Large tower crane task scheduling time prediction method based on gated multi-attribute transformer
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN MUNICIPAL CONSTR GROUP
- Filing Date
- 2026-05-06
- Publication Date
- 2026-08-07
AI Technical Summary
当前塔吊调度多依赖人工经验，存在任务执行时间预估偏差大、设备空载与等待率高、异构工况下特征融合效果差等问题；传统的时序预测模型难以有效捕获塔吊运行过程中设备状态、环境因素、任务属性等多维度异构特征的长短期依赖关系，且在样本有限、工况多变的实际场景中，模型精度和泛化能力不足，无法实现塔吊任务时间的端到端精准预测，制约了大型塔吊设备的综合使用效率提升
本发明设计的门控多属性融合模块通过时序演化门控、跨属性注意力门控和双层残差融合的三级门控机制，实现了异构特征的自适应动态加权融合，有效解决了多属性特征耦合性差的问题，输出稳定的高层特征；
Smart Images
Figure CN122529286A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to artificial intelligence technology, and more particularly to a method for predicting the scheduling time of large tower crane tasks based on a gated multi-attribute Transformer. Background Technology
[0002] Large tower cranes are core heavy equipment in construction, port logistics, and other scenarios. Their utilization efficiency and scheduling rationality directly affect the progress of construction projects, equipment operating costs, and operational safety. Currently, tower crane scheduling relies heavily on manual experience, resulting in problems such as large deviations in task execution time estimation, high equipment idle and waiting rates, and poor feature fusion effects under heterogeneous operating conditions. Traditional time series prediction models struggle to effectively capture the long- and short-term dependencies of multi-dimensional heterogeneous features such as equipment status, environmental factors, and task attributes during tower crane operation. Furthermore, in real-world scenarios with limited samples and variable operating conditions, the model accuracy and generalization ability are insufficient, failing to achieve end-to-end accurate prediction of tower crane task times, thus hindering the improvement of the overall utilization efficiency of large tower crane equipment.
[0003] To address the aforementioned issues, there is an urgent need for a method to predict the utilization efficiency and scheduling time of large tower cranes. This method should accurately predict the execution time of tower crane tasks under complex working conditions and generate scheduling optimization instructions, thereby significantly reducing equipment idle and waiting rates and improving the overall utilization efficiency of large tower cranes. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer, which addresses the deficiencies in the prior art.
[0005] The technical solution adopted by this invention to solve its technical problem is: a method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer, comprising the following steps: 1) Construct a multi-attribute operation dataset for tower cranes as the training set for the model. The multi-attribute operation dataset includes equipment status parameters, environmental data, task attributes, and historical time series data. 2) Design a gated multi-attribute fusion method, and dynamically weight fused heterogeneous features through an adaptive weight allocation mechanism; The heterogeneous features include device status features, environmental features, task attribute features, and historical time series features; Historical temporal characteristics are the temporalized set of the above three types of characteristics in the historical operation process; 3) Construct a Transformer temporal coding network, input gated fusion features into the temporal coding network to capture long-short-term dependencies, and output deep temporal representations to predict the execution time of the tower crane's current task; 4) Use validation data to perform online evaluation and incremental fine-tuning of the prediction model until it meets the preset accuracy threshold; 5) Output the task execution time prediction results and generate scheduling optimization instructions.
[0006] According to the above scheme, step 1) includes: 1.1) Real-time collection of multi-dimensional operational data, including lifting capacity, slewing angle, wind speed, task floor, operation type, and historical task completion time; 1.2) The collected raw data is timestamped to form a time-series record with a uniform frequency; 1.3) The time series records are embedded and encoded according to the discrete task attributes, and the continuous state variables are segmented, quantized, and concatenated into a multi-dimensional tensor to form a multi-attribute time series sample. 1.4) Divide the multi-attribute time series samples into training set, validation set and test set according to time order, and keep the working condition distribution of each set consistent to ensure the effectiveness of model training.
[0007] According to the above scheme, in step 1.3), the discrete task attributes include the following categories: Job types: including different categories of work actions such as material lifting, equipment transfer, empty hook return, and floor unloading; Material types: This includes categories of materials that are hoisted, such as steel bars, concrete, scaffolding, precast building components, and construction equipment. Work area: This includes the markings for the construction area, main building, podium building, and work surface; Operating modes: These include tower crane operating modes such as manual operation, semi-automatic scheduling, and emergency operation. Task Priority: This refers to the priority level of tasks.
[0008] According to the above scheme, the gated multi-attribute fusion in step 2) completes the mapping from the original multi-attribute sequence to the unified prediction representation in three steps, specifically including: 2.1) Temporal evolution gating: The temporal gate is independently calculated for each feature attribute to determine the weight allocation of the current observation and historical observation, generating an attribute-specific evolutionary representation; 2.2) Cross-attribute attention gating: Based on scaled dot product attention, weights are assigned among all attributes to obtain a unified representation of cross-attribute coupling. The originally scattered and different-dimensional attribute features (such as wind speed being numerical and job type being categorical) are transformed into feature vectors with a unified format and containing cross-attribute association information. 2.3) The dual-layer residual fusion output performs gating fusion again with the cross-attribute representation and the hidden state of the previous time step, and introduces a regularization term to prevent gating extremes, outputting stable high-level features.
[0009] According to the above scheme, the Transformer temporal coding network constructed in step 3) is the Aug-Transformer model. The Aug-Transformer model is based on the standard Transformer architecture and improves feature diversity by adding enhanced shortcut paths. The overall structure of the Aug-Transformer model is a deep stacked encoder, including an input layer, a core module, and an output layer. The core modules are the AugMSA module and the AugMLP module.
[0010] According to the above scheme, step 3) includes the following sub-steps: 3.1) Input sequence representation: The time series is divided into fixed-length windows. Each time point or window segment is projected onto a d-dimensional embedding vector through a linear projection layer, forming the input sequence. , where N is the sequence length and d is the feature dimension; 3.2) Construct the AugMSA module. The AugMSA module adds an enhanced shortcut path to the standard MSA module. MSA captures the temporal dependencies within a sequence, while the enhanced shortcut path provides an additional path to bypass MSA and directly pass and transform features. 3.3) Construct the AugMLP module; the AugMLP module adds enhanced shortcut paths to the standard MLP; 3.4) Use block circular matrices to implement enhanced shortcut paths; 3.5) The model is obtained by stacking L layers. Each layer contains AugMSA and AugMLP modules. The output layer adopts a fully connected regression layer. The [CLS] position of the output feature of the last layer is taken, and the predicted value of the tower crane task execution time is generated through the fully connected layer.
[0011] According to the above scheme, in step 3.2), the standard MSA is represented as: ; ; in, H represents the input features of the l-th layer; H is the number of attention heads. It is a query, key, and value projection matrix; It is an attention map that captures the relationships between points in time; The AugMSA module is represented as follows: ; in, It is the i-th enhanced shortcut path, defined as:
[0012] It is an activation function. It is a learnable weight matrix.
[0013] According to the above scheme, in step 3.4), an enhanced shortcut path is implemented using a block circular matrix, as detailed below: The dense weight matrix is decomposed into block cyclic matrices. Utilizing the mathematical properties of cyclic matrices, matrix multiplication is transformed into element-wise multiplication through Fourier transform, reducing the computational complexity to O(dlogd). The projection complexity is also reduced through Fourier transform, making it suitable for high-dimensional time series data. The original dense matrix is decomposed into multiple cyclic submatrices, and the computation is performed in the frequency domain. weight matrix Decompose into block loop form:
[0014] Each of them It is a circular matrix, formed by vectors generate; The projection is calculated as follows: ; Among them, FFT and IFFT are Fast Fourier Transform and Inverse Fourier Transform, respectively.
[0015] According to the above scheme, in step 4), for the time prediction task, the loss function uses the mean squared error (MSE) to calculate the predicted value. and the true value The difference, and minimize the loss function, The formula for calculating the MSE loss function is as follows:
[0016] in, It is the batch sample size. This is the model output.
[0017] The beneficial effects of this invention are: The gated multi-attribute fusion module designed in this invention achieves adaptive dynamic weighted fusion of heterogeneous features through a three-level gating mechanism of temporal evolution gating, cross-attribute attention gating, and two-layer residual fusion, effectively solving the problem of poor coupling of multi-attribute features and outputting stable high-level features. The Aug-Transformer temporal coding network constructed in this invention, by adding enhanced shortcut paths and using block cyclic matrices to reduce computational complexity, can accurately capture the long-term and short-term dependencies of tower crane operation data, improve feature diversity and model generalization ability, and adapt to the real-time prediction needs of high-dimensional time series. Attached Figure Description
[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0020] like Figure 1 As shown, a method for predicting the scheduling time of large tower crane tasks based on a gated multi-attribute Transformer includes the following steps: 1) Construct a multi-attribute operation dataset for tower cranes as the training set for the model. The multi-attribute operation dataset includes equipment status parameters, environmental data, task attributes, and historical time series data. 1.1) Real-time collection of multi-dimensional operational data, including lifting capacity, slewing angle, wind speed, task floor, operation type, and historical task completion time, through tower crane logic controller PLC, sensor array, and construction management platform; 1.2) Perform data preprocessing on the collected raw data, including missing value imputation, anomaly removal, normalization and timestamp alignment, to form time series records with a uniform frequency; 1.3) The time series records are embedded and encoded according to the discrete task attributes, and the continuous state variables are segmented, quantized, and concatenated into a multi-dimensional tensor to form a multi-attribute time series sample. Discrete task attributes refer to non-continuous attribute data in tower crane operation tasks that have no numerical correlation and only represent category type characteristics. This type of data cannot be directly represented by continuous numerical values and needs to be converted into a vector form that the model can recognize through embedding encoding. It is one of the important heterogeneous features in the multi-attribute operation dataset of tower cranes.
[0021] Based on the actual operation scenarios of tower cranes, discrete task attributes mainly include the following categories: Job types: such as material lifting, equipment transfer, empty hook return, floor unloading, and other different types of work actions; Material types: such as steel bars, concrete, scaffolding, precast building components, construction equipment, and other types of materials that are hoisted; Work areas: such as non-numerical area markings like construction zones A / B / C, main building / podium building, east / west work areas; Operating modes: such as manual operation, semi-automatic scheduling, emergency operation, and other tower crane operating mode categories; Task priority: such as first-level, second-level, and third-level priority, and the classification of tasks into normal or urgent tasks.
[0022] Continuous state variables (such as lifting weight, wind speed, and slewing angle) directly input into the model can lead to unstable training due to large differences in their numerical ranges. Piecewise quantization maps continuous numerical values to finite discrete intervals, transforms them into fixed-dimensional vectors, and finally concatenates them with the encoded discrete attributes to form a unified multidimensional tensor. In other words, encoding is accomplished through interval numbering, and the number of intervals can be adjusted according to business needs. The core principle is to discretize continuous numerical values while retaining their business meaning.
[0023] 1.4) Divide the samples into training set, validation set and test set in chronological order, and keep the working condition distribution of each set consistent to ensure the effectiveness of model training; 2) Design a gated multi-attribute fusion module to dynamically weight and fuse heterogeneous features through an adaptive weight allocation mechanism; Heterogeneous features include equipment status features, environmental features (such as wind speed), task attribute features, and historical time series features; Historical time-series features are the time-series set of the above three types of features in the historical operation process, and exist in the form of a sequence of timestamps and feature values; Step 2) of the gated multi-attribute fusion completes the mapping from the original multi-attribute sequence to a unified predictive representation in three steps, specifically including: 2.1) Temporal evolution gating is a temporal gate that is independently computed for each task attribute in the sequence, generating an attribute-specific evolutionary representation; the temporal gate retains how much current observation and references how much history; Each feature attribute refers to the parameters of four major categories of features of the tower crane: equipment status, environment, task, and historical time series. Each individual parameter under each feature category (such as wind speed, lifting capacity, and operation type) is an independent attribute. First, regarding the first The attributes of time Calculate independently:
[0024] Generate gated representation: ; By calculating the time series gate independently for each attribute, the importance of a single attribute at different time points can be accurately captured, avoiding interference between the weights of different attributes. Finally, through gated fusion, the time series weights of all attributes are integrated to obtain a more realistic fusion feature and improve prediction accuracy.
[0025] 2.2) Cross-attribute attention gating: Based on scaled dot product attention, weights are assigned among all attributes to obtain a unified representation of cross-attribute coupling; Will Align to a uniform dimension and stack as ; Scale the dot product attention calculation, calculate the attention weight of each attribute relative to the "time prediction" target, and then calculate the attention weights between attributes:
[0026] Where softmax represents normalization. And mapped to a gated vector ; Weighted fusion yields: ; 2.3) The dual-layer residual fusion output performs gated fusion again with the cross-attribute representation and the hidden state of the previous time step, and uses regularization terms to prevent the gating from becoming extreme, outputting stable high-level features. First of all and Perform gating fusion:
[0027] Introducing regularization terms To prevent saturation; 3) Construct a Transformer temporal coding network, input gated fusion features into the temporal coding network to capture long-short-term dependencies, and output deep temporal representations to predict the execution time of the tower crane's current task; This invention uses the Aug-Transformer model, which is based on the standard Transformer architecture, but improves feature diversity by adding augmented shortcuts.
[0028] The overall structure of the model is a deep stacked encoder, consisting of multiple modules: input layer, core modules: AugmentedMulti-headSelf-Attention (AugMSA) module and AugmentedMultilayerPerceptron (AugMLP) module, and output layer.
[0029] Step 3) includes the following sub-steps: 3.1) Input sequence representation: The time series is divided into fixed-length windows (e.g., each window contains T time points). Each time point or window segment is projected onto a d-dimensional embedding vector. Where N is the sequence length (number of time steps) and d is the feature dimension. For example, for a time series window, the embedding for each time point t is calculated as follows: ; in It is the original value at time point t, and LinearEmbedding is a linear projection layer.
[0030] 3.2) Construct the AugMSA module. The AugMSA module adds an enhanced shortcut path to the standard (Multi-head Self-Attention) MSA module to improve feature diversity. MSA captures temporal dependencies within a sequence, while the enhanced shortcut path provides an additional path to bypass MSA, directly passing and transforming features; The standard MSA calculation method is as follows:
[0031]
[0032] in, H represents the input features of the l-th layer; H is the number of attention heads. It is the projection matrix of query Q, key K, and value V of the h-th attention head; It is an attention map that captures the relationships between points in time; Let L be the projection weight matrix of the l-th layer; The enhanced AugMSA module is represented as follows:
[0033] in, It is the i-th enhanced shortcut path, usually defined as:
[0034] It is an activation function. It is a learnable weight matrix. In time-based prediction, it allows features to transform independently at each time point, increasing diversity.
[0035] 3.3) Constructing the AugMLP module (feature transformation module); The AugMLP module adds enhanced shortcut paths to the standard MLP to process local features at each time point. Similar to AugMSA, but applied to the MLP module. The MLP handles nonlinear transformations, while the enhanced shortcut paths provide additional paths to maintain feature diversity; The AugMLP module is represented as:
[0036] Among them, enhance shortcut paths Similar to AugMSA, it uses the same form. This ensures that the MLP module can also avoid feature crashes.
[0037] 3.4) Use block-circulant matrices to implement enhanced shortcut paths; The original dense matrix is decomposed into multiple cyclic submatrices, and the computation is performed in the frequency domain. weight matrix Decompose into block loop form:
[0038] Each of them It is a circular matrix, formed by vectors generate.
[0039] The projection is calculated as follows: ; Among them, FFT and IFFT are Fast Fourier Transform and Inverse Fourier Transform, respectively; Represents element-wise multiplication; complexity is O(n log n). Much lower than the original dense matrix Suitable for real-time time prediction. Among them, The complexity of the projection of the original dense matrix is given by... The complexity after dimensionality reduction by combining the block cyclic matrix with FFT is denoted as d, where d is the feature dimension.
[0040] In the enhanced shortcut path, the original dense weight matrix (also often called "dense weight matrix") refers to the fully connected weight matrix in the deep learning model that is used to perform linear transformation on the feature vector without any sparsification. It is a two-dimensional matrix with independent learnable parameters at each position. In Aug-Transformer's enhanced shortcut path, it was originally a core parameter used to perform projection transformation on the feature vectors of tower crane operation (such as the attribute features of wind speed and lifting capacity after gating fusion); Suppose the enhanced shortcut path needs to process the 128-dimensional gated fusion features of the tower crane (input d) in =128), outputting the same 128-dimensional features (d out =128): The original dense weight matrix W is 128×128 in size, containing 128×128=16384 independent learnable parameters; The input 128-dimensional tower crane feature vector X (containing fused information such as wind speed, lifting capacity, and operation type) is transformed into a new feature vector through matrix multiplication X×W, realizing the linear transformation of features and preparing for subsequent nonlinear activation.
[0041] In tower crane task scheduling and prediction, high-dimensional time-series data (such as long-sequence tower crane operation parameters) are processed. The original dense weight matrix has the drawbacks of high computational complexity and large number of parameters.
[0042] Therefore, decomposing the original dense weight matrix in the enhanced shortcut path into multiple cyclic submatrices is essentially replacing the dense matrix with a block cyclic matrix, reducing the computational complexity from O(d2) to O(dlogd) (accelerated by FFT). This retains the feature transformation capability while improving the computational efficiency of the model, making it suitable for real-time prediction scenarios for tower cranes.
[0043] 3.5) The model is obtained by stacking L layers, each layer containing AugMSA and AugMLP modules.
[0044] For time prediction tasks, the output layer is a fully connected regression layer. The [CLS] position of the output feature Z of the last layer is taken and the predicted value is generated through the fully connected layer. Final prediction layer: ; in, LinearOutput is the predicted value (e.g., the next point in a time series) and is a fully connected layer.
[0045] 4) Use validation data to perform online evaluation and incremental fine-tuning of the prediction model until it meets the preset accuracy threshold; For time-based forecasting tasks, the loss function uses the mean squared error (MSE) to calculate the predicted value. and the true value The difference, and minimize the loss function, The formula for calculating the MSE loss function is as follows: ; in, It is the batch sample size. This is the model output.
[0046] 5) Output the task execution time prediction results and generate scheduling optimization instructions; Based on the predicted task execution times output by the Transformer temporal coding network, a dynamic scheduling model is constructed with the goal of maximizing overall equipment utilization. An integer programming method is used to solve the task allocation problem, generating a globally optimal task instruction set while satisfying equipment physical constraints and temporal feasibility.
[0047] 5.1) Input the multi-attribute data of the tower crane collected in real time at the construction site into the optimal prediction model, and output the k predicted execution times of the current and subsequent tasks to be executed, where k is the number of tasks to be executed; 5.2) Construct a dynamic scheduling model with the goal of maximizing the comprehensive utilization rate of tower crane equipment. The objective function is: ; Where η is the overall utilization rate of the equipment, and T totalt represents the total duration of the scheduling cycle. wait(i) Let i be the waiting time between the i-th task and the (i+1)-th task. 5.3) Set physical constraints for tower crane equipment (such as lifting capacity ≤ rated lifting capacity, slewing angle ≤ maximum slewing angle, etc.) and temporal feasibility constraints (no conflict in task execution order, task start time ≥ completion time of the previous task, etc.), and solve the task allocation problem using integer programming to obtain the globally optimal task execution order and time arrangement; 5.4) Generate a scheduling optimization instruction set based on the solution results, including task execution order, start time of each task, and suggestions for adjusting equipment operating parameters, and send the instruction set to the tower crane operation control console to realize intelligent scheduling of tower cranes.
[0048] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.
Claims
1. A method for predicting the scheduling time of large tower crane tasks based on a gated multi-attribute Transformer, characterized in that, Includes the following steps: 1) Construct a multi-attribute operation dataset for tower cranes as the training set for the model; the multi-attribute operation dataset includes equipment status parameters, environmental data, task attributes, and historical time series data; 2) A gated multi-attribute fusion strategy is adopted, which dynamically weights and fuses heterogeneous features through an adaptive weight allocation mechanism to output gated fused features; The heterogeneous features include device status features, environmental features, task attribute features, and historical time series features; 3) Construct a Transformer temporal coding network, input the gated fusion features into the Transformer temporal coding network to capture the long-term and short-term dependencies of the tower crane operation data, and output a deep temporal representation to predict the execution time of the current task of the tower crane; 4) Use validation data to evaluate and fine-tune the prediction model online until the model's prediction accuracy meets the preset accuracy threshold; 5) Output the tower crane task execution time prediction results and generate scheduling optimization instructions based on the prediction results.
2. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 1, characterized in that, Step 1) includes: 1.1) Real-time collection of multi-dimensional operational data, including lifting capacity, slewing angle, wind speed, task floor, job type, and historical completion time; 1.2) The collected raw data is timestamped to form a time-series record with a uniform frequency; 1.3) The time series records are embedded and encoded according to the discrete task attributes, and the continuous state variables are segmented, quantized, and concatenated into a multi-dimensional tensor to form a multi-attribute time series sample. 1.4) Divide the multi-attribute time series samples into training set, validation set and test set according to time order, and keep the working condition distribution of each set consistent.
3. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 2, characterized in that, In step 1.3), the discrete task attributes include job type, material type, job area, operation mode, and task priority, as detailed below: Job types: including different categories of work actions such as material lifting, equipment transfer, empty hook return, and floor unloading; Material types: This includes categories of materials that are hoisted, such as steel bars, concrete, scaffolding, precast building components, and construction equipment. Work area: This includes the markings for the construction area, main building, podium building, and work surface; Operating modes: These include tower crane operating modes such as manual operation, semi-automatic scheduling, and emergency operation. Task Priority: Used to classify the priority level of tasks.
4. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 1, characterized in that, The gated multi-attribute fusion strategy in step 2) completes the mapping from the original multi-attribute sequence to the unified prediction representation through the following steps: 2.1) Temporal evolution gating; For each feature attribute, a time-series gate is independently calculated to determine the weight allocation of the current and historical observations, generating an attribute-specific evolutionary representation. 2.2) Cross-attribute attention gating: Based on scaled dot product attention, weights are assigned among all attributes to obtain a unified representation of cross-attribute coupling, transforming the originally scattered and multi-dimensional attribute features into a feature vector with a unified format and containing cross-attribute association information. 2.3) The dual-layer residual fusion output performs gating fusion again with the cross-attribute representation and the hidden state of the previous time step, and introduces a regularization term to prevent gating extremes, outputting stable high-level features.
5. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 1, characterized in that, The Transformer temporal coding network constructed in step 3) is the Aug-Transformer model, which adds an enhanced shortcut path based on the standard Transformer architecture. The Aug-Transformer model structure includes an input layer, a core module, and an output layer. The core modules are the AugMSA module and the AugMLP module.
6. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 5, characterized in that, Step 3) predicting the execution time of the tower crane's current task includes the following steps: 3.1) Input sequence representation: The time series is divided into fixed-length windows. Each time point or window segment is projected onto a d-dimensional embedding vector through a linear projection layer, forming the input sequence. Where N is the sequence length and d is the feature dimension; 3.2) Construct the AugMSA module, which adds an enhanced shortcut path to the standard MSA module; the MSA module captures the temporal dependencies within the sequence, while the enhanced shortcut path provides an additional path to bypass the MSA module and directly pass and transform features; 3.3) Construct the AugMLP module; the AugMLP module adds enhanced shortcut paths to the standard MLP; 3.4) Use block circular matrices to implement enhanced shortcut paths; 3.5) The model is obtained by stacking L network layers. Each network layer contains AugMSA and AugMLP modules. The output layer adopts a fully connected regression layer. The [CLS] position of the output feature of the last layer is taken, and the predicted value of the tower crane task execution time is generated through the fully connected layer.
7. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 6, characterized in that, In step 3.2), the standard MSA is represented as: ； ； in, Here, H represents the input features of the l-th layer, and H is the number of attention heads. It is the projection matrix of query Q, key K, and value V of the h-th attention head. It is an attention map that captures the relationships between points in time; Let L be the projection weight matrix of the l-th layer; The AugMSA module is represented as follows: ； in, It is the i-th enhanced shortcut path, defined as: ； It is an activation function. It is a learnable weight matrix.
8. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 1, characterized in that, In step 3.4), an enhanced shortcut path is implemented using a block cyclic matrix, as detailed below: The original dense matrix is decomposed into multiple cyclic submatrices and computed in the frequency domain. weight matrix Decompose into block loop form: Each of them It is a cyclic submatrix, formed by vectors generate; The projection calculation is performed based on the decomposed block cyclic matrix, as follows: ； Among them, FFT and IFFT are Fast Fourier Transform and Inverse Fourier Transform, respectively.
9. The method for predicting the scheduling time of large tower crane tasks based on gated multi-attribute Transformer according to claim 1, characterized in that, In step 4), for time prediction tasks, the loss function uses the mean squared error (MSE) to calculate the predicted value. and the true value The difference, and minimize the loss function, The formula for calculating the MSE loss function is as follows: in, It is the batch sample size. This is the model output.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 9.