CEEMDAN-CNN-LSTM-based heavy-load train long and large uphill working condition track quality index prediction method

Abstract

The invention discloses a CEEMDAN-CNN-LSTM-based heavy haul train long and large uphill working condition track quality index prediction method, which comprises the following steps: taking collected uphill TQI data as time sequence data, and predicting TQI information of a long and large uphill by utilizing a deep learning model and fully utilizing TQI collected data of the long and large uphill of a heavy haul train; the nonlinear TQI data is decomposed into relatively simple time sequence data based on the CEEMDAN; a hybrid convolutional neural network (CNN) and a long short-term memory neural network (LSTM) perform training learning on historical wind speed time series data of an uphill working condition, extract a relationship between a historical TQI value and a future TQI value, and verify model performance. According to the method, the prediction precision of the heavy-duty train long and large uphill working condition track quality index is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rail transit data analysis and prediction, and in particular, relates to a method for predicting track quality index of a heavy-load train in a long uphill condition based on CEEMDAN-CNN-LSTM. Background Art

[0002] Heavy-haul railway line operation and maintenance managers and intelligent operation and maintenance systems can quickly obtain the line track quality index (TQI), which is crucial to ensure the safe operation of heavy-haul trains. When a train is running under heavy load, the uphill section will be affected by a variety of forces, which will be reflected in the TQI value of the track.

[0003] In the actual operation and maintenance of heavy-load lines, the collection of track quality index mainly includes two methods: sensor measurement by inspection vehicle and manual measurement. The sensors of the inspection vehicle collect geometric status data of uphill sections by setting positions, while manual measurement mainly performs regular inspection and maintenance on sections with high detection results of the inspection vehicle. However, sensor measurement is easily disturbed by the location of line damage and the mud or dust generated by the inspection vehicle during operation, resulting in the inability of sensors to accurately measure track quality; manual measurement is difficult to cover all uphill sections, especially for sections with low values ​​or long distances, it is difficult to obtain accurate TQI values ​​in time. The combined influence of these factors restricts the comprehensive grasp of line safety operation information. Both methods can collect sufficient TQI data of uphill sections on a time scale during implementation. The long uphill condition has obvious working condition characteristics, and is affected by production and transportation activities. The collected TQI series data show nonlinear characteristics, and traditional statistical models are difficult to deal with such nonlinear time series data.

[0004] The existing patent publication number CN111310990A discloses a track quality prediction method and system based on an improved grey combination model, wherein the method includes raw data preprocessing, and also includes the following steps: constructing an improved non-equidistant grey IGM prediction model; optimizing the weights and thresholds of the Elman network using the PSO algorithm, and constructing a PSO-Elman neural network correction model; and constructing an IGM-PSO-Elman model by combining the advantages of the improved non-equidistant grey IGM model, the particle swarm optimization PSO algorithm, and the Elman neural network. The track quality prediction method and system based on the improved grey combination model proposed in the present invention are improved on the basis of the traditional grey model, and a first-order cumulative generation process of weight coefficient optimization is introduced, and a weight matrix is ​​added to optimize the model parameter solution. The integral area is used to optimize the calculation of the background value of the differential equation. The PSO-Elman residual correction model can reduce data fluctuations and reduce the improved non-equidistant grey prediction residual. The shortcoming of this patent is that the grey combination model is usually suitable for small samples and poor information situations, but has high requirements on the quality of the data (such as the continuity and regularity of the data). If the track quality data does not have a certain regularity or there is a lot of noise, the prediction accuracy of the model may be affected. Summary of the invention

[0005] The present invention is mainly aimed at the problem that the detection vehicle sensor in the prior art collects the geometric state data of the uphill section by setting the position, while the manual measurement is mainly for the sections with higher detection result values ​​of the detection vehicle for regular inspection and maintenance. However, the sensor measurement is easily interfered by the position of line damage and the mud or dust generated by the operation of the detection vehicle, resulting in the sensor being unable to accurately measure the track quality; manual measurement is difficult to cover the entire uphill section, especially for the sections with lower values ​​or longer distances, it is difficult to obtain the accurate TQI value in time. A method for predicting the track quality index of a heavy-load train long uphill condition based on CEEMDAN-CNN-LSTM is proposed.

[0006] In view of the above technical problems, the technical solutions of the present invention are as follows:

[0007] A method for predicting track quality index of a heavy-load train long uphill condition based on CEEMDAN-CNN-LSTM comprises the following steps:

[0008] S1. Input the historical data of TQI values ​​of long and steep slopes, and use the historical observation data of TQI values ​​of long and steep slopes as the original data for decomposition and model training;

[0009] S2. CEEMDAN decomposition of the time series data of TQI values ​​of long and steep ramps: Decompose the historical time series data of the original long and steep ramp TQI values ​​into multiple modal components IMF s ;

[0010] S3, normalizing the decomposed data of the long and steep slope time series TQI value;

[0011] S4, training set and test set division: the decomposed long and steep slope TQI data is divided into a training set and a test set;

[0012] S5, inputting each decomposed long and steep ramp TQI time series data into a convolutional neural network for convolution processing and extracting feature values;

[0013] S6. Input each extracted decomposed long and steep ramp TQI feature value into the long short-term memory neural network to train and predict the long and steep ramp TQI value, and simultaneously superimpose the prediction results to output the final result.

[0014] Furthermore, in step S2, the CEEMDAN decomposition of the long ramp TQI value time series data includes the following steps:

[0015] S2.1. Add a Gaussian white noise value ε with an amplitude of Δ to the original long and steep ramp TQI time series data. i (t), the expression is:

[0016] v i (t) = v (t) + Δε i (t)

[0017] In the above formula, v(t) is the original TQI value, v i (t) is the TQI value after the i-th processing;

[0018] S2.2, v i (t) Perform empirical mode decomposition and average all the first modal components obtained by decomposition to serve as the first modal decomposition of CEEMDAN decomposition. The expression of the average operation is:

[0019]

[0020] In the above formula, E 1 [v 1 (t)] is the first modal component obtained by empirical mode decomposition;

[0021] S2.3. Subtract IMF from the original uphill time series TQI data v(t) 1 (t), and the uphill residual TQI value is obtained, which is expressed as:

[0022] v res (t) = v(t) - IMF 1 (t)

[0023] In the above formula, v res (t) Uphill residual TQI value;

[0024] S2.4. Use the uphill residual TQI data as the original data and repeat steps S2.1 to S2.4 until the number of extreme value points of the uphill residual TQI does not exceed 2.

[0025] Furthermore, in step S3, the specific form of normalization of the uphill time series TQI decomposition data is as follows:

[0026]

[0027] In the above formula, v norm (t) is the normalized uphill TQI value.

[0028] Furthermore, in step S4, the training set is used for model training, and the test set is used for model verification.

[0029] Furthermore, the ratio of the training set to the test set is 9:1.

[0030] Furthermore, in model training, a sliding window method is used to generate training samples with a window size of 8, each sliding step is 1 step, and the next step is predicted.

[0031] Furthermore, in step S5, the convolutional neural network includes multiple one-dimensional convolutional layers and multiple one-dimensional MaxPooling pooling layers.

[0032] Furthermore, the long short-term memory neural network includes multiple recurrent units, each of which is composed of a forget gate, an input gate and an output gate.

[0033] Furthermore, the specific expressions of the forget gate, input gate, and output gate are:

[0034] f t =σ(w f [h t-1 ,v(t)]+b f )

[0035] i t =σ(w i [h t-1 ,v(t)]+b i )

[0036] C t '=tanh(w C [h t-1 ,v(t)]+b C )

[0037] C t =f t ×C t-1 +I t ×Ct '

[0038] o t =σ(w o [h t-1 ,v(t)]+b o )

[0039] h t =o t ×tanh(C t )

[0040] In the above formula, f t is the output of the forget gate; i t is the output of the input gate; C t ' is the unit information value; C t is the cell state; o t is the output of the output gate; σ is the sigmoid function; w f 、w i 、w C and w o are the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; h t-1 Output value for the previous unit; b f 、b i 、b C and b o is the corresponding bias term; h t Outputs a value for each cell.

[0041] Furthermore, in step S6, the uphill TQI prediction model is trained using the decomposed TQI time series data based on back propagation using the Adam optimizer.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention regards the collected TQI data of the long uphill working condition of heavy-duty trains as time series data, and uses a deep learning model to make full use of the TQI data of the uphill section to predict the TQI value of the slope. The long uphill working condition has obvious working condition characteristics, and affected by production and transportation activities, the collected TQI sequence data shows nonlinear characteristics; the CEEMDAN method is used to decompose the nonlinear TQI data into relatively simple time series data, and then the hybrid convolutional neural network (CNN) and the long short-term memory neural network (LSTM) are used to train and learn the historical TQI data of the uphill working condition, and extract the relationship between the historical TQI value and the future TQI value. Compared with the traditional statistical model, which is difficult to deal with such nonlinear time series data, the present invention improves the prediction accuracy of the track quality index of the long uphill working condition of heavy-duty trains; the deep learning model combining the hybrid convolutional neural network (CNN) and the long short-term memory neural network (LSTM) can make full use of the historical TQI data of the uphill working condition. This model performs well in processing time series data and can further improve the accuracy of prediction.

[0044] 2. The present invention can provide a basis for early warning and maintenance for railway operating departments by accurately predicting the TQI value under long uphill conditions. The proposed method for predicting the track quality index of heavy-load trains under long uphill conditions based on CEEMDAN-CNN-LSTM is of great significance to the safety management, production and intelligent development of heavy-load lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is the time series data of track quality status index of long uphill condition based on CEEMDAN decomposition;

[0047] Figure 3 This paper selects the prediction results of TQI measured data of some uphill sections of Shenshuo Line of Baoshen Railway in China. DETAILED DESCRIPTION

[0048] In order to clearly illustrate the technical features of the application scheme of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0050] In the present invention, unless otherwise clearly specified and limited, the first feature "on" or "under" the second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. In the description of this specification, reference terms; descriptions of "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" etc. mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples.

[0051] Example 1

[0052] like Figure 1 As shown, a method for predicting track quality index of a heavy-load train long uphill condition based on CEEMDAN-CNN-LSTM includes the following steps:

[0053] S1. Input the historical data of TQI values ​​of long and steep slopes, and use the historical observation data of TQI values ​​of long and steep slopes as the original data for decomposition and model training. In this embodiment, the TQI time series data actually measured on the uphill section of Shenshuo Line (with a slope greater than 6% and a length of more than 10 kilometers) are used. There are 3641 time series data, and some experimental data are shown in Table 1.

[0054] S2, such as Figure 2 As shown, the CEEMDAN decomposition of the long and steep ramp TQI value time series data: the original long and steep ramp TQI value historical time series data is decomposed into multiple modal components IMF s ; including the following steps:

[0055] S2.1. Add a Gaussian white noise value ε with an amplitude of Δ=0.05 to the original long and steep slope TQI time series data. i (t) is used to increase the complexity of the data and improve the decomposition effect. Its expression is:

[0056] v i (t) = v (t) + Δε i (t)

[0057] In the above formula, v(t) is the original TQI value, v i (t) is the TQI value after the i-th processing;

[0058] S2.2, v i(t) Perform empirical mode decomposition, extract the first modal component, and repeat this process 100 times. Perform an average operation on all the first modal components obtained by decomposition to serve as the first modal decomposition of CEEMDAN decomposition. The expression of the average operation is:

[0059]

[0060] In the above formula, E 1 [v 1 (t)] is the first modal component obtained by empirical mode decomposition;

[0061] S2.3. Subtract IMF from the original uphill time series TQI data v(t) 1 (t), and the uphill residual TQI value is obtained, which is expressed as:

[0062] v res (t) = v(t) - IMF 1 (t)

[0063] In the above formula, v res (t) Uphill residual TQI value;

[0064] S2.4. Using the uphill residual TQI data as the original data, repeat steps S2.1 to S2.4 until the number of extreme value points of the uphill residual TQI is no more than 2, and finally multiple modal components are obtained.

[0065] S3. Use the linear normalization method to normalize the decomposed data of the long and steep slope time series TQI value, and scale the data of each modal component to the range of [0, 1]. The specific form of normalization of the uphill time series TQI decomposed data is as follows:

[0066]

[0067] In the above formula, v norm (t) is the normalized uphill TQI value.

[0068] S4, training set and test set division: the decomposed long and steep slope TQI data is divided into a training set and a test set, the training set is used for model training, and the test set is used for model verification. The ratio of the training set to the test set is 9:1, that is, 3277 data points are used for training and 364 data points are used for testing. The training samples are generated by sliding window method, the window size is at least greater than 7, and the preferred window size is 8. Each sliding step is 1 step to predict the TQI value of the next step.

[0069] S5. Input each decomposed long ramp TQI time series data into the convolutional neural network for convolution processing and extracting feature values. The specific process is as follows:

[0070] The convolutional neural network includes two one-dimensional convolutional layers and two one-dimensional MaxPooling layers for extracting important features of uphill time series TQI. The convolutional layer is responsible for extracting effective features from the uphill TQI decomposition data through convolution operations. The pooling layer is responsible for extracting higher-dimensional features after the convolutional layer to alleviate the limitations of the feature map. The tanh activation function is used to enhance the nonlinear ability, and the feature values ​​of the uphill TQI data are extracted through convolution and pooling.

[0071] S6, inputting the extracted TQI feature value of each decomposed long and short ramp into the long short-term memory neural network to train and predict the TQI value of the long and long ramp, and superimposing the prediction results to output the final result, such as Figure 3 shown.

[0072] In this embodiment, the track quality status index data for a heavy-load train in a long uphill condition is shown in Table 1 below.

[0073] Table 1 Track quality status index data for long uphill conditions

[0074] Collection time point TQI value 1 6.81 2 9.47 3 9.74 4 5.74 5 8.25 6 4.65 7 5.73 8 5.35 …… …… 3636 4.71 3637 4.64 3638 4.24 3639 5.65 3640 5.59 3641 6.75

[0075] The error on the test set in the above example is shown in Table 2.

[0076] Table 2 Prediction index values

[0077] Prediction indicators Index value RMSE 1.132 MAE 0.847 <![CDATA[R 2 ]]> 0.788

[0078] In this embodiment, in step S6, each extracted decomposed uphill TQI feature value is input into the long short-term memory neural network for uphill TQI training and prediction, and the prediction results are superimposed and output as the final result. The specific process includes:

[0079] Construct a long short-term memory neural network (LSTM). The long short-term memory neural network contains multiple recurrent units, specifically 128 recurrent units, each of which consists of a forget gate, an input gate, and an output gate. The sigmoid function is used as the activation function of the gate function. The specific expressions of the forget gate, input gate, and output gate are:

[0080] f t =σ(w f [h t-1 ,v(t)]+b f )

[0081] i t =σ(w i [h t-1 ,v(t)]+b i )

[0082] C t'=tanh(w C [h t-1 ,v(t)]+b C )

[0083] C t =f t ×C t-1 +I t ×C t '

[0084] o t =σ(w o [h t-1 ,v(t)]+b o )

[0085] h t =o t ×tanh(C t )

[0086] In the above formula, f t is the output of the forget gate; i t is the output of the input gate; C t ' is the unit information value; C t is the cell state; o t is the output of the output gate; σ is the sigmoid function; w f 、w i 、w C and w o are the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; h t-1 Output value for the previous unit; b f 、b i 、b C and b o is the corresponding bias term; h t Outputs a value for each cell.

[0087] The Adam optimizer was used with the learning efficiency set to 0.001. Based on the back propagation algorithm, the decomposed TQI time series data was used to train the uphill TQI prediction model. During the training process, a sliding window method was used to generate training samples with a window size of 8. Each sliding step was used to predict the TQI value of the next step. The TQI prediction values ​​of each modal component were superimposed to obtain the final TQI prediction value.

[0088] Results verification: By comparing the predicted TQI values ​​on the test set with the actual observed values, the root mean square error (RMSE), mean absolute error (MAE) and absolute coefficient (R 2 ), verifying the prediction performance of the model. As shown in Table 2, it shows that the model has good prediction accuracy.

[0089] Example 2

[0090] like Figure 1 , Figure 2 as well as Figure 3 As shown, a method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM is provided. Based on Example 1, this embodiment adjusts some parameters and adds some detailed descriptions.

[0091] Steps S1 to S4: Same as in Example 1, data is acquired, CEEMDAN decomposition is performed, normalization is performed, and training and test sets are divided.

[0092] In this embodiment, step S5, hybrid convolutional neural network processing: adjust the hybrid convolutional neural network (CNN) structure, add three one-dimensional convolutional layers, the convolution kernel sizes are 3, 5 and 7, the step size is 1, and the tanh activation function is used. Each convolutional layer is followed by a one-dimensional MaxPooling pooling layer, the pooling window sizes are 2, 3 and 4, and the step size is 2. Add a Dropout layer after the last pooling layer, and the dropout rate is 0.5 to reduce overfitting. The ReLU activation function can also be used, keeping other parameters unchanged, and compared with the previous results.

[0093] In this embodiment, step S6, long short-term memory neural network prediction: keep the long short-term memory neural network (LSTM) structure unchanged, add regularization terms to prevent overfitting. During the training process, early stopping is used, and training is stopped when the loss on the validation set does not decrease for 5 consecutive epochs to avoid overfitting. The sliding window size is adjusted to 10, and each sliding step is 1 step, and the TQI value of the next 3 steps is predicted to explore the model's prediction ability for longer time series.

[0094] Calculate RMSE and MAE and introduce R 2 The score is used as an evaluation indicator to evaluate the model performance. The results show that both RMSE and MAE have decreased, while R 2 This indicates that the model prediction performance has been significantly improved by adjusting the CNN structure and adding regularization measures.

[0095] Example 3

[0096] In this embodiment, before CNN feature extraction, feature selection is performed on the decomposed modal components, and the most representative features are screened out using methods such as correlation analysis or principal component analysis (PCA). The screened features are input into CNN for further feature extraction.

[0097] In the LSTM prediction stage, model fusion technology, such as Bagging or Boosting methods in ensemble learning, is used to fuse the prediction results of multiple LSTM models to improve the accuracy of the prediction. The fused model is trained and tested to evaluate its performance and applied to the actual TQI value prediction.

[0098] Obviously, the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM, characterized in that: The following steps are involved: S1. Input the historical data of TQI values ​​of long and steep slopes, and use the historical observation data of TQI values ​​of long and steep slopes as the original data for decomposition and model training; S2. CEEMDAN decomposition of the time series data of TQI values ​​of long and steep ramps: Decompose the historical time series data of the original long and steep ramp TQI values ​​into multiple modal components IMF s ; S3, normalizing the decomposed data of the long and steep slope time series TQI value; S4, training set and test set division: the decomposed long and steep slope TQI data is divided into a training set and a test set; S5, inputting each decomposed long and steep ramp TQI time series data into a convolutional neural network for convolution processing and extracting feature values; S6. Input each extracted decomposed long and steep ramp TQI feature value into the long short-term memory neural network to train and predict the long and steep ramp TQI value, and simultaneously superimpose the prediction results to output the final result.

2. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1 is characterized in that: In step S2, the CEEMDAN decomposition of the long ramp TQI value time series data includes the following steps: S2.

1. Add a Gaussian white noise value ε with an amplitude of Δ to the original long and steep ramp TQI time series data. i (t), the expression is: v i (t)=v(t)+Δε i (t) In the above formula, v(t) is the original TQI value, v i (t) is the TQI value after the i-th processing; S2.2, v i (t) Perform empirical mode decomposition and average all the first modal components obtained by decomposition to serve as the first modal decomposition of CEEMDAN decomposition. The expression of the average operation is: In the above formula, E1[v1(t)] is the first modal component obtained by empirical mode decomposition; S2.

3. Subtract IMF1(t) from the original uphill time series TQI data v(t) to obtain the uphill residual TQI value, expressed as: v res (t)=v(t)-IMF1(t) In the above formula, v res (t) Uphill residual TQI value; S2.

4. Use the uphill residual TQI data as the original data and repeat steps S2.1 to S2.4 until the number of extreme value points of the uphill residual TQI does not exceed 2.

3. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1 is characterized in that: In step S3, the specific form of normalization of the uphill time series TQI decomposition data is as follows: In the above formula, v norm (t) is the normalized uphill TQI value.

4. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1 is characterized in that: In step S4, the training set is used for model training, and the test set is used for model verification.

5. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 4 is characterized in that: The ratio of the training set to the test set is 9:

1.

6. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 5 is characterized in that: In model training, a sliding window method is used to generate training samples. The window size is 8, and each sliding step is 1 step to predict the future.

7. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1 is characterized in that: In step S5, the convolutional neural network includes multiple one-dimensional convolutional layers and multiple one-dimensional MaxPooling pooling layers.

8. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1 is characterized in that: In step S6, the long short-term memory neural network includes a plurality of recurrent units, each of which is composed of a forget gate, an input gate and an output gate.

9. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 7, characterized in that: The specific expressions of the forget gate, input gate, and output gate are: f t =σ(w f [h t-1 ,v(t)]+b f ) i t =σ(w i [h t-1 ,v(t)]+b i ) C t '=tanh(w C [h t-1 ,v(t)]+b C ) C t =f t ×C t-1 +I t ×C t ' o t =σ(w o [h t-1 ,v(t)]+b o ) h t =o t ×tanh(C t ) In the above formula, f t is the output of the forget gate; i t is the output of the input gate; C t ' is the unit information value; C t is the cell state; o t is the output of the output gate; σ is the sigmoid function; w f 、w i 、w C and w o are the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; h t-1 Output value for the previous unit; b f 、b i 、b C and b o is the corresponding bias term; h t Outputs a value for each cell.

10. The method for predicting track quality index of heavy-load train long uphill conditions based on CEEMDAN-CNN-LSTM according to claim 1, characterized in that: In step S6, the uphill TQI prediction model is trained using the decomposed TQI time series data based on back propagation using the Adam optimizer.

Images