Trailing suction dredger load prediction method based on cavity LSTM algorithm

Abstract

The invention belongs to the technical field of ship intelligent control and energy management, and particularly relates to a drag suction dredger load prediction method based on a cavity LSTM algorithm, and the method comprises the steps: collecting the operation data of a dredger, and carrying out the sliding window preprocessing of the collected data, and obtaining a data set; establishing a load prediction model of multi-scale time characteristics by adopting an LSTM (Long Short Term Memory) algorithm improved by cavity convolution; performing layer normalization optimization on the load prediction model, and training the optimized model through a training set; and inputting the data set into the optimized load prediction model, comparing a prediction result with an actual result, adjusting the number of iterations and the learning rate of the model according to the deviation between a prediction value and actual power, and then testing and comparing again until the prediction result tends to be consistent with the actual result. Through the parallel LSTM structure with the voidage of [1, 3, 7], the long-time-sequence power prediction precision is remarkably improved, and dredger energy-saving optimization control can be effectively guided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent control and energy management of ships, and particularly relates to a drag suction dredger load prediction method based on a cavity LSTM algorithm. BACKGROUND

[0002] As important equipment in the fields of water conservancy projects, port construction and environmental governance, the drag suction dredger is of great importance to the improvement of operation efficiency and the reduction of energy consumption. The traditional prediction methods such as empirical formula and linear regression are difficult to effectively capture the nonlinear dynamic characteristics and time sequence dependence in the operation process. The existing prediction methods based on LSTM still have the limitations of gradient disappearance, gradient explosion and single-scale time sequence modeling, resulting in insufficient prediction accuracy under complex working conditions. In particular, the standard LSTM model cannot simultaneously consider multi-scale features such as instantaneous fluctuations, short-term trends and long-term changes, which seriously restricts the prediction performance. In view of these problems, it is urgent to develop an innovative prediction method that can effectively model nonlinear dynamic characteristics, enhance long-term sequence dependence learning ability and consider multi-scale time features to meet the needs of precise operation and energy saving optimization of the dredger. SUMMARY

[0003] The purpose of the application is to solve the problem of insufficient prediction accuracy under complex working conditions caused by the limitations of gradient disappearance, gradient explosion and single-scale time sequence modeling in the existing drag suction dredger load prediction method, and to propose a drag suction dredger load prediction method based on a cavity LSTM algorithm.

[0004] In order to achieve the above purpose, the application adopts the following technical solutions:

[0005] The drag suction dredger load prediction method based on the cavity LSTM algorithm comprises the following steps:

[0006] S1: Collecting the parameters of the dredge pump power, propulsion power, speed, water depth, wave height and wind speed in the operation process of the drag suction dredger;

[0007] S2: Preprocessing the collected data by sliding window to obtain a data set, which is divided into a training set and a test set. The sliding window method is a common time sequence data processing technology in the prior art, which can divide the time sequence data into multiple fixed-length subsequences for subsequent analysis and modeling;

[0008] S3: Using a parallel cavity LSTM network architecture to establish a load prediction model with multi-scale time features. The model effectively expands the receptive field of the network through the cavity convolution technology to capture longer time range dependencies, thereby improving the prediction accuracy of load changes under complex operation environments;

[0009] S4: Perform layer normalization optimization on the load prediction model, and train the optimized model with the training set. Through these optimization methods, the model can maintain high efficiency in a variable working environment, enhance the robustness of the model, and improve the stability and accuracy of the prediction;

[0010] S5: Input the test set into the optimized load prediction model, compare the predicted results with the actual results, and adjust the training strategy to improve the model. The predicted value can be used to dynamically adjust the operation parameters, thereby achieving more accurate control and energy optimization of the dredger operation load.

[0011] As a further preferred solution, in S2, the sliding window preprocessing includes noise enhancement processing, which adds random noise to the already obtained data set. This noise simulates the inevitable mechanical noise fluctuations during ship operation, improving the robustness and accuracy of the model in the actual environment. The noise formula is as follows:

[0012] Y(k,l) = Y(k,l) + beta * avgY(l) * 2random

[0013] Where k is the number of samples, l is the number of thermal parameters, beta is used to control the strength of random noise, beta is 0.01, avgY represents the average value of the lth parameter, and 2random is a random number of 2 fluctuations.

[0014] As a further preferred solution, in S2, the data set obtained by preprocessing is divided into 85% of the training set and 15% of the test set according to the time sequence.

[0015] As a further preferred solution, in S3, first introduce a hole convolution processing based on the LSTM model to obtain a hole LSTM model, enhance the learning ability of multi-scale time features, and then pass the hole LSTM model through layer normalization processing to obtain a fusion layer normalization hole LSTM optimization model to improve the stability of the training process and speed up the convergence speed.

[0016] As a further preferred solution, introduce a hole convolution processing based on the existing LSTM model, including: insert a fixed interval inside the convolution kernel to expand the input feature map area covered by the convolution kernel, so as to capture more wide context information, and the calculation formula is:

[0017] k' = k + (d-1) * (k+1)

[0018] Where k' is the convolution kernel size after the convolution kernel is expanded by the hole, d is the expansion rate, and k is the initial convolution kernel size.

[0019] As a further preferred solution, a dilated convolution process is introduced on the basis of the existing LSTM model, which also includes: introducing a dilated factor in the standard LSTM so that the current time step t depends on the time step C. t-d The hidden state H t-d and cell state C t-d , d is called the void factor, and the improved void LSTM input gate calculation formula is:

[0020] i t =σ(W i ·[H t-d ,X t ]+b i )

[0021] Where: i t is the input gate, σ is the Sigmoid function, W i is the weight, X t is the current input, H t-d is the unit output at the previous td moment, b i is the bias term.

[0022] As a further preferred solution, layer normalization processing is introduced on the basis of the void LSTM model, including: introducing layer normalization processing before each gating unit of the void LSTM model, and normalizing the input data of each time step along the feature dimension, so that the input of each time step maintains a stable mean and variance, effectively preventing the values ​​of certain dimensions from being too large or too small, causing the sigmoid / tanh activation function to enter the saturation region.

[0023] As a further preferred option, the test set is input into the optimized model, and the model automatically outputs the future change trends of relevant parameters (mud pump power, propulsion power, speed, water depth, wave height, and wind speed parameters), and compares the performance of the model based on the deviation between the predicted results and the actual results of the test set; if the deviation is large, adjust the number of iterations and learning rate of the model, and then retest and compare until the predicted results are consistent with the actual results, so as to improve the prediction accuracy of the model and further enhance the performance of the model in the actual operating environment.

[0024] Compared with the existing technology, the present invention proposes a load prediction method for a trailing suction dredger based on the void LSTM algorithm, and proposes an improved void convolution LSTM method, which can effectively model nonlinear dynamic characteristics, enhance long time series dependency learning capabilities, and take into account multi-scale time characteristics, thereby improving the accuracy of load prediction, specifically including the following advantages.

[0025] 1. Multi-scale time series feature extraction capability: by adopting a parallel LSTM structure with dilated rates of [1, 3, 7], the application can simultaneously capture multi-scale time features such as instantaneous fluctuations (dilated rate 1), short-term trends (dilated rate 3) and long-term changes (dilated rate 7), compared with traditional single-scale LSTM model, the prediction accuracy is significantly improved, which can effectively solve the problem of insufficient adaptability of standard LSTM under complex working conditions.

[0026] 2. Innovatively introducing layer normalization processing before the LSTM gate unit, cooperating with the dynamic noise suppression algorithm, the gradient vanishing and gradient explosion problems of the model in long time series training are significantly improved, the training convergence speed is improved by 40% compared with the traditional LSTM, and the prediction accuracy can be maintained above 98% under different noise intensities. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a receptive field comparison chart, wherein (a) is ordinary convolution, (b) is dilated convolution;

[0028] Figure 2 is a dilated LSTM model structure diagram;

[0029] Figure 3 is a power prediction flowchart. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all.

[0031] The improved LSTM neural network model fusing dilated convolution and layer normalization of the application is used for accurate prediction of the working load of a drag suction dredger, as shown in Figure 3 The specific implementation steps are as follows:

[0032] S1: For the research object of the application, the drag suction dredger, collect its running data and perform sliding window preprocessing to obtain a data set;

[0033] In real ship navigation, there are various noise interferences, such as vibrations and noises generated by mechanical parts such as generators and pumps. In order to make the data obtained by power prediction more real, random noise is added on the basis of the data obtained, and the noise formula is as follows:

[0034] Y(k,l)=Y(k,l)+β·avgY(l)·2random

[0035] In the formula, k is the number of samples, l is the number of thermal parameters, β is used to control the strength of random noise, β is 0.01 in this section, avgY represents the average value of the lth parameter, and 2random is a random number of 2 up and down fluctuations.

[0036] S2: On the basis of the LSTM model, the hole convolution and layer normalization processing method are introduced, and a new Dilated LSTM model is designed, as shown in Figure 2 , so as to improve the prediction accuracy and robustness of the algorithm. The specific improvement steps are:

[0037] S2-1: Perform hole convolution operation, insert fixed interval (hole rate) in the convolution kernel to expand the input feature map region covered by the convolution kernel, so as to capture more wide context information. While keeping the invariance of model parameters, the range of receptive field is significantly increased. The receptive field of ordinary convolution and hole convolution is as shown in Figure 1 , it can be seen that the hole mechanism enables neurons to skip intermediate time steps and directly connect to more distant historical information. The calculation formula is:

[0038] k' = k + (d-1) x (k+1)

[0039] Where: k' is the convolution kernel size after the convolution kernel is expanded by holes, d is the expansion rate, and k is the initial convolution kernel size.

[0040] S2-2: Improve the structure of the LSTM network model. In the standard LSTM, the hidden state H t and cell state C t of the current time step t depend on the H t and C t of the previous time step t-1. The standard LSTM input gate formula is:

[0041] i t = σ(W i ·[H t-1 ,X t ]+b i )

[0042] Where: i t is the input gate, σ is the Sigmoid function, W i is the weight, X t is the current input, H t-d is the unit output at time t-d, and b i is the bias term.

[0043] In order to enable the network to directly learn more distant historical information, the hole mechanism is introduced in the standard LSTM, so that the current time step t depends on the hidden state H t-dand cell state C t-d Instead of t-1. Here d is called the dilation factor. The improved dilated LSTM input gate calculation formula is:

[0044] i t = σ(W i · [H t-d , X t ] + b i )

[0045] Instead of stacking L standard LSTM layers to obtain L-step historical dependencies, the dilated mechanism enables neurons to skip intermediate time steps and directly connect to more distant historical information. A layer with a dilation factor of d has an effective receptive field size of (kernel1_size-1)*dilation_rate+1. In LSTM, the "kernel size" is implicitly 1 (only dependent on the previous step), but through the skip connection, the receptive field of the d layer is significantly expanded to d*L (L is the number of layers).

[0046] The data at each time step is processed in parallel through three layers of LSTM units with different dilation rates (1, 3, 7), as shown in Figure 1 This multi-scale architecture enables the network to simultaneously capture local details (dilation rate 1), short-term patterns (dilation rate 3), and long-term trends (dilation rate 7), forming a pyramid-like time perception ability, solving the single-scale time series modeling limitations of standard LSTM models that cannot simultaneously consider multi-scale features such as instantaneous fluctuations, short-term trends, and long-term changes.

[0047] After S2-1 and S2-2 dilated convolution operations, the dilated LSTM model is obtained.

[0048] S2-3: Introduce layer normalization (Layer Normalization, LN) processing before each gating unit (input gate, forget gate, output gate) of the dilated LSTM model. This technology normalizes the input data along the feature dimension to obtain a dilated LSTM optimization model with layer normalization, keeping the mean and variance of each time step input stable, effectively preventing the values of some dimensions from being too large or too small, causing the sigmoid / tanh activation function to enter the saturation zone, and solving the problem of gradient disappearance and gradient explosion in LSTM-based prediction methods.

[0049] S3: Model training. The preprocessed data is divided into the first 85% as the training set and the remaining 15% as the test set. Then use the training set to train the dilated LSTM optimization model with layer normalization. When the model meets the accuracy requirement or reaches the maximum number of iterations, the model is output and saved.

[0050] S4: model testing. The test set is input into the fused layer normalized hollow LSTM optimization model, the model automatically outputs the future trend of the relevant parameters, and according to the deviation between the prediction result and the actual result of the test set, the performance of the model is compared; if the deviation is large, the iteration number and learning rate of the model are adjusted, and then the test comparison is retested until the prediction result and the actual result tend to be consistent, so as to improve the prediction accuracy of the model and further improve the performance of the model in the actual working environment.

[0051] The present application is a load prediction method for a trailing suction dredger based on a hollow LSTM algorithm. The dredger operation mud pump power, propulsion power and environmental working condition data are preprocessed by a sliding window to construct a training sample set as the model input, and the predicted value of the load is output. On the basis of the LSTM model, the hollow convolution, layer normalization processing and dynamic noise suppression method are introduced, thereby improving the prediction accuracy and robustness of the algorithm.

[0052] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for predicting the load of a trailing suction hopper dredger based on a void LSTM algorithm, characterized in that: The following steps are involved: S1: Collect the mud pump power, propulsion power, speed, water depth, wave height and wind speed parameter data of the trailing suction hopper dredger during operation; S2: Perform sliding window preprocessing on the collected data to obtain a data set, which is divided into a training set and a test set; S3: Using a parallel atrous LSTM network architecture, a load forecasting model with multi-scale temporal features is established. S4: Perform layer normalization optimization on the load forecasting model, and train the optimized model using the training set; S5: Input the test set into the optimized load forecasting model, compare the predicted results with the actual results, and adjust the training strategy to improve the model.

2. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 1, characterized in that: In S2, the sliding window preprocessing includes noise enhancement processing. Random noise is added to the acquired data set. The noise formula is as follows: Y(k,l)=Y(k,l)+β·avgY(l)·2random Where k is the number of samples, l is the number of thermal parameters, β is used to control the intensity of random noise, β takes a value of 0.01, avgY represents the average value of the lth parameter, and 2random is a random number that fluctuates up and down by 2.

3. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 2, characterized in that: In S2, the preprocessed dataset is used as the training set according to the first 85% of the time series, and the remaining 15% is used as the test set.

4. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 3, characterized in that: In S3, a dilated convolution process is first introduced on the basis of the LSTM model to obtain a dilated LSTM model, which enhances the learning ability of multi-scale temporal features. Then, the dilated LSTM model is processed by layer normalization to obtain a fusion layer normalized dilated LSTM optimization model, which improves the stability of the training process and accelerates the convergence speed.

5. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 4, characterized in that: Based on the existing LSTM model, we introduce dilated convolution processing, which includes inserting fixed intervals inside the convolution kernel to expand the input feature map area covered by the convolution kernel, thereby capturing broader context information. The calculation formula is: k′=k+(d-1)×(k+1) Where: k′ is the size of the convolution kernel after the convolution kernel is expanded with the void, d is the expansion rate, and k is the initial convolution kernel size.

6. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 5, characterized in that: The dilated convolution processing is introduced on the basis of the existing LSTM model, which also includes: introducing a dilated factor in the standard LSTM so that the current time step t depends on the time step C t-d The hidden state H t-d and cell state C t-d , d is called the void factor, and the improved void LSTM input gate calculation formula is: i t =σ(W i ·[H t-d ,X t ]+b i ) Where: i t is the input gate, σ is the Sigmoid function, W i is the weight, X t is the current input, H t-d is the unit output at the previous td moment, b i is the bias term.

7. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 6, characterized in that: Layer normalization processing is introduced on the basis of the hollow LSTM model, including: introducing layer normalization processing before each gating unit of the hollow LSTM model, and normalizing the input data of each time step along the feature dimension.

8. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 7, characterized in that: The input of each time step maintains a stable mean and variance, effectively preventing the sigmoid / tanh activation function from entering the saturation region due to the values ​​of certain dimensions being too large or too small.

9. The method for predicting the load of a trailing suction hopper dredger based on the void LSTM algorithm according to claim 7, characterized in that: The test set is input into the optimized model, and the model automatically outputs the future change trend of relevant parameters. The performance of the model is compared based on the deviation between the predicted results of the test set and the actual results.

10. The method for load prediction of a trailing suction hopper dredger based on a void LSTM algorithm according to claim 9, characterized in that: If the deviation is large, adjust the number of model iterations and learning rate, and then retest and compare until the predicted results are consistent with the actual results to improve the model's prediction accuracy and further enhance the model's performance in the actual operating environment.

Images