A method and system for intelligent speed control of traction winch for paving on slopes.

By combining multi-level signal preprocessing and long short-term memory networks with engineering parameters to calculate correction factors, the problem of dynamic adjustment of traction winch speed control in slope paving construction was solved, achieving precise speed control in strong vibration and noise environments, and ensuring balanced construction progress and safety.

CN123061883APending Publication Date: 2026-11-10LIAONING YONGSHENG RUNDA CONSTRUCTION ENGINEERING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611246708.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing technologies, the speed control of the traction winch during slope paving construction cannot be dynamically adjusted, lacks real-time response to changes in the load, position, and slope inclination of the feeding trolley, and the control accuracy decreases under strong vibration and noise environments, and lacks global optimization capabilities.

Method used

By employing multi-level signal preprocessing and a long short-time memory (LSTM) network, combined with engineering parameters to calculate correction factors, intelligent control of the traction winch speed is achieved. Specific steps include signal filtering, noise reduction, outlier removal, time synchronization, spatial alignment, and engineering parameter calculation. Finally, the optimal speed command is output through the LSTM network.

Benefits of technology

It improved the quality of sensor signals, enabled optimal speed decision-making based on global engineering information, reduced computational latency, ensured a balance between control accuracy and construction progress, avoided misjudgment of sudden slope changes, and improved construction quality and safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123061883A_ABST
    Figure CN123061883A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent speed control method and system for a traction winch in slope paving, relating to the field of engineering control technology. The method includes: acquiring raw signals from the feeding trolley in real time at a preset sampling period; performing multi-level intelligent signal preprocessing on the raw signals; reading slope engineering parameters and calculating a global speed adjustment factor, a finishing speed attenuation factor, a local slope correction factor, a load correction factor, and a position period compensation factor based on the preprocessed position and inclination signals; inputting the preprocessed signals and the five correction factors into a pre-trained long short-term memory network to directly output the optimal target speed command for the current moment; sending the optimal target speed command to a frequency converter to adjust the drive motor speed; repeating the above steps at the next sampling moment to achieve real-time rolling updates of the target speed command. This invention can automatically adjust the traction winch's operating speed according to real-time working conditions and global engineering information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of engineering control technology, specifically to an intelligent speed control method and system for traction winches used in slope paving. Background Technology

[0002] In slope paving construction, a traction winch is used to pull the feeding trolley or paver along the slope. Taking the construction of pumped storage power station dams as an example, steep and long slopes with a slope ratio of 1.1 to 1.3 and a single slope length of more than 120 m are commonly used. The total weight of the construction equipment reaches 120 tons, and the travel speed is required to be precisely adjustable within the range of 0 to 10 m / min.

[0003] In existing technologies, the speed control of traction winches mostly relies on manual operation or simple PID constant speed control, which has the following prominent problems: it cannot dynamically adjust the traction speed according to the real-time load changes, running position changes, and slope inclination changes of the feeding trolley; existing control strategies are mostly simple setpoint tracking control, lacking the ability to optimize the overall conditions of slope construction; existing systems have a single signal processing method, lacking effective preprocessing means for sensor signals in strong vibration and high noise environments, resulting in decreased control accuracy; existing control methods do not consider the impact of the overall parameters of the slope project and the cumulative construction progress on traction demand, and lack intelligent decision-making capabilities based on global project information.

[0004] Therefore, there is an urgent need for an intelligent speed control method and system for traction winches that integrates advanced signal preprocessing, adaptive correction of engineering parameters, and intelligent decision-making based on neural networks. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing an intelligent speed control method and system for traction winches in paving on slopes.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent speed control of a traction winch for paving on a slope includes the following steps: Step S1: Collect raw signals in real time at a preset sampling period. The raw signals include the current load mass signal of the feeding trolley, the current position signal, the current slope inclination signal, the current actual speed signal of the traction winch, the current tension signal of the wire rope, and the acceleration signal of the feeding trolley along the slope direction. Step S2: Perform multi-level intelligent signal preprocessing on the original signal to obtain a preprocessed signal. The preprocessed signal includes a preprocessed load mass signal, a preprocessed position signal, a preprocessed tilt angle signal, a preprocessed velocity signal, a preprocessed tension signal, and a preprocessed acceleration signal. Step S3: Read the engineering parameters of the slope project. Based on the preprocessed position signal, the preprocessed inclination angle signal and the engineering parameters, calculate the global speed adjustment factor, the final speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor. Step S4: Input the preprocessed signal, along with the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor, into the pre-trained long short-term memory network. The long short-term memory network then directly outputs the optimal target speed command for the current moment. Step S5: The optimal target speed command is sent directly to the frequency converter of the traction winch, and the frequency converter adjusts the speed of the traction winch drive motor according to the optimal target speed command. Step S6: At the next sampling time, repeat steps S1 to S5 to realize the real-time rolling update of the target speed command of the traction winch.

[0007] Furthermore, step S2 specifically includes the following sub-steps: Step S2.1: Perform anti-aliasing filtering on the original signal using a hardware low-pass filter to filter out high-frequency noise higher than the Nyquist frequency; Step S2.2: The signal after hardware filtering is processed by wavelet transform threshold denoising to obtain the denoised signal; Step S2.3: The denoised signal is processed segment by segment using a sliding window method. The Grubbs statistic of each data point in the window is calculated. When the Grubbs statistic exceeds the adaptively adjusted Grubbs threshold, the data point is determined to be an outlier and is removed. The removed data points are filled in using cubic spline interpolation. Step S2.4: Unify the data from each sensor to the same time reference using linear interpolation, perform spatial alignment, and transform them into a unified feed trolley centroid coordinate system; Step S2.5: Perform sliding window normalization on the synchronized and aligned signal to obtain the preprocessed signal.

[0008] Furthermore, in step S3, the engineering parameters include at least the total length of the inclined track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch.

[0009] Furthermore, step S3 specifically includes the following sub-steps: Step S3.1: Based on the preprocessed location signal and slope distribution data of each segment, calculate the cumulative design workload of the remaining road segment from the current position to the endpoint and the total design workload of the entire slope, and at the same time calculate the remaining distance from the current position to the endpoint; Step S3.2: Calculate the global speed adjustment factor based on the remaining workload, total workload, remaining distance, and total length of the ramp track; Step S3.3: Calculate the final velocity attenuation factor based on the preprocessed position signal and the total length of the ramp track; Step S3.4: Calculate the local slope correction factor based on the preprocessed tilt angle signal; Step S3.5: Calculate the load correction factor based on the preprocessed load mass signal, the empty load mass of the feeding trolley, and the maximum allowable load mass; Step S3.6: Calculate the position period compensation factor based on the preprocessed position signal and the total length of the ramp track.

[0010] Further, in step S3.1, the cumulative design workload of the remaining road segments from the current position to the endpoint is the weighted sum of the products of the length of each remaining road segment, its corresponding sine value of the inclination angle, and the design load mass, multiplied by the gravitational acceleration; the total design workload of the entire slope is the weighted sum of the products of the length of all road segments on the entire slope, its corresponding sine value of the inclination angle, and the design load mass, multiplied by the gravitational acceleration; the remaining distance is the total length of the slope track minus the current position; In step S3.2, the global speed adjustment factor is 1 plus the workload distance deviation adjustment coefficient multiplied by the difference between the remaining workload ratio and the remaining distance ratio; the remaining workload ratio is the remaining workload divided by the total workload, and the remaining distance ratio is the remaining distance divided by the total length of the ramp track. In step S3.3, when the position corresponding to the preprocessed position signal does not exceed 85% of the total length of the ramp track, the closing speed attenuation factor is 1; when the position corresponding to the preprocessed position signal exceeds 85% of the total length of the ramp track, the closing speed attenuation factor is 1 minus the closing attenuation coefficient multiplied by the square of the normalized excess ratio, where the normalized excess ratio is the difference between the current position percentage and 0.85 divided by 0.15. In step S3.4, the local slope correction factor is a negative exponent of the natural constant. The exponent is the sum of the slope amplitude influence coefficient multiplied by the square of the current inclination angle sine and the slope change rate influence coefficient multiplied by the absolute value of the difference between the current inclination angle and the inclination angle at the previous moment, divided by the sampling period. In step S3.5, the load correction factor is 1 minus the load influence coefficient multiplied by the load influence nonlinear exponent of the normalized load ratio. The normalized load ratio is the difference between the current load mass and the unloaded mass divided by the difference between the maximum allowable load mass and the unloaded mass. When the current load mass is not greater than the unloaded mass, the load correction factor is 1. When the current load mass is not less than the maximum allowable load mass, the load correction factor is 1 minus the load influence coefficient. In step S3.6, the position period compensation factor is 1 plus the sum of the first harmonic term and the second harmonic term; the first harmonic term is the sine function value calculated by multiplying the first harmonic amplitude coefficient by 2π times the current position divided by the total length plus the first initial phase as the independent variable; the second harmonic term is the cosine function value calculated by multiplying the second harmonic amplitude coefficient by 4π times the current position divided by the total length plus the second initial phase as the independent variable.

[0011] Further, in step S4, the input vector of the Long Short-Term Memory network has a dimension of 11, and is composed of 11 features: preprocessed load mass signal, preprocessed position signal, preprocessed tilt angle signal, preprocessed velocity signal, preprocessed tension signal, preprocessed acceleration signal, and the global velocity adjustment factor, the tail velocity attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor; the Long Short-Term Memory network simultaneously receives the historical input sequence of the current moment and the previous 15 consecutive moments, forming a 15x11 temporal input tensor; The Long Short-Term Memory (LSTM) network comprises two stacked LSTM hidden layers and a fully connected output layer. The first LSTM hidden layer contains 64 hidden units, which receive the temporal input tensor and output the complete time series. The second LSTM hidden layer contains 32 hidden units, which receive the complete sequence output by the first layer and output only the hidden state of the last time step. Layer normalization and Dropout regularization with a dropout rate of 0.2 are set between the two LSTM hidden layers. The fully connected output layer maps the 32-dimensional hidden state to a 1-dimensional velocity command output.

[0012] Furthermore, in step S4, the Long Short-Term Memory Network is trained using supervised learning, and the training labels are human expert speed commands recorded in historical construction data or verified historical optimal speed commands; the overall loss function of the Long Short-Term Memory Network is a weighted sum of speed tracking loss, tension safety loss, smoothing loss, load-speed coordination loss and physical feasibility loss.

[0013] Furthermore, the speed tracking loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network and the corresponding training label is calculated, and then the average value of all samples is taken. The tension safety loss is calculated as follows: the static tension component is calculated based on the current load mass, gravitational acceleration, current tilt angle sine value, and tension safety factor. Then, the tension dynamic fluctuation information extracted from the internal state of the long short-term memory network is superimposed to obtain the tension estimate. Within a training batch, the tension estimate of each sample is compared with the safety tension limit. If the tension estimate does not exceed the safety tension limit, the loss is 0. If it exceeds the limit, the square of the excess is taken, and then the average of all samples is taken. The smoothing loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network at the current moment and the speed command output by the network at the previous moment is calculated, and then the average value of all samples is taken. The load-speed coordination loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network and the product of the current load weight divided by the maximum allowable load weight, minus the load-speed coordination target value, is calculated, and then the average value of all samples is taken. The physical feasibility loss is calculated as follows: within a training batch, the network output speed command of each sample is checked. If the speed command exceeds the maximum allowable speed, the square of the excess is taken. If the speed command is lower than the minimum allowable speed, the square of the lower part is taken. If the speed command is between the maximum allowable speed and the minimum allowable speed, the loss is 0. The above losses of all samples are summed and then averaged.

[0014] An intelligent speed control system for a traction winch in paving on a slope, used to implement any one of the intelligent speed control methods for a traction winch in paving on a slope, comprising: The multi-source sensor signal acquisition module is used to acquire in real time the load mass signal, position signal, tilt angle signal, speed signal of the traction winch, wire rope tension signal and acceleration signal of the feeding trolley along the slope direction; The engineering parameter storage module is used to store the total length of the inclined track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch. The signal processing and intelligent decision-making module is connected to the multi-source sensor signal acquisition module and the engineering parameter storage module, respectively. It is used to preprocess the acquired raw signals, calculate the correction factor based on the engineering parameters and the preprocessed position and tilt signals, and output the optimal target speed command using the built-in long short-term memory network. The traction speed execution module, connected to the signal processing and intelligent decision-making module, is used to receive the optimal target speed command and adjust the speed of the traction winch drive motor.

[0015] Furthermore, the signal processing and intelligent decision-making module specifically includes: The data acquisition submodule is connected to each sensor of the multi-source sensing signal acquisition module and is used to synchronously acquire the raw signals of each sensor at a preset sampling period. The intelligent signal preprocessing submodule is connected to the data acquisition submodule and is used to perform hardware filtering, wavelet transform denoising, outlier removal and interpolation filling, time synchronization and spatial alignment and normalization on the original signal, and output the preprocessed signal. The engineering parameter correction submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter storage module, respectively, and is used to calculate the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor based on the preprocessed position signal, the preprocessed tilt angle signal and the engineering parameters. The deep decision submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter correction submodule, respectively. The deep decision submodule has a built-in long short-term memory network, which is used to directly output the optimal target speed command at the current moment based on the preprocessed signal and the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position period compensation factor. The drive control interface submodule, connected to the depth decision submodule, is used to send the optimal target speed command to the traction speed execution module.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention overcomes the adverse effects of strong vibration and noise on sensor signals at the paving site by introducing multi-level intelligent signal preprocessing, thereby improving the quality of input data. The mechanism of adaptive adjustment of the Grubbs critical value according to the degree of data fluctuation avoids misjudging real working condition changes such as sudden slope changes as outliers.

[0017] 2. This invention calculates five correction factors based on engineering parameters and uses these five factors as input features of a long short-term memory network, enabling the network to perceive the overall state and cumulative construction progress of the slope project, thus achieving optimal speed decision based on global project information.

[0018] 3. This invention constructs an end-to-end long short-term memory network deep decision model, which takes multi-source sensor signals and engineering correction factors as inputs and directly outputs the optimal target speed command, thereby achieving end-to-end optimal decision-making and reducing computational latency.

[0019] 4. This invention designs a comprehensive loss function that includes speed tracking loss, tension safety loss, smoothing loss, load speed coordination loss, and physical feasibility loss, ensuring that the speed command output by the network achieves an optimal balance among multiple targets, tracking the training label, ensuring tension safety, and simultaneously satisfying smoothness and physical feasibility.

[0020] 5. This invention achieves intelligent balanced control of construction progress by introducing a global speed adjustment factor based on the deviation of the workload distance; ensures the paving quality of the final stage by introducing a speed attenuation factor; and compensates for the periodic additional resistance fluctuations caused by the wire rope suspension effect and track irregularities by introducing a double harmonic position compensation factor. Attached Figure Description

[0021] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2 This is a system module architecture diagram of an embodiment of the present invention; Figure 3 This is a diagram of the LSTM network structure according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, a method for intelligent speed control of a traction winch for paving on a slope includes the following steps: Step S1: Collect raw signals in real time at a preset sampling period. The raw signals include the current load mass signal of the feeding trolley, the current position signal, the current slope inclination signal, the current actual speed signal of the traction winch, the current tension signal of the wire rope, and the acceleration signal of the feeding trolley along the slope direction. Step S2: Perform multi-level intelligent signal preprocessing on the original signal to obtain a preprocessed signal. The preprocessed signal includes a preprocessed load mass signal, a preprocessed position signal, a preprocessed tilt angle signal, a preprocessed velocity signal, a preprocessed tension signal, and a preprocessed acceleration signal. Step S3: Read the engineering parameters of the slope project. Based on the preprocessed position signal, the preprocessed inclination angle signal and the engineering parameters, calculate the global speed adjustment factor, the final speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor. Step S4: Input the preprocessed signal, along with the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor, into the pre-trained long short-term memory network. The long short-term memory network then directly outputs the optimal target speed command for the current moment. Step S5: The optimal target speed command is sent directly to the frequency converter of the traction winch, and the frequency converter adjusts the speed of the traction winch drive motor according to the optimal target speed command. Step S6: At the next sampling time, repeat steps S1 to S5 to realize the real-time rolling update of the target speed command of the traction winch.

[0024] Data acquisition is performed periodically with a preset sampling period of 50ms, adjustable range of 10~100ms. The raw signals are acquired in real time through a multi-source sensor signal acquisition module. The raw signals include the current load mass signal of the feeding trolley, the current position signal, the current slope inclination signal, the current actual speed signal of the traction winch, the current tension signal of the wire rope, and the acceleration signal of the feeding trolley along the slope direction.

[0025] The acquired raw signal undergoes multi-level intelligent signal preprocessing to obtain a preprocessed signal, specifically including the following sub-steps: The original signal is subjected to anti-aliasing filtering through a hardware low-pass filter to filter out high-frequency noise above the Nyquist frequency. The cutoff frequency is set according to the sampling period. The signal after hardware filtering is processed by wavelet transform threshold denoising method. The Daubechies wavelet basis db4 is selected to perform multi-level wavelet decomposition on the signal. The high-frequency coefficients of each level are subjected to soft threshold shrinkage processing using a general threshold. Then, the processed wavelet coefficients are used to perform inverse wavelet transform to reconstruct the signal and obtain the denoised signal. The denoised signal is processed segment by segment using a sliding window approach. Within each sliding window, the mean and standard deviation of the data are calculated, and the Grubbs statistic is calculated for each data point within the window. The sliding window length is set to 50 sampling points, which can be adjusted within the range of 30 to 100 depending on the signal noise level. The specific formula is as follows:

[0026] in, Let Grubbs' statistic represent the i-th data point, reflecting the degree to which that data point deviates from the mean. This represents the magnitude of the i-th data point within the window. This represents the arithmetic mean of the data within the sliding window. This represents the sample standard deviation of the data within the sliding window; when When it exceeds the Grubbs threshold, determine Outliers are identified and removed. The threshold value in this method is dynamically adjusted adaptively based on the range of data variation within the window. The specific formula is as follows:

[0027] in, This represents the adaptively adjusted Grubbs critical value. This represents the basic Grubbs' critical value, calculated using a significance level of 0.05 and with degrees of freedom equal to the sliding window length minus 2. This represents the adaptive adjustment coefficient, ranging from 0.1 to 0.5, used to control the sensitivity of the critical value to the degree of data fluctuation. This represents the difference between the maximum and minimum values ​​of the data within the window, reflecting the fluctuation range of the data in the current window. The reference value representing the rate of change is taken as twice the standard deviation of the data within the window, or pre-calibrated according to the signal amplitude range, and used for normalization processing; When the data within the window fluctuates drastically Automatically increasing the threshold and relaxing the judgment threshold avoids misclassifying real-world changes such as sudden slope changes as outliers; when the data is stable... Approaching the baseline value, the judgment threshold is tightened to effectively eliminate abnormal jump points. The eliminated data points are filled by cubic spline interpolation. The interpolation function is constructed using the effective data points adjacent to the eliminated points to ensure the continuity and integrity of the signal on the time axis. Since different sensors have different sampling frequencies and response delays, linear interpolation is used to unify all sensor data to the same time base, eliminating the timing deviation caused by inconsistent sampling times. The synchronized data is spatially aligned, and the measurement values ​​of each sensor are converted to a unified feed trolley centroid coordinate system to eliminate the measurement error introduced by installation position deviation. The synchronized signals are normalized using a sliding window Z-score to unify the dimensions of each sensor. After the above preprocessing, the preprocessed signals are obtained, including preprocessed load mass signal, position signal, tilt angle signal, velocity signal, tension signal and acceleration signal.

[0028] Read the engineering parameters of the slope project. Based on the preprocessed position signal, preprocessed inclination angle signal, and engineering parameters, calculate five key correction factors. The engineering parameters include at least the total length of the slope track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch. The specific steps include the following: Based on the preprocessed location signal and slope distribution data for each segment, calculate the cumulative design workload of the remaining road segment from the current location to the endpoint, as well as the total design workload for the entire slope. Simultaneously, calculate the remaining distance from the current location to the endpoint. The formula for calculating the cumulative design workload of the remaining road segment from the current location to the endpoint is:

[0029] in, This represents the remaining workload, that is, the cumulative design workload for the remaining road segment from the current location to the destination, measured in terms of work done to overcome gravity. This indicates a slope segment index, dividing the entire slope into several segments based on its length. Indicates the end position of segment j. This represents the current position at step k, obtained in real time from a position sensor and after preprocessing. This represents the length of the j-th segment. Indicates the slope angle of the j-th segment. The design load capacity of segment j is provided by the engineering parameter storage module. Represents gravitational acceleration; The formula for calculating the total design workload of the entire slope is:

[0030] in, This indicates the total design workload for the entire slope. Indicates the total number of slope segments; The remaining distance is the sum of the total length of the ramp track and the current position at step k. The difference; The global speed adjustment factor is used to adjust the speed benchmark based on the deviation between the remaining workload and the remaining distance, thereby achieving intelligent balancing of the construction progress. The calculation formula is as follows:

[0031] in, This represents the global velocity adjustment factor at step k. This represents the adjustment coefficient for workload distance deviation, with a value range of 0.05 to 0.30, used to control the adjustment range. This indicates the remaining distance from the current location to the destination. Indicates the total length of the ramp track; When the first item of the remaining workload ratio is higher than the second item of the remaining distance ratio, it indicates that the subsequent road section is steeper or the load is heavier. The speed benchmark should be appropriately increased to allow more time to deal with the subsequent high-load section. Conversely, when the remaining workload ratio is lower than the remaining distance ratio, the speed benchmark should be appropriately decreased to maintain a balanced construction pace. The finishing speed attenuation factor is used to gradually reduce the speed as paving enters the final stage, ensuring the construction quality of the finishing section. The calculation formula is as follows:

[0032] in, This represents the rate of decrease at the end of step k. This represents the attenuation coefficient, ranging from 0.05 to 0.15, which controls the severity of attenuation. When the trolley is less than 85% of the total length, the attenuation factor is 1 and no attenuation is performed at the end; when the trolley enters the last 15% of the journey, the attenuation factor gradually decreases from 1 according to a quadratic curve, so as to achieve a smooth speed reduction. The local slope correction factor adjusts the speed based on the real-time slope at the current location. It reduces speed on steep slopes to ensure traction safety and temporarily reduces speed during sudden slope changes to mitigate impact. The calculation formula is as follows:

[0033] in, This represents the local slope correction factor at step k. The slope angle at step k is represented by the slope angle data collected in real time by the tilt sensor and obtained through preprocessing. This represents the slope angle at step k-1. Indicates the sampling period. This represents the slope amplitude influence coefficient, with a value ranging from 0.2 to 0.8, controlling the intensity of the slope's influence on speed. This represents the influence coefficient of slope change rate, with a value range of 0.5~3.0s, which controls the intensity of speed suppression during sudden slope changes; The load correction factor adjusts the speed based on the current load of the feeding trolley; the greater the load, the lower the speed, to avoid overload. The calculation formula is as follows:

[0034] in, This represents the load correction factor at step k. This represents the current load mass at step k, obtained by real-time data acquisition from the load sensor and preprocessing. The unloaded mass of the feeding trolley is indicated by the engineering parameter storage module. The maximum permissible load capacity is indicated by the engineering parameter storage module. This represents the load influence coefficient, with a value ranging from 0.05 to 0.25, which controls the maximum impact of load on speed. The nonlinearity index representing the load effect, ranging from 0.5 to 1.5, controls the degree of nonlinearity of the load effect as it changes with load; when Time to take =1; when hour Pick ; The position period compensation factor is used to compensate for the periodic additional resistance caused by wire rope sag and track irregularities. It employs a double harmonic compensation method, and the calculation formula is as follows:

[0035] in, This represents the position period compensation factor at step k. This represents the amplitude coefficient of the first harmonic, with a value ranging from -0.03 to 0.03. This represents the second harmonic amplitude coefficient, with a value range of -0.02 to 0.02. This represents the initial phase of the first harmonic. Indicates the initial phase of the second harmonic; The five correction factors mentioned above, together with the preprocessed signal, constitute the input feature vector of the subsequent Long Short-Term Memory network.

[0036] The preprocessed signal, along with the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor, is input into a pre-trained long short-term memory network. The long short-term memory network then directly outputs the optimal target speed command for the current moment. The input vector of the Long Short-Term Memory (LSTM) network has a dimension of 11 and consists of 11 features, including preprocessed load mass signal, preprocessed position signal, preprocessed tilt angle signal, preprocessed velocity signal, preprocessed tension signal, preprocessed acceleration signal, global velocity adjustment factor, tail velocity attenuation factor, local slope correction factor, load correction factor, and position period compensation factor. The LSM network simultaneously receives the historical input sequence of the current moment and the previous 15 consecutive moments, forming a 15×11 temporal input tensor. The Long Short-Term Memory (LSTM) network consists of two stacked LSTM hidden layers and a fully connected output layer. The first LSTM hidden layer contains 64 hidden units, which receive the temporal input tensor and output the complete time series. The second LSTM hidden layer contains 32 hidden units, which receive the complete sequence output from the first layer and output only the hidden state of the last time step. Layer normalization and Dropout regularization with a dropout rate of 0.2 are applied between the two LSTM hidden layers to prevent overfitting. The fully connected output layer maps the 32-dimensional hidden states to a 1-dimensional velocity command output.

[0037] The Long Short-Term Memory Network is trained using supervised learning. The training labels are human expert speed instructions recorded in historical construction data or verified historical best speed instructions. The training labels are strictly aligned with the network input in time. The loss function used in training is a weighted sum of velocity tracking loss, tension safety loss, smoothing loss, load-speed coordination loss, and physical feasibility loss. The specific formula is as follows:

[0038] in, Represents the overall loss function. Indicates speed tracking loss, Indicates the loss of tension safety. Indicates smoothing loss. This indicates the load-speed coordination loss. Indicates the physical feasibility loss. These represent the weighting coefficients for each loss, all of which are positive numbers and determined through cross-validation; The specific operation for determining weight coefficients through cross-validation is as follows: the historical construction dataset is divided into 5 parts according to time sequence, with 4 parts used as the training set and 1 part as the validation set, and 5-fold cross-validation is performed in rotation 5 times. Candidate weight combinations are generated by sampling at logarithmic intervals within a preset range, and the candidate range for each weight is... The values ​​are [0.1, 10.0], [0.01, 1.0], [0.01, 1.0], and [1.0, 100.0], respectively, and are fixed. As a benchmark, for each group of candidate weights, the network is trained in 5-fold cross-validation, and the comprehensive performance index on each fold validation set is calculated. The comprehensive performance index is the weighted sum of speed tracking error, total number of tension overruns, average speed command jump amplitude, and average speed under heavy load conditions. The average value of the 5 folds is taken as the final score of the group of weights, and the group with the best score is selected as the final weight coefficient. The velocity tracking loss aims to make the network output approximate the training label. Within a training batch, the square of the difference between the network output velocity command and the corresponding training label is calculated, and then the average of all samples is taken. The specific formula is as follows:

[0039] in, Indicates the training batch size. Indicates the index of a sample within a training batch. This represents the network output speed command for the nth sample. This represents the training label corresponding to the nth sample; Tension safety loss penalizes speed decisions that may lead to excessive wire rope tension. The static tension component is calculated based on the current load mass, gravitational acceleration, current inclination angle sine, and tension safety factor. This static component is then superimposed with dynamic tension fluctuation information extracted from the long short-term memory network to obtain the tension estimate. The specific formula is as follows:

[0040] in, This represents the tension estimate for the nth sample. This represents the tension safety factor, with a value ranging from 0.05 to 0.15. This represents the current load capacity of the nth sample, which is obtained in real time from the load sensor and after preprocessing. This represents the current slope inclination angle of the nth sample, which is obtained in real time by an inclination sensor and after preprocessing. This represents the tension dynamic fluctuation information of the nth sample extracted from the internal state of the Long Short-Term Memory network. During training, the hidden state output by the second LSTM layer is obtained through a linear projection. Within a training batch, the tension estimate for each sample is compared with the safe tension limit. If the tension estimate does not exceed the safe tension limit, the loss for that sample is 0. If it exceeds the limit, the square of the excess is taken, and then the average of all samples is calculated. The specific formula is as follows:

[0041] in, This indicates the safety tension limit, which is 80% of the rated traction force of the winch. For example, with the rated traction force of the feeding trolley winch being approximately 78,400 N, the safety tension limit is 62,720 N. The specific value is set according to the actual equipment parameters. The smoothing loss constrains the rate of change of velocity commands between adjacent time steps to prevent drastic jumps in velocity commands. Within a training batch, the square of the difference between the velocity command output by the network at the current time step and the velocity command output by the network at the previous time step is calculated, and then the average of all samples is taken. The specific formula is as follows:

[0042] in, This represents the network output speed command for the (n-1)th sample, i.e., the previous time step. The load-speed coordination loss causes the network to actively reduce its speed under heavy load. Within a training batch, the square of the difference between the network's output speed command and the product of the current load weight divided by the maximum allowable load weight, minus the load-speed coordination target value, is calculated, and then the average of all samples is taken. The specific formula is as follows:

[0043] in, This represents the current load capacity of the nth sample. Indicates the maximum permissible load capacity. This represents the target value for load-speed coordination, which is a preset constant with a value range of 0.01~0.05 m / s, preferably 0.02 m / s; The physical feasibility loss ensures that the output speed command is within the physically permissible range. Within a training batch, the network output speed command for each sample is checked. If the speed command exceeds the maximum permissible speed, the square of the excess is taken; if the speed command is below the minimum permissible speed, the square of the deficiency is taken; if the speed command is between the maximum and minimum permissible speeds, the loss is 0. The above losses for all samples are summed and then averaged. The specific formula is as follows:

[0044] in, This represents the network output speed command for the nth sample. This represents the maximum permissible speed, taken as 10 m / min = 0.167 m / s. This represents the minimum permissible speed, taken as 0 m / s; Of the five losses mentioned above, velocity tracking loss causes the network output to converge toward the training label; tension safety loss ensures that speed decisions do not cause the wire rope to exceed its limits; smoothing loss ensures that speed changes are smooth and avoids impacting the mechanical system; load speed coordination loss guides the network to actively reduce speed under heavy load; and physical feasibility loss ensures that the output is always within the speed range allowed by the equipment. The network employs a two-stage training approach. The offline pre-training stage utilizes historical construction data, including sensor records and corresponding human expert speed commands under various slope and load conditions, for supervised learning to optimize the overall loss function. The online fine-tuning stage involves continuously collecting data during the actual operation of the system after commissioning and using mini-batch stochastic gradient descent to update the network weights online. This allows the network to gradually adapt to the personalized characteristics of specific projects, such as the actual slope distribution, material properties, and equipment wear status of the construction site.

[0045] The optimal target speed command output by the Long Short-Term Memory (LSTM) network is directly sent to the variable frequency drive (VFD) of the traction winch through the drive control interface submodule. Based on the received target speed command, the VFD adjusts the actual speed of the drive motor to the target value corresponding to the optimal target speed command through its inherent speed loop and current loop, thereby driving the traction winch drum to rotate and achieving precise speed control. The generation of the speed command is entirely determined by the LSTM network in one go after considering the current sensor status, historical timing information, and global engineering parameters, without the need for subsequent PID closed-loop correction, thus avoiding the inherent lag and overshoot problems of traditional PID control.

[0046] At the next sampling moment, based on the new sensor observation data, steps S1 to S5 are repeated to realize real-time rolling updates and dynamic decision-making of the traction winch target speed command, ensuring that the system has a rapid response capability to changes in working conditions such as slope changes and wire rope tension fluctuations.

[0047] like Figure 2 As shown, an intelligent speed control system for a traction winch in sloping paving is used to implement any one of the intelligent speed control methods for a traction winch in sloping paving, comprising: The multi-source sensor signal acquisition module is used to acquire in real time the load mass signal, position signal, tilt angle signal, speed signal of the traction winch, wire rope tension signal and acceleration signal of the feeding trolley along the slope direction; The engineering parameter storage module is used to store the total length of the inclined track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch. The signal processing and intelligent decision-making module is connected to the multi-source sensor signal acquisition module and the engineering parameter storage module, respectively. It is used to preprocess the acquired raw signals, calculate the correction factor based on the engineering parameters and the preprocessed position and tilt signals, and output the optimal target speed command using the built-in long short-term memory network. The traction speed execution module, connected to the signal processing and intelligent decision-making module, is used to receive the optimal target speed command and adjust the speed of the traction winch drive motor.

[0048] like Figure 3 As shown, the Long Short-Term Memory (LSTM) network comprises an input layer, two stacked LSTM hidden layers, and a fully connected output layer. The input layer receives a 15×11 temporal input tensor composed of 11-dimensional feature vectors from the current time step and the previous 15 consecutive time steps. These 11-dimensional feature vectors consist of 11 preprocessed features: load mass signal, position signal, tilt angle signal, velocity signal, tension signal, acceleration signal, and global velocity adjustment factor, tail velocity decay factor, local slope correction factor, load correction factor, and position period compensation factor. The first LSTM hidden layer contains 64 hidden units, which receive the temporal input tensor and output the complete time series to the second layer. The second LSTM hidden layer contains 32 hidden units, which receive the complete sequence output by the first layer and output only the hidden state of the last time step. Layer normalization and Dropout regularization with a dropout rate of 0.2 are applied between the two LSTM hidden layers to prevent overfitting. The fully connected output layer maps the 32-dimensional hidden state to a 1-dimensional velocity command output.

[0049] The multi-source sensor signal acquisition module specifically includes: The load sensor is installed between the frame and wheels of the feeding trolley to collect the load mass signal of the feeding trolley in real time. A position sensor, installed on the end of the traction winch drum shaft, is used to collect the position signal of the feeding trolley along the slope in real time; Inclination sensors, installed on the feeding trolley or the base of the inclined track, are used to collect incline signals in real time. A speed sensor, installed on the drum of the traction winch, is used to collect the actual speed signal of the traction winch in real time; Tension sensors are installed in series at the fixed end of the wire rope to collect the wire rope tension signal in real time. An accelerometer, installed on the feeding trolley, is used to collect the acceleration signal of the feeding trolley along the slope in real time.

[0050] The load sensor is a pin-type pressure sensor or a strain gauge load cell, with a range covering from no-load to 1.5 times the maximum permissible load mass; the position sensor is a rotary encoder, which indirectly calculates the position signal of the feeding trolley by measuring the number of rotations of the drum; the tilt sensor is a MEMS accelerometer-type tilt sensor or an electrolyte-type tilt sensor, with a measurement range covering 0° to 45° and an accuracy of not less than 0.1°; the tension sensor is a pin-type tension sensor or a piezomagnetic tension sensor, with a range of not less than 1.5 times the rated traction force of the winch.

[0051] The signal processing and intelligent decision-making module specifically includes: The data acquisition submodule is connected to each sensor of the multi-source sensing signal acquisition module and is used to synchronously acquire the raw signals of each sensor at a preset sampling period. The intelligent signal preprocessing submodule is connected to the data acquisition submodule and is used to perform hardware filtering, wavelet transform denoising, outlier removal and interpolation filling, time synchronization and spatial alignment and normalization on the original signal, and output the preprocessed signal. The engineering parameter correction submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter storage module, respectively, and is used to calculate the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor based on the preprocessed position signal, the preprocessed tilt angle signal and the engineering parameters. The deep decision submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter correction submodule, respectively. The deep decision submodule has a built-in long short-term memory network, which is used to directly output the optimal target speed command at the current moment based on the preprocessed signal and the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position period compensation factor. The drive control interface submodule, connected to the depth decision submodule, is used to send the optimal target speed command to the traction speed execution module.

[0052] The traction speed execution module specifically includes: The variable frequency speed controller is connected to the drive control interface submodule of the signal processing and intelligent decision-making module, and is used to receive the optimal target speed command and generate the corresponding drive frequency and voltage signal according to the optimal target speed command. A drive motor is connected to the frequency converter, and its output shaft is connected to the drum drive of the traction winch. It is used to output the corresponding speed according to the drive frequency and voltage signal to drive the drum of the traction winch to rotate.

[0053] The variable frequency drive is a vector control type variable frequency drive, which contains a dual closed-loop control structure of speed loop and current loop, and supports speed regulation within the range of 0 to 10 m / min; after receiving the optimal target speed command, the variable frequency drive adjusts the actual speed of the drive motor to the target value corresponding to the optimal target speed command through the speed loop and current loop.

[0054] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A method for intelligent speed control of a traction winch for paving on a slope, characterized in that, Includes the following steps: Step S1: Collect raw signals in real time at a preset sampling period. The raw signals include the current load mass signal of the feeding trolley, the current position signal, the current slope inclination signal, the current actual speed signal of the traction winch, the current tension signal of the wire rope, and the acceleration signal of the feeding trolley along the slope direction. Step S2: Perform multi-level intelligent signal preprocessing on the original signal to obtain a preprocessed signal. The preprocessed signal includes a preprocessed load mass signal, a preprocessed position signal, a preprocessed tilt angle signal, a preprocessed velocity signal, a preprocessed tension signal, and a preprocessed acceleration signal. Step S3: Read the engineering parameters of the slope project. Based on the preprocessed position signal, the preprocessed inclination angle signal and the engineering parameters, calculate the global speed adjustment factor, the final speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor. Step S4: Input the preprocessed signal, along with the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor, into the pre-trained long short-term memory network. The long short-term memory network then directly outputs the optimal target speed command for the current moment. Step S5: The optimal target speed command is sent directly to the frequency converter of the traction winch, and the frequency converter adjusts the speed of the traction winch drive motor according to the optimal target speed command. Step S6: At the next sampling time, repeat steps S1 to S5 to realize the real-time rolling update of the target speed command of the traction winch.

2. The method according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: Step S2.1: Perform anti-aliasing filtering on the original signal using a hardware low-pass filter to filter out high-frequency noise higher than the Nyquist frequency; Step S2.2: The signal after hardware filtering is processed by wavelet transform threshold denoising to obtain the denoised signal; Step S2.3: The denoised signal is processed segment by segment using a sliding window method. The Grubbs statistic of each data point in the window is calculated. When the Grubbs statistic exceeds the adaptively adjusted Grubbs threshold, the data point is determined to be an outlier and is removed. The removed data points are filled in using cubic spline interpolation. Step S2.4: Unify the data from each sensor to the same time reference using linear interpolation, perform spatial alignment, and transform them into a unified feed trolley centroid coordinate system; Step S2.5: Perform sliding window normalization on the synchronized and aligned signal to obtain the preprocessed signal.

3. The method according to claim 1, characterized in that, In step S3, the engineering parameters include at least the total length of the inclined track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch.

4. The method according to claim 3, characterized in that, Step S3 specifically includes the following sub-steps: Step S3.1: Based on the preprocessed location signal and slope distribution data of each segment, calculate the cumulative design workload of the remaining road segment from the current position to the endpoint and the total design workload of the entire slope, and at the same time calculate the remaining distance from the current position to the endpoint; Step S3.2: Calculate the global speed adjustment factor based on the remaining workload, total workload, remaining distance, and total length of the ramp track; Step S3.3: Calculate the final velocity attenuation factor based on the preprocessed position signal and the total length of the ramp track; Step S3.4: Calculate the local slope correction factor based on the preprocessed tilt angle signal; Step S3.5: Calculate the load correction factor based on the preprocessed load mass signal, the empty load mass of the feeding trolley, and the maximum allowable load mass; Step S3.6: Calculate the position period compensation factor based on the preprocessed position signal and the total length of the ramp track.

5. The method according to claim 4, characterized in that, In step S3.1, the cumulative design workload of the remaining road segments from the current position to the endpoint is the weighted sum of the products of the length of each remaining road segment, its corresponding sine value of the inclination angle, and the design load mass, multiplied by the gravitational acceleration; the total design workload of the entire slope is the weighted sum of the products of the length of all road segments of the entire slope, its corresponding sine value of the inclination angle, and the design load mass, multiplied by the gravitational acceleration; the remaining distance is the total length of the slope track minus the current position; In step S3.2, the global speed adjustment factor is 1 plus the workload distance deviation adjustment coefficient multiplied by the difference between the remaining workload ratio and the remaining distance ratio; the remaining workload ratio is the remaining workload divided by the total workload, and the remaining distance ratio is the remaining distance divided by the total length of the ramp track. In step S3.3, when the position corresponding to the preprocessed position signal does not exceed 85% of the total length of the ramp track, the closing speed attenuation factor is 1; when the position corresponding to the preprocessed position signal exceeds 85% of the total length of the ramp track, the closing speed attenuation factor is 1 minus the closing attenuation coefficient multiplied by the square of the normalized excess ratio, where the normalized excess ratio is the difference between the current position percentage and 0.85 divided by 0.

15. In step S3.4, the local slope correction factor is a negative exponent of the natural constant. The exponent is the sum of the slope amplitude influence coefficient multiplied by the square of the current inclination angle sine and the slope change rate influence coefficient multiplied by the absolute value of the difference between the current inclination angle and the inclination angle at the previous moment, divided by the sampling period. In step S3.5, the load correction factor is 1 minus the load influence coefficient multiplied by the load influence nonlinear exponent of the normalized load ratio. The normalized load ratio is the difference between the current load mass and the unloaded mass divided by the difference between the maximum allowable load mass and the unloaded mass. When the current load mass is not greater than the unloaded mass, the load correction factor is 1. When the current load mass is not less than the maximum allowable load mass, the load correction factor is 1 minus the load influence coefficient. In step S3.6, the position period compensation factor is 1 plus the sum of the first harmonic term and the second harmonic term; the first harmonic term is the sine function value calculated by multiplying the first harmonic amplitude coefficient by 2π times the current position divided by the total length plus the first initial phase as the independent variable; the second harmonic term is the cosine function value calculated by multiplying the second harmonic amplitude coefficient by 4π times the current position divided by the total length plus the second initial phase as the independent variable.

6. The method according to claim 5, characterized in that, In step S4, the input vector of the Long Short-Term Memory (LSTM) network has a dimension of 11 and consists of 11 features: preprocessed load mass signal, preprocessed position signal, preprocessed tilt angle signal, preprocessed velocity signal, preprocessed tension signal, preprocessed acceleration signal, and the global velocity adjustment factor, the tail velocity attenuation factor, the local slope correction factor, the load correction factor, and the position period compensation factor. The LSM network simultaneously receives the historical input sequence of the current moment and the previous 15 consecutive moments, forming a 15x11 temporal input tensor. The Long Short-Term Memory (LSTM) network comprises two stacked LSTM hidden layers and a fully connected output layer. The first LSTM hidden layer contains 64 hidden units, which receive the temporal input tensor and output the complete time series. The second LSTM hidden layer contains 32 hidden units, which receive the complete sequence output by the first layer and output only the hidden state of the last time step. Layer normalization and Dropout regularization with a dropout rate of 0.2 are set between the two LSTM hidden layers. The fully connected output layer maps the 32-dimensional hidden state to a 1-dimensional velocity command output.

7. The method according to claim 6, characterized in that, In step S4, the Long Short-Term Memory (LSTM) network is trained using supervised learning, with training labels being human expert speed commands recorded in historical construction data or verified historical optimal speed commands. The overall loss function of the LSM network is a weighted sum of speed tracking loss, tension safety loss, smoothing loss, load-speed coordination loss, and physical feasibility loss.

8. The method according to claim 7, characterized in that, The speed tracking loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network and the corresponding training label is calculated, and then the average value of all samples is taken. The tension safety loss is calculated as follows: the static tension component is calculated based on the current load mass, gravitational acceleration, current tilt angle sine value, and tension safety factor. Then, the tension dynamic fluctuation information extracted from the internal state of the long short-term memory network is superimposed to obtain the tension estimate. Within a training batch, the tension estimate of each sample is compared with the safety tension limit. If the tension estimate does not exceed the safety tension limit, the loss is 0. If it exceeds the limit, the square of the excess is taken, and then the average of all samples is taken. The smoothing loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network at the current moment and the speed command output by the network at the previous moment is calculated, and then the average value of all samples is taken. The load-speed coordination loss is calculated as follows: within a training batch, the square of the difference between the speed command output by the network and the product of the current load weight divided by the maximum allowable load weight, minus the load-speed coordination target value, is calculated, and then the average value of all samples is taken. The physical feasibility loss is calculated as follows: within a training batch, the network output speed command of each sample is checked. If the speed command exceeds the maximum allowable speed, the square of the excess is taken. If the speed command is lower than the minimum allowable speed, the square of the lower part is taken. If the speed command is between the maximum allowable speed and the minimum allowable speed, the loss is 0. The above losses of all samples are summed and then averaged.

9. An intelligent speed control system for a traction winch in paving on a slope, used to implement the intelligent speed control method for a traction winch in paving on a slope as described in any one of claims 1-8, characterized in that, include: The multi-source sensor signal acquisition module is used to acquire in real time the load mass signal, position signal, tilt angle signal, speed signal of the traction winch, wire rope tension signal and acceleration signal of the feeding trolley along the slope direction; The engineering parameter storage module is used to store the total length of the inclined track, the slope distribution data of each section, the design paving speed, the unloaded mass of the feeding trolley, the maximum allowable load mass, and the rated traction force of the winch. The signal processing and intelligent decision-making module is connected to the multi-source sensor signal acquisition module and the engineering parameter storage module, respectively. It is used to preprocess the acquired raw signals, calculate the correction factor based on the engineering parameters and the preprocessed position and tilt signals, and output the optimal target speed command using the built-in long short-term memory network. The traction speed execution module, connected to the signal processing and intelligent decision-making module, is used to receive the optimal target speed command and adjust the speed of the traction winch drive motor.

10. The system according to claim 9, characterized in that, The signal processing and intelligent decision-making module specifically includes: The data acquisition submodule is connected to each sensor of the multi-source sensing signal acquisition module and is used to synchronously acquire the raw signals of each sensor at a preset sampling period. The intelligent signal preprocessing submodule is connected to the data acquisition submodule and is used to perform hardware filtering, wavelet transform denoising, outlier removal and interpolation filling, time synchronization and spatial alignment and normalization on the original signal, and output the preprocessed signal. The engineering parameter correction submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter storage module, respectively, and is used to calculate the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position cycle compensation factor based on the preprocessed position signal, the preprocessed tilt angle signal and the engineering parameters. The deep decision submodule is connected to the intelligent signal preprocessing submodule and the engineering parameter correction submodule, respectively. The deep decision submodule has a built-in long short-term memory network, which is used to directly output the optimal target speed command at the current moment based on the preprocessed signal and the global speed adjustment factor, the tail speed attenuation factor, the local slope correction factor, the load correction factor and the position period compensation factor. The drive control interface submodule, connected to the depth decision submodule, is used to send the optimal target speed command to the traction speed execution module.