A large caisson excavation instruction prediction method based on a multi-label classification algorithm

Abstract

The application discloses a large-scale sinking well excavation instruction prediction method based on a multi-label classification algorithm, and comprises the following steps: step one, multiple data collection devices are arranged in the sinking well, all sinking well hole positions are covered, and data in a set time interval is collected, so that required samples for modeling are obtained, and a sample set is generated; step two, data analysis and visual analysis are performed on the sample set, and data cleaning is realized; step three, feature selection is performed on the cleaned data, and a feature subset required for modeling is obtained; step four, modeling is performed according to the feature subset, and multiple multi-label classification algorithm models for large-scale sinking well excavation instruction prediction are constructed; step five, the established models are selected and optimized, and an optimal prediction model for large-scale sinking well excavation instruction is obtained; and step six, the optimal prediction model is combined to predict the sinking well excavation instruction, and a decision basis for construction instruction measures is provided. The prediction method promotes accurate and effective large-scale sinking well excavation instruction, and reduces human risks.

Description

Technical Field

[0001] This invention relates to the field of large caisson construction technology. More specifically, this invention relates to a method for predicting excavation commands for large caissons based on a multi-label classification algorithm, applicable to the prediction of excavation commands for large caissons with arbitrary multi-hole locations. Background Technology

[0002] During caisson construction, soil is excavated or sludge is removed from the caisson, and the caisson sinks to the design elevation by its own weight, overcoming friction against the caisson walls and resistance at the cutting edge. Due to its good seismic performance, small footprint, high load-bearing capacity, and strong overall integrity, it has been widely used in the construction of large bridges such as the Changtai Yangtze River Bridge, the Shanghai-Nantong Yangtze River Bridge, and the Wufengshan Yangtze River Bridge. However, complex geological conditions and increasingly larger caisson foundations pose greater challenges to the precise and effective execution of excavation commands for large caissons.

[0003] Currently, there are two main methods for controlling the construction of large caissons: numerical analysis using finite element analysis (FEM) and decision-making based on human experience. FEM has long been one of the most effective methods for studying the stress and deformation during caisson sinking, and most current FEM analyses for caisson foundation construction focus on the stress and strain of the caisson structure under specific construction conditions, such as the characteristics of soil resistance and deformation during sinking. However, due to the complexity of geological conditions and actual working conditions, the results of numerical analysis often differ from the actual sinking situation. Decision-making based on human experience relies heavily on past construction experience, and the judgments of different managers vary and are difficult to quantify. Furthermore, this method lacks reliable reference data. Other methods, such as the automated caisson construction method (Super Open Caisson System), employ technologies such as prefabricated segment assembly of the caisson, automatic excavation, soil removal, and control of the caisson's attitude, greatly improving the automation level of the construction process. However, the decision-making process for borehole excavation still relies on subjective human judgment. Therefore, given the rapid development of intelligent construction technology, and considering the lack of reliability in existing borehole excavation decision-making methods and the actual needs of various engineering projects for caisson construction technology, there is an urgent need to find a decision-making method that combines statistical analysis with human experience. Therefore, this application proposes a large-scale caisson excavation command prediction method based on a multi-label classification algorithm to solve the problems existing in the prior art. Summary of the Invention

[0004] One objective of this invention is to provide a method for predicting large-scale caisson excavation commands based on a multi-label classification algorithm, so as to promote the accuracy and effectiveness of large-scale caisson excavation commands and reduce human risk.

[0005] To achieve these objectives and other advantages according to the present invention, a method for predicting large-scale caisson excavation instructions based on a multi-label classification algorithm is provided, comprising the following steps:

[0006] Step 1: Deploy multiple data collection devices inside the caisson, covering all caisson openings and collecting data at set time intervals to obtain the samples required for modeling and generate a sample set;

[0007] Step 2: Perform data analysis and visualization analysis on the sample set to achieve data cleaning;

[0008] Step 3: Perform feature selection on the cleaned data to obtain the feature subset required for modeling;

[0009] Step 4: Model based on feature subsets and construct a multi-label classification algorithm model for predicting multiple large-scale caisson excavation commands;

[0010] Step 5: Select and optimize the established model to obtain the best prediction model for large caisson excavation commands;

[0011] Step Six: Combine the best prediction model to predict the caisson excavation instructions, and provide a decision-making basis for the implementation of construction instructions.

[0012] Preferably, the data collection device in step one includes several stress sensors installed on the inclined surface and bottom surface of the caisson cutting edge, and a Beidou and GPS positioning system installed on the caisson wall, thereby acquiring several stress data and several caisson subsidence data as feature counts, forming a matrix of several samples, and generating a sample set.

[0013] Preferably, the data cleaning method in step two is as follows:

[0014] First, perform statistical analysis on the sample set obtained in step one to obtain information on missing values, duplicate values, and outliers in the sample set;

[0015] Secondly, samples with missing or duplicate values ​​are removed, and outliers are identified using box plots to obtain the final normal samples, thus achieving data cleaning.

[0016] Preferably, the specific method for feature selection of the cleaned data in step three is as follows:

[0017] First, based on human experience, sample characteristics corresponding to data collection devices damaged during on-site construction are eliminated;

[0018] Secondly, the feature importance index based on the random forest algorithm is used to evaluate the importance of the remaining features, thereby obtaining the relative importance and ranking of each feature. According to the design requirements, the top-ranked features are selected.

[0019] Finally, the top-ranked features are searched in reverse using a sequence reverse selection algorithm. Based on the change in the accuracy index of the base model, one feature is removed at a time. The importance of the feature is determined by the change in model performance before and after the removal of the feature, until the model performance is stable and the feature subset required for modeling is obtained.

[0020] Preferably, the specific modeling method in step four is as follows: model construction is carried out based on three strategies: problem transformation method, adaptive algorithm and ensemble method, and five types of models are established: based on the Binary Relevance framework, modeling with Naive Bayes, SVM and KNN as base models respectively; based on the Classifier Chains framework, modeling with Naive Bayes, SVM and KNN as base models respectively; based on the Label Powerset framework, modeling with Naive Bayes, SVM and KNN as base models respectively; based on the adaptive algorithm framework, modeling using MLKNN and MLTSVM algorithms respectively; and based on the Bagging ensemble learning framework, modeling using the Random Forests algorithm.

[0021] Preferably, the model selection and optimization in step five includes:

[0022] 1) Model evaluation and performance comparison: Select the model with the best overall performance from step four. The specific method is as follows:

[0023] Based on K-fold cross-validation and multi-label classification algorithm model evaluation metrics, we evaluated and compared the performance of various algorithm models to determine the model with the best overall performance. The evaluation metrics for the multi-label classification algorithm used were accuracy, F1 score, Jaccard similarity coefficient, and Hamming loss. The model with relatively high accuracy, F1 score, and Jaccard similarity coefficient, and relatively low Hamming loss was selected as the model with the best overall performance.

[0024] 2) Based on the model with the best overall performance obtained above, the features are normalized using MinMaxScaler, StandardScaler, RobustScaler and feature scaling methods respectively. Combining the model evaluation index in step 1), the normalization method that is most helpful to improve the model performance in the current task is selected, and the model features are normalized.

[0025] 3) Based on the Bayesian optimization method, the hyperparameters of the above model are optimized to obtain the optimal parameters of the model, and the final optimal prediction model for large caisson excavation commands is obtained.

[0026] Preferably, the model selection and optimization in step five, between steps 2) and 3), also includes model improvement, specifically: combining different distance metrics, such as Euclidean distance, Manhattan distance, and Chebyshev distance, to improve the above-mentioned optimal model, and obtain the final evaluation index of the model on the test set, and select an appropriate distance metric to improve the model.

[0027] The present invention has at least the following beneficial effects:

[0028] (1) This invention accurately considers the various characteristic factors that affect the accuracy of the large caisson excavation command prediction algorithm and their linear or nonlinear relationships; combined with exploratory data analysis, the dataset is cleaned and transformed in many ways, eliminating redundancy and noise in the original data, and laying the data foundation for data-driven methods.

[0029] (2) This invention performs efficient feature selection on the collected feature set, proposes and implements a combined feature selection method, which fully integrates and utilizes the combined feature selection method of human experience + feature ranking + search strategy, reduces data dimensionality, and improves modeling efficiency and feature stability.

[0030] (3) This invention takes into account the complexity of the implementation of excavation command for large caissons with multiple holes and the interaction between holes. It adopts a multi-label classification algorithm for modeling and prediction, taking into account the inherent relationship between each label (hole), achieving overall consideration and fully improving the accuracy and reliability of large caisson excavation command prediction.

[0031] (4) This invention combines cross-validation and multiple model evaluation indicators to compare and evaluate multiple models and determine the model with the best overall performance; and further conducts model hyperparameter tuning and model optimization, which greatly improves the model prediction accuracy and the stability and reliability of practical application, provides accurate and effective auxiliary decision-making basis for the implementation of large caisson excavation instructions, helps to greatly reduce the impact of human factors, and provides strong guarantee for construction quality and construction safety.

[0032] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the overall modeling and prediction process of the present invention.

[0034] Figure 2 This is a schematic diagram of the stress sensor arrangement according to the present invention;

[0035] Figure 3 This is a schematic diagram of the well hole distribution of the present invention;

[0036] Figure 4 This is a schematic diagram of the data sample of the present invention;

[0037] Figure 5 This is a schematic diagram of the three-step combined feature selection of the present invention;

[0038] Figure 6 This is a schematic diagram of the sequence reverse selection algorithm based on the search strategy of the present invention;

[0039] Figure 7 This is a schematic diagram illustrating the effect of normalization processing in this invention;

[0040] Figure 8 The true value is the sample for the model-assisted decision-making of this invention.

[0041] Figure 9 This is a predicted value for the model-assisted decision-making sample of the present invention;

[0042] Figure 10 The true value of sample two for model-assisted decision making in this invention;

[0043] Figure 11 The predicted value for sample two of the model-assisted decision-making in this invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0045] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified. In the description of this invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0046] Example

[0047] like Figure 1 As shown, a method for predicting large-scale caisson excavation instructions based on a multi-label classification algorithm includes the following steps:

[0048] Step 1: Deploy multiple data collection devices within the caisson, covering all caisson openings and collecting data at set time intervals to obtain the samples needed for modeling and generate a sample set. A schematic diagram of the caisson opening distribution is shown below. Figure 3As shown; each data collection device serves as a feature number, and multiple data collection devices correspond to multiple feature numbers; the data collection device includes several stress sensors installed on the cutting edge of the caisson, as well as Beidou and GPS positioning systems. Stress data is acquired using the stress sensors to monitor the reaction force of the soil at the bottom of the caisson and the stress condition of the caisson structure itself during the caisson sinking process. This helps to understand the soil resistance encountered during the caisson sinking process and ensure the caisson structure's attitude balance. A schematic diagram of one specific arrangement of the stress sensors is shown below. Figure 2 As shown, the three-dimensional geometric coordinates of several points on the caisson are monitored in real time using the BeiDou and GPS positioning systems. The program automatically calculates and thus monitors the changes in the caisson's tilt, horizontal displacement, relative elevation, and horizontal torsion, among other aggregate attitude information. Taking 5 subsidence data points and 125 cutting edge stress data points as an example, 38 caisson boreholes are set, corresponding to 38 label columns. A schematic diagram of the data sample is shown below. Figure 4 As shown. The purpose of the modeling is to predict the 38 label values ​​corresponding to each sample, thus it is a multi-label classification problem. Before step one, it also includes determining the task type of predicting the excavation command for a large caisson. Based on machine learning theory and engineering business scenarios, the problem is analyzed and abstracted, and it is transformed into a common task type of machine learning. Considering the task requirements and the difference between regression and classification tasks, the current task is determined to be a classification task, so a machine learning classification algorithm is chosen. After on-site investigation and task scenario analysis, it is known that the excavation command for 38 holes in a large caisson needs to be predicted simultaneously, i.e., whether to excavate or not (labeled as 1 and 0 respectively). Therefore, the current task is further determined to be a multi-label classification task, so a multi-label classification algorithm is chosen.

[0049] Step 2: Perform data analysis and visualization analysis on the sample set to achieve data cleaning;

[0050] First, perform statistical analysis on the sample set obtained in step one to obtain information on missing values, duplicate values, and outliers in the sample set; then perform exploratory data analysis on the sample set, including data statistical description (obtaining mean, extreme values, variance, quantiles, data types, etc.) and visualization analysis, and obtain information on missing values, duplicate values, and outliers in the sample set through statistical analysis.

[0051] Secondly, samples with missing or duplicate values ​​were removed, and outliers were identified using box plots to obtain the final normal samples, thus achieving data cleaning. Considering the complex on-site environment and the significant noise present during data acquisition, samples with missing or duplicate values ​​were removed; outliers were identified using box plots to obtain normal samples whose values ​​fall between the minimum and maximum estimates, thus achieving data cleaning.

[0052] Step 3: A combined feature selection method integrating human experience, feature ranking, and search strategies was used to reduce data dimensionality and improve modeling efficiency and feature stability. Figure 5 As shown.

[0053] First, based on human experience, sample characteristics corresponding to data collection devices damaged during on-site construction are eliminated;

[0054] Secondly, the importance of the remaining features is evaluated using the feature importance index based on the Random Forests algorithm, thereby obtaining the relative importance and ranking of each feature. According to the design requirements, the top-ranked features are selected to identify the relatively important features.

[0055] Finally, a sequential backward selection algorithm was implemented by combining a greedy algorithm and following Ticham's razor principle. It performs a backward search on the top-ranking features, removing one feature at a time based on changes in the base model accuracy. The importance of a feature is determined by the change in model performance before and after its removal, continuing until the model performance stabilizes. This process selects fewer features while ensuring model accuracy, reducing computational overhead by lowering data dimensionality and improving model timeliness. A curve showing the change in base model accuracy with the number of features is plotted to obtain the final feature subset required for modeling. Figure 6 As shown, the feature set selected through manual experience and feature ranking is used as input. Using model accuracy as a reference metric, one feature is removed from the input feature set each time, and the change in model accuracy is observed and recorded. If the model accuracy loss is zero or the relative change is small, it indicates that the feature's importance is relatively low and it is removed. This process is repeated until all features have been traversed. The final number of expected feature values ​​is equivalent to the minimum number of features required for a model to achieve the same high level of accuracy. Figure 6 For example, the expected number of features is 3; finally, combining the expected number of features with the feature importance evaluation ranking based on the random forest algorithm, the corresponding number of features with the highest ranking are selected.

[0056] Step 4: Model based on feature subsets and construct a multi-label classification algorithm model for predicting multiple large-scale caisson excavation commands;

[0057] Models are constructed based on three different approaches to multi-label classification problems: Binary Relevance (BR), Classifier Chains (CC), and Label Powerset (LP), adaptive algorithms (MLKNN, MLTSVM, etc.), and ensemble methods (Random Forests, RF). Considering the relatively small amount of data, the base models for the BP, CC, and LP frameworks are Naive Bayes. Bayes (NB), Support Vector Machine (SVM), and K Nearest Neighbors (KNN).

[0058] (1) Based on the BR framework, modeling is performed using NB, SVM, and KNN as base models respectively. The BR framework-based method, also known as binary association, utilizes all sample sets X to model each label Y. i Prediction is performed on (i∈{1, 2, ..., 38}), therefore 38 base models are established respectively. Finally, the results are summarized and analyzed to obtain a comprehensive result. Let the models be BR_NB, BR_SVM, and BR_KNN;

[0059] (2) Based on the CC framework, modeling is performed using NB, SVM, and KNN as base models respectively. The CC framework-based method is also known as a classifier chain. The first classifier uses the sample geometry X to predict the first label Y1, represented as (X, Y1); the second classifier also uses the input and output of the previous classifier in the classifier chain, represented as (X, Y1, Y2); and so on, until all labels are included, i.e., (X, Y1, Y2, ..., Y1). 38 Similar to (1), the difference is that the CC framework-based method preserves the correlation between the labels. Let the models be CC_NB, CC_SVM, and CC_KNN;

[0060] (3) Based on the LP framework, modeling is performed using NB, SVM, and KNN as base models respectively. The method based on the LP framework, also known as the LP method, assigns a single label to samples with the same label combination value. That is, it transforms the multi-label classification problem into a multi-class classification problem, achieving the assignment of a unique class label to each possible label combination. Let the models be LP_NB, LP_SVM, and LP_KNN;

[0061] (4) Based on the adaptive algorithm framework, MLKNN and MLTSVM algorithms are used for modeling respectively;

[0062] (5) Based on the Bagging ensemble learning framework, the Random Forests algorithm is used for modeling.

[0063] Step 5: Select and optimize the established model to obtain the optimal prediction model for large caisson excavation commands; including the following:

[0064] 1) Model evaluation and performance comparison: Select the model with the best overall performance from step four. The specific method is as follows:

[0065] Based on K-fold cross-validation and multi-label classification algorithm model evaluation metrics, we evaluated and compared the performance of various algorithm models to determine the model with the best overall performance. The evaluation metrics used for the multi-label classification algorithm were accuracy, F1 score, Jaccard similarity coefficient, and Hamming loss. The model with relatively high accuracy, F1 score, and Jaccard similarity coefficient, and relatively low Hamming loss was selected as the model with the best overall performance.

[0066] The indicators are defined as follows:

[0067]

[0068] Where I is the indicator function, when the model predicts the value f(x) i ) and the true value y i The value is 1 when they are equal (i.e., the excavation command predicted by the model is consistent with the actual excavation command), otherwise it is 0.

[0069]

[0070] Where P is precision, which represents the proportion of samples that are actually true out of all samples that are predicted to be true (i.e., the proportion of samples that are actually excavated out of all samples that are predicted to be excavated); R is recall, which represents the proportion of samples that are exactly predicted to be true out of all samples that are actually true (i.e., the proportion of samples that are exactly excavated out of all samples that are actually excavated).

[0071]

[0072]

[0073] Where I is the indicator function, the same as (4); n is the number of tags (i.e., the total number of holes, 38).

[0074] 2) Based on the model with the best overall performance obtained above, the features are normalized using MinMaxScaler, StandardScaler, RobustScaler, and feature scaling methods respectively. Combining this with model evaluation metrics, the normalization method most conducive to improving model performance in the current task is selected, and feature normalization is then performed; for example... Figure 7 The diagram shows the effect of a normalization method before and after processing.

[0075] Normalization methods are used to standardize each feature to a uniform scale, eliminating the influence of different dimensions between features, while also accelerating computation, speeding up gradient descent, and promoting model convergence. Several normalization methods are compared. The normalization methods used include: MinMaxScaler, StandardScaler, RobustScaler, and custom feature scaling. The definitions of each normalization method are as follows:

[0076] MinMaxScaler:

[0077] Where, x min and x max These represent the minimum and maximum feature values ​​of the current data, respectively; MinMaxScaler performs a linear transformation on the original data, mapping the result to the range [0, 1], thus achieving proportional scaling of the original data;

[0078] StandardScaler:

[0079] Where μ and σ represent the mean and standard deviation of the current feature, respectively, and StandardScaler maps the data to a distribution with a mean of 0 and a standard deviation of 1.

[0080] The RobustScaler method removes the median and scales the data according to the quantile interval (IQR by default). IQR is defined as... and The difference in quantiles. This is achieved by calculating relevant statistics on the training sample set, and then independently centering and scaling on each feature.

[0081] Custom feature scaling is similar to formula (8), but this method scales the data to the range [-1, 1], and its definition is as follows:

[0082]

[0083] 3) The hyperparameters of the above model are optimized using Bayesian optimization and random search parameter optimization methods to obtain the optimal parameters. A random search method is used to optimize some key hyperparameters of the optimal model obtained in step eight. To verify the reliability of the random search parameter tuning method, Bayesian optimization is used as a comparison. During the parameter tuning process, the parameters to be tuned and candidate parameter values ​​are used as the key and value of a dictionary data structure, respectively, and the entire dictionary data structure is used as the elements of a list; that is, the parameters to be tuned are defined as a list dictionary to speed up the search. Cross-validation is also used during the parameter tuning process to ensure the stability and reliability of the results. Example code for defining the list dictionary is as follows:

[0084]

[0085]

[0086] The model selection and optimization in step five above, between steps 2) and 3), also includes model improvement. Specifically, it involves improving the optimal model by combining different distance metrics, such as Euclidean distance, Manhattan distance, and Chebyshev distance, and obtaining the final evaluation index of the model on the test set. A suitable distance metric is then selected for model improvement. The multi-label classification algorithm based on an adaptive algorithm framework needs to calculate the distance between samples for sample classification. By employing multiple distance metrics to reflect sample differences at different levels, the optimal metric is sought to improve model performance. The distance metrics used are Euclidean distance, Manhattan distance, and Chebyshev distance, etc., and the data points are denoted as... The distances are defined as follows:

[0087]

[0088]

[0089]

[0090] Optimize the model based on the distance metrics listed above, and select the most suitable distance metric for model optimization.

[0091] Step Six: Combine the best prediction model to predict the caisson excavation command, providing a decision-making basis for construction command implementation. Test the best prediction model obtained in the above steps. For example, if the model predicts the result of 38 labels for a sample as [0,1,0,,,1], then the hole corresponding to the label with a prediction result of 0 will not be excavated, and the hole corresponding to the label with a prediction result of 1 will be excavated. At this point, the prediction model for large caisson excavation commands based on a multi-label classification algorithm has been constructed and tested, achieving a high degree of integration between statistical analysis and human experience, thus providing reliable auxiliary decision-making for on-site construction command implementation. A schematic diagram of the model's auxiliary decision-making is shown below. Figures 8 to 11 As shown in the figure, the actual sample labels and the prediction results of the multi-label classification model are examples. Policymakers can use the model prediction results as an auxiliary reference for decision-making. Figure 8 For sample one, the true value, Figure 9 For sample 1, the predicted value is... Figure 10 This represents the true value of sample two. Figure 11 For example, the dark gray mark indicates the hole location with a label value of 1, indicating that all model predictions are correct. Therefore, all corresponding holes need to be excavated. The personnel implementing the measures can combine engineering experience with the model prediction results for comprehensive evaluation, thereby achieving precise control over the on-site construction measures.

[0092] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A large caisson excavation instruction prediction method based on a multi-label classification algorithm, characterized in that, It comprises the following steps: Step one, a plurality of data collection devices are arranged in the open caisson, covering all the open caisson hole positions and collecting data within a set time interval, so as to obtain the required samples for modeling and generate a sample set; the data collection device comprises a plurality of stress sensors arranged on the inclined surface and the bottom surface of the open caisson blade, and a Beidou and GPS positioning system arranged on the well wall, so as to obtain a plurality of stress data and a plurality of open caisson sinking amount data as characteristic numbers, form a plurality of matrix samples, and generate a sample set; Step two, data analysis and visualization analysis are performed on the sample set to realize data cleaning; Step three, feature selection is performed on the cleaned data to obtain a feature subset required for modeling; Step four, modeling is performed according to the feature subset to construct a plurality of large open caisson excavation instruction prediction multi-label classification algorithm models for predicting the excavation instruction of the large open caisson hole, i.e., the label is dig or not dig; Step five, the established model is selected and optimized to obtain the best prediction model of the large open caisson excavation instruction; Step six, the best prediction model is combined to predict the open caisson excavation instruction, providing a decision basis for construction instruction strategy. 2.The large caisson excavation instruction prediction method based on a multi-label classification algorithm of claim 1, wherein, The data cleaning method in step two is specifically as follows: First, statistical analysis is performed on the sample set obtained in step one to obtain information of missing values, repeated values and abnormal values in the sample set; Second, the samples with missing values and repeated values are removed, and the abnormal values are identified and processed by using a box plot, so as to obtain the final normal samples and realize data cleaning. 3.The large caisson excavation instruction prediction method based on multi-label classification algorithm of claim 1, wherein, The specific method of feature selection on the cleaned data in step three is as follows: First, the sample features corresponding to the damaged data collection devices in the field construction process are removed in combination with artificial experience; Second, the importance of each feature is evaluated by using the feature importance index based on the random forest algorithm, so as to obtain the relative importance and ranking of each feature, and a plurality of features with high rankings are selected according to the design requirements; Finally, the selected features with high rankings are searched in reverse by using a sequential reverse selection algorithm, and the change of the base model accuracy index is taken as the basis, one feature is removed each time, and whether the feature is important is determined according to the change of the model performance before and after the removal of the feature, until the model performance is stable, and a feature subset required for modeling is obtained. 4.The method of claim 1, wherein the method is characterized by, The specific method of modeling in step four is as follows: based on the problem conversion method, the adaptive algorithm and the integration method, five types of models are constructed, which are based on the Binary Relevance framework, the Naive Bayes, the SVM and the KNN as the base model for modeling; based on the Classifier Chains framework, the Naive Bayes, the SVM and the KNN as the base model for modeling; based on the Label Powerset framework, the Naive Bayes, the SVM and the KNN as the base model for modeling; based on the adaptive algorithm framework, the MLKNN and the MLTSVM algorithms are used for modeling; based on the Bagging ensemble learning framework, the Random Forests algorithm is used for modeling. 5.The large caisson excavation instruction prediction method based on multi-label classification algorithm of claim 4, wherein, The model selection and optimization in the fifth step includes: 1) Model evaluation and performance comparison, select the model with the best comprehensive performance in step four, the specific method is: Based on K-fold cross-validation and multi-label classification algorithm model evaluation index, the model evaluation and performance of each algorithm model are compared, and a model with the best comprehensive performance is determined. The multi-label classification algorithm evaluation indexes used are accuracy, F1 value, Jaccard similarity coefficient and Hamming loss. Select the model with relatively high accuracy, F1 value and Jaccard similarity coefficient and relatively small Hamming loss as the model with the best comprehensive performance; 2) Based on the above obtained model with the best comprehensive performance, normalize the features by using MinMaxScaler, StandardScaler, RobustScaler and feature scaling method respectively. Combine the model evaluation indexes in step 1) above to select the normalization method that is most helpful to improve the model performance in the current task, and perform normalization processing on the model features; 3) Based on the Bayesian optimization method, the hyperparameters of the above model are optimized to obtain the best parameters of the model, and the best prediction model of large caisson excavation instruction is obtained.

Images