Method for predicting dredging difficulty in excavating process of cutter suction dredger in real time

Abstract

The invention relates to a real-time prediction method for dredging difficulty in the excavating process of a cutter suction dredger, and the method comprises the following steps: classifying construction parameters in a monitoring system, and dividing controllable parameters as construction data; selecting parameters capable of reflecting the excavation difficulty degree in real time, namely reamer cutting torque and winch transverse movement torque; preprocessing the construction data; respectively screening out key characteristics which have obvious influence on the cutting torque of the reamer and the transverse torque of the winch; creating a CNN-LSTM multivariable time sequence model, dividing construction data into a training set and a prediction set, and training the construction data; the key features related to the reamer cutting torque and the winch transverse movement torque are used as the input of the model, the reamer cutting torque and the winch transverse movement torque are used as the output, and real-time prediction is achieved. The method can assist an operator to evaluate the dredging operation difficulty in real time according to the dredging difficulty level real-time prediction value, and correspondingly adjust an excavation strategy and operation parameters, so that the control precision of dredging construction is improved.

Description

Technical Field

[0001] The invention relates to the technical field of dredging engineering, and in particular to a real-time prediction method for dredging difficulty during an dredging process of a cutter suction dredger. Background Art

[0002] Dredging is an important underwater earthwork excavation project, which uses specific machinery to excavate silt, gravel, rocks and other geotechnical materials at the bottom of rivers, lakes, waters and other water bodies, and transport them to designated areas. In this field, cutter suction dredgers have been widely used in port and waterway construction, seawall and beach reclamation, and marine environmental restoration due to their high efficiency, safety, continuous operation capability and good adaptability. However, the complex and changeable underwater soil characteristics often make it difficult for operators to accurately predict the difficulty of dredging operations before excavation sections, and determining appropriate excavation control parameters (such as traverse speed, auger speed, etc.) becomes a challenge. Excessive reliance on the operator's empirical judgment may lead to insufficient or excessive dredging operations, affecting the accuracy and efficiency of the project, and sometimes even damage mechanical equipment.

[0003] In current practical projects, underwater soil information is usually obtained through water geological survey drilling before construction to predict the difficulty of future dredging operations. However, due to the high geological costs, drilling construction is usually carried out at intervals of hundreds of meters, and the number of samples obtained is limited, making it difficult to fully and accurately describe the difficulty of the entire dredging area. In addition, human experts are somewhat subjective when classifying different types of soil, and the diverse naming methods also limit the accuracy of soil information, making it difficult to provide effective guidance for construction. Therefore, there is an urgent need to develop a new tool that can easily and accurately predict the difficulty of dredging operations in real time, providing a new way to solve the above problems. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the deficiencies existing in the prior art and provide a real-time prediction method for the dredging difficulty during the excavation process of a cutter suction dredger, so as to assist operators in optimizing the operation instructions during the construction process in advance.

[0005] The present invention is achieved through the following technical solutions:

[0006] A real-time prediction method for dredging difficulty during the excavation process of a cutter suction dredger comprises the following steps:

[0007] S1. Classify the construction parameters in the monitoring system and divide the controllable parameters as construction data;

[0008] S2. Select parameters that can reflect the difficulty of excavation in real time, including reamer cutting torque and winch traverse torque;

[0009] S3. Preprocessing the construction data;

[0010] S4. By using a hybrid method of the maximum mutual information coefficient and the Pearson correlation coefficient, the key features that significantly affect the reamer cutting torque and the winch traverse torque are respectively selected from the high-dimensional construction data;

[0011] S5. Create a CNN-LSTM multivariate time series model, divide the construction data into a training set and a prediction set, and use it to predict the reamer cutting torque and winch traverse torque;

[0012] S6. Through the CNN-LSTM multivariate time series model, the key features related to the reamer cutting torque and the winch lateral torque are respectively used as the input of the model, and the reamer cutting torque and the winch lateral torque are used as outputs to achieve real-time prediction.

[0013] According to the above technical solution, preferably, in step S1, the construction parameters in the monitoring system are classified into three categories: controllable parameters, feedback controllable parameters, and uncontrollable parameters, wherein the controllable parameters are used as construction data for data analysis.

[0014] According to the above technical solution, preferably, in step S3, the data preprocessing includes: using the Z-score method to detect outliers in the construction data; using the Lagrange interpolation method to interpolate missing values ​​and outliers in the construction data; and using the Savitzky-Golay filter to smooth the features.

[0015] According to the above technical solution, preferably, step S4 includes: using the maximum mutual information coefficient method to thin out all characteristic parameters of the cutter suction dredger and retaining the maximum relevant characteristic parameters; selecting the reamer cutting torque and the winch lateral torque as target vectors respectively, and using the Pearson correlation coefficient method to calculate and select the features with the highest correlation for the selected maximum relevant characteristic parameters, eliminating redundant parameters, and screening out key features that have a significant impact on the reamer cutting torque and the winch lateral torque.

[0016] According to the above technical solution, preferably, in step S5, CNN, RNN and LSTM models are used to compare the CNN-LSTM multivariate time series model, and RMSE, MAPE and R 2 The indicators are used to evaluate the prediction effect of the model.

[0017] According to the above technical solution, preferably, in step S6, the reamer cutting torque and the winch lateral torque are divided into different dredging difficulty levels according to their numerical ranges; the predicted values ​​of the reamer cutting torque and the winch lateral torque are mapped to different dredging difficulty levels to achieve real-time prediction of dredging difficulty.

[0018] According to the above technical solution, preferably, in step S6, for different dredging difficulties, specific operation suggestions for key control parameters of reamer rotation speed and traverse speed are given according to the numerical ranges of reamer cutting torque and winch traverse torque.

[0019] The beneficial effects of the present invention are:

[0020] The present invention combines the working principle of the excavation process of a cutter suction dredger in dredging engineering with a deep learning algorithm. It does not rely on soil information. It utilizes the internal relationship between high-dimensional construction parameters and constructs a CNN-LSTM multivariate time series prediction model to propose a real-time prediction method for the dredging difficulty of the excavation process of a cutter suction dredger.

[0021] On the one hand, the present invention solves the problem that it is difficult to fully and accurately describe the difficulty of the entire dredging area due to sparse drilling, effectively saving the construction cost of dredging geological exploration; on the other hand, it avoids the limitations brought by the subjectivity of human experts in classifying different soil types, such as the difficulty in considering the spatial variability of soil, and provides effective guidance for construction;

[0022] In addition, specific suggestions are provided for the detailed construction of different dredging difficulty levels, clarifying the adaptability of adjusting key construction parameters such as cutter speed and traverse speed under different dredging difficulty levels. Operators can instantly evaluate the difficulty of dredging operations based on the real-time prediction value of the dredging difficulty level, and adjust the excavation strategy and operating parameters accordingly, thereby improving the control accuracy of dredging construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of real-time prediction of dredging difficulty during the excavation process of the cutter suction dredger in the present invention.

[0024] Figure 2 It is a schematic diagram of the principle of construction parameter preprocessing in the present invention.

[0025] Figure 3 It is a schematic diagram of the neural network of the CNN-LSTM model in the present invention.

[0026] Figure 4 It is a schematic diagram of the prediction results of the CNN-LSTM model in the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and the best embodiment. Based on the embodiments in the invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the invention.

[0028] Embodiment 1: The present invention comprises the following steps:

[0029] S1. Classify the construction parameters in the monitoring system and divide the controllable parameters as construction data. Specifically, classify the construction parameters in the monitoring system into three categories: controllable parameters, feedback controllable parameters, and uncontrollable parameters, where the controllable parameters are used as construction data for data analysis.

[0030] S2. The excavation mechanism of the cutter suction dredger is explored, and parameters that can reflect the difficulty of excavation in real time are selected, including the reamer cutting torque and the winch lateral torque.

[0031] S3. Perform data preprocessing on the construction data. The data preprocessing includes:

[0032] S31. Using the Z-score method to detect outliers in the construction data;

[0033] S32. Using Lagrange interpolation method to interpolate missing values ​​and abnormal values ​​in the construction data;

[0034] S33. The Savitzky-Golay filter is used to smooth the features, improve the smoothness of the data, and reduce the noise interference caused by the instability of the environment and the instrument itself.

[0035] S4. Through the hybrid method of the maximum mutual information coefficient and the Pearson correlation coefficient, the key features that significantly affect the reamer cutting torque and the winch traverse torque are respectively selected from the high-dimensional construction data. Specifically:

[0036] S41. The maximum mutual information coefficient method is used to thin out all characteristic parameters of the cutter suction dredger and retain the maximum relevant characteristic parameters;

[0037] S42. Select the reamer cutting torque and winch lateral torque as target vectors respectively, use the Pearson correlation coefficient method to calculate and select the features with the highest correlation for the selected maximum correlated feature parameters, eliminate redundant parameters, and screen out the key features that have a significant impact on the reamer cutting torque and winch lateral torque.

[0038] S5. Create a CNN-LSTM multivariate time series model and divide the construction data into a training set and a prediction set for predicting the reamer cutting torque and winch traverse torque.

[0039] Among them, convolutional neural network (CNN) has advantages in identifying and capturing short-term patterns or local fluctuations in sequence data with its excellent feature extraction ability and powerful information mining ability. Long short-term memory (LSTM) neural network is one of the popular methods, which solves the gradient vanishing problem in the recurrent neural network (RNN) training process by introducing interaction gates. Preferably, the division ratio of the training set and the test set in this example is 8:2.

[0040] In addition, CNN, RNN, and LSTM models are used to compare the CNN-LSTM multivariate time series model, and RMSE, MAPE, and R 2 The indicators are used to evaluate the prediction effect of the model.

[0041] S6. Through the CNN-LSTM multivariate time series model, the key features related to the reamer cutting torque and the winch lateral torque are respectively used as the input of the model, and the reamer cutting torque and the winch lateral torque are used as outputs to achieve real-time prediction.

[0042] In the real-time prediction process, the reamer cutting torque and winch traverse torque are divided into different dredging difficulty levels according to their numerical ranges; the predicted values ​​of the reamer cutting torque and winch traverse torque are mapped to different dredging difficulty levels to achieve real-time prediction of dredging difficulty. At the same time, specific operation suggestions for the key control parameters of the reamer speed and traverse speed can be given for different dredging difficulties according to the numerical ranges of the reamer cutting torque and winch traverse torque.

[0043] Example 2: Figure 1 As shown in the figure, a real-time prediction method for dredging difficulty in the excavation process of a cutter suction dredger based on a CNN-LSTM neural network includes the following steps: (1) Analyze the working principle of the excavation process of the cutter suction dredger to obtain two construction parameters that can reflect the dredging difficulty, namely, the reamer cutting torque and the winch lateral torque. (2) The operation data collected during the construction process by the automatic monitoring system of the cutter suction dredger is interpolated and completed by the Lagrange interpolation method to ensure the integrity of the data information, and then the data is smoothed and denoised by a filter. (3) The two construction parameters, reamer cutting torque and winch lateral torque, are used as target values, and the optimal feature set is selected from the high-dimensional monitoring data by a hybrid method of the maximum mutual information coefficient and the Pearson correlation coefficient, and the importance of the features is analyzed. (4) A CNN-LSTM multivariate time series prediction model is constructed for prediction, and evaluation indicators such as RMSE, MAPE, and R2 are used to compare the prediction effects of other algorithms on the target value. (5) The predicted values ​​of the reamer cutting torque and winch lateral torque are mapped to different dredging difficulty levels to achieve real-time prediction of dredging difficulty, and specific operational suggestions for the key control parameters of the reamer speed and lateral speed are given respectively.

[0044] Project Overview: A waterway excavation project was selected as a case project. According to the project data, the underwater soil conditions in the project area are complex, involving various soil types such as silt, clay, solid sandy soil, gravel, medium-weak weathered rock and strongly weathered rock. At the same time, it is affected by water scouring and sedimentation, and the spatial variability is large. Therefore, when carrying out waterway deepening and dredging operations, the cutter suction dredger faces great challenges. In the selected case project, a large cutter suction dredger was used for dredging construction. The cutter suction dredger has a cutter power of up to 4200kw and a maximum dredging capacity of up to 4500m3 / h, ranking among the top in the world. The construction parameters are automatically collected and saved every 2 seconds through multi-source sensors. Considering the differences in construction habits of different operators, data from an experienced operator within a 4-hour construction cycle was collected, totaling 7200 sets of 256-dimensional offline construction parameters.

[0045] (I) Analysis of excavation mechanism:

[0046] The excavation process of a cutter suction dredger mainly includes three key steps: cutting, traverse and stepping. First, the hard soil or rock is cut into pieces by rotating the cutter disc, and then the dredging pump sucks the material away. Some of the soil may overflow and be discharged to the sedimentation area through the pipeline. At the same time, by adjusting the speed and stop of the left and right traverse winches, the reamer is moved from one side of the hull to the other side, thereby changing the cutting position of the reamer on the cross section. This excavation process is carried out in layers, and the reamer needs to be raised and lowered with the help of a bridge. When the area to be excavated reaches the predetermined dredging depth, the stepping motion is achieved by the movement of the trolley. When the trolley reaches the maximum displacement position, the auxiliary pile is released to fix the hull position, and then the working pile is lifted by the sling device. Then the trolley is moved back to the minimum displacement position, the working pile is released, and the auxiliary pile is lifted again by the sling device to complete the entire step. The synergy of these technical means enables the cutter suction dredger to carry out excavation operations efficiently.

[0047] During the entire excavation process, the two steps of cutting and traversing need to adjust the construction parameters in real time according to different soil conditions. In contrast, the stepping process does not require real-time operation control and is generally not affected by soil conditions. The stepping distance and frequency are pre-set before construction. Therefore, the reamer cutting torque and winch traversing torque can effectively reflect the load capacity of the cutter suction dredger within a certain range. There are significant differences in different types of soil and their physical and mechanical properties at different positions and depths. During operation, the numerical values ​​of the reamer cutting torque and winch traversing torque will change in real time with the change of position. For example, for soft soils such as silt and fine sand, the torque that the cutter suction dredger needs to withstand is small, while when facing strongly weathered rocks, the torque that needs to be borne is larger. In addition, due to the spatial variability of soil, the torque required for the same soil at different depths and positions will also be different. The cutting process refers to the reamer destroying the cohesion of the soil through rotation, while the traversing process requires overcoming the lateral resistance of the soil, and the traversing winch swings left and right through the anchors fixed on both sides of the hull. Therefore, it is reasonable and feasible to evaluate the difficulty of dredging operations based on the size of the reamer cutting torque and the winch lateral torque.

[0048] (II) Data preprocessing:

[0049] Affected by various factors such as waves, gusts, forward momentum, mechanical vibration, etc., the sensor output results show extremely unstable characteristics. For example, parameters such as reamer speed, underwater pump suction vacuum, mud pump speed, etc. may have missing values, abnormal values ​​and noise due to factors such as monitoring instrument device layout, ambient temperature, sensitivity, etc. Before performing correlation analysis, the data must be processed to ensure the accuracy and reliability of the experimental results.

[0050] First, the missing values ​​in the original data set were analyzed and checked. Considering the complexity of the construction data, in order to preserve the integrity and time relevance of the data, the Lagrange interpolation method was chosen for processing instead of simply deleting outliers. The Lagrange interpolation method estimates missing data or outliers by constructing a polynomial that satisfies the interpolation conditions and accurately passes through all given data points, as shown in formula (1).

[0051]

[0052] In the formula, l i (x) is the Lagrangian basis function, defined as shown in formula (2).

[0053]

[0054] Each L i (x) all satisfy formula (3).

[0055]

[0056] This ensures that L n (x) At each data point x i The value at is equal to the corresponding (y i ), satisfying the interpolation conditions. The principle of Lagrange interpolation method is to use the properties of these basis functions to construct a polynomial that satisfies the interpolation conditions through linear combination to perform interpolation estimation of data.

[0057] In addition, the Z-score method is used to predict outliers in the original data set. Z-score is a commonly used statistical method to evaluate the degree of deviation between a data point and the mean. The data is judged as an outlier by calculating the standard deviation multiple between the data point and the mean, as shown in formula (4).

[0058]

[0059] Where x represents a single raw data value, μ represents the population mean, and σ represents the population standard deviation. If the Z-score is positive, it means that the data point is above the mean; if the Z-score is negative, it means that the data point is below the mean. Z-scores with larger absolute values ​​mean that the data point deviates more from the mean. In this study, the anomaly detection criterion is defined as follows: if Z-score<-3 or Z-score>3, the data point is considered an outlier. Outliers can be treated as missing values ​​and the same method as for missing values ​​is used, that is, Lagrange interpolation. The advantage of this method is that it uses information from existing variables to fill in outliers.

[0060] After processing missing values ​​and outliers, noise reduction is required. The noise in the original data usually appears as random noise. Savitzky-Golay filter (SG) is used for smoothing to improve the smoothness of the data and reduce noise interference. SG is a time domain filtering method based on local polynomial least squares fitting. Its advantage is that it keeps the shape and width of the sequence unchanged while filtering out noise. iΔz |-n￡i￡n} is considered, the fitting polynomial is shown in formula (5).

[0061]

[0062] In the formula, m is the coefficient of the fitting polynomial, b is k is the fitting coefficient, k is the order of the fitting term, and the residual sum of squares of the least squares fitting is shown in formula (6).

[0063]

[0064] In the formula, n is the number of samples, P m(i) is the fitted value, and T(i) is the actual value.

[0065] Taking the reamer motor voltage construction parameter as an example, the filtering process of all original construction parameters is introduced. Figure 2 This is a comparison of the reamer motor voltage before and after processing. After inspection, no missing values ​​were found in the 7200 data sets. In order to ensure the accuracy and reliability of the data, the Z-score method was used to detect and mark 65 groups of outliers, and the Lagrange interpolation method was used for processing, which effectively corrected these outliers and eliminated the possible interference. In addition, after filtering, it was observed that the change trend of the data remained basically unchanged, indicating that the filter used effectively retained the original characteristics of the data and made the data smoother, which was conducive to subsequent data analysis and information extraction, and further improved the data quality and reliability. Overall, by comprehensively processing various data anomalies, the data set was successfully optimized, providing a reliable basis for subsequent in-depth analysis and decision-making.

[0066] (III) Feature selection:

[0067] Feature selection refers to separating an optimal feature subset from the original feature space to achieve the purpose of dimensionality reduction. The present invention adopts the maximum mutual information coefficient (MIC)-Pearson correlation coefficient (PCC) hybrid method to select feature vectors. In view of the diversity and complexity of high-dimensional construction parameters in the CSD monitoring system, the MIC method is first used to thin out all the feature parameters of the CSD, and the largest relevant feature parameters are retained. Then, the reamer cutting torque and winch lateral torque are selected as target vectors respectively, and the Pearson correlation coefficient (PCC) method is used to calculate and select the features with the highest correlation, and redundant parameters are eliminated. This hybrid method is used to accurately extract the key parameters related to the target variable and eliminate redundant parameters.

[0068] MIC can measure the linear and nonlinear, functional and non-functional dependencies between features, and mine more dependencies between categories and features. The MIC method can effectively remove the redundancy between features. For a given feature set and target attribute, the MIC method measures the correlation between them by calculating the maximum information coefficient between each pair of features. The maximum information coefficient ranges from [0,1]. The larger the value, the stronger the correlation between the two features and the stronger the redundancy, as shown in formula (7).

[0069]

[0070] Where: MIC(x;y) is the maximum mutual information coefficient between variables; a and b are the number of grids divided in the x and y directions, respectively, which is essentially a grid distribution; B is a variable, B <n 0.6 , n is the amount of data.

[0071] Considering the high dimensionality of construction parameters, the selected features may be redundant during the MIC feature selection process, which will increase the inaccuracy of the training model. To solve this problem, the Pearson correlation coefficient (PCC) method is used to further eliminate the redundancy between features. PCC can measure the linear correlation between two random variable features and mine more dependencies between categories and features. For a given feature set and target attribute, the PCC method measures the correlation between them by calculating the normalized covariance between each pair of features. Its value range is [-1,1]. The larger the value, the stronger the correlation between the two features and the stronger the redundancy. The calculation principle of PCC is shown in formula (8).

[0072]

[0073] In the formula, x and y represent each pair of features, and They represent the sample means of the two respectively.

[0074] In this case, the monitoring system involves 256 dimensions of parameter indicators. Through the MIC algorithm, 15 features were selected for the reamer cutting torque and winch traverse torque under sparse conditions. However, among these features, there are some closely related situations, resulting in information redundancy. In order to solve this problem, the PCC method was used to measure the linear relationship between features. Taking the reamer cutting torque as an example, after verification, it was found that among the 15 features, there was an extremely high correlation coefficient between "left bridge winch motor current" and "left bridge winch torque", "concentration", "flow" and "hourly production rate", which was close to or even reached 1, so these redundant features were eliminated. In addition, features directly related to the reamer cutting torque were selected, such as "reamer power" and "reamer speed". However, from a professional perspective, these features are not suitable as input features because some data can be calculated by fixed mathematical formulas, such as the reamer cutting torque can be calculated by formula (9).

[0075]

[0076] Where T is torque, P is power, and N is speed; power is obtained by multiplying current and voltage. Therefore, these related variables (|r xy |>0.9) were eliminated. Considering the computational efficiency of the model, 6 features related to the reamer cutting torque were finally selected. A similar method was also used to select 15 features of the winch traverse torque, and 6 relevant features were finally determined. These features are mainly related to mechanics and do not involve mathematical formula conversion relationships between features, thereby improving the reliability of feature selection.

[0077] (IV) Model construction:

[0078] In this example, a CNN-LSTM multivariate time series model is constructed ( Figure 3 ), and use the corresponding feature sets to predict the reamer cutting torque and winch traverse torque in real time.

[0079] CNN has the ability to capture temporal and spatial dependencies, and can automatically learn and extract features that are closely related to the reamer cutting torque and winch traverse torque. The typical CNN architecture consists of multiple layers, including convolutional layers, pooling layers, dropout layers, and fully connected layers. The convolutional layer plays a vital role in feature extraction, capturing relevant features from the input data through convolution kernels, and the abstraction level of features increases with the increase in the number of convolution kernels. The use of maximum pooling technology and ReLU activation function helps to discard piecewise linear irrelevant features. The fully connected layer expands the pooled neurons into a one-dimensional vector form to facilitate more efficient management of data processing. However, CNN faces challenges in capturing the temporal dependencies of time series data in prediction tasks. Therefore, combining the recurrent neural network (RNN) technology and combining CNN with the long short-term memory (LSTM) network can improve performance.

[0080] LSTM overcomes the long-term dependency problem in RNN and can effectively alleviate the gradient vanishing problem. Therefore, in terms of time series feature prediction, LSTM can usually achieve significantly better results. The core of LSTM lies in its storage unit. The input signal is controlled by the input gate, forget gate and output gate. The forget gate selectively forgets the information of the previous unit through a 0-1 function, thereby retaining important information and ignoring unimportant information. The input gate determines which related signals should be added to the current process, while the output gate determines which signals should be used as the output of the current process. The specific calculation formulas are shown in (10)-(15).

[0081] f t =s(W f [h t-1 ,x t ]+b f )(10)

[0082] i t =s(W i [h t-1 ,x t ]+b i )(11)

[0083] c t ￠=tanh(W c [h t-1 ,x t ]+b c )(12)

[0084] o t =s(W o [h t-1 ,x t ]+b o )(13)

[0085] c t =c t-1 ×f t +i t ×c t ￠(14)

[0086] h t =tanh(c t )×o t (15)

[0087] In the formula, x is the input of the LSTM unit. h is the output of the LSTM unit. t is the LSTM

[0088] ￠

[0089] The unit number. t is the output of the forget gate. t and c t is the output and current state of the input gate. o t and c t is the output and current state of the output gate. W o and b o are the corresponding structural weights and biases. σ represents the sigmoid activation function, and tanh represents the hyperbolic tangent activation function.

[0090] (V) Prediction effect and comparison:

[0091] Considering that the time transmission of the monitoring system of the suction dredger selected in this case is 2s, it is necessary to lightweight the multivariate time series prediction model. The present invention sets the number of feature points close to the value to be predicted to 2, and the number of historical points of the input target value is 10. This is because the farther away from the point to be predicted, the smaller the utilization value may be, especially the multi-dimensional historical features. For multi-step long time series prediction, a recursive multi-step prediction strategy is adopted. The specific implementation of this strategy is as follows: First, a single-step prediction is performed using known historical data and models to obtain the predicted value of the next time step. Secondly, the prediction result is added to the historical data to form a new input sequence. Finally, the updated input sequence is used to make a prediction again to obtain the predicted value of the next time step. This cycle is iterated until the required multi-step prediction length is obtained.

[0092] Table 1 shows the comparison of evaluation indicators of different prediction models on the test set. The results show that RMSE, MAPE and R 2 These three performance indicators show consistent trends and magnitudes in all models. Among them, the CNN-LSTM model has the highest performance in RMSE, MAPE, and R 2 The performances in all aspects are the best, which are 293.12 and 38.80, 204.36 and 30.15, 0.9879 and 0.9898 respectively. By integrating the characteristics of CNN and LSTM, the CNN-LSTM model can more effectively capture the long-term dependencies and short-term patterns in time series data and achieve more accurate prediction performance.

[0093] Table 1 Rating indicators of different models

[0094]

[0095] Note: The data in the table are the evaluation indexes of reamer cutting torque and winch lateral torque.

[0096] Figure 4 The prediction results of the CNN-LSTM model are shown. In general, the prediction trend of the model is similar to the true value, showing good stability, among which the prediction effect of the winch lateral torque is better than the reamer cutting torque. Especially when the reamer cutting torque and winch lateral torque change suddenly, the CNN-LSTM model can still accurately capture the changes, indicating that the designed CNN-LSTM neural network model can effectively learn complex features.

[0097] (VI) Classification and prediction results:

[0098] According to the data distribution characteristics of reamer cutting torque and winch lateral torque, they are divided into four data intervals, and the numerical values ​​correspond to dredging difficulty levels I, II, III and IV respectively (as shown in Table 2). Then, the predicted values ​​of reamer cutting torque and winch lateral torque are mapped to the corresponding dredging difficulty level in real time. The analysis shows that the prediction accuracy of dredging difficulty level is 95.85%, which shows that CNN-LSTM can effectively predict the dredging difficulty of the excavation process of cutter suction dredger.

[0099] Table 2 Classification of dredging difficulty

[0100]

[0101] According to different dredging difficulty levels, operators can take corresponding operating instructions to cope with the challenges of different dredging operation difficulties. At level I, operators can follow the normal operating procedures for excavation and earthwork operations. At this time, the reamer speed and traverse speed can be set at a higher level to improve construction efficiency. At level II, considering that the reamer cutting torque range is expanded but the winch traverse torque does not change much, the operator should appropriately reduce the reamer speed and traverse speed to adapt to the characteristics of conventional soil. For level III, the operator needs to increase the winch power, significantly reduce the traverse speed, and keep the reamer speed close to level II. At level IV, the operator needs to be more cautious, keep the reamer speed and traverse speed within a relatively minimum range, regularly check the wear of the reamer, and should not pay too much attention to the construction progress because it may face challenges such as strongly weathered rocks. For emergencies, data quality checks should be carried out, special factors should be found out in combination with on-site conditions and other monitoring data, and early warnings and improvement measures should be put forward in a timely manner to ensure safe operation of equipment and construction efficiency.

[0102] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A real-time prediction method for dredging difficulty during the excavation process of a cutter suction dredger, characterized in that: The steps include: S1. Classify the construction parameters in the monitoring system and divide the controllable parameters as construction data; S2. Select parameters that can reflect the difficulty of excavation in real time, including reamer cutting torque and winch traverse torque; S3. Preprocessing the construction data; S4. By using a hybrid method of the maximum mutual information coefficient and the Pearson correlation coefficient, the key features that significantly affect the reamer cutting torque and the winch traverse torque are respectively selected from the high-dimensional construction data; S5. Create a CNN-LSTM multivariate time series model, divide the construction data into a training set and a prediction set, and use it to predict the reamer cutting torque and winch traverse torque; S6. Through the CNN-LSTM multivariate time series model, the key features related to the reamer cutting torque and the winch lateral torque are respectively used as the input of the model, and the reamer cutting torque and the winch lateral torque are used as outputs to achieve real-time prediction.

2. According to claim 1, a real-time prediction method for dredging difficulty during the excavation process of a cutter suction dredger is characterized in that: In step S1, the construction parameters in the monitoring system are classified into three categories: controllable parameters, feedback controllable parameters, and uncontrollable parameters, wherein the controllable parameters are used as construction data for data analysis.

3. The method for real-time prediction of dredging difficulty during excavation by a cutter suction dredger according to claim 1, characterized in that: In step S3, the data preprocessing includes: Using the Z-score method to detect outliers in the construction data; The Lagrange interpolation method is used to interpolate and complete the missing values ​​and abnormal values ​​in the construction data; The Savitzky-Golay filter is used to smooth the features.

4. The method for real-time prediction of dredging difficulty during excavation of a cutter suction dredger according to any one of claims 1 to 3, characterized in that: Step S4 includes: The maximum mutual information coefficient method is used to thin out all characteristic parameters of the cutter suction dredger and retain the maximum relevant characteristic parameters; The reamer cutting torque and winch lateral torque are selected as target vectors respectively. The Pearson correlation coefficient method is used to calculate and select the features with the highest correlation for the selected maximum correlated feature parameters, eliminate redundant parameters, and screen out the key features that have a significant impact on the reamer cutting torque and winch lateral torque.

5. The method for real-time prediction of dredging difficulty during excavation by a cutter suction dredger according to claim 4, characterized in that: In step S5, the CNN, RNN, and LSTM models are used to compare the CNN-LSTM multivariate time series model, and RMSE, MAPE, and R 2 The indicators are used to evaluate the prediction effect of the model.

6. The method for real-time prediction of dredging difficulty during excavation by a cutter suction dredger according to claim 5, characterized in that: In step S6, according to the numerical ranges of the reamer cutting torque and the winch traverse torque, they are respectively divided into different dredging difficulty levels; The predicted values ​​of reamer cutting torque and winch traverse torque are mapped to different dredging difficulty levels to achieve real-time prediction of dredging difficulty.

7. The method for real-time prediction of dredging difficulty during excavation by a cutter suction dredger according to claim 6, characterized in that: In step S6, specific operation suggestions for the key control parameters of the reamer rotation speed and the traverse speed are given according to the numerical ranges of the reamer cutting torque and the winch traverse torque for different dredging difficulties.

Images