Track surface identification and idling state cooperative identification method based on ensemble learning

Abstract

The application discloses a track surface identification and idling state cooperative identification method based on integrated learning, first, train operation state data is acquired, and a multidimensional physical feature tensor containing a motion state and a transient mutation is extracted through a sliding window; then, the multidimensional physical feature tensor is input into a heterogeneous base learner containing a convolutional neural network, a gated recurrent unit and a long short-term memory network in parallel, under the common supervision of a label mask mechanism and a focal loss function, and a primary classification probability is output; finally, the outputs of the three heterogeneous base learners are spliced into multidimensional element features, which are input into a random forest element learner to be fused and corrected through a voting mechanism, and a final track surface identification and idling recognition cooperative result is output. The application innovatively introduces an event-triggered binary label mask mechanism and a focal loss function, effectively overcomes the discrimination bias of class imbalance and a feature fuzzy area, and realizes accurate cooperative identification of track surface states and idling evolution trends under complex working conditions.

Description

Technical Field

[0001] This invention relates to the fields of rail transit train operation status monitoring and artificial intelligence technology, and in particular to a method for collaborative identification of track surface status and idling risk based on a multi-model integrated learning architecture. Background Technology

[0002] Wheel-rail adhesion is the sole driving force during train traction. The wheel-rail adhesion characteristic curves under different rail surfaces are as follows: Figure 1 As shown, the adhesion characteristic curves differ on different rail surfaces, but all have a maximum adhesion point, with the adhesion zone on the left and the slip zone on the right. When a train encounters severe weather such as rain, snow, or frost that deteriorates the rail surface condition, the available adhesion force can easily drop precipitously. If control measures are not taken in time, it can easily induce large-scale train slippage, seriously threatening operational safety and reducing traction efficiency. Therefore, under complex operating conditions, real-time and accurate identification of the current rail surface contact state and early identification of the wheel slippage evolution trend are key prerequisites for optimizing train adhesion control strategies and ensuring operational safety.

[0003] Existing methods for train track surface identification and idle detection mainly face the following technical bottlenecks: (1) The commonly used idling detection methods are shown in Figure 2. They are mainly based on setting fixed thresholds for single or combined physical quantities such as creep speed and wheel circumference acceleration. These methods rely heavily on engineering experience, have a slow response, and are extremely sensitive to changes in operating conditions. Under different operating speed ranges or variable rail surface conditions, fixed thresholds are easily affected by noise interference, resulting in false alarms (frequent torque reduction affects traction) or missed alarms (idling deterioration cannot be suppressed), making it difficult to balance robustness and sensitivity.

[0004] (2) There are currently few methods for real-time identification of the wheel-rail contact surface of trains. On the one hand, traditional visual or optical sensors are easily affected by the operating environment in open and complex track environments; at the same time, the installation space at the bottom of the train bogie is extremely small and accompanied by high-frequency and severe vibrations, making it difficult for optical lenses to keep clean and focus stably, and thus unable to continuously provide high signal-to-noise ratio rail surface image data under harsh conditions. On the other hand, the wheel-rail contact patch is extremely small and is in a dynamic compression state of high-speed rolling and sliding, making it difficult for existing physical sensors to directly penetrate this confined area for real-time, non-destructive interface property measurement without damaging the wheel or rail structure.

[0005] (3) In recent years, data-driven recognition models have gradually emerged, but single neural network structures (such as simple CNN or RNN) often have blind spots in feature extraction. The dynamic response of the train wheel-rail system includes high-frequency transient changes (such as the moment of idling) and long-term temporal evolution (such as the slow accumulation of creep state) on the time scale. It is difficult for a single model to capture these complex multi-scale features comprehensively and stably.

[0006] (4) In actual train operation data, idling and warnings are extremely rare abnormal events, and the train is in a normal sticking state most of the time. Traditional model training methods perform global loss calculation on the entire time domain sequence, which causes the gradient update signal to be "submerged" or dominated by a large number of featureless normal samples. The model cannot fully focus on learning the key transition features before and after idling, resulting in a significant drop in prediction performance.

[0007] In the field of train track surface identification and slip detection, traditional physical thresholding methods heavily rely on engineering experience and are highly susceptible to noise interference, leading to false alarms or missed alarms under complex and variable operating conditions. Traditional visual or physical sensors, on the other hand, are limited by confined spaces and harsh environments, failing to provide consistently stable data with a high signal-to-noise ratio. With the rise of data-driven approaches, deep learning has been introduced into this field, but single neural network structures (such as standard CNNs or RNNs) often have blind spots in feature extraction. The dynamic response of train wheel-rail systems encompasses both high-frequency transient changes at the moment slip occurs and long-term temporal evolution of the creep state, making it difficult for a single model to comprehensively and stably capture these complex multi-scale features. A more significant problem is that slip and warnings are extremely rare abnormal events in actual operation. Traditional global loss calculation methods result in the model gradient update signal being "overwhelmed" by massive amounts of featureless normal samples, preventing the model from fully focusing on learning the key transitional features before and after slip, leading to a significant drop in prediction performance. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention proposes a collaborative recognition method for track surface identification and idling status based on ensemble learning. It combines the advantages of Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Long Short-Term Memory Networks (LSTM) in multi-scale feature extraction, thus compensating for the limitations of single-model feature capture. Addressing the extreme data imbalance caused by scarce events, this invention improves upon conventional global training methods by innovatively introducing an event-triggered binary label masking mechanism and a Focal Loss function. The model extracts multi-dimensional physical features of train operation through a sliding window and feeds them into a heterogeneous base network. Under the filtering of the masking mechanism, the model is forced to focus on abnormal transition features. Finally, the initial predicted probabilities are fed into a Random Forest (RF) meta-learner for fusion and error correction, effectively overcoming class imbalance and discriminative bias in feature ambiguity areas, and achieving accurate collaborative recognition of track surface status and idling evolution trends under complex operating conditions.

[0009] The technical solution adopted by this invention to solve its technical problem is: a method for collaborative identification of track surface recognition and idling state based on ensemble learning, comprising the following steps: Step 1: Obtain train operation status data and extract multidimensional physical feature tensors containing motion state and transient changes through a sliding window; Step 2: Input the multidimensional physical feature tensor in parallel into a heterogeneous base learner containing a convolutional neural network, a gated recurrent unit, and a long short-term memory network. Under the joint supervision of the label masking mechanism and the focus loss function, the primary classification probability is output. Step 3: Concatenate the outputs of the three heterogeneous base learners into multidimensional meta-features, input them into the random forest meta-learner, and perform fusion and error correction through a voting mechanism to output the final collaborative result of orbital plane recognition and idle spin recognition.

[0010] Compared with the prior art, the positive effects of the present invention are: This invention employs an ensemble learning network based on a stacking architecture, combining the advantages of Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Long Short-Term Memory Networks (LSTM) in complex multi-scale feature extraction. In terms of model structure and training optimization, by introducing a label masking mechanism and Focal Loss, the model is forced to focus on anomalous transition features, effectively overcoming the problem of extreme data class imbalance caused by scarce idling events. The model first processes the multi-dimensional physical state data of the train during operation, constructing a three-dimensional temporal feature tensor and feeding it into a parallel base learner for inference. Subsequently, the primary classification probabilities output by the network are concatenated into meta-features, which are directly input into a Random Forest (RF) meta-learner for fusion and error correction through a voting mechanism. Finally, the model collaboratively outputs highly robust track surface identification and idling state classification results. Specific advantages include: 1) The introduction of a masking mechanism to filter out interference from a large number of normal samples, combined with the Focal Loss function to force the model to focus on difficult-to-distinguish samples, effectively solves the pain point of extremely unbalanced idle data unique to the rail transit field.

[0011] 2) Heterogeneous integration of Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU) and Long Short-Term Memory Network (LSTM) makes up for the blind spot of single neural network in multi-scale temporal dynamic feature extraction.

[0012] 3) Without increasing extremely high computational overhead, efficient and parallel collaborative identification of environmental track surface status and vehicle idling trend is achieved through joint decoding and error correction of shared backbone network and random forest (RF) meta-learner. Attached Figure Description

[0013] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 Wheel-rail adhesion characteristic curves under different rail surfaces; Figure 2 This is a schematic diagram of an existing idling detection method; Figure 3 A schematic diagram of feature extraction using a sliding window; Figure 4 This is a schematic diagram of the idle marking strategy; Figure 5 for Mask Schematic diagram of tag extraction; Figure 6 This is a diagram of the ensemble learning architecture. Detailed Implementation

[0014] This invention proposes a collaborative method for train track surface identification and idling detection based on a stacking architecture and ensemble learning. First, train operation status data is acquired, and a multi-dimensional physical feature tensor containing motion state and transient changes is extracted using a sliding window. Second, considering the rarity of idling events, a forward-looking warning label is defined, and an event-triggered binary label masking mechanism is constructed to filter interference from a large number of normal samples. Then, the feature tensor is input in parallel into a heterogeneous base learner comprising a convolutional neural network (CNN), a gated recurrent unit (GRU), and a long short-term memory network (LSTM). Under the joint supervision of the label masking mechanism and the focal loss function, a primary classification probability is output. Finally, the outputs of the base learners are concatenated into meta-features, which are then input into a random forest (RF) meta-learner for fusion and error correction, outputting the final collaborative result for track surface identification and idling detection. This invention effectively overcomes the problem of extremely imbalanced data categories, improves the accuracy and robustness of state recognition under complex operating conditions, and achieves collaborative perception of track surface operation status and idling status.

[0015] The method of the present invention will be described in detail below with reference to the accompanying drawings: 1. Time series data preprocessing and physical feature tensor extraction Raw sensor data of train operation (including vehicle speed, wheel circumference speed, and wheelset acceleration at different axles) is acquired. To deeply mine the dynamic characteristics hidden in the time-series signals, a physical feature vector containing motion states and transient changes is constructed. Based on creep as the physical basis for wheel-rail adhesion transmission, creep velocity is extracted. and creep rate To capture the evolution trend of the creep state, a first-order difference feature of the creep velocity is introduced: Meanwhile, when wheel-rail interface spins, the violent fluctuations in the net external force will manifest in higher-order changes of acceleration before changes in velocity. This leads to the introduction of wheelset acceleration. And calculate the first and second derivatives of the acceleration: Based on the above characteristics, a multi-dimensional feature vector is constructed at a single time step: like Figure 3 (Diagram of sliding window feature extraction) As shown, a sliding window mechanism is used to convert single-time-point data into continuous segments. The sliding window parameters are set as follows: length L=50, sliding window step size S=10, forming an 80% overlap region to enhance the ability to capture details of state transitions. The final result is a three-dimensional tensor with shape (M, 50, 6). Due to the significant differences in different physical dimensions, a standardization method is used to process the input feature vector. in, The mean, The standard deviation is denoted as .

[0016] 2. Event-driven dual-task label definition and mask generation To achieve dual-task collaboration, two types of tag systems are first defined: (1) Rail surface condition label: The rail surface contact conditions are divided into three categories: dry, wet and ice / snow; (2) Idle status label: Regarding the definition of the idle label, such as Figure 4(Diagram of idling marking strategy) As shown, a forward-looking warning marking strategy is adopted: areas with a creep speed greater than 3 km / h and an acceleration greater than 2 m / s² are defined as idling; areas with a creep speed greater than 1.5 km / h or an acceleration greater than 1 m / s² are defined as warning areas; and all other cases are defined as normal areas.

[0017] To address the issue of class imbalance caused by idling being a rare event in real-world data, such as... Figure 5 ( Mask As shown in the diagram of tag extraction, a tag masking mechanism is introduced. A binary mask is constructed. Mask k When an idling event is detected, the normal window for the idling segment, the warning segments before and after it, and the beginning and end of the event (70 sampling points each) is set to [normal]. Mask k = 1, but those skilled in the art can make adaptive adjustments within a reasonable range (e.g., 50-100 sampling points) based on the actual train operating speed and sampling frequency; the remaining long-term periods far from the normal event period are set to Mask k = 0. This mechanism forces the model to focus on anomalous transition features when calculating the loss.

[0018] 3. Parallel Inference and Cross-Validation of Base Learners Based on Heterogeneous Feature Extraction like Figure 6 (Ensemble learning architecture diagram) The upper part shows the construction of the first layer of ensemble learning network, which includes three heterogeneous base learners: convolutional neural network (CNN), gated recurrent unit (GRU), and long short-term memory network (LSTM).

[0019] Each base learner adopts a network topology structure with a shared backbone and dual-task branches. Temporal features are extracted through the shared backbone network and divided into a track surface identification branch and an idle state identification branch. The idle state identification branch concatenates the shared layer features with the intermediate features of the track surface branch and then classifies them, outputting the class probabilities of track surface state and idle state in parallel. Each base learner model has a Dropout layer after its corresponding layer, with a value between 0.28 and 0.35, to suppress overfitting. During the model training phase, a K-fold cross-validation strategy (K=5) is used to train the base learners to obtain the bias-free out-of-fold prediction probability distribution to avoid data leakage and overfitting. Simultaneously, the track surface identification task uses standard cross-entropy loss, while the idle state identification task uses the label mask generated in step S2, only targeting specific tasks. Mask k For samples with a value of 1, the loss is calculated and the gradient is backpropagated. The FocalLoss function is introduced to address the class imbalance problem.

[0020] The core idea of ​​focus loss is to introduce a modulation factor on top of standard cross-entropy, automatically reducing the weight of easily classified samples, so that model training focuses more on difficult-to-classify samples. It is defined as: The formula combining the mask cross-entropy and focus loss is as follows: In the formula, N This represents the total amount of data from all sampling points in the current batch. K This indicates the total number of classification categories for the idle recognition task. For numerically stable terms, This represents the model's predicted probability of the true class. To focus parameters, Weights are balanced for each category. The focus parameter is set to... The Effective Number method is used to automatically calculate the class balance weights.

[0021] During training, the model outputs the predicted probabilities for three track surface states: dry, wet, and icy / snowy. Its satisfaction The masked cross-entropy loss (Masked CE) during training is as follows: in, This represents the mask coefficient of the k-th sample, with a value of 0 or 1. A value of 1 indicates that the sample is valid and will be included in the calculation of the loss function. Indicates the first k The actual track surface state label corresponding to each sample For numerical stability, Single-point cross-entropy: Indicator function Defined as: The network parameters are updated using the Adam optimization algorithm. Finally, the losses from the two tasks are summed in equal weights, and a masking mechanism is used to calculate the idle loss and backpropagate the gradient only for samples within the event window. The total loss function is: In the formula, This represents the loss function for the track surface identification task. The superscript "rail" indicates the track surface identification and classification task, and the subscript "CE" represents Cross-Entropy. This represents the loss function for the idle recognition task. The superscript slip indicates the idle recognition task, and the subscript FL indicates that it uses Focal Loss. yes The basic unit, The single-point cross-entropy of all valid samples The overall batch loss is obtained by performing a mask-weighted average.

[0022] 4. Meta-learning fusion decision based on random forest (RF).

[0023] like Figure 6 As shown in the lower part, the second layer based on the stacking architecture uses a random forest (RF) as the meta-learner. The predicted probabilities (track surface identification probability and idle speed identification probability) output in parallel by the three heterogeneous base learners (CNN, GRU, and LSTM) in step S3 are extracted and concatenated to form a multi-dimensional meta-feature input. A joint label encoding mechanism combines the track surface state task and the idle speed identification task into a joint category for unified classification. The classification results are then decoded and separated to output the final independent track surface state category and idle speed state category. This fusion error correction mechanism, combined with out-of-bag (OOB) sample estimation, effectively corrects the discrimination bias of a single model in feature ambiguity areas (such as the transition boundary between the warning state and the normal state).

Claims

1. A method for collaborative identification of track surface recognition and idling state based on ensemble learning, characterized in that: Includes the following steps: Step 1: Obtain train operation status data and extract multidimensional physical feature tensors containing motion state and transient changes through a sliding window; Step 2: Input the multidimensional physical feature tensor in parallel into a heterogeneous base learner containing a convolutional neural network, a gated recurrent unit, and a long short-term memory network. Under the joint supervision of the label masking mechanism and the focus loss function, the primary classification probability is output. Step 3: Concatenate the outputs of the three heterogeneous base learners into multidimensional meta-features, input them into the random forest meta-learner, and perform fusion and error correction through a voting mechanism to output the final collaborative result of orbital plane recognition and idle spin recognition.

2. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 1, characterized in that: The multidimensional physical feature tensor mentioned in step one includes creep velocity and its first-order difference, creep rate, wheelset acceleration and its first and second derivatives.

3. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 1, characterized in that: The heterogeneous base learners described in step two all adopt a network topology structure with a shared backbone and dual task branches. After the temporal features are extracted by the shared backbone network, they are divided into a track surface identification task branch and an idle state identification task branch. The idle state identification task branch concatenates the shared layer features with the intermediate features of the track surface identification task branch and then classifies them, and outputs the class probabilities of track surface state and idle state in parallel.

4. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 3, characterized in that: The loss function used for the idle recognition task is: ， In the formula, N This represents the total amount of data from all sampling points in the current batch. K This indicates the total number of classification categories for the idle recognition task. For numerically stable terms, This represents the model's predicted probability of the true class. To focus parameters, To balance the weights for each category, Mask k For binary masking: When an idle event is detected, the normal window is set to backtrack 50-100 sampling points for the idle segment, the warning segments before and after it, and the beginning and end of the event. Mask k = 1, and the remaining long-term periods far removed from the normal event time are set to 1. Mask k = 0.

5. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 4, characterized in that: Focusing parameters The Effective Number method is used to automatically calculate the class balance weights.

6. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 4, characterized in that: The area with a creep speed greater than 3 km / h and an acceleration greater than 2 m / s² is defined as the idle zone; the area with a creep speed greater than 1.5 km / h or an acceleration greater than 1 m / s² is defined as the warning zone; all other situations are defined as normal periods.

7. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 4, characterized in that: The loss function used for track surface identification is: ， The model output predicts the probabilities for three track surface conditions: dry, wet, and icy / snowy. Its satisfaction ; Indicates the first k Mask coefficients for each sample; Indicates the first k The actual track surface status label corresponding to each sample; It is a numerically stable term; The single-point cross-entropy is calculated using the following formula: Indicator function Defined as: 。 8. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 3, characterized in that: Each heterogeneous base learner model has a Dropout layer set after the corresponding layer.

9. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 8, characterized in that: The dropout layer value ranges from 0.28 to 0.

35.

10. The method for collaborative identification of track surface recognition and idling state based on ensemble learning according to claim 8, characterized in that: During the model training phase, a K-fold cross-validation strategy is used to train the base learner.

Images