A high-speed motor train unit coasting trend prediction method

Abstract

The application discloses a high-speed motor train unit idling and sliding trend prediction method, which comprises the following steps: step one, screening and standardizing the model input data; step two, obtaining a feature matrix containing position information through value encoding and position encoding; step three, inputting the feature matrix obtained in step two into a ProbSparse self-attention network to obtain the attention relationship of different feature vectors at different positions, and performing parallel calculation through a multi-head mode; step four, using a Mask self-attention network for the decoder to associate the sliding parameter information at different moments with the sliding state, and finally performing classification prediction on the future moment through a full connection layer and a Sigmoid activation function. The application can predict the idling and sliding trend of the train more than 1 second in advance, the prediction accuracy meets the train operation demand, the prediction speed is fast, and the idling and sliding trend states of multiple axles of the train can be simultaneously and on-line predicted in parallel.

Description

Technical Field

[0001] This invention relates to traction control technology for high-speed trains, and more particularly to a method for predicting slip / coasting. Background Technology

[0002] In the railway field, 'adhesion' refers to the phenomenon where the wheel-rail contact point remains relatively stationary without relative slippage during train operation. When a train is running, under the influence of axle load, the wheel-rail contact surface undergoes relative deformation, forming an elliptical contact area. This contact area is divided into a slip zone and an adhesion zone. The relative velocity is zero in the adhesion zone, while micro-slippage occurs in the slip zone. When the adhesion conditions between the wheel and rail deteriorate, and the train's traction / braking force exceeds the maximum usable adhesion force between the wheel and rail, wheel spin / slippage will occur. This reduces the effectiveness of traction / braking force, increases wear and even damage between the wheel and rail, and seriously threatens the safe operation of the train. Therefore, identifying wheel spin / slippage is of great significance.

[0003] Existing methods for identifying idle / coasting mainly include:

[0004] (1) Combined correction and identification method

[0005] This method first calculates creep speed, acceleration, and differential acceleration benchmarks, then compares them with pre-set thresholds to determine whether train slippage has occurred based on numerical values ​​and conditions. However, this method is susceptible to wheel-rail noise interference, and the parameters are manually set, introducing human error and affecting the accuracy of the detection.

[0006] (2) Idle Identification Method Based on Unscented Kalman Filter (UFK)

[0007] The adhesive state is estimated by estimating the error between the relative adhesive force and the desired relative adhesive force. UFK is suitable for time-domain filtering of the state of recursively Gaussian linear systems and performs well in state estimation of dynamic linear systems following a Gaussian distribution. However, UFK performs poorly in complex and variable environments, exhibiting weak robustness.

[0008] (3) Idle recognition method based on binary classification SVM

[0009] This method detects slippage by adjusting training data and hyperplane. It has a fast detection rate, but it does not consider feature extraction and cannot accurately measure the intrinsic features of the data, resulting in low detection accuracy.

[0010] (4) Train slip recognition method based on empirical wavelet transform (EWT), fuzzy entropy (FE) and SVM

[0011] This algorithm uses EWT to decompose velocity data, extracts slip-related features using the fuzzy entropy algorithm, and then uses particle swarm optimization to train and optimize the SVM. This method has high accuracy, but the computational process is complex, the detection rate is low, and it cannot perform real-time online prediction of train data.

[0012] (5) Slide recognition method based on deep denoising autoencoder

[0013] This method combines an autoencoder with a classification neural network, using the autoencoder to extract features, which are then fed into the classification network for prediction. However, this method performs poorly with small sample sizes and may fail to make predictions.

[0014] (6) Sliding recognition method based on recurrent neural network (RNN)

[0015] This method feeds time series data into an RNN. The current input is based on the output of previous times, which can extract the correlation features in the time series. However, the output speed is slow, and it is necessary to calculate the weighted sum of the feature values ​​of previous times. It can only decode and predict step by step and cannot predict the results of long sequences.

[0016] In deep learning, classic networks such as fully connected neural networks and convolutional neural networks possess powerful feature extraction capabilities. However, each layer leads to feature loss, and the receptive field grows larger and larger, causing the final layer to fail to represent all the features of the original data. This results in training performance deteriorating with the increase in the number of network layers. The introduction of ResNet in 2016 effectively solved the feature loss problem by concatenating and fusing features. However, as the number of network layers increases, the computational cost also increases, along with the number of training parameters and memory usage. Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTMs) are widely used in Natural Language Processing (NLP) due to their powerful ability to process temporal text data. However, RNNs and LSTMs lack parallel output capabilities, and decoding relies on information from previous time steps; as the amount of text information increases, the output efficiency slows down.

[0017] The introduction of the Transformer in 2017 significantly challenged the dominance of previous feature extraction networks. The Transformer employs an encoder-decoder structure, allowing for parallel output of predictions in the decoder. It introduces a multi-head attention mechanism, ensuring contextual relevance in textual data and providing correlation scores between the current time step and all other time steps in time-series data, enabling parallel training of multiple feature matrices. Furthermore, the Transformer structure effectively extracts features from the original data without requiring deep network stacks. Compared to CNNs, the Transformer avoids deep network stacks, and each layer outputs feature maps of the same size and number as the original, ensuring each layer captures feature information from the original maps. In 2021, the Informer further improved upon the attention mechanism of the Transformer, filtering out unnecessary attention information and significantly optimizing network speed and memory usage. The Informer also introduces timestamp encoding, enhancing the temporal order of the predicted time-series data. The Informer's decoder uses a one-step decoder; by changing the decoder's input, all predictions can be obtained through a single fully connected layer at the output. Summary of the Invention

[0018] To overcome the aforementioned shortcomings of existing technologies, this invention proposes a method for predicting the idling and taxiing trends of high-speed trains. Combining the advantages of Probattention and one-step Decoder in the Transformer, the output structure of the Decoder in the Transformer is improved. Without increasing the classification branches, the prediction model is transformed into a prediction classification model. The model performs position encoding and feature extraction on the train operation data of the previous few seconds, and then feeds it into the Transformer. The Decoder in the network directly outputs the prediction results of the idling and taxiing trends for the next few seconds.

[0019] The technical solution adopted by this invention to solve its technical problem is: a method for predicting the idling and coasting trend of high-speed trains, comprising the following steps:

[0020] Step 1: Filter and standardize the input data for the model;

[0021] Step 2: Obtain the feature matrix containing location information through value encoding and position encoding;

[0022] Step 3: Input the feature matrix obtained in Step 2 into the ProbSparse self-attention network to obtain the attention relationship between different positions of different feature vectors, and perform parallel computation in a multi-head manner;

[0023] Step 4: The decoder uses a Mask self-attention network to associate the sliding parameter information at different times with the sliding state, and finally uses a fully connected layer and a Sigmoid activation function to directly classify and predict future times.

[0024] Compared with the prior art, the positive effects of the present invention are:

[0025] 1) It can predict the train's idling / gliding trend more than 1 second in advance, and the prediction accuracy meets the train operation requirements.

[0026] 2) It has a fast prediction speed and can simultaneously perform online prediction of the idling / slipping status of multiple axles of the train.

[0027] 3) It can accurately predict the idling / coasting state by integrating multiple parameters such as vehicle speed, axle speed, traction torque of each axle, and traction motor voltage / current. Attached Figure Description

[0028] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0029] Figure 1 A flowchart of a method for predicting the idling and coasting trend of high-speed trains;

[0030] Figure 2 This is a flowchart of a multi-head self-attention mechanism. Detailed Implementation

[0031] The creep classification prediction method based on Informer is a method that obtains a multi-axis binary classification prediction model by training train operation data to predict the stickiness state for several future time steps. This invention combines the advantages of Transformer and Informer, improving the network's output structure. For example... Figure 1 As shown, this invention employs the same Encoder-Decoder structure as the Transformer. The input is encoded by value and position to obtain a feature matrix containing positional information. The feature matrix is ​​processed through a ProbSparse self-attention mechanism to obtain the attention relationships between different positions of different feature vectors, and is computed in parallel using a multi-head approach. After the self-attention mechanism, distillation pooling is performed to reduce the number of network parameters, followed by another multi-head ProbSparse self-attention mechanism. The input to the Decoder consists of the prediction results from some time steps prior to the time step to be predicted and the length of the prediction target being L. predThe padding values ​​are concatenated. To prevent autoregression, a joint attention mechanism is used with the temporal feature matrix of the Encoder after Mask Self-Attention, associating the sliding parameter information at different time steps with the sliding state. Finally, a fully connected layer and a sigmoid activation function are used to directly classify and predict future time steps. Specific content includes:

[0032] 1. Data Processing

[0033] Considering that slippage occurs at a very small percentage of the total data, we screened the dataset to prevent overfitting to normal states during uneven training. For model input selection, we considered multiple parameters, including speed, wheel speed, and axle traction. Due to the uneven magnitude of these parameters, we standardized them to scale them proportionally into a small, specific range, removing unit limitations and converting them into dimensionless pure numerical values. This facilitates the simultaneous input of parameters with different units or magnitudes into the network model for training.

[0034] 2. Encoding

[0035] (1) Value Encoding

[0036] First, input X en Features are extracted using a one-dimensional convolution to obtain the feature matrix V. emb ∈R Leq×d_mod , where d_mod is the input dimension of the model.

[0037] V emb =Conv1d([X en ])

[0038] (2) Location coding

[0039] To avoid losing the time series X en The temporal information of each row vector needs to be analyzed for X. en Each row of vectors is positionally encoded. There are many ways to perform positional encoding; this invention uses relative positional encoding based on trigonometric functions.

[0040]

[0041] Finally, the feature matrix V emb The input X of the Attention network is obtained by adding the positional encoding PE. attention ∈R Leq ×d_mod X attention By performing only one convolution, almost all the features of the original data are preserved, and the temporal information between the feature vectors in the original data is also included.

[0042] Xattention =V emb +PE

[0043] 3. Multi-head self-attention mechanism

[0044] like Figure 2 As shown, the self-attention mechanism first focuses X attention The computation is performed in parallel by dividing the computation into multiple heads, with each head obtaining the result through three fully connected layers FC(x). Three matrices.

[0045]

[0046] L head =d_mod / n

[0047] FC(X)=XW+b

[0048] Where W is the weight matrix, b is the bias, and L is the weight matrix. head For each head dimension, σ(·) is the ReLU activation function. The dot product of Qurries and Keys follows a long-tailed distribution, with only a small portion generating primary attention and the rest generating secondary attention, resulting in wasted computational resources. To reduce computation, the top_n Keys are randomly selected to obtain RandomKeys, and then Max-Mean Measurement is performed with Qurries to obtain the distributional difference between Keys and Qurries, as shown in the following formula, where Qi∈Qurries, L k Let L be the length of the keys in the self-attention mechanism. k With L seq same.

[0049]

[0050] We expect the attention distribution to be highly varied, with each Qurries paying attention to relevant keys, rather than following a uniform distribution. We select the top_n Qurries with the largest differences from the RandomKeys distribution, and then perform attention calculations with the keys and values. Finally, we average the remaining sparse attention components in the values ​​to obtain the final Attention_V.

[0051]

[0052] Attention_V=replace[Mean(V),Attention(Q,K,V)]

[0053] 4. Distillation

[0054] If a multi-layer self-attention mechanism is required, distillation pooling can be used between each layer to reduce the number of parameters that need to be computed and speed up the forward propagation of the network.

[0055] X l+1 =MaxPool(X l )

[0056] Where X l Let `Attention_V` be the `l`-th layer's `Attention_V` after convolution and dimension transformation operations. `MaxPool(x)` is a pooling downsampling operation with a stride of 2. Simultaneously, feature fusion is introduced before distillation, adding the input from the previous layer to the output of the next layer, thus ensuring that important features are not lost as the network deepens.

[0057] 5. Decoder

[0058] This invention makes some improvements to the input of the Decoder for classification and prediction tasks. The input of the Decoder is the prediction results of some time steps before the time step to be predicted and the length of the time step to be predicted is L. pred The padding values ​​are concatenated. This allows the known sliding states to guide the sequence to be predicted. The original Informer paper uses all-zero padding. In sliding state recognition, the classification prediction value is 0 or 1. Both all-zero padding and all-1 padding will interfere with our prediction results. We use the Sigmoid function on the model output and calculate the index by setting an adjustable threshold 'thus'. Here, we use the same value as the threshold for padding, such as 0.5, which effectively eliminates the interference caused by all-zero padding. In the Decoder's self-attention layer, we do not want future results to affect previous results. To prevent autoregression, we use a Masked self-attention mechanism. In the Decoder's attention layer, the attention relationship between the self-attention layer and the Encoder output is calculated, linking the parameter features affecting the sliding state to the sliding state. When analyzing the attention relationship between the two, the distribution difference is large, and the sliding state parameters are few. Therefore, we do not use ProbSparse Attention, but Fully Attention.

[0059] 6. Prediction

[0060] After the Encoder-Decoder is completed, a fully connected layer maps the output to a Y-shape of the same size as the sliding state. out Since we directly select the sliding state as the Decoder input, we do not need to add a new classification branch on the output. Instead, we use the Sigmoid activation function to map the output between 0 and 1 and set a threshold to perform classification prediction.

[0061]

[0062] Y pred =Sigmoid(Y out )

[0063] 7. Adam Backpropagation Training

[0064] The cross-entropy loss function is widely used in classification tasks, and this invention uses this function as a loss estimate for the model:

[0065]

[0066] Gradient descent (SGD) is widely used in various neural network learning tasks. However, since the gradient term of this method does not include any momentum coefficient, it is prone to getting trapped in local optima, as shown in the following equation:

[0067] θ t+1 =θ t -η·g t

[0068] Where θ t Let g be the weight parameter matrix at time t, η be the learning rate, and g be the weight parameter matrix at time t. t This represents the gradient of the loss function with respect to the weights. This invention compares Adam (Adaptive Momentum Estimation) with other mainstream optimization methods, and it is evident that Adam outperforms other optimizers in both convergence speed and final convergence performance. Adam adds a first-order momentum m to SGD. t and second momentum v t First, calculate the first-order momentum m. t As shown in the following formula:

[0069] m t =β1m t-1 +(1-β1)·g t

[0070] Calculate the second momentum v t The method is the same as RMSProp and Adagrad:

[0071]

[0072] Where β1 and β2 are the estimated exponential decay rates in the two tests. To prevent the gradient weights from being too small at the start of training, a correction factor β, representing the exponentially weighted average, is added to each of the two momentum values. t1 and β t2 This ensures that the sum of the weights at each time step during training is 1. The formula derivation is as follows (v0 is initialized to 0):

[0073] From the first-order momentum formula, we get:

[0074] m1=β1m0+(1-β1)g0

[0075] =(1-β1)g0

[0076] From the above formula, we can deduce:

[0077] m2=β1m1+(1-β1)g1

[0078] =β1(1-β1)g0+(1-β1)g1

[0079] From the above formula, we can derive m. t The relationship between gradient and gradient:

[0080]

[0081] The weighting correction factor is derived as follows:

[0082]

[0083] β t2 The calculation process and β t1 The same applies, so it will not be repeated. Consider the two momentum m... t and v t Dividing by the corresponding correction factor, we obtain the corrected momentum:

[0084]

[0085]

[0086] Finally, the parameters of the Adam optimization algorithm are updated as follows:

[0087]

Claims

1. A method for predicting the idling and coasting trend of a high-speed train, characterized in that: Includes the following steps: Step 1: Filter and standardize the input data for the model; Step 2: Obtain the feature matrix containing location information through value encoding and position encoding: (1) For input X en Features are extracted using a one-dimensional convolution to obtain the feature matrix. Where d_mod is the input dimension of the model; (2) For X en Each row vector is encoded using relative position encoding based on trigonometric functions: (3) The characteristic matrix V emb The input to the Attention network is obtained by adding the positional encoding (PE). Step 3: Input the feature matrix obtained in Step 2 into the ProbSparse self-attention network to obtain the attention relationships at different positions of different feature vectors, and perform parallel computation using a multi-head approach. The method for parallel computation using a multi-head approach is as follows: Step 1, X attention The computation is performed in parallel by dividing the computation into multiple heads, with each head obtaining the result through three fully connected layers FC(x). Three matrices: L head =d_mod / n FC(X)=XW+b Where W is the weight matrix, b is the bias, and L is the weight matrix. head For each head dimension, σ(·) is the ReLU activation function; The second step is to randomly select the top_n keys to obtain RandomKeys, and then perform Max-Mean Measurement with the Qurries to obtain the distribution difference between the keys and the Qurries: Where Qi ∈ Qurries, L k L is the length of the keys. k =L seq ; The third step is to select the top_n Queries that differ the most from the RandomKeys distribution, then perform attention calculations with the Keys and Values, and finally average the remaining sparse attention components in the Values ​​to obtain the final Attention_V: Attention_V=replace[Mean(V),Attention(Q,K,V)]; Step 4: The decoder uses a Mask self-attention network to associate the sliding parameter information at different times with the sliding state, and finally uses a fully connected layer and a Sigmoid activation function to directly classify and predict future times.

2. The method for predicting the idling and coasting trend of a high-speed train according to claim 1, characterized in that: When using a multi-layer ProbSparse self-attention network, a distillation pooling operation is performed between each layer.

3. The method for predicting the idling and coasting trend of a high-speed train according to claim 2, characterized in that: Perform distillation in a pooling process using the following formula: X l+1 =MaxPool(X l ) Where X l Let Attention_V be the result of convolution and dimension transformation operations in the l-th layer, and MaxPool(x) be a pooling downsampling operation with a stride of 2.

4. The method for predicting the idling and coasting trend of a high-speed train according to claim 2, characterized in that: Feature fusion is introduced before performing distillation pooling, which adds the input of the previous layer to the output of the next layer.

5. The method for predicting the idling and coasting trend of a high-speed train according to claim 1, characterized in that: The decoder takes as input the prediction results from some time steps before the time step to be predicted and the length of the time step to be predicted as L. pred The concatenation of fill values.

6. The method for predicting the idling and coasting trend of a high-speed train according to claim 5, characterized in that: The fill value is the same as the adjustable threshold value.

7. The method for predicting the idling and coasting trend of a high-speed train according to claim 1, characterized in that: The method of directly classifying and predicting future time steps using a fully connected layer and a sigmoid activation function is as follows: the output is mapped to a Y-shape of the same size as the sliding state through a fully connected layer. out The sliding state is directly selected as the input to the decoder. The Sigmoid activation function is used to map the output to between 0 and 1, and a threshold is set for classification prediction. Y pred =Sigmoid(Y out )。 8. The method for predicting the idling and coasting trend of a high-speed train according to claim 1, characterized in that: The filtered model input data includes the rotational speed of each axle, vehicle speed, traction torque of each axle, and traction motor voltage / current.

Images