Sleeper labeling method based on deep learning algorithm

Abstract

The invention discloses a deep learning algorithm-based sleeper labeling method, which comprises the following steps of: S1, acquiring sleeper image data, and generating an initial coordinate of a sleeper position; s2, acquiring real-time position information of the equipment based on a mileage measuring device; s3, constructing a dynamic curve section marking optimization algorithm, calculating the shortening amount of each marking point, and dynamically adjusting the spacing of the marking points; s4, optimizing a spraying path based on a reinforcement learning algorithm, and generating an optimal spraying action sequence; s5, in combination with a multi-sensor fusion technology, the position and the distance of spraying nozzles are adjusted in real time; s6, displaying the operation state of the equipment at the handheld end by using a remote control system; s7, based on a self-learning optimization model built in the equipment, dynamically adjusting spraying parameters according to a classification result; and S8, through real-time pressure monitoring of the spraying module, the spraying width and length meet preset requirements, and periodic correction is conducted on the spraying device. The method has the advantages of being high in labeling precision, high in efficiency and high in environmental adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of sleeper labeling, and in particular to a sleeper labeling method based on a deep learning algorithm. Background Art

[0002] With the rapid development of rail transit, sleeper marking plays an important role in track laying and maintenance. Especially in complex curved sections and large-scale track construction scenarios, how to efficiently and accurately complete sleeper marking has become a key technical issue. The accuracy of sleeper marking is directly related to the stability and safety of the track, but traditional sleeper marking technology has many shortcomings in practical applications, which limits the improvement of construction efficiency and quality.

[0003] In the existing technology, sleeper marking usually relies on manual operation and simple mechanical auxiliary equipment. These methods are incapable of meeting the requirements of complex environments and high precision. Specifically, the traditional sleeper marking technology has the following major defects:

[0004] 1. Limited marking accuracy: Traditional mechanically assisted marking equipment lacks intelligent recognition and optimization capabilities, making it difficult to achieve consistency in the spacing between marking points on curves, especially in track areas with large changes in curvature.

[0005] 2. Inefficiency: Manual labeling methods require a lot of manpower input, are complex and time-consuming to operate, and the physical limitations of construction workers further affect construction efficiency.

[0006] 3. Poor environmental adaptability: Traditional methods have insufficient response capabilities to changes in the external environment (such as curve radius and uneven distribution of sleepers), and are unable to adjust annotation parameters in real time to adapt to the construction environment.

[0007] 4. Lack of intelligent optimization and feedback mechanism: Traditional equipment usually does not have self-learning capabilities and cannot use historical annotation data for dynamic optimization, which leads to repeated errors in the annotation process and is difficult to improve.

[0008] 5. Safety hazards: During manual operation, the construction environment is complex and high-risk. The remote control and real-time monitoring capabilities of traditional mechanical equipment are limited, making it difficult to effectively ensure the safety of construction workers.

[0009] Therefore, how to provide a sleeper labeling method based on deep learning algorithm is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0010] One object of the present invention is to propose a sleeper labeling method based on a deep learning algorithm. The present invention makes full use of deep learning models, dynamic optimization algorithms, multi-sensor fusion technology and intelligent control systems, and describes in detail the implementation process of automatic sleeper identification, curve section labeling optimization, intelligent spray path planning and labeling error self-learning optimization, and has the advantages of high labeling accuracy, high efficiency and strong environmental adaptability.

[0011] A sleeper labeling method based on a deep learning algorithm according to an embodiment of the present invention is characterized by comprising the following steps:

[0012] S1. Collect sleeper image data, extract features from the image through a convolutional neural network, use multi-layer convolution operations to obtain the position information of the edge and center of the sleeper, and use a non-maximum suppression algorithm to filter interference points to generate the initial coordinates of the sleeper position;

[0013] S2. Based on the mileage measurement device, the photoelectric encoder and the inertial measurement unit are used for data fusion, and the travel position and posture of the equipment are corrected through the Kalman filter algorithm. At the same time, the laser ranging device is combined to compensate for the data error and obtain the real-time position information of the equipment;

[0014] S3. Construct a dynamic curve section marking optimization algorithm to calculate the shortening of each marking point and dynamically adjust the spacing between marking points according to the track curve radius, transition curve parameters and equipment mileage;

[0015] S4. Optimize the spraying path based on the reinforcement learning algorithm. By building a path planning model based on the deep Q network, train the spraying path of the equipment under different track characteristics, adjust the spraying start and stop points and path based on real-time track information, and generate the optimal spraying action sequence;

[0016] S5. Combine multi-sensor fusion technology, use the data from mileage sensors, laser ranging devices and attitude sensors, adopt weighted average method to fuse the measurement results of each data source, detect abnormal points on the fusion results, and adjust the position and spacing of the spray nozzles in real time;

[0017] S6. Using the remote control system, the annotation device is connected via the narrowband communication module, and the handheld terminal displays the device operation status;

[0018] S7, based on the built-in self-learning optimization model of the equipment, collect historical annotation data, use the support vector machine algorithm to classify and analyze the spraying errors, and dynamically adjust the spraying parameters according to the classification results;

[0019] S8. Through real-time pressure monitoring of the spraying module, the working pressure of the spraying device is adjusted based on the PID control algorithm so that the spraying width and length meet the preset requirements, and the spraying device is periodically calibrated.

[0020] Optionally, the S1 specifically includes:

[0021] S11, collecting real-time image data of the track area through a camera device, inputting the image data into a preprocessing module, graying the image, removing irrelevant color information, and enhancing the edge features of the target area;

[0022] S12, perform Gaussian filtering on the preprocessed image to reduce noise interference:

[0023]

[0024] The size of the convolution kernel used is k×k, where k is an odd number, x and y are the coordinates of the pixel points, and σ is the standard deviation, which determines the degree of smoothing of the filter on the edge details of the sleeper;

[0025] S13, inputting the filtered image data into a convolutional neural network, and extracting features of the edge and center images of the sleeper by setting multiple convolutional layers:

[0026]

[0027] Among them, f ij is the convolution output, which represents the feature value of the current pixel, I(i+m,j+n) is the pixel value of the input image, and W mn is the convolution kernel weight, b is the bias value, and k is the convolution kernel size;

[0028] S14, performing a pooling operation on the convolution output result, using a maximum pooling method, by setting a fixed-size pooling window to reduce the size of the feature map while retaining the maximum value in each pooling window;

[0029] S15, processing the pooling result through a fully connected layer, mapping it to the target classification space, and generating a classification result to determine the probability of the sleeper position;

[0030] S16. Use the non-maximum suppression algorithm to filter the interference points on the classification results, set the threshold T, remove the detection boxes with scores lower than T, retain the sleeper position with the highest score, and output the final initial coordinates of the sleeper position.

[0031] Optionally, the S2 specifically includes:

[0032] S21, collecting the number of pulses per unit time during the travel of the device through a photoelectric encoder, and calculating the travel mileage of the device based on the number of pulses of the encoder;

[0033] S22, collecting acceleration data and angular velocity data of the device through an inertial measurement unit, and calculating the attitude angle of the device according to the angular velocity data;

[0034] S23, synchronously processing the data of the photoelectric encoder and the inertial measurement unit, and using the Kalman filter algorithm to correct the travel position and posture of the device, including predicting the device state and updating the state estimate;

[0035] S24, obtaining ranging data on the track in combination with a laser ranging device, and marking measurement points exceeding the threshold as abnormal according to a set ranging accuracy threshold;

[0036] S25, integrate the data of the photoelectric encoder, inertial measurement unit and laser ranging device to calculate the final device position:

[0037] X f =w 1 X e +w 2 X i +w 3 X l ;

[0038] Among them, X f is the final position, X e , X i , X l are the data of photoelectric encoder, inertial measurement unit and laser ranging, respectively. 1 、w 2 、w 3 is the weight of the corresponding data source, satisfying w 1 +w 2 +w 3 =1;

[0039] S26. Output the real-time location information of the device, including mileage, attitude angle, and coordinates of the marked abnormal area.

[0040] Optionally, the S3 specifically includes:

[0041] S31. Collect track geometry parameters, including track curve radius R, transition curve length L h and the length of the straight line segment L s , and associate it with the mileage data D of the device;

[0042] S32. Calculate the shortening amount ΔL of the curve segment c The shortening of the transition curve ΔL h :

[0043]

[0044] Where, ΔL c is the shortening of the circular curve segment, ΔL h S is the shortening of the transition curve segment. 1is the distance between the center lines of the two rails, R is the track curve radius, L c is the length of the circular curve, L h is the length of the transition curve;

[0045] S33. According to the travel position of the device, determine the type of track segment to which the current point belongs. If it is a transition curve segment, calculate the dynamic shortening amount of each point:

[0046]

[0047] Where, ΔL h,i is the shortening of the i-th point in the transition curve, d i is the distance from the starting point of the transition curve to the i-th point;

[0048] S34, based on the calculation results of the straight line segment, transition curve segment and circular curve segment, dynamically adjust the spacing of the sleeper marking points, the spacing of the marking points ΔD i Calculated according to the following rules:

[0049] ΔD i =D 0 -ΔL;

[0050] Where, ΔD i is the adjusted spacing of the ith annotation point, D 0 is the initial spacing, ΔL is the shortening of the corresponding segment;

[0051] S35, transmitting the dynamically adjusted marking point spacing to the spraying control module, updating the mileage position data of each marking point, and generating complete marking path data.

[0052] Optionally, the S4 specifically includes:

[0053] S41, collecting geometric parameters of the track environment and spraying requirement data, including the straight length, curve radius, transition curve length and spacing requirements of the marked points of the track, and constructing the state space of the spraying path;

[0054] S42. Design a reinforcement learning framework for the spraying path, define the state set as the combination of the current position of the device and the spraying state, define the action set as the spraying action of the device, define the reward function according to the path optimization degree of the device to complete the current spraying task, and the state transition function describes the path evolution process through the probability of the device reaching the next state through a certain action from the current state;

[0055] S43. Optimize the spraying path based on the deep Q network, and use the Q value update method to adjust the spraying strategy of the equipment. The Q value update method makes corrections by adjusting the difference between the current Q value and the maximum Q value of the subsequent state:

[0056] Q(s,a)=Q(s,a)+α[r+γmax a Q(s ′ ,a)-Q(s,a)];

[0057] Among them, Q(s,a) represents the Q value of the current state and action, α is the learning rate, which is used to control the update amplitude, r is the immediate reward of the current action, γ is the discount factor, which is used to balance the weight of short-term rewards and long-term benefits, and max a Q(s ′ ,a) represents the action value with the highest expected reward in the subsequent state;

[0058] S44. The Q value function is fitted by a deep neural network. The input of the neural network is the combined state of the current position of the device and the track characteristics. The output is the Q value of all possible spraying actions. The mean square error is used as the loss function to minimize the difference between the predicted Q value and the target Q value:

[0059] L(θ)=E[(r+γmax a Q(s′,a;θ′)-Q(s,a;θ)) 2 ];

[0060] Among them, θ represents the parameters of the current deep neural network, θ′ is the parameter of the target network, and the loss function is optimized through multiple iterations to enhance the adaptability of the model to complex orbit paths;

[0061] S45, through multiple rounds of training iterations, dynamically update the Q value according to the characteristics of the track environment and the equipment spraying task, and generate the optimal spraying action sequence;

[0062] S46, combining the track information obtained by the device in real time, dynamically adjusting the spraying start and stop points and the path in the path planning model;

[0063] Optionally, the S5 specifically includes:

[0064] S51, collecting travel distance data of the device through a mileage sensor, and dynamically calibrating the device in combination with operating parameters of the device to record the real-time travel distance;

[0065] S52, obtaining real-time distance measurement data between the track and the equipment through a laser distance measuring device, and performing outlier detection in combination with a deep learning model;

[0066] S53, collect the acceleration and angular velocity data of the device through the attitude sensor, and calculate the attitude angle after real-time calibration based on the inertial navigation algorithm:

[0067] θ c =θ+k a ·a t ;

[0068] Among them, θ c is the calibrated attitude angle, θ is the initial attitude angle, k a is the acceleration correction factor, a t is the current acceleration;

[0069] S54, integrates the data of the mileage sensor, laser ranging device and attitude sensor, and uses a dynamic fusion algorithm to calculate the final position of the device:

[0070]

[0071] Among them, X f is the final position, X is the current device position, is the weight ratio of laser ranging and mileage sensor, D t It is the distance measurement data after dynamic calibration;

[0072] S55, combined with track characteristics, outlier detection is performed on sensor fusion data, and the outliers are marked and processed using a dynamic threshold adjustment mechanism;

[0073] S56. Adjust the position of the spray nozzle and the spacing between marking points in real time according to the sensor fusion results, generate dynamic spray instructions, and synchronize the adjusted spray instructions to the spray control module through the intelligent decision-making module on the device side.

[0074] Optionally, the S7 specifically includes:

[0075] S71. Collecting real-time annotation data generated by the device during the spraying process, including the actual position, spacing, and spraying width of the annotation points, and building an annotation error sample set in combination with historical annotation data;

[0076] S72, extracting features from the collected error samples, wherein the extracted features include the marking point offset Δx, the marking point spacing deviation Δd, and the spraying width deviation Δw of the sleeper;

[0077] S73. Based on the support vector machine algorithm, the extracted error samples are classified and analyzed to build a classification model:

[0078]

[0079] Among them, f(x) is the classification decision function, α i is the Lagrange multiplier, y i is the category label of the sample, x i and x are support vectors and input samples respectively, K(x i ,x) is the kernel function, b is the bias;

[0080] S74, dividing the error samples into normal areas and abnormal areas according to the classification results, and extracting error correction parameters through error features of the abnormal areas;

[0081] S75, using error correction parameters to dynamically adjust spraying parameters, including spraying spacing Δd s , spraying width Δw s and spraying time t s :

[0082] Δd s =Δd-k d ·e d ;

[0083] Δw s =Δw-k w ·e w ;

[0084] t s =tk t ·e t ;

[0085] Where, Δd s , Δw s ,t s are the adjusted spraying spacing, width and time respectively, Δd, Δw, t are the initial spraying parameters, e d 、e w 、e t is the corresponding error value, k d , k w , k t is the error correction coefficient;

[0086] S76, transmitting the adjusted spraying parameters to the equipment spraying control module in real time, and updating the equipment spraying instructions to correct the errors in the current operation.

[0087] Optionally, the S8 specifically includes:

[0088] S81, collect the working pressure data of the spraying device in real time through the pressure sensor in the spraying module, and record the current pressure P t and target pressure P r ;

[0089] S82, input the collected pressure data into the PID control algorithm, and calculate the adjustment signal according to the pressure error:

[0090]

[0091] Where u(t) is the adjustment signal, e(t) = P r -P t is the current pressure error, K p, K i , K d are proportional, integral and derivative gains respectively;

[0092] S83, based on the adjustment signal u(t), the pressure regulating valve of the spraying device is adjusted in real time to make the working pressure of the spraying module gradually approach the target pressure P r ;

[0093] S84, real-time monitoring of the spraying width W through the width sensor and length sensor t and length L t , and compare it with the preset width W r and length L r Compare and calculate the width deviation ΔW = W r -W t and length deviation ΔL=L r -L t ;

[0094] S85, dynamically adjust the pressure target value P according to the width deviation and length deviation r , update the target value:

[0095] P r =P r +k w ΔW+k l ΔL;

[0096] Among them, k w and k l are the correction factors for width and length deviations, respectively;

[0097] S86, setting the calibration cycle T of the spraying device c , calibrate the spray device in each cycle, including cleaning the spray nozzle, checking the accuracy of the pressure sensor, and recalibrating the target pressure P r and the actual pressure P t Correction factor for

[0098] S87, the corrected pressure target value P r The real-time adjustment results are transmitted to the spray control module to make the spray width and length meet the preset requirements, and the correction log is recorded for subsequent optimization.

[0099] The beneficial effects of the present invention are:

[0100] (1) The present invention provides automated, intelligent and efficient processing of the sleeper marking process by combining deep learning models, dynamic optimization algorithms, multi-sensor fusion technology and intelligent control systems, realizes automatic sleeper recognition, dynamic marking optimization of curved sections, intelligent spray path planning and error self-learning optimization, and effectively improves the accuracy and adaptability of sleeper marking.

[0101] (2) The present invention achieves the effect of real-time monitoring and dynamic adjustment of labeling through real-time data monitoring and PID-based spray pressure control algorithm, combined with a remote intelligent control system, which significantly improves the efficiency and consistency of labeling and reduces the risk of manual operation.

[0102] (3) The present invention optimizes and adjusts the spraying parameters by using the historical marking data through self-learning optimization model and error classification analysis, thereby reducing the occurrence of repeated errors and improving the quality and intelligence level of the sleeper marking operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0104] Figure 1 This is a system overall framework diagram of a sleeper labeling method based on a deep learning algorithm proposed in the present invention. DETAILED DESCRIPTION

[0105] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0106] refer to Figure 1 A sleeper labeling method based on deep learning algorithm, characterized in that it comprises the following steps:

[0107] S1. Collect sleeper image data, extract features from the image through a convolutional neural network, use multi-layer convolution operations to obtain the position information of the edge and center of the sleeper, and use a non-maximum suppression algorithm to filter interference points to generate the initial coordinates of the sleeper position;

[0108] In this implementation, S1 specifically includes:

[0109] S11, collecting real-time image data of the track area through a camera device, inputting the image data into a preprocessing module, graying the image, removing irrelevant color information, and enhancing the edge features of the target area;

[0110] S12, perform Gaussian filtering on the preprocessed image to reduce noise interference:

[0111]

[0112] The size of the convolution kernel used is k×k, where k is an odd number, x and y are the coordinates of the pixel points, and σ is the standard deviation, which determines the degree of smoothing of the filter on the edge details of the sleeper;

[0113] S13, inputting the filtered image data into a convolutional neural network, and extracting features of the edge and center images of the sleeper by setting multiple convolutional layers:

[0114]

[0115] Among them, f ij is the convolution output, which represents the feature value of the current pixel, I(i+m,j+n) is the pixel value of the input image, and W mn is the convolution kernel weight, b is the bias value, and k is the convolution kernel size;

[0116] S14, performing a pooling operation on the convolution output result, using a maximum pooling method, by setting a fixed-size pooling window to reduce the size of the feature map while retaining the maximum value in each pooling window;

[0117] S15, processing the pooling result through a fully connected layer, mapping it to the target classification space, and generating a classification result to determine the probability of the sleeper position;

[0118] S16. Use the non-maximum suppression algorithm to filter the interference points on the classification results, set the threshold T, remove the detection boxes with scores lower than T, retain the sleeper position with the highest score, and output the final initial coordinates of the sleeper position.

[0119] S2. Based on the mileage measurement device, the photoelectric encoder and the inertial measurement unit are used for data fusion, and the travel position and posture of the equipment are corrected through the Kalman filter algorithm. At the same time, the laser ranging device is combined to compensate for the data error and obtain the real-time position information of the equipment;

[0120] In this implementation, S2 specifically includes:

[0121] S21, collecting the number of pulses per unit time during the travel of the device through a photoelectric encoder, and calculating the travel mileage of the device based on the number of pulses of the encoder;

[0122] S22, collecting acceleration data and angular velocity data of the device through an inertial measurement unit, and calculating the attitude angle of the device according to the angular velocity data;

[0123] S23, synchronously processing the data of the photoelectric encoder and the inertial measurement unit, and using the Kalman filter algorithm to correct the travel position and posture of the device, including predicting the device state and updating the state estimate;

[0124] S24, obtaining ranging data on the track in combination with a laser ranging device, and marking measurement points exceeding the threshold as abnormal according to a set ranging accuracy threshold;

[0125] S25, integrate the data of the photoelectric encoder, inertial measurement unit and laser ranging device to calculate the final device position:

[0126] X f =w 1 X e +w 2 X i +w 3 X l ;

[0127] Among them, X f is the final position, X e , X i , X l are the data of photoelectric encoder, inertial measurement unit and laser ranging, respectively. 1 、w 2 、w 3 is the weight of the corresponding data source, satisfying w 1 +w 2 +w 3 =1;

[0128] S26. Output the real-time location information of the device, including mileage, attitude angle, and coordinates of the marked abnormal area.

[0129] S3. Construct a dynamic curve section marking optimization algorithm to calculate the shortening of each marking point and dynamically adjust the spacing between marking points according to the track curve radius, transition curve parameters and equipment mileage;

[0130] In this implementation, S3 specifically includes:

[0131] S31. Collect track geometry parameters, including track curve radius R, transition curve length L h and the length of the straight line segment L s , and associate it with the mileage data D of the device;

[0132] S32. Calculate the shortening amount ΔL of the curve segment c The shortening of the transition curve ΔL h :

[0133]

[0134] Where, ΔL c is the shortening of the circular curve segment, ΔL h S is the shortening of the transition curve segment. 1 is the distance between the center lines of the two rails, R is the track curve radius, L c is the length of the circular curve, L h is the length of the transition curve;

[0135] S33. According to the travel position of the device, determine the type of track segment to which the current point belongs. If it is a transition curve segment, calculate the dynamic shortening amount of each point:

[0136]

[0137] Where, ΔL h,i is the shortening of the i-th point in the transition curve, d i is the distance from the starting point of the transition curve to the i-th point;

[0138] S34, based on the calculation results of the straight line segment, transition curve segment and circular curve segment, dynamically adjust the spacing of the sleeper marking points, the spacing of the marking points ΔD i Calculated according to the following rules:

[0139] ΔD i =D 0 -ΔL;

[0140] Where, ΔD i is the adjusted spacing of the ith annotation point, D 0 is the initial spacing, ΔL is the shortening of the corresponding segment;

[0141] S35, transmitting the dynamically adjusted marking point spacing to the spraying control module, updating the mileage position data of each marking point, and generating complete marking path data.

[0142] S4. Optimize the spraying path based on the reinforcement learning algorithm. By building a path planning model based on the deep Q network, train the spraying path of the equipment under different track characteristics, adjust the spraying start and stop points and path based on real-time track information, and generate the optimal spraying action sequence;

[0143] In this implementation, S4 specifically includes:

[0144] S41, collecting geometric parameters of the track environment and spraying requirement data, including the straight length, curve radius, transition curve length and spacing requirements of the marked points of the track, and constructing the state space of the spraying path;

[0145] S42. Design a reinforcement learning framework for the spraying path, define the state set as the combination of the current position of the device and the spraying state, define the action set as the spraying action of the device, define the reward function according to the path optimization degree of the device to complete the current spraying task, and the state transition function describes the path evolution process through the probability of the device reaching the next state through a certain action from the current state;

[0146] S43. Optimize the spraying path based on the deep Q network, and use the Q value update method to adjust the spraying strategy of the equipment. The Q value update method makes corrections by adjusting the difference between the current Q value and the maximum Q value of the subsequent state:

[0147] Q(s,a)=Q(s,a)+α[r+γmax a Q(s ′ ,a)-Q(s,a)];

[0148] Among them, Q(s,a) represents the Q value of the current state and action, α is the learning rate, which is used to control the update amplitude, r is the immediate reward of the current action, γ is the discount factor, which is used to balance the weight of short-term rewards and long-term benefits, and max a Q(s ′ ,a) represents the action value with the highest expected reward in the subsequent state;

[0149] S44. The Q value function is fitted by a deep neural network. The input of the neural network is the combined state of the current position of the device and the track characteristics. The output is the Q value of all possible spraying actions. The mean square error is used as the loss function to minimize the difference between the predicted Q value and the target Q value:

[0150] L(θ)=E[(r+γmax a Q(s′,a;θ′)-Q(s,a;θ)) 2 ];

[0151] Among them, θ represents the parameters of the current deep neural network, θ′ is the parameter of the target network, and the loss function is optimized through multiple iterations to enhance the adaptability of the model to complex orbit paths;

[0152] S45, through multiple rounds of training iterations, dynamically update the Q value according to the characteristics of the track environment and the equipment spraying task, and generate the optimal spraying action sequence;

[0153] S46, combining the track information obtained by the device in real time, dynamically adjusting the spraying start and stop points and the path in the path planning model;

[0154] S5. Combine multi-sensor fusion technology, use the data from mileage sensors, laser ranging devices and attitude sensors, adopt weighted average method to fuse the measurement results of each data source, detect abnormal points on the fusion results, and adjust the position and spacing of the spray nozzles in real time;

[0155] In this implementation, S5 specifically includes:

[0156] S51, collecting travel distance data of the device through a mileage sensor, and dynamically calibrating the device in combination with operating parameters of the device to record the real-time travel distance;

[0157] S52, obtaining real-time distance measurement data between the track and the equipment through a laser distance measuring device, and performing outlier detection in combination with a deep learning model;

[0158] S53, collect the acceleration and angular velocity data of the device through the attitude sensor, and calculate the attitude angle after real-time calibration based on the inertial navigation algorithm:

[0159] θ c =θ+k a ·a t ;

[0160] Among them, θ c is the calibrated attitude angle, θ is the initial attitude angle, k a is the acceleration correction factor, a t is the current acceleration;

[0161] S54, integrates the data of the mileage sensor, laser ranging device and attitude sensor, and uses a dynamic fusion algorithm to calculate the final position of the device:

[0162]

[0163] Among them, X f is the final position, X is the current device position, is the weight ratio of laser ranging and mileage sensor, D t It is the distance measurement data after dynamic calibration;

[0164] S55, combined with track characteristics, outlier detection is performed on sensor fusion data, and the outliers are marked and processed using a dynamic threshold adjustment mechanism;

[0165] S56. Adjust the position of the spray nozzle and the spacing between marking points in real time according to the sensor fusion results, generate dynamic spray instructions, and synchronize the adjusted spray instructions to the spray control module through the intelligent decision-making module on the device side.

[0166] S6. Using the remote control system, the annotation device is connected via the narrowband communication module, and the handheld terminal displays the device operation status;

[0167] S7, based on the built-in self-learning optimization model of the equipment, collect historical annotation data, use the support vector machine algorithm to classify and analyze the spraying errors, and dynamically adjust the spraying parameters according to the classification results;

[0168] In this implementation, S7 specifically includes:

[0169] S71. Collecting real-time annotation data generated by the device during the spraying process, including the actual position, spacing, and spraying width of the annotation points, and building an annotation error sample set in combination with historical annotation data;

[0170] S72, extracting features from the collected error samples, wherein the extracted features include the marking point offset Δx, the marking point spacing deviation Δd, and the spraying width deviation Δw of the sleeper;

[0171] S73. Based on the support vector machine algorithm, the extracted error samples are classified and analyzed to build a classification model:

[0172]

[0173] Among them, f(x) is the classification decision function, α i is the Lagrange multiplier, y i is the category label of the sample, x i and x are support vectors and input samples respectively, K(x i ,x) is the kernel function, b is the bias;

[0174] S74, dividing the error samples into normal areas and abnormal areas according to the classification results, and extracting error correction parameters through error features of the abnormal areas;

[0175] S75, using error correction parameters to dynamically adjust spraying parameters, including spraying spacing Δd s , spraying width Δw s and spraying time t s :

[0176] Δd s =Δd-k d ·e d ;

[0177] Δw s =Δw-k w ·e w ;

[0178] t s =tk t ·e t ;

[0179] Where, Δd s , Δw s ,t s are the adjusted spraying spacing, width and time respectively, Δd, Δw, t are the initial spraying parameters, e d 、e w 、e t is the corresponding error value, k d , k w , k t is the error correction coefficient;

[0180] S76, transmitting the adjusted spraying parameters to the equipment spraying control module in real time, and updating the equipment spraying instructions to correct the errors in the current operation.

[0181] S8. Through real-time pressure monitoring of the spraying module, the working pressure of the spraying device is adjusted based on the PID control algorithm so that the spraying width and length meet the preset requirements, and the spraying device is periodically calibrated.

[0182] In this implementation, S8 specifically includes:

[0183] S81, collect the working pressure data of the spraying device in real time through the pressure sensor in the spraying module, and record the current pressure P t and target pressure P r ;

[0184] S82, input the collected pressure data into the PID control algorithm, and calculate the adjustment signal according to the pressure error:

[0185]

[0186] Where u(t) is the adjustment signal, e(t) = P r -P t is the current pressure error, K p , K i , K d are proportional, integral and derivative gains respectively;

[0187] S83, based on the adjustment signal u(t), the pressure regulating valve of the spraying device is adjusted in real time to make the working pressure of the spraying module gradually approach the target pressure P r ;

[0188] S84, real-time monitoring of the spraying width W through the width sensor and length sensor t and length L t , and compare it with the preset width W r and length L r Compare and calculate the width deviation ΔW = W r -W t and length deviation ΔL=L r -L t ;

[0189] S85, dynamically adjust the pressure target value P according to the width deviation and length deviation r , update the target value:

[0190] P r =P r +k w ΔW+k l ΔL;

[0191] Among them, k w and k l are the correction factors for width and length deviations, respectively;

[0192] S86, setting the calibration cycle T of the spraying device c , calibrate the spray device in each cycle, including cleaning the spray nozzle, checking the accuracy of the pressure sensor, and recalibrating the target pressure P r and the actual pressure P t Correction factor for

[0193] S87, the corrected pressure target value P r The real-time adjustment results are transmitted to the spray control module to make the spray width and length meet the preset requirements, and the correction log is recorded for subsequent optimization.

[0194] Embodiment 1:

[0195] In order to verify the feasibility and practical effect of the present invention, the present invention is applied to the sleeper marking task of a certain city railway freight line. The line is 2.5 kilometers long, including 2 gentle curve sections, 3 circular curve sections with different curvatures and multiple straight sections. The construction goal is to accurately mark the sleepers of the rails on both sides. A total of 3,920 sleepers need to be marked, and the spacing between the sleeper marking points should be controlled within the range of ±1 mm.

[0196] The construction environment is complex, and the accuracy of sleeper marking in curved sections is extremely high. However, traditional manual methods face many challenges such as low efficiency, high error, and high safety risks. In this scenario, manual marking requires the collaboration of 10 people, takes more than 10 hours, and the marking error is more than ±5 mm. Especially in gentle curve sections and areas with large curvature changes, the spacing between marking points is prone to significant deviations, which cannot meet the high-precision requirements. In response to the above problems, the sleeper marking method based on deep learning algorithm described in the present invention is adopted to comprehensively improve the marking efficiency and accuracy.

[0197] During the construction preparation phase, sleeper image data and track geometry parameters were first collected, including sleeper distribution, track curve radius, and transition curve length. The collected sleeper images were processed using a deep learning model, and sleeper features were extracted through a convolutional neural network to accurately identify sleeper positions and generate coordinates for initial annotation points. Subsequently, a dynamic curve segment annotation optimization algorithm was constructed based on track geometry parameters, and the spacing between annotation points was dynamically adjusted in combination with real-time mileage data to keep it consistent between different curvature segments. In a curve with a curvature radius of 300 meters, the spacing between annotation points was stable within the range of 568 mm ± 0.8 mm, while the spacing error of the traditional method was as high as ± 5 mm.

[0198] During the construction process, the spraying path planning is implemented based on the reinforcement learning algorithm. The equipment optimizes the spraying path through the deep Q network, completing the marking at a speed of 3 km / h in the straight section, and the marking speed of the curved section is 2.7 km / h. Compared with traditional manual operation, the overall construction efficiency is improved by 2.8 times. Real-time path optimization effectively reduces redundant actions and ensures that the spraying path is compact and efficient. At the same time, the working pressure of the spraying module is monitored in real time by the built-in pressure sensor, and the spraying pressure is dynamically adjusted in combination with the PID control algorithm to keep the spraying width within the range of 15 mm ± 1 mm, which is significantly better than the 3 mm error in the traditional method.

[0199] In order to further verify the accuracy of the present invention, multi-sensor fusion technology was used to detect the marking results after the construction was completed. By fusing the data of the photoelectric encoder, laser ranging device and inertial measurement unit, the position of the marking point was measured with high precision. The test results show that the longitudinal error between two points of a single sleeper is controlled within 0.04 mm, and the cumulative error of the sleeper marking points is less than 1.2 mm / km, while the cumulative error of the traditional method is usually more than 6 mm.

[0200] In the labeling operation, the self-learning optimization model described in the present invention plays an important role. The equipment classifies the labeling errors in real time through the support vector machine algorithm and dynamically adjusts the spraying parameters according to the classification results. In the initial stage, the spraying width deviation is ±1.5 mm. Through self-learning optimization, the spraying width deviation is gradually reduced to ±0.5 mm. For the optimization of the labeling point spacing of the curve segment, by adjusting the spraying time and the nozzle position, the labeling point spacing error is reduced from the initial ±3 mm to ±0.8 mm, which significantly improves the labeling accuracy.

[0201] In addition, the present invention ensures the safety and controllability of the construction process through a remote intelligent control system. Operators can monitor the operating status of the equipment in real time at the remote control center, including the travel speed, the number of marking points, and the adjustment of spraying parameters, avoiding the safety risks of personnel approaching the track construction. After the construction is completed, the spraying module is periodically calibrated using the correction log data. In this construction case, the actual application data of the present invention is as follows:

[0202] Table 1 Performance comparison of different sleeper marking methods

[0203]

[0204]

[0205] As can be seen from the above table, in the embodiment, the traditional sleeper marking method requires more than 10 hours to complete the track marking operation, while the intelligent sleeper marking system of the present invention shortens the construction time to 3 hours, and the efficiency is improved by nearly 3 times. During the marking process, the spacing error of the marking points of the traditional method is as high as ±5 mm, while the method of the present invention controls the spacing error of the marking points within ±0.8 mm through a dynamic optimization algorithm and a self-learning mechanism. At the same time, the spray width deviation is reduced from ±3 mm of the traditional method to ±0.5 mm, and the marking accuracy is significantly improved.

[0206] In the specific construction of curved sections, it was found through testing that the method of the present invention can dynamically adjust the spacing of the marking points under complex track geometry conditions, avoiding the problem of uneven marking caused by changes in curvature. After the construction is completed, the marking results are tested using multi-sensor fusion technology. The longitudinal error of two points on a single sleeper is less than 0.04 mm, and the cumulative position error of the marking points is less than 1.2 mm / km, while the cumulative error of the traditional method is usually more than 6 mm.

[0207] In addition, the remote intelligent control system of the present invention significantly improves construction safety. The traditional method requires 10 people to work manually at close range, which poses a high safety risk. However, the method of the present invention only requires one operator to remotely control the equipment, avoiding exposure of personnel to construction risks.

[0208] These data show that the present invention is not only significantly superior to traditional methods in terms of efficiency and accuracy, but also effectively solves the construction problems in complex track environments, while reducing the safety risks and labor intensity caused by manual operations. Through these improvements, the present invention has achieved technological innovation in the field of sleeper marking and met the comprehensive requirements of track construction for efficiency, accuracy and safety.

[0209] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A sleeper labeling method based on deep learning algorithm, characterized in that: The steps include: S1. Collect sleeper image data, extract features from the image through a convolutional neural network, use multi-layer convolution operations to obtain the position information of the edge and center of the sleeper, and use a non-maximum suppression algorithm to filter interference points to generate the initial coordinates of the sleeper position; S2. Based on the mileage measurement device, the photoelectric encoder and the inertial measurement unit are used for data fusion, and the travel position and posture of the equipment are corrected through the Kalman filter algorithm. At the same time, the laser ranging device is combined to compensate for the data error and obtain the real-time position information of the equipment; S3. Construct a dynamic curve section marking optimization algorithm to calculate the shortening of each marking point and dynamically adjust the spacing between marking points according to the track curve radius, transition curve parameters and equipment mileage; S4. Optimize the spraying path based on the reinforcement learning algorithm. By building a path planning model based on the deep Q network, train the spraying path of the equipment under different track characteristics, adjust the spraying start and stop points and path based on real-time track information, and generate the optimal spraying action sequence; S5. Combine multi-sensor fusion technology, use the data from mileage sensors, laser ranging devices and attitude sensors, adopt weighted average method to fuse the measurement results of each data source, detect abnormal points on the fusion results, and adjust the position and spacing of the spray nozzles in real time; S6. Using the remote control system, the annotation device is connected via the narrowband communication module, and the handheld terminal displays the device operation status; S7, based on the built-in self-learning optimization model of the equipment, collect historical annotation data, use the support vector machine algorithm to classify and analyze the spraying errors, and dynamically adjust the spraying parameters according to the classification results; S8. Through real-time pressure monitoring of the spraying module, the working pressure of the spraying device is adjusted based on the PID control algorithm so that the spraying width and length meet the preset requirements, and the spraying device is periodically calibrated.

2. A sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S1 specifically includes: S11, collecting real-time image data of the track area through a camera device, inputting the image data into a preprocessing module, graying the image, removing irrelevant color information, and enhancing the edge features of the target area; S12, perform Gaussian filtering on the preprocessed image to reduce noise interference: The size of the convolution kernel used is k×k, where k is an odd number, x and y are the coordinates of the pixel points, and σ is the standard deviation, which determines the degree of smoothing of the filter on the edge details of the sleeper; S13, inputting the filtered image data into a convolutional neural network, and extracting features of the edge and center images of the sleeper by setting multiple convolutional layers: Among them, f ij is the convolution output, which represents the feature value of the current pixel, I(i+m,j+n) is the pixel value of the input image, and W mn is the convolution kernel weight, b is the bias value, and k is the convolution kernel size; S14, performing a pooling operation on the convolution output result, using a maximum pooling method, by setting a fixed-size pooling window to reduce the size of the feature map while retaining the maximum value in each pooling window; S15, processing the pooling result through a fully connected layer, mapping it to the target classification space, and generating a classification result to determine the probability of the sleeper position; S16. Use the non-maximum suppression algorithm to filter the interference points on the classification results, set the threshold T, remove the detection boxes with scores lower than T, retain the sleeper position with the highest score, and output the final initial coordinates of the sleeper position.

3. A sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S2 specifically includes: S21, collecting the number of pulses per unit time during the travel of the device through a photoelectric encoder, and calculating the travel mileage of the device based on the number of pulses of the encoder; S22, collecting acceleration data and angular velocity data of the device through an inertial measurement unit, and calculating the attitude angle of the device according to the angular velocity data; S23, synchronously processing the data of the photoelectric encoder and the inertial measurement unit, and using the Kalman filter algorithm to correct the travel position and posture of the device, including predicting the device state and updating the state estimate; S24, obtaining ranging data on the track in combination with a laser ranging device, and marking measurement points exceeding the threshold as abnormal according to a set ranging accuracy threshold; S25, integrate the data of the photoelectric encoder, inertial measurement unit and laser ranging device to calculate the final device position: X f =w1X e +w2X i +w3X l ; Among them, X f is the final position, X e , X i , X l are the data of the photoelectric encoder, inertial measurement unit and laser ranging, respectively. w1, w2 and w3 are the weights of the corresponding data sources, satisfying w1+w2+w3=1; S26. Output the real-time location information of the device, including mileage, attitude angle, and coordinates of the marked abnormal area.

4. A sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S3 specifically includes: S31. Collect track geometry parameters, including track curve radius R, transition curve length L h and the length of the straight line segment L s , and associate it with the mileage data D of the device; S32. Calculate the shortening amount ΔL of the curve segment c The shortening of the transition curve ΔL h : Where, ΔL c is the shortening of the circular curve segment, ΔL h The shortening of the easing curve section, S1 is the distance between the center lines of the two rails, R is the track curve radius, L is c is the length of the circular curve, L h is the length of the transition curve; S33. According to the travel position of the device, determine the type of track segment to which the current point belongs. If it is a transition curve segment, calculate the dynamic shortening amount of each point: Where, ΔL h,i is the shortening of the i-th point in the transition curve, d i is the distance from the starting point of the transition curve to the i-th point; S34, based on the calculation results of the straight line segment, transition curve segment and circular curve segment, dynamically adjust the spacing of the sleeper marking points, the spacing of the marking points ΔD i Calculated according to the following rules: ΔD i =D0-ΔL; Where, ΔD i is the adjusted spacing of the i-th annotation point, D0 is the initial spacing, and ΔL is the shortening of the corresponding segment; S35, transmitting the dynamically adjusted marking point spacing to the spraying control module, updating the mileage position data of each marking point, and generating complete marking path data.

5. The sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S4 specifically includes: S41, collecting geometric parameters of the track environment and spraying requirement data, including the straight length, curve radius, transition curve length and spacing requirements of the marked points of the track, and constructing the state space of the spraying path; S42. Design a reinforcement learning framework for the spraying path, define the state set as the combination of the current position of the device and the spraying state, define the action set as the spraying action of the device, define the reward function according to the path optimization degree of the device to complete the current spraying task, and the state transition function describes the path evolution process through the probability of the device reaching the next state through a certain action from the current state; S43. Optimize the spraying path based on the deep Q network, and use the Q value update method to adjust the spraying strategy of the equipment. The Q value update method makes corrections by adjusting the difference between the current Q value and the maximum Q value of the subsequent state: Q(s,a)=Q(s,a)+α[r+γmax a Q(s′,a)-Q(s,a)]; Among them, Q(s,a) represents the Q value of the current state and action, α is the learning rate, which is used to control the update amplitude, r is the immediate reward of the current action, γ is the discount factor, which is used to balance the weight of short-term rewards and long-term benefits, and max a Q(s′,a) represents the action value with the highest expected return in the subsequent state; S44. The Q value function is fitted by a deep neural network. The input of the neural network is the combined state of the current position of the device and the track characteristics. The output is the Q value of all possible spraying actions. The mean square error is used as the loss function to minimize the difference between the predicted Q value and the target Q value: L(θ)=E[(r+γmax a Q(s′,a;θ′)-Q(s,a;θ)) 2 ]; Among them, θ represents the parameters of the current deep neural network, θ′ is the parameter of the target network, and the loss function is optimized through multiple iterations to enhance the adaptability of the model to complex orbit paths; S45, through multiple rounds of training iterations, dynamically update the Q value according to the characteristics of the track environment and the equipment spraying task, and generate the optimal spraying action sequence; S46. Combined with the track information obtained by the device in real time, the spraying start and stop points and the path are dynamically adjusted in the path planning model.

6. A sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S5 specifically includes: S51, collecting travel distance data of the device through a mileage sensor, and dynamically calibrating the device in combination with operating parameters of the device to record the real-time travel distance; S52, obtaining real-time distance measurement data between the track and the equipment through a laser distance measuring device, and performing outlier detection in combination with a deep learning model; S53, collect the acceleration and angular velocity data of the device through the attitude sensor, and calculate the attitude angle after real-time calibration based on the inertial navigation algorithm: i c =θ+k a ·a t ; Among them, θ c is the calibrated attitude angle, θ is the initial attitude angle, k a is the acceleration correction factor, a t is the current acceleration; S54, integrates the data of the mileage sensor, laser ranging device and attitude sensor, and uses a dynamic fusion algorithm to calculate the final position of the device: Among them, X f is the final position, X is the current device position, is the weight ratio of laser ranging and mileage sensor, D t It is the distance measurement data after dynamic calibration; S55, combined with track characteristics, outlier detection is performed on sensor fusion data, and the outliers are marked and processed using a dynamic threshold adjustment mechanism; S56. Adjust the position of the spray nozzle and the spacing between marking points in real time according to the sensor fusion results, generate dynamic spray instructions, and synchronize the adjusted spray instructions to the spray control module through the intelligent decision-making module on the device side.

7. The sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S7 specifically includes: S71. Collecting real-time annotation data generated by the device during the spraying process, including the actual position, spacing, and spraying width of the annotation points, and building an annotation error sample set in combination with historical annotation data; S72, extracting features from the collected error samples, wherein the extracted features include the marking point offset Δx, the marking point spacing deviation Δd, and the spraying width deviation Δw of the sleeper; S73. Based on the support vector machine algorithm, the extracted error samples are classified and analyzed to build a classification model: Among them, f(x) is the classification decision function, α i is the Lagrange multiplier, y i is the category label of the sample, x i and x are support vectors and input samples respectively, K(x i ,x) is the kernel function, b is the bias; S74, dividing the error samples into normal areas and abnormal areas according to the classification results, and extracting error correction parameters through error features of the abnormal areas; S75, using error correction parameters to dynamically adjust spraying parameters, including spraying spacing Δd s , spraying width Δw s and spraying time t s : Δd s =Δd-k d ·e d ; Δw s =Δw-k w ·and w ; t s =t-k t ·e t ; Where, Δd s , Δw s ,t s are the adjusted spraying spacing, width and time respectively, Δd, Δw, t are the initial spraying parameters, e d 、e w 、e t is the corresponding error value, k d , k w , k t is the error correction coefficient; S76, transmitting the adjusted spraying parameters to the equipment spraying control module in real time, and updating the equipment spraying instructions to correct the errors in the current operation.

8. The sleeper labeling method based on deep learning algorithm according to claim 1, characterized in that: The S8 specifically includes: S81, collect the working pressure data of the spraying device in real time through the pressure sensor in the spraying module, and record the current pressure P t and target pressure P r ; S82, input the collected pressure data into the PID control algorithm, and calculate the adjustment signal according to the pressure error: Where u(t) is the adjustment signal, e(t) = P r -P t is the current pressure error, K p , K i , K d are proportional, integral and derivative gains respectively; S83, based on the adjustment signal u(t), the pressure regulating valve of the spraying device is adjusted in real time to make the working pressure of the spraying module gradually approach the target pressure P r ; S84, real-time monitoring of the spraying width W through the width sensor and length sensor t and length L t , and compare it with the preset width W r and length L r Compare and calculate the width deviation ΔW = W r -W t and length deviation ΔL=L r -L t ; S85, dynamically adjust the pressure target value P according to the width deviation and length deviation r , update the target value: P r =P r +k w ·ΔW+k l ·ΔL; Among them, k w and k l are the correction factors for width and length deviations, respectively; S86, setting the calibration period T of the spraying device c , calibrate the spray device in each cycle, including cleaning the spray nozzle, checking the accuracy of the pressure sensor, and recalibrating the target pressure P r and the actual pressure P t Correction factor for S87, the corrected pressure target value P r The real-time adjustment results are transmitted to the spray control module to make the spray width and length meet the preset requirements, and the correction log is recorded for subsequent optimization.

Images