A train dynamics model construction method and device

Abstract

The application discloses a train dynamics model construction method and device, comprising: obtaining a plurality of first running data sets corresponding to a target train, the first running data set comprising a running speed and a basic resistance deceleration of the target train in at least one running cycle corresponding to a target speed in a target speed interval under the condition that the target train enters an idle state; training a dedicated dynamics model for each speed interval based on the first running data set corresponding to the speed interval; constructing a second running data set based on the dedicated dynamics model for each speed interval, the second running data set comprising a dynamics correlation feature between a speed and a basic resistance deceleration predicted by the dedicated dynamics model for each speed interval based on the same sample speed; training a target dynamics model based on the second running data set; and combining the target dynamics model and the dedicated dynamics model for each speed interval to obtain a train dynamics model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of train control, in particular to a train dynamics model construction method and device. BACKGROUND

[0002] In the automatic train driving system, the dynamic correlation between the speed and the basic resistance deceleration is one of the core parameters to ensure the safe and efficient operation of the train, and directly affects the formulation of control strategies such as traction, braking and energy consumption optimization.

[0003] At present, the traditional prediction method of the dynamic correlation between the speed and the basic resistance deceleration mainly relies on statistical regression based on empirical formula or historical data. However, the basic resistance is affected by the nonlinear characteristics of wheel-rail contact, aerodynamic effects and other multivariate coupling, and the traditional prediction method is difficult to accurately depict the dynamic change rule between the speed and the basic resistance deceleration, resulting in deviation between the prediction result and the actual working condition. The deviation will affect the decision-making accuracy of the automatic driving system, and may cause unnecessary acceleration or braking, thereby reducing the operation efficiency, increasing the energy consumption and even threatening the driving safety.

[0004] Therefore, it is urgent to construct a high-precision train dynamics model to break through the limitations of the traditional prediction method and realize high-precision prediction of the dynamic correlation between the speed and the basic resistance deceleration, thereby providing a reliable control basis for the automatic driving system. SUMMARY

[0005] The present application provides a train dynamics model construction method and device, which is mainly aimed at constructing a high-precision train dynamics model to realize high-precision prediction of the dynamic correlation between the speed and the basic resistance deceleration, thereby providing a reliable control basis for the automatic driving system.

[0006] In order to achieve the above purpose, the present application mainly provides the following technical solutions:

[0007] In the first aspect, the present application provides a train dynamics model construction method, and the train dynamics model construction method provided in the embodiment can at least include:

[0008] Obtaining a plurality of first running data sets corresponding to a target train, wherein different first running data sets correspond to different speed intervals, and the first running data set includes the running speed and the basic resistance deceleration of the target train in at least one running cycle corresponding to the target train in the target line entering the idle state under the condition of the target speed in the corresponding speed interval;

[0009] Training a dedicated dynamics model for each speed interval based on the first running data set corresponding to the speed interval to obtain the dedicated dynamics model for the speed interval;

[0010] construct a plurality of second running data sets based on the exclusive dynamic model of each speed interval, wherein the second running data set comprises a dynamic correlation feature between the speed and the basic resistance deceleration predicted by the exclusive dynamic model of each speed interval based on the same sample speed, and each dynamic correlation feature is labeled with a corresponding speed interval identifier;

[0011] training the target dynamic model based on the plurality of second running data sets;

[0012] combining the target dynamic model and the exclusive dynamic model of each speed interval to obtain a train dynamic model.

[0013] In a second aspect, the present application provides a train dynamic model construction device, and the train dynamic model construction device provided in the embodiment can at least include:

[0014] The acquisition module is configured to acquire a plurality of first running data sets corresponding to a target train, wherein different first running data sets correspond to different speed intervals, and the first running data set comprises a running speed and a basic resistance deceleration of the target train in at least one running cycle when the target train enters an idle state at a target speed in a target speed interval.

[0015] The first training module is configured to train an exclusive dynamic model of a corresponding speed interval based on the first running data set corresponding to each speed interval, respectively.

[0016] The construction module is configured to construct a plurality of second running data sets based on the exclusive dynamic model of each speed interval, wherein the second running data set comprises a dynamic correlation feature between the speed and the basic resistance deceleration predicted by the exclusive dynamic model of each speed interval based on the same sample speed, and each dynamic correlation feature is labeled with a corresponding speed interval identifier.

[0017] The second training module is configured to train a target dynamic model based on the plurality of second running data sets.

[0018] The combination module is configured to combine the target dynamic model and the exclusive dynamic model of each speed interval to obtain a train dynamic model.

[0019] In a third aspect, the present application provides a computer readable storage medium, wherein the storage medium comprises a stored program, and when the program runs, the device where the storage medium is located executes the train dynamic model construction method of the first aspect.

[0020] In a fourth aspect, the present application provides an electronic device, comprising: a memory for storing a program; and a processor coupled to the memory and configured to execute the program to perform the train dynamics model construction method of the first aspect.

[0021] In a fifth aspect, the present application provides a computer program product, comprising: computer programs / computer executable instructions for performing the train dynamics model construction method described above.

[0022] The train dynamics model construction method and device provided by the present application, in the case of needing to construct a train dynamics model, acquires a plurality of first running data sets corresponding to a target train, different first running data sets correspond to different speed intervals, and the first running data set includes the running speed and the basic resistance deceleration of the target train in at least one running cycle corresponding to the target train in the case of entering the coasting state at a target speed in the target line in the corresponding speed interval. Then, the exclusive dynamics model of the corresponding speed interval is respectively trained based on the first running data set corresponding to each speed interval, and the plurality of second running data sets are constructed based on the exclusive dynamics model of each speed interval, the second running data set includes the dynamics correlation characteristics between the speed and the basic resistance deceleration predicted and output by the exclusive dynamics model of each speed interval based on the same sample speed, and each dynamics correlation characteristic is labeled with the corresponding speed interval identifier. Finally, the target dynamics model is trained based on the plurality of second running data sets, and the target dynamics model and the exclusive dynamics model of each speed interval are combined to obtain the train dynamics model. As can be seen, the train dynamics model constructed by the above-mentioned scheme of the present embodiment is obtained by combining the exclusive dynamics model of different speed intervals and the target dynamics model, which accurately learns the dynamics correlation relationship between the speed and the basic resistance deceleration in the corresponding speed interval through the exclusive dynamics model, and fuses the corresponding dynamics correlation relationship of each speed interval through the target dynamics model to obtain the dynamics correlation relationship in the full speed domain. Such a train dynamics model not only retains the dynamics correlation relationship between the speed and the basic resistance deceleration exclusive to each speed interval, but also realizes the fusion of the dynamics correlation relationship in the full speed domain through global optimization, so that the train dynamics model has stronger full-speed working condition coverage capability and abnormal robustness, and can more accurately predict the dynamics correlation relationship between the speed and the basic resistance deceleration of the train, thereby providing reliable control basis for the automatic driving system, so that the automatic driving system can more accurately plan the speed curve and the braking strategy.

[0023] The above description is only a summary of the technical solutions of the present application. In order to enable a more thorough understanding of the technical means of the present application, the present application can be implemented according to the content of the specification, and in order to enable the above and other purposes, characteristics and advantages of the present application to be more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0025] Figure 1 A flow chart of a train dynamics model construction method provided by an embodiment of the present application is shown;

[0026] Figure 2 A structural schematic diagram of a train dynamics model construction device provided by an embodiment of the present application is shown;

[0027] Figure 3 A structural schematic diagram of a train dynamics model construction device provided by another embodiment of the present application is shown. DETAILED DESCRIPTION

[0028] The exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0029] The dynamic relationship between the speed of the train and the basic resistance deceleration is reflected in that the basic resistance will change with the change of the speed of the train, thereby causing the dynamic change of the deceleration generated by the basic resistance. Specifically, when the speed of the train increases, the basic resistance it receives will generally increase significantly. According to the principle of dynamics, the increase of resistance directly causes the deceleration of the train in the unpowered state to increase accordingly, i.e. the deceleration process of the train will be more obvious due to the increase of resistance. Conversely, when the speed of the train decreases, the basic resistance decreases, and the deceleration generated thereby also decreases simultaneously, and the deceleration effect tends to be mitigated. This process of changing resistance due to speed change, and further causing dynamic change of deceleration, is a specific manifestation of the dynamic relationship between speed and basic resistance deceleration. This dynamic relationship is crucial in train dynamics analysis and operation control, and directly affects the calculation of braking distance, energy consumption optimization and operation safety.

[0030] In an actual running environment, the traditional prediction method is difficult to accurately depict the dynamic change law between the speed and the basic resistance deceleration of the train, resulting in a deviation between the prediction result and the actual working condition. The deviation will affect the decision accuracy of the automatic driving system, and may cause unnecessary acceleration or braking, thereby reducing the running efficiency, increasing the energy consumption, and even threatening the train safety.

[0031] It is found through research that a train dynamics model with a two-layer architecture can be constructed. Specifically, the first layer of the train dynamics model includes dedicated dynamics models independently trained for different speed intervals, which accurately learn the dynamics correlation between the speed and the basic resistance deceleration in the corresponding speed interval through the dedicated dynamics models. The second layer of the train dynamics model includes a target dynamics model trained based on the dynamics correlation predicted and output by the dedicated dynamics models of the first layer, which fuses the dynamics correlations of the respective speed intervals through the target dynamics model to obtain the dynamics correlation in the full speed domain. The dedicated dynamics model of each speed interval in the first layer is integrated in parallel into the target dynamics model in the second layer, and finally forms the train dynamics model. As can be seen, the train dynamics model not only retains the dynamics correlation between the speed and the basic resistance deceleration specific to each speed interval, but also realizes the fusion of the dynamics correlation in the full speed domain through global optimization. Such a train dynamics model has stronger full-speed working condition coverage capability and abnormal robustness (for example, when the model of a certain speed interval fails, the models of other speed intervals can still support), and can more accurately predict the dynamics correlation between the speed and the basic resistance deceleration of the train, providing a reliable control basis for the automatic driving system, so that the automatic driving system can more accurately plan the speed curve and the braking strategy.

[0032] Based on the above findings, the embodiment specifically provides a technical scheme for constructing a train dynamics model. Specifically, a plurality of first running data sets corresponding to a target train are obtained, wherein different first running data sets correspond to different speed intervals, and the first running data set includes the running speed and the basic resistance deceleration of the target train in at least one running cycle under the condition that the target train enters the idle state at a target speed in the target line in the corresponding speed interval. A dedicated dynamics model of the corresponding speed interval is respectively trained based on the first running data set corresponding to each speed interval. A plurality of second running data sets are constructed based on the dedicated dynamics model of each speed interval, wherein the second running data set includes the dynamics correlation features between the speed and the basic resistance deceleration predicted and output by the dedicated dynamics model of each speed interval based on the same sample speed, and each dynamics correlation feature is labeled with a corresponding speed interval identifier. A target dynamics model is trained based on the plurality of second running data sets. The target dynamics model and the dedicated dynamics model of each speed interval are combined to obtain a train dynamics model.

[0033] The train dynamics model provided by the embodiment can be used to support any type of train running on any line, and the type of the train and the line are not limited by the embodiment and can be selected flexibly based on business needs.

[0034] Based on the above train dynamics model construction technical solution, the embodiment specifically provides a train dynamics model construction method and device, and the train dynamics model construction method and device provided by the embodiment are specifically described below.

[0035] The embodiment of the application provides a train dynamics model construction method, as shown in Figure 1 The dynamics model construction method provided by the embodiment of the train can at least include the following steps 101 to 105.

[0036] 101, obtaining a plurality of first running data sets corresponding to a target train, wherein different first running data sets correspond to different speed intervals, and the first running data set includes the running speed and the basic resistance deceleration of the target train in at least one running cycle corresponding to the target train in the target line entering the idle state at the target speed in the corresponding speed interval.

[0037] In some embodiments, the data for constructing the train dynamics model is the running data of the target train in the target line entering the idle state. Based on this, the train dynamics model construction method provided by the embodiment can further include the step of selecting the target train and the target line, and the implementation method of the step can at least include the following methods A1 to A3.

[0038] Method A1, when it is determined that the train dynamics model for supporting the target type of train running on the target type of line needs to be constructed, the train of the target type is selected as the target train, and the line of the target type is selected as the target line.

[0039] In order to realize the flexibility of model construction, the model construction instruction issuing authority can be opened to the user, so that the user can issue the model construction instruction flexibly based on the model training needs of the user. The model construction instruction is used to specify the train type and the line type. The train type refers to a specific code used to distinguish and identify different types or series of trains. Generally, the performance and basic attributes such as length and weight of the same train type are basically the same. The line type refers to different track categories divided by the railway or rail transit system according to the operation needs, technical standards and functions. Generally, the performance and attributes such as slope and line environment of the same line type are basically the same.

[0040] Upon receiving the model construction instruction, a train model specified by the model construction instruction is determined as a target model, a line type specified by the model construction instruction is determined as a target type, and it is determined that a train dynamics model for supporting the target model train to run on the target type line needs to be constructed, at this time, a train of the target model is selected as a target train and a line of the target type is selected as a target line, so as to obtain running data of the target train entering the coasting state on the target line, and train dynamics model construction is performed.

[0041] Method A2, when it is determined that a train dynamics model for supporting a specified train to run on a specified line needs to be constructed, the specified train is selected as a target train and the specified line is selected as a target line.

[0042] Considering that different trains have individual differences, and different lines also have individual differences, when it is determined that a train dynamics model for supporting a specified train to run on a specified line needs to be constructed, the specified train is selected as a target train and the specified line is selected as a target line, so as to construct a train dynamics model that is adapted to the specified train running on the specified line, so as to more accurately predict the dynamics correlation between the speed and the basic resistance deceleration of the specified train running on the specified line.

[0043] Method A3, for the train dynamics model that has been constructed for supporting a specified train to run on a specified line, the following is performed: judging whether the specified train and the specified line corresponding to the current train dynamics model enter a new wear period; if it is determined that at least one of the specified train and the specified line enters a new wear period, the specified train is selected as a target train and the specified line is selected as a target line.

[0044] Both the train and the line will be worn to varying degrees over time, and wear affects the dynamics correlation of the train running on the line. Based on this, in order to realize that the train dynamics model can be adapted to the wear condition of the train and the line in real time, when it is determined that at least one of the specified train and the specified line corresponding to the train dynamics model enters a new wear period, the specified train is selected as a target train and the specified line is selected as a target line, so as to obtain the latest running data of the target train entering the coasting state on the target line, and the train dynamics model construction is constructed. Here, the construction specifically refers to updating the train dynamics model, so as to realize that the train dynamics model is adapted to the wear condition of the train and the line.

[0045] The above-mentioned methods A1 to A3 can be flexibly selected based on business needs, and the present embodiment does not limit the use of at least one of the above-mentioned methods.

[0046] In some embodiments, the target train and the target line selected by the embodiments can at least exist in the following two cases: one is that the target train and the target line are both simulation objects in the simulation system, so that the construction of the train dynamics model can be completed without occupying real trains and lines in the real world, thereby reducing the construction cost. The other is that the target train is a real train in the real world and the target line is a real line in the real world, and the model is constructed by using the real train and the real line in the real world, which can improve the accuracy of the train dynamics model. The above two cases can be flexibly selected based on business needs, and the embodiments do not limit them.

[0047] In some embodiments, after the target train and the target line are selected, the step of obtaining a plurality of first running data sets corresponding to the target train is performed. It should be noted that different first running data sets correspond to different speed intervals, and the first running data sets are used to support training of corresponding dedicated dynamics models for the speed intervals. The implementation method of the step of obtaining a plurality of first running data sets corresponding to the target train can at least include the following steps 101A to 101C.

[0048] 101A, selecting at least one target speed from each speed interval.

[0049] After the target train and the target line are selected, all speeds that the target train can cover when running on the target line are determined to obtain a full speed interval, and then the full speed interval is split to obtain a plurality of speed intervals. There can be overlapping speeds between the speed intervals or there can be no overlapping speeds, and the specific case can be selected based on business needs. It should be noted that in the case where there is no overlapping speed, each speed interval cannot miss the speed, and each speed interval can cover all speeds involved in the full speed interval. For example, the set speed intervals include: [120km / h - 100km / h), [100km / h - 80km / h), [80km / h - 60km / h), [60km / h - 40km / h].

[0050] After the speed intervals are determined, at least one target speed is selected from each speed interval, so that the target speed is used as a basis for controlling the target train to enter the coasting state on the target line. The implementation method of selecting at least one target speed from each speed interval can at least include the following two methods: one is that for each speed interval, at least one target speed is selected from the current speed interval with a preset step size, and the preset step size can be determined based on business needs. The other is that for each speed interval, a preset number of target speeds are randomly extracted from the current speed interval. The above two methods can be flexibly selected based on business needs, and the embodiments do not limit them.

[0051] 101B, respectively for each target speed: in the case of entering the target train into the coasting state on the target line at the current target speed, collecting the slope value and the running speed of the target train on the road section corresponding to at least one running cycle, and respectively determining the basic resistance deceleration corresponding to the corresponding running cycle based on the slope value and the running speed of the corresponding running cycle.

[0052] Specifically, after selecting the target speed, the following steps 101B1 to 101B2 are respectively performed for each target speed.

[0053] 101B1, in the case of entering the target train into the coasting state on the target line at the current target speed, collecting the slope value and the running speed of the target train on the road section corresponding to at least one running cycle.

[0054] After determining the current target speed, the step of controlling the target train to enter the coasting state on the target line at the current target speed is first performed, and the implementation method of the step includes: on the target line, the target train is accelerated to the current target speed by traction, and then the traction is cut off to make the target train run in the coasting state on the target line.

[0055] In the case of entering the target train into the coasting state on the target line at the current target speed, the slope value and the running speed of the target train on the road section corresponding to at least one running cycle are collected. The slope value is the average slope value of the road section traveled by the target train in the corresponding running cycle, and the running speed is the instantaneous speed at the end time point of the corresponding running cycle. For example, the running cycle is 5 seconds, and the slope value and the running speed are collected every 5 seconds.

[0056] 101B2, respectively for each running cycle, based on the slope value and the running speed corresponding to the corresponding running cycle, respectively determining the basic resistance deceleration corresponding to the corresponding running cycle.

[0057] The implementation method of determining the basic resistance deceleration corresponding to each running cycle can include the following steps one to three for each running cycle.

[0058] Step one, based on the running speed corresponding to the current running cycle, determining the train acceleration of the target train.

[0059] Specifically, the running speed is differentiated with respect to the time length corresponding to the current running cycle, and the differentiation result is determined as the train acceleration of the target train. In this embodiment, it can be represented by the formula "a=dv / dt", wherein dv / dt represents the differentiation of the running speed of the target train on the target line with respect to the time length corresponding to the current running cycle, a represents the train acceleration of the target train on the target line in the current running cycle, and a is the differentiation result.

[0060] Step two, based on the slope value of the section corresponding to the current running period, determine the slope acceleration caused by the slope of the section traveled by the target train in the current running period.

[0061] Specifically, the product between the slope value of the section corresponding to the current running period and the gravitational acceleration is determined as the slope acceleration caused by the slope of the section traveled by the target train in the current running period. In this embodiment, it can be represented by the formula "ω = g x i", where g represents the gravitational acceleration, i represents the slope value of the section corresponding to the current running period (for example, i can be the slope in thousandths), and ω represents the slope acceleration caused by the slope of the section traveled by the target train in the current running period.

[0062] Step three, based on the train acceleration and the slope acceleration, determine the basic resistance deceleration corresponding to the current running period.

[0063] Specifically, after the train acceleration and the slope acceleration are determined, the basic resistance deceleration corresponding to the current running period can be determined by the formula "ω' = -a - ω". In this formula, ω' represents the basic resistance deceleration corresponding to the current running period, a represents the train acceleration of the target train on the target line in the current running period, and ω represents the slope acceleration caused by the slope of the section traveled by the target train in the current running period.

[0064] After performing steps one to three on all target speeds, the basic resistance deceleration corresponding to each running period for all target speeds is obtained.

[0065] 101C, for each speed interval, respectively: aggregate the running speed and the basic resistance deceleration corresponding to each running period of all target speeds in the current speed interval, to form a first running data set corresponding to the current speed interval.

[0066] After the aggregation operation of this step, the first running data set corresponding to each speed interval is obtained, which includes the running speed and the basic resistance deceleration corresponding to each running period of the target train in the case of entering the coasting state on the target line for each target speed in the corresponding speed interval. Based on the first running data set, a dedicated dynamic model for the corresponding speed section can be independently trained.

[0067] 102, based on the first running data set corresponding to each speed interval, a dedicated dynamic model for the corresponding speed interval is trained.

[0068] The main purpose of this step is to train a dedicated dynamic model for each speed interval. The implementation process of training a dedicated dynamic model for each speed interval based on the corresponding first running data set can at least include the following steps 102A to 102B.

[0069] 102A, based on the corresponding first running data set of each speed interval, the dedicated dynamic model of each speed interval is trained in parallel.

[0070] Specifically, for each speed interval, the first running data set is executed respectively: based on the first running data set corresponding to the current speed interval, supervised learning is performed using a preset model, and the dedicated dynamic model of the current speed interval is independently trained. The type of dedicated dynamic model can be flexibly selected based on business requirements, and the present embodiment is not limited thereto. For example, the type of dedicated dynamic model can include but is not limited to a neural network model based on a dense layer.

[0071] For example, based on the first running data set corresponding to the speed interval [120km / h - 100km / h), a neural network model based on a dense layer is used for supervised learning, and a dedicated dynamic model 1 corresponding to the current speed interval is independently trained. The dedicated dynamic model 1 describes the corresponding speed and the full connection network topology relationship based on the resistance deceleration, i.e. the dynamic correlation relationship of the speed interval [120km / h - 100km / h). Based on the first running data set corresponding to the speed interval [100km / h - 80km / h), a neural network model based on a dense layer is used for supervised learning, and a dedicated dynamic model 2 corresponding to the current speed interval is independently trained. The dedicated dynamic model 2 describes the corresponding speed and the full connection network topology relationship based on the resistance deceleration, i.e. the dynamic correlation relationship of the speed interval [100km / h - 80km / h). Based on the first running data set corresponding to the speed interval [80km / h - 60km / h), a neural network model based on a dense layer is used for supervised learning, and a dedicated dynamic model 3 corresponding to the current speed interval is independently trained. The dedicated dynamic model 3 describes the corresponding speed and the full connection network topology relationship based on the resistance deceleration, i.e. the dynamic correlation relationship of the speed interval [80km / h - 60km / h). Based on the first running data set corresponding to the speed interval [60km / h - 40km / h], a neural network model based on a dense layer is used for supervised learning, and a dedicated dynamic model 4 corresponding to the current speed interval is independently trained. The dedicated dynamic model 4 describes the corresponding speed and the full connection network topology relationship based on the resistance deceleration, i.e. the dynamic correlation relationship of the speed interval [60km / h - 40km / h].

[0072] 102B, after each round of parallel iteration training, the following steps 102B1-102B4 are performed.

[0073] 102B1, determine the training index data of the current parallel iteration training of the dedicated dynamic model.

[0074] Each speed interval has its own dedicated dynamic model. After the current parallel iteration training, the step of determining the training index data of the current parallel iteration training of the dedicated dynamic model needs to be performed to determine the training index data as the basis for determining whether the iteration training can be terminated. The implementation method of this step can at least include the following methods B1-B4.

[0075] Method B1, the implementation process of determining the training index data of the current parallel iteration training of the dedicated dynamic model can include: obtaining the weight corresponding to each speed interval, and obtaining the loss value corresponding to each dedicated dynamic model after the current parallel iteration training; weight the loss value of each dedicated dynamic model; the weighted processing result is used as the training index data.

[0076] The speed interval has a corresponding weight, and the weight is used to reflect the importance of the corresponding dedicated dynamic model of the speed interval to the dynamic correlation of the predicted train, and the weights of different speed intervals are different. After obtaining the loss value corresponding to each dedicated dynamic model after the current parallel iteration training, the loss value of each dedicated dynamic model is weighted based on the weight, and the weighted processing result is used as the training index data. This training index data is used to reflect the convergence of each dedicated dynamic model from the collaborative dimension of each dedicated dynamic model.

[0077] Method B2, the implementation process of determining the training index data of the current parallel iteration training of the dedicated dynamic model can include: taking the loss value corresponding to each dedicated dynamic model after the current parallel iteration training as the training index data. This training index data is used to reflect the convergence of each dedicated dynamic model from the loss of each dedicated dynamic model itself.

[0078] Method B3, determining the training index data of the current parallel iteration training of the dedicated dynamic model includes: for each speed interval, respectively: determine the adjacent speed interval of the current speed interval, in the dedicated dynamic model after the current parallel iteration training, determine the similarity between the model parameters of the dedicated dynamic model of the current speed interval and the model parameters of the dedicated dynamic model of the adjacent speed interval; the similarity corresponding to each speed interval is used as the training index data.

[0079] One of the manifestations of model convergence is that the parameter update amplitude becomes smaller and tends to be stable value, so it can be considered that the adjacent speed interval of the exclusive dynamic model will produce similar optimal parameters when converging. Based on this, for each speed interval, the adjacent speed interval of the current speed interval is determined, and in the exclusive dynamic model after the current parallel iteration training, the similarity between the model parameters of the exclusive dynamic model of the current speed interval and the model parameters of the exclusive dynamic model of the adjacent speed interval is determined through a preset similarity algorithm. Subsequently, the similarity corresponding to each speed interval is used as the training index data. The preset similarity algorithm can be flexibly selected based on business needs, and the present embodiment is not limited in this regard.

[0080] For example, the adjacent speed interval of [120km / h - 100km / h) is [100km / h - 80km / h), and the adjacent speed interval of [100km / h - 80km / h) is [80km / h - 60km / h). As can be seen, when determining the similarity interval, only one limit value (such as the upper limit value or the lower limit value) in the speed interval is used as the basis to avoid repeated determination.

[0081] Method B4, determining the training index data of the exclusive dynamic model after the current parallel iteration training, comprises: using each exclusive dynamic model after the current parallel iteration training to respectively output the dynamic correlation characteristics between the speed and the basic resistance deceleration based on the same speed, and taking the difference value between the dynamic correlation characteristics output by any exclusive dynamic model and the dynamic correlation characteristics output by other exclusive dynamic models as the training index data.

[0082] One of the manifestations of model convergence is that the output is accurate prediction result, so if the dynamic correlation characteristics between the speed and the basic resistance deceleration output by each exclusive dynamic model based on the same speed are basically consistent, it can be determined that all exclusive dynamic models converge. Based on this, each exclusive dynamic model after the current parallel iteration training is used to respectively output the dynamic correlation characteristics between the speed and the basic resistance deceleration based on the same speed, and the difference value between the dynamic correlation characteristics output by any exclusive dynamic model (which can be randomly selected from all exclusive dynamic models, can be specified by a user, or can be the exclusive dynamic model closest to convergence in the last iteration training) and the dynamic correlation characteristics output by other exclusive dynamic models is taken as the training index data.

[0083] The above-mentioned methods B1 to B4 can be flexibly selected based on business needs, and the present embodiment is not limited in this regard. When two or more are selected and used in combination, it can be more accurately determined whether the iteration training can be terminated.

[0084] 102B2, determine whether all the specific kinetic models meet the convergence condition based on the training index data. If yes, execute 102B3. If no, execute 102B4.

[0085] The implementation process of determining whether all the specific kinetic models meet the convergence condition based on the training index data can include: determining whether the training index data meets the preset value requirement. If yes, it is determined that the convergence condition is met. If no, it is determined that the convergence condition is not met.

[0086] The process of determining whether the training index data meets the preset value requirement is related to the type of the training index data. Specifically, in the case of the weighted processing result, it is determined whether the weighted processing result meets the value requirement. If yes, it is determined that the convergence condition is met. If no, it is determined that the convergence condition is not met. Specifically, in the case of the loss value corresponding to each specific kinetic model, for the loss value corresponding to each specific kinetic model, the following is performed respectively: determining whether the loss value meets the value requirement of the current specific kinetic model; if the loss values corresponding to all the specific kinetic models meet the value requirement, it is determined that the convergence condition is met; if there is a loss value corresponding to a specific kinetic model that does not meet the value requirement, it is determined that the convergence condition is not met. Specifically, in the case of the similarity, it is determined whether all the similarities meet the value requirement. If yes, it is determined that the convergence condition is met. If no, it is determined that the convergence condition is not met. Specifically, in the case of the difference value, it is determined whether all the difference values meet the value requirement. If yes, it is determined that the convergence condition is met. If no, it is determined that the convergence condition is not met.

[0087] 102B3, terminate the training, and determine the specific kinetic model after the current parallel iteration training as the trained specific kinetic model of the corresponding speed interval.

[0088] In the case where it is determined that the convergence condition is met, it means that all the specific kinetic models have found the optimal solution of the parameters. In order to avoid overfitting, the specific kinetic model after the current parallel iteration training is determined as the trained specific kinetic model of the corresponding speed interval.

[0089] 102B4, adjust the parameters of the specific kinetic model after the current parallel iteration training, and continue the next round of parallel iteration training.

[0090] In the case where it is determined that the convergence condition is not met, it means that the specific kinetic model has not found the optimal solution of the parameters, and needs to be further optimized. Therefore, the parameters of the specific kinetic model after the current parallel iteration training need to be adjusted, and the next round of parallel iteration training is continued to further optimize the model.

[0091] The implementation method for adjusting the parameters of the current parallel iteration trained dedicated dynamic model can at least include the following method C1 to method C2.

[0092] Method C1, adjust the parameters of all dedicated dynamic models, and continue the next round of parallel iteration training. Such global adjustment ensures the synergy between the dedicated dynamic models and achieves the best balance of the dedicated dynamic models on the same task as much as possible.

[0093] Method C2, determine whether there is a converged dedicated dynamic model in the current parallel iteration trained dedicated dynamic model; if there is, fix the parameters of the converged dedicated dynamic model, and adjust the parameters of the non-converged dedicated dynamic model; if there is not, adjust the parameters of all dedicated dynamic models.

[0094] If it is determined that there is a converged dedicated dynamic model, at this time, in order to significantly reduce the calculation overhead and memory occupation, while effectively preventing overfitting, the parameters of the converged dedicated dynamic model are fixed, and the parameters of the non-converged dedicated dynamic model are adjusted. If it is determined that there is no converged dedicated dynamic model, the parameters of all dedicated dynamic models are adjusted to further optimize each dedicated dynamic model in the next iteration training.

[0095] The above method C1 and method C2 can be flexibly selected based on business needs, which is not limited in this embodiment.

[0096] 103、Based on the dedicated dynamic model of each speed interval, a plurality of second running data sets are constructed, wherein the second running data set includes the dynamic correlation characteristics between the speed and the basic resistance deceleration predicted and output by the dedicated dynamic model of each speed interval based on the same sample speed, and each dynamic correlation characteristic is labeled with the corresponding speed interval identifier.

[0097] The specific implementation method for constructing a plurality of second running data sets based on the dedicated dynamic model of each speed interval can at least include the following steps 103A to 103B.

[0098] 103A, select a plurality of sample speeds.

[0099] The selection method of the sample speed can include: one is to determine all the speeds that the target train can cover on the target line, and select the sample speed from all the speeds with a target step. Another is to determine all the speeds that the target train can cover on the target line, and select the sample speed by including the experience speed that can cause the train to enter the idle state in actual operation. The above two methods can be flexibly selected based on business needs, and at least one of them is used, which is not limited in this embodiment.

[0100] 103B, for each sample speed, respectively: taking the current speed as an input of each speed interval's exclusive dynamic model, respectively, to obtain a dynamic correlation feature between the speed and the basic resistance deceleration predicted by the exclusive dynamic model of each speed interval based on the current speed, and aggregating the dynamic correlation features predicted by each exclusive dynamic model into a second running data set corresponding to the current sample speed, and respectively labeling each dynamic correlation feature with a speed interval identifier of the corresponding speed interval.

[0101] After performing the above 103A to 103B on all sample speeds, a plurality of second running data sets are obtained.

[0102] 104, training a target dynamic model based on the plurality of second running data sets.

[0103] Each exclusive dynamic model respectively captures and learns the dynamic correlation between the speed and the basic resistance deceleration in the corresponding speed interval. Based on this, in order to fuse the dynamic correlation between the speed and the basic resistance deceleration of each exclusive dynamic model, and eliminate the defects that there is redundant information between each exclusive dynamic model and each exclusive dynamic model is inconsistent in predicting at some speeds, the second running data set constructed by the above step 103 is used to train the target dynamic model, so as to fuse the corresponding dynamic correlation of each exclusive dynamic model through the target dynamic model, obtain the dynamic correlation in the full speed range, and eliminate the above defects.

[0104] The implementation process of training the target dynamic model based on the plurality of second running data sets can include: iteratively training the target dynamic model based on the plurality of second running data sets, and after each iteration, determining whether the target dynamic model after the current iteration converges, if it converges, determining the target dynamic model after the current iteration as the trained target dynamic model, if it does not converge, adjusting the parameters of the target dynamic model after the current iteration, and continuing the next iteration.

[0105] In the process of iteratively training the target dynamic model based on the plurality of second running data sets, based on the speed interval identifier labeled by each dynamic correlation feature, the weight of each exclusive dynamic model in prediction is captured and learned, and the corresponding dynamic correlation of each exclusive dynamic model is adjusted and fused based on the weight. In addition, the type of the target dynamic model can be flexibly selected based on business needs, which is not limited by the present embodiment. Illustratively, the target dynamic model can include but is not limited to a neural network model based on a fully connected layer.

[0106] The method for determining whether the target dynamic model after the current iteration training is converged can include the following two methods: one is to obtain the loss value of the target dynamic model after the current iteration training, and if the loss value meets the value condition, it is determined to be converged, and if the loss value does not meet the value condition, it is determined to be not converged. The other is to obtain the loss value of the target dynamic model after the current iteration training and the loss value of the continuous specified number of iterations before the current iteration, and determine whether the difference between these loss values is stable within a preset numerical range. If yes, it is determined to be converged, and if no, it is determined to be not converged. The above two methods can be selected based on business needs, and this embodiment does not limit the use of at least one of the above two methods.

[0107] 105. Combining the target dynamic model and the dedicated dynamic model of each speed interval to obtain a train dynamic model.

[0108] Each dedicated dynamic model respectively captures and learns the dynamic correlation between the speed and the basic resistance deceleration in the corresponding speed interval. The target dynamic model fuses the corresponding dynamic correlation of each dedicated dynamic model and captures and learns the dynamic correlation in the full speed domain. Based on this, in order to realize high-precision prediction of the dynamic correlation between the speed and the basic resistance deceleration, the target dynamic model and the dedicated dynamic model of each speed interval are combined to obtain a train dynamic model.

[0109] The specific implementation method of combining the target dynamic model and the dedicated dynamic model of each speed interval to obtain a train dynamic model can include: based on the combination strategy of taking the output of the dedicated dynamic model as the input of the target dynamic model, the dedicated dynamic model of each speed interval is integrated in parallel into the target dynamic model to form a train dynamic model. In this way, when it is necessary to predict the dynamic correlation between the speed and the basic resistance deceleration of the train, the speed of the train is first input into each dedicated dynamic model in the train dynamic model, and then the dynamic correlation characteristics between the speed and the basic resistance deceleration predicted and output by each dedicated dynamic model are provided to the target dynamic model, and then the target dynamic model outputs the dynamic correlation between the speed and the basic resistance deceleration of the train based on the input dynamic correlation characteristics.

[0110] The train dynamics model construction method provided in the embodiments of the present application comprises the following steps: in the case of needing to construct a train dynamics model, a plurality of first running data sets corresponding to a target train are acquired, different first running data sets correspond to different speed intervals, and the first running data sets comprise running speeds and basic resistance decelerations of the target train in at least one running cycle corresponding to the target train in an inert state at a target speed in a target speed interval on a target line. Then, a dedicated dynamics model of each speed interval is respectively trained based on the first running data set corresponding to the speed interval, and a plurality of second running data sets are constructed based on the dedicated dynamics model of each speed interval, the second running data sets comprise dynamics correlation characteristics between speeds and basic resistance decelerations respectively predicted and output by the dedicated dynamics model of each speed interval based on the same sample speed, and each dynamics correlation characteristic is labeled with a corresponding speed interval identifier. Finally, a target dynamics model is trained based on the plurality of second running data sets, and the target dynamics model and the dedicated dynamics model of each speed interval are combined to obtain a train dynamics model. As can be seen, the train dynamics model constructed by the above scheme is obtained by combining the dedicated dynamics model of different speed intervals and the target dynamics model, which accurately learns the dynamics correlation between the speed and the basic resistance deceleration of the corresponding speed interval through the dedicated dynamics model, and fuses the corresponding dynamics correlation of each speed interval through the target dynamics model to obtain the dynamics correlation in the full speed domain. Such a train dynamics model not only retains the dynamics correlation between the speed and the basic resistance deceleration of each speed interval, but also realizes the fusion of the dynamics correlation in the full speed domain through global optimization, so that the train dynamics model has stronger full-speed working condition coverage capability and abnormal robustness, and can more accurately predict the dynamics correlation between the speed and the basic resistance deceleration of the train, thereby providing a reliable control basis for the automatic driving system, so that the automatic driving system can more accurately plan the speed curve and the braking strategy.

[0111] In some embodiments of the present application, the running environment of the train is considered to affect the dynamics correlation between the speed and the basic resistance deceleration of the train. Based on this, in order to more accurately predict the dynamics correlation, the train dynamics model construction method provided in the embodiments of the present application can further comprise the following steps (1) to (2) before performing the above step 101B.

[0112] Step (1), at least one train running environment to which the train dynamics model needs to be applied is determined.

[0113] The train running environment is a running environment that the target train can encounter on the target line, which can be set based on the actual situation of the target line and the target train. The train running environment can be expressed by at least one of the following factors: train load, geographical location of the train, line type, train model, and weather data.

[0114] Step (2), sequentially select a train running environment from the determined at least one train running environment, and for each selected train running environment, control the target train to enter the idle state on the target line with the current selected train running environment at the current target speed, and perform the steps of collecting the gradient value and running speed of the target train on the section corresponding to at least one running period and determining the basic resistance deceleration corresponding to the corresponding running period based on the gradient value and running speed of the section corresponding to each running period, which are involved in 101B1-101B2 included in step 101B.

[0115] After determining the train running environment, sequentially select a train running environment from the determined train running environment, and for each selected train running environment, control the target train to enter the idle state on the target line with the current selected train running environment at the current target speed, and after entering the idle state, perform steps 101B1-101B2 to collect the gradient value and running speed of the target train on the section corresponding to at least one running period under the current running environment, and determine the basic resistance deceleration corresponding to the corresponding running period based on the gradient value and running speed of the section corresponding to each running period, to form a corresponding first running data set.

[0116] It should be noted that if the running environment factor is used to obtain the first running data set, the speed and basic resistance deceleration in the first running data set will have a label of the corresponding train running environment, so that the subsequent dedicated dynamic model can capture and learn the dynamic correlation between the speed and the basic resistance deceleration based on the train running environment, and predict the corresponding dynamic correlation characteristics. In addition, the dedicated dynamic model trained based on this condition can predict the dynamic correlation characteristics based on the train speed alone, or in combination with the train speed and the train running environment, which is not limited by this embodiment.

[0117] In some embodiments of the present application, the running environment of the train is considered to affect the dynamic correlation between the speed and the basic resistance deceleration of the train. Based on this, in order to more accurately predict the dynamic correlation, before the above step 104 trains the target dynamic model based on the plurality of second running data sets, the train dynamic model construction method provided by the present embodiment can further include: for each second running data set, respectively performing: adding corresponding running condition data for each dynamic correlation characteristic in the current second running data set, so that the target dynamic model is adapted to the target train running under the running condition corresponding to the running condition data, and the running condition data includes at least one of the following: gradient value, train load, geographical location of the train, line type, train model, weather data, sample speed.

[0118] Each dynamic correlation feature in the second operation data set adds corresponding operation condition data, so that in the subsequent target dynamic model training process, the target dynamic model can combine the operation condition data to capture, fuse and learn the dynamic correlation relationship between the speed and the basic resistance deceleration, and accurately predict the corresponding dynamic correlation feature. In addition, the target dynamic model trained based on this condition can predict the dynamic correlation feature based on the dynamic correlation feature predicted by the exclusive dynamic model alone, or can predict the dynamic correlation feature in combination with the dynamic correlation feature predicted by the exclusive dynamic model and the train operation condition. This embodiment is not limited.

[0119] In some embodiments of the present application, the exclusive dynamic model and the target dynamic model in the embodiment can be neural network models based on full connection layers. Based on this, the model architecture of the exclusive dynamic model and the target dynamic model can be as follows: the model architecture includes an input layer, a convolution layer, a feature fusion layer, a pooling layer and an output layer. For example, the number of convolution layers can be selected based on business needs, and the embodiment is described with three convolution layers. First, the input layer obtains the input required by the training model. Then, the first convolution layer extracts the dynamic correlation feature, the second convolution layer deepens the dynamic correlation feature, and the third convolution layer represents the deep features of the dynamic correlation feature. Then, the feature fusion layer fuses the dynamic correlation features. Then, the global pooling layer globally pools the dynamic correlation features, and finally the output layer outputs the prediction result.

[0120] In some embodiments of the present application, the train dynamic model constructed by the train dynamic model construction method provided by the embodiment can be applied in the train automatic driving scene to predict the dynamic correlation relationship between the speed and the basic resistance deceleration of the train to be appeared based on a speed of the train, so as to provide a reliable control basis for the automatic driving system through the specific trend of the speed and the basic resistance deceleration in the correlation relationship, thereby enabling the automatic driving system to more accurately plan the speed curve and the braking strategy.

[0121] Further, an embodiment of the present application also provides a train dynamic model construction device, as shown in Figure 2 The train dynamic model construction device provided by the embodiment can at least include:

[0122] The acquisition module 21 is configured to acquire a plurality of first running data sets corresponding to a target train, wherein different first running data sets correspond to different speed intervals, and the first running data set includes a running speed and a basic resistance deceleration of the target train in at least one running cycle in a case where the target train enters an idle state at a target speed in a target speed interval.

[0123] The first training module 22 is configured to respectively train a dedicated dynamic model of each speed interval based on the first running data set corresponding to the speed interval.

[0124] The construction module 23 is configured to construct a plurality of second running data sets based on the dedicated dynamic model of each speed interval, wherein the second running data set includes a dynamic correlation feature between a speed and a basic resistance deceleration predicted and output by the dedicated dynamic model of each speed interval based on the same sample speed, and each dynamic correlation feature is labeled with a corresponding speed interval identifier.

[0125] The second training module 24 is configured to train a target dynamic model based on the plurality of second running data sets.

[0126] The combination module 25 is configured to combine the target dynamic model and the dedicated dynamic model of each speed interval to obtain a train dynamic model.

[0127] The train dynamics model construction device provided in the embodiments of the present application comprises: a first obtaining module configured to obtain a plurality of first running data sets corresponding to a target train in a case where a train dynamics model needs to be constructed, wherein different first running data sets correspond to different speed intervals, and each first running data set comprises running speeds and basic resistance decelerations of the target train in at least one running cycle corresponding to a target speed in a target speed interval in a case where the target train enters an idle state on the target line at the target speed; a first training module configured to respectively train exclusive dynamics models of the speed intervals based on the first running data sets corresponding to the speed intervals; a second obtaining module configured to obtain a plurality of second running data sets based on the exclusive dynamics models of the speed intervals, wherein each second running data set comprises dynamics correlation characteristics between speeds and basic resistance decelerations predicted and output by the exclusive dynamics model of each speed interval based on the same sample speed, and each dynamics correlation characteristic is labeled with a corresponding speed interval identifier; and a second training module configured to train a target dynamics model based on the second running data sets, and combine the target dynamics model and the exclusive dynamics models of the speed intervals to obtain a train dynamics model. As can be seen, the train dynamics model constructed by the above scheme is obtained by combining the exclusive dynamics models of different speed intervals and the target dynamics model. The train dynamics model accurately learns the dynamics correlation between the speeds and the basic resistance decelerations in the corresponding speed interval through the exclusive dynamics model, and fuses the dynamics correlation of each speed interval through the target dynamics model to obtain the dynamics correlation in the full speed domain. Such a train dynamics model not only retains the dynamics correlation between the speeds and the basic resistance decelerations in each speed interval, but also realizes the fusion of the dynamics correlation in the full speed domain through global optimization. Therefore, the train dynamics model has stronger full-speed working condition coverage capability and abnormal robustness, and can more accurately predict the dynamics correlation between the speeds and the basic resistance decelerations of the train, thereby providing reliable control basis for the automatic driving system, so that the automatic driving system can more accurately plan the speed curve and the braking strategy.

[0128] In some embodiments of the present application, as shown in Figure 3 The obtaining module 21 can comprise:

[0129] The selecting unit 211 is configured to select at least one target speed from each speed interval.

[0130] The first determining unit 212 is configured to, for each target speed, perform the following: in a case where the target train is controlled to enter an idle state on the target line at the current target speed, collect the slope values and running speeds of the target train in at least one running cycle corresponding to the target speed, and determine the basic resistance deceleration corresponding to each running cycle based on the slope value and the running speed corresponding to each running cycle.

[0131] The aggregation unit 213 is configured to, for each speed interval, respectively aggregate the running speed corresponding to each running period and the basic resistance deceleration of all target speeds in the current speed interval, to form a first running data set corresponding to the current speed interval.

[0132] In some embodiments of the present application, as shown in Figure 3 The first determining unit 212 can determine the basic resistance deceleration corresponding to each running period based on the running speed corresponding to each running period and the slope value of the track section, and the process can include: for each running period, respectively determining the train acceleration of the target train based on the running speed corresponding to the current running period; determining the slope acceleration caused by the slope of the track section traveled by the target train in the current running period based on the slope value of the track section corresponding to the current running period; and determining the basic resistance deceleration corresponding to the current running period based on the train acceleration and the slope acceleration.

[0133] In some embodiments of the present application, as shown in Figure 3 The first determining unit 212 can also be configured to determine at least one train running environment to which the train dynamics model needs to be applied; sequentially select train running environments from the at least one train running environment, and for each selected train running environment, control the target train to enter an idle state at the target speed on the target line with the current selected train running environment, and perform the steps of collecting the slope value of the track section and the running speed corresponding to at least one running period of the target train, and determining the basic resistance deceleration corresponding to each running period based on the slope value of the track section corresponding to each running period.

[0134] In some embodiments of the present application, as shown in Figure 3 The first training module 22 can include:

[0135] The training unit 221 is configured to train the exclusive dynamics model of each speed interval in parallel based on the first running data set corresponding to each speed interval.

[0136] The second determining unit 222 is configured to determine the training index data of the exclusive dynamics model after each round of parallel iterative training of the training unit 221.

[0137] The determination unit 223 is configured to determine whether all the exclusive kinetic models meet a convergence condition based on the training index data determined by the second determination unit 222. If yes, the termination unit 224 is triggered to terminate the training, and the exclusive kinetic model after the current parallel iteration training is determined as the trained exclusive kinetic model of the corresponding speed interval. If no, the adjustment unit 225 is triggered to adjust the parameters of the exclusive kinetic model after the current parallel iteration training, and the next round of parallel iteration training is continued.

[0138] In some embodiments of the present application, as shown in Figure 3 The implementation process of the second determination unit 222 for determining the training index data of the exclusive kinetic model after the current parallel iteration training can include at least one of the following:

[0139] The weight corresponding to each speed interval is obtained, and the loss value corresponding to each exclusive kinetic model after the current parallel iteration training is obtained. The loss value of each exclusive kinetic model is weighted based on the weight. The weighted processing result is taken as the training index data.

[0140] And / or, for each speed interval, the following is performed respectively: determining the adjacent speed interval of the current speed interval, and determining the similarity between the model parameters of the exclusive kinetic model of the current speed interval and the model parameters of the exclusive kinetic model of the adjacent speed interval among the exclusive kinetic models after the current parallel iteration training. The similarity corresponding to each speed interval is taken as the training index data.

[0141] And / or, each exclusive kinetic model after the current parallel iteration training is used to output the dynamic correlation characteristics between the corresponding speed and the basic resistance deceleration based on the same speed prediction, and the difference value between the dynamic correlation characteristics output by any exclusive kinetic model and the dynamic correlation characteristics output by other exclusive kinetic models is taken as the training index data.

[0142] In some embodiments of the present application, as shown in Figure 3 The determination unit 223 is specifically configured to determine whether the training index data meets the preset value requirement. If yes, it is determined that the convergence condition is met. If no, it is determined that the convergence condition is not met.

[0143] In some embodiments of the present application, as shown in Figure 3 The adjustment unit 225 is specifically configured to determine whether there is a converged exclusive kinetic model in the exclusive kinetic model after the current parallel iteration training. If yes, the parameters of the converged exclusive kinetic model are fixed, and the parameters of the exclusive kinetic model that has not converged are adjusted. If no, the parameters of all the exclusive kinetic models are adjusted.

[0144] In some embodiments of the present application, asFigure 3 As shown in the figure, the train dynamics model construction device provided in the embodiment can further include an adding module 26 configured to, before the second training module 24 trains the target dynamics model based on the plurality of second running data sets, respectively perform, for each of the second running data sets, adding corresponding running condition data to each dynamics-related feature in the current second running data set, so as to adapt the target dynamics model to a target train running under a corresponding running condition of the running condition data, the running condition data including at least one of the following: a gradient value, a train load, a geographical location where the train is located, a line type, a train model, weather data, and a sample speed.

[0145] In some embodiments of the present application, as shown in the figure, Figure 3 As shown in the figure, the combining module 25 is specifically configured to combine the output of the dedicated dynamics model as the input of the target dynamics model, and integrate the dedicated dynamics model of each speed interval into the target dynamics model in parallel, to form the train dynamics model.

[0146] In some embodiments of the present application, as shown in the figure, Figure 3 As shown in the figure, the train dynamics model construction device provided in the embodiment can further include a selecting module 27 configured to, when it is determined that a train dynamics model for supporting a target model train running on a target type line needs to be constructed, select the target model train as the target train, and select the target type line as the target line.

[0147] In some embodiments of the present application, the dedicated dynamics model in the embodiment is a neural network model based on a full connection layer, and the target dynamics model is a neural network model based on a full connection layer.

[0148] In the train dynamics model construction device provided in the embodiment of the present application, the details used in the running process of each functional module can refer to the corresponding details of the train dynamics model construction method described above, and will not be described here.

[0149] Further, one embodiment of the present application further provides a computer readable storage medium, the storage medium includes a stored program, wherein when the program runs, the device where the storage medium is located is controlled to execute the train dynamics model construction method described above.

[0150] Further, one embodiment of the present application further provides an electronic device, the electronic device includes: a memory for storing a program; a processor coupled to the memory for running the program to execute the train dynamics model construction method described above.

[0151] Further, an embodiment of the present application also provides a computer program product, which comprises computer programs / computer executable instructions for implementing the above train dynamics model construction method.

[0152] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0153] It can be understood that the related features in the above method and device can be mutually referred. In addition, the "first", "second" and the like in the above embodiments are used to distinguish the embodiments, and do not represent the advantages and disadvantages of the embodiments.

[0154] It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0155] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with these teachings, with or without accompanying programs. Those skilled in the art will recognize that the required structure for a variety of these systems will appear from the description above. In addition, the present application is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages can be used to implement the teachings of the application as described above, and any references above to specific languages are provided for disclosure of preferred embodiments.

[0156] In addition, the memory can include non-persistent memory in computer readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one memory chip.

[0157] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROM, optical memory, etc.) containing computer usable program code.

[0158] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0159] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0160] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0161] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0162] The memory can include non-persistent memory and / or persistent memory, such as flash memory, read-only memory (ROM), and / or volatile or non-volatile random access memory (RAM), among others. The memory is an example of computer readable media.

[0163] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0164] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusions, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0165] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system or computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0166] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A method for constructing a train dynamics model, characterized in that, The method includes: Multiple first operation datasets corresponding to the target train are obtained, wherein different first operation datasets correspond to different speed ranges, and the first operation datasets include the operating speed and basic resistance deceleration of the target train in at least one operation cycle when the target train enters the coasting state on the target line at the target speed within the corresponding speed range. Based on the first running dataset corresponding to each velocity range, a dedicated dynamic model for each velocity range is trained. Based on the dedicated dynamic model for each speed range, multiple second running datasets are constructed. The second running datasets involve multiple sample velocities and include dynamic correlation features between the velocity and the basic drag deceleration predicted by the dedicated dynamic model for each speed range based on each corresponding sample velocity. Each dynamic correlation feature is labeled with the corresponding speed range identifier. The target dynamics model is trained based on the multiple second running datasets; The target dynamics model and the specific dynamics model for each speed range are combined to obtain the train dynamics model.

2. The method according to claim 1, characterized in that, Obtain multiple first-run datasets corresponding to the target train, including: Select at least one target speed from each speed range; For each target speed, the following steps are performed: while controlling the target train to enter a coasting state on the target line at the current target speed, the slope value and running speed of the target train in at least one operating cycle are collected, and based on the slope value and running speed of the segment corresponding to each operating cycle, the basic resistance deceleration corresponding to the corresponding operating cycle is determined respectively. For each speed range, the following steps are performed: summarize the running speed and basic drag deceleration corresponding to all target speeds in the current speed range in each running cycle to form the first running dataset corresponding to the current speed range.

3. The method according to claim 2, characterized in that, Based on the road segment gradient and operating speed corresponding to each operating cycle, the basic resistance deceleration corresponding to the corresponding operating cycle is determined, including: for each operating cycle, the following steps are performed: based on the operating speed corresponding to the current operating cycle, the train acceleration of the target train is determined; based on the road segment gradient corresponding to the current operating cycle, the gradient acceleration caused by the gradient of the road segment traveled by the target train in the current operating cycle is determined; based on the train acceleration and the gradient acceleration, the basic resistance deceleration corresponding to the current operating cycle is determined. And / or, The method further includes: determining at least one train operating environment that the train dynamics model needs to be applicable to; sequentially selecting train operating environments from the at least one train operating environment, and for each selected train operating environment, controlling the target train to enter a coasting state at the current target speed on the target line with the currently selected train operating environment, and performing the steps of collecting the section gradient value and operating speed of the target train in at least one operating cycle, and determining the basic resistance deceleration corresponding to the corresponding operating cycle based on the section gradient value and operating speed corresponding to each operating cycle.

4. The method according to claim 1, characterized in that, Based on the first running dataset corresponding to each velocity range, a dedicated dynamic model for each velocity range is trained, including: Based on the first running dataset corresponding to each speed range, a dedicated dynamic model for each speed range is trained in parallel. After each round of parallel iterative training, the training index data of the dedicated dynamics model after the current parallel iterative training is determined. Based on the training index data, it is determined whether all dedicated dynamics models meet the convergence condition. If they do, the training is terminated, and the dedicated dynamics model after the current parallel iterative training is determined as the trained dedicated dynamics model for the corresponding velocity range. If they do not meet the condition, the parameters of the dedicated dynamics model after the current parallel iterative training are adjusted, and the next round of parallel iterative training continues.

5. The method according to claim 4, characterized in that, Determine the training metric data of the dedicated dynamics model after the current parallel iterative training, including: obtaining the weights corresponding to each velocity range, and obtaining the loss value corresponding to each dedicated dynamics model after the current parallel iterative training; weighting the loss value of each dedicated dynamics model based on the weights; and using the weighted result as the training metric data. And / or, Determine the training metric data of the dedicated dynamics model after the current parallel iterative training, including: for each velocity interval, perform the following: determine the velocity intervals adjacent to the current velocity interval; in the dedicated dynamics model after the current parallel iterative training, determine the similarity between the model parameters of the dedicated dynamics model of the current velocity interval and the model parameters of the dedicated dynamics models of the adjacent velocity intervals; use the similarity corresponding to each velocity interval as the training metric data. And / or, Determine the training index data of the dedicated dynamics model after the current parallel iterative training, including: using each dedicated dynamics model after the current parallel iterative training to predict and output the corresponding dynamic correlation features between the velocity and the basic drag deceleration based on the same velocity, and using the difference between the dynamic correlation features output by any dedicated dynamics model and the dynamic correlation features output by other dedicated dynamics models as the training index data.

6. The method according to claim 4, characterized in that, Determining whether all dedicated dynamic models meet the convergence condition based on the training index data includes: determining whether the training index data meets the preset value requirements; if it does, the convergence condition is met; if it does not, the convergence condition is not met. And / or, The parameters of the dedicated dynamics model after parallel iterative training are adjusted, including: determining whether there is a converged dedicated dynamics model among the dedicated dynamics models after parallel iterative training; if there is, fixing the parameters of the converged dedicated dynamics model and adjusting the parameters of the non-converged dedicated dynamics model; if there is no converged dedicated dynamics model, adjusting the parameters of all dedicated dynamics models.

7. The method according to any one of claims 1-6, characterized in that, Before training the target dynamics model based on the multiple second running datasets, the method further includes: for each second running dataset, performing the following: adding corresponding running condition data for each dynamic correlation feature in the current second running dataset, so that the target dynamics model is adapted to the target train running under the corresponding running conditions of the running condition data, wherein the running condition data includes at least one of the following: gradient value, train load, train geographical location, line type, train model, weather data, and sample speed; And / or, The target dynamics model and the dedicated dynamics model for each speed range are combined to obtain the train dynamics model, including: a combination strategy based on the output of the dedicated dynamics model as the input of the target dynamics model, which integrates the dedicated dynamics model for each speed range in parallel into the target dynamics model to form the train dynamics model; And / or, The method further includes: when it is determined that a train dynamics model for supporting the operation of a target type train on a target type line needs to be constructed, selecting the target type train as the target train and selecting the target type line as the target line; And / or, The proprietary dynamics model is a neural network model based on fully connected layers, and the target dynamics model is a neural network model based on fully connected layers.

8. A train dynamics model construction device, characterized in that, The device includes: The acquisition module is used to acquire multiple first operation datasets corresponding to the target train, wherein different first operation datasets correspond to different speed ranges, and the first operation datasets include the operating speed and basic resistance deceleration of the target train in at least one operation cycle when the target train enters a coasting state at the target speed within the corresponding speed range on the target line. The first training module is used to train the specific dynamic model for each speed range based on the first running dataset corresponding to each speed range. The construction module is used to construct multiple second running datasets based on the dedicated dynamic model for each speed range. The second running dataset involves multiple sample velocities and includes dynamic correlation features between the velocity and the basic drag deceleration predicted by the dedicated dynamic model for each speed range based on each corresponding sample velocity. Each dynamic correlation feature is labeled with the corresponding speed range identifier. The second training module is used to train the target dynamics model based on the multiple second running datasets; The combination module is used to combine the target dynamics model and the dedicated dynamics model for each speed range to obtain the train dynamics model.

9. A computer-readable storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to execute the train dynamics model construction method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: a memory for storing a program; and a processor coupled to the memory for running the program to execute the train dynamics model construction method according to any one of claims 1 to 7.

Images