Fusion time-frequency domain feature-based switch point rail crack detection method and device

Abstract

The invention provides a switch point rail crack detection method and device based on fusion time-frequency domain characteristics, and relates to the technical field of point rail crack detection.The method comprises the steps that vibration signals of different knocking points of a switch point rail are obtained, and the vibration signals comprise force signals and acceleration signals; performing synchronous compression wavelet transform on the vibration signal to obtain a synchronous compression wavelet transform result; calculating a frequency response function of the vibration signal; constructing a multi-modal vibration signal fusion model; based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result, training the multi-modal vibration signal fusion model to obtain a crack damage identification model; and detecting the actual crack of the switch rail through the crack damage identification model to obtain a damage result of the actual switch rail. The problems that an existing detection method is low in sensitivity and high in measurement cost, and long-time real-time detection cannot be achieved are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of switch rail crack detection, and in particular to a switch rail crack detection method and device based on fusion of time-frequency domain features. Background Art

[0002] In recent years, the volume of railway transportation has also increased. The cyclic effects of wheel loads and temperature loads will accelerate the fracture of rails. This will cause train derailment and cause huge casualties. In the railway track structure, turnouts are an indispensable device. Railway turnouts are composed of point rails, heart rails, slide bed plates and fasteners. Compared with ordinary rails, the cross-section of the turnout is very thin. Due to the special function of its structure, the contact track is discontinuous when the vehicle passes through the turnout switch. This will cause the impact force between the wheel and rail to be greater than that of ordinary lines, which will cause the point rail to crack more easily than ordinary equal-section rails. However, more and more turnouts are laid in uninhabited areas at high altitudes, and it is difficult to regularly detect their status. Cracks further lead to rail fractures, which will directly affect the safety of train operation.

[0003] Traditional track crack detection is usually done manually and cannot detect tiny cracks or damage in hidden locations. Acoustic wave real-time broken rail detection has poor detection sensitivity at mechanical insulation and is not suitable for turnout damage detection. Ultrasonic guided wave detection has high measurement costs and cannot achieve long-term real-time detection at turnouts due to issues such as measurement equipment and sensor size. Summary of the invention

[0004] The purpose of the present invention is to provide a switch rail crack detection method and device based on the fusion of time-frequency domain features to improve the above problems. In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present application provides a switch rail crack detection method based on fusion of time-frequency domain features, comprising:

[0006] Acquire vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal;

[0007] Performing synchronous compression wavelet transform on the vibration signal to obtain a synchronous compression wavelet transform result;

[0008] Calculating a frequency response function of the vibration signal;

[0009] Construct a multimodal vibration signal fusion model;

[0010] Based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result, the multimodal vibration signal fusion model is trained to obtain a crack damage identification model;

[0011] The crack damage identification model is used to detect cracks in actual switch point rails to obtain damage results of actual switch point rails.

[0012] In a second aspect, the present application also provides a switch rail crack detection device based on fusion of time-frequency domain features, comprising:

[0013] An acquisition module, used for acquiring vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal;

[0014] A transformation module, used for performing synchronous compression wavelet transformation on the vibration signal to obtain a synchronous compression wavelet transformation result;

[0015] A calculation module, used for calculating the frequency response function of the vibration signal;

[0016] A building module for building a multi-modal vibration signal fusion model;

[0017] A training module, used for training the multimodal vibration signal fusion model based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result to obtain a crack damage identification model;

[0018] The prediction module is used to detect cracks in actual switch point rails through the crack damage identification model to obtain damage results of actual switch point rails.

[0019] The beneficial effects of the present invention are as follows: the vibration signal is converted by calculating the frequency response function and synchronous compression wavelet transform, the characteristics of the multimodal data are identified by a multimodal vibration signal fusion model, cracks can be identified and judged only by the vibration signal, and the present invention can be applied to complex turnout structures with good effects, and can effectively avoid the poor sensitivity and reliability in the detection of complex variable-section rails such as turnouts. In addition, the energy consumption of the measuring sensor is reduced by collecting medium and low frequency vibration signals, and the measurement cost and data transmission cost are significantly reduced, thereby realizing long-term real-time detection in the turnout.

[0020] Other features and advantages of the present invention will be set forth in the following description, and in part will become apparent from the description, or may be understood by practicing embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 It is a schematic flow chart of a switch rail crack detection method based on fusion of time-frequency domain features described in an embodiment of the present invention;

[0023] Figure 2 Schematic diagram of the structure of the crack damage identification model described in the embodiment of the present invention;

[0024] Figure 3 Schematic diagram of the arrangement of some measuring points and striking points in an embodiment of the present invention.

[0025] Figure 4 Schematic diagram of a time-frequency vibration signal in an embodiment of the present invention;

[0026] Figure 5 It is a time-frequency diagram after synchronous compression wavelet transformation in an embodiment of the present invention;

[0027] Figure 6 A schematic diagram of a frequency response function in an embodiment of the present invention;

[0028] Figure 7 It is a time domain comparison diagram of vibration signals in different damage states in an embodiment of the present invention;

[0029] Figure 8 A comparison diagram of confusion matrices for classification using different data set training models in an embodiment of the present invention;

[0030] Fig. 9 It is a schematic diagram of the structure of the switch point rail crack detection device based on the fusion of time-frequency domain features described in an embodiment of the present invention.

[0031] Markings in the figure: 800, switch point rail crack detection equipment based on fusion of time-frequency domain features; 801, processor; 802, memory; 803, multimedia component; 804, I / O interface; 805, communication component. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0034] Embodiment 1:

[0035] This embodiment provides a switch point rail crack detection method based on the fusion of time-frequency domain features.

[0036] See also Figure 1 , the figure shows that the method includes step S100, step S200, step S300, step S400, step S500, and step S600.

[0037] Step S100: Acquire vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal;

[0038] In this embodiment, the acceleration signal is a vibration acceleration signal.

[0039] The step S100 includes:

[0040] Arrange sensors at different positions on the switch rail;

[0041] Setting a plurality of striking points, striking each striking point multiple times by a force hammer, wherein the plurality of striking points are located at different positions of the switch point rail;

[0042] The acceleration signals and force signals of the multiple sensors are acquired when each tapping point is tapped.

[0043] In this embodiment, crack damage with different characteristics is produced on the complete switch point rail, and the hammering is performed on the sleeper and in the middle of the span, and the hammering is performed with random force. The uneven hammering method can avoid hitting the modal nodes of the turnout rail structure.

[0044] Step S200: performing synchronous compression wavelet transform on the vibration signal to obtain a synchronous compression wavelet transform result;

[0045] In this embodiment, the vibration signal is subjected to wavelet transform, instantaneous frequency estimation, compression redistribution and signal reconstruction operations in sequence through synchronous compression wavelet transform.

[0046] The concentration of the time-frequency diagram is enhanced by redistributing the wavelet coefficients. Specifically, after calculating the wavelet transform of the signal, the instantaneous frequency is estimated using the phase information of the wavelet coefficients, and then the wavelet coefficients are redistributed to the corresponding frequency positions, thereby compressing the time-frequency diagram and making it more concentrated.

[0047] The step S200 includes:

[0048] Performing continuous wavelet transform on the vibration signal to obtain wavelet coefficients at different scales and at different times;

[0049] The calculation formula of the wavelet coefficient is:

[0050]

[0051] Where W x (·) represents the wavelet coefficient, a represents the scale parameter, b represents the translation parameter, express The complex conjugate of Represents the mother wavelet function, and x(t) represents the signal at time t, that is, the vibration signal.

[0052] Calculating the phase derivative of the wavelet coefficient, and obtaining the instantaneous frequency of the vibration signal based on the phase derivative;

[0053] The expression of the instantaneous frequency is:

[0054]

[0055] In the formula, ω x (·) represents the instantaneous frequency, a represents the scale parameter, b represents the translation parameter, arg(W x (a, b)) represents the phase of the wavelet coefficients, W x (·) represents the wavelet coefficient, It means partial derivative.

[0056] Based on the wavelet coefficients and the instantaneous frequency, compressing and redistributing the wavelet coefficients to obtain a compressed time-frequency diagram;

[0057] In this embodiment, the wavelet coefficients are compressed and redistributed to obtain a compressed time-frequency diagram, that is, a time-frequency representation after synchronous compression wavelet transformation. The process of compression and redistribution can be regarded as an optimization problem, the goal of which is to minimize the objective function.

[0058] The time-frequency representation and objective function formula after the synchronous compression wavelet transform are as follows:

[0059]

[0060] Where, T x (·) represents the time-frequency representation after synchronous compression wavelet transform, ω represents the angular frequency, a represents the scale parameter, b represents the translation parameter, W x (·) represents the wavelet coefficient, A represents the range of the scale parameter, ω x(·) represents the instantaneous frequency, δ(·) represents the Dirac delta function, J(a,b) represents the objective function with respect to a and b, arg(W x (a,b)) represents the phase of the wavelet coefficients, represents partial derivative, and |·| represents absolute value.

[0061] The compressed time-frequency diagram is inversely transformed to obtain a synchronous compression wavelet transform result.

[0062] In this embodiment, a synchronous compression wavelet transform result is obtained through inverse transformation, that is, the original vibration signal is obtained, and the synchronous compression wavelet transform result is a two-dimensional image.

[0063] The calculation formula of the synchronous compression wavelet transform result is:

[0064]

[0065] In the formula, x'(t) represents the result of synchronous compression wavelet transform, t represents time, T x (·) represents the time-frequency representation after synchronous compression wavelet transform, ω represents the angular frequency, a represents the scale parameter, b represents the translation parameter, represents the mother wavelet function, Represents the normalization constant of the wavelet function, which is determined by the characteristics of the mother wavelet.

[0066] Step S300: calculating the frequency response function of the vibration signal;

[0067] The step S300 includes:

[0068] Calculating a frequency response function curve of the vibration signal;

[0069] Performing Fourier transform on the vibration signal and the frequency response function curve respectively to obtain a frequency domain representation of the vibration signal and a frequency domain representation of the frequency response function curve;

[0070] The frequency response function of the vibration signal is calculated through the frequency domain representation of the vibration signal and the frequency domain representation of the frequency response function curve.

[0071] The calculation formula of the frequency response function is:

[0072]

[0073] Where H(·) represents the frequency response function, X(·) represents the frequency domain representation of the vibration signal, Y(·) represents the frequency domain representation of the frequency response function curve, and ω represents the angular frequency.

[0074] Frequency response function plays an important role in modal analysis and can be used to identify the natural frequency, damping ratio and modal shape of the system. By analyzing the amplitude-frequency characteristics and phase-frequency characteristics of the frequency response function, we can gain a deep understanding of the dynamic behavior of the system and conduct effective analysis and design.

[0075] The essence of the frequency response function is the specific performance of the system transfer function in the frequency domain, which reflects the response characteristics of the system to input signals of different frequencies. In order to calculate the frequency response function from the time domain signal, the common method is to convert the input and output signals from the time domain to the frequency domain through Fourier transform, and then calculate using the above formula. In experimental analysis, the frequency response function is usually obtained by applying a known excitation (such as an impact hammer or vibrator) and measuring the corresponding response signal.

[0076] Step S400: constructing a multi-modal vibration signal fusion model;

[0077] Step S500: training the multimodal vibration signal fusion model based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result to obtain a crack damage identification model;

[0078] In this embodiment, before model training, each vibration signal is labeled, and the label indicates the damage status of the switch point rail.

[0079] Step S600: Detect cracks on actual switch point rails using the crack damage identification model to obtain damage results of actual switch point rails.

[0080] The step S600 includes:

[0081] Convolve the actual acceleration signal and the actual frequency response function through the convolution kernel to obtain the convolution feature;

[0082] Encoding the convolutional features through a weight coding layer to obtain coded features, wherein the weight coding layer includes a first branch, a second branch, and a feature vector dot product layer;

[0083] Performing vector transformation on the feature code through a vector flattening layer to obtain a transformed feature, and inputting the transformed feature into a long short-term memory network to obtain a first output feature;

[0084] The actual synchronous compression wavelet transform result is subjected to feature extraction through a feature extraction layer to obtain a second output feature;

[0085] Performing feature fusion on the first output feature and the second output feature through a feature vector fusion layer to obtain a fused feature;

[0086] The fused features are input into the fully connected layer and the classifier in sequence to obtain the damage results of the actual switch rail.

[0087] The convolutional features are input into the weight coding layer for encoding to obtain the encoded features, including:

[0088] The convolutional features are subjected to batch normalization, a first activation operation, a maximum pooling operation, a fully connected layer connection, and a second activation operation in sequence through the first branch to obtain the first branch features, wherein the first activation operation and the second activation operation both use the relu activation function;

[0089] The convolutional features are sequentially subjected to average pooling operations, full connection layer connections, and mapping operations through the second branch to obtain second branch features, wherein the mapping operation uses a sigmoid function;

[0090] The first branch feature and the second branch feature are subjected to a dot product operation through a feature vector dot product layer to obtain a coding feature.

[0091] In this embodiment, the trained multimodal vibration signal fusion model, i.e., the crack damage recognition model, combines a convolutional neural network, a long short-term memory network, and a weight coding layer designed by the present invention, and uses multiple parallel branches to process multimodal inputs.

[0092] The initial data set is constructed by tapping the acceleration signals and force signals with different crack defect time courses. Then the single time series signal is subjected to two different feature enhancement methods, namely, the synchronous compression wavelet transform is applied to convert the vibration signal generated by the tapping into a time-frequency spectrum based on the 2D image and a force signal combined with the hammer vibration, and the ordinary single force signal is given its physical meaning and converted into a frequency response function, making its physical properties more interpretable. These two methods not only expand the single time domain to the frequency domain, or even the combination of the time domain and the frequency domain, but also greatly increase the interpretability of the training data itself, and make it easier to identify smaller crack damage defects.

[0093] In this embodiment, the specific structure of the crack damage identification model is as follows: Figure 2 As shown, it includes a convolution kernel, a weight coding layer, a vector flattening layer, a long short-term memory network LSTM, a feature extraction layer, a feature vector fusion layer, a fully connected layer and a classifier, wherein the convolution kernel, the weight coding layer, the vector flattening layer, the long short-term memory network LSTM, the feature vector fusion layer, the fully connected layer and the classifier are connected in sequence, and the feature vector fusion layer is also connected to the feature extraction layer.

[0094] The weight coding layer includes a first branch, a second branch and a feature vector dot product layer. The first branch includes a batch normalization processing layer, a relu activation function, a maximum pooling layer, a fully connected layer connection and a relu activation function connected in sequence. The second branch includes an average pooling layer, a fully connected layer and a sigmoid function connected in sequence.

[0095] The feature extraction layer includes a convolution kernel, a batch normalization processing layer, a relu activation function, a maximum pooling layer, a convolution kernel, a batch normalization processing layer, a relu activation function, a maximum pooling layer and a vector flattening layer connected in sequence.

[0096] In this embodiment, multiple convolutional layers are set to extract complex key features from complex time-frequency images after synchronous compression wavelet transformation. Pooling layers are constructed to reduce the dimension of extracted features while retaining key information. On the basis of convolutional layers and pooling layers, fully connected layers are further introduced to integrate the extracted features to form a more compact feature representation.

[0097] The output is mapped to a range between 0 and 1 using the Sigmoid function, and the self-attention mechanism of the weight coding layer is introduced. This mechanism effectively enhances the model's ability to capture key information by assigning different weights to each time point in the input sequence, while reducing the impact of irrelevant or redundant information. This not only improves the model's ability to understand multimodal input data globally, but also improves the accuracy and robustness of crack detection, facilitating subsequent classification tasks.

[0098] In addition, in order to process time series information, the model integrates a long short-term memory network to capture the long-term dependencies and temporal dynamic characteristics in time series data. This step is crucial for the time domain analysis of vibration signals and can effectively identify small changes in the time dimension.

[0099] Embodiment 2:

[0100] In this embodiment, the vibration characteristics of the combined rails in the switch area under healthy and crack damaged unhealthy conditions are investigated, and a hammer excitation test is carried out on the combined rails in a high-speed switch area for a whole set of complete switch point rails with different crack damage conditions.

[0101] A plurality of high-precision vibration acceleration sensors are installed on the rail surface, and a crack damage feature information library of the vibration signal of the turnout variable-section rail is established based on the test data of this embodiment. Through the vibration hammer test, a large number of vibration signals of different damage degrees and different damage locations of the turnout rail are obtained, and the mapping relationship between the damage features and the vibration characteristics of the turnout rail is obtained.

[0102] In this embodiment, crack damage of different damage degrees and positions was artificially created on the complete turnout point rail, considering different positions (4.2m, 10m, 15m from the tip) and different crack damage degrees (1cm, 2cm, 4cm and 6cm respectively). The rail crack was set on the working side of the rail bottom. A complete force hammer vibration test was carried out for each damage state, and the vibration response characteristics of each position were obtained in detail.

[0103] This embodiment uses a high-performance data acquisition instrument, namely a PCB three-axis acceleration sensor, to obtain vibration signals. The sensor has a sensitivity of 50mV / g, a range of ±100g, and a resolution of 0.0002grms, and is suitable for measuring vibration signals in the range of 1Hz to 5kHz.

[0104] The PCB three-axis acceleration sensor measures the time-series vibration signals of the vertical, horizontal and vertical directions of the pointed rail, which are located at the tip rail waist, the heel rail waist and the middle rail bottom, respectively. Together with the force signal of the hammer, a total of 10 channels are occupied, that is, the hammer is used for one excitation, and the collector will simultaneously collect the time-series data of 10 channels, namely the force signal, the tip signal in three directions, the heel signal and the middle signal, among which the tip signal, the heel signal and the middle signal are the acceleration signals in the vibration signal. At the same time, in order to enrich the source of the acquired vibration signal and increase the sample, each knocking point is repeated 20 times, and the positions of some knocking points are shown in Table 1.

[0105] Table 1

[0106]

[0107] As can be seen from Table 1, the striking points are not evenly distributed. On the one hand, this can prevent the striking points from being exactly at the vibration nodes. On the other hand, the excitation points of the switch rail assembly are both at the turnout sleeper position and the mid-span position, so that the switch rail can be excited more randomly. The layout of some measuring points and striking points is shown in Figure 3 .

[0108] In this embodiment, the original signals of the hammer striking corresponding to different rail crack defects and the time-frequency domain results after the calculation of the frequency response function and the synchronous compression wavelet transform are as follows: Figure 4-6 As shown, Figure 4 represents the time-frequency vibration signal, Figure 5 represents the time-frequency diagram after synchronous compression wavelet transform, Figure 6 represents the frequency response function.

[0109] like Figure 4-6 As shown in Figure 2, there is no obvious difference in the distribution of the original vibration signal in each case. Although the decay trend after the peak shows some differences, no obvious pattern is shown, and it is not practical to manually identify and analyze these features. Figure 7 As shown in the figure, it is a time domain comparison diagram of vibration signals in different damage states. Figure 7 The time domain of vibration signals under different damage states is shown, that is, the time domain of vibration acceleration. Therefore, it is difficult to visually distinguish and identify key features from the original time domain signal.

[0110] The present invention integrates one-dimensional features and two-dimensional features, and integrates features in the time domain and frequency domain. The frequency response function in the frequency domain transformed from the most basic acceleration signal in the time domain and the force signal of the hammer is used as one of the sources. Then, the two-dimensional image in the time-frequency domain after synchronous compression wavelet transformation is used as another part of the multimodal source, and accurate feature extraction and recognition are performed through the established model.

[0111] In this embodiment, low-frequency vibration signals are used for identification, and 500 data points are extracted from each time-series vibration signal sample, which are converted into 500*1 data points after frequency response function calculation. After synchronous compression wavelet transform calculation, the data points are 224*500. In order to improve efficiency, each image is compressed into 32×32 by reducing the resolution ratio and color channels, and the processed images are saved as data sets.

[0112] Before entering the model, the dataset was standardized and randomly divided into training, validation, and test sets. 70% of the data was selected for training, and the remaining 30% of the data was used for testing. 80% of the data in the training set was used for training, and the remaining 20% ​​of the data was used for validation. The Adam algorithm with a learning rate of 0.001 and a batch size of 70 was used in the training process of the model. The number of iterations during the training process was set to 20.

[0113] In this embodiment, the classification results of different degrees of crack damage at 4.2m are taken as an example to calculate the evaluation indicators, including accuracy, precision, recall and F1 score. Among them, all evaluation indicators of the tip signal are 100% regardless of whether it is single-axis or three-axis. The three-axis evaluation indicators of the follow-end signal are all 100%, and the accuracy, precision, recall and F1 score of the single-axis indicators are all over 99.6%. It can be seen that the classification effect of the results of both the single-axis sensor and the three-axis sensor is very good. And it is judged that the difficulty of identifying tiny cracks by the follow-end signal is higher than that of the tip signal, but the vibration signal of the three-axis sensor can well improve the various classification indicators under the fusion of the detection method of the present invention.

[0114] As for the classification results of different crack damage positions, the model proposed in the present invention shows extremely high accuracy in classifying different damage positions, whether it is the tip or the heel, the data of the single-axis sensor or the three-axis sensor, all indicators are 100%.

[0115] At the same time, under mixed conditions, that is, the classification results of different damage degrees and position combinations, all evaluation indicators of the tip signal are 100% whether it is single-axis or three-axis. For the single-axis signal, the heel vibration signal has all indicators exceeding 99.9% in 2270 samples, and the accuracy of the three-axis signal has improved. It also proves that the effect of the tip signal is slightly better than that of the heel signal, and the fusion of the three-axis vibration signal can significantly improve the classification and recognition of the vibration signal.

[0116] This embodiment organizes the vibration signal data set to obtain three data sets, namely, a data set with only time series vibration response, a data set with only synchronous compression wavelet transform, and a multimodal data set of this embodiment. The model proposed in the present invention is trained by the three data sets respectively, such as Figure 8 As shown, from left to right are respectively a dataset with only time series vibration response, a dataset with only synchronous compression wavelet transform, and the confusion matrix of the model classification result trained by the multimodal dataset of this embodiment. It can be seen that the training effect of the model of the present invention using multimodal datasets at the same time is much better than that of the model trained using single modal data alone.

[0117] like Figure 8 As shown in the figure, the amount of data on the diagonal from the upper left corner to the lower right corner of the confusion matrix can intuitively represent the classification effect. The horizontal and vertical axes represent the predicted labels and true labels of the health status, respectively. The larger the value, the more accurate the classification. From the confusion matrix, the classification effect using multimodality is the best, and the classification accuracy of all categories is 100%, showing a very strong classification ability. In contrast, the model trained with single modality showed lower classification accuracy in some categories, and the classification accuracy of the overall test set was 85% and 93% respectively. The two working conditions with obvious classification errors are 1cm and 2cm in damage degree. This is mainly because the degree of crack is small, and the difference in its physical properties is not very large. Identification through low-frequency vibration signals must require the model to have a more powerful feature recognition ability.

[0118] Therefore, the detection method proposed in the present invention can identify tiny cracks without manual visual inspection. It can identify and judge only by inputting vibration signals into the training model. It can be applicable to complex turnout structures with good effects, and can effectively avoid the poor sensitivity and reliability in the detection of complex variable-section rails such as turnouts. In addition, by collecting medium and low frequency vibration signals, specifically an analysis frequency of 1600 Hz, the advantage is that the energy consumption of the measuring sensor can be reduced, and the measurement cost and data transmission cost are significantly reduced. Therefore, long-term real-time detection can be achieved in the turnout.

[0119] Embodiment 3:

[0120] This embodiment provides a switch rail crack detection device based on the fusion of time-frequency domain features, the device comprising:

[0121] An acquisition module, used for acquiring vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal;

[0122] A transformation module, used for performing synchronous compression wavelet transformation on the vibration signal to obtain a synchronous compression wavelet transformation result;

[0123] A calculation module, used for calculating the frequency response function of the vibration signal;

[0124] A building module for building a multi-modal vibration signal fusion model;

[0125] A training module, used for training the multimodal vibration signal fusion model based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result to obtain a crack damage identification model;

[0126] The prediction module is used to detect cracks in actual switch point rails through the crack damage identification model to obtain damage results of actual switch point rails.

[0127] The acquisition module comprises:

[0128] A first setting unit is used to arrange sensors at different positions of the switch rail;

[0129] A second setting unit is used to set a plurality of striking points, and strike each striking point multiple times by a force hammer, wherein the plurality of striking points are located at different positions of the switch point rail;

[0130] The first acquisition unit acquires acceleration signals and force signals of the multiple sensors when each tapping point is tapped.

[0131] The transformation module comprises:

[0132] A first transformation unit is used to perform continuous wavelet transformation on the vibration signal to obtain wavelet coefficients at different scales and different times;

[0133] a first calculation unit, configured to calculate a phase derivative of the wavelet coefficient, and obtain an instantaneous frequency of the vibration signal based on the phase derivative;

[0134] A reallocation unit, configured to compress and reallocate the wavelet coefficients based on the wavelet coefficients and the instantaneous frequency to obtain a compressed time-frequency diagram;

[0135] The second transformation unit is used to perform inverse transformation on the compressed time-frequency diagram to obtain a synchronous compression wavelet transformation result.

[0136] The prediction module comprises:

[0137] A convolution unit is used to convolve the actual acceleration signal and the actual frequency response function through a convolution kernel to obtain a convolution feature;

[0138] An encoding unit, used for encoding the convolutional features through a weight coding layer to obtain encoded features, wherein the weight coding layer includes a first branch, a second branch and a feature vector dot product layer;

[0139] A third transformation unit is used to perform vector transformation on the feature code through a vector flattening layer to obtain a transformation feature, and input the transformation feature into a long short-term memory network to obtain a first output feature;

[0140] A feature extraction unit, used for extracting features from the actual synchronous compression wavelet transform result through a feature extraction layer to obtain a second output feature;

[0141] A feature fusion unit, used for fusing the first output feature and the second output feature through a feature vector fusion layer to obtain a fused feature;

[0142] The prediction unit is used to input the fused features into the fully connected layer and the classifier in sequence to obtain the damage results of the actual switch rail.

[0143] The encoding unit comprises:

[0144] A first encoding subunit is used to perform batch normalization, a first activation operation, a maximum pooling operation, a fully connected layer connection, and a second activation operation on the convolutional features in sequence through the first branch to obtain a first branch feature, wherein the first activation operation and the second activation operation both use a relu activation function;

[0145] The second encoding subunit is used to perform an average pooling operation, a fully connected layer connection and a mapping operation on the convolutional features in sequence through the second branch to obtain second branch features, wherein the mapping operation adopts a sigmoid function;

[0146] The feature vector dot product unit is used to perform a dot product operation on the first branch feature and the second branch feature through a feature vector dot product layer to obtain a coding feature.

[0147] It should be noted that, regarding the device in the above embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0148] Embodiment 4:

[0149] Corresponding to the above method embodiment, this embodiment also provides a switch point rail crack detection device based on the fusion of time-frequency domain features. The switch point rail crack detection device based on the fusion of time-frequency domain features described below and the switch point rail crack detection method based on the fusion of time-frequency domain features described above can be referenced to each other.

[0150] Fig. 9 FIG. 8 is a block diagram of a switch rail crack detection device 800 based on the fusion of time-frequency domain features according to an exemplary embodiment. Fig. 9 As shown, the switch rail crack detection device 800 based on the fusion of time-frequency domain features may include: a processor 801, a memory 802. The switch rail crack detection device 800 based on the fusion of time-frequency domain features may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0151] The processor 801 is used to control the overall operation of the switch point rail crack detection device 800 based on the fusion of time-frequency domain features to complete all or part of the steps in the above-mentioned switch point rail crack detection method based on the fusion of time-frequency domain features. The memory 802 is used to store various types of data to support the operation of the switch point rail crack detection device 800 based on the fusion of time-frequency domain features. These data may include, for example, instructions for any application or method operating on the switch point rail crack detection device 800 based on the fusion of time-frequency domain features, as well as application-related data, such as contact data, sent and received messages, pictures, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 802 or sent via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules can be keyboards, mice, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the switch point rail crack detection device 800 based on fusion of time-frequency domain features and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 805 can include: Wi-Fi module, Bluetooth module, NFC module.

[0152] In an exemplary embodiment, the switch point rail crack detection device 800 based on the fusion of time and frequency domain features can be implemented by one or more application specific integrated circuits (Application Specific Integrated Circuit, referred to as ASIC), digital signal processors (Digital Signal Processor, referred to as DSP), digital signal processing devices (Digital Signal Processing Device, referred to as DSPD), programmable logic devices (Programmable Logic Device, referred to as PLD), field programmable gate arrays (Field Programmable Gate Array, referred to as FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned switch point rail crack detection method based on the fusion of time and frequency domain features.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0154] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A switch rail crack detection method based on fusion of time-frequency domain features, characterized in that: include: Acquire vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal; Performing synchronous compression wavelet transform on the vibration signal to obtain a synchronous compression wavelet transform result; Calculating a frequency response function of the vibration signal; Construct a multimodal vibration signal fusion model; Based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result, the multimodal vibration signal fusion model is trained to obtain a crack damage identification model; The crack damage identification model is used to detect cracks in actual switch point rails to obtain damage results of actual switch point rails.

2. The switch rail crack detection method based on the fusion of time-frequency domain features according to claim 1 is characterized in that , obtain the vibration signals of different striking points of the switch rail, including: Arrange sensors at different positions on the switch rail; Setting a plurality of striking points, striking each striking point multiple times by a force hammer, wherein the plurality of striking points are located at different positions of the switch point rail; The acceleration signals and force signals of the multiple sensors are acquired when each tapping point is tapped.

3. The switch rail crack detection method based on the fusion of time-frequency domain features according to claim 1 is characterized in that , performing synchronous compression wavelet transform on the vibration signal to obtain a synchronous compression wavelet transform result, including: Performing continuous wavelet transform on the vibration signal to obtain wavelet coefficients at different scales and at different times; Calculating the phase derivative of the wavelet coefficient, and obtaining the instantaneous frequency of the vibration signal based on the phase derivative; Based on the wavelet coefficients and the instantaneous frequency, compressing and redistributing the wavelet coefficients to obtain a compressed time-frequency diagram; The compressed time-frequency diagram is inversely transformed to obtain a synchronous compression wavelet transform result.

4. The switch rail crack detection method based on the fusion of time-frequency domain features according to claim 1 is characterized in that ,The crack damage identification model is used to detect cracks in the actual switch rail, and the damage results of the actual switch rail are obtained, including: Convolve the actual acceleration signal and the actual frequency response function through the convolution kernel to obtain the convolution feature; Encoding the convolutional features through a weight coding layer to obtain coded features, wherein the weight coding layer includes a first branch, a second branch, and a feature vector dot product layer; Performing vector transformation on the feature code through a vector flattening layer to obtain a transformed feature, and inputting the transformed feature into a long short-term memory network to obtain a first output feature; The actual synchronous compression wavelet transform result is subjected to feature extraction through a feature extraction layer to obtain a second output feature; Performing feature fusion on the first output feature and the second output feature through a feature vector fusion layer to obtain a fused feature; The fused features are input into the fully connected layer and the classifier in sequence to obtain the damage results of the actual switch rail.

5. The switch rail crack detection method based on the fusion of time-frequency domain features according to claim 4 is characterized in that , the convolutional features are input into the weight coding layer for encoding to obtain the encoded features, including: The convolutional features are subjected to batch normalization, a first activation operation, a maximum pooling operation, a fully connected layer connection, and a second activation operation in sequence through the first branch to obtain the first branch features, wherein the first activation operation and the second activation operation both use the relu activation function; The convolutional features are sequentially subjected to average pooling operations, full connection layer connections, and mapping operations through the second branch to obtain second branch features, wherein the mapping operation uses a sigmoid function; The first branch feature and the second branch feature are subjected to a dot product operation through a feature vector dot product layer to obtain a coding feature.

6. A switch rail crack detection device based on the fusion of time-frequency domain features, characterized in that: include: An acquisition module, used for acquiring vibration signals of different striking points of the switch rail, wherein the vibration signals include a force signal and an acceleration signal; A transformation module, used for performing synchronous compression wavelet transformation on the vibration signal to obtain a synchronous compression wavelet transformation result; A calculation module, used for calculating the frequency response function of the vibration signal; A building module for building a multi-modal vibration signal fusion model; A training module, used for training the multimodal vibration signal fusion model based on the acceleration signal, the frequency response function and the synchronous compression wavelet transform result to obtain a crack damage identification model; The prediction module is used to detect cracks in actual switch point rails through the crack damage identification model to obtain damage results of actual switch point rails.

7. The switch point rail crack detection device based on fusion of time-frequency domain features according to claim 6 is characterized in that: The acquisition module comprises: A first setting unit is used to arrange sensors at different positions of the switch rail; A second setting unit is used to set a plurality of striking points, and strike each striking point multiple times by a force hammer, wherein the plurality of striking points are located at different positions of the switch point rail; The first acquisition unit acquires acceleration signals and force signals of the multiple sensors when each tapping point is tapped.

8. The switch rail crack detection device based on the fusion of time-frequency domain features according to claim 6 is characterized in that: The transformation module comprises: A first transformation unit is used to perform continuous wavelet transformation on the vibration signal to obtain wavelet coefficients at different scales and different times; a first calculation unit, configured to calculate a phase derivative of the wavelet coefficient, and obtain an instantaneous frequency of the vibration signal based on the phase derivative; A reallocation unit, configured to compress and reallocate the wavelet coefficients based on the wavelet coefficients and the instantaneous frequency to obtain a compressed time-frequency diagram; The second transformation unit is used to perform inverse transformation on the compressed time-frequency diagram to obtain a synchronous compression wavelet transformation result.

9. The switch point rail crack detection device based on fusion of time-frequency domain features according to claim 6 is characterized in that: The prediction module comprises: A convolution unit is used to convolve the actual acceleration signal and the actual frequency response function through a convolution kernel to obtain a convolution feature; An encoding unit, used for encoding the convolutional features through a weight coding layer to obtain encoded features, wherein the weight coding layer includes a first branch, a second branch and a feature vector dot product layer; A third transformation unit is used to perform vector transformation on the feature code through a vector flattening layer to obtain a transformation feature, and input the transformation feature into a long short-term memory network to obtain a first output feature; A feature extraction unit, used for extracting features from the actual synchronous compression wavelet transform result through a feature extraction layer to obtain a second output feature; A feature fusion unit, used for fusing the first output feature and the second output feature through a feature vector fusion layer to obtain a fused feature; The prediction unit is used to input the fused features into the fully connected layer and the classifier in sequence to obtain the damage results of the actual switch rail.

10. The switch rail crack detection device based on the fusion of time-frequency domain features according to claim 9 is characterized in that: The encoding unit comprises: A first encoding subunit is used to perform batch normalization, a first activation operation, a maximum pooling operation, a fully connected layer connection, and a second activation operation on the convolutional features in sequence through the first branch to obtain a first branch feature, wherein the first activation operation and the second activation operation both use a relu activation function; The second encoding subunit is used to perform an average pooling operation, a fully connected layer connection and a mapping operation on the convolutional features in sequence through the second branch to obtain second branch features, wherein the mapping operation adopts a sigmoid function; The feature vector dot product unit is used to perform a dot product operation on the first branch feature and the second branch feature through a feature vector dot product layer to obtain a coding feature.

Images