A method for identifying rail corrugation in rail transit based on the combined characteristics of train vibration and acoustics.

Abstract

This invention belongs to the technical field of railway tracks, specifically disclosing a method for identifying rail corrugation in rail transit based on the composite characteristics of train vibration and noise. The method includes the following steps: acquisition and processing of train vibration, noise, and related signals; enhancement and fusion of vibration and noise signals; mapping of vibration and noise composite data to corrugation relationships and establishment of a sample set; design and training of a convolutional neural network structure for the vibration and noise composite sample set; and identification of rail corrugation status. This invention uses a one-dimensional convolutional neural network to adaptively extract features from the sample set of vibration fusion data based on vibration and noise fusion data. Simultaneously, a random search method is used to determine the optimal parameters, shortening the sample training and identification time, and meeting the requirements for accuracy and timeliness in rail corrugation detection and monitoring.

Description

Technical Field

[0001] This invention belongs to the technical field of railway tracks, and in particular to a method for identifying rail corrugation in rail transit based on the composite characteristics of train vibration and sound. Background Technology

[0002] In recent years, my country's urban rail transit construction has been booming, entering and continuing a period of rapid development. The increasing operational mileage has led to the exposure of numerous problems at the operational sites, among which rail corrugation can affect train operation safety and cause vibration and noise complaints from residents along the line. Therefore, timely detection and early warning of rail corrugation can effectively prevent safety issues and avoid complaints.

[0003] The difficulty in solving the above technical problems lies in:

[0004] Currently, the mainstream detection methods for rail corrugation can be broadly categorized into two types: direct detection and indirect detection. Direct detection involves applying measuring equipment directly to the rail surface when the train is stationary. Portable equipment such as manual calipers and corrugation trolleys offer high accuracy but are inefficient. Indirect detection methods, due to the non-steady-state, nonlinear, and strongly coupled characteristics of the measured dynamic response data, traditional digital signal processing methods are difficult to use for data analysis. They often fail to effectively extract the true fault information from the signal and mostly rely on a single signal input, resulting in large models and low accuracy.

[0005] The significance of solving the above-mentioned technical problems lies in the fact that this invention, based on vibration and acoustic fusion data, uses a one-dimensional convolutional neural network to adaptively extract features from the sample set of vibration fusion data. Simultaneously, a random search method is employed to determine the optimal parameters, shortening the sample training and recognition time, and meeting the requirements for accuracy and timeliness in rail corrugation detection and monitoring. Summary of the Invention

[0006] In response to the problems mentioned above, this invention proposes a method for identifying rail corrugation in rail transit based on the composite characteristics of train vibration and acoustics.

[0007] The technical solution of this invention is:

[0008] A method for identifying rail corrugation in rail transit based on the combined characteristics of train vibration and acoustics includes the following steps:

[0009] Step S101: Acquisition and processing of train vibration, noise and related signals;

[0010] Step S102: Vibration and noise signal enhancement and fusion;

[0011] Step S103: Mapping the relationship between vibration and noise composite data and wave wear, and establishing a sample set;

[0012] Step S104: Design and training of convolutional neural network structure for vibration and noise composite sample set;

[0013] Step S105: Identification of rail corrugation condition.

[0014] Preferably, in step S101, an acceleration sensor is set outside the vehicle to collect vibration signals and train speed signals, and rail corrugation status data is collected simultaneously. A sound pressure sensor is set inside the vehicle to collect noise signals. The collected speed signals are integrated over time to obtain train displacement information. The speed signals are converted into displacement signals. Based on the one-to-one correspondence between the displacement signals and time signals, the time coordinates of the vibration signals and noise signals are replaced.

[0015] More preferably, the acceleration sensor is installed at the axle box or bogie of the train, and the sound pressure sensor is installed inside the train above the bogie, at a height of 1.2m from the bottom of the carriage.

[0016] More preferably, in step S102, a spatial sliding window with an overlap coefficient between 0 and 1 is used to cut the vibration signal and noise signal in the spatial domain obtained in step S101, and the vibration data and noise data in the same sliding window are connected in parallel to form dual-channel data.

[0017] Preferably, step S103 includes the following steps:

[0018] Step S301: Define the principle of excessive rail corrugation. It is agreed that if the roughness exceeds the limit by 3dB at three consecutive 1 / 3 octave bands or exceeds the limit by 6dB at a single frequency, the corrugation of that section is judged to be excessive.

[0019] Step S302: Assign corresponding composite data response labels for vibration and noise based on whether the corrugation exceeds the limit. The corrugation exceeding the limit label is defined as "1", otherwise it is "0".

[0020] Step S303: Divide the mapping sample set into a training set and a test set, accounting for p% and (1-p)% respectively.

[0021] Preferably, in step S104, the composite sample set of vibration and noise is determined based on the amount of data using a one-dimensional convolutional neural network structure, and its layer number, function, and effect are determined. The optimal parameters are then determined using a random search method.

[0022] More preferably, the one-dimensional convolutional neural network structure has a total of 7 layers, including 1 input layer, which is used to feed the input data into the convolutional neural network for feature extraction;

[0023] The second layer is a large convolutional kernel layer, used for short-term feature extraction of vibration and noise fusion samples;

[0024] Layers 3-6 are small convolutional kernel layers, used for capturing local and global features of samples;

[0025] Layer 7 is a sigmoid classification layer used to determine sample labels.

[0026] More preferably, the one-dimensional convolutional neural network structure uses functions including:

[0027] The same zero-padding command is used in convolutional layers to make the input and output lengths of the convolutional layer equal.

[0028] Adam optimizes the training network by iteratively updating the neural network weights based on sample data, setting the learning rate to 0.0001 and using binary_crossentropy as the loss function;

[0029] Batch processing: Training is performed using batches of samples to avoid gradient vanishing and gradient explosion.

[0030] Five-fold cross-validation is used to train the model to improve its generalization ability and enhance the robustness of the trained model.

[0031] Further preferred, the optimal parameters determined by the random search method are: a large convolutional kernel size of 128, a small convolutional kernel size of 8, a stride of 3, a batch size of 512, and a patience value of 100 in the Early Stopping technique of 5-fold cross-validation.

[0032] More preferably, in step S105, the threshold of the sigmoid classification layer is set to 0.5. If the output is lower than this number, the sample label is determined to be 0, indicating no ripple; otherwise, the label is determined to be 1, indicating ripple.

[0033] The beneficial effects of this invention are:

[0034] 1. Compared with indirect detection methods, the technical solution of the present invention can effectively extract real fault information, and the identification result based on multi-data fusion is not less than 95%, while the time consumption meets the requirements of real-time monitoring.

[0035] 2. Compared with the direct detection method, the technical solution of this invention realizes real-time online monitoring of rail corrugation, avoiding repeated complaints that require a lot of manpower and resources, and saving expenses.

[0036] 3. Through this invention, rail grinding can be transformed from condition-based maintenance to preventive maintenance, avoiding vehicle and rail damage and vibration noise caused by excessive rail corrugation. Attached Figure Description

[0037] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that these drawings are designed for illustrative purposes only and are not intended to limit the scope of the present invention. Furthermore, unless specifically indicated, these drawings are intended only to conceptually illustrate the structural construction described herein and are not necessarily drawn to scale.

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 It is a schematic diagram of acceleration, noise, and rail corrugation tests;

[0040] Figure 3 This is a schematic diagram of vibration and noise data fusion;

[0041] Figure 4 This is a schematic diagram of the data mapping between rail corrugation and vibration / acoustic fusion;

[0042] Figure 5 This is a schematic diagram illustrating the definition and signal characteristics of excessive rail corrugation.

[0043] Figure 6 This is a design diagram of a convolutional neural network structure for a vibration and acoustic fusion sample set;

[0044] Figure 7 This is a schematic diagram of the training results of the convolutional neural network structure for the vibration and acoustic fusion sample set. Detailed Implementation

[0045] First, it should be noted that the specific structure, features, and advantages of the present invention will be described in detail below by way of examples. However, all descriptions are for illustrative purposes only and should not be construed as limiting the present invention in any way. Furthermore, any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the accompanying drawings, can still be arbitrarily combined or deleted among these technical features (or their equivalents) to obtain more other embodiments of the present invention that may not be directly mentioned herein. Additionally, for the sake of simplifying the drawings, the same or similar technical features may be indicated only in one place in the same drawing.

[0046] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "screw connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0047] The following is in conjunction with the appendix Figure 1-7 This invention will be described in detail.

[0048] Example 1:

[0049] like Figure 1 As shown, a method for identifying rail corrugation in rail transit based on the composite characteristics of train vibration and acoustics includes the following steps:

[0050] Step S101: Acquisition and processing of train vibration, noise and related signals;

[0051] Step S102: Vibration and noise signal enhancement and fusion;

[0052] Step S103: Mapping the relationship between vibration and noise composite data and wave wear, and establishing a sample set;

[0053] Step S104: Design and training of convolutional neural network structure for vibration and noise composite sample set;

[0054] Step S105: Identification of rail corrugation condition.

[0055] Working principle:

[0056] In the technical solution of this invention, data acquisition is the foundation of this embodiment, including two steps: testing and processing. The processed data is used to create a mapping sample set, which is divided into a training set and a test set. Then, a one-dimensional convolutional neural network is used to adaptively extract features from the sample set of vibration fusion data. At the same time, a random search method is used to determine the optimal parameters, shortening the sample training and recognition time. Through the above operations, the rail corrugation state can be accurately identified, monitored, and located.

[0057] Furthermore, in the embodiments, it can be considered that in step S101, an acceleration sensor is set outside the vehicle to collect vibration signals and train speed signals, and rail corrugation status data is collected at the same time; a sound pressure sensor is set inside the vehicle to collect noise signals; the collected speed signals are integrated over time to obtain train displacement information; the speed signals are converted into displacement signals; and the time coordinates of the vibration signals and noise signals are replaced according to the one-to-one correspondence between the displacement signals and time signals.

[0058] Data collection and testing: such as Figure 2 As shown, the acceleration sensor is placed outside the vehicle and the sound pressure sensor is placed inside the vehicle. A multi-channel data acquisition system is used to simultaneously collect acceleration, noise sound pressure and speed signals, while track irregularity test is carried out to collect rail corrugation data.

[0059] Data processing: By integrating the velocity signal v(t) over time t, the train displacement information s(t) is obtained. Based on the one-to-one correspondence between time t and displacement information s, the independent variable time t of the time-domain acceleration signal a(t) and the noise sound pressure signal p(t) is replaced with displacement information s, thus obtaining the spatial domain acceleration signal a(s) and the spatial domain noise sound pressure signal p(s). This realizes the conversion of time-domain acceleration and noise sound pressure signals into spatial domain signals, unifies the displacement information of spatial domain acceleration, sound pressure, and wave-erosion signals, and obtains the correspondence between acceleration, sound pressure, and wave-erosion signals and the same displacement information s.

[0060] Furthermore, in another embodiment, the acceleration sensor can be positioned at the axle box or bogie, while the sound pressure sensor is positioned inside the train above the bogie, at a height of 1.2m from the bottom of the carriage. Verification has shown that the positions of the acceleration sensor and sound pressure sensor in this embodiment are optimally selected.

[0061] Furthermore, in the embodiments, it can be considered that in step S102, a spatial sliding window with an overlap coefficient between 0 and 1 is used to cut the vibration signal and noise signal in the spatial domain obtained in step S101, and the vibration data and noise data in the same sliding window are connected in parallel to form dual-channel data.

[0062] like Figure 3 As shown, step S102 requires two steps: First, the spatial vibration and noise wave signals obtained in S1 are simultaneously segmented using a spatial sliding window with a size of 5m; then, overlapping sampling is used to improve the data sample utilization rate with a displacement overlap coefficient of 0.875; finally, the vibration and noise signals enhanced by overlapping using the same spatial sliding window are used as a data sample, so that a sample set contains both vibration and noise data, forming parallel dual-channel data.

[0063] Furthermore, in the embodiments, step S103 may include the following steps:

[0064] Step S301: Define the principle of excessive rail corrugation. It is agreed that if the roughness exceeds the limit by 3dB at three consecutive 1 / 3 octave bands or exceeds the limit by 6dB at a single frequency, the corrugation of that section is judged to be excessive.

[0065] Step S302: Assign corresponding composite data response labels for vibration and noise based on whether the corrugation exceeds the limit. The corrugation exceeding the limit label is defined as "1", otherwise it is "0".

[0066] Step S303: Divide the mapping sample set into a training set and a test set, accounting for p% and (1-p)% respectively.

[0067] like Figure 4 As shown, step S103 consists of three steps:

[0068] First, the principle of rail corrugation exceeding the limit is defined as follows: if the roughness level of three consecutive center wavelengths in 1 / 3 octave band exceeds the limit of ISO 3095 by 3 dB, or the roughness level of a single center wavelength exceeds the limit of ISO 3095 by 6 dB, then the corrugation of that section is considered to be exceeding the limit.

[0069] Secondly, such as Figure 5 As shown, depending on whether the corrugation exceeds the limit, the corresponding vibration and noise composite data response label is assigned. That is, if the corrugation exceeds the limit, the label is defined as "1", otherwise it is "0".

[0070] Finally, the mapping sample set is divided into a training set and a test set, accounting for p% and (1-p)% respectively, for training and testing of the structured network.

[0071] Furthermore, in the embodiments, it can be considered that in step S104, the composite sample set of vibration and noise is determined according to the amount of data using a one-dimensional convolutional neural network structure, the number of layers, their functions, and the optimal parameters are determined using a random search method.

[0072] Furthermore, in the embodiments, the one-dimensional convolutional neural network structure has a total of 7 layers, including 1 layer as an input layer, which is used to feed the input data into the convolutional neural network for feature extraction.

[0073] The second layer is a large convolutional kernel layer, used for short-term feature extraction of vibration and noise fusion samples;

[0074] Layers 3-6 are small convolutional kernel layers, used for capturing local and global features of samples;

[0075] Layer 7 is a sigmoid classification layer used to determine sample labels.

[0076] Furthermore, in the embodiments, the one-dimensional convolutional neural network structure may utilize functions including:

[0077] The same zero-padding command is used in convolutional layers to make the input and output lengths of the convolutional layer equal.

[0078] Adam optimizes the training network by iteratively updating the neural network weights based on sample data, setting the learning rate to 0.0001 and using binary_crossentropy as the loss function;

[0079] Batch processing: Training is performed using batches of samples to avoid gradient vanishing and gradient explosion.

[0080] Five-fold cross-validation is used to train the model to improve its generalization ability and enhance the robustness of the trained model.

[0081] Furthermore, in the embodiments, the optimal parameters determined by the random search method are: a large convolutional kernel size of 128, a small convolutional kernel size of 8, a stride of 3, a batch size of 512, and a patience of 100 in the EarlyStopping technique of 5-fold cross-validation.

[0082] The design and training results of the convolutional neural network structure for vibration-acoustic fusion samples are as follows: Figure 7 As shown, Figure 7 Images (a) and (b) illustrate the training process of the convolutional neural network on the vibration signal dataset. It is evident that as the number of training epochs increases, the training and testing accuracy of the model continuously improves, while the loss level decreases. When the accuracy on the test set no longer improves, training is stopped by the overfitting technique Dropout. Figure 7 As shown in (c), the recognition accuracy of the test set and the test time of each sample are used as evaluation criteria to verify the model's recognition rate and timeliness. It can be seen that the diagnostic rate is not less than 95% after 10 trials, the average accuracy is 96.66%, the standard deviation is 0.0055, and the test sample time is less than 0.2ms.

[0083] Furthermore, in the embodiments, it can be considered that in step S105, the threshold of the sigmoid classification layer is agreed to be 0.5. If the output is lower than this number, the sample label is determined to be 0, which means there is no ripple; otherwise, the label is determined to be 1, which means there is ripple.

[0084] State recognition is the goal of this embodiment. For rail corrugation state recognition, a normal rail is labeled as 0, and a corrugated rail is labeled as 1. It is stipulated that if the output of the sigmoid activation layer is less than 0.5, the sample is labeled as 0; otherwise, it is labeled as 1. Through the above operations, the rail corrugation state can be accurately identified and located.

[0085] In summary, this invention provides a method for identifying rail corrugation in rail transit based on the composite characteristics of train vibration and sound, which can effectively extract real fault information and meets the requirements for real-time monitoring in terms of time consumption.

[0086] The above embodiments have provided a detailed description of the present invention, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A rail transit rail corrugation identification method based on train vibration-sound composite features, characterized in that, The method comprises the following steps: Step S101: collecting and processing train vibration, noise and related signals; Step S102: enhancing and fusing vibration and noise signals; Step S103: mapping vibration and noise composite data and wave abrasion relationship and establishing a sample set; Step S104: designing and training a convolutional neural network structure of the vibration and noise composite sample set; Step S105: identifying a rail wave abrasion state; In the step S101, an acceleration sensor is arranged outside the train to collect vibration signals and train speed signals, and rail wave abrasion state data is collected at the same time, and an acoustic pressure sensor is arranged inside the train to collect noise signals; the speed signal is time-integrated to obtain train displacement information, the speed signal is converted into a displacement domain signal, and the displacement signal is used to replace the time coordinates of the vibration signal and the noise signal according to the one-to-one correspondence between the train displacement and the time; In the step S102, a spatial sliding window with an overlap coefficient between 0 and 1 is used to cut the vibration signal and the noise signal in the spatial domain obtained in the step S101, and the vibration data and the noise data in the same sliding window are connected in parallel to form double-channel data; The step S103 comprises the following steps: Step S301: defining a rail wave abrasion overrun principle, and agreeing that the roughness at three 1 / 3 octave bands is overrun by 3 dB or a single frequency is overrun by 6 dB, and judging whether the wave abrasion of the section is overrun; Step S302: according to whether the wave abrasion is out of limit, a corresponding vibration and noise composite data response label is given, the wave abrasion overrun label is defined as "1", and otherwise, it is "0"; Step S303: dividing the mapping sample set into a training set and a test set, and the proportions are p% and (1-p)%, respectively; The acceleration sensor is arranged at a train axle box or a bogie, and the acoustic pressure sensor is arranged above the bogie in the train, at a height of 1.2 m from the bottom of the carriage; In the step S104, according to the data quantity, a one-dimensional convolutional neural network structure is used for the vibration and noise composite sample set, the number of layers, the function and the function are determined, and the random search method is used to determine the optimal parameters; The one-dimensional convolutional neural network structure has 7 layers, including 1 input layer for inputting data into the convolutional neural network for feature extraction; 2 layers of large convolution kernel convolutional layers are used for short-time feature extraction of vibration and noise fusion samples; Layers 3-6 are small convolution kernel convolutional layers for capturing local and global features of samples; Layer 7 is a sigmoid classification layer for determining sample labels; The functions used in the one-dimensional convolutional neural network structure include: The same zero padding command is used in the convolutional layer, so that the length of the input and the output of the convolutional layer is equal; Adam optimization is used to train the network, so that the neural network weight is updated based on the sample data iteration, the learning rate is set to 0.0001, and the binary_crossentropy loss function is used; Batch processing is used to train the sample to avoid gradient dispersion and gradient explosion; 5-fold cross-validation is used to train the model to improve its generalization ability and improve the model robustness of the trained model.

2. The rail transit rail corrugation identification method based on train vibration-sound combined features according to claim 1, characterized in that: The optimal parameters determined by the random search method are: a large convolution kernel size of 128, a small convolution kernel size of 8, a step of 3, a batch size of 512, and a patience of 100 in the Early Stopping technology of the 5-fold cross-validation method.

3. The rail transit rail corrugation identification method based on train vibration-sound combined characteristics according to claim 2, characterized in that: In the step S105, the sigmoid classification layer threshold value is conventionally 0.5, and if the output is lower than the number, it is determined that the sample label is 0, that is, there is no burnishing, otherwise it is determined that the label is 1, which represents that there is burnishing.

Images