatially continuous features of nodes within the path. Different paths are coded independently, outputting the coded feature set of all paths. }; The perception attention sublayer uses standardized BIM features. As the query vector, the CNN-encoded features along the same path are used as key / value vectors. Path-attribute attention masks (MIDs) are used to mask cross-path attention interactions, ultimately outputting a global node feature set. The mathematical expression is as follows: In the formula, Q is the query vector. Let i be the key / value vector of the i-th path. For a single-head dimension of multi-head attention, the path-attribute attention mask (MID) is determined by the following rule: if the node for which attention is calculated matches the query vector across paths, the value is... ∞, the value of which is 0 on the same path. Let be the affiliation-aware attention feature of the i-th path.

[0012] Preferably, the CNN branch in the branch feature fusion layer calculates candidate pose angles that satisfy collision constraints for each node, ensuring that the device's contour at that node matches the spatial constraints; The Transformer branch models the switching logic between adjacent states and autonomously learns the temporal logic characteristics. A gated feature fusion mechanism is set in the branch feature fusion layer to allow the model to automatically adjust the contribution of spatial matching features and temporal optimization features. The mathematical expression is as follows: In the formula, and For gated learnable weight matrices and biases, σ is used to replace the sigmoid activation function. The fused feature vector represents the final output. and These are the output feature vectors of the two branches, respectively.

[0013] Preferably, the state sequence generation layer first maps the fused features into a pose angle sequence matching the task length through two fully connected layers, as follows: In the formula, ReLU represents the activation function. , These are the bias vectors corresponding to the two fully connected layers; Secondly, the generated attitude angle sequence is spatially similar to the nodes of all candidate paths to obtain the original path to which the state sequence belongs. The mathematical expression is as follows: In the formula, The coordinates of the node to which the t-th attitude angle is bound are represented. The attitude angle is bound to the node on the same path by the relationship between the node and the path and the order of the path nodes.

[0014] This application also provides a BIM-based method for controlling hoisting operations of ethylene plants, including the following steps: S100: BIM Space Simulation Module: Digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space; S200: Lifting path generation module, obtains the standardized dataset Ds of the BIM virtual work space, and generates a set of alternative paths Cr; S300: Lifting Strategy Generation Module: Utilizes a neural network to decompose and filter the lifting state of the path elements in the candidate path set Cr, and outputs the optimal lifting path and lifting state sequence.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: (1) This application obtains basic data through BIM spatial modeling, then performs path planning through the hoisting path generation module, and then performs hoisting state (attitude control) decomposition and optimization of the alternative paths through the hoisting strategy generation module, thus completing the closed-loop hoisting strategy generation, realizing the effective strategy generation for the overall hoisting operation, improving hoisting efficiency, and enhancing the safety of the operation.

[0016] (2) This application generates multiple alternative paths with a modified A* path generation algorithm, abandoning the shortest path principle in path optimization, and provides rich alternative samples and spatial constraints for subsequent neural network optimization of hoisting state; through a neural network model based on CNN-Transformer fusion architecture, the optimal hoisting state sequence and optimal hoisting path are directly output end-to-end. The model integrates the spatial feature extraction capability of CNN and the sequence generation and sequential dependency modeling capability of Transformer, and learns autonomously to balance collision safety, attitude switching efficiency and path smoothness, and generates the optimal hoisting state sequence end-to-end, realizing effective planning of hoisting strategy. Attached Figure Description

[0017] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a schematic diagram of the system composition of this application. Figure 2 This is a schematic diagram of the method flow of this application. Detailed Implementation

[0019] The specification and claims use certain terms to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The term "comprising" throughout the specification and claims is an open-ended term and should be interpreted as "comprising but not limited to." "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error.

[0020] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0021] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0022] Example 1 like Figures 1 to 2 As shown, this application provides a BIM-based hoisting operation control system for ethylene plants, including a BIM spatial simulation module, a hoisting path generation module, and a hoisting strategy generation module. The functions of each module are as follows: The BIM space simulation module digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space. The hoisting path generation module obtains the standardized dataset Ds of the BIM virtual work space and generates a set of alternative paths Cr; The hoisting strategy generation module uses a neural network to decompose and filter the hoisting states of the path elements in the candidate path set Cr, and outputs the optimal hoisting path and hoisting state sequence.

[0023] Specifically, the BIM spatial simulation module provides 1:1 spatial location data for the entire system, completing the detailed contour modeling of the work area boundary, buildings or other obstacles, hoisting targets (various ethylene production line accessories), and hoisting positions (the starting and ending poses of the hoisting target equipment), providing digital support for subsequent path and state generation.

[0024] Specifically, the standardized dataset Ds of the BIM virtual work space obtained by the hoisting path generation module includes: global coordinate information of the work area, coordinates of obstacle node areas, hoisting target outline information, hoisting position, collision detection information, etc. Among them, except for the collision detection information, all can be directly obtained from the BIM virtual work space. The collision detection information is obtained based on the Separated Axis Theorem (SAT), and the specific method is as follows: S201: Construct OBB bounding boxes for lifting targets and obstacles in the work area; S202: Extract potential separation axes for each combination of lifting targets and obstacles, including the bounding box axes of 3 lifting targets, the bounding box axes of 3 obstacles, and the cross-product separation axes of 9 axes; S203: For each separation axis, calculate whether the projections of the two bounding boxes on the separation axis have an intersection area. If the projections on any separation axis do not have an intersection, it is determined that the corresponding hoisting target and obstacle combination has no collision. If there is a separation axis with a non-empty projection intersection, it is determined that the corresponding combination has a collision risk. Record the relevant collision detection results and relevant collision detection information such as the bounding box coordinates of the obstacle, the collision separation axis, and the feasible space boundary.

[0025] The final standardized dataset Ds is stored and transmitted in JSON format. The hoisting path generation module uses an improved A* path generation algorithm to receive the standardized dataset Ds and generate a set of alternative paths Cr. It optimizes the cost function and shrinkage logic and outputs multiple alternative paths that meet the requirements of equipment differentiation, collision-free operation, and short path length.

[0026] First, the hoisting path generation module loads differentiated preset constraints based on the hoisting target classification. The hoisting targets are classified based on dimensions such as weight, size, and attitude sensitivity. To ensure that the hoisting path conforms to the actual hoisting requirements of the equipment, hard constraints are set for each type of hoisting target. For example, different constraints are set for large key equipment such as towers and reactors, medium-sized general equipment such as heat exchangers and separators, and small auxiliary equipment such as pipe fittings and pump bodies, based on dimensions such as total path length, Z-axis height, number of continuous rotation nodes, single rotation angle, and channel width.

[0027] Secondly, an improved cost function is designed, introducing a contour safety distance penalty term and a pose change penalty term, so that the algorithm can balance path length and safety during the search process. The improved cost function is as follows: In the formula, Let n be the total cost of the current node n. Let S be the cumulative three-dimensional path length from the starting point S to the current node n. This is the heuristic distance (3D Euclidean distance) from node n to the endpoint T. It has no actual cost and only guides the search direction. This is a penalty term for attitude change, which is related to the cumulative change in attitude angle from the starting point S to the current node n. As a contour safety distance penalty term, it refers to the shortest distance between the lifting target contour surface and the obstacle when the lifting target posture corresponds to the minimum angle change from the previous node to node n. , , The penalty item weighting coefficient is set with different values ​​based on the different hoisting types.

[0028] Specifically, posture change penalty item The typical mathematical expression is as follows: Where m is the number of path nodes from the starting point to node n. This represents the maximum allowable cumulative change in attitude angle of the hoisting target. , , The three-dimensional attitude angles of node k (corresponding to the x, y, and z axes); Outline safety distance penalty The typical mathematical expression is as follows: In the formula, The standard threshold for the outline safety distance of the corresponding type of lifting target. Let ε be the shortest distance between the target contour surface and the obstacle at node n, and let ε be the minimum value to prevent the denominator from being 0.

[0029] The improved cost function balances path length and safety, adding two independent penalty terms to be finely tuned separately for "obstacle avoidance" and "attitude control". The contour penalty term retains the basic smoothness and strong constraint logic, while the attitude penalty term directly quantifies the cumulative change. This ensures the flexibility of the algorithm search and avoids the dual high risks of near obstacles and large attitude changes from the root.

[0030] Finally, the execution steps of the improved A* path generation algorithm are designed as follows: Algorithm initialization: Extract the collision-free feasible space from the standardized dataset Ds, extract the branch node set N from it, initialize the open list OpenList, the closed list CloseList, and the parent node mapping table to be empty, and initialize the cumulative actual cost g(S) of the starting point S to 0; Loop search: S211: Select the node with the smallest total cost f(n) from OpenList as the current node n, move the current node n to CloseList and mark it as traversed. If the current node n is the destination T, perform path backtracking, generate and save a complete feasible path, but do not break out of the loop and continue to process the remaining nodes in OpenList. S212: Traverse all valid adjacent nodes m of the current node n, and execute the following steps in sequence: If m is already in CloseList and is not the endpoint T, skip directly; if it is the endpoint T, proceed with the following steps. Calculate the actual 3D distance d(n, m) from node n to m; If m is not in OpenList, or g(m){temp} < the recorded g(m): Update the cumulative actual cost g(m) = g(m){temp}, set the parent node parent(m) = n, calculate the total cost f(m), and add m to OpenList; Repeat the above steps until OpenList is empty; Path backtracking starts from the endpoint T and backtracks to the starting point S through the parent node mapping table to obtain all possible paths. A preset number of alternative paths are obtained through differential preset constraints and the principle of minimizing total cost, generating a set of alternative paths Cr.

[0031] Taking a node set N={A, B, C, D, E} as an example, with a starting point S=A, an ending point T=E, and C and D being parallel branch nodes, during the initialization phase, OpenList = {A}, CloseList = g(A)=0, the parent node mapping table is empty; First processing: Take A with the smallest f value in OpenList, move A into CloseList, traverse the valid adjacent nodes of A: only B, calculate g(B) = g(A) + d(A, B) of B, then calculate h(B) and penalty term to get f(B), calculate g(B) = g(A) + d(A→B) of B, then calculate h(B) and penalty term to get f(B), add B to OpenList; Second processing: Take B with the smallest f value in OpenList, move B into CloseList, traverse the valid adjacent nodes of B: C and D, calculate the g value, h value and f value of C and D respectively, and add C and D to OpenList in turn; The third processing step: Take the C with the smallest f value in OpenList (assuming that C's f value is smaller than D's), move C into CloseList, traverse the effective adjacent nodes of C: E (the end point), calculate the g value, h value (0), and f value of E, and add E to OpenList; Fourth processing: Take the E with the smallest f value in OpenList, determine that E is the endpoint, first backtrack to generate the first path A→B→C→E, do not terminate the global search (the key to improving A*), and continue to process the remaining nodes in OpenList; Fifth processing step: Take the only remaining D in OpenList, traverse the valid adjacent nodes of D: E (although it is in CloseList, it does not affect it and is still used as the endpoint), calculate the temporary g value of E, generate an independent parent node mapping based on D, backtrack to generate the second path A→B→D→E, and move D into CloseList.

[0032] After finding the destination T, the hoisting path generation module of this application only saves the path without terminating the loop, and continues to process the remaining nodes of the OpenList to generate multiple differentiated feasible paths. This provides multiple path outputs for subsequent multi-objective optimization, and an improved cost function is designed to balance path length and safety.

[0033] Specifically, the hoisting strategy generation module is a neural network model based on a CNN-Transformer fusion architecture. It takes the standardized dataset Ds from the BIM space simulation module, the alternative path set Cr from the hoisting path generation module, and the environmental parameters of the work area as inputs. It directly outputs the optimal hoisting state sequence (main output) and the optimal hoisting path (secondary output) from end to end. The model integrates the spatial feature extraction capability of CNN and the sequence generation and sequential dependency modeling capability of Transformer. It autonomously learns to balance collision safety, attitude switching efficiency, and path smoothness, and generates the optimal hoisting state sequence from end to end. The single alternative path to which it is attached is the optimal path.

[0034] In this application, the hoisting state is defined as the instantaneous spatial pose of the hoisting target at a certain moment during the hoisting process. Changes in the hoisting state are associated with the start point, end point, and branch nodes of the hoisting path, and are mathematically expressed as follows: In the formula, Let i represent the hoisting status of the j-th node in the i-th alternative path. , , This represents the three-dimensional coordinates of the hoisting equipment at the corresponding node. , , The pitch angle, yaw angle, and roll angle of the hoisting target at the corresponding switching node are rotated around the x, y, and z axes, respectively.

[0035] Each alternative route can be broken down into an ordered sequence of multiple hoisting states based on time and work order. Unlike other path optimization problems, the length of the hoisting path is not the core of the overall hoisting operation process. Instead, the safety and overall efficiency associated with the state transition of the hoisting target are more important. Therefore, this application provides differentiated multi-path alternatives that take into account path length, contour safety and attitude change, and then selects the hoisting path with the optimal state sequence.

[0036] Specifically, the hoisting strategy generation module includes an input layer, a path-node encoding layer, an attention layer, a dual-branch feature fusion layer, and a state sequence generation layer. The functions of each layer are as follows: The input layer standardizes and concatenates the input feature data. The path-node encoding layer adds a path affiliation identifier to each node to ensure that the state sequence belongs to the same path. The attention layer extracts the spatial and temporal features of nodes within the same path and uses an attention mask to shield the feature interactions between nodes across paths. The branch feature fusion layer extracts spatial matching features between nodes and hoisting states through CNN branches and extracts temporal optimization features of the state sequence through Transformer branches. The two branch features are then fused and output. The state sequence generation layer maps the fused features into a sequence of pose angles and assigns each state to a node on the same path, outputting the optimal state sequence and the optimal path.

[0037] Specifically, the input layer receives the standardized dataset Ds from the BIM spatial simulation module, the alternative path set Cr from the hoisting path generation module, and the environmental parameters of the work area. It then extracts the necessary feature parameters, mainly including spatial constraint features, path features, and environmental features. Spatial constraint features include hoisting target parameters (outline, center of gravity, etc.) and hoisting requirements (attitude angle variation range, height threshold, etc.). Path features include path number, path node sequence, path spatial range, and obstacle coordinates. Environmental features mainly include wind parameters. The input layer first eliminates the dimensions of each feature parameter and standardizes them. The standardization algorithm uses the Z-Score standardization algorithm, concatenating the standardized features according to the path node dimension to obtain the original feature set. .

[0038] Specifically, the path-node encoding layer uses one-hot encoding and feature fusion to add a path attribution identifier to each node, outputting node features carrying path attribution information. The mathematical expression is as follows: In the formula, Let be the one-hot encoded vector of the i-th path, with the dimension being the number of candidate paths. Emb is a learnable fully connected layer used to map the low-dimensional one-hot encoding to a high-dimensional attribution embedding vector. Let be the embedding vector of the i-th path, be the unique identifier of the path, and ⊕ represent the element-wise addition operation.

[0039] Specifically, the attention layer includes a path space encoding sublayer and a perceptual attention sublayer. The path space encoding sublayer is used for feature extraction of nodes along the same path, while the perceptual attention sublayer is used for masking features across paths.

[0040] Furthermore, the path spatial coding sublayer performs 1D-CNN coding on the node features within a single path, extracting the spatially continuous features of nodes within the path (such as node spacing and obstacle distance variations). Different paths are coded independently, outputting the coded feature set for all paths.}

[0041] Furthermore, the perception attention sublayer standardizes BIM features. As the query vector, the CNN-encoded features along the same path are used as key / value vectors. Path-attribute attention masks (MIDs) are used to mask cross-path attention interactions, ultimately outputting a global node feature set. The mathematical expression is as follows: In the formula, Q is the query vector. Let i be the key / value vector of the i-th path. For a single-head dimension of multi-head attention, the path-attribute attention mask (MID) is determined by the following rule: if the node for which attention is calculated matches the query vector across paths, the value is... ∞, the value of which is 0 on the same path. Let be the affiliation-aware attention feature of the i-th path.

[0042] It is important to note that standardized BIM features This refers to the spatial constraint features in the input layer, after being encoded by a CNN. The global spatial constraints are localized, losing their role as a global unified anchor point. This is addressed by introducing [a new feature] into the perceptual attention sublayer. The query vector essentially addresses the problem caused by global feature duplication. The attention calculation process involves matching global requirements with local path features. The model calculates Q and... The similarity is used to select the local path features that best fit the global constraints. If an independent Q is not introduced, the model can only perform self-attention within the local path features, which is equivalent to matching local features with local features and cannot anchor the global constraints.

[0043] Specifically, the CNN branch in the branch feature fusion layer calculates candidate pose angles that satisfy collision constraints for each node, ensuring that the device's contour at that node matches the spatial constraints. The mathematical expression is as follows: In the formula, It is a 1D convolutional layer with a kernel size of 3. For bias vectors, This represents the spatial matching feature of the j-th node in the i-th path. This is the candidate attitude angle vector for the j-th node of the i-th path, providing attitude options for state generation.

[0044] Specifically, the Transformer branch models the switching logic between adjacent states and autonomously learns temporal logic features, such as small-amplitude pose switching (high efficiency) and large obstacle spacing (low risk), which can be expressed mathematically as follows: In the formula, Attention represents the multi-head attention mechanism. The constraint mask is the sum of the lifting process constraints for the corresponding lifting target, including the attitude switching range, collision risk (distance between the outline and obstacles) mask, operation sequence mask (such as the adjustment order of attitude due to process constraints), and path attribution mask. The decoder is a Transformer decoder, and the final output of the formula is the temporal feature optimization vector. .

[0045] Specifically, a gated feature fusion mechanism is set in the branch feature fusion layer to allow the model to automatically adjust the contribution of spatial matching features and temporal optimization features, which is expressed mathematically as follows: In the formula, and For gated learnable weight matrices and biases, σ is used to replace the sigmoid activation function. This represents the fused feature vector of the final output.

[0046] Specifically, the state sequence generation layer first maps the fused features into a pose angle sequence that matches the task length through two fully connected layers, as follows: In the formula, ReLU represents the activation function. , These are the bias vectors corresponding to the two fully connected layers; Secondly, the generated attitude angle sequence is spatially similar to the nodes of all candidate paths to obtain the original path to which the state sequence belongs. The mathematical expression is as follows: In the formula, The coordinates of the node to which the t-th attitude angle is bound are represented. The attitude angle is bound to the node on the same path by the relationship between the node and the path and the order of the path nodes.

[0047] It should be noted that the length of the attitude angle sequence is a fixed value set based on the maximum number of nodes of all alternative paths and the process requirements of the corresponding hoisting target. That is, the number of attitude angles in the attitude angle sequence is greater than or equal to the number of path nodes to which they are bound. The excess attitude angles are evenly inserted between the path nodes according to their order in the sequence.

[0048] The essence of the state sequence generation layer in this application is a constraint-driven path-state selection model, rather than an unconstrained new path generation model. The core goal is not to create new paths, but to select the optimal hoisting state sequence that can be directly executed from the existing set of candidate paths by learning spatial constraints, process requirements and efficiency goals. When setting the loss function, it is necessary to take into account safety loss (collision loss), state switching loss and state continuity loss through weights.

[0049] Example 2 Based on Example 1, this application also provides a BIM-based method for controlling the hoisting operation of an ethylene plant, which specifically includes the following steps: S100: BIM Space Simulation Module: Digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space; S200: Lifting path generation module, obtains the standardized dataset Ds of the BIM virtual work space, and generates a set of alternative paths Cr; S300: Lifting Strategy Generation Module: Utilizes a neural network to decompose and filter the lifting state of the path elements in the candidate path set Cr, and outputs the optimal lifting path and lifting state sequence.

[0050] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A BIM-based hoisting operation control system for an ethylene plant, characterized in that, It includes a BIM space simulation module, a hoisting path generation module, and a hoisting strategy generation module. The functions of each module are as follows: The BIM space simulation module digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space. The hoisting path generation module obtains the standardized dataset Ds of the BIM virtual work space and generates a set of alternative paths Cr; The hoisting strategy generation module uses a neural network to decompose and filter the hoisting state of the path elements in the candidate path set Cr, and outputs the optimal hoisting path and hoisting state sequence. The hoisting state refers to the instantaneous spatial pose of the hoisting target at a certain moment during the hoisting process; The hoisting strategy generation module is a neural network model based on a CNN-Transformer fusion architecture. It takes the standardized dataset Ds from the BIM space simulation module, the alternative path set Cr from the hoisting path generation module, and the environmental parameters of the work area as input. It directly outputs the optimal hoisting state sequence and the optimal hoisting path end-to-end. The module includes: an input layer, a path-node encoding layer, an attention layer, a dual-branch feature fusion layer, and a state sequence generation layer. The functions of each layer are as follows: The input layer standardizes the input feature data and concatenates features according to the path node dimension, outputting the original feature set. (i,j) represents the j-th node of the i-th path; The path-node encoding layer adds a path attribution identifier to each node and outputs node features carrying path attribution information. ; The attention layer extracts spatial and temporal features of nodes within the same path, masks feature interactions between nodes across paths using an attention mask, and outputs a global node feature set. ; The branch feature fusion layer extracts spatial matching features between nodes and hoisting states through CNN branches, and extracts temporal optimization features of the state sequence through Transformer branches. The two branch features are fused to output a fused feature vector. ; The state sequence generation layer maps the fused features into a sequence of pose angles and assigns each state to a node on the same path, outputting the optimal state sequence and the optimal path.

2. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The standardized dataset Ds includes collision detection information, which is obtained based on the Separating Axis Theorem (SAT).

3. The BIM-based hoisting operation control system for ethylene plants according to claim 2, characterized in that: The hoisting path generation module uses the improved A* path generation algorithm to receive the standardized dataset Ds and generate a set of alternative paths Cr, including the following steps: Load differentiated preset constraints based on the hoisting target classification; Secondly, an improved cost function is designed, introducing a contour safety distance penalty term and an attitude change penalty term: In the formula, Let n be the total cost of the current node n. Let S be the cumulative three-dimensional path length from the starting point S to the current node n. Let n be the heuristic distance from node n to the endpoint T. This is a penalty term for attitude change, which is related to the cumulative change in attitude angle from the starting point S to the current node n. As a contour safety distance penalty term, it refers to the shortest distance between the lifting target contour surface and the obstacle when the lifting target posture corresponds to the minimum angle change from the previous node to node n. , , The penalty item weighting coefficient is set with different values ​​based on different hoisting types; When executing the search loop, after finding the endpoint T, only the path is saved without terminating the loop. The remaining nodes of the OpenList are processed to obtain all possible paths. A preset number of candidate paths are obtained through differential preset constraints and the principle of minimizing total cost, and a candidate path set Cr is generated.

4. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The input layer first removes the dimensions and standardizes each feature parameter. The standardization algorithm uses the Z-Score standardization algorithm, and the standardized features are concatenated according to the path node dimension to obtain the original feature set. .

5. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The path-node encoding layer uses one-hot encoding and feature fusion to add a path attribution identifier to each node, outputting node features carrying path attribution information. The mathematical expression is as follows: In the formula, Let be the one-hot encoded vector of the i-th path, with the dimension being the number of candidate paths. Emb is a learnable fully connected layer used to map the low-dimensional one-hot encoding to a high-dimensional attribution embedding vector. Let be the embedding vector belonging to the i-th path, be the unique identifier of the path, and ⊕ represent the element-wise addition operation. Let j represent the j nodes on the i-th alternative path.

6. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The attention layer includes a path space coding sublayer and a perceptual attention sublayer; The path spatial coding sublayer performs 1D-CNN coding on the node features within a single path, extracting the spatially continuous features of nodes within the path. Different paths are coded independently, outputting the coded feature set of all paths. }; The perception attention sublayer uses standardized BIM features. As the query vector, the CNN-encoded features along the same path are used as key / value vectors. Path-attribute attention masks (MIDs) are used to mask cross-path attention interactions, ultimately outputting a global node feature set. The mathematical expression is as follows: In the formula, Q is the query vector. Let i be the key / value vector of the i-th path. For a single-head dimension of multi-head attention, the path-attribute attention mask (MID) is determined by the following rule: if the node for which attention is calculated matches the query vector across paths, the value is... ∞, the value of the same path is 0. Let be the affiliation-aware attention feature of the i-th path.

7. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The CNN branch in the branch feature fusion layer calculates the candidate pose angle that satisfies the collision constraint for each node, ensuring that the device's contour at that node matches the spatial constraints; The Transformer branch models the switching logic between adjacent states and autonomously learns the temporal logic characteristics. A gated feature fusion mechanism is set in the branch feature fusion layer to allow the model to automatically adjust the contribution of spatial matching features and temporal optimization features. The mathematical expression is as follows: In the formula, and For gated learnable weight matrices and biases, σ is used to replace the sigmoid activation function. The fused feature vector represents the final output. and These are the output feature vectors of the two branches, respectively.

8. The BIM-based hoisting operation control system for ethylene plants according to claim 1, characterized in that: The state sequence generation layer first maps the fused features into a pose angle sequence that matches the task length through two fully connected layers, as follows: In the formula, ReLU represents the activation function. , These are the bias vectors corresponding to the two fully connected layers; Secondly, the generated attitude angle sequence is spatially similar to the nodes of all candidate paths to obtain the original path to which the state sequence belongs. The mathematical expression is as follows: In the formula, The coordinates of the node to which the t-th attitude angle is bound are represented. The attitude angle is bound to the node on the same path by the relationship between the node and the path and the order of the path nodes.

9. A BIM-based method for controlling hoisting operations in ethylene plants, characterized in that, Includes the following steps: S100: BIM Space Simulation Module: Digitally models the work area, hoisting target, and hoisting location to generate a BIM virtual work space; S200: Lifting path generation module, obtains the standardized dataset Ds of the BIM virtual work space, and generates a set of alternative paths Cr; S300: Lifting Strategy Generation Module: Uses a neural network to decompose and filter the lifting state of the path elements in the candidate path set Cr, and outputs the optimal lifting path and lifting state sequence. In step S300, the hoisting strategy generation module takes the standardized dataset Ds from the BIM space simulation module, the alternative path set Cr from the hoisting path generation module, and the environmental parameters of the work area as input, and outputs the optimal hoisting state sequence and the optimal hoisting path end-to-end, as follows: The input layer standardizes the input feature data and concatenates features according to the path node dimension, outputting the original feature set. (i,j) represents the j-th node of the i-th path; The path-node encoding layer adds a path attribution identifier to each node and outputs node features carrying path attribution information. ; The attention layer extracts spatial and temporal features of nodes within the same path, masks feature interactions between nodes across paths using an attention mask, and outputs a global node feature set. ; The branch feature fusion layer extracts spatial matching features between nodes and hoisting states through CNN branches, and extracts temporal optimization features of the state sequence through Transformer branches. The two branch features are fused to output a fused feature vector. ; The state sequence generation layer maps the fused features into an attitude angle sequence and assigns each state to a node on the same path, outputting the optimal state sequence and the optimal path.

Images