A rail grinding control method

Abstract

The application discloses a rail polishing control method, characterized in that the method comprises the following two stages: a starting polishing stage, in which a polishing amount is determined according to a rail corrugation amount, standard polishing parameters are obtained according to the polishing amount, and the rail is polished by using the standard polishing parameters; and a stable polishing stage, in which a rail corrugation ratio before and after polishing is obtained, the standard polishing parameters are corrected according to the corrugation ratio, optimized polishing parameters are obtained, and the rail is polished by using the optimized polishing parameters. The rail polishing control method disclosed by the application can automatically adjust polishing parameters, and has the advantages of high polishing precision, good polishing effect, strong polishing stability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to a rail grinding control method, belonging to the technical field of track maintenance. BACKGROUND

[0002] Rail grinding is an effective means of preventing and repairing rail diseases. Through rail grinding, rail surface damage can be repaired or reduced, contact fatigue and wave abrasion diseases can be prevented, wheel-rail matching relationship can be effectively improved, train operation quality can be improved, and rail service life can be extended.

[0003] In the repair grinding, the existing grinding method is to measure the rail wave abrasion amount, then calculate the grinding amount according to the rail wave abrasion amount, and then control the grinder to grind according to the grinding amount.

[0004] Due to the stability of the grinder itself and the wear of the consumables during grinding, the grinder is difficult to accurately achieve the predetermined grinding amount. After grinding is completed, rework, secondary grinding and even over-grinding at individual positions may occur.

[0005] In the face of this situation, the method of grinding and detecting at the same time is generally adopted. The grinding effect is detected in time after a period of grinding, and the grinding angle and grinding pressure of the grinder are adjusted according to the grinding effect. However, this method still has a lag, and the adjustment of the grinder completely depends on the experience of the operator, and the precision is poor.

[0006] In addition, the maintenance time of the rail is less. If the grinding effect is poor, secondary grinding is needed, which will seriously affect the normal use of the railway.

[0007] Due to the above reasons, it is necessary to propose a rail grinding control method to accurately adjust the grinding angle and pressure of the grinder in real time, thereby improving the accuracy of grinding. SUMMARY

[0008] In order to overcome the above problems, the present application has been designed. The rail grinding control method includes the following two stages:

[0009] In the starting grinding stage, the grinding amount is determined according to the rail wave abrasion amount, the standard grinding parameters are obtained according to the grinding amount, and the rail is ground by using the standard grinding parameters;

[0010] In the stable grinding stage, the rail wave abrasion ratio before and after grinding is obtained, the standard grinding parameters are corrected according to the wave abrasion ratio, the optimized grinding parameters are obtained, and the rail is ground by using the optimized grinding parameters.

[0011] Further, the grinding parameters include the grinding angle and the grinding pressure.

[0012] In a preferred embodiment, the correction is achieved by a correction matrix, and the optimized polishing parameters are represented as:

[0013] M = P x L

[0014] wherein P represents the standard polishing parameters, L represents the correction matrix, and M represents the optimized polishing parameters.

[0015] In a preferred embodiment, the correction matrix is obtained by a spatio-temporal neural network and a BP neural network, the spatio-temporal neural network is used to predict the wave abrasion ratio after polishing, and the BP neural network obtains the correction matrix according to the predicted wave abrasion ratio.

[0016] In a preferred embodiment, the spatio-temporal neural network comprises a feature embedding module, a spatio-temporal graph module, and a prediction module, the input of the spatio-temporal neural network is the historical wave abrasion ratio, and the output is the predicted wave abrasion ratio.

[0017] In a preferred embodiment, the predicted wave abrasion ratio obtaining comprises the following steps:

[0018] inputting the historical wave abrasion ratio into the feature embedding module to obtain a graph state and a graph signal;

[0019] inputting the graph state and the graph signal into the spatio-temporal graph module, updating the graph state and the graph signal by the spatio-temporal graph module,

[0020] inputting the updated graph state and the updated graph signal into the prediction module, and converting the updated graph state and the updated graph signal into a wave abrasion ratio output by the prediction module, wherein the wave abrasion ratio of the output is the predicted wave abrasion ratio after polishing.

[0021] In a preferred embodiment, the feature embedding module comprises two fully connected layers, the input of the two fully connected layers is the historical wave abrasion ratio, the output value of one fully connected layer is converted into a graph state by a matrix transformation function, and the output value of the other fully connected layer is connected to a time sequence embedding matrix to serve as a graph signal.

[0022] In a preferred embodiment, the spatio-temporal graph module has a plurality of spatio-temporal graph modules, the plurality of spatio-temporal graph modules are connected in sequence, the next spatio-temporal graph module receives the updated graph state and the updated graph signal output by the previous spatio-temporal graph module, and the spatio-temporal graph module performs spatio-temporal feature capture and updating again and then outputs.

[0023] In a preferred embodiment, the input data of the BP neural network is the wave abrasion ratio and the standard polishing parameters, and the output data is the correction matrix.

[0024] On the other hand, the present application also provides an electronic device comprising:

[0025] at least one processor; and

[0026] A memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0027] The present application has the beneficial effects including:

[0028] (1) The polishing parameters can be automatically adjusted to improve the polishing effect;

[0029] (2) High polishing precision and good polishing effect;

[0030] (3) Strong polishing stability, no need for secondary rework. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A flowchart of a steel rail polishing control method according to a preferred embodiment of the present application is shown;

[0032] Figure 2 A schematic diagram of a corrugation measurement position in a steel rail polishing control method according to a preferred embodiment of the present application is shown;

[0033] Figure 3 A schematic diagram of a feature embedding module structure in a steel rail polishing control method according to a preferred embodiment of the present application is shown;

[0034] Figure 4 A schematic diagram of a space-time diagram module structure in a steel rail polishing control method according to a preferred embodiment of the present application is shown;

[0035] Figure 5 A schematic diagram of a prediction module structure in a steel rail polishing control method according to a preferred embodiment of the present application is shown. DETAILED DESCRIPTION

[0036] The present application will be further described in detail by the accompanying drawings and examples. Through these descriptions, the features and advantages of the present application will become more apparent.

[0037] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Unless specifically stated otherwise, the drawings are not drawn to scale and are shown for purposes of explanation only.

[0038] A steel rail polishing control method according to the present application is provided, as shown in Figure 1 includes the following two stages:

[0039] The grinding process is started, the grinding amount is determined according to the rail corrugation amount, the standard grinding parameters are obtained according to the grinding amount, and the rail is ground by using the standard grinding parameters;

[0040] The grinding process is started, the grinding amount is determined according to the rail corrugation amount, the standard grinding parameters are obtained according to the grinding amount, and the rail is ground by using the standard grinding parameters;

[0041] Preferably, the length of the rail ground in the starting grinding stage is not more than 1000 meters.

[0042] Preferably, the corrugation amount is obtained by measuring different positions of the rail section, such as Figure 2 As shown, the detection results of different positions are recorded in the form of a matrix.

[0043] Further, in the present application, the specific method of determining the grinding amount according to the rail corrugation amount in the starting grinding stage is not limited, and those skilled in the art can determine it according to experience.

[0044] In a preferred embodiment, the grinding parameters include the grinding angle and the grinding pressure.

[0045] Further, in the starting grinding stage, the method of obtaining the standard grinding parameters from the grinding amount can be obtained from the factory instructions of the grinder.

[0046] The grinding effect is not only related to the corrugation amount before grinding, but also related to the material quality of the rail, the state of the grinder, the grinding parameters, and the wear degree of the grinder consumables. In the present application, how to obtain suitable grinding parameters according to the corrugation amount before grinding to achieve excellent grinding effect is the difficulty of the present application.

[0047] In the stable grinding stage, the corrugation ratio refers to the ratio of the corrugation value after grinding to the corrugation value before grinding. Preferably, the corrugation ratio is the average corrugation ratio of a certain length of rail, and the length of the rail is preferably 30-300 meters, and more preferably 60-100 meters.

[0048] Preferably, when the corrugation amount is a matrix obtained by detecting different positions of the rail section, the corrugation ratio is also in the form of a matrix, and the elements in the matrix are the ratios of the corrugation values before and after grinding at the same position.

[0049] According to the present application, the correction is realized by a correction matrix, and the optimized grinding parameters can be represented as:

[0050] M=P×L

[0051] Wherein, P represents the standard grinding parameters, L represents the correction matrix, and M represents the optimized grinding parameters.

[0052] Further, in the present application, the correction matrix is obtained by a space-time neural network and a BP neural network, the space-time neural network is used to predict the rail corrugation ratio after grinding, and the BP neural network obtains the correction matrix according to the predicted rail corrugation ratio.

[0053] The inventors find that the change of the rail corrugation ratio has a strong time sequence change rule, which is jointly affected by the material difference of the continuous rail section, the gradual wear of the grinding consumables and the fluctuation of the running state of the grinding machine, and forms a complex local space-time correlation. With the wear of the grinding consumables during the use of the grinding machine, the traditional neural network has less consideration of the time sequence, and it is difficult to achieve efficient prediction of the rail corrugation ratio. In the present application, the space-time neural network is used to effectively capture the complex local space-time correlation, and then obtain a more accurate prediction result.

[0054] According to the present application, the space-time neural network comprises a feature embedding module, a space-time graph module and a prediction module, the input of the space-time neural network is a historical rail corrugation ratio, and the output is a predicted rail corrugation ratio.

[0055] The historical rail corrugation ratio refers to the rail corrugation ratio of a section of rail or several sections of rail just after grinding, and the predicted rail corrugation ratio refers to the rail corrugation ratio of a section of rail to be ground after grinding.

[0056] Specifically, the predicted rail corrugation ratio comprises the following steps:

[0057] The historical rail corrugation ratio is input into the feature embedding module to obtain a graph state and a graph signal;

[0058] The graph state and the graph signal are input into the space-time graph module, and the graph state and the graph signal are updated by the space-time graph module,

[0059] The updated graph state and the graph signal are input into the prediction module, and the updated graph state and the graph signal are converted into a rail corrugation ratio output by the prediction module. The rail corrugation ratio of the output is the predicted rail corrugation ratio after grinding.

[0060] Further, the space-time graph module is a neural network for predicting a space-time neural network sequence of a future time length T' using a known space-time neural network sequence of a time length T,

[0061] According to the present application, the known time length T is the ratio of the length of a section of rail to the advancing speed of the grinding machine.

[0062] Further, in the present application, the known space-time neural network sequence of the time length T is represented as

[0063] The space-time neural network sequence of the future time length T' is represented as: where G=(V, E, A) represents the spatial network, V is the set of nodes in the network, E is the set of edges in the network, and A is the adjacency matrix of the network.

[0064] Preferably, the feature embedding module comprises two fully connected layers, and the input of the two fully connected layers is the historical wave-ratio, and the fully connected layers are configured so that the spatio-temporal graph module can accept input data with multiple node features, and the structure is as shown in Figure 3 .

[0065] In the present application, the graph state at time t is represented as The graph-level feature for describing the spatial network G;

[0066] The graph signal at time t is represented as The node-level feature for describing the spatial network G.

[0067] Further, in the feature embedding module, the output value of one fully connected layer is converted into a graph state through a matrix transformation function, and preferably, the matrix transformation function is a reshape function.

[0068] Further, the generation of the graph state can be represented as:

[0069] X state = σ(X in W state ) (1)

[0070] where X state represents the graph state, σ() represents an activation function, preferably a ReLU activation function, X in represents the input historical wave-ratio, and the matrix dimension is T*N*C in , T represents the time step length of the input historical wave-ratio, N is the number of nodes in the graph, and C in represents the input feature dimension;

[0071] W state is a trainable state parameter matrix,

[0072] In the feature embedding module, the output value of the other fully connected layer is connected to the time sequence embedding matrix as a graph signal.

[0073] The addition of the time sequence embedding matrix enables the prediction process to utilize not only the historical wave-ratio data but also the time sequence information.

[0074] Further, the generation of the graph signal can be represented as:

[0075]

[0076] where X signalgraph signal, whose matrix dimension is T*N*d signal , d signal is a constant, which can be obtained by multiple experiments by those skilled in the art, σ() represents an activation function, preferably a ReLu function, X in represents the input historical polishing ratio, W signal is a trainable signal parameter matrix,

[0077] W T is a trainable time sequence embedding parameter matrix,

[0078] According to the application, the input data of the space-time graph module is a graph state and a graph signal, and the space-time graph module captures and updates the space-time features of the graph state and the graph signal.

[0079] Further, the data input into the space-time graph module is fused in terms of time correlation, spatial correlation and space-time correlation in the space-time graph module, so that the prediction value can take into account the running state of the grinding machine, the wear degree of the grinding consumables, the material of the steel rail and other related influencing factors.

[0080] Specifically, as Figure 4 shown, two linear layers are arranged in the space-time graph module, which are a first linear layer and a second linear layer, the first linear layer linearly transforms the graph state input into the space-time graph module to obtain a Q vector, the second linear layer linearly transforms the graph state input into the space-time module to obtain a K vector, and the multiplication of the Q vector and the K vector can obtain an attention matrix in the time dimension, and the time correlation of the input data is realized through the attention matrix.

[0081] The space-time graph module is provided with a time-by-time graph convolution, and the graph signal input into the space-time graph module obtains a V vector through the time-by-time graph convolution operation, the V vector contains the spatial correlation, and the spatial correlation of the input data is realized.

[0082] Further, in the space-time graph module, the attention matrix and the V vector are multiplied to obtain a space-time correlation matrix X update , so as to realize the correlation fusion of time and space, and further realize the global space-time feature update.

[0083] In a preferred embodiment, before the attention matrix and the V vector are multiplied, the attention matrix is also subjected to regression processing, preferably using a softmax regressor for regression processing.

[0084] Further, the space-time graph module also has a third linear layer and a fourth linear layer, which linearly transform the features of the space-time correlation matrix.

[0085] In the present application, the Q and K vectors are obtained by linearly transforming the graph state, the V vector is obtained by feature dimension increasing and time-by-time graph convolution of the graph signal, so that the time correlation of the data can be obtained when Q and K are matched with each other in the time dimension, and the attention score containing the time correlation can be fully fused with the V vector containing the spatial correlation when the V is weighted and summed, so as to realize the global spatio-temporal feature updating.

[0086] Further, the above process can be represented as an attention model:

[0087]

[0088] More preferably, the attention model has multiple attention heads, and the multiple heads make the model simultaneously pay attention to information in different feature spaces, preferably 8 attention heads, and more preferably, the feature dimension of each attention head is 32.

[0089] Further, the features linearly transformed by the third linear layer are connected with the residual of the graph state, and are output as the updated graph state; the features linearly transformed by the fourth linear layer are connected with the residual of the graph signal, and are output as the updated graph signal.

[0090] The residual connection is a classic structure in machine learning models, which is not particularly described in the present application.

[0091] In a preferred embodiment, the updated graph state and the updated graph signal are also normalized before being output, so as to ensure the stability during model training.

[0092] Further, the updated graph state and the graph signal can be represented as:

[0093]

[0094] where LayerNorm represents normalization, and are trainable parameter matrices, and d v is a constant, which can be freely set by those skilled in the art as needed.

[0095] According to the present application, the time-by-time graph convolution is obtained by improving the time graph convolution. In the traditional time graph convolution network, the same convolution kernel is often used to extract spatial features in different time steps, which ignores the time heterogeneity of spatio-temporal data. In order to overcome this limitation, the time-by-time graph convolution operation is proposed in the present application.

[0096] The output X of the time-wise graph convolution out may be represented as:

[0097]

[0098] wherein TimeWiseGC(X in-sig ) represents the time-wise graph convolution, X in-sig represents the input of the time-wise graph convolution, σ() represents an activation function, preferably a ReLU function, and the operator represents the multiplication of matrices of the same time dimension under the time dimension, W represents the convolution kernel of the time-wise graph convolution, A represents the adjacency matrix of X in-sig , represents the matrix after the adjacency matrix and self-connection, and represents the degree matrix.

[0099] In a preferred embodiment, the spatio-temporal graph module has multiple spatio-temporal graph modules connected in sequence, the next spatio-temporal graph module receives the updated graph state and updated graph signal output by the previous spatio-temporal graph module, and the spatio-temporal feature is captured and output again after being updated. According to the present application, using multiple modules can extract deeper and more complex spatio-temporal features. Too few modules will result in too simple spatio-temporal features or an inability to completely describe the spatio-temporal features of the data, and too many will increase the time complexity of the calculation. Preferably, 1-3 modules are provided, and more preferably, 3 modules are provided.

[0100] In the prediction module, as shown in Figure 5 , the graph state input into the module is processed through a fully connected layer and a softmax activation function in sequence to obtain the time dimension attention, and further, the time dimension attention is multiplied by the graph signal input into the module to obtain the historical graph signal. The historical graph signal passes through a double-layer fully connected network to obtain the predicted roughness ratio.

[0101] In a preferred embodiment, the obtaining of the historical graph signal can be represented as:

[0102]

[0103] wherein X his represents the historical graph signal, σ() represents an activation function, preferably a ReLU function,

[0104] W1∈N*1, W2∈d signal *d his , d his is a constant which can be freely set by those skilled in the art as needed, W1 and W2 are two trainable parameter matrices, and the superscript represents transposition.

[0105] In a preferred embodiment, the double-layer fully connected network through which the historical graph signal output final prediction result passes is a time-by-time double-layer fully connected network, so that the spatio-temporal heterogeneity of the historical graph signal is solved.

[0106] The time-by-time is sequentially performed one time step by one time step.

[0107] Further, the time-by-time double-layer fully connected network comprises a fully connected layer and a time feedforward neural network, and the time-by-time double-layer fully connected network can be expressed as:

[0108]

[0109] wherein y (i) represents the output of the time-by-time double-layer fully connected network, i represents different time steps, i = 1, 2, …, T', T' represents the length of the time step to be predicted, represents the matrix parameter at the i-th time step, d future is a constant value for representing the matrix dimension, is the two trainable matrix parameters at the i-th time step.

[0110] The BP neural network is a commonly used neural network, and the BP neural network is a concept proposed by scientists led by Rumelhart and McClelland in 1986, is a multi-layer feedforward neural network trained according to the error back propagation algorithm, and is one of the most widely used neural networks.

[0111] Further, the input data of the BP neural network is the ratio of the wave and the standard polishing parameter, and the output data is the correction matrix.

[0112] In the present application, the specific structure of the BP neural network is not particularly limited, preferably, the hidden layer of the BP neural network is 3 layers, the number of hidden layer neurons is 4-13, preferably 6, and the excitation function is a tansig function.

[0113] According to the present application, before use, the spatio-temporal neural network and the BP neural network are respectively subjected to sample training.

[0114] Further, when the polisher or the consumable type is changed, the spatio-temporal neural network needs to be retrained.

[0115] Further, the sample training mode of the spatio-temporal neural network is the same as the conventional neural network training mode, which is not described herein.

[0116] When the BP neural network is subjected to sample training, the sample is obtained by the following method:

[0117] S11, divide the steel rail to be ground into continuous groups, each group containing a standard grinding section and an optimized grinding section, and detect the corrugation depth of each group before grinding;

[0118] S12, using standard grinding parameters to grind the standard grinding section in a group of steel rails, detecting the corrugation depth after grinding, obtaining the corrugation ratio under the standard grinding parameters, and taking the corrugation ratio and the standard parameters as input data of a group of data in the sample;

[0119] S13, adjusting the grinding parameters according to experience, grinding the optimized grinding section, so that the corrugation depth of the optimized grinding section after grinding is lower than the expected corrugation depth, dividing the grinding parameters by the standard grinding parameters to obtain a correction matrix, and taking the correction matrix as the output data of the group of data in the sample;

[0120] Repeat steps S12-S13 to obtain a sample with multiple groups of data.

[0121] Preferably, the length of the standard grinding section and the optimized grinding section in each group is the same, preferably 1-3 meters.

[0122] It should be understood that various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present disclosure can be executed in parallel, sequentially or in different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, which is not limited herein.

[0123] On the other hand, the present application also provides an electronic device comprising:

[0124] at least one processor; and

[0125] a memory in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any of the above methods.

[0126] The present application also provides a computer readable storage medium storing computer instructions, wherein the computer instructions are used to make the computer execute any of the above methods.

[0127] Embodiments

[0128] Embodiment 1

[0129] A certain grinding car is used to repair a certain high-speed railway, and the following method is used for rail grinding control, including the following two stages:

[0130] A grinding stage is started, a grinding amount is determined according to a rail corrugation amount, standard grinding parameters are obtained according to the grinding amount, and the rail is ground by using the standard grinding parameters;

[0131] A stable grinding stage is started, a rail corrugation ratio before and after grinding is obtained, the standard grinding parameters are corrected according to the corrugation ratio, optimized grinding parameters are obtained, and the rail is ground by using the optimized grinding parameters.

[0132] In the starting grinding stage, the length of the ground rail is 1000 meters, and the corrugation amount is obtained at three positions on the surface of the rail, which are the central position and the positions close to the two sides of the end of the rail.

[0133] The corrugation ratio is the average corrugation ratio of a 60-meter-long rail.

[0134] The correction is realized by a correction matrix, and the optimized grinding parameters are:

[0135] M = P x L

[0136] The correction matrix is obtained by a space-time neural network and a BP neural network,

[0137] The predicted corrugation ratio after grinding includes the following steps:

[0138] The historical corrugation ratio is input into a feature embedding module to obtain a graph state and a graph signal;

[0139] The graph state and the graph signal are input into a space-time graph module, and the graph state and the graph signal are updated by the space-time graph module,

[0140] The updated graph state and graph signal are input into a prediction module, and the updated graph state and graph signal are converted into a corrugation ratio output by the prediction module, which is the predicted corrugation ratio after grinding.

[0141] In the feature embedding module, the output value of a fully connected layer is converted into a graph state by a matrix transformation function, and the matrix transformation function is a reshape function.

[0142] The generation of the graph state can be represented as:

[0143] X state =σ(X in W state )

[0144] In the feature embedding module, the output value of another fully connected layer is connected to a time sequence embedding matrix to serve as a graph signal, and the generation of the graph signal is:

[0145]

[0146] σ() is a ReLu function.

[0147] In the spatio-temporal graph module, two linear layers are arranged, which are a first linear layer and a second linear layer. The first linear layer performs linear transformation on the graph state input into the spatio-temporal graph module to obtain a Q vector. The second linear layer performs linear transformation on the graph state input into the spatio-temporal module to obtain a K vector. The Q vector and the K vector are multiplied to obtain an attention matrix in the time dimension. The attention matrix is processed by a softmax regressor.

[0148] In the spatio-temporal graph module, a time-wise graph convolution is arranged. The graph signal input into the spatio-temporal graph module is processed by the time-wise graph convolution to obtain a V vector.

[0149] In the spatio-temporal graph module, the attention matrix is multiplied by the V vector to obtain a spatio-temporal correlation matrix X. update .

[0150] In the spatio-temporal graph module, a third linear layer and a fourth linear layer are further arranged. The third linear layer and the fourth linear layer perform linear transformation on the features of the spatio-temporal correlation matrix. The features linearly transformed by the third linear layer are connected with the residual of the graph state and output as an updated graph state. The features linearly transformed by the fourth linear layer are connected with the residual of the graph signal and output as an updated graph signal.

[0151] The updated graph state and the graph signal are:

[0152]

[0153] The output X of the time-wise graph convolution out is:

[0154]

[0155] The spatio-temporal graph module has three.

[0156] In the prediction module, the graph state input into the module is processed by a fully connected layer and a softmax activation function in sequence to obtain attention in the time dimension. Further, the attention in the time dimension is multiplied by the graph signal input into the module to obtain a historical graph signal. The historical graph signal is processed by a double-layer fully connected network to obtain a predicted roughness ratio. The obtaining of the historical graph signal can be represented as:

[0157]

[0158] The time-wise double-layer fully connected network includes a fully connected layer and a time feedforward neural network. The time-wise double-layer fully connected network can be represented as:

[0159]

[0160] The hidden layer of the BP neural network is 3 layers, the number of hidden layer neurons is 6, and the excitation function is tansig function.

[0161] The track co-grinding repair length is 200km, the specified maximum position peak amplitude after grinding is 0.06mm, the peak amplitude greater than 0.05mm accounts for less than 5%, the grinding effect is good, and the use index is completely met, without secondary grinding.

[0162] In the description of the present application, it should be pointed out that the terms "upper", "lower", "inner", "outer", "front", "rear" and the like indicate the orientation or positional relationship in the working state of the present application, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third", "fourth" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0163] In the description of the present application, it should be pointed out that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be direct connection, or indirect connection through intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0164] The above describes the present application in combination with the preferred embodiments, but these embodiments are only exemplary and serve only to illustrate. On this basis, various substitutions and improvements can be made to the present application, which all fall within the scope of protection of the present application.

Claims

1. A rail grinding control method, characterized by, The method comprises the following two stages: A grinding stage is started, a grinding amount is determined according to a rail corrugation amount, a standard grinding parameter is obtained according to the grinding amount, and the rail is ground by using the standard grinding parameter; A stable grinding stage is started, a rail corrugation ratio before and after grinding is obtained, the standard grinding parameter is corrected according to the corrugation ratio, an optimized grinding parameter is obtained, and the rail is ground by using the optimized grinding parameter, The grinding parameter comprises a grinding angle and a grinding pressure, The correction is realized by a correction matrix, and the optimized grinding parameter is represented as: M = P x L Wherein, P represents the standard grinding parameter, L represents the correction matrix, and M represents the optimized grinding parameter, The correction matrix is obtained by a space-time neural network and a BP neural network, the space-time neural network is used for predicting a corrugation ratio after grinding, and the BP neural network obtains the correction matrix according to the predicted corrugation ratio, The space-time neural network comprises a feature embedding module, a space-time graph module and a prediction module, the input of the space-time neural network is a historical corrugation ratio, and the output is a predicted corrugation ratio, The predicted corrugation ratio acquisition comprises the following steps: The historical corrugation ratio is input into the feature embedding module to obtain a graph state and a graph signal; The graph state and the graph signal are input into the space-time graph module, and the graph state and the graph signal are updated by the space-time graph module, The updated graph state and the updated graph signal are input into the prediction module, and the updated graph state and the updated graph signal are converted into a corrugation ratio output by the prediction module, and the corrugation ratio of the output is the predicted corrugation ratio after grinding, Two linear layers are arranged in the space-time graph module, which are a first linear layer and a second linear layer, the first linear layer performs linear transformation on the graph state input into the space-time graph module to obtain a Q vector, the second linear layer performs linear transformation on the graph state input into the space-time module to obtain a K vector, and the Q vector and the K vector are multiplied to obtain a time dimension attention matrix, and the time correlation of the input data is fused through the attention matrix, a time graph convolution is arranged in the space-time graph module, and a V vector is obtained after the graph signal input into the space-time graph module is operated through the time graph convolution, the V vector contains spatial correlation, the attention matrix and the V vector are multiplied to obtain a space-time correlation matrix, The space-time graph module also has a third linear layer and a fourth linear layer, and the space-time correlation matrix is linearly transformed by the third linear layer and the fourth linear layer.

2. The rail grinding control method according to claim 1, wherein The feature embedding module comprises two fully connected layers, the input of the two fully connected layers is the historical corrugation ratio, the output value of one fully connected layer is converted into a graph state by a matrix transformation function, and the output value of the other fully connected layer is connected to a time sequence embedding matrix to serve as a graph signal.

3. The rail grinding control method according to claim 1, wherein The space-time graph module has a plurality of space-time graph modules, the plurality of space-time graph modules are connected in sequence, the next space-time graph module receives the updated graph state and the updated graph signal output by the previous space-time graph module, and the space-time graph module outputs the updated graph state and the updated graph signal after the space-time feature is captured and updated again.

4. The rail grinding control method of claim 1 wherein, The input data of the BP neural network is the corrugation ratio and the standard grinding parameter, and the output data is the correction matrix.

5. An electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-4.

Images