A train close tracking adaptive control method and device and computing equipment

Abstract

The application provides a train close tracking adaptive control method and device and a computing device. The method comprises the following steps: acquiring the position and reference speed curve of a front vehicle in a virtual marshalling train at a current time; combining the reference speed curve of the front vehicle at the current time to calculate a reference speed curve of the train; combining the position of the front vehicle at the current time and the position of the train to calculate an adjustment speed, and correcting a network error index by using the adjustment speed; training a radial basis function neural network by using the network error index to obtain a control force of the train; and controlling the train to run according to the control force. The method can realize adaptive control of the train, and further realize close tracking control of the marshalling train, and has the advantages of high real-time performance, high safety, high punctuality and high anti-interference performance.

Description

Technical Field

[0001] This invention relates to the field of train control technology for urban rail transit, and in particular to virtual train formation control methods, devices, and computing equipment. Background Technology

[0002] In urban rail transit operation control systems, headway control is one of the fundamental functions. Currently, the spatial spacing method is a widely used headway control method. In this method, a strict distance is maintained between the preceding and following trains, effectively preventing head-on collisions and rear-end collisions. While ensuring operational safety, shortening headway is an effective way to improve transport capacity. To this end, many subway lines in China have completed the transformation from fixed block signaling to moving block signaling, and high-speed railways on main lines have also chosen to use quasi-moving block signaling instead of fixed block signaling, resulting in a significant improvement in operating efficiency. However, due to the absolute braking distance tracking interval limitation—that is, the train interval should not be less than the braking distance of the following train—the method of shortening headway to improve transport efficiency has reached a bottleneck under the existing block signaling system.

[0003] To break free from the constraints of existing traffic block systems and further shorten train intervals, virtual train formation has gradually become a research hotspot. The concept of virtual train formation originates from highway traffic, where vehicles maintain a safe distance from the vehicle in front, a distance far lower than the braking distance required to come to a complete stop. The driver reacts to the brake lights of the vehicle in front. In virtual train formation, trains use wireless communication to establish connections with adjacent trains within the section, exchanging information such as speed and position. Combining reference signals from ground equipment and train-to-train communication data, trains forming a convoy operate in close tracking, significantly reducing train intervals. This interval is not constrained by absolute braking distances, further improving transportation efficiency.

[0004] The implementation of virtual train formation involves two essential steps: dynamic formation and close tracking. Dynamic formation refers to the process by which trains transition from a moving block operation to a close formation during operation. After dynamic formation is completed, the train enters the close tracking phase. Maintaining a constant ideal distance is a crucial objective of train control during this process. However, external disturbances and differences in train traction and braking performance pose significant challenges to train following control. Therefore, the control method must meet the control objectives of high real-time performance, high safety, high punctuality, and high anti-interference capabilities. Existing control methods struggle to simultaneously address all these challenges. Summary of the Invention

[0005] To address the aforementioned technical deficiencies, embodiments of this application provide a train close tracking adaptive control method, apparatus, and computing device.

[0006] The first aspect of this application provides a train close-track adaptive control method, comprising: acquiring the position and reference speed curve of the preceding vehicle in a virtual train formation at the current moment; calculating the reference speed curve of the current vehicle in combination with the reference speed curve of the preceding vehicle at the current moment; calculating an adjustment speed in combination with the position of the preceding vehicle and the position of the current vehicle, and using the adjustment speed to correct a network error index; training a radial basis function neural network using the network error index to obtain the control force of the current vehicle; and controlling the operation of the current vehicle according to the control force.

[0007] In one possible implementation, calculating the adjustment speed based on the current position of the vehicle ahead and the current position of the vehicle includes: calculating the train interval between the vehicle ahead and the vehicle based on the current position of the vehicle ahead and the current position of the vehicle; and calculating the adjustment speed based on the train interval and a preset error adjustment time.

[0008] In one possible implementation, training the radial basis function neural network using the network error index includes online training and offline training; the online training is: training the radial basis function neural network according to the reference speed curve of the vehicle ahead; the offline training is: performing network training according to a predetermined reference curve before the train runs, or performing simulation training on the onboard computer during operation.

[0009] In one possible implementation, the vehicle in front is the vehicle in the virtual convoy that is in front of and closest to the vehicle.

[0010] In one possible implementation, the method further includes: if it fails to obtain the position and reference speed curve of the vehicle ahead in the virtual formation convoy at the current moment within a number of consecutive communication cycles, then controlling the vehicle to run at a constant speed and broadcasting communication failure information to the vehicles behind.

[0011] In one possible implementation, the method further includes: when the deviation between the actual speed of the vehicle and the reference speed curve of the vehicle reaches a preset threshold, updating the reference speed curve of the vehicle according to the actual speed of the vehicle, and sending the updated reference speed curve to the following vehicle.

[0012] A second aspect of this application also provides a train close-tracking adaptive control device, comprising: a communication module adapted to acquire the position and reference speed curve of the preceding vehicle in a virtual train formation at the current moment; an information preprocessing module adapted to calculate the reference speed curve of the current vehicle in combination with the reference speed curve of the preceding vehicle at the current moment, and to calculate the adjustment speed in combination with the position of the preceding vehicle and the position of the current vehicle, and to correct the network error index using the adjustment speed; a radial basis function neural network module adapted to train the radial basis function neural network using the network error index to obtain the control force of the current vehicle; and a control module adapted to control the operation of the current vehicle according to the control force.

[0013] In one possible implementation, the information preprocessing module calculates the adjustment speed by: calculating the train interval between the preceding vehicle and the current vehicle based on the current position of the preceding vehicle and the current vehicle's position; and calculating the adjustment speed based on the train interval and a preset error adjustment time.

[0014] A third aspect of this application also provides an electronic device, including: at least one processor and a memory storing a computer program; when the computer program is read and executed by the processor, the electronic device performs the above-described train close-tracking adaptive control method.

[0015] The fourth aspect of this application also provides a readable storage medium storing a computer program, which, when read and executed by an electronic device, causes the electronic device to perform the above-described train close-tracking adaptive control method.

[0016] The train close-tracking adaptive control method and device provided in this application embodiment is a model-free adaptive train platoon control method. This method has the following advantages: Based on following a reference speed curve, it ensures a constant and ideal interval between trains by introducing adjustment speed into the network error index. It utilizes an RBF neural network to achieve adaptive control of the train based on the real-time position of the preceding train and the reference speed curve, thereby realizing close-tracking control of the train platoon. This method boasts advantages such as high anti-interference capability, high real-time performance, and high punctuality. Furthermore, the above method also provides a control method for unexpected situations in the train platoon, effectively reducing the risk of accidents and offering high safety. Attached Figure Description

[0017] Figure 1 A flowchart of a train close-track adaptive control method according to an embodiment of the present invention is shown;

[0018] Figure 2 This is a reference speed curve for a virtual train formation crossing from a high-speed zone to a low-speed zone according to an embodiment of the present invention;

[0019] Figure 3 This is a reference speed curve of the preceding vehicle that does not consider the following vehicle, according to an embodiment of the present invention.

[0020] Figure 4 A reference speed curve for a virtual train formation crossing from a low-speed zone to a high-speed zone according to an embodiment of the present invention. Figures 2 to 4 The reference velocity curve has been converted from a velocity-time curve to a velocity-position curve.

[0021] Figure 5 This is a schematic diagram of the train operation under a tight formation according to an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of an RBF neural network structure according to an embodiment of the present invention;

[0023] Figure 7 This is a schematic diagram of an RBF neural network follower control structure according to an embodiment of the present invention;

[0024] Figure 8 This is a schematic diagram of a virtual train speed following control architecture according to an embodiment of the present invention;

[0025] Figure 9 This is a schematic diagram of a speed following control structure based on an RBF neural network according to an embodiment of the present invention;

[0026] Figure 10 This is a schematic diagram of a speed following control structure based on an RBF neural network according to an embodiment of the present invention;

[0027] Figure 11 This is a schematic diagram illustrating the network training effect under different initial network parameters according to an embodiment of the present invention;

[0028] Figure 12 This is a schematic diagram illustrating the principle of adaptive control based on an RBF neural network according to an embodiment of the present invention;

[0029] Figure 13 This is a schematic diagram of the platoon control system framework under a lead vehicle following strategy according to an embodiment of the present invention;

[0030] Figure 14 This is a schematic diagram of the platoon control system framework under a leading vehicle following strategy according to an embodiment of the present invention;

[0031] Figure 15 A schematic diagram of a train close-track adaptive control device according to an embodiment of the present invention is shown;

[0032] Figure 16This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0033] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0034] Figure 1 A flowchart of a train close-track adaptive control method 100 according to an embodiment of the present invention is shown. Figure 1 As shown, method 100 begins with step S110.

[0035] In step S110, the position and reference speed curve of the vehicle in front in the virtual convoy at the current moment are obtained.

[0036] Next, in step S120, the reference speed curve of this vehicle is calculated by combining the reference speed curve of the vehicle ahead at the current moment.

[0037] The principle of virtual train formation control is as follows: After dynamic formation, trains maintain a tight formation. The lead car operates according to the speed curve generated by its onboard equipment (i.e., the lead car's reference speed curve), and following cars adopt an active following strategy to maintain synchronous operation with the lead car. Ideally, the train can obtain the reference speed curve of the lead car through car-to-car communication. After adjustment, the reference speed curve of the lead car can be used as the reference speed curve of the train itself, and speed synchronization can be achieved by operating according to the reference speed.

[0038] The reference speed curve provides guidance for the train's operating speed over a future period and is crucial for ensuring train speed synchronization. During train speed curve planning, the train is not treated as a single point mass. For example, when crossing from a smaller speed-limited area to a higher speed-limited area, the train length factor must be considered. That is, the train is only allowed to accelerate after the entire train has passed through the lower speed-limited area; this principle is to maintain the rear end of the train.

[0039] Virtual train formations have longer car lengths than single trains. Following the principle of maintaining the rear of the train, the running curves of the preceding and following trains need to be adjusted to generate reference speed curves to ensure the entire train formation operates within the speed limit. The following two special operating scenarios for virtual train formations are considered.

[0040] (1) The convoy crosses from a high-speed-limit area to a low-speed-limit area.

[0041] Figure 2This describes the reference speed curves for a virtual convoy crossing from a high-speed zone to a low-speed zone. When transitioning from a high-speed-limit area to a low-speed-limit area, the convoy should reduce speed; the reference speed curve of the lead vehicle (i.e., the lead car in the virtual convoy in the diagram) is the driving curve calculated by its onboard equipment. For example... Figure 2 As shown, the vehicle in front brakes at point S1, at which point the front of the vehicle behind is at point S2. If the vehicle behind follows the speed curve of the vehicle in front exactly, its speed will be greater than that of the vehicle in front, which violates the principle of synchronized operation. Therefore, in this situation, the speed curve of the vehicle in front should be shifted to the left by a distance l. head This serves as a reference speed curve for the following vehicle. Where l head This is the distance between the front and rear locomotives, specifically the sum of the length of the preceding locomotive and the train interval.

[0042] (2) The convoy crosses from a low-speed-limit zone to a high-speed-limit zone.

[0043] Figure 3 This represents the speed curve of the preceding train without considering the following train. When the rear of the train passes through the low-speed-limit area, the train begins to accelerate. At this point, the rear of the following train is at point S2. According to the principle of synchronized operation, the following train should accelerate at this point, but this would cause the following train's speed to exceed the speed limit. If the following train chooses to continue running at a constant speed, its speed will be less than that of the preceding train, which violates the principle of synchronized operation. Therefore, the speed curves of the preceding and following trains are corrected as follows:

[0044] When a virtual train formation transitions from a lower speed limit area to a higher speed limit area, the reference speed curves of the preceding and following trains are as follows: Figure 4 As shown. The following train maintains its rear-end alignment, and the entire train begins to accelerate as it passes through the low-speed zone. The following train is shifted a distance l to the right on the curve. head The curve of the preceding vehicle is obtained, where l head The specific value is the sum of the length of the following vehicle and the ideal train formation interval.

[0045] Next, in step S130, the adjustment speed is calculated by combining the position of the vehicle in front and the position of the vehicle at the current moment, and the network error index is corrected by the adjustment speed.

[0046] In a tight formation, the goal of platoon control is to maintain an ideal formation interval d between the following and preceding vehicles. vcm and at the same speed.

[0047] Figure 5 This describes the operational status of the train formation. At time 1, the positions of the preceding and following trains are as follows: and The speeds of the front and rear trains are respectively and The distance between the two cars is d1. Assume the convoy is in an ideal formation, i.e., d1 = d... vcm and Under a distributed control structure, the on-board equipment of the preceding and following trains each control the operation of their own trains. After time k, the positions of the preceding and following trains respectively reach... and The speeds of the trains before and after reached respectively and According to the control objectives, the train status should satisfy the following relationship:

[0048]

[0049]

[0050] Where time k represents the end time of the train formation. Equations (1) and (2) indicate that during the entire formation process, all trains in the convoy should maintain the same speed and constant spacing. Since the reference speed curve is a way to constrain the position and speed relationship between trains, if the train formation can strictly follow the reference speed curve, as shown in equation (3), the speed synchronization control objective can be achieved.

[0051]

[0052] Where v x (s) represents the actual train operation curve. The above formula shows that the train control objective can also be expressed in another way: during formation operation, the train follows the reference speed curve.

[0053] Regardless of whether equation (1) or equation (3) is used as the control target, the synchronization error between the two vehicles needs to be monitored at all times during the control process. The synchronization error is a parameter used to measure the synchronization effect, including the speed error Δv. x and spacing error Δs x The closer the synchronization error is to zero, the better the synchronization effect. The maximum allowable speed error is set at 2 km / h, or 0.56 m / s; the maximum allowable spacing error is 5 meters. The specific calculation method for synchronization error is as follows:

[0054]

[0055]

[0056] To achieve the aforementioned control objectives, a common control approach is to abstract the train motion process into a mathematical model. This means that, given the current speed, position, and control force, the train's state can be predicted using mathematical expressions. By optimizing the control strategy, the train can achieve the aforementioned objectives within the prediction time domain. This method requires a high degree of accuracy in establishing the train motion model table. However, external disturbances and differences in train traction and braking performance during actual operation are difficult to quantify using standardized mathematical formulas. Furthermore, computational efficiency decreases as the prediction time increases. These factors pose significant obstacles to applying this method to tracking control.

[0057] Radial basis function (RBF) neural network control is a model-free adaptive control method based on neural networks, and it is an effective means to achieve the above-mentioned vehicle control objectives.

[0058] Neural networks possess advantages such as highly parallel structure, powerful learning ability, and the ability to approximate nonlinear functions, which has promoted the application of neural network technology in the field of nonlinear control. Currently, the most widely used feedforward network based on backpropagation suffers from slow convergence speed, hindering its application in online control. In contrast, RBF neural networks exhibit better convergence ability than feedforward networks, and also have a simpler network structure and lower computational cost.

[0059] RBF neural networks have a three-layer structure: input layer, hidden layer, and output layer. The topology is as follows: Figure 6 As shown in the figure. In the figure, x = [x i ] T The input layer contains n neurons; h = [h j ] T h is the output of the hidden layer of the network. j The output of the j-th neuron in the hidden layer is given. The activation function of the neurons in the hidden layer is composed of radial basis functions. The operational units formed by the hidden layer are called hidden layer neurons. The number of hidden layer neurons is m. The output of the hidden layer is selected from the Gaussian function, a non-linear activation function.

[0060]

[0061] Where ||·|| is the Euclidean norm, representing the distance between two vectors, and c is an n×m vector matrix. j This represents the coordinate vector of the center point of the Gaussian function of the j-th neuron in the hidden layer, which has the same dimension as the input vector x, and b = [b1, ..., bj]. m ] T The width of the Gaussian function for the hidden layer, w = [w1, ..., w m ] T It is the weight matrix from the hidden layer to the output layer, w jThis represents the weights from the j-th hidden layer neuron to the output layer. The output layer has one neuron, and the network output is:

[0062] y(t)=w T h = w1h1 + w2h2 + ... + w m h m (Equation 7)

[0063] In the above formula, t represents time. Assume the output of the ideal reference model is y. m After the neural network is built, training the network parameters w can approximate the model. Gradient descent is used to train the network parameters. Gradient descent is a supervised training method and also the most common training method. It uses iterative calculations to adjust the network parameters to minimize the training error function. The training error index function is E = (yy) / (yt). m ) 2 / 2, this metric represents the error between the network output (i.e., the model's actual output) and the reference output (i.e., the ideal output). The training objective is to reduce this error value. The parameters are trained as follows:

[0064] Δw j (t)=σ(y(t)-y m (t))h j (Equation 8)

[0065] w j (t)=w j (t-1)+Δw j (t)+α(w j (t-1)-w j (t-2)) (Equation 9)

[0066] In the above formula, Δw j (t) represents the weight change, σ is the learning rate, and α is the momentum factor. By repeatedly correcting the network parameters using the above formula, a network that meets the error requirements can be obtained. The same method can be used to train network parameters c and b; however, training multiple network parameters simultaneously will increase computation time and reduce convergence speed.

[0067] Based on the principles of neural networks, this can be applied to trajectory following control processes. Consider a single-input, single-output controlled object: y(k+1) = g[y(k), u(k)]. Here, u(k) is the input of the controlled object, and y(k) is the output of the controlled object. Let y m (k) represents the ideal following output of the system. In actual engineering, g[·] is often unknown. In this case, it is difficult to calculate the control input. However, by reducing the following error through iterative training of the RBF neural network, following control can be achieved under the premise that the control system model is unknown. Figure 7This describes the framework of an RBF neural network follower control system, where the output u of the neural network is the control input to the controlled object. The following error and network parameter training error indices are designed as follows:

[0068] e(k) = |y m (k)-y(k)|(Equation 10)

[0069]

[0070] The output of the neural network is the control input of the controlled object:

[0071] u(k)=w T h = w1h1 + w2h2 + ... + w m h m (Equation 12)

[0072] Where m is the number of nodes in the hidden layer, m is the weight vector of the neural network, and h is the output vector of the Gaussian function. In the RBF network, x = [x i ] T Given network input, the Gaussian function output is:

[0073]

[0074] Where, the center vector c = [c 11 …c 1m The width of the Gaussian function is b = [b1, ..., b]. m ] T The training mechanism for the network based on gradient descent is as follows:

[0075]

[0076] w j (t)=w j (t-1)+Δw j (t)+α(w j (t-1)-w j (t-2)) (Equation 15)

[0077] Since the following train calculates its own reference speed curve based on the reference speed curve of the preceding train, and obtains the status information of the preceding train in real time through vehicle-to-vehicle communication during operation, a train speed following control method based on RBF neural network is adopted. This method aims to follow the reference speed curve and calculates the optimal instantaneous traction / braking force for the train at the current moment. The virtual train speed following control architecture used is as follows: Figure 8 As shown.

[0078] The onboard equipment modules involved in train control mainly consist of vehicle-to-vehicle communication, information preprocessing, and an RBF neural network controller. The vehicle-to-vehicle communication module is responsible for acquiring information from the preceding vehicle, communicating every 0.1 seconds to ensure precise train interval control. The information acquired through vehicle-to-vehicle communication includes the preceding vehicle's speed curve and position information. The speed curve data is relatively large, so it is acquired every minute, while position information needs to be acquired in each communication cycle. Information preprocessing is the process of processing communication data and the train's own status data, based on the preceding vehicle's position s. p and the position of this car s f The real-time interval d between the two trains is calculated, and based on the principle of generating the train reference speed curve, combined with the mathematical relationship between speed, distance, and travel time, a reference speed curve is calculated for this train. During train formation operation, discrete time values ​​are represented by integer symbols k. The RBF neural network controller calculates the optimal control force u based on the current train-to-train interval, reference speed curve, and the speed of the following train. * The controller is designed as follows.

[0079] The train is considered as a single-input, single-output system, where the input is the train's control force u and the output is the train's speed v. f The system satisfies the relationship: v f (k+1)=g[v f [(k),u(k)], this formula represents an abstract model of train motion, where the train speed and input control force exhibit a nonlinear relationship. (Reference speed curve for this train) As an ideal following instruction, taking equation (3) as the target of train convoy control, under the condition that g[·] is known, the train control force u can be calculated so that the train running state satisfies the following relationship: In the above formula, g[·] is related to the train's traction / braking performance, track conditions, and train speed. Since the model cannot completely replicate the actual train operation process, and it is difficult to describe it using a unified mathematical expression under different environments, this application's embodiment uses the RBF neural network control method to achieve train reference speed curve following. The design of a virtual train following control system based on the RBF neural network is as follows: Figure 9 .

[0080] exist Figure 9 In this context, the train is treated as a black box system, meaning the control force u... * After inputting the data into the train, only the train's actual speed v at the next moment can be observed. f The speed change process cannot be described by a mathematical model; the actual speed is obtained through train speed sensors. The neural network built through iterative calculation and training in the RBF network module works by adjusting network parameters to reduce error metrics. The following mechanism is based on the train's actual speed v. f and reference speed The difference provides a parameter adjustment strategy for the neural network. (In the reference velocity curve) Given the given information, the challenge in calculating the optimal control input lies in designing and training the neural network to achieve high-quality following control.

[0081] according to Figure 9 The network's input layer consists of three neurons, representing the train's actual speed, reference speed, and the difference between the two speeds, respectively. The number of neurons in the hidden layer affects the network's convergence speed and control performance; the number of neurons needs to be determined through continuous testing in actual simulation experiments, and is generally chosen to be around ten. The output layer consists of a single neuron, representing the reference control force u. * .

[0082] Design the Gaussian function as follows: in, c is the center vector, b is the basis function width vector, and w is the neural network weight vector. Under this network structure, the network output is:

[0083] u * (k)=w T h (Equation 16)

[0084] Network training can be performed on multiple network parameters, such as the center vector c, the width vector b, and the weight vector w. Multi-parameter training can improve control accuracy, but computational efficiency will relatively decrease. For close-following control of virtual train formations, simulation experiments demonstrate that, with appropriate initial parameter values, training only two parameters, w and c, can achieve good speed following while maintaining high computational efficiency. Therefore, the selection of initial parameter values ​​is crucial.

[0085] Parameter b j and c ij The design principles for initial values ​​are as follows: (1)b j The larger the value, the wider the Gaussian function. This value is an important parameter affecting the network's mapping range. The larger the value, the stronger the network's mapping ability to the input; otherwise, the weaker the network's mapping ability to the input. Generally, this value is taken as a moderate value. (2) Center vector coordinate value c ij The closer the basis function is to the input value, the more sensitive it is to the input. The design of this value should ensure that the basis function is within the effective input mapping range, generally taking a value near the range of the input value.

[0086] The goal of network training is to make the actual train speed approximate the reference speed, therefore, a following speed error is established: Establish training error metrics: Using gradient descent, the strategy for adjusting neural network parameters is designed as follows:

[0087]

[0088] w j (k)=w j (k-1)+Δw j (k)+α(w j (k-1)-w j (k-2)) (Equation 18)

[0089]

[0090] c ij (k)=c ij (k-1)+Δc ij (k)+α(c ij (k-1)-c ij (k-2)) (Equation 20)

[0091] By continuously training the neural network using the above parameter adjustment strategy, it is theoretically possible to make the network error index converge to zero. This indicates that the actual train speed is equal to the reference speed. However, even after multiple iterations of neural network training, a small network approximation error will still exist. Therefore, in real-world scenarios, the actual speed and the reference speed are difficult to guarantee absolute equality, and instead, they have the following relationship:

[0092]

[0093] Δv max This is the maximum difference between the actual speed and the reference speed after network training. Since the reference speed is obtained from the speed curve of the preceding train, the reference speed of the following train is the same as the speed of the preceding train at any given moment. Therefore, the speed difference between the two trains also satisfies the constraint relationship above. Assuming that the distance between the two trains is an ideal train interval at a certain moment, considering the most unfavorable control situation, i.e., the speed difference between the two trains is Δv... max Furthermore, since the speed of the vehicle in front is greater than that of the vehicle behind, after time t, the distance error will reach Δv. max The distance between trains is ×t, and this value gradually increases over time. When the distance between two trains increases to a certain extent, the convoy will enter an unexpected dispersal state. If the speed of the preceding train is less than that of the following train, the distance between trains will gradually decrease. In both of these cases, the convoy operation may produce a large spacing error, which is also an important indicator for measuring the synchronization effect of the preceding and following trains.

[0094] In summary, the above-mentioned control method has the following problem: it cannot guarantee that the train spacing error is controlled within the allowable range. If the train spacing deviates significantly from the ideal train interval, this method cannot control the train to return to the ideal train interval, and there is a risk of rear-end collision. Therefore, this application's embodiments improve the above-mentioned control method, achieving adaptive adjustment of train spacing while ensuring speed following.

[0095] Because of speed errors during the following train's operation, and the cumulative operational errors that occur over time, the two trains cannot maintain an ideal train interval. To achieve adaptive adjustment of train interval, this embodiment starts with a reference speed and makes appropriate corrections to the reference speed based on the real-time distance between the two trains. This changes the error index E(k) of the neural network, and a new network parameter adjustment strategy is designed using the corrected error index. This adjustment strategy takes into account the train-to-train spacing factor, thus the trained neural network will have the function of adaptive interval adjustment.

[0096] Figure 10 This is the improved control structure. For example... Figure 10 As shown, in addition to the actual train speed and the reference speed, another important indicator that the following mechanism needs to refer to is the adjustment speed v. ε Speed ​​adjustment is used to adjust the real-time spacing between trains, bringing the train spacing back to the ideal train interval. The value of the speed adjustment, v, is... ε The difference between the real-time interval d between trains and the ideal interval is related to the real-time interval d. D-ideal =dd vcm In the formula, the train interval is the real-time interval considering communication delay. The design adjusts the speed v. ε The value can be:

[0097]

[0098] The above formula represents the adjustment speed and the difference d D-ideal They show a linear positive correlation, t ε The preset error adjustment time is a crucial factor affecting the train interval return rate. A longer adjustment time indicates a slower rate of return to the ideal interval; conversely, a shorter adjustment time indicates a faster rate of return. According to the formula above, when the train interval equals the ideal train interval, the train only needs to continue following the reference speed curve, and the adjustment speed is 0. When the real-time interval is greater than the ideal train interval, the train needs to increase its operating speed based on the reference speed, and the adjustment speed should be positive. When the real-time interval is less than the ideal train interval, the train needs to decrease its operating speed based on the reference speed, and the adjustment speed should be negative.

[0099] The network error metric after incorporating speed adjustment is designed as follows:

[0100]

[0101] Next, in step S140, the radial basis function neural network is trained using the network error index to obtain the optimal control force of the vehicle.

[0102] Based on the network error index shown in equation (23), the network training method is designed as follows:

[0103]

[0104]

[0105] The parameter update method is still carried out according to Equations (18) and (20). Using the above method to train the neural network, as the number of training times increases, the error index will gradually decrease, and the actual curve will gradually approach the reference speed curve.

[0106] However, choosing different initial network parameters may result in different convergence speeds for network training, or even lead to failure to converge.

[0107] Figure 11 The training performance of the network is shown on the horizontal axis, representing the number of training iterations, under different initial network parameter values ​​(including center vector c, basis function width b, and weights w). The training performance is measured by the following effect of the actual curve and the reference velocity curve. In the figure, the following effect decreases sequentially from (a) to (d). Figure 11 (a) After about 200 training iterations, the actual curve converges to the vicinity of the reference velocity curve, and the subsequent approximation effect is good. Figure 11 (b) After about 1500 training iterations, the actual curve converges to the vicinity of the reference velocity curve, with a convergence speed slower than 11(a). Figure 11 (c) The curve converges after 200 training iterations, but diverges after 2300 training iterations. Figure 11 (d) No convergence was achieved at all.

[0108] Based on the experimental results above, the number of neurons in the hidden layer, the center vector, the width of the basis functions, and the initial values ​​of the weights all affect the approximation effect and convergence speed of the network. Without choosing appropriate parameters, the neural network may require a long training process to converge to within the allowable error range. However, the virtual convoy control process has high requirements for control error; therefore, to improve iterative efficiency, the network needs to be trained offline for a period of time.

[0109] This embodiment uses two methods to achieve offline training: first, network training is performed before train operation based on a predetermined reference speed curve; second, simulation training is conducted on the onboard computer during train operation. By obtaining network parameters with good training results through these two offline training methods, the rationality of neural network parameters during online control is ensured. In actual operation, the network can achieve the ideal approximation effect in a short time.

[0110] Based on the above analysis, the adaptive control principle based on the RBF neural network is obtained as follows: Figure 12 As shown. The current train operating status includes the train's current traction force, braking performance, track condition, train formation, and decoupling status. If the current train operating status remains unchanged, the RBF neural network is trained according to equations (18) and (20). If the train operating status changes, offline training parameters are selected based on the train status. Then, the control force u is calculated based on the weight matrix w output by the RBF neural network. * (k).

[0111] Finally, in step S150, the vehicle is controlled to operate according to the control force.

[0112] This application embodiment is based on RBF neural network control. In the process of calculating the train control force, the adjustment of speed is taken into account, so that the train and the preceding train can maintain speed synchronization and constant distance. Even if there are external interferences and differences in the traction and braking performance of the train, the method 100 can achieve adaptive control according to the change of the distance between the train and the preceding train. It has the characteristics of high anti-interference, high real-time performance and high punctuality.

[0113] In method 100, the vehicle ahead can be either the lead car or the train located in front of and closest to this car.

[0114] In one implementation, a lead-car following strategy is adopted. The lead car sends its reference speed curve to all following cars via car-to-car communication. The following cars calculate their own reference speed curves based on the lead car's speed curve, then control their train operation according to these curves, and adaptively adjust the distance between the following cars and the lead car. For example... Figure 13As shown, A, B, and C represent the lead car, first following car, and second following car in a train formation, respectively. Trains transmit data via car-to-car communication. The first following car B and the second following car C both follow the lead car A, theoretically allowing for synchronous following of both cars. Although the lead car following strategy can only control the distance between the second following car C and the lead car A, and not its distance from the first following car B, when the number of trains in the formation is small, the lead car following strategy can ensure that the distance between all adjacent trains in the formation is controlled within the allowable range, and according to... Figure 13 As can be seen, the offline neural network simulation process only requires calculating parameters, which involves little computation and has a fast processing speed.

[0115] In another implementation, a leading-train-following strategy is adopted. That is, all adjacent trains in the platoon form a leading-train-following relationship. The leading train sends its reference speed curve to the following train, which then generates its own reference speed curve and follows it, adaptively adjusting the distance between the leading and following trains. For example... Figure 14 As shown, train A and train B are in a front-to-back relationship, and train B and train C are in a front-to-back relationship. The lead car (i.e., train A) follows its own reference speed curve. During operation, train B obtains the reference speed curve of train A through the train-to-train communication module. and train position Calculate the reference speed curve for this vehicle. Combined with the vehicle's real-time speed and location Calculate the optimal control force for this vehicle Train C follows Train B. The reference speed curve data of the preceding train is large in volume and transmitted at long intervals, typically once per minute, resulting in poor real-time performance. However, the interval between the preceding train transmitting position information to the following train is much shorter. Each time the following train receives position information from the preceding train, its RBF neural network performs a calculation. Therefore, when the RBF neural network calculates based on the reference speed curve transmitted from the preceding train, the output usually deviates significantly from the ideal output. Thus, in this implementation, the following train's RBF neural network calculates its own reference speed curve offline; this reference speed curve is called the simulated following curve. (Reference speed curve of Train C) It is based on the simulated following curve of train B. Calculated and simulated following curve It is an RBF neural network operating offline based on the reference speed curve of train A. The predicted follow-up curve.

[0116] Compared to the lead car following strategy, the preceding car following strategy provides more direct control over the spacing between adjacent trains, making the operation of the convoy safer and particularly suitable for controlling convoys with multiple cars.

[0117] This application also provides control principles for abnormal train formation status, so as to effectively deal with emergencies during train operation, reduce the risk of accidents, and make method 100 have the advantage of high safety.

[0118] (1) Train-to-train communication failure. In actual virtual train formation operation, due to the existence of slight communication delays, the last following train may not be able to achieve synchronized operation with the lead train. However, the leading train following strategy improves the synchronization effect of adjacent trains and ensures the safety of train operation. Therefore, synchronization errors between trains that are far apart in the same train formation are acceptable. However, when train-to-train communication fails or the communication delay is large, this problem cannot be ignored.

[0119] For example, if a following train fails to receive position information from the preceding train for several consecutive communication cycles, a communication failure can be identified. To reduce the risk of accidents, this train is controlled to maintain a constant speed, and an attempt is made to broadcast a communication failure message to following trains. If the train's car-to-car communication module fails, the following trains are responsible for broadcasting the communication failure information. Upon receiving the message, the last train in the convoy begins to decelerate and brake. When the distance between the last train and the preceding train exceeds the absolute braking distance, the preceding train also begins to decelerate, and this process is repeated to complete the formation and disbanding of the convoy.

[0120] (2) The train does not follow the reference speed curve. In an emergency, the train deviates significantly from its reference speed curve. Even if the following train strictly follows its own reference speed curve, a large gap error will occur in a short period of time. At this time, it is difficult to correct the gap by simply adjusting the speed.

[0121] In such scenarios, to reduce the risk of accidents, trains that deviate from the reference speed curve immediately send new reference speed curve information to the following trains via train-to-train communication. Upon receiving this information, the following trains immediately update their own reference speed curve and sequentially transmit the updated reference speed curve information to the following trains in the convoy. Then, the train operation is controlled according to the process from steps S110 to S150.

[0122] (3) Unexpected Decoupling of Trains: When the distance between two cars in a train platoon is too large and cannot be adjusted back to the ideal distance by the control system, the train will be in an unexpected decoupling state. At this time, the original train platoon will be broken up into two train platoons. The two train platoons still share the same section of track ahead, which meets the requirements of dynamic formation. In this state, the two train platoons should apply to the ground equipment for re-formation. The ground equipment plans the formation area according to the track conditions and train conditions, and then the two train platoons complete the dynamic formation process.

[0123] The dynamic formation control method for two train sets is as follows: First, ensure communication between the lead car of the following train set and the last car of the preceding train set; then, treat the lead car of the following train set and the last car of the preceding train set as trains to be formed virtually. Next, all following trains in the following train set except for the lead car adopt the lead car following strategy and run synchronously with the lead car.

[0124] This application also provides a virtual train formation control device based on model prediction, such as... Figure 15 As shown, the device includes a communication module 1510, an information preprocessing module 1520, a radial basis function neural network module 1530, and a control module 1540.

[0125] The communication module 1510 is adapted to acquire the position and reference speed curve of the vehicle ahead in the virtual convoy at the current moment.

[0126] The information preprocessing module 1520 is adapted to calculate the reference speed curve of the vehicle in combination with the reference speed curve of the vehicle in front at the current moment, calculate the adjustment speed in combination with the position of the vehicle in front and the position of the vehicle in front at the current moment, and use the adjustment speed to correct the network error index.

[0127] The radial basis function neural network module 1530 is adapted to train the radial basis function neural network using the network error index to obtain the control force of the vehicle.

[0128] The control module 1540 is adapted to control the operation of the vehicle according to the control force.

[0129] In a preferred embodiment of the present invention, the information preprocessing module calculates the adjustment speed in the following manner:

[0130] Calculate the train interval between the vehicle ahead and the current vehicle based on the current position of the vehicle ahead and the current position of the current vehicle; and

[0131] The adjustment speed is calculated based on the train interval and the preset error adjustment time.

[0132] In a preferred embodiment of the present invention, the step of training the radial basis function neural network using the network error index includes online training and offline training.

[0133] The online training involves training the radial basis function neural network based on the reference speed curve of the vehicle ahead.

[0134] The offline training refers to: conducting network training based on a predetermined reference curve before the train runs, or conducting simulation training on the onboard computer during operation.

[0135] In a preferred embodiment of the present invention, the leading vehicle is the lead vehicle in a virtual convoy.

[0136] In a preferred embodiment of the present invention, the device further includes a first anomaly control module.

[0137] The first anomaly control module is adapted to control the vehicle to run at a constant speed and broadcast communication fault information to the following vehicles if it fails to obtain the position and reference speed curve of the vehicle in front of it in the virtual formation convoy at the current moment within a number of consecutive communication cycles.

[0138] In a preferred embodiment of the present invention, the device further includes a second anomaly control module.

[0139] The second anomaly control module is adapted to update the reference speed curve of the vehicle according to the actual speed of the vehicle when the deviation between the actual speed of the vehicle and the reference speed curve of the vehicle reaches a preset threshold, and send the updated reference speed curve to the following vehicle.

[0140] This application also provides an electronic device, such as... Figure 16 As shown. When the electronic device starts running, it communicates with other vehicles through the communication device 1630. The processor 1620 reads program instructions from the memory 1610 for implementing the train close tracking adaptive control method 100 and executes them. These program instructions can instruct the processor 1620 to execute the train close tracking adaptive control method of the present invention.

[0141] The various techniques described herein can be implemented in combination with hardware or software, or a combination thereof. Thus, the methods and apparatus of the present invention, or certain aspects or portions thereof, can take the form of program code (i.e., instructions) embedded in a tangible medium, such as a removable hard disk, USB flash drive, floppy disk, CD-ROM, or any other machine-readable storage medium, wherein when the program is loaded into and executed by a machine such as a computer, the machine becomes an apparatus for practicing the present invention.

[0142] In the specification provided herein, the algorithms and displays are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used with the examples of this invention. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing preferred embodiments of the invention.

[0143] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and edibility purposes, and not for the purpose of explaining or limiting the subject matter of the invention.

[0144] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for adaptive control of close train following, characterized in that, The method comprises the following steps: acquiring the position and reference speed curve of a front vehicle in a virtual marshalling train at a current time; calculating the reference speed curve of the train according to the reference speed curve of the front vehicle at the current time; calculating the adjustment speed according to the position of the front vehicle and the position of the train at the current time, and correcting the network error index by using the adjustment speed; the adjustment speed is calculated according to the position of the front vehicle and the position of the train at the current time, which comprises the following steps: calculating the train interval between the front vehicle and the train according to the position of the front vehicle and the position of the train at the current time; and calculating the adjustment speed according to the train interval and a preset error adjustment time; the radial basis function neural network is trained by using the network error index to obtain the control force of the train; the training of the radial basis function neural network by using the network error index comprises online training and offline training; the online training is training the radial basis function neural network according to the reference speed curve of the front vehicle; the offline training is network training according to a predetermined reference curve before the train runs, or simulation training in the on-board computer during the running process; and controlling the train to run according to the control force.

2. The method of claim 1, wherein, The front vehicle is the vehicle located in front of the train and closest to the train in the virtual marshalling train.

3. The method of claim 1, wherein: The method further comprises the following steps: if the position and reference speed curve of the front vehicle in the virtual marshalling train at the current time cannot be acquired for continuous communication periods, controlling the train to run at a constant speed, and broadcasting communication failure information to a rear vehicle.

4. The method of claim 1, wherein: The method further comprises the following steps: when the deviation between the actual speed of the train and the reference speed curve of the train reaches a preset threshold, updating the reference speed curve of the train according to the actual speed of the train, and sending the updated reference speed curve to the rear vehicle.

5. A model prediction based virtual marshalling device, characterized in that, The system comprises the following modules: a communication module adapted to acquire the position and reference speed curve of a front vehicle in a virtual marshalling train at a current time; an information preprocessing module adapted to calculate the reference speed curve of the train according to the reference speed curve of the front vehicle at the current time, and calculate the adjustment speed according to the position of the front vehicle and the position of the train at the current time, and correct the network error index by using the adjustment speed; the adjustment speed is calculated according to the position of the front vehicle and the position of the train at the current time, which comprises the following steps: calculating the train interval between the front vehicle and the train according to the position of the front vehicle and the position of the train at the current time; and calculating the adjustment speed according to the train interval and a preset error adjustment time; a radial basis function neural network module adapted to train the radial basis function neural network by using the network error index, which comprises online training and offline training; the online training is training the radial basis function neural network according to the reference speed curve of the front vehicle; the offline training is network training according to a predetermined reference curve before the train runs, or simulation training in the on-board computer during the running process; obtaining the control force of the train; a control module adapted to control the train to run according to the control force.

6. The apparatus of claim 5, wherein: The information preprocessing module calculates the adjustment speed by the following ways: calculating the train interval between the front vehicle and the host vehicle according to the position of the front vehicle and the position of the host vehicle at the current time; and calculating the adjustment speed according to the train interval and a preset error adjustment time.

7. An electronic device, comprising: comprise: a memory; a processor; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method of any one of claims 1 to 4.

8. A computer readable storage medium, characterized in that, having a computer program stored thereon; the computer program is executed by a processor to implement the method of any one of claims 1 to 4.

Images