Cutter suction dredger productivity prediction method and system based on dynamic Transform model

By using dynamic Transformer model preprocessing and feature selection, combined with local and global attention mechanisms, the complex working conditions and long sequence forgetting problems in cutter suction dredger capacity prediction are solved, achieving high-precision capacity prediction.

CN121615829APending Publication Date: 2026-03-06SANYA SCI & EDUCATION INNOVATION PARK WUHAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511631133.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-10
Publication Date
2026-03-06

AI Technical Summary

Technical Problem

Existing methods for predicting the capacity of cutter suction dredgers cannot provide high-precision predictions when faced with complex nonlinear factors and equipment interactions. Data-driven methods struggle to capture long-term series information, and feature selection methods are not sufficiently relied upon, affecting the accuracy and reliability of the prediction results.

Method used

By employing a dynamic Transformer model-based approach, and combining sensor data preprocessing, feature selection, and multi-layer encoder stacking modules with local context and global associative attention, along with a memory enhancement submodule, high-precision prediction of cutter suction dredger capacity is achieved.

Benefits of technology

It achieves high-precision prediction of cutter suction dredger capacity, solves the problem of insufficient prediction under complex working conditions by traditional methods and the long-sequence forgetting problem of data-driven methods, ensures the uniqueness and relevance of feature selection, and provides reliable capacity prediction support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121615829A_ABST
    Figure CN121615829A_ABST
Patent Text Reader

Abstract

The invention discloses a cutter suction dredger productivity prediction method based on a dynamic Transform model, and the method comprises the steps: firstly cleaning the original data of a sensor, and eliminating noise and abnormal values to guarantee the validity of the data (the cleaned data can accurately reflect the normal working condition performance of a CSD); then carrying out feature analysis, and screening out features which are highly related to the target variable and have differences with each other; after a time sequence sample is generated, model training is started, and after data are interactively processed by encoders of all layers, model performance is evaluated through goodness of fit, and prediction is completed. The method can solve the technical problem that a high-precision prediction result cannot be provided when an existing prediction method of a traditional mechanical model faces interaction between complex nonlinear factors and equipment, and an existing machine learning method based on data driving is difficult to capture an important dependency relationship in long-time sequence information. The technical problem that long sequence forgetting exists when super-long historical data of the cutter suction dredger is processed, and then prediction precision is affected is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of machine learning and digital science processing technology in artificial intelligence, and more specifically, relates to a method and system for predicting the production capacity of cutter suction dredgers based on dynamic transformers. Background Technology

[0002] Dredging is a crucial component of waterway infrastructure construction and maintenance, playing a vital role in channel dredging, port construction, coastal protection, land reclamation, and marine engineering. With the continuous growth of global port traffic and coastal development demands, the scale and complexity of dredging operations are constantly increasing, placing higher demands on the operational efficiency and reliability of construction equipment. Cutter suction dredgers (CSDs), as the most widely used and primary equipment in the dredging industry, have a significant impact on project progress and economic benefits when their capacity is increased. Therefore, accurately assessing and predicting the capacity of cutter suction dredgers is of great importance for optimizing the construction process and making scientific decisions.

[0003] Currently, there are three main types of methods for predicting the production capacity of cutter suction dredgers. The first type is based on traditional mechanical models, which are based on physical principles and mainly use mechanical formulas to derive operational performance, but the accuracy of the results is not high. The second type is based on data-driven machine learning methods, which rely on a large amount of historical sensor data and use machine learning models to make predictions. The third type extracts feature variables that are highly correlated with the objective function through feature selection and studies their linear or nonlinear relationship with the objective function.

[0004] However, the three existing methods for predicting the production capacity of cutter suction dredgers all have some significant drawbacks: First, existing prediction methods based on traditional mechanical models rely too heavily on static physical assumptions and are ill-equipped to handle dynamic changes in operating conditions during actual operation. Therefore, these methods cannot provide high-precision predictions when faced with complex nonlinear factors and interactions between equipment. Second, while existing data-driven machine learning methods can be trained on large-scale datasets, they lack effective memory mechanisms and struggle to capture important dependencies in long-term time series information. This leads to long-sequence forgetting (i.e., the model cannot remember dependencies between information with significant time differences) when processing the extremely long historical data of cutter suction dredgers, thus affecting the accuracy of predictions. Third, existing feature selection methods often rely on a single algorithm for feature screening, which can easily lead to feature redundancy or omissions. This results in insufficient reliance of the prediction model on key features, further impacting the accuracy and reliability of production capacity predictions. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for predicting the production capacity of cutter suction dredgers based on a dynamic Transformer model. The purpose is to solve the technical problems of existing prediction methods based on traditional mechanical models failing to provide high-precision prediction results when faced with complex nonlinear factors and interactions between equipment; the technical problems of existing data-driven machine learning methods struggling to capture important dependencies in long-term data series, leading to long-sequence forgetting when processing ultra-long historical data of cutter suction dredgers, thus affecting prediction accuracy; and the technical problems of existing feature selection methods' prediction models not relying sufficiently on key features, further affecting the accuracy and reliability of production capacity prediction results.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for predicting the production capacity of a cutter suction dredger based on a dynamic Transformer model is provided, comprising the following steps: (1) Collect multiple sensor data of the cutter suction dredger by the sensors installed on the cutter suction dredger, and preprocess all sensor data to obtain preprocessed multiple sensor data; (2) Perform feature selection processing on all the preprocessed sensor data obtained in step (1) to obtain feature matrix B; (3) Input the feature matrix B obtained in step (2) into the pre-trained dynamic Transformer model to obtain the production capacity prediction result D of the cutter suction dredger.

[0007] Preferably, the sensor data includes mud flow rate, pump speed, vacuum level, pipeline pressure, cutting power, and vessel position speed; The preprocessing process in step (1) is as follows: First, outlier detection, missing value imputation and data smoothing are performed on all sensor data in sequence to obtain the sensor data after preliminary processing; then, sliding window processing is performed on all the sensor data after preliminary processing to obtain multiple sensor data after preprocessing. The feature selection process in step (2) includes collinearity removal, automatic matching selection, and manual screening. Collinearity removal is the variance inflation factor (VIF) test method. Automatic matching selection is the SelectKBest algorithm based on the F test. Manual screening is to supplement important but missing feature variables based on the operating principle and production impact mechanism of the cutter suction dredger.

[0008] Preferably, the dynamic Transformer model includes an input feature embedding module, a position encoding module, a multi-layer encoder stacking module, an adaptive pooling module, and a fully connected prediction module; The input to the input feature embedding module is a multi-dimensional temporal feature matrix with dimensions 64*15*197. This layer maps this multi-dimensional temporal feature matrix into a unified hidden layer space to obtain a temporal feature matrix with dimensions 64*15*197. The input to the position encoding module is a 64*15*197-dimensional temporal feature matrix output from the feature embedding module. This layer generates a fixed position vector using a sine-cosine function and adds this fixed position vector element-wise to the temporal feature matrix to obtain the sequence order information matrix. , where R represents the set of real numbers.

[0009] Preferably, the multi-layer encoder stacking module includes two encoder layers, each of which includes a local context attention submodule, a global association attention submodule, a memory enhancement submodule, and a feedforward neural network submodule, with the following specific structure: The local context attention submodule consists of six layers: a first linear projection layer, a first multi-head segmentation layer, a local window extraction layer, a local attention weight calculation layer, a local context aggregation layer, and a first multi-head fusion and linear projection layer. The first linear projection layer takes as input the sequence order information matrix output by the position encoding module. This layer first applies the matrix through three independent sets of linear mapping layers (fully connected layers). Perform mapping processing to obtain the query matrix separately. Key matrix and value matrix ; The first multi-head segmentation layer takes as input the query matrix output by the first linear projection layer. Key matrix and value matrix This layer divides each matrix into 4 subspaces along the feature dimension (each subspace has a dimension of 8, and the shape of the matrix after division is as follows). ), to obtain the query matrix after multi-head partitioning Key matrix and value matrix ; The local window extraction layer takes as input the query matrix after multi-head partitioning output from the first multi-head partitioning layer. Key matrix and value matrix This layer is for each time step in each matrix. In other words, define a local context window Extract the local key matrix corresponding to each time step within this window. With local value matrix To obtain the query vector corresponding to each time step. Local bond matrix and local value matrix ,in Represents window size, in this invention =15; The local attention weight calculation layer takes the query vector corresponding to each time step as input from the local window extraction layer. and local bond matrix This layer is based on the query vector. and local bond matrix Calculate the attention score matrix ( The attention score matrix is ​​then processed using the Softmax function. Normalization is performed to obtain the attention weight matrix for each time step. ; The local context aggregation layer takes as input the attention weights at each time step, output by the local attention weight calculation layer. The local value matrix corresponding to each time step output by the local window extraction layer. This layer performs a weighted sum of the two to obtain the context representation and local context feature set corresponding to each time step. ; The first multi-head fusion and linear projection layer takes as input the set of local context features corresponding to each time step output by the local context aggregation layer. This layer is the set of local context features corresponding to all time steps. The data is then spliced ​​and reshaped, and the results are linearly mapped to obtain the local enhanced feature matrix. And output it.

[0010] Preferably, the global association attention submodule comprises eight layers: a second linear projection layer, a second multi-head segmentation layer, a rotating bucket partitioning layer, a mask generation layer, an attention score calculation layer, an attention weight calculation layer, a context feature generation layer, and a second multi-head fusion and linear projection layer. The second linear projection layer takes as input the local enhancement feature matrix from the first multi-head fusion and the output of the linear projection layer in the local context attention submodule. This layer enhances the local feature matrix. The processing method is the same as that of the first linear projection layer in the local context attention submodule, and the final output is a query matrix. Key matrix and value matrix ; The second multi-head segmentation layer takes the query matrix output by the second linear projection layer as its input. Key matrix and value matrix Its query matrix Key matrix and value matrix The processing method is exactly the same as that in the first multi-head segmentation layer of the local context attention submodule, and the final output is the query matrix after multi-head segmentation. Key matrix and value matrix ; The rotating bucket partitioning layer takes the query matrix output by the second multi-head partitioning layer as its input. Bond matrix This layer first performs the query matrix Key matrix Matrix multiplication is performed with the random rotation matrix initialized by the rotating bucket partitioning layer to obtain the query matrix and key matrix after random projection. Then, the query matrix and key matrix after random projection are respectively subjected to bucket partitioning to obtain the query bucket number matrix. Key bucket numbering matrix ; The mask generation layer takes as input the query bucket number matrix output by the rotating bucket partitioning layer. Key bucket numbering matrix This layer uses the unsqueeze function to process the query bucket number matrix. Key bucket numbering matrix Expand to obtain the attention mask attn_mask; The attention score calculation layer takes the query matrix output by the second multi-head segmentation layer as its input. Bond matrix And the attention mask attn_mask output by the mask generation layer, which first processes the query matrix. AND key matrix The transpose of the matrix is ​​multiplied to obtain the original attention score; then, the original attention score is multiplied by a scaling factor; finally, the masked_fill function is used to set the score of the position with a mask value of 0 in the multiplication result to -1e9 to obtain the masked attention score. The attention weight calculation layer takes the masked attention score as input from the attention score calculation layer and normalizes the score to obtain the attention weight. Generate a context feature layer, whose input is the value matrix output by the second multi-head segmentation layer. The attention weights are calculated from the output of the attention weight calculation layer, which is a value matrix. Matrix multiplication is performed with the attention weights to obtain contextual features. ; The second multi-head fusion and linear projection layer takes as input the context features output by the context feature generation layer. The second multi-head fusion and linear projection layer on this contextual feature The processing method is the same as that of the first multi-head fusion and linear projection layer in the local context attention submodule, and the output is a feature matrix. .

[0011] Preferably, the memory enhancement submodule comprises four layers: a third linear projection layer, a memory attention score calculation layer, a memory attention weight calculation layer, and a memory output generation layer. The third linear projection layer takes as input the feature matrix from the second multi-head fusion and the output of the linear projection layer in the global association attention submodule. This layer has this feature matrix The processing method is the same as that of the first linear projection layer in the local context attention submodule, and the final output is a query matrix. Key matrix and value matrix ; The memory attention score calculation layer takes as input the query matrix output by the third linear projection layer. The layer consists of a memory matrix with dimensions of number of memory slots * 197. This layer is used for querying the matrix. Matrix multiplication is performed with the transpose of the memory matrix to obtain the memory attention score; The memory attention weight calculation layer takes the memory attention score output by the memory attention score calculation layer as its input and normalizes the memory attention score to obtain the memory attention weight. A memory output layer is generated, whose inputs are the memory attention weights output by the memory attention weight calculation layer and a memory matrix with dimensions of the number of memory slots * 197. This layer performs matrix multiplication on the memory attention weights and the memory matrix to obtain the context feature matrix. And output it.

[0012] Preferably, the input to the feedforward neural network submodule is the context feature matrix generated by the memory output layer in the memory enhancement submodule. This layer first examines the context feature matrix. A linear transformation is performed to obtain the linearly transformed feature matrix. Then, the nn.GELU() activation function is used to perform a non-linear mapping on each element of the linearly transformed feature matrix to obtain a feature matrix with non-linear expressive power. Subsequently, the nn.Dropout function is used to randomly set a portion of the elements in the feature matrix to 0 to obtain a deactivated feature matrix. Finally, a second linear transformation is performed on the deactivated feature matrix, and it is mapped back to the original dimension to obtain the depth-processed feature matrix. And output; The input to the adaptive pooling module is the depth-processed feature matrix output from the feedforward neural network submodule in the multi-layer encoder stack module. This layer performs average pooling on the last dimension of the feature matrix to obtain the pooled feature; then, it removes the last dimension of size 1 from the feature to obtain the compressed feature vector and outputs it. The fully connected prediction module consists of two layers, and its specific structure is as follows: The first layer takes the compressed feature vector output by the adaptive pooling module as input. This layer performs a linear transformation on the compressed feature vector to obtain a linearly transformed feature vector. Then, the nn.GELU() activation function is used to perform a non-linear mapping on each element of the linearly transformed feature vector to obtain a feature vector with non-linear expressive power. Subsequently, the nn.Dropout function is used to randomly set some elements to 0 to obtain a deactivated feature vector. Finally, a second linear transformation is performed on the deactivated feature vector, and it is mapped back to the original dimensional space to obtain the feature vector processed by the first layer, which is then output. The second layer takes the feature matrix processed by the first layer as its input and processes the feature matrix in the same way as the first layer to obtain the final predicted value and output it.

[0013] Preferably, the dynamic Transformer model is obtained through the following steps: (3-1) Collect multiple sensor data of the cutter suction dredger under different working conditions from actual engineering projects, perform data cleaning and sliding window processing on the multiple sensor data to obtain multiple time series samples, and divide all time series samples into training set, validation set and test set in a ratio of 7:1.5:1.5; (3-2) For each time series sample in the training set obtained in step (3-1), the StandardScaler function is used to normalize it to obtain the normalized time series sample corresponding to the time series sample, and feature selection processing is performed on the normalized time series sample to obtain the feature matrix B corresponding to the time series sample. (3-3) For each time series sample in the training set obtained in step (3-1), the feature matrix B corresponding to the time series sample obtained in step (3-2) is input into the input feature embedding module, and the feature matrix B is mapped to a unified hidden layer space through a linear projection layer to obtain the time series feature matrix corresponding to the time series sample. (3-4) For each time series sample in the training set obtained in step (3-1), the time series feature matrix corresponding to the time series sample obtained in step (3-3) is input into the position encoding module, and the time series feature matrix is ​​encoded in time position to obtain the sequence order information matrix corresponding to the time series sample. (3-5) For each time series sample in the training set obtained in step (3-1), the sequence order information matrix corresponding to the time series sample obtained in step (3-4) is input into the multilayer encoder stacking module to obtain the deep-processed feature matrix corresponding to the time series sample. ; (3-6) For each time series sample in the training set obtained in step (3-1), the feature matrix corresponding to that time series sample obtained in step (3-5) is... Input adaptive pooling module to the feature matrix Perform average pooling along the time step dimension to obtain the compressed feature vector; (3-7) For each time series sample in the training set obtained in step (3-1), the compressed feature vector obtained in step (3-6) is input into the fully connected prediction module to obtain the final prediction value corresponding to the time series sample. (3-8) For each time series sample in the training set obtained in step (3-1), the coefficient of determination R is obtained based on the final predicted value corresponding to that time series sample obtained in step (3-7). 2 and the prediction loss function, which includes the mean absolute error (MAE) and the root mean square error (RMSE); (3-9) For each time series sample in the training set obtained in step (3-1), based on the determination coefficient and prediction loss function obtained in step (3-8), and using the Adam optimizer to perform gradient backpropagation and parameter update on the dynamic Transformer model until the number of iterations reaches the preset value, the optimal parameters of the dynamic Transformer model at this time are obtained, thereby obtaining the dynamic Transformer model that has been initially trained. (3-10) Use the test set obtained in step (3-1) to test the dynamic Transformer model initially trained in step (3-9) until the prediction accuracy reaches the optimal level, thereby obtaining the final trained dynamic Transformer model.

[0014] Preferably, step (3-8) involves calculating the coefficient of determination R using the following formula. 2 Mean Absolute Error (MAE) and Root Mean Square Error (RMSE):

[0015] in, This represents the total number of time-series samples in the training set; This represents the true value of the i-th time series sample; This represents the predicted value of the i-th time series sample; This represents the average of the true values ​​of all time-series samples. This represents the average of the predicted values ​​for all time-series samples.

[0016] According to another aspect of the present invention, a cutter suction dredger capacity prediction system based on a dynamic Transformer model is provided, comprising: The first module is used to collect multiple sensor data of the cutter suction dredger through the sensors installed on the cutter suction dredger, and to preprocess all the sensor data to obtain preprocessed multiple sensor data. The second module is used to perform feature selection processing on all the preprocessed sensor data obtained by the first module to obtain feature matrix B. The third module is used to input the feature matrix B obtained from the second module into the pre-trained dynamic Transformer model to obtain the capacity prediction result D of the cutter suction dredger.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. Because the present invention adopts step (3-3), its multi-layer encoder's local context attention submodule and global correlation attention submodule simultaneously pay attention to global and local operating features, realizing hierarchical modeling of the dynamic working state of CSD, which can more accurately reflect the equipment capacity change characteristics under different working conditions, and make the model have dynamic adjustment capabilities. Therefore, it can solve the problem that the existing traditional prediction methods cannot face complex nonlinear factors and the interaction between equipment, thus affecting the prediction accuracy. 2. Because the present invention adopts step (3-3), its memory enhancement submodule of the multi-layer encoder can focus on long-term sequence information while acquiring short-term dependencies. Therefore, it can solve the technical problem that existing data-driven machine learning methods have difficulty capturing important dependencies in long-term sequence information, which leads to the problem of long-sequence forgetting when processing ultra-long historical data of cutter suction dredgers, thus affecting the prediction accuracy. 3. Because of the adoption of step (2), this invention proposes a multi-level feature engineering method that combines domain knowledge and algorithm screening. By combining collinearity elimination, automatic feature matching and manual feature screening, it effectively ensures the uniqueness and high correlation of input features. Therefore, it can solve the technical problem that existing feature selection methods rely on a single algorithm for feature screening, resulting in feature redundancy or omission, which leads to insufficient reliance of the prediction model on key features, thereby affecting the accuracy of production capacity prediction results. 4. The production capacity prediction framework established by this invention can achieve high-precision hourly production prediction under real engineering data, providing reliable theoretical and technical support for the intelligent operation and decision optimization of large-scale dredging equipment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the structure of the dynamic Transformer model of this invention; Figure 2 This is a comparison chart of the distribution of the objective function before and after data preprocessing; Figure 3 It is a heatmap showing the correlation between each feature and the objective function; Figure 4 This is a schematic diagram of the capacity prediction results from the dynamic Transformer model; Figure 5 This is a schematic diagram illustrating the error in the capacity prediction results of the dynamic Transformer model; Figure 6 This is a flowchart illustrating the method for predicting the production capacity of cutter suction dredgers based on a dynamic Transformer model, as described in this 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 the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The basic idea of ​​this invention is to provide a method for predicting the production capacity of cutter suction dredgers based on a dynamic Transformer model. First, the raw sensor data is cleaned to eliminate noise and outliers to ensure data validity (the cleaned data can accurately reflect the normal operating performance of the CSD). Then, feature analysis is carried out to select features that are highly correlated with the target variable and have differences from each other. After generating time series samples, model training is started. After the data is processed by the encoders of each layer, the model performance is evaluated by the goodness of fit and the prediction is completed.

[0021] like Figure 6 As shown, this invention provides a method for predicting the production capacity of a cutter suction dredger based on a dynamic Transformer model, comprising the following steps: (1) Collect multiple sensor data of the cutter suction dredger by the sensors installed on the cutter suction dredger, and preprocess all sensor data to obtain preprocessed multiple sensor data; Specifically, the sensor data acquired in this step includes, but is not limited to, 256-dimensional feature information such as mud flow rate, pump speed, vacuum degree, pipeline pressure, cutting power, and vessel position speed. The preprocessing process in this step is as follows: First, outlier detection (i.e., using box plots to identify extreme outliers in the sensor data and removing values ​​exceeding a preset range to ensure the quality and validity of the sensor data), missing value imputation (i.e., using the median to imput missing numerical values ​​and the mode to imput missing non-numerical values ​​to ensure data integrity), and data smoothing (i.e. using Savgol filtering to remove data noise) are performed on all the preprocessed sensor data to ensure the continuity and accuracy of the sensor data, thus obtaining the preprocessed sensor data; then, sliding window processing is performed on all the preprocessed sensor data (i.e., a fixed-size window slides through the time series or signal in each sensor data to process all segments of the sensor data step by step) to obtain multiple preprocessed sensor data. (2) Perform feature selection processing on all the preprocessed sensor data obtained in step (1) to obtain feature matrix B; Specifically, the feature selection process in this step includes collinearity removal (in this invention, the variance inflation factor (VIF) test), automatic matching selection (in this invention, the SelectKBest algorithm based on the F-test), and manual screening (i.e., supplementing important but missing feature variables, including trolley travel and cutter speed, based on the operating principle and production impact mechanism of the cutter suction dredger), to ensure the differences between features and their correlation with the objective function (hourly production rate). Figure 3 A heatmap showing the correlation of feature selection results is presented.

[0022] The advantage of this step (2) is that the feature selection results are obtained through a process of selection and screening, which ensures that the selected features are highly correlated with the objective function. At the same time, the use of the variance inflation factor avoids the similarity between the selected features, thus ensuring the accuracy of the prediction results.

[0023] (3) Input the feature matrix B obtained in step (2) into the pre-trained dynamic Transformer model to obtain the production capacity prediction result D of the cutter suction dredger.

[0024] like Figure 1 As shown, the dynamic Transformer model of this invention includes an input feature embedding module, a position encoding module, a multi-layer encoder stacking module, an adaptive pooling module, and a fully connected prediction module.

[0025] The input to the input feature embedding module is a multi-dimensional temporal feature matrix with dimensions of 64*15*197. This layer maps the multi-dimensional temporal feature matrix to a unified hidden layer (hidden_dim, hidden_dim=32 in this invention) space to obtain a temporal feature matrix with dimensions of 64*15*197. The input to the position encoding module is a 64*15*197-dimensional temporal feature matrix output from the feature embedding module. This layer generates a fixed position vector using a sine-cosine function and adds this fixed position vector element-wise to the temporal feature matrix to obtain the sequence order information matrix. , where R represents the set of real numbers; The multi-layer encoder stacking module consists of two encoder layers. Each encoder layer includes a local context attention submodule, a global association attention submodule, a memory enhancement submodule, and a feedforward neural network submodule. Its specific structure is as follows: The local context attention submodule consists of six layers: a first linear projection layer, a first multi-head segmentation layer, a local window extraction layer, a local attention weight calculation layer, a local context aggregation layer, and a first multi-head fusion and linear projection layer. The first linear projection layer takes as input the sequence order information matrix output by the position encoding module. This layer first applies the matrix through three independent sets of linear mapping layers (fully connected layers). Perform mapping processing to obtain the query matrix separately. Key matrix and value matrix ; The first multi-head segmentation layer takes as input the query matrix output by the first linear projection layer. Key matrix and value matrix This layer divides each matrix into 4 subspaces along the feature dimension (each subspace has a dimension of 8, and the shape of the matrix after division is as follows). ), to obtain the query matrix after multi-head partitioning Key matrix and value matrix ; The local window extraction layer takes as input the query matrix after multi-head partitioning output from the first multi-head partitioning layer. Key matrix and value matrix This layer is for each time step in each matrix. In other words, define a local context window Extract the local key matrix corresponding to each time step within this window. With local value matrix ( ), to obtain the query vector corresponding to each time step. Local bond matrix and local value matrix ,in Represents window size, in this invention =15; The local attention weight calculation layer takes the query vector corresponding to each time step as input from the local window extraction layer. and local bond matrix This layer is based on the query vector. and local bond matrix Calculate the attention score matrix ( The attention score matrix is ​​then processed using the Softmax function. Normalization is performed to obtain the attention weight matrix for each time step. ; The local context aggregation layer takes as input the attention weights at each time step, output by the local attention weight calculation layer. The local value matrix corresponding to each time step output by the local window extraction layer. This layer performs a weighted sum of the two to obtain the context representation and local context feature set corresponding to each time step. ; The first multi-head fusion and linear projection layer takes as input the set of local context features corresponding to each time step output by the local context aggregation layer. This layer is the set of local context features corresponding to all time steps. The model undergoes splicing and reshaping, followed by linear mapping of the results to obtain a locally enhanced feature matrix. And output; The global correlation attention submodule consists of eight layers: the second linear projection layer, the second multi-head segmentation layer, the rotating bucket partitioning layer, the mask generation layer, the attention score calculation layer, the attention weight calculation layer, the context feature generation layer, and the second multi-head fusion and linear projection layer. The second linear projection layer takes as input the local enhancement feature matrix from the first multi-head fusion and the output of the linear projection layer in the local context attention submodule. This layer enhances the local feature matrix. The processing method is the same as that of the first linear projection layer in the local context attention submodule, and the final output is a query matrix. Key matrix and value matrix ; The second multi-head segmentation layer takes the query matrix output by the second linear projection layer as its input. Key matrix and value matrix Its query matrix Key matrix and value matrix The processing method is exactly the same as that in the first multi-head segmentation layer of the local context attention submodule, and the final output is the query matrix after multi-head segmentation. Key matrix and value matrix ; The rotating bucket partitioning layer takes the query matrix output by the second multi-head partitioning layer as its input. Bond matrix This layer first performs the query matrix Key matrix Matrix multiplication is performed with the random rotation matrix initialized by the rotating bucket partitioning layer to obtain the query matrix and key matrix after random projection. Then, the query matrix and key matrix after random projection are respectively subjected to bucket partitioning to obtain the query bucket number matrix. Key bucket numbering matrix ; The mask generation layer takes as input the query bucket number matrix output by the rotating bucket partitioning layer. Key bucket numbering matrix This layer uses the unsqueeze function to process the query bucket number matrix. Key bucket numbering matrix Expand to obtain the attention mask attn_mask; The attention score calculation layer takes the query matrix output by the second multi-head segmentation layer as its input. Bond matrix And the attention mask attn_mask output by the mask generation layer, which first processes the query matrix. AND key matrix The transpose of the matrix is ​​multiplied to obtain the original attention score; then, the original attention score is multiplied by a scaling factor (to avoid the score being too large and causing softmax saturation, the scaling factor in this invention is 8). -0.5Finally, the masked_fill function sets the score of the positions with a mask value of 0 in the multiplication result to -1e9 (masking the attention of elements from different buckets) to obtain the masked attention score. The attention weight calculation layer takes the masked attention score as input from the attention score calculation layer and normalizes the score to obtain the attention weight. Generate a context feature layer, whose input is the value matrix output by the second multi-head segmentation layer. The attention weights are calculated from the output of the attention weight calculation layer, which is a value matrix. Matrix multiplication is performed with the attention weights to obtain contextual features. ; The second multi-head fusion and linear projection layer takes as input the context features output by the context feature generation layer. The second multi-head fusion and linear projection layer on this contextual feature The processing method is the same as that of the first multi-head fusion and linear projection layer in the local context attention submodule, and the output is a feature matrix. .

[0026] The memory enhancement submodule consists of four layers: the third linear projection layer, the memory attention score calculation layer, the memory attention weight calculation layer, and the memory output generation layer. The third linear projection layer takes as input the feature matrix from the second multi-head fusion and the output of the linear projection layer in the global association attention submodule. This layer has this feature matrix The processing method is the same as that of the first linear projection layer in the local context attention submodule, and the final output is a query matrix. Key matrix and value matrix ; The memory attention score calculation layer takes as input the query matrix output by the third linear projection layer. A memory matrix with dimensions of number of memory slots * 197 (in this invention, the number of memory slots = 32) is used for querying the matrix. Matrix multiplication is performed with the transpose of the memory matrix to obtain the memory attention score; The memory attention weight calculation layer takes the memory attention score output by the memory attention score calculation layer as its input and normalizes the memory attention score to obtain the memory attention weight. A memory output layer is generated, whose inputs are the memory attention weights output by the memory attention weight calculation layer and a memory matrix with dimensions of the number of memory slots * 197. This layer performs matrix multiplication on the memory attention weights and the memory matrix (i.e., weighted aggregation of the features of the memory matrix through the weights) to obtain the context feature matrix. And output; The feedforward neural network submodule takes as input the context feature matrix the output of the memory output layer generated in the memory enhancement submodule. This layer first examines the context feature matrix. A linear transformation is performed to obtain the linearly transformed feature matrix. Then, the nn.GELU() activation function is used to perform nonlinear mapping on each element of the linearly transformed feature matrix to obtain a feature matrix with nonlinear expressive power. Subsequently, the nn.Dropout(dropout) function is used to randomly set a portion of the elements in the feature matrix to 0 (to prevent model overfitting; in this invention, the probability of setting to zero is dropout=0.3) to obtain a deactivated feature matrix (i.e., all elements in the feature matrix are 0). Finally, a second linear transformation is performed on the deactivated feature matrix and mapped back to the original dimension (i.e., the dimension of the input features of this module) to obtain the deep-processed feature matrix. And output it.

[0027] The input to the adaptive pooling module is the depth-processed feature matrix output from the feedforward neural network submodule in the multi-layer encoder stack module. This layer performs average pooling on the last dimension of the feature matrix (that is, compresses a sequence of arbitrary length into a feature of length 1) to obtain the pooled feature; then, it removes the last dimension of size 1 from the feature to obtain the compressed feature vector and outputs it.

[0028] The fully connected prediction module consists of two layers, and its specific structure is as follows: The first layer takes the compressed feature vector output by the adaptive pooling module as input. This layer performs a linear transformation on the compressed feature vector to obtain a linearly transformed feature vector. Then, the nn.GELU() activation function is used to perform a non-linear mapping on each element of the linearly transformed feature vector to obtain a feature vector with non-linear expressive power. Subsequently, the nn.Dropout(dropout=0.2) function is used to randomly set some elements to 0 to obtain a deactivated feature vector (i.e., all elements in the feature matrix are 0). Finally, a second linear transformation is performed on the deactivated feature vector, and it is mapped back to the original dimension (i.e., the dimension of the input features of this module) to obtain the feature vector processed by the first layer, which is then output. The second layer takes the feature matrix processed by the first layer as its input and processes the feature matrix in the same way as the first layer to obtain the final predicted value and output it.

[0029] Specifically, the dynamic Transformer model of this invention is obtained through the following steps: (3-1) Collect multiple sensor data (including pump speed, flow rate, pressure, power, concentration, displacement, time series, etc.) of the cutter suction dredger under different working conditions from actual engineering projects. Perform data cleaning (including outlier handling, missing value handling, time alignment handling) and sliding window processing on the multiple sensor data to obtain multiple time series samples. Divide all time series samples into training set, validation set and test set in a ratio of 7:1.5:1.5 (for subsequent model training, tuning and performance evaluation). It should be noted that the data preprocessing process in this step is exactly the same as in step (1), and will not be repeated here.

[0030] (3-2) For each time series sample in the training set obtained in step (3-1), the StandardScaler function is used to normalize it (to make the dimensions of different physical quantities consistent) to obtain the normalized time series sample corresponding to the time series sample, and feature selection processing is performed on the normalized time series sample to obtain the feature matrix B corresponding to the time series sample.

[0031] (3-3) For each time series sample in the training set obtained in step (3-1), the feature matrix B corresponding to the time series sample obtained in step (3-2) is input into the input feature embedding module, and the feature matrix B is mapped to a unified hidden layer space through a linear projection layer to obtain the time series feature matrix corresponding to the time series sample.

[0032] (3-4) For each time series sample in the training set obtained in step (3-1), the time series feature matrix corresponding to the time series sample obtained in step (3-3) is input into the position encoding module, and the time series feature matrix is ​​encoded in time position to obtain the sequence order information matrix corresponding to the time series sample (which can capture time order information).

[0033] Specifically, the positional encoding is constructed based on sine and cosine functions, and the encoding result is directly added to the temporal feature matrix (ensuring that the model captures temporal sequence information).

[0034] (3-5) For each time series sample in the training set obtained in step (3-1), the sequence order information matrix corresponding to the time series sample obtained in step (3-4) is input into the multilayer encoder stacking module to obtain the deep-processed feature matrix corresponding to the time series sample. (It includes both short-term and long-term operating condition dependencies).

[0035] Specifically, this module consists of stacked multi-layer encoders, each layer including a local contextual attention submodule, a global relational attention submodule, a memory enhancement submodule, and a feedforward neural network submodule. First, local contextual attention is calculated on the input sequence order information matrix, focusing on short-term operational changes to improve the model's responsiveness to dynamic operations, and outputting locally enhanced features. Then on Perform global correlation attention calculation to capture the global dependencies of the entire working condition sequence and output the feature matrix. Subsequently, a memory enhancement submodule is introduced to store and update the hidden states of historical time steps, retain long-term dependency information, and output a context feature matrix. Finally, the deep-processed feature matrix is ​​obtained through a feedforward neural network module. .

[0036] Furthermore, the local context attention submodule only calculates the attention weights for 15 time steps before and after each time step (i.e., window size = 15); the global association attention submodule reduces computation through random rotation and bucket matching (i.e., bucket size = 64); and the memory enhancement submodule sets the memory matrix size to 32, updates the coefficient to 0.1, and dynamically adjusts the memory matrix.

[0037] The advantage of this step (3-5) is that, through the memory enhancement submodule, the model can accurately capture important dependencies of long-term series information while focusing on short-term operating condition changes, thus solving the problem of insufficient long-term series capture in the field; at the same time, through the local context attention submodule and the global association attention submodule, the model can focus on global and local operating features, adapt to dynamic operating condition changes, and has dynamic adjustment capabilities.

[0038] (3-6) For each time series sample in the training set obtained in step (3-1), the feature matrix corresponding to that time series sample obtained in step (3-5) is... Input adaptive pooling module to the feature matrix Average pooling is performed along the time step dimension to obtain the compressed feature vector.

[0039] (3-7) For each time series sample in the training set obtained in step (3-1), the compressed feature vector obtained in step (3-6) is input into the fully connected prediction module to obtain the final prediction value corresponding to the time series sample.

[0040] Specifically, the compressed feature vector is passed through two linear layers, interspersed with GELU activation function and dropout layer (dropout probabilities of 0.3 and 0.2, respectively), to output the final predicted value.

[0041] (3-8) For each time series sample in the training set obtained in step (3-1), the coefficient of determination (also known as the goodness of fit, abbreviated as R) is calculated based on the final predicted value corresponding to the time series sample obtained in step (3-7). 2 ) and prediction loss functions (which include mean absolute error (MAE) and root mean square error (RMSE). Specifically, this step involves calculating the coefficient of determination and the prediction loss function using the following formulas:

[0042] in, This represents the total number of time-series samples in the training set; This represents the true value of the i-th time series sample; This represents the predicted value of the i-th time series sample; This represents the average of the true values ​​of all time-series samples. This represents the average of the predicted values ​​for all time-series samples.

[0043] (3-9) For each time series sample in the training set obtained in step (3-1), based on the determination coefficient and prediction loss function obtained in step (3-8), and using the Adam optimizer to perform gradient backpropagation and parameter update on the dynamic Transformer model until the number of iterations reaches a preset value (preferably 40 times in this invention), and obtain the optimal parameters of the dynamic Transformer model at this time, thereby obtaining the initially trained dynamic Transformer model; (3-10) Use the test set obtained in step (3-1) to test the dynamic Transformer model initially trained in step (3-9) until the prediction accuracy reaches the optimal level, thereby obtaining the final trained dynamic Transformer model.

[0044] Experimental results The CSD capacity prediction algorithm proposed in this invention demonstrates a significant performance improvement compared to traditional models. In this experiment, a dynamic Transformer framework based on memory enhancement and dual attention mechanisms was adopted. This framework outperforms the traditional Transformer model on multiple metrics and exhibits more accurate and robust prediction capabilities.

[0045] During the experiment, the improved model proposed in this invention was compared with the traditional Transformer model. By predicting CSD capacity under different operating conditions, the model of this invention demonstrated superior performance across all performance metrics. The specific prediction accuracy is shown in the table below:

[0046] As can be seen from the table above, the model of this invention outperforms the traditional Transformer model in all metrics, especially in R² and RMSE, where it improves by 12.2% and reduces by 17.2%, respectively. In addition, the model of this invention also shows a significant reduction in MAE. Figure 2 The changes in the objective function before and after data preprocessing are shown. Figure 3 The results of multiple feature selection are shown. Figure 4 and Figure 5 The prediction results processed by the algorithm of this invention are shown. Experimental results and image displays demonstrate that the method of this invention outperforms traditional prediction models in both accuracy and stability of capacity prediction. Specifically, the memory enhancement submodule helps the model retain historical equipment data, while the dual attention mechanism optimizes the model's ability to recognize various dynamic operating conditions of the equipment, thereby significantly improving the overall performance of the model. Accurate prediction results provide real-time decision support for equipment managers, significantly improving work efficiency.

[0047] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the production capacity of a cutter suction dredger based on a dynamic Transformer model, characterized in that, The method comprises the following steps: (1) Collecting multiple sensor data of the cutter suction dredger through sensors arranged on the cutter suction dredger, and preprocessing all the sensor data to obtain multiple preprocessed sensor data; (2) Processing the preprocessed sensor data obtained in step (1) to obtain a feature matrix B; (3) Inputting the feature matrix B obtained in step (2) into a pre-trained dynamic Transformer model to obtain a production capacity prediction result D of the cutter suction dredger.

2. The cutter suction dredger production capacity prediction method based on the dynamic Transformer model according to claim 1, wherein the sensor data comprises slurry flow, pump speed, vacuum degree, pipeline pressure, cutting power, and ship position speed; The preprocessing process in step (1) is specifically as follows: first, all the sensor data are subjected to abnormal value detection, missing value filling and data smoothing processing in sequence to obtain preliminarily processed sensor data; and then, the preliminarily processed sensor data are subjected to sliding window processing to obtain multiple preprocessed sensor data; The feature selection processing in step (2) comprises, in sequence, collinearity removal processing, automatic matching selection method, and manual screening, wherein the collinearity removal processing is a variance inflation factor VIF test method, the automatic matching selection method is a SelectKBest algorithm based on F test, and the manual screening is to supplement important but missing feature variables according to the operation principle and yield influence mechanism of the cutter suction dredger.

3. The cutter suction dredger production capacity prediction method based on the dynamic Transformer model according to claim 1 or 2, wherein the dynamic Transformer model comprises an input feature embedding module, a position encoding module, a multi-layer encoder stacking module, an adaptive pooling module, and a full-connection prediction module; The input of the input feature embedding module is a multi-dimensional time series feature matrix with a dimension of 64*15*197, and this layer maps the multi-dimensional time series feature matrix into a unified hidden layer space to obtain a time series feature matrix with a dimension of 64*15*197; 4. The cutter suction dredger production capacity prediction method based on the dynamic Transformer model according to any one of claims 1 to 3, wherein the multi-layer encoder stacking module comprises two layers of encoders, and each layer of encoder comprises a local context attention sub-module, a global correlation attention sub-module, a memory enhancement sub-module, and a feedforward neural network sub-module, and the specific structure is as follows: The local context attention sub-module comprises six layers, namely a first linear projection layer, a first multi-head segmentation layer, a local window extraction layer, a local attention weight calculation layer, a local context aggregation layer, and a first multi-head fusion and linear projection layer: . The input of the position encoding module is a time sequence feature matrix with a dimension of 64*15*197 output by the feature embedding module, the layer generates a fixed position vector in the form of a sine-cosine function, and the fixed position vector is added to the time sequence feature matrix element by element to obtain a sequence order information matrix where R denotes the set of real numbers.

5. The cutter suction dredger production capacity prediction method based on the dynamic Transformer model according to claim 4, wherein ​ ​ The first linear projection layer, the input is the sequence order information matrix output by the position encoding module The layer firstly maps the matrix through three independent linear mapping layers to obtain a query matrix , a key matrix , and a value matrix respectively; A first multi-head split layer, which inputs the query matrix output by the first linear projection layer , the key matrix , and the value matrix , divides each matrix into 4 subspaces along the feature dimension, and the dimension of each subspace is 8, so that the shape of the divided matrix is , to obtain the multi-head divided query matrix , the key matrix , and the value matrix ; The local window extraction layer takes as input the query matrix after multi-head partitioning output from the first multi-head partitioning layer. Key matrix and value matrix This layer is for each time step in each matrix. In other words, define a local context window Extract the local key matrix corresponding to each time step within this window. With local value matrix To obtain the query vector corresponding to each time step. Local bond matrix and local value matrix ,in Represents window size, in this invention =15; The local attention weight calculation layer takes the query vector corresponding to each time step as input from the local window extraction layer. and local bond matrix This layer is based on the query vector. and local bond matrix Calculate the attention score matrix And the attention score matrix is ​​processed using the Softmax function. Normalization is performed to obtain the attention weight matrix for each time step. ; a local context aggregation layer that inputs the attention weights corresponding to each time step output by the local attention weight computation layer and the local value matrix corresponding to each time step output by the local window extraction layer and performs weighted summation of the two to obtain a context representation corresponding to each time step and a set of local context features corresponding to each time step ; a first multi-head fusion and linear projection layer, which takes as input the local context feature set corresponding to each time step of the output of the local context aggregation layer This layer splices and reshapes the local context feature set corresponding to all time steps , and performs linear mapping processing on the processing result to obtain a local enhanced feature matrix and outputs. ​ The global correlation attention sub-module includes eight layers, which are a second linear projection layer, a second multi-head segmentation layer, a rotating barrel division layer, a mask generation layer, an attention score calculation layer, an attention weight calculation layer, a context feature generation layer, and a second multi-head fusion and linear projection layer: A second linear projection layer, which takes the local enhanced feature matrix output by the first multi-head fusion in the local context attention sub-module and the linear projection layer as input , and performs linear projection on the local enhanced feature matrix to obtain a query matrix , a key matrix , and a value matrix ; A second multi-head split layer, input of which is the query matrix output by the second linear projection layer , the key matrix , and the value matrix , processes the query matrix , the key matrix , and the value matrix in the same way as the first multi-head split layer in the local context attention sub-module, and finally outputs the multi-head split query matrix , the key matrix , and the value matrix ; A rotation bucket division layer, which takes as input the query matrix output by the second multi-head split layer and the key matrix The layer first performs matrix multiplication on the query matrix , the key matrix and the random rotation matrix initialized by the rotation bucket division layer to obtain the randomly projected query matrix and the randomly projected key matrix, and then performs bucket division processing on the randomly projected query matrix and the randomly projected key matrix respectively to obtain the query bucket number matrix and the key bucket number matrix respectively; a mask generation layer that takes as input the query bucket number matrix output by the rotated bucket division layer and the key bucket number matrix The layer expands the query bucket number matrix and the key bucket number matrix by an unsqueeze function to obtain an attention mask attn_mask; a attention score calculation layer, which inputs the query matrix output by the second multi-head segmentation layer and the key matrix , and an attention mask attn_mask output by a mask generation layer, which first performs matrix multiplication processing on the query matrix and the transpose matrix of the key matrix to obtain an original attention score; then, multiplies the original attention score by a scaling factor; Finally, the masked_fill function is used to set the score of the position with a mask value of 0 in the multiplication result to -1e9 to obtain the masked attention score; The attention weight calculation layer takes the masked attention score output by the attention score calculation layer as input, and normalizes the score to obtain the attention weight; generate a context feature layer, which inputs a value matrix output by the second multi-head segmentation layer and attention weights output by an attention weight calculation layer, which performs matrix multiplication on the value matrix and the attention weights to obtain a context feature ; a second multi-head fusion and linear projection layer, which inputs the context features generated by the context feature layer and outputs a feature matrix . The second multi-head fusion and linear projection layer processes the context features in the same way as the first multi-head fusion and linear projection layer in the local context attention sub-module and outputs a feature matrix .

6. The dynamic Transformer model-based cutter suction dredger productivity prediction method according to claim 5, characterized in that, The memory enhancer module includes four layers, which are a third linear projection layer, a memory attention score calculation layer, a memory attention weight calculation layer, and a memory output generation layer: The third linear projection layer takes as input the feature matrix from the second multi-head fusion and the output of the linear projection layer in the global association attention submodule. This layer has this feature matrix The processing method is the same as that of the first linear projection layer in the local context attention submodule, and the final output is a query matrix. Key matrix and value matrix ; a memory attention score calculation layer, which inputs the query matrix output by the third linear projection layer , and a memory matrix with a dimension of a number of memory slots * 197, which performs matrix multiplication processing on the query matrix and the transposed matrix of the memory matrix to obtain a memory attention score; The memory attention weight calculation layer takes the memory attention score output by the memory attention score calculation layer as input, and normalizes the memory attention score to obtain the memory attention weight; A memory output layer is generated, which takes as input the memory attention weight output by the memory attention weight calculation layer and a memory matrix with a dimension of the number of memory slots * 197, and performs matrix multiplication processing on the memory attention weight and the memory matrix to obtain a context feature matrix and output.

7. The dynamic Transformer model-based cutter suction dredger productivity prediction method according to claim 6, characterized in that, The input of the feedforward neural network submodule is the context feature matrix generated by the memory enhancer submodule The layer first performs linear transformation on the context feature matrix to obtain a linearly transformed feature matrix Then, the nn.GELU() activation function is used to perform non-linear mapping processing on each element in the linearly transformed feature matrix to obtain a feature matrix with non-linear expression capability; subsequently, a part of elements in the feature matrix are randomly set to 0 by the nn.Dropout function to obtain a deactivated feature matrix; finally, the deactivated feature matrix is subjected to a second linear transformation processing and mapped back to the original dimension to obtain a deeply processed feature matrix and output; The input of the adaptive pooling module is a deep-processed feature matrix output by a feedforward neural network submodule in the multi-layer encoder stacking module The layer performs average pooling processing on the last dimension feature in the feature matrix to obtain a pooled feature, and then removes the last dimension with a size of 1 in the feature to obtain a compressed feature vector and output The fully connected prediction module includes two layers, and the specific structure is as follows: The first layer takes the compressed feature vector output by the adaptive pooling module as input, performs linear transformation on the compressed feature vector to obtain a linearly transformed feature vector, then uses the nn.GELU() activation function to perform nonlinear mapping on each element in the linearly transformed feature vector to obtain a feature vector with nonlinear expression capability, then randomly sets some elements to 0 through the nn.Dropout function to obtain an inactivated feature vector, and finally performs a second linear transformation on the inactivated feature vector and maps it back to the original dimension space to obtain the first layer processed feature vector and output; The second layer takes the first layer processed feature matrix as input, and the processing method of the second layer is the same as that of the first layer to obtain the final prediction value and output.

8. The cutter suction dredger production capacity prediction method based on the dynamic Transformer model according to claim 7, characterized in that, The dynamic Transformer model is obtained by the following steps: (3-1) Collect multiple sensor data of the cutter suction dredger under different working conditions from actual engineering projects, perform data cleaning and sliding window processing on the multiple sensor data to obtain multiple time series samples, and divide all the time series samples into a training set, a validation set and a test set according to a ratio of 7:1.5:1.5; (3-2) For each time series sample in the training set obtained in step (3-1), use the StandardScaler function to normalize it to obtain a normalized time series sample corresponding to the time series sample, and perform feature selection processing on the normalized time series sample to obtain a feature matrix B corresponding to the time series sample; (3-3) For each time series sample in the training set obtained in step (3-1), input the feature matrix B corresponding to the time series sample obtained in step (3-2) into the input feature embedding module, and map the feature matrix B to a unified hidden layer space through a linear projection layer to obtain a time series feature matrix corresponding to the time series sample; (3-4) For each time series sample in the training set obtained in step (3-1), input the time series feature matrix corresponding to the time series sample obtained in step (3-3) into the position encoding module, and perform time position encoding on the time series feature matrix to obtain a sequence order information matrix corresponding to the time series sample; (3-5) For each time series sample in the training set obtained in step (3-1), input the sequence order information matrix corresponding to the time series sample obtained in step (3-4) into the multi-layer encoder stack module to obtain the deep-processed feature matrix corresponding to the time series sample ; (3-6) For each time series sample in the training set obtained in step (3-1), the feature matrix corresponding to the time series sample obtained in step (3-5) is input into an adaptive pooling module to perform average pooling in the time step dimension, thereby obtaining a compressed feature vector. The input adaptive pooling module to perform average pooling in the time step dimension on the feature matrix , thereby obtaining a compressed feature vector. (3-7) For each time series sample in the training set obtained in step (3-1), input the compressed feature vector obtained in step (3-6) into the full connection prediction module to obtain a final prediction value corresponding to the time series sample; (3-8) For each time series sample in the training set obtained in step (3-1), obtain the decision coefficient R of the final prediction value corresponding to the time series sample according to step (3-7) 2 and a prediction loss function, which includes mean absolute error (MAE) and root mean square error (RMSE); (3-9) For each time series sample in the training set obtained in step (3-1), according to the determination coefficient and the prediction loss function obtained in step (3-8), and using the Adam optimizer, gradient back propagation and parameter updating are performed on the dynamic Transformer model until the iteration number reaches a preset value, and the optimal parameters of the dynamic Transformer model at this time are obtained, thereby obtaining a dynamic Transformer model preliminarily trained; (3-10) Test the dynamic Transformer model preliminarily trained in step (3-9) using the test set obtained in step (3-1) until the prediction accuracy obtained reaches the optimal value, thereby obtaining a final trained dynamic Transformer model.

9. The cutter suction dredger production capacity prediction method based on the dynamicized Transformer model according to claim 8, characterized in that, Step (3-8) is to calculate the determination coefficient R using the following formula 2 , mean absolute error MAE and root mean square error RMSE: , wherein, represents the total number of time series samples in the training set; represents the true value of the i-th time series sample; represents the predicted value of the i-th time series sample; represents the average value of all time series sample true values; represents the average value of all time series sample predicted values.

10. A dynamic Transformer model-based cutter suction dredger production capacity prediction system, characterized in that, Comprise: A first module for collecting a plurality of sensor data of a cutter suction dredger through sensors arranged on the cutter suction dredger, and preprocessing all the sensor data to obtain preprocessed sensor data; A second module for performing feature selection processing on all the preprocessed sensor data obtained by the first module to obtain a feature matrix B; A third module for inputting the feature matrix B obtained by the second module into a pre-trained dynamic Transformer model to obtain a production capacity prediction result D of the cutter suction dredger.