Deep learning-based timing coordination control method for self-cleaning of vibrating plate operation gap

Abstract

The present application relates to the technical field of device intelligent cooperation. Specifically, it is a method for timing coordination control of self-cleaning of vibrating plate operation gap based on deep learning, which reuses construction vibrating motor as a cleaning vibration source, realizes accurate perception of bonding state through an improved segmentation network guided by time-varying damping, realizes space-time decoupling cooperation of brush and vibration relying on wave propagation delay compensation mechanism, dynamically plans cleaning timing combining reinforcement learning model with hard constraint parameters, and completes the whole process of self-cleaning in the operation gap with the supporting double-scale closed-loop correction and precision and life double-domain calibration mechanism. The present application does not need to add new hardware and does not occupy construction time, significantly improves the cleaning thoroughness and scene adaptability, at the same time offsets the precision drift and fatigue loss caused by motor reuse, effectively reduces the risk of cracking and breaking of vibrating plate, prolongs the service life of the device, and improves the continuous construction efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent collaborative control technology for equipment, specifically a time-series collaborative control method for self-cleaning during the operation interval of a vibrating tamping plate based on deep learning. Background Technology

[0002] Concrete floor leveling robots rely on vibratory slabs to complete concrete vibration and leveling operations. During the operation, concrete easily clumps and adheres to the surface of the slab. Manual cleaning is time-consuming and labor-intensive, and residual clumps can cause uneven stress on the slab surface, corrosion and aging, significantly increasing the risk of cracking and breakage of the slab. Existing leveling robots lack in-situ fully automatic self-cleaning control systems. A few devices equipped with cleaning mechanisms only use fixed serial timing control, failing to combine the vibration dynamics of the slab itself to achieve multi-process coordination. At the same time, conventional control algorithms only focus on cleanliness as a single objective, without considering the negative impact of cleaning operations on the vibration construction accuracy and fatigue life of the slab, making it difficult to complete slab surface cleaning efficiently and without damage during construction breaks. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based method for the self-cleaning timing coordination control of the working gap of vibratory tamping slabs, so as to solve the problems mentioned in the background art.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a self-cleaning timing collaborative control method for the working gap of a vibratory plate based on deep learning, which is applied to a concrete floor leveling robot equipped with a robot body, a leveling head rotating device, a vibratory plate, a vibratory motor that has both construction vibration and cleaning and desorption functions, a spraying device, a water storage device, a cleaning brush structure and an embedded control system. The robot collects operational data and uses a time-series prediction model to predict the total duration and confidence interval of the current work interval. The operational data includes walking speed, start / stop status of the vibrating motor, and historical process time-series data. A comprehensive determination is made, considering the remaining fatigue life margin of the vibrating plate, to determine whether to trigger the self-cleaning mode. Only when the interval duration meets the minimum cleaning threshold and the life margin is sufficient, a start command and hard constraint parameters for available time are output. The available time is the maximum allowable cleaning execution time obtained by deducting the mechanism reset reserve time, construction command response buffer time, and abnormal emergency margin from the total work interval duration. Upon receiving the start command, the vibrating motor is controlled to vibrate for a short time, and images and vibration data of the working surface of the vibrating plate are collected simultaneously. The distribution and strength of concrete bonding are evaluated through an improved segmentation network, and bonding status data is output. Based on the total duration of the work interval, the remaining fatigue life margin of the vibrating plate, and the bonding status data, a reinforcement learning model with hard constraint parameters is used to generate time-series collaborative operation parameters for each actuator to complete the four-stage process in stages, including low-frequency micro-vibration spraying for wetting, rated strong vibration intermittent spraying for desorption, synchronous continuous spraying and brushing with brushes, and rotating and tilting pulse vibration for slag discharge; real-time operation feedback data is collected synchronously; the collaborative operation parameters include spray flow rate, staged vibration frequency of the vibrating motor, brush travel rhythm, and leveling head tilt angle; Based on real-time operational feedback data, the cleaning parameters are dynamically adjusted using a fast and slow dual-scale MPC algorithm. After cleaning is completed, the construction accuracy is calibrated, and the construction accuracy drift and fatigue damage of the vibrating plate caused by cleaning vibration are compensated.

[0005] Furthermore, the specific method for predicting the total duration of the current work gap and the confidence interval using a time-series prediction model is as follows: The system collects real-time operational status data of the leveling robot, including walking speed, vibratory motor working status, leveling head posture, and current working surface position coordinates. It also retrieves historical process interval data and on-site environmental temperature and humidity data from the same batch of construction. All data undergo outlier removal, timestamp alignment, and normalization to generate a standardized time-series feature sequence with a fixed time step, which serves as input data for the prediction model. The standardized temporal feature sequence is input into the temporal fusion Transformer model. First, the gating feature selection layer performs adaptive weight allocation on the features of each dimension to filter out redundant and interfering features. Then, the multi-head self-attention layer models the temporal dependency relationship of historical construction procedures, captures the periodic pattern of procedure switching and the on-site scheduling correlation features, and outputs the encoded deep temporal feature vector. The deep temporal feature vector is input into the quantile regression output layer of the model. At the same time, the predicted duration values ​​corresponding to multiple preset quantiles are calculated, and the predicted total duration of the current work gap is output, as well as the upper and lower limits of the duration corresponding to the 95% confidence level, forming the gap duration prediction result with confidence assessment. The predicted lower limit of the gap duration is used as a hard constraint on the available duration and compared with the minimum cleaning time threshold corresponding to the current construction condition. At the same time, the cumulative fatigue damage data of the vibrating plate is retrieved to calculate the remaining life margin. The self-cleaning mode is triggered only when the lower limit of the available duration is greater than or equal to the minimum cleaning time and the remaining life margin meets the preset safety threshold, and the available duration constraint parameters are passed to the downstream timing planning step. If the constraint conditions are not met, the robot's operating status is continuously monitored and the gap duration prediction results are dynamically updated.

[0006] Furthermore, the method for calculating the remaining life margin by retrieving the cumulative fatigue damage data of the vibrating plate is as follows: For the stress concentration area of ​​the vibratory tamping plate, namely the connection area between the vibratory motor mounting base and the plate surface that is prone to cracking and breakage, the stress amplitude-fatigue life SN curve of the critical section is calibrated by combining the material properties, structural dimensions and finite element modal analysis results of the vibratory tamping plate, and the limit number of vibration cycles that the material can withstand under different vibration stress amplitudes is determined. Based on the stress amplitude-fatigue life SN curve, the benchmark total life threshold of the vibrating plate of the leveling robot under the rated construction vibration parameters is calibrated. It can be expressed as the equivalent total number of vibration cycles or the equivalent total working time, corresponding to the critical value of damage d=1 (when the damage amount accumulates to 1, it is determined that the life is exhausted). The operating parameters of the vibratory motor during each construction operation are recorded in real time, including vibration frequency, amplitude, and duration of a single continuous vibration. Through a pre-calibrated mapping relationship between motor vibration parameters and stress amplitude at the critical section, the construction vibration parameters are converted into the working stress amplitude and corresponding number of vibration cycles at the critical section. The mapping relationship between motor vibration parameters and stress amplitude at the critical section needs to be combined with the geometry of the vibratory slab, material properties, installation position of the vibratory motor, and boundary constraints, and is obtained through finite element dynamics simulation or on-site strain calibration. This addresses the question of how much stress fluctuation the vibratory slab actually experiences at the critical location given a set of vibration parameters.

[0007] According to Miner's linear accumulation criterion, the fatigue damage generated by each construction operation is accumulated successively to obtain the total cumulative construction damage D=∑ni=1(Ni / N0i), where Ni is the actual number of vibration cycles under the corresponding stress amplitude during the i-th construction, and N0i is the limit number of cycles under the stress amplitude. For the four vibration stages of the self-cleaning process, including low-frequency micro-vibration wetting, rated strong vibration desorption, intermittent vibration anti-sticking, and pulse vibration slag discharge, the vibration frequency, amplitude, duty cycle, and duration of the vibrating motor for each stage are matched. Through the same set of motor vibration parameters-critical section stress amplitude mapping relationship, the fatigue damage is calculated and summed stage by stage to obtain the fatigue damage increment Dc of a single self-cleaning cycle. In the time sequence planning stage, the PPO decision network can pre-calculate the damage increment of the corresponding scheme based on the candidate cleaning time sequence parameters, and use it as a hard constraint condition for decision-making to avoid outputting high-intensity cleaning schemes that exceed the lifespan tolerance. The remaining fatigue life margin is represented by a normalized value, ranging from 0 to 1. The higher the value, the more sufficient the remaining life. The calculation formula is R=d-(D+∑Dc), where ∑Dc is the sum of damage increments from each self-cleaning operation.

[0008] Furthermore, the specific method for evaluating the distribution and strength of concrete bonds using an improved segmentation network and outputting bond state data is as follows: Upon receiving the self-cleaning mode start command, the vibrating motor is controlled to perform short-term vibration detection at a fixed low frequency; high-definition images of the working surface of the vibrating plate and vibration acceleration and operating current time-series data of the vibrating motor are acquired simultaneously; the images are normalized in size and grayscale calibrated, and the vibration time-series data are filtered, denoised and converted in spectrum to obtain standardized visual input data and vibration input data respectively. Visual input data is fed into the encoder of the improved semantic segmentation network. Features are extracted layer by layer through a multi-layer convolutional downsampling structure, generating shallow feature maps containing surface texture information and deep feature maps containing block semantic information in sequence, thus completing the hierarchical extraction of multi-scale visual features. The vibration input data is input into a bidirectional temporal feature extraction network to extract the damping features of each region of the slab surface that characterize the bond strength of the concrete. The damping features are converted into weight matrices corresponding to the spatial locations. The visual feature maps of each level of the encoder are then modulated region by region through a feature weighting module to increase the feature weights of high bond strength regions and decrease the feature weights of low bond strength regions, resulting in a modulated feature map that integrates bond strength information. The modulated multi-scale feature map is input into the decoder, and the modulation features of the corresponding level are fused through skip connections. After upsampling, the image is restored to the original image size, and a pixel-level concrete bonding area segmentation mask is output. Based on the segmentation mask, the bonding area ratio, the spatial distribution of agglomerates, and the estimated average thickness are calculated. Combined with the damping features, complete bonding state data containing distribution and strength information is generated.

[0009] Furthermore, the specific method for generating time-series coordinated operation parameters for each actuator to complete the four-stage process in stages is as follows: The remaining fatigue life margin of the vibratory plate is set as a hard constraint for fatigue loss, and the rated parameters of the vibratory motor, spraying device, cleaning brush structure, and leveling head rotation device are set as hard constraints. The bonding area ratio, agglomerate thickness distribution, and plate surface damping matrix in the bonding state data are used as the initial state. All hard constraints limit the feasible domain of the reinforcement learning action space and directly filter out invalid action parameters that exceed the constraint boundaries. A temporal decision network is constructed using a near-end strategy optimization algorithm. The initial state and hard constraint boundary are used as inputs, and time slices are divided with a granularity of hundreds of milliseconds. The time-series collaborative operation parameters of each actuator are output slice by slice. The network has built-in rules for compensating for the propagation delay of the plate vibration wave and rules for synchronizing the phase of the brush vibration. The parameters are automatically generated to meet the requirements of spatiotemporal decoupling, so as to avoid the mutual cancellation of vibration and brushing effects. Based on the time-series collaborative operation parameters, four parallel cleaning processes are automatically mapped: the first stage is mapped to a wetting process with continuous water supply from the spray device and low-frequency intermittent vibration of the vibrating motor; the second stage is mapped to a desorption process with the rated frequency vibration of the vibrating motor and intermittent water spray from the spray device; the third stage is mapped to a brushing process with reciprocating brushing, continuous water supply from the spray device, and intermittent anti-sticking vibration of the vibrating motor; and the fourth stage is mapped to a sewage discharge process with the rotating and tilting of the leveling head and pulse vibration of the vibrating motor. Overlapping transition periods are set between adjacent processes to achieve smooth connection of actions. The time-series parameters are scored and verified by a reward function weighted by three factors: cleanliness, total time, and fatigue loss. The parameters are iteratively adjusted until all hard constraints are met and the reward value reaches the target. The final time-series collaborative operation parameters are output to drive each actuator to complete the cleaning operation in sequence.

[0010] Furthermore, based on real-time operational feedback data, the specific method for dynamically adjusting cleaning parameters using a fast and slow dual-scale MPC algorithm is as follows: Execution is carried out in seconds. The real-time collected board surface cleanliness feedback data and remaining cleaning time are input. The pre-planned baseline time parameters are used as a reference trajectory. Under the constraint of the remaining available time, the global parameter adjustment is solved by rolling optimization. The total water flow of the spray device, the overall vibration intensity of the vibrating motor, and the total number of back-and-forth brushing times of the cleaning brush structure are adjusted accordingly to ensure that the cleaning effect matches the time constraint. Execution is performed in milliseconds. The system takes into account the vibration feedback data of the vibrating motor and the real-time position data of the cleaning brush, and calculates the real-time vibration phase of the cleaning brush at the current position point by point based on the pre-constructed vibration wave propagation delay matrix of the board surface. The system dynamically fine-tunes the up-and-down movement rhythm and travel speed of the brush so that the moment when the brush bristles contact the surface of the vibrating board always corresponds to the vibration wave trough at that position, thus avoiding the vibration from canceling out the brushing effect.

[0011] Furthermore, the specific methods for calibrating construction accuracy and compensating for construction accuracy drift and fatigue damage to the vibratory tamping plate caused by cleaning vibration are as follows: The system starts at the end of the cleaning process. It inputs the running time, average power and temperature rise data of the vibrating motor during this cleaning process. Based on the pre-built mapping relationship between motor thermal deformation and parameter drift, it calculates the offset of construction vibration parameters caused by cleaning vibration. It then compensates and corrects the frequency and amplitude parameters of the vibrating motor in the construction mode to offset the accuracy drift caused by the cleaning conditions and ensure the leveling accuracy of subsequent operations. Input the vibration parameters and duration of this cleaning operation, calculate the fatigue damage increment based on the fatigue damage accumulation model of the vibratory tamping plate, dynamically optimize the continuous vibration duration and vibration frequency range of subsequent construction, compensate for the life loss caused by cleaning vibration, and reduce the risk of cracking and breakage of the vibratory tamping plate.

[0012] Compared with the prior art, the beneficial effects achieved by the present invention are: in the process of multiple cleaning processes running in parallel, the present invention eliminates the mutual interference between high-frequency vibration and brush washing action, ensures the stable effect of each cleaning mechanism, adapts to the vibration propagation characteristics of large-size vibrating plates, and makes the cleaning effect uniform and consistent on all parts of the plate surface. This invention utilizes a multi-stage composite cleaning process involving spraying, vibration, brushing, and rotating waste removal to effectively remove concrete residue adhering to the surface of the vibratory tamping plate. This reduces damage to the vibratory tamping plate caused by concrete corrosion and stress concentration, lowers the risk of cracking and deformation, extends the service life of the equipment, and eliminates the need for manual cleaning, thus reducing the labor intensity of operators and the downtime of the equipment. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of the self-cleaning timing collaborative control method for the working interval of vibratory tamping plates based on deep learning, as proposed in this invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please see Figure 1 The present invention provides a technical solution: a deep learning-based self-cleaning timing collaborative control method for the working gap of a vibratory plate, which is applied to a concrete floor leveling robot equipped with a robot body, a leveling head rotating device, a vibratory plate, a vibratory motor that combines construction vibration and cleaning and desorption functions, a spraying device, a water storage device, a cleaning brush structure and an embedded control system. The robot collects operational data and uses a time-series prediction model to predict the total duration and confidence interval of the current work interval. The operational data includes walking speed, start / stop status of the vibrating motor, and historical process time-series data. A comprehensive determination is made, considering the remaining fatigue life margin of the vibrating plate, to determine whether to trigger the self-cleaning mode. Only when the interval duration meets the minimum cleaning threshold and the life margin is sufficient, a start command and hard constraint parameters for available time are output. The available time is the maximum allowable cleaning execution time obtained by deducting the mechanism reset reserve time, construction command response buffer time, and abnormal emergency margin from the total work interval duration. Upon receiving the start command, the vibrating motor is controlled to vibrate for a short time, and images and vibration data of the working surface of the vibrating plate are collected simultaneously. The distribution and strength of concrete bonding are evaluated through an improved segmentation network, and bonding status data is output. Based on the total duration of the work interval, the remaining fatigue life margin of the vibrating plate, and the bonding status data, a reinforcement learning model with hard constraint parameters is used to generate time-series collaborative operation parameters for each actuator to complete the four-stage process in stages, including low-frequency micro-vibration spraying for wetting, rated strong vibration intermittent spraying for desorption, synchronous continuous spraying and brushing with brushes, and rotating and tilting pulse vibration for slag discharge; real-time operation feedback data is collected synchronously; the collaborative operation parameters include spray flow rate, staged vibration frequency of the vibrating motor, brush travel rhythm, and leveling head tilt angle; Based on real-time operational feedback data, the cleaning parameters are dynamically adjusted using a fast and slow dual-scale MPC algorithm. After cleaning is completed, the construction accuracy is calibrated, and the construction accuracy drift and fatigue damage of the vibrating plate caused by cleaning vibration are compensated.

[0016] Furthermore, the specific method for predicting the total duration and confidence interval of the current work gap using a time-series prediction model is as follows: The system collects real-time operational status data of the leveling robot, including walking speed, vibratory motor working status, leveling head posture, and current working surface position coordinates. It also retrieves historical process interval data and on-site environmental temperature and humidity data from the same batch of construction. All data undergo outlier removal, timestamp alignment, and normalization to generate a standardized time-series feature sequence with a fixed time step, which serves as input data for the prediction model. The standardized temporal feature sequence is input into the temporal fusion Transformer model. First, the gating feature selection layer performs adaptive weight allocation on the features of each dimension to filter out redundant and interfering features. Then, the multi-head self-attention layer models the temporal dependency relationship of historical construction procedures, captures the periodic pattern of procedure switching and the on-site scheduling correlation features, and outputs the encoded deep temporal feature vector. The deep temporal feature vector is input into the quantile regression output layer of the model. At the same time, the predicted duration values ​​corresponding to multiple preset quantiles are calculated, and the predicted total duration of the current work gap is output, as well as the upper and lower limits of the duration corresponding to the 95% confidence level, forming the gap duration prediction result with confidence assessment. Specifically, after the standardized temporal feature sequence is input into the temporal fusion Transformer model, it sequentially passes through a gated feature selection layer, a multi-head self-attention layer, and a quantile regression output layer to output the final gap duration prediction result. The gated feature selection layer uses a gated linear unit structure with a 32-dimensional output. It generates a channel-wise feature weight mask using the Sigmoid function to adaptively assign weights to the original features and filter out redundant interference features. The multi-head self-attention layer uses four parallel attention heads, each with an 8-dimensional feature dimension, resulting in a total output dimension of 32. It employs a scaled dot product attention mechanism to analyze historical construction procedures. The temporal dependencies are modeled to capture the periodic patterns of process switching and the correlation features with on-site scheduling, and the encoded deep temporal feature vector is output. The quantile regression output layer first passes through a fully connected hidden layer with a 16-dimensional output dimension. The ReLU activation function is used and L2 regularization is used to constrain the weights to avoid overfitting. Finally, the quantile prediction output layer is used to calculate the predicted duration values ​​corresponding to the three preset quantiles of 0.025, 0.5, and 0.975, and outputs the predicted total duration of the current work gap, as well as the upper and lower limits of the duration corresponding to the 95% confidence level, forming a gap duration prediction result with confidence assessment.

[0017] Preferably, the training set of the temporal fusion Transformer model consists of historical process time series samples covering different construction scenarios, different concrete grades, and different scheduling rhythms. Quantile loss is used as the loss function, Adam optimizer is used, the initial learning rate is set to 0.001, and the number of training epochs is set to 80.

[0018] The predicted lower limit of the gap duration is used as a hard constraint on the available duration and compared with the minimum cleaning time threshold corresponding to the current construction condition. At the same time, the cumulative fatigue damage data of the vibrating plate is retrieved to calculate the remaining life margin. The self-cleaning mode is triggered only when the lower limit of the available duration is greater than or equal to the minimum cleaning time and the remaining life margin meets the preset safety threshold, and the available duration constraint parameters are passed to the downstream timing planning step. If the constraint conditions are not met, the robot's operating status is continuously monitored and the gap duration prediction results are dynamically updated.

[0019] Furthermore, the method for calculating the remaining life margin by retrieving the cumulative fatigue damage data of the vibrating plate is as follows: For the stress concentration area of ​​the vibratory tamping plate, namely the connection area between the vibratory motor mounting base and the plate surface that is prone to cracking and breakage, the stress amplitude-fatigue life SN curve of the critical section is calibrated by combining the material properties, structural dimensions and finite element modal analysis results of the vibratory tamping plate, and the limit number of vibration cycles that the material can withstand under different vibration stress amplitudes is determined. Based on the stress amplitude-fatigue life SN curve, the benchmark total life threshold of the vibrating plate of the leveling robot under the rated construction vibration parameters is calibrated. It can be expressed as the equivalent total number of vibration cycles or the equivalent total working time, corresponding to the critical value of damage d=1 (when the damage amount accumulates to 1, it is determined that the life is exhausted). The operating parameters of the vibratory motor during each construction operation are recorded in real time, including vibration frequency, amplitude, and duration of a single continuous vibration. Through a pre-calibrated mapping relationship between motor vibration parameters and stress amplitude at the critical section, the construction vibration parameters are converted into the working stress amplitude and corresponding number of vibration cycles at the critical section. The mapping relationship between motor vibration parameters and stress amplitude at the critical section needs to be combined with the geometry of the vibratory slab, material properties, installation position of the vibratory motor, and boundary constraints, and is obtained through finite element dynamics simulation or on-site strain calibration. This addresses the question of how much stress fluctuation the vibratory slab actually experiences at the critical location given a set of vibration parameters.

[0020] According to Miner's linear accumulation criterion, the fatigue damage generated by each construction operation is accumulated successively to obtain the total cumulative construction damage D=∑ni=1(Ni / N0i), where Ni is the actual number of vibration cycles under the corresponding stress amplitude during the i-th construction, and N0i is the limit number of cycles under the stress amplitude. For the four vibration stages of the self-cleaning process, including low-frequency micro-vibration wetting, rated strong vibration desorption, intermittent vibration anti-sticking, and pulse vibration slag discharge, the vibration frequency, amplitude, duty cycle, and duration of the vibrating motor for each stage are matched. Through the same set of motor vibration parameters-critical section stress amplitude mapping relationship, the fatigue damage is calculated and summed stage by stage to obtain the fatigue damage increment Dc of a single self-cleaning cycle. In the time sequence planning stage, the PPO decision network can pre-calculate the damage increment of the corresponding scheme based on the candidate cleaning time sequence parameters, and use it as a hard constraint condition for decision-making to avoid outputting high-intensity cleaning schemes that exceed the lifespan tolerance. The remaining fatigue life margin is represented by a normalized value, ranging from 0 to 1. The higher the value, the more sufficient the remaining life. The calculation formula is R=d-(D+∑Dc), where ∑Dc is the sum of damage increments from each self-cleaning operation.

[0021] The margin adopts a dual-node update mechanism: the construction damage item is updated after each construction operation is completed; the cleaning damage item is updated after each self-cleaning process is completed and the dual-domain calibration is completed; the current real-time margin value is directly called in the cleaning access judgment stage. When the margin is lower than the preset minimum threshold, the high-damage cleaning mode is automatically restricted, and only light cleaning or cleaning is performed, thereby reducing the risk of stress concentration and fracture of the vibratory slab from the control level.

[0022] Furthermore, the specific method for evaluating the distribution and strength of concrete bonds using an improved segmentation network and outputting bond state data is as follows: Upon receiving the self-cleaning mode start command, the vibrating motor is controlled to perform short-term vibration detection at a fixed low frequency; high-definition images of the working surface of the vibrating plate and vibration acceleration and operating current time-series data of the vibrating motor are acquired simultaneously; the images are normalized in size and grayscale calibrated, and the vibration time-series data are filtered, denoised and converted in spectrum to obtain standardized visual input data and vibration input data respectively. Visual input data is fed into the encoder of the improved semantic segmentation network. Features are extracted layer by layer through a multi-layer convolutional downsampling structure, generating shallow feature maps containing surface texture information and deep feature maps containing block semantic information in sequence, thus completing the hierarchical extraction of multi-scale visual features. Specifically, the encoder adopts a lightweight four-level downsampling architecture based on depthwise separable convolution. Each downsampling unit achieves half the resolution through a depthwise convolution with a stride of 2. Each downsampling unit is followed by a spatial dimension damped feature modulation module. The first-level downsampling unit outputs a feature map with 32 channels. It uses a 3×3 depthwise separable convolution to extract the bottom edge texture features and employs the ReLU6 activation function. It also uses a batch normalization layer to accelerate network convergence. The corresponding damping modulation module performs pixel-wise spatial weighting on the feature map based on the real-time vibration damping space matrix to accurately match regional damping differences under high resolution. The second-level downsampling unit outputs a feature map with 64 channels. It uses an inverted residual structure to enhance feature representation capabilities and employs the ReLU6 activation function. L2 regularization constraints are uniformly applied to the convolution kernel weights across all levels to avoid overfitting. The corresponding damping modulation module dynamically adjusts the feature response weights of different block regions based on the damping partitioning results. The third-level downsampling unit outputs a feature map with 128 channels. Downsampling is completed through a depthwise convolution with a stride of 2. The ReLU6 activation function is used to extract the abstract semantic features of the agglomeration region. The corresponding damping modulation module increases the weight of the high consolidation strength category features from the semantic channel dimension. The fourth-level downsampling unit outputs a feature map with 256 channels. It is then connected to the hollow spatial pyramid pooling module to fuse multi-scale receptive field information and output a deep semantic feature map, which is activated by the ReLU6 activation function. The modulated feature maps output from each downsampling stage are passed to the corresponding levels of the decoder via skip connections. The decoder performs four levels of bilinear interpolation upsampling combined with convolution operations to fuse with shallow features layer by layer. Finally, the output layer generates a pixel-level concrete bonding region segmentation mask and discrete bonding strength level distribution results that match the size of the input image.

[0023] The vibration input data is input into a bidirectional temporal feature extraction network to extract the damping features of each region of the slab surface that characterize the bond strength of the concrete. The damping features are converted into weight matrices corresponding to the spatial locations. The visual feature maps of each level of the encoder are then modulated region by region through a feature weighting module to increase the feature weights of high bond strength regions and decrease the feature weights of low bond strength regions, resulting in a modulated feature map that integrates bond strength information. Specifically, the vibration input data is fed into a bidirectional temporal feature extraction network, then passes through two layers of bidirectional long short-term memory networks and a spatial mapping module to output damping weight matrices for each region of the board surface. The first layer of the bidirectional long short-term memory network has a hidden layer dimension of 64 dimensions, uses the ReLU activation function, and outputs a 128-dimensional temporal feature that is spliced ​​forward and backward. The second layer of the bidirectional long short-term memory network has a hidden layer dimension of 32 dimensions, uses the ReLU activation function, and uses L2 regularization to constrain the weights to avoid overfitting, outputting a 64-dimensional deep temporal feature that is spliced ​​forward and backward. Subsequently, a global damping feature vector is output through a fully connected layer, and combined with a pre-constructed board surface vibration wave propagation delay reference matrix, a two-dimensional damping weight matrix aligned with the spatial size of the visual feature map is obtained. Finally, the feature weighting module performs region-by-region weighted modulation on the visual feature maps of each level of the encoder, increasing the feature weights of high adhesion strength regions and decreasing the feature weights of low adhesion strength regions, to obtain a modulated feature map that integrates adhesion strength information. The weights output by the feature weighting module are normalized to the 0-1 range by the Sigmoid function.

[0024] Preferably, the training set of the time-varying damping guided U-Net network consists of images of vibrating slabs with different bonding levels and corresponding vibration time-series paired samples. The loss function is a weighted combination of cross-entropy loss and Dice loss. The Adam optimizer is used, with an initial learning rate of 0.001 and 80 training epochs. During the training phase, data augmentation strategies such as random cropping and brightness perturbation are used to improve the model's generalization ability.

[0025] The modulated multi-scale feature map is input into the decoder, and the modulation features of the corresponding level are fused through skip connections. After upsampling, the image is restored to the original image size, and a pixel-level concrete bonding area segmentation mask is output. Based on the segmentation mask, the bonding area ratio, the spatial distribution of agglomerates, and the estimated average thickness are calculated. Combined with the damping features, complete bonding state data containing distribution and strength information is generated.

[0026] Specifically, the modulated multi-scale feature map is input into the decoder, and the modulation features of the corresponding level are fused through skip connections. After upsampling, the image is restored to its original size, and a pixel-level concrete bonding region segmentation mask is output. The decoder contains four feature restoration levels. The first level first enlarges the feature map size by 2 times through bilinear upsampling, and then concatenates it with the damping modulation features output from the corresponding level of the encoder. The fused features are then extracted by a 3×3 depthwise separable convolutional layer with 128 channels. The ReLU activation function is used, and L2 regularization is used to constrain the weights to avoid overfitting. The second level also first performs upsampling by 2 times, concatenates the modulation features of the corresponding level, and then inputs a 6-channel layer. A 3×3 depthwise separable convolutional layer with a depth of 4 is used, employing the ReLU activation function and incorporating L2 regularization constraints. After a 2x upsampling in the third stage, a 3×3 depthwise separable convolutional layer with 32 channels is connected, also employing the ReLU activation function. After a 2x upsampling to restore the original image size in the fourth stage, a 1×1 convolutional output layer with 1 channel is connected. The output is then normalized using the Sigmoid function to obtain the probability value of the concrete bonding area corresponding to each pixel, generating the final binary segmentation mask. Based on the segmentation mask, the bonding area ratio, agglomerate spatial distribution, and average thickness estimate are calculated. Combined with damping features, complete bonding state data containing distribution and strength information is generated.

[0027] Preferably, the training set of the adhesive region segmentation model is a sample set of images of the working surface of the vibratory treadmill covering three levels of adhesiveness: light, moderate and heavy. The model uses binary cross-entropy loss as the loss function, Adam optimizer, initial learning rate set to 0.001, training epochs set to 80, and online data augmentation strategies such as random flipping and brightness perturbation are introduced during the training phase to improve the model's generalization ability.

[0028] Furthermore, the specific method for generating time-series coordinated operation parameters for each actuator to complete the four-stage process in stages is as follows: The remaining fatigue life margin of the vibratory plate is set as a hard constraint for fatigue loss, and the rated parameters of the vibratory motor, spraying device, cleaning brush structure, and leveling head rotation device are set as hard constraints. The bonding area ratio, agglomerate thickness distribution, and plate surface damping matrix in the bonding state data are used as the initial state. All hard constraints limit the feasible domain of the reinforcement learning action space and directly filter out invalid action parameters that exceed the constraint boundaries. A temporal decision network is constructed using a near-end strategy optimization algorithm. The initial state and hard constraint boundary are used as inputs, and time slices are divided with a granularity of hundreds of milliseconds. The time-series collaborative operation parameters of each actuator are output slice by slice. The network has built-in rules for compensating for the propagation delay of the plate vibration wave and rules for synchronizing the phase of the brush vibration. The parameters are automatically generated to meet the requirements of spatiotemporal decoupling, so as to avoid the mutual cancellation of vibration and brushing effects. Based on the time-series collaborative operation parameters, four parallel cleaning processes are automatically mapped: the first stage is mapped to a wetting process with continuous water supply from the spray device and low-frequency intermittent vibration of the vibrating motor; the second stage is mapped to a desorption process with the rated frequency vibration of the vibrating motor and intermittent water spray from the spray device; the third stage is mapped to a brushing process with reciprocating brushing, continuous water supply from the spray device, and intermittent anti-sticking vibration of the vibrating motor; and the fourth stage is mapped to a sewage discharge process with the rotating and tilting of the leveling head and pulse vibration of the vibrating motor. Overlapping transition periods are set between adjacent processes to achieve smooth connection of actions. The time-series parameters are scored and verified by a reward function weighted by three factors: cleanliness, total time, and fatigue loss. The parameters are iteratively adjusted until all hard constraints are met and the reward value reaches the target. The final time-series collaborative operation parameters are output to drive each actuator to complete the cleaning operation in sequence.

[0029] Furthermore, based on real-time operational feedback data, the specific method for dynamically adjusting cleaning parameters using a fast and slow dual-scale MPC algorithm is as follows: Execution is carried out in seconds. The real-time collected board surface cleanliness feedback data and remaining cleaning time are input. The pre-planned baseline time parameters are used as a reference trajectory. Under the constraint of the remaining available time, the global parameter adjustment is solved by rolling optimization. The total water flow of the spray device, the overall vibration intensity of the vibrating motor, and the total number of back-and-forth brushing times of the cleaning brush structure are adjusted accordingly to ensure that the cleaning effect matches the time constraint. Execution is performed in milliseconds. The system takes into account the vibration feedback data of the vibrating motor and the real-time position data of the cleaning brush, and calculates the real-time vibration phase of the cleaning brush at the current position point by point based on the pre-constructed vibration wave propagation delay matrix of the board surface. The system dynamically fine-tunes the up-and-down movement rhythm and travel speed of the brush so that the moment when the brush bristles contact the surface of the vibrating board always corresponds to the vibration wave trough at that position, thus avoiding the vibration from canceling out the brushing effect.

[0030] Furthermore, the specific methods for calibrating construction accuracy and compensating for construction accuracy drift and fatigue damage to the vibratory tamping plate caused by cleaning vibration are as follows: The system starts at the end of the cleaning process. It inputs the running time, average power and temperature rise data of the vibrating motor during this cleaning process. Based on the pre-built mapping relationship between motor thermal deformation and parameter drift, it calculates the offset of construction vibration parameters caused by cleaning vibration. It then compensates and corrects the frequency and amplitude parameters of the vibrating motor in the construction mode to offset the accuracy drift caused by the cleaning conditions and ensure the leveling accuracy of subsequent operations. Input the vibration parameters and duration of this cleaning operation, calculate the fatigue damage increment based on the fatigue damage accumulation model of the vibratory tamping plate, dynamically optimize the continuous vibration duration and vibration frequency range of subsequent construction, compensate for the life loss caused by cleaning vibration, and reduce the risk of cracking and breakage of the vibratory tamping plate.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for the time-series collaborative control of self-cleaning during the operation interval of a vibrating tamping plate, characterized in that: Collect robot operation data, predict the total duration and confidence interval of the current work interval through a time-series prediction model, and comprehensively determine whether to trigger the self-cleaning mode by combining the remaining fatigue life margin of the vibrating plate, and output the start command and available time hard constraint parameters. Upon receiving the start command, the vibrating motor is controlled to vibrate for a short time, and images and vibration data of the working surface of the vibrating plate are collected simultaneously. The distribution and strength of concrete bonding are evaluated through an improved segmentation network, and bonding status data is output. Based on the total working interval time, the remaining fatigue life margin of the vibrating plate, and the bonding status data, a reinforcement learning model with hard constraint parameters is used to generate time-series collaborative operation parameters for each actuator to complete the four-stage process in stages, while simultaneously collecting real-time operation feedback data. Based on real-time operational feedback data, the cleaning parameters are dynamically adjusted using a fast and slow dual-scale MPC algorithm. After cleaning is completed, the construction accuracy is calibrated, and the construction accuracy drift and fatigue damage of the vibrating plate caused by cleaning vibration are compensated.

2. The self-cleaning timing collaborative control method for vibratory tamping plate operation intervals based on deep learning according to claim 1, characterized in that: The specific method for predicting the total duration and confidence interval of the current work gap using a time-series prediction model is as follows: The system collects real-time operational status data of the leveling robot, including walking speed, vibratory motor working status, leveling head posture, and current working surface position coordinates. It also retrieves historical process interval data and on-site environmental temperature and humidity data from the same batch of construction. All data undergo outlier removal, timestamp alignment, and normalization to generate a standardized time-series feature sequence with a fixed time step, which serves as input data for the prediction model. The standardized temporal feature sequence is input into the temporal fusion Transformer model. First, the gating feature selection layer performs adaptive weight allocation on the features of each dimension to filter out redundant and interfering features. Then, the multi-head self-attention layer models the temporal dependency relationship of historical construction procedures, captures the periodic pattern of procedure switching and the on-site scheduling correlation features, and outputs the encoded deep temporal feature vector. The deep temporal feature vector is input into the quantile regression output layer of the model. At the same time, the predicted duration values ​​corresponding to multiple preset quantiles are calculated. The predicted total duration of the current work gap is output, as well as the upper and lower limits of the duration corresponding to the confidence level, forming the gap duration prediction result with confidence assessment. The predicted lower limit of the gap duration is used as a hard constraint on the available time, and compared with the minimum cleaning time threshold corresponding to the current construction conditions. At the same time, the cumulative fatigue damage data of the vibrating plate is retrieved to calculate the remaining life margin. The self-cleaning mode is triggered only when the lower limit of the available time is greater than or equal to the minimum cleaning time and the remaining life margin meets the preset safety threshold. If the constraint conditions are not met, the robot's operating status is continuously monitored and the gap duration prediction results are dynamically updated.

3. The deep learning-based self-cleaning timing collaborative control method for the working interval of a vibratory tamping plate according to claim 2, characterized in that: The method for calculating the remaining life margin by retrieving the cumulative fatigue damage data of the vibratory tamping plate is as follows: By combining the material properties, structural dimensions and finite element modal analysis results of the vibratory tamping plate in the stress concentration area, the stress amplitude-fatigue life curve of the critical section is calibrated. Based on the stress amplitude-fatigue life curve, the benchmark total life threshold of the vibrating plate of the leveling robot under the rated construction vibration parameters is determined. It can be expressed as the equivalent total number of vibration cycles or the equivalent total working time, corresponding to the critical value of damage d=1. The operating parameters of the vibrating motor are recorded in real time during each construction operation. Through the pre-calibrated mapping relationship between motor vibration parameters and stress amplitude of critical sections, the construction vibration parameters are converted into the working stress amplitude of the critical section and the corresponding number of vibration cycles. According to Miner's linear accumulation criterion, the fatigue damage generated by each construction operation is accumulated successively to obtain the total cumulative construction damage D=∑ni=1(Ni / N0i), where Ni is the actual number of vibration cycles under the corresponding stress amplitude during the i-th construction, and N0i is the limit number of cycles under the stress amplitude. For the four vibration stages of the self-cleaning process, the vibration frequency, amplitude, duty cycle and duration of the vibrating motor in each stage are matched respectively. Through the same set of motor vibration parameters-critical section stress amplitude mapping relationship, the fatigue damage is calculated and summed stage by stage to obtain the fatigue damage increment Dc of a single self-cleaning. The remaining fatigue life margin is expressed as a normalized value and is calculated using the formula R = d - (D + ∑Dc), where ∑Dc is the sum of damage increments from each self-cleaning operation.

4. The deep learning-based self-cleaning timing collaborative control method for the working interval of a vibrating tamping plate according to claim 3, characterized in that: The specific method for evaluating the distribution and strength of concrete bonds using an improved segmentation network and outputting bond state data is as follows: Upon receiving the self-cleaning mode start command, the vibrating motor is controlled to perform short-term vibration detection at a fixed low frequency; high-definition images of the working surface of the vibrating plate and vibration acceleration and operating current time-series data of the vibrating motor are acquired simultaneously; the images are normalized in size and grayscale calibrated, and the vibration time-series data are filtered, denoised and converted in spectrum to obtain standardized visual input data and vibration input data respectively. Visual input data is fed into the encoder of the improved semantic segmentation network. Features are extracted layer by layer through a multi-layer convolutional downsampling structure, generating shallow feature maps containing surface texture information and deep feature maps containing block semantic information in sequence, thus completing the hierarchical extraction of multi-scale visual features. The vibration input data is input into a bidirectional temporal feature extraction network to extract the damping features of each region of the slab surface that characterize the bond strength of the concrete. The damping features are converted into weight matrices corresponding to the spatial locations. The visual feature maps of each level of the encoder are then modulated region by region through a feature weighting module to increase the feature weights of high bond strength regions and decrease the feature weights of low bond strength regions, resulting in a modulated feature map that integrates bond strength information. The modulated multi-scale feature map is input into the decoder, and the modulation features of the corresponding level are fused through skip connections. After upsampling, the image is restored to the original image size, and a pixel-level concrete bonding area segmentation mask is output. Based on the segmentation mask, the bonding area ratio, the spatial distribution of agglomerates, and the estimated average thickness are calculated. Combined with the damping features, complete bonding state data containing distribution and strength information is generated.

5. The deep learning-based self-cleaning timing collaborative control method for the working interval of a vibrating tamping plate according to claim 4, characterized in that: The specific method for generating time-series coordinated operation parameters for each actuator to complete the four-stage process in stages is as follows: The remaining fatigue life margin of the vibratory plate is set as a hard constraint of fatigue loss, and the rated parameters of the vibratory motor, spraying device, cleaning brush structure, and leveling head rotation device are set as hard constraints of hardware. Using the bonding area ratio, agglomerate thickness distribution, and plate damping matrix in the bonding state data as the initial state, all hard constraints limit the feasible region of the reinforcement learning action space. A temporal decision network is constructed using a proximal policy optimization algorithm. The initial state and hard constraint boundary are used as inputs. The network is divided into time slices, and the time-series collaborative operation parameters of each actuator are output slice by slice. The network incorporates a propagation delay compensation rule for board surface vibration waves and a phase synchronization rule for brush vibration. Based on the time-series collaborative operation parameters, the system automatically corresponds to four parallel cleaning processes. The time-series parameters are scored and verified by a reward function that weights cleaning cleanliness, total time consumption, and fatigue loss. The parameters are iteratively adjusted until all hard constraints are met and the reward value reaches the target. The final time-series collaborative operation parameters are then output to drive each actuator to complete the cleaning operation in sequence.

6. The deep learning-based self-cleaning timing collaborative control method for the working interval of a vibratory tamping plate according to claim 5, characterized in that: Based on real-time operational feedback data, the specific method for dynamically adjusting cleaning parameters using a fast and slow dual-scale MPC algorithm is as follows: Execution is carried out in seconds. The real-time collected board surface cleanliness feedback data and remaining cleaning time are input. The pre-planned baseline time parameters are used as a reference trajectory. Under the constraint of the remaining available time, the global parameter adjustment is solved by rolling optimization. The total water flow of the spray device, the overall vibration intensity of the vibrating motor, and the total number of back-and-forth brushing times of the cleaning brush structure are adjusted accordingly to ensure that the cleaning effect matches the time constraint. Execution is performed in milliseconds. The system takes into account the vibration feedback data of the vibrating motor and the real-time position data of the cleaning brush, and calculates the real-time vibration phase of the cleaning brush at the current position point by point based on the pre-constructed vibration wave propagation delay matrix of the board surface. The system dynamically fine-tunes the up-and-down movement rhythm and travel speed of the brush so that the moment when the brush bristles contact the surface of the vibrating board always corresponds to the vibration wave trough at that position.

7. The deep learning-based self-cleaning timing collaborative control method for the working interval of a vibrating tamping plate according to claim 6, characterized in that: The specific methods for calibrating construction accuracy and compensating for construction accuracy drift and fatigue damage to vibratory tampers caused by cleaning vibration are as follows: The system starts at the end of the cleaning process, inputs the running time, average power and temperature rise data of the vibrating motor during this cleaning process, calculates the offset of construction vibration parameters caused by cleaning vibration based on the pre-built mapping relationship between motor thermal deformation and parameter drift, and compensates and corrects the frequency and amplitude parameters of the vibrating motor in the construction mode to offset the accuracy drift caused by the cleaning conditions. Input the vibration parameters and duration of this cleaning operation, calculate the fatigue damage increment based on the fatigue damage accumulation model of the vibratory tamping plate, dynamically optimize the continuous vibration duration and vibration frequency range of subsequent construction, and compensate for the life loss caused by cleaning vibration.

8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images