Bridge rotation method based on monocular vision and robust optimization algorithm

Abstract

The present invention discloses a bridge rotation method based on monocular vision and robust optimization algorithm, comprising: installing a monocular depth vision camera on the top of a rotating device to capture the position of a sensor target on a traction device, and transmitting the collected image signal to a main controller; the main controller calculates a series of parameters of the rotating device in real time according to the parameters of the sensor in the image; the main operating platform screens a series of parameters based on a robust optimization algorithm and controls them within a safe and reasonable rotation range; the main operating platform transmits the screened optimal rotation parameter range to the sensor target to control the voltage and current required to be loaded on the traction device, thereby further controlling the travel change of the steel strand and the rotation arc length; the above process is repeated to complete the entire process of the bridge rotation; the present invention effectively avoids unsafe conditions caused by human errors and insufficient monitoring frequency, and greatly saves the cost of manpower and material resources.

Description

Technical Field

[0001] The invention relates to the technical field of automatic control of bridge rotation, in particular to a bridge rotation method based on monocular vision and a robust optimization algorithm. Background Art

[0002] When the existing railway or highway lines that the new bridge needs to cross are wide and have heavy traffic, the general construction method cannot be completed in a short time. Therefore, in order to reduce the impact of the new bridge on the traffic operation of the existing line, the bridge structure is manufactured at a non-design axis position, and then the bridge structure is rotated and installed to the designed position with the help of a rotation system to achieve crossing the existing route. The rotation system consists of a lower turntable, an upper turntable, a ball joint, a slideway, and a traction system. The rotation process is generally achieved by continuously pulling the traction cable with a jack to form a rotating force couple.

[0003] The rotation method is used for construction. The rotation has the characteristics of heavy weight, short time window, high construction difficulty and complex bridge type. How to accurately position and smoothly rotate the rotating bridge is the main difficulty in the rotation. Therefore, the data records during the rotation process play a very important role in the optimization and supplement of the rotating bridge system. It is the data reserve for the later rotation construction and the basis for the revision of the technical standards of the rotating bridge. At present, the monitoring method during the bridge rotation construction is mainly to use the trial rotation method for manual measurement and manual calculation. The degree of automation of the rotation construction monitoring is low, the coordination cycle between the construction steps is long, and the degree of visualization is low. Conventional rotation operations usually require two stages: trial rotation and formal rotation. The trial rotation stage also requires a lot of manpower and material costs, and the operation process is complicated. Safety accidents are prone to occur during the construction process, resulting in the failure of the bridge rotation. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a bridge rotation method based on monocular vision and robust optimization algorithm. The present invention effectively avoids unsafe conditions caused by human errors and insufficient monitoring frequency, and greatly saves the cost of manpower and material resources.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a bridge rotation method based on monocular vision and robust optimization algorithm, comprising the following steps:

[0006] Step 1: Install a monocular depth vision camera on the top of the rotating device. When the traction device starts to work during the rotation process, the monocular depth vision camera is used to capture the position of the sensor target on the traction device and transmit the collected relevant image signals to the main controller of the rotating device in real time and quickly;

[0007] Step 2, the main controller calculates a series of parameters of the rotating device in real time according to the parameters of the sensor in the image;

[0008] Step 3: The main controller sends the calculated real-time data to the main operating platform of the traction device, and the main operating platform selects a series of parameters based on a robust optimization algorithm and controls them within a safe and reasonable rotation range;

[0009] Step 4: The main operating platform transmits the optimal rotation parameter range obtained by screening to the sensor target to control the voltage and current required to be loaded by the traction device, thereby further controlling the change in the steel strand stroke and the rotation arc length;

[0010] Step 5: Repeat the above process to complete the entire process of bridge rotation.

[0011] As a further improvement of the present invention, if an abnormal situation occurs during the rotation process, that is, a series of parameters calculated by the main operating platform exceed the safe and reasonable rotation range, the main operating platform issues a warning and takes emergency braking measures, and stops the bridge rotation process.

[0012] As a further improvement of the present invention, a total of six monocular depth vision wide-angle cameras are arranged on the top of the rotating device to provide complete 360-degree full-range coverage, and the monocular depth vision wide-angle cameras are connected to the main controller through a built-in wireless signal transmitter to quickly transmit the collected image data to the main controller.

[0013] As a further improvement of the present invention, in step 2, the parameters of the sensor include the position, speed and acceleration of the target object, and a series of parameters of the rotating device include the horizontal coordinate position, speed and acceleration of the target object of the rotating device in the x and y directions, the rotation distance, the rotation angular velocity and the rotation linear velocity.

[0014] As a further improvement of the present invention, in step 3, the robust optimization algorithm adjusts the connection weights between the multi-layer neurons in the reinforcement learning neural network model by adopting an improved particle swarm optimization algorithm to achieve adaptive learning of the reinforcement neural network for sensor target trajectory tracking control, specifically comprising the following steps:

[0015] a. Convert the problem of selecting the connection weights between multiple layers of neurons into an optimization problem. The objective function of the optimization problem is the L2 norm of the output vector and the input signal vector, as shown in formula (1). The input signal is the sensor target coordinates, and the output vector is the threshold value of the loading current and voltage required by the traction device.

[0016]

[0017] In formula (1), Error is the objective function of the optimization problem, xd(k) and yd(k) are the expected speed and expected acceleration of the sensor target at the kth step, respectively, and y x (k), yy (k) is the actual speed and actual acceleration value of the kth step;

[0018] b. Use previous experience to determine the value range of each connection weight, that is, determine the optimization range;

[0019] c. Randomly initialize a group of particles within the optimization range, namely the particle swarm, including the initial position and initial velocity of the initialized particles. The particle characteristics are represented by the three indicators of position, velocity and fitness. The position represents the value of all connection weights in the reinforcement learning neural network model, the velocity represents the direction of evolution of each particle, and the fitness value is obtained by the fitness function, namely the objective function corresponding to each particle.

[0020] The speed of the particle is updated based on the current position of the particle, the current speed, the particle's historical best position Pbest and the position of the best particle in the particle group Gbest. The update formula is shown in formula (2):

[0021]

[0022] In formula (2), id is the number of the particle in the particle group, is the velocity of the i-th generation particle, is the best historical position of the i-th generation particle before the i-th generation, is the position of the optimal particle in the i-th generation particle group; ω(i) is the inertia weight of the i-th generation particle, and its size determines to what extent the velocity inherits the motion velocity of the previous generation of particles; c1, c2 are acceleration factors, which are non-negative constants; r1, r2 are random numbers between 0 and 1; is the position of the particle of the i-th generation; the value of ω at the time of initialization is ω start is 0.9, and at the end of the iteration, ω start The value of ω end is 0.01. In the iterative process, the inertia weight ω decays rapidly. In the initial stage, the optimization speed is prioritized, and in the later stage, the optimization accuracy is emphasized. The update formula of the inertia weight ω is shown in formula (3):

[0023]

[0024] In formula (3), maxgen is the maximum number of iterations; the updated particle speed is obtained After that, update the position of the particle. The update formula is shown in formula (4):

[0025]

[0026] If the formula (4) is obtained The corresponding objective function is less than The corresponding fitness function is on the contrary, At the same time, update the position of the optimal particle in the population and get

[0027] In this way, multiple iterations can obtain an approximate optimal solution, that is, the approximate optimal weight with the smallest control error of the reinforcement learning neural network model, thereby screening out a safe and reasonable rotation range.

[0028] The beneficial effects of the present invention are:

[0029] The present invention avoids the trial rotation process during the rotation of the bridge. During the rotation process, the main controller can calculate a series of parameters such as the rotation distance, rotation angle, angular velocity, linear velocity, etc. of the rotating bridge in real time, and send the calculated real-time parameters to the main operating platform of the traction device. The main operating platform screens the calculated parameters based on the robust optimization algorithm and controls them within a reasonable rotation range. The main operating platform transmits the screened optimal rotation parameter range to the sensor target to control the voltage and current required to be loaded by the traction device, thereby further controlling the stroke change of the steel strand and the rotation arc length. Finally, the algorithm is used to realize the function of positioning while rotating during the rotation process to perform automatic positioning during the rotation construction process, effectively avoiding unsafe conditions caused by human errors and insufficient monitoring frequency, and greatly saving manpower and material costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0032] Example

[0033] like Figure 1 As shown, a bridge rotation method based on monocular vision and robust optimization algorithm includes the following steps:

[0034] Step 1: Install a monocular depth vision camera on the top of the rotating device. When the traction device starts to work during the rotation process, the monocular depth vision camera will immediately capture the position of the sensor target on the traction device and quickly transmit the collected relevant image signals to the main controller of the rotating device in real time;

[0035] Step 2: The main controller calculates a series of parameters such as the rotation angle, rotation distance, rotation angular velocity and rotation linear velocity of the rotating device in real time according to the change in the size of the sensor target speed in the image, the displacement distance per millisecond, and the rotation arc;

[0036] Step 3: The main controller sends the calculated real-time data to the main operating platform of the traction device, and the main operating platform selects a series of parameters based on a robust optimization algorithm and controls them within a safe and reasonable rotation range;

[0037] Step 4: The main operating platform transmits the optimal rotation parameter range obtained by screening to the sensor target to control the voltage and current required to be loaded by the traction device, thereby further controlling the change in the steel strand stroke and the rotation arc length;

[0038] Step 5: Repeat the above process to complete the entire process of bridge rotation. If an abnormal situation occurs during the rotation, that is, a series of parameters calculated by the main operating platform exceed the safe and reasonable rotation range, the main operating platform will issue a warning and take emergency braking measures to stop the bridge rotation process.

[0039] In this embodiment, a total of six monocular depth vision wide-angle cameras are arranged on the top of the rotating device, which can provide a complete 360-degree full-range coverage, and the monocular depth vision wide-angle camera is connected to the main controller through a built-in wireless signal transmitter, which has extremely strong real-time performance and can quickly transmit the collected image data to the main controller.

[0040] Specifically, the main controller in this embodiment integrates image recognition conversion, high-speed storage calculation, real-time data transmission, and rapid prototype control, and can respond quickly and accurately according to the state of the traction device.

[0041] The parameters obtained by the main controller are the position, speed and acceleration of the target object. The series of parameters output include the horizontal coordinate position, speed and acceleration of the target object in the x and y directions of the rotating device, the rotation distance, the rotation angular velocity and the rotation linear velocity.

[0042] In step 3, a series of parameters for screening and controlling a safe and reasonable rotation range are performed based on a robust optimization algorithm, and the connection weights between multiple layers of neurons in the reinforcement learning neural network model are adjusted by using an improved particle swarm optimization algorithm to achieve adaptive learning of the reinforcement neural network for sensor target trajectory tracking control, including:

[0043] a. Convert the problem of selecting the connection weights between multiple layers of neurons into an optimization problem. The objective function of the optimization problem is the L2 norm of the output vector and the input signal vector, as shown in formula (1). The input signal is the sensor target coordinates, and the output vector is the threshold value of the loading current and voltage required by the traction device.

[0044]

[0045] In formula (1), Error is the objective function of the optimization problem, xd(k) and yd(k) are the expected speed and expected acceleration of the sensor target at the kth step, respectively, and y x (k), y y (k) is the actual speed and actual acceleration value of the kth step;

[0046] b. Use previous experience to determine the value range of each connection weight, that is, determine the optimization range;

[0047] c. Randomly initialize a group of particles within the optimization range, namely the particle swarm, including the initial position and initial velocity of the initialized particles. The particle characteristics are represented by the three indicators of position, velocity and fitness. The position represents the value of all connection weights in the reinforcement learning neural network model, the velocity represents the direction of evolution of each particle, and the fitness value is obtained by the fitness function, namely the objective function corresponding to each particle.

[0048] The speed of the particle is updated based on the current position of the particle, the current speed, the particle's historical best position Pbest and the position of the best particle in the particle group Gbest. The update formula is shown in formula (2):

[0049]

[0050] In formula (2), id is the number of the particle in the particle group, is the velocity of the i-th generation particle, is the best historical position of the i-th generation particle before the i-th generation, is the position of the optimal particle in the i-th generation particle group; ω(i) is the inertia weight of the i-th generation particle, and its size determines to what extent the velocity inherits the motion velocity of the previous generation of particles; c1, c2 are acceleration factors, which are non-negative constants; r1, r2 are random numbers between 0 and 1; is the position of the particle of the i-th generation; the value of ω at the time of initialization is ω start is 0.9, and at the end of the iteration, ω start The value of ω end is 0.01. In the iterative process, the inertia weight ω decays rapidly. In the initial stage, the optimization speed is prioritized, and in the later stage, the optimization accuracy is emphasized. The update formula of the inertia weight ω is shown in formula (3):

[0051]

[0052] In formula (3), maxgen is the maximum number of iterations; the updated particle speed is obtained After that, update the position of the particle. The update formula is shown in formula (4):

[0053]

[0054] If the formula (4) is obtained The corresponding objective function is less than The corresponding fitness function is on the contrary, At the same time, update the position of the optimal particle in the population and get

[0055] In this way, multiple iterations can obtain an approximate optimal solution, that is, the approximate optimal weight with the smallest control error of the reinforcement learning neural network model, thereby screening out a safe and reasonable rotation range.

[0056] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A bridge rotation method based on monocular vision and robust optimization algorithm, characterized in that: The following steps are involved: Step 1: Install a monocular depth vision camera on the top of the rotating device. When the traction device starts to work during the rotation process, the monocular depth vision camera is used to capture the position of the sensor target on the traction device and transmit the collected relevant image signals to the main controller of the rotating device in real time and quickly; Step 2, the main controller calculates a series of parameters of the rotating device in real time according to the parameters of the sensor in the image; Step 3: The main controller sends the calculated real-time data to the main operating platform of the traction device, and the main operating platform selects a series of parameters based on a robust optimization algorithm and controls them within a safe and reasonable rotation range; Step 4: The main operating platform transmits the optimal rotation parameter range obtained by screening to the sensor target to control the voltage and current required to be loaded by the traction device, thereby further controlling the change in the steel strand stroke and the rotation arc length; Step 5: Repeat the above process to complete the entire process of bridge rotation.

2. The bridge rotation method based on monocular vision and robust optimization algorithm according to claim 1 is characterized in that: If an abnormal situation occurs during the rotation process, that is, a series of parameters calculated by the main operating platform exceed the safe and reasonable rotation range, the main operating platform will issue a warning and take emergency braking measures, and stop the bridge rotation process.

3. The bridge rotation method based on monocular vision and robust optimization algorithm according to claim 1 is characterized in that: A total of six monocular depth vision wide-angle cameras are arranged on the top of the rotating device to provide complete 360-degree full-range coverage, and the monocular depth vision wide-angle cameras are connected to the main controller through a built-in wireless signal transmitter to quickly transmit the collected image data to the main controller.

4. The bridge rotation method based on monocular vision and robust optimization algorithm according to claim 1 is characterized in that: In step 2, the sensor parameters include the position, velocity and acceleration of the target object, and a series of parameters of the rotating device include the horizontal coordinate position, velocity and acceleration of the target object in the x and y directions of the rotating device, the rotation distance, the rotation angular velocity and the rotation linear velocity.

5. The bridge rotation method based on monocular vision and robust optimization algorithm according to claim 1 is characterized in that: In step 3, the robust optimization algorithm adjusts the connection weights between the multi-layer neurons in the reinforcement learning neural network model by using an improved particle swarm optimization algorithm to achieve adaptive learning of the reinforcement neural network for sensor target trajectory tracking control, specifically including the following steps: a. Convert the problem of selecting the connection weights between multiple layers of neurons into an optimization problem. The objective function of the optimization problem is the L2 norm of the output vector and the input signal vector, as shown in formula (1). The input signal is the sensor target coordinates, and the output vector is the threshold value of the loading current and voltage required by the traction device. In formula (1), Error is the objective function of the optimization problem, xd(k) and yd(k) are the expected speed and expected acceleration of the sensor target at the kth step, respectively, and y x (k), y y (k) is the actual speed and acceleration value of the kth step; b. Use previous experience to determine the value range of each connection weight, that is, determine the optimization range; c. Randomly initialize a group of particles within the optimization range, namely the particle swarm, including the initial position and initial velocity of the initialized particles. The particle characteristics are represented by the three indicators of position, velocity and fitness. The position represents the value of all connection weights in the reinforcement learning neural network model, the velocity represents the direction of evolution of each particle, and the fitness value is obtained by the fitness function, namely the objective function corresponding to each particle. The speed of the particle is updated based on the current position of the particle, the current speed, the particle's historical best position Pbest and the position of the best particle in the particle group Gbest. The update formula is shown in formula (2): In formula (2), id is the number of the particle in the particle group, is the velocity of the i-th generation particle, is the best historical position of the i-th generation particle before the i-th generation, is the position of the optimal particle in the i-th generation particle group; ω(i) is the inertia weight of the i-th generation particle, and its size determines to what extent the velocity inherits the motion velocity of the previous generation of particles; c1, c2 are acceleration factors, which are non-negative constants; r1, r2 are random numbers between 0 and 1; is the position of the particle of the i-th generation; the value of ω at initialization is ω start is 0.9, and at the end of the iteration, ω start The value of ω end is 0.

01. In the iterative process, the inertia weight ω decays rapidly. In the initial stage, the optimization speed is prioritized, and in the later stage, the optimization accuracy is emphasized. The update formula of the inertia weight ω is shown in formula (3): In formula (3), maxgen is the maximum number of iterations; the updated particle speed is obtained After that, update the position of the particle. The update formula is shown in formula (4): If the formula (4) is obtained The corresponding objective function is less than The corresponding fitness function is on the contrary, At the same time, update the position of the optimal particle in the population and get In this way, multiple iterations can obtain an approximate optimal solution, that is, the approximate optimal weight with the smallest control error of the reinforcement learning neural network model, thereby screening out a safe and reasonable rotation range.

Images