Bridge cutting dynamic deviation correction method and system based on multi-modal sensing

Abstract

The invention discloses a bridge cutting dynamic deviation correction method and system based on multi-modal sensing, relates to the technical field of bridge construction and demolition, and solves the technical problems of realizing high-precision self-adaptive cutting under complex working conditions, reducing rope breakage rate and track deviation and improving construction efficiency and safety. By fusing the fiber grating, ultrasonic wave and machine vision technologies, the monitoring limitation of a single sensor is broken through, multi-mode synchronous sensing of a stress field, a temperature field, a reinforcing steel bar position and geometric deformation is realized, a complete data basis is provided for intelligent decision making, and multi-source data is fused and denoised by adopting wavelet transform and principal component analysis, so that the accuracy and the reliability of the system are improved. The robustness of the system in vibration, chipping interference and temperature drift environments is improved, and the deviation correction efficiency is continuously improved by predicting a reinforcing steel bar dense area and a stress sudden change point in real time, dynamically adjusting cutting parameters, prolonging the service life of a rope saw, reducing construction interruption and introducing a residual feedback and iterative learning mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bridge construction and demolition technology, in particular to a bridge cutting dynamic deviation correction method and system based on multi-modal sensing. BACKGROUND

[0002] In bridge demolition, reconstruction or emergency rescue engineering, the rope saw cutting technology is widely used due to its advantages of high efficiency, low vibration and adaptability to complex structure. The traditional rope saw cutting method mainly relies on the experience of operators or single sensor for cutting trajectory control.

[0003] The prior art has the following common defects: The traditional system mainly relies on a single type of sensor, and cannot synchronously obtain multi-modal data such as stress field, temperature field and spatial position of steel bars, resulting in incomplete decision basis; The existing deviation correction algorithm is mainly based on static model or empirical formula, and does not fuse real-time deep learning framework, so it is difficult to realize the collaborative optimization and millisecond-level dynamic adjustment of cutting parameters; The heterogeneous interface between steel bars and concrete is easy to cause cutting vibration, debris accumulation and sensor signal drift, and the traditional method lacks multi-source data fusion and noise suppression mechanism, so the deviation correction stability is poor; Because the position of the steel bar and the stress concentration area cannot be accurately predicted, the rope saw frequently bears impact load during cutting, resulting in a high rope breaking rate, which greatly increases the construction cost and safety risk Therefore, there is an urgent need for a bridge cutting dynamic deviation correction method and system based on multi-modal sensing, which breaks through the perception limitations and control bottlenecks of traditional technology through multi-dimensional information fusion and intelligent decision-making, realizes high-precision adaptive cutting under complex working conditions, reduces the rope breaking rate and trajectory deviation, and improves the construction efficiency and safety. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a bridge cutting dynamic deviation correction method and system based on multi-modal sensing, which solves the problems of realizing high-precision adaptive cutting under complex working conditions, reducing the rope breaking rate and trajectory deviation, and improving the construction efficiency and safety.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a bridge cutting dynamic deviation correction method based on multi-modal sensing, which comprises the following steps: Step one, arrange a fiber grating sensor array along the inner side of the cutting path, synchronously configure a high-frequency ultrasonic flaw detection device, use pulse reflection method to scan the three-dimensional coordinates of the steel bars to generate a steel bar distribution thermal map, and use an infrared thermal imager and a laser profiler to capture the temperature field and trajectory geometric deformation of the cutting surface, respectively; Step two, collect the three-dimensional coordinate data of the steel bar, the thermal imaging map and the three-dimensional point cloud data, use wavelet transform for noise reduction, use principal component analysis method to extract core feature vector, including mechanical characteristics, structural characteristics, thermal characteristics and geometric characteristics, generate 12-dimensional fusion feature data set; Step three, construct a DQN model based on deep reinforcement learning, the input layer receives 12-dimensional feature data, the output layer defines the action space, including rope saw tension increment, cutting speed correction coefficient, cooling liquid flow level, etc. Step four, the servo hydraulic system constructs the real-time control link of instruction-execution-feedback, the variable frequency driving speed regulation system adopts space vector pulse width modulation technology, according to the speed correction coefficient The duty cycle is adjusted dynamically, and the cooling system controls the flow according to the thermodynamic feedback through the electromagnetic pulse valve in a segmented manner. Step five, after the deviation is corrected, the laser profiler scans the track again, extracts the residual error between the actual and designed paths, if the residual error is not up to standard, triggers the iterative learning, updates the DQN model weight online and performs three iterations, automatically generates an efficiency report containing the number of broken ropes, average deviation and energy consumption indicators.

[0006] As a further scheme of the application, the wavelet transform for noise reduction includes 3-layer wavelet transform noise reduction using db4 wavelet basis to eliminate 50Hz power frequency interference and environmental vibration noise ≤0.5Hz; the missing data is completed by K-nearest neighbor interpolation method.

[0007] As a further scheme of the application, the mechanical characteristics include the position coordinates of the stress mutation region, the mutation duration and the peak stress; The position coordinates of the stress mutation region are identified by gradient threshold method, the mutation region boundary is fitted by least square method, and the center coordinates of the boundary rectangle are output; The mutation duration specifically represents the time interval from the first time the stress gradient exceeds the threshold to the time when it returns to below the threshold; The peak stress specifically represents the maximum stress value in the mutation region.

[0008] As a further scheme of the application, the structural characteristics include the distribution range of the steel bar dense area, the average diameter of the steel bar and the minimum protective layer thickness; The distribution range of the steel bar dense area is merged by morphological dilation operation, and the radius and center coordinates of the circumscribed circle of the dense area are output; The average diameter of the steel bar is calculated by the peak distance of the ultrasonic echo signal, and the calculation formula is diameter=wave speed×time difference / 2, and the arithmetic mean value is taken; The minimum protective layer thickness is the shortest distance from the outer edge of the steel bar to the surface of the concrete, which is calculated by three-dimensional coordinate difference.

[0009] As a further scheme of the present application, the thermal features include the center coordinates of the temperature abnormal hot spot, the maximum temperature, and the thermal diffusion coefficient; The center coordinates of the temperature abnormal hot spot are obtained by using 8-neighborhood connected domain analysis on the infrared thermal image, identifying the area with a temperature rise rate ≥ 5℃ / s through two-frame thermal image temperature difference calculation rate, and outputting the center coordinates of the hot spot area. The maximum temperature represents the maximum temperature value in the hot spot area. The thermal diffusion coefficient is obtained by fitting the heat conduction equation with the temperature field distribution of the hot spot area, and the specific equation is , and the diffusion coefficient is calculated.

[0010] As a further scheme of the present application, the geometric features include the flatness of the cutting surface, the edge perpendicularity, and the real-time trajectory curvature. The flatness of the cutting surface is obtained by fitting an ideal plane with the least square method on the laser point cloud data, and calculating the arithmetic mean of the vertical distance of each point to the plane. The edge perpendicularity is the angle between the cutting surface edge line and the design vertical line, which is calculated after the edge line is extracted by Hough transform. The real-time trajectory curvature is obtained by fitting a circular arc on the cutting trajectory point cloud, and calculating the curvature value.

[0011] As a further scheme of the present application, the DQN model based on deep reinforcement learning includes that the input layer receives 12-dimensional normalized feature data, the hidden layer adopts a 3-layer fully connected network, and the output layer defines a three-dimensional action space including the increment of the rope saw tension, the correction coefficient of the cutting speed, and the flow level of the cooling liquid.

[0012] As a further scheme of the present application, the servo hydraulic system constructs a real-time control link of instruction-execution-feedback, which includes using a load-sensitive servo hydraulic system, the core execution element is a double-acting hydraulic cylinder, the inner ring uses a high-precision pressure sensor to feedback the tension in real time, and the outer ring is integrated with the hydraulic cylinder piston rod through an absolute linear encoder, and adopts position-speed double-loop regulation.

[0013] As a further scheme of the present application, the specific way of segmented flow control through the electromagnetic pulse valve is: Extract the maximum temperature T of the cutting surface based on the temperature field data of the infrared thermal imager max As a control reference, when T max ≤ 60℃, preferentially maintain low flow to reduce cooling liquid consumption, and adopt 5Hz intermittent injection mode, specifically 100ms valve opening / 100ms valve closing cycle; When 60℃ < T max≤80℃, real-time switching according to temperature fluctuation, using 10Hz high-frequency intermittent injection; When T max >80℃, the electromagnetic pulse valve is opened without interval, and the standby pump is started, and the total flow is increased to 3.0L / s.

[0014] A bridge cutting dynamic deviation correction system based on multi-modal sensing, comprising: The multi-modal sensing module is composed of a fiber grating sensor array, an ultrasonic flaw detection device, an infrared thermal imager and a laser profiler, and realizes full-dimensional perception of stress, temperature, reinforcement position and geometric deformation of a cutting surface; The edge computing and fusion module: deploy FPGA chips and GPU acceleration units to complete real-time fusion of multi-source data, noise suppression and feature extraction; The intelligent decision module: contains a deep reinforcement learning prediction model and a fuzzy PID controller, and outputs dynamic deviation correction instructions; The actuator module: integrates a servo hydraulic tension regulator, a variable frequency motor driver and an intelligent cooling system, and realizes millisecond-level collaborative regulation of cutting parameters; The closed-loop verification module: through machine vision and fiber grating sensing feedback, an evaluation system for deviation correction effect is constructed to drive model self-optimization.

[0015] The present application provides a kind of bridge cutting dynamic deviation correction method and system based on multi-modal sensing.Compared with prior art, it has the following beneficial effects: The present application breaks through the limitation of single sensor monitoring by fusing fiber grating, ultrasonic and machine vision technology, realizes multi-modal synchronous perception of stress field, temperature field, reinforcement position and geometric deformation, provides complete data basis for intelligent decision, realizes millisecond-level collaborative optimization of rope saw tension, cutting speed and cooling parameters through the coupling of deep reinforcement learning model and fuzzy PID controller, effectively suppresses the accumulation of cutting deviation, and improves the deviation correction accuracy.

[0016] The present application uses wavelet transform and principal component analysis to fuse and denoise multi-source data, improves the robustness of the system in vibration, debris interference and temperature drift environment, dynamically adjusts the cutting parameters by real-time prediction of reinforcement dense area and stress mutation point, prolongs the service life of rope saw and reduces construction interruption, introduces residual feedback and iterative learning mechanism, so that the system can adapt to different bridge structures, material properties and environmental conditions, and continuously improve the deviation correction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is a step method diagram; Figure 2 It is a multi-modal sensing network deployment schematic diagram of the present application; Figure 3 It is a data fusion and feature extraction flowchart; Figure 4 is a deep reinforcement learning model architecture diagram; Figure 5 is a cutting trajectory deviation contrast schematic diagram. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0019] First embodiment Please refer to Figure 1 The present application provides a bridge cutting dynamic deviation correction method based on multi-modal sensing, comprising the following steps: Step one, in the bridge cutting operation area, a fiber grating sensor array is arranged along the inside of the cutting path at a longitudinal interval of 10 cm. Specifically, the sensor nodes are integrated into a single armored optical fiber using WDM (wavelength division multiplexing technology), and each meter of optical fiber contains 8 monitoring nodes. The nodes are embedded in the concrete surface layer and deep layer with a grid density of 5 cm x 5 cm, where the surface layer represents a depth of 0-10 cm, and the deep layer represents a depth of 10-50 cm. The deep layer nodes are implanted through special drilling tools to ensure no contact interference with the steel bars. The sensor synchronously monitors the axial stress and environmental temperature; A high-frequency ultrasonic flaw detection device is synchronously configured. Specifically, the transmitting and receiving probes are symmetrically distributed on both sides of the cutting surface at an angle of 30°, with a spacing of 50 cm. The pulse-echo method is used to scan the three-dimensional coordinates of the steel bars. Specifically, the scanning step is 0.5 mm, and a steel bar distribution thermal map is generated every 10 seconds. The machine vision module is composed of an infrared thermal imager and a laser profiler. The parameters of the infrared thermal imager include resolution 640x480, thermal sensitivity 0.05℃, and frame rate 30fps. The parameters of the laser profiler include scanning frequency 100Hz and point cloud density 100 points / mm. The former captures the temperature field distribution of the cutting surface in real time, and the latter dynamically records the three-dimensional geometric deformation data of the cutting trajectory.

[0020] Step two, through data preprocessing and feature extraction, heterogeneous sensor data is converted into high-dimensional feature vectors that can be used for risk decision-making. The stress-temperature time series data of the fiber grating sensor and the three-dimensional coordinate data set of the ultrasonic flaw detection device are transmitted in real time to the edge computing unit through industrial Ethernet. Next, the acquired data is preprocessed to obtain preprocessed data. Specifically, a 3-level wavelet transform is performed using the db4 wavelet basis for noise reduction to eliminate 50Hz power frequency interference and environmental vibration noise ≤0.5Hz. Missing data is filled using K-nearest neighbor (K=5) interpolation. Simultaneously, feature extraction is performed on the obtained preprocessed data, using principal component analysis to extract core feature vectors, specifically including mechanical, structural, thermal, and geometric features. The specific methods are as follows: Mechanical characteristics include the location coordinates of the stress abrupt change region, the duration of the abrupt change, and the peak stress; The gradient threshold method is used to identify mutation centers, and the least squares method is used to fit the boundary of the mutation region to output the center coordinates of the boundary rectangle. The duration of the mutation is specifically the time interval from when the stress gradient first exceeds the threshold to when it recovers to below the threshold. Peak stress, specifically representing the maximum stress value within the abrupt change region; Structural features include the distribution range of densely reinforced areas, the average diameter of the reinforcing bars, and the minimum thickness of the protective layer; The distribution range of densely reinforced areas is defined as areas with a reinforcing bar density of ≥3 bars / 10cm. 2 As the criterion, adjacent dense points are merged through morphological dilation operation, and the radius and center coordinates of the circumcircle of the dense region are output. The average diameter of the reinforcing bars is calculated by taking the diameter of all reinforcing bars in the dense area through the peak spacing of the ultrasonic echo signal. The calculation formula is: diameter = wave velocity × time difference / 2, and the arithmetic mean is taken. The minimum protective layer thickness, specifically the shortest distance from the outer edge of the reinforcing bar to the concrete surface, is calculated using three-dimensional coordinate differences and retained to the nearest integer. Thermal characteristics include the center coordinates of the temperature anomaly hotspot, the highest temperature, and the diffusion coefficient; The center coordinates of the temperature anomaly hotspots are obtained by performing 8-neighbor connected component analysis on the infrared thermal image and identifying regions with a temperature rise rate ≥ 5℃ / s by calculating the temperature difference between two thermal images, and then outputting the center coordinates of the hotspot regions. Maximum temperature, specifically referring to the highest temperature value within the hotspot area; The thermal diffusivity is specifically determined by fitting the heat conduction equation to the temperature field distribution in the hot spot region, and the specific equation is as follows: The diffusion coefficient is calculated. Geometric features include cut surface flatness, edge perpendicularity, and real-time trajectory curvature; The flatness of the cut surface is determined by fitting an ideal plane to the laser point cloud data using the least squares method, and calculating the arithmetic mean of the vertical distances from each point to the plane. Edge perpendicularity: The angle between the edge line of the cut surface and the design perpendicular line. The angle difference is calculated after extracting the edge line through Hough transform and is retained to two decimal places. Real-time trajectory curvature: Perform circular arc fitting on the point cloud of the cutting trajectory, calculate the curvature value, and retain 4 decimal places; Finally, a 12-dimensional fused feature dataset is generated. The feature vectors are normalized and mapped to the interval [0, 1] to obtain 12-dimensional normalized feature vectors, which are then input into the decision model.

[0021] Step 3: Construct a DQN model based on deep reinforcement learning. It receives a 12-dimensional normalized feature vector, uses a linear activation function, and employs a 3-layer fully connected network for the hidden layers: layer 1: 256 neurons, layer 2: 128 neurons, and layer 3: 64 neurons. The output layer uses a Q-value matrix in a 3D action space, with dimensions including the selectable values ​​for the wire saw tension increment ΔF and the cutting speed correction coefficient. Optional values ​​and coolant flow rate L values; Construct an experience-replay-driven deep reinforcement learning training system with a capacity of 10. 6 The replay buffer stores historical interaction data for each sample, and random sampling is performed according to a TD error priority ratio of 0.6. Then, a distribution offset suppression module is introduced. Through adversarial training, the distribution offset is controlled to keep the JS divergence < 0.1. At the same time, the learning rate is initialized to 0.001, and segmentation is adopted. - Greedy strategy, specifically The training time was reduced from 0.9 to 0.1 after 200,000 steps, then fixed. The value is 0.1, and the training duration is 200,000 steps. Next, a training set covering 500 sets of differentiated working conditions was obtained, specifically including 1-5 roots / 10cm. 2 The data collection process for each working condition included the gradient of rebar density, the strength of concrete from C30 to C60, and the ambient temperature range of 20-50℃. A three-layer convergence verification standard was established, including core criteria, process indicators, and generalization indicators. Specifically, model convergence was defined as follows: prediction error ≤ 5% as the core criterion; Q-value loss function change rate < 0.1% over 5000 consecutive steps as the process indicator; and action strategy entropy value decreasing to below 1 / 3 of the initial value as the generalization indicator.

[0022] Step 4: Dynamically implement decision parameters through a high-precision actuator. The servo hydraulic system constructs a real-time control link of command-execution-feedback. A load-sensitive servo hydraulic system is adopted, with a double-acting hydraulic cylinder as the core actuator. A high-precision pressure sensor in the inner ring provides real-time tension feedback. Simultaneously, a PID control algorithm is used for parameter adjustment. The specific parameter configuration is K. p =1.2, Ki =0.8, K d =0.3, and introduce feedforward compensation to offset the influence of elastic deformation of the wire saw, with a compensation coefficient of 0.15. The outer ring is integrated into the hydraulic cylinder piston rod through an absolute linear encoder, and position-speed dual-ring adjustment is adopted. The position ring achieves smooth feed through trapezoidal acceleration and deceleration planning. The variable frequency drive speed regulation system is equipped with a vector-type variable frequency motor with a rated power of 22kW and a speed range of 0-5m / s. It adopts space vector pulse width modulation technology and adjusts the speed according to the speed correction coefficient. Dynamically adjust the PWM duty cycle, and The value ranges from 0.8 to 1.2, with a speed range of 0-5 m / s, corresponding to a motor speed of 0-1500 rpm, and a speed fluctuation of ≤±2%. The cooling system consists of three sets of electromagnetic pulse valves, spiral nozzles, and stainless steel piping. The electromagnetic pulse valves are model 2W-250-25, with a response time of <10ms and an operating pressure of 0.1-1.0MPa. The spiral nozzles have a spray angle of 60° and an atomized particle size of 50-100μm. The highest temperature T on the cut surface is extracted based on temperature field data from an infrared thermal imager. max As a control benchmark, segmented flow control is adopted, and the specific control method is as follows: When T max At ≤60℃, it is classified as a low temperature stable zone. For flow level one or two, low flow rate is prioritized to reduce coolant consumption. Specifically, the flow rate value for level one is 0.5L / s, and the flow rate value for level two is 1.0-1.5L / s. A 5Hz intermittent injection mode is adopted, with the valve opening 100ms / closing 100ms cycle. When 60℃ <T max When the temperature is ≤80℃, it is classified as a medium temperature control zone. For flow levels of three or four, it is switched in real time according to temperature fluctuations. Specifically, the flow value of level three is 1.5L / s and the flow value of level four is 2.0L / s, using 10Hz high-frequency intermittent injection. When T max When the temperature is above 80℃, it is classified as a high-temperature forced zone. For flow rate level 5, the electromagnetic pulse valve is fully open without interruption, and the standby pump is started, increasing the total flow rate to 3.0L / s. Step 5: The laser profilometer completes a second scan within 100ms after the correction is executed. Based on the point cloud data, it extracts the 3D residual between the actual path and the designed path. When the residual on any axis is >0.5mm, an iterative learning mechanism is immediately triggered to ensure timely correction of the deviation. The specific correction method is as follows: A median filtering + Gaussian filtering combined algorithm is used to remove measurement noise ≤0.1mm and retain the true trajectory deviation characteristics. The filtered residual data is added to the training set, and the weights of the DQN model are updated by online incremental learning. 50 sets of simulated control schemes are generated in each iteration. After 3 iterations, the residual converges to within 0.3mm. A correction efficiency report is automatically generated every 10m of cutting length, which includes safety indicators, accuracy indicators and energy efficiency indicators. The safety indicators include the number of rope breaks and the duration of overheating. The number of rope breaks is statistically analyzed by the fluctuation curve of the tension sensor. The warning value is 0 times. When a rope breaks, the root cause analysis is triggered. The duration of overheating is based on the temperature of the cutting surface monitored by the infrared thermal imager. The cumulative duration of temperature > 80℃ is counted. The accuracy indicators include average trajectory deviation and verticality error. Specifically, the average trajectory deviation is the average value of the trajectory residual at 10cm intervals within a 10m cutting segment. The verticality error is obtained by scanning the normal vector of the cutting surface with a laser profilometer. The angle between the normal vector and the designed vertical surface is ≤0.1°. When the deviation exceeds the limit, path planning correction is triggered, and the cutting trajectory is refitted. The energy consumption indicators include energy consumption per unit length and coolant consumption. The specific energy consumption per unit length is calculated by statistically analyzing the power consumption of the frequency converter and hydraulic system to determine the average energy consumption of a 10m cutting section, and then compared with its target value, which is set at 5kWh / m. If the energy consumption per unit length exceeds the target, the speed-tension matching coefficient is optimized to reduce idling losses. Coolant consumption is accumulated and statistically analyzed by the flow sensor, and then compared with its target value, which is set at 25L / m. If the consumption is abnormal, faults such as nozzle blockage and pulse valve leakage are checked.

[0023] Second Embodiment This application provides a bridge cutting dynamic correction system based on multimodal sensing, including a multimodal sensing module, an edge computing and fusion module, an intelligent decision-making module, an actuator module, and a closed-loop verification module; Multimodal sensing module: Composed of fiber optic grating sensor array, ultrasonic flaw detector, infrared thermal imager and laser profilometer, it realizes full-dimensional perception of cutting surface stress, temperature, rebar position and geometric deformation, and the specific processing method is the same as the processing process in step one. Edge computing and fusion module: Deploys FPGA chip and GPU acceleration unit to complete real-time fusion of multi-source data, noise suppression and feature extraction, and the specific processing method is the same as the processing process in step two; Intelligent decision-making module: includes a deep reinforcement learning prediction model and a fuzzy PID controller, outputs dynamic correction instructions, and the specific processing method is the same as the processing in step three; Actuator module: integrates servo hydraulic tension regulator, variable frequency motor driver and intelligent cooling system to achieve millisecond-level coordinated control of cutting parameters, and the specific processing method is the same as the processing process in step four; Closed-loop verification module: Through machine vision and fiber optic grating sensing feedback, a correction effect evaluation system is constructed to drive model self-optimization, and the specific processing method is the same as the processing process in step five.

[0024] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0025] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for dynamic correction of bridge cutting based on multimodal sensing, characterized in that, The method includes the following steps: Step 1: Deploy a fiber optic grating sensor array along the inner side of the cutting path, simultaneously configure a high-frequency ultrasonic flaw detection device, use the pulse reflection method to scan the three-dimensional coordinates of the steel bar to generate a thermal map of the steel bar distribution, and use an infrared thermal imager and a laser profilometer to capture the temperature field and trajectory geometric deformation of the cutting surface, respectively. Step 2: Collect 3D coordinate data, thermal imaging images and 3D point cloud data of steel bars, use wavelet transform for noise reduction, and use principal component analysis to extract core feature vectors, including mechanical features, structural features, thermal features and geometric features, to generate a 12-dimensional fusion feature dataset. Step 3: Construct a DQN model based on deep reinforcement learning. The input layer receives 12-dimensional feature data, and the output layer defines the action space, including the wire saw tension increment, cutting speed correction coefficient, and coolant flow level. Step 4: The servo hydraulic system constructs a real-time control link of command-execution-feedback. The variable frequency drive speed regulation system adopts space vector pulse width modulation technology to dynamically adjust the duty cycle according to the speed correction coefficient. The cooling system uses thermodynamic feedback to perform segmented flow control through electromagnetic pulse valves. Step 5: After correction, the laser profilometer scans the trajectory a second time to extract the residual between the actual and designed paths. If the residual does not meet the standard, iterative learning is triggered, the DQN model weights are updated online and three iterations are performed to automatically generate an efficiency report, including the number of rope breaks, average deviation, and energy consumption indicators.

2. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The noise reduction using wavelet transform includes performing three-level wavelet transform noise reduction using the db4 wavelet basis to eliminate 50Hz power frequency interference and environmental vibration noise ≤0.5Hz; and using K-nearest neighbor interpolation to complete missing data.

3. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The mechanical characteristics include the location coordinates of the stress abrupt change region, the duration of the abrupt change, and the peak stress. The location coordinates of the force mutation region are determined, the mutation center is identified using the gradient threshold method, the boundary of the mutation region is fitted using the least squares method, and the center coordinates of the boundary rectangle are output. The duration of the mutation is specifically the time interval from when the stress gradient first exceeds the threshold to when it recovers to below the threshold. Peak stress specifically represents the maximum stress value within the abrupt change region.

4. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The structural features include the distribution range of the densely reinforced area, the average diameter of the reinforcing bars, and the minimum protective layer thickness; The distribution range of the densely reinforced area is calculated by merging adjacent dense points through morphological dilation, and the radius and center coordinates of the circumcircle of the dense area are output. The average diameter of the reinforcing bars is calculated by taking the diameter of all reinforcing bars in the dense area through the peak spacing of the ultrasonic echo signal. The calculation formula is: diameter = wave velocity × time difference / 2, and the arithmetic mean is taken. The minimum protective layer thickness, specifically the shortest distance from the outer edge of the reinforcing bar to the concrete surface, is calculated using three-dimensional coordinate differences.

5. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The thermal characteristics include the center coordinates of the temperature anomaly hotspot, the highest temperature, and the diffusion coefficient; The center coordinates of the temperature anomaly hotspots are obtained by performing 8-neighbor connected component analysis on the infrared thermal image and identifying regions with a temperature rise rate ≥5℃ / s by calculating the temperature difference between two thermal images, and then outputting the center coordinates of the hotspot regions. Maximum temperature indicates the highest temperature value within the hotspot area; The thermal diffusivity is determined by fitting the heat conduction equation to the temperature field distribution in the hotspot region, and the specific equation is as follows: The diffusion coefficient was calculated.

6. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The geometric features include the flatness of the cut surface, the perpendicularity of the edge, and the curvature of the real-time trajectory; Cutting surface flatness: The least squares method is used to fit an ideal plane to the laser point cloud data, and the arithmetic mean of the vertical distance from each point to the plane is calculated. Edge perpendicularity: The angle between the edge line of the cut surface and the design perpendicular line, calculated by extracting the edge line through Hough transform; Real-time trajectory curvature: Perform circular arc fitting on the point cloud of the cutting trajectory and calculate the curvature value.

7. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The construction of the DQN model based on deep reinforcement learning includes an input layer that receives 12-dimensional normalized feature data, a hidden layer that uses a 3-layer fully connected network, and an output layer that defines a three-dimensional action space, including wire saw tension increment, cutting speed correction coefficient, and coolant flow rate level.

8. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The servo hydraulic system constructs a real-time control link of command-execution-feedback, which includes adopting a load-sensitive servo hydraulic system. The core actuator is a double-acting hydraulic cylinder. The inner loop uses a high-precision pressure sensor to provide real-time tension feedback, and the parameters are adjusted by a PID control algorithm. The outer loop uses an absolute linear encoder integrated into the hydraulic cylinder piston rod, and the position-speed dual-loop adjustment is adopted.

9. The method for dynamic correction of bridge cutting based on multimodal sensing according to claim 1, characterized in that, The specific method for segmented flow control via electromagnetic pulse valve is as follows: Extracting the highest temperature T of the cut surface based on temperature field data from an infrared thermal imager. max As a control benchmark, when T max When the temperature is ≤60℃, prioritize maintaining a low flow rate to reduce coolant consumption, and adopt a 5Hz intermittent injection mode, specifically with a valve opening cycle of 100ms / closing cycle; When 60℃ When T max At temperatures above 80℃, the electromagnetic pulse valve opens fully without interruption, and the standby pump starts, increasing the total flow rate to 3.0L / s.

10. A bridge cutting dynamic correction system based on multimodal sensing, executed by the correction method according to any one of claims 1-9, characterized in that, include: Multimodal sensing module: Composed of fiber optic grating sensor array, ultrasonic flaw detector, infrared thermal imager and laser profilometer, it realizes full-dimensional perception of cutting surface stress, temperature, rebar position and geometric deformation. Edge computing and fusion module: Deploys FPGA chips and GPU acceleration units to complete real-time fusion of multi-source data, noise suppression and feature extraction; Intelligent decision-making module: includes a deep reinforcement learning prediction model and a fuzzy PID controller, outputting dynamic correction commands; Actuator module: integrates servo hydraulic tension regulator, variable frequency motor driver and intelligent cooling system to achieve millisecond-level coordinated control of cutting parameters; Closed-loop verification module: Through machine vision and fiber optic grating sensing feedback, a correction effect evaluation system is constructed to drive model self-optimization.

Images