Method for monitoring damage of vibrating plate based on double-channel attention time convolution network

Abstract

The application discloses a vibrating plate damage monitoring method based on a double-channel attention time convolution network, which comprises the following steps: based on the original finite state machine control system of a leveling robot, matching self-cleaning process node trigger data acquisition, reusing a vibrating motor as an excitation source to collect vibration time sequence signals, and reusing a leveling head rotating device to collect clean plate surface images; after preprocessing, the double-channel attention time convolution network is input, internal implicit and surface apparent damage features are extracted respectively, and a damage grading result is output through cross-modal fusion; based on the grading result, a parameter adjustment strategy is generated through a deep Q network, the original control system is reused for execution, and the model is iteratively optimized through semi-supervised incremental learning. The scheme can realize non-inductive online monitoring, has the characteristics of strong anti-interference, active protection closed loop and online self-optimization.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and intelligent operation and maintenance technology for engineering machinery, specifically a method for monitoring damage to vibratory tamping slabs based on a dual-channel attention temporal convolutional network. Background Technology

[0002] The vibratory slab is the core operating component of the concrete leveling robot. It is subjected to alternating vibration loads and concrete corrosion over a long period of time, which can easily cause internal structural damage and surface corrosion deterioration. Accurate monitoring of its damage status is crucial to ensuring construction quality, ensuring equipment operation safety, and extending the service life of the equipment.

[0003] Existing technologies have two main shortcomings: First, most are stand-alone offline detection solutions that require additional dedicated vibration equipment and independent control units, resulting in high equipment modification costs. Furthermore, the detection process requires interrupting construction operations and cannot rely on the equipment's original self-cleaning process to achieve seamless online monitoring. Second, single-modal detection methods are weak against on-site interference. Vibration detection is affected by fluctuations in excitation parameters caused by switching vibration levels, leading to poor consistency of detection benchmarks. Visual detection is easily affected by residual water stains on the board surface, resulting in false damage misjudgments. At the same time, existing solutions generally only have damage identification functions and lack active protection and control mechanisms, which cannot slow down the evolution of equipment damage.

[0004] To address the aforementioned two technical shortcomings, a vibratory tamping plate damage monitoring solution is proposed. This solution can reuse the original hardware and control system of the equipment, embed a self-cleaning process to achieve seamless online monitoring, and simultaneously possess dual-modal anti-interference detection capabilities and an active protection and control mechanism. This reduces deployment costs while improving detection accuracy, achieving an upgrade from passive identification to active prevention and control. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring damage to vibratory treads based on a dual-channel attention temporal convolutional network, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network, comprising: Based on the existing finite state machine control system of the leveling robot, the data acquisition is triggered by matching the self-cleaning process nodes. The vibrating motor of the vibration desorption process is reused as the excitation source to collect vibration timing signals. The clean window period after brush cleaning and the leveling head rotation device are reused to collect images of the stain-free board surface to complete multimodal data preprocessing. The preprocessed vibration time series signal and the plate image are input into a dual-channel attention temporal convolutional network to extract internal latent damage features and surface apparent damage features respectively. After cross-modal feature fusion, the damage grading result of the vibrating plate is output. Based on the damage grading results, a deep Q-network decision model is constructed to adaptively generate self-cleaning parameters and vibration operation parameter adjustment strategies, which are then executed by reusing the original whole machine control system. Based on unlabeled samples continuously collected during the self-cleaning cycle, a dual-channel attention temporal convolutional network is iteratively optimized through semi-supervised incremental learning.

[0007] Furthermore, all self-cleaning processes, including spraying and wetting, vibration desorption, brush washing, and rotary sewage discharge, are set as independent state nodes of a finite state machine. The process state signals output by the original control system of the leveling robot are directly read as trigger commands for data acquisition actions, without the need to add an independent detection and control unit. When the finite state machine jumps to the vibration desorption state node, the vibration sensor deployed on the vibrating plate is simultaneously activated to collect vibration timing signals, and the original vibrating motor is reused as a common excitation source for the detection process. When the finite state machine jumps to the clean window state node before the brush washing ends and the rotary sewage discharge starts, the leveling head rotation device is driven to rotate the vibrating plate intermittently at a fixed angle, and the image acquisition device is simultaneously triggered to collect multi-view images of the vibrating plate surface.

[0008] Furthermore, a deep Q-network decision model is constructed, incorporating real-time damage level, damage area distribution, concrete grade, ambient temperature and humidity, and cumulative construction time into the state space input. A self-cleaning parameter action space is set, including spraying duration, desorption intensity, and fixed-point scrubbing frequency, while a vibration operation parameter action space is set, including vibration amplitude, vibration frequency, and travel speed. A multi-objective weighted reward function is used for iterative optimization, outputting an optimized parameter combination adapted to the current working conditions. Differentiated control logic is executed for different damage levels: standard parameters are maintained under normal conditions; cleaning intensity is enhanced for minor damage; cleaning is simultaneously enhanced and vibration load is reduced for moderate damage; and a shutdown warning command is triggered for severe damage. All parameter adjustments and warning commands are then sent to the existing overall control system of the leveling robot for execution.

[0009] Furthermore, a channel attention-enhanced temporal convolutional network is constructed, stacking multiple causal dilated convolutional layers with different dilation coefficients to expand the feature receptive field without increasing the network depth. This captures the vibration response variation patterns over a long temporal range at multiple scales, extracting latent damage features such as structural stiffness degradation and frequency shift contained in the vibration signal. A channel attention module is embedded at the back end of each causal dilated convolutional layer. Global average pooling aggregates the global information of each feature channel, and the importance weights of each channel are learned through a fully connected layer. The weights are then weighted and mapped to the corresponding feature channels, adaptively strengthening the contribution of feature channels strongly correlated with structural damage, suppressing interfering feature channels corresponding to environmental noise, and outputting a standardized temporal damage feature sequence.

[0010] Furthermore, a lightweight convolutional network with spatial attention is constructed, using depthwise separable convolution as the basic convolutional unit. The network backbone is built through the separation structure of depthwise convolution and pointwise convolution, and multi-scale apparent damage spatial features of the board surface image are extracted layer by layer. A spatial attention module is embedded in the back end of the network, and the channel dimension information of the feature map is aggregated by average pooling and max pooling to generate a pixel-level spatial weight map. Higher weights are assigned to the candidate pixel regions of damage in the image, weakening the irrelevant regional features corresponding to the basic texture of the board surface and residual water stains, and outputting a standardized apparent damage feature sequence.

[0011] Furthermore, a dimension-unified mapping process is performed on temporal damage features and apparent damage features to ensure that the two features have the same vector dimension and sequence length. Bidirectional cross-attention computation is used to achieve mutual verification and enhancement of the two types of features, using visual features to guide the positioning accuracy of temporal features and using temporal features to filter out false damage interference in visual features. Process status labels are introduced to adaptively adjust the fusion weights of the two features to generate a comprehensive damage feature vector.

[0012] Furthermore, the output of the damage grading result of the vibratory tamping plate after cross-modal feature fusion includes: constructing a multi-index damage evaluation system that includes vibration and visual dimensions; selecting stiffness degradation rate, harmonic distortion rate, and damping ratio change rate as evaluation indicators for the vibration dimension, and selecting crack length ratio, corrosion area ratio, and plate surface deformation as evaluation indicators for the visual dimension; constructing a corresponding hierarchical structure model and constructing a judgment matrix using the analytic hierarchy process (AHP); determining the weight coefficients corresponding to each evaluation indicator after consistency verification; constructing membership functions corresponding to different damage levels for each evaluation indicator; calculating the comprehensive membership degree of the sample to be tested corresponding to each damage level through fuzzy transformation; and outputting the final damage grading result of the vibratory tamping plate according to the principle of maximum membership degree.

[0013] Furthermore, a multivariable mapping calibration algorithm is constructed. The real-time speed, operating current, and set amplitude of the vibratory motor are read as inputs. The vibration acceleration signals collected under different excitation conditions are converted into equivalent structural response signals under the standard excitation reference through a pre-trained mapping model. This eliminates the vibration characteristic offset caused by the difference in excitation parameters and unifies the vibration signal reference under different operating modes. The calibrated vibration time series signal is processed by sliding window slicing. The long time series signal is divided into multiple time series samples of equal length according to a fixed time window. Then, amplitude normalization is performed on all samples to map the signal amplitude to the standard numerical range and generate standardized time series samples.

[0014] Furthermore, morphological filtering and grayscale correction were used to process the acquired multi-view images of the slab surface. Top-hat transformation was used to suppress abrupt grayscale changes in the bright water stain areas within the image, and adaptive grayscale equalization was used to adjust the image grayscale distribution to enhance texture details. Perspective transformation was used to correct the viewing angle of the oblique-view images acquired at different rotation angles, converting the obliquely acquired local images into a standard frontal view. Based on the image feature matching algorithm, feature points of each frontal local image were extracted, and feature matching and image alignment fusion were completed to stitch together a complete panoramic image covering the entire surface of the vibrating slab.

[0015] Furthermore, a teacher-student semi-supervised incremental learning framework is constructed. The currently deployed dual-channel attention temporal convolutional network is used as the teacher model to infer unlabeled multimodal samples collected during self-cleaning cycles. Samples with confidence scores higher than a preset threshold are selected to generate corresponding pseudo-labels. The original labeled dataset and the newly added pseudo-label samples are combined to form a mixed training set, which is then input into the student model for incremental fine-tuning training. During the training process, knowledge distillation loss is introduced to constrain the consistency of feature distribution between the student model and the teacher model. An interpretable algorithm is used to analyze the feature contribution of the damage determination results, statistically analyze the influence of each dimension of features on the damage classification, and combine the re-examination data to correct the damage classification determination threshold and membership function parameters, thus completing the iterative optimization of the dual-channel attention temporal convolutional network.

[0016] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention reuses native hardware and self-cleaning process windows in the original finite state machine control system of the leveling robot, achieving seamless embedding of monitoring functions without the need for additional independent detection and control units, significantly reducing modification costs and not occupying effective construction time; the dual-channel attention network customized for vibration and visual modalities, combined with a bidirectional cross-attention fusion mechanism, can achieve bidirectional verification enhancement of the two types of damage characteristics, and with process adaptive weights adapting to the data reliability of different acquisition stages, it effectively improves the damage detection accuracy and anti-interference capability in low signal-to-noise ratio environments at construction sites; the adaptive parameter control closed loop based on deep Q-networks can dynamically optimize cleaning and vibration operation parameters according to the damage state, realizing an upgrade from passive damage detection to active structural protection, delaying the evolution of equipment damage while taking into account construction efficiency and operational economy. Attached Figure Description

[0017] 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 a method for monitoring damage to tamping plates based on a dual-channel attention temporal convolutional network, provided by an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 The flowchart below illustrates a method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network, as provided in an embodiment of the present invention. Figure 1 It can be seen that the method for monitoring damage to vibrating slabs based on a dual-channel attention temporal convolutional network includes: Based on the existing finite state machine control system of the leveling robot, the data acquisition is triggered by matching the self-cleaning process nodes. The vibrating motor of the vibration desorption process is reused as the excitation source to collect vibration timing signals. The clean window period after brush cleaning and the leveling head rotation device are reused to collect images of the stain-free board surface to complete multimodal data preprocessing. The preprocessed vibration time series signal and the plate image are input into a dual-channel attention temporal convolutional network to extract internal latent damage features and surface apparent damage features respectively. After cross-modal feature fusion, the damage grading result of the vibrating plate is output. Based on the damage grading results, a deep Q-network decision model is constructed to adaptively generate self-cleaning parameters and vibration operation parameter adjustment strategies, which are then executed by reusing the original whole machine control system. Based on unlabeled samples continuously collected during the self-cleaning cycle, a dual-channel attention temporal convolutional network is iteratively optimized through semi-supervised incremental learning.

[0020] Furthermore, all self-cleaning processes, including spraying and wetting, vibration desorption, brush washing, and rotary sewage discharge, are set as independent state nodes of a finite state machine. The process state signals output by the original control system of the leveling robot are directly read as trigger commands for data acquisition actions, without the need to add an independent detection and control unit. When the finite state machine jumps to the vibration desorption state node, the vibration sensor deployed on the vibrating plate is simultaneously activated to collect vibration timing signals, and the original vibrating motor is reused as a common excitation source for the detection process. When the finite state machine jumps to the clean window state node before the brush washing ends and the rotary sewage discharge starts, the leveling head rotation device is driven to rotate the vibrating plate intermittently at a fixed angle, and the image acquisition device is simultaneously triggered to collect multi-view images of the vibrating plate surface.

[0021] Furthermore, a deep Q-network decision model is constructed, incorporating real-time damage level, damage area distribution, concrete grade, ambient temperature and humidity, and cumulative construction time into the state space input. A self-cleaning parameter action space is set, including spraying duration, desorption intensity, and fixed-point scrubbing frequency, while a vibration operation parameter action space is set, including vibration amplitude, vibration frequency, and travel speed. A multi-objective weighted reward function is used for iterative optimization, outputting an optimized parameter combination adapted to the current working conditions. Differentiated control logic is executed for different damage levels. Under normal conditions, standard parameters are maintained; for minor damage, cleaning intensity in the damaged area is enhanced; for moderate damage, cleaning is simultaneously enhanced while vibration load is reduced; and for severe damage, a shutdown warning command is triggered. All parameter adjustments and warning commands are then sent to the existing overall control system of the leveling robot for execution.

[0022] Preferably, the deep Q-network adopts a dual-network structure with the current evaluation network and the target network running in parallel. An experience replay pool is set up to store interactive samples, and the training stability is improved by periodically synchronizing the target network parameters, adapting to the embedded computing power constraints of the leveling robot and the fluctuating working conditions at the construction site. The model input layer directly connects to the preceding damage grading results and the equipment's native working condition data, covering scene features such as damage level, damage area distribution, concrete grade, ambient temperature and humidity, and cumulative construction time. The output layer maps to the equipment's natively adjustable self-cleaning parameters and vibration operation parameters, and commands can be directly issued to the existing whole-machine control system for execution. The model first completes offline pre-training based on simulation samples, and after deployment, it performs online incremental fine-tuning using field samples accumulated during the self-cleaning cycle. The inference process is embedded within the self-cleaning clean window and executed serially, without additionally occupying effective construction time.

[0023] The multi-objective weighted reward function is constructed around the core scenario requirements of active protection using vibratory tamping slabs. It integrates three optimization sub-objectives—damage propagation rate, construction efficiency loss, and clean energy consumption—using a weighted summation method. These sub-objectives correspond to three engineering demands: delaying structural failure, ensuring operational capacity, and optimizing operational economy. Damage propagation rate is the core optimization objective, with reward and penalty rules set based on the magnitude of damage level changes over a continuous period. Construction efficiency loss and clean energy consumption are constraining objectives, with corresponding rewards and penalties set based on the magnitude of efficiency decreases and energy consumption increases resulting from parameter adjustments. Each sub-objective has a differentiated weight coefficient, with the core objective having the highest weight. The weights can be flexibly adjusted according to different construction scenarios such as rushing work and maintenance, guiding the model to output parameter combinations that balance protective effectiveness and engineering practicality.

[0024] Furthermore, a channel attention-enhanced temporal convolutional network is constructed, stacking multiple causal dilated convolutional layers with different dilation coefficients to expand the feature receptive field without increasing the network depth. This captures the vibration response variation patterns over a long temporal range at multiple scales, extracting latent damage features such as structural stiffness degradation and frequency shift contained in the vibration signal. A channel attention module is embedded at the back end of each causal dilated convolutional layer. Global average pooling aggregates the global information of each feature channel, and the importance weights of each channel are learned through a fully connected layer. The weights are then weighted and mapped to the corresponding feature channels, adaptively strengthening the contribution of feature channels strongly correlated with structural damage, suppressing interfering feature channels corresponding to environmental noise, and outputting a standardized temporal damage feature sequence.

[0025] Preferably, the channel attention-enhanced temporal convolutional network is the temporal extraction branch in the dual-channel feature extraction architecture, customized for working conditions where the vibration signal of the vibrating plate has a low signal-to-noise ratio, latent damage evolves gradually over a long time, and the computing power of embedded devices is limited. The network adopts a structure of four-level causal dilated convolutional units and channel attention modules stacked in series. The input is a standardized vibration acceleration time-series sample that has been sliced ​​by sliding windows and normalized, and the output is a multi-scale temporal damage feature sequence that retains the correspondence of the time dimension. Among them, the causal convolution satisfies the temporal causality requirement of online monitoring, and the progressively increasing dilation coefficient can expand the feature receptive field without increasing the network depth, covering multiple vibration cycles to capture progressive damage such as stiffness degradation and frequency shift. The channel attention module adaptively strengthens the weights of damage-sensitive channels, suppresses noise such as construction noise and motor electromagnetic interference, and adapts to the low signal-to-noise ratio signal extraction requirements under complex on-site working conditions.

[0026] The model adopts a two-stage training mode of offline pre-training plus online incremental fine-tuning. In the offline stage, it is trained based on multi-condition vibration samples with damage labels, and the generalization is improved by simulating field condition fluctuations through data augmentation. In the online stage, the top-level parameters are fine-tuned by combining field samples accumulated during the self-cleaning cycle. While retaining the basic feature extraction capability, it adapts to the exclusive vibration characteristics of a single device. The inference process is triggered synchronously with the self-cleaning vibration desorption process, which can meet the real-time requirements on the embedded end.

[0027] Furthermore, a lightweight convolutional network with spatial attention is constructed, using depthwise separable convolution as the basic convolutional unit. The network backbone is built through the separation structure of depthwise convolution and pointwise convolution, and multi-scale apparent damage spatial features of the board surface image are extracted layer by layer. A spatial attention module is embedded in the back end of the network, and the channel dimension information of the feature map is aggregated by average pooling and max pooling to generate a pixel-level spatial weight map. Higher weights are assigned to the candidate pixel regions of damage in the image, weakening the irrelevant regional features corresponding to the basic texture of the board surface and residual water stains, and outputting a standardized apparent damage feature sequence.

[0028] Preferably, the spatial attention lightweight convolutional network is the visual branch of the dual-channel feature extraction architecture, custom-designed for conditions where water stains easily remain on the surface of the vibratory tamping plate, embedded computing power is limited, and damage scale varies greatly. The network constructs four levels of downsampling units with depthwise separable convolutions as the backbone, significantly compressing the number of parameters to adapt to edge computing power, while extracting multi-scale apparent damage features from micro-cracks to large-area corrosion layer by layer; the back-end of the network embeds a spatial attention module, which generates pixel-level spatial weight maps through pooling aggregation, adaptively strengthening the features of the damage candidate region and weakening background interference such as plate texture and residual water stains. Its input is a preprocessed frontal panoramic grayscale image of the vibratory tamping plate, and the output is an apparent damage feature map that retains the complete spatial position mapping relationship, which can directly correspond to specific areas of the vibratory tamping plate.

[0029] The model employs a training mode combining offline pre-training and online incremental fine-tuning. In the offline phase, training is conducted using labeled damage samples under various lighting and cleanliness levels, incorporating simulated water stains and lighting disturbances to enhance environmental robustness. In the online phase, top-level network parameters are fine-tuned based on field samples collected during the self-cleaning cycle, adapting to the specific panel characteristics of each device. The inference process is embedded within the self-cleaning cleanliness window, meeting the real-time requirements of embedded systems. The output spatial location features provide accurate visual positioning data for subsequent targeted brushing and graded load control.

[0030] Furthermore, a dimension-unified mapping process is performed on temporal damage features and apparent damage features to ensure that the two features have the same vector dimension and sequence length. Bidirectional cross-attention computation is used to achieve mutual verification and enhancement of the two types of features, using visual features to guide the positioning accuracy of temporal features and using temporal features to filter out false damage interference in visual features. Process status labels are introduced to adaptively adjust the fusion weights of the two features to generate a comprehensive damage feature vector.

[0031] Preferably, bidirectional cross-attention computation is the core unit of cross-modal feature fusion. A dual-path parallel interactive structure is constructed based on a scaled dot product attention mechanism to achieve bidirectional verification and enhancement of temporal vibration features and visual appearance features. Before computation, a fully connected projection layer unifies the two heterogeneous features to the same vector dimension. Then, two attention operations are performed: one uses temporal damage features as the query vector and visual features as the key and value vectors, utilizing structural response patterns to filter out false damage such as water stains and dirt without corresponding vibration anomalies in the visual representation; the other uses visual damage features as the query vector and temporal features as the key and value vectors, using spatial location information to calibrate the damage localization accuracy of the temporal features. Both operations output the corresponding modally enhanced global feature vectors.

[0032] This computational module features an adaptive weighting mechanism based on process status. It dynamically adjusts the fusion ratio of the two enhanced features according to different acquisition processes, such as vibration desorption and cleanroom windows, matching the signal-to-noise ratio differences between the two types of data under different processes and aligning with the operational mode of process-linked acquisition. During training, it is jointly optimized end-to-end with a dual-branch feature extraction network and a damage grading module. Cross-entropy loss serves as the core supervision, supplemented by modal consistency loss to enhance the feature correlation of real damage. Random modal mask data enhancement further improves on-site robustness. Finally, it outputs a comprehensive damage feature vector fusing structural response and spatial location information, supporting subsequent precise grading and targeted control.

[0033] Furthermore, the output of the damage grading result of the vibratory tamping plate after cross-modal feature fusion includes: constructing a multi-index damage evaluation system that includes vibration and visual dimensions; selecting stiffness degradation rate, harmonic distortion rate, and damping ratio change rate as evaluation indicators for the vibration dimension, and selecting crack length ratio, corrosion area ratio, and plate surface deformation as evaluation indicators for the visual dimension; constructing a corresponding hierarchical structure model and constructing a judgment matrix using the analytic hierarchy process (AHP); determining the weight coefficients corresponding to each evaluation indicator after consistency verification; constructing membership functions corresponding to different damage levels for each evaluation indicator; calculating the comprehensive membership degree of the sample to be tested corresponding to each damage level through fuzzy transformation; and outputting the final damage grading result of the vibratory tamping plate according to the principle of maximum membership degree.

[0034] Preferably, the Analytic Hierarchy Process (AHP) is the core method for weight allocation in the multi-index damage evaluation system of vibratory tamping slabs, used to transform qualitative engineering experience into quantitative index weight coefficients. This method constructs a three-level progressive hierarchical structure: "Target Layer - Criterion Layer - Index Layer." The target layer corresponds to the comprehensive damage level determination; the criterion layer is divided into two evaluation dimensions: vibration and visual; and the index layer corresponds to six quantifiable evaluation indicators: stiffness degradation rate, harmonic distortion rate, damping ratio change rate, crack length ratio, corrosion area ratio, and slab surface deformation. During calculation, a 1-9 scaling method is used to construct pairwise comparison judgment matrices for indicators at the same level. After solving the weight vector using the sum-product method and completing the consistency check, the comprehensive weight coefficients of each indicator are output and directly input into the subsequent fuzzy comprehensive evaluation module to support damage grading calculation.

[0035] This method follows the engineering principle of "prioritizing structural safety and supplementing with apparent damage" to set initial weights, while also supporting scenario-based adaptation and adjustment: in high-grade heavy-load construction scenarios, the weight priority of vibration dimension and stiffness degradation rate indicators is increased to ensure structural safety assessment; in coastal high-corrosion environments, the weight of visual dimension and rust area ratio indicators is increased to enhance the monitoring sensitivity of corrosion damage; in long-cycle continuous operation scenarios, the weight of damping ratio change rate and plate surface deformation is emphasized to match the assessment needs of fatigue cumulative damage, so that the weight system fits the damage assessment focus of different working conditions.

[0036] Furthermore, a multivariable mapping calibration algorithm is constructed. The real-time speed, operating current, and set amplitude of the vibratory motor are read as inputs. The vibration acceleration signals collected under different excitation conditions are converted into equivalent structural response signals under the standard excitation reference through a pre-trained mapping model. This eliminates the vibration characteristic offset caused by the difference in excitation parameters and unifies the vibration signal reference under different operating modes. The calibrated vibration time series signal is processed by sliding window slicing. The long time series signal is divided into multiple time series samples of equal length according to a fixed time window. Then, amplitude normalization is performed on all samples to map the signal amplitude to the standard numerical range and generate standardized time series samples.

[0037] Preferably, the multivariable mapping calibration algorithm is the core of vibration signal preprocessing. It is designed to address the pain points of excitation parameter fluctuations and inconsistent vibration response benchmarks caused by multiple gear switching in vibration operation. Its inputs are three types of original control parameters of the vibratory motor: real-time speed, operating current, and set amplitude, as well as the original vibration acceleration time-series signal of the corresponding time window. The core of the algorithm is a lightweight three-layer fully connected mapping network. By fitting the nonlinear correlation between control parameters and vibration response, it outputs amplitude correction coefficients and baseline offsets. It performs baseline translation and amplitude scaling correction on the original signal, converting it into an equivalent structural response signal under a standard excitation benchmark. After benchmark consistency verification, sliding window slicing, and amplitude normalization, it generates standardized time-series samples and outputs them to the subsequent time-series damage feature extraction network.

[0038] This algorithm employs a two-stage approach: offline calibration followed by on-site fine-tuning. In the offline stage, under structurally sound conditions, it traverses all motor speeds to construct a calibration dataset and trains a general mapping model covering the entire range of operating parameters. In the on-site stage, a small number of samples are collected for specific scenarios such as high-speed heavy-load and low-temperature operation to fine-tune the top-level parameters, quickly adapting to actual load characteristics. After deployment, the algorithm updates the correction coefficients in real time as the motor speed changes, operating automatically without manual intervention. This completely eliminates the interference of excitation parameter fluctuations on damage assessment, enabling a single detection model to adapt to all operating scenarios and improving the consistency and accuracy of damage detection results under different operating conditions.

[0039] Furthermore, morphological filtering and grayscale correction were used to process the acquired multi-view images of the slab surface. Top-hat transformation was used to suppress abrupt grayscale changes in the bright water stain areas within the image, and adaptive grayscale equalization was used to adjust the image grayscale distribution to enhance texture details. Perspective transformation was used to correct the viewing angle of the oblique-view images acquired at different rotation angles, converting the obliquely acquired local images into a standard frontal view. Based on the image feature matching algorithm, feature points of each frontal local image were extracted, and feature matching and image alignment fusion were completed to stitch together a complete panoramic image covering the entire surface of the vibrating slab.

[0040] Furthermore, a teacher-student semi-supervised incremental learning framework is constructed. The currently deployed dual-channel attention-based temporal convolutional network is used as the teacher model to infer information from unlabeled multimodal samples collected during self-cleaning cycles. Samples with confidence scores higher than a preset threshold are selected to generate corresponding pseudo-labels. The original labeled dataset and the newly added pseudo-labeled samples form a mixed training set, which is then input into the student model for incremental fine-tuning training. Knowledge distillation loss is introduced during training to constrain the consistency of feature distributions between the student and teacher models. An interpretable algorithm is used to analyze the feature contribution of the damage assessment results, statistically analyzing the influence of each dimension of features on damage grading. The damage grading threshold and membership function parameters are then corrected based on re-examination data, completing the iterative optimization of the dual-channel attention-based temporal convolutional network.

[0041] 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.

[0042] 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 method for monitoring damage of a vibrating plate based on a dual-channel attention time convolutional network, characterized in that, include: Based on the existing finite state machine control system of the leveling robot, the data acquisition is triggered by matching the self-cleaning process nodes. The vibrating motor of the vibration desorption process is reused as the excitation source to collect vibration timing signals. The clean window period after brush cleaning and the leveling head rotation device are reused to collect images of the stain-free board surface to complete multimodal data preprocessing. The preprocessed vibration time series signal and the plate image are input into a dual-channel attention temporal convolutional network to extract internal latent damage features and surface apparent damage features respectively. After cross-modal feature fusion, the damage grading result of the vibrating plate is output. Based on the damage grading results, a deep Q-network decision model is constructed to adaptively generate self-cleaning parameters and vibration operation parameter adjustment strategies, which are then executed by reusing the original whole machine control system. Based on unlabeled samples continuously collected during the self-cleaning cycle, a dual-channel attention temporal convolutional network is iteratively optimized through semi-supervised incremental learning.

2. The method of claim 1, wherein the method is based on a dual-channel attention time convolutional network. The matching of data collection triggered at each process node of the self-cleaning process includes: The self-cleaning processes of spraying and wetting, vibration desorption, brush washing, and rotary sewage discharge are each set as independent state nodes of the finite state machine. The process state signals output by the original control system of the leveling robot are directly read as the trigger commands for data acquisition actions, without the need to add an independent detection and control unit. When the finite state machine jumps to the vibration desorption state node, the vibration sensor deployed on the vibrating plate is activated simultaneously to collect the vibration timing signal, and the original vibrating motor is reused as the common excitation source for the detection process. When the finite state machine jumps to the clean window state node before the brush washing ends and the rotary sewage discharge starts, the leveling head rotation device is driven to drive the vibrating plate to rotate intermittently at a fixed angle, and the image acquisition device is triggered simultaneously to collect multi-view images of the vibrating plate surface.

3. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 1, characterized in that, The adaptive generation and adjustment strategy for self-cleaning parameters and vibration operation parameters includes: A deep Q-network decision-making model is constructed, incorporating real-time damage level, damage area distribution, concrete grade, ambient temperature and humidity, and cumulative construction time into the state space input. A self-cleaning parameter action space is set, including spraying duration, desorption intensity, and fixed-point scrubbing frequency, while a vibration operation parameter action space is set, including vibration amplitude, vibration frequency, and travel speed. A multi-objective weighted reward function is used for iterative optimization, outputting an optimized parameter combination adapted to the current working conditions. Differentiated control logic is executed for different damage levels. Under normal conditions, standard parameters are maintained; for minor damage, cleaning intensity in the damaged area is enhanced; for moderate damage, cleaning is simultaneously enhanced while reducing vibration load; and for severe damage, a shutdown warning command is triggered. All parameter adjustments and warning commands are sent to the existing overall control system of the leveling robot for execution.

4. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 1, characterized in that, The extraction of internal latent damage features and surface apparent damage features includes: A channel attention-enhanced temporal convolutional network is constructed by stacking multiple causal dilated convolutional layers with different dilation coefficients. This expands the receptive field of features without increasing the network depth, capturing the vibration response variation patterns over a long temporal range at multiple scales, and extracting latent damage features such as structural stiffness degradation and frequency shift contained in the vibration signal. A channel attention module is embedded at the back end of each causal dilated convolutional layer. Global average pooling is used to aggregate the global information of each feature channel. The importance weights of each channel are learned through a fully connected layer, and the weights are weighted and mapped to the corresponding feature channels. This adaptively strengthens the contribution of feature channels that are strongly correlated with structural damage, suppresses the interference feature channels corresponding to environmental noise, and outputs a standardized temporal damage feature sequence.

5. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 1, characterized in that, The extracted surface apparent damage features include: A lightweight convolutional network with spatial attention is constructed, using depthwise separable convolution as the basic convolutional unit. The network backbone is built through the separation structure of depthwise convolution and pointwise convolution, and multi-scale apparent damage spatial features of the board surface image are extracted layer by layer. A spatial attention module is embedded in the back end of the network. The channel dimension information of the feature map is aggregated by average pooling and max pooling to generate a pixel-level spatial weight map. Higher weights are assigned to the candidate pixel regions of damage in the image, weakening the irrelevant regional features corresponding to the basic texture of the board surface and residual water stains, and outputting a standardized apparent damage feature sequence.

6. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 1, characterized in that, The damage grading results of the vibratory tamping plate output after cross-modal feature fusion include: A unified dimensional mapping process is performed on temporal damage features and apparent damage features to ensure that the two features have the same vector dimension and sequence length. Bidirectional cross-attention calculation is used to achieve mutual verification and enhancement of the two types of features. Visual features guide the positioning accuracy of temporal features, and temporal features filter out false damage interference in visual features. Process status labels are introduced to adaptively adjust the fusion weight of the two features and generate a comprehensive damage feature vector.

7. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 6, characterized in that, The damage grading results of the vibratory tamping plate output after cross-modal feature fusion include: The output of the damage grading result of the vibratory tamping plate after cross-modal feature fusion includes: constructing a multi-index damage evaluation system that includes vibration and visual dimensions. The vibration dimension selects stiffness degradation rate, harmonic distortion rate, and damping ratio change rate as evaluation indicators, while the visual dimension selects crack length ratio, corrosion area ratio, and plate surface deformation as evaluation indicators; constructing a corresponding hierarchical structure model and constructing a judgment matrix using the analytic hierarchy process (AHP), and determining the weight coefficients corresponding to each evaluation indicator after consistency verification; constructing membership functions corresponding to different damage levels for each evaluation indicator, calculating the comprehensive membership degree of the sample to be tested for each damage level through fuzzy transformation, and outputting the final damage grading result of the vibratory tamping plate according to the principle of maximum membership degree.

8. The method for monitoring damage to tamping slabs based on a dual-channel attention temporal convolutional network according to claim 2, characterized in that, The completion of multimodal data preprocessing includes: A multivariable mapping calibration algorithm is constructed, which reads the real-time speed, operating current, and set amplitude of the vibratory motor as inputs. Through a pre-trained mapping model, the vibration acceleration signals collected under different excitation conditions are converted into equivalent structural response signals under a standard excitation reference, eliminating the vibration characteristic offset caused by the difference in excitation parameters and unifying the vibration signal reference under different operating modes. The calibrated vibration time series signal is processed by sliding window slicing, which divides the long time series signal into multiple time series samples of equal length according to a fixed time window. Then, amplitude normalization is performed on all samples to map the signal amplitude to a standard numerical range and generate standardized time series samples.

9. The method for monitoring damage to a vibrating slab based on a dual-channel attention temporal convolutional network according to claim 2, characterized in that, The completion of multimodal data preprocessing includes: The acquired multi-view slab images were processed using morphological filtering and grayscale correction. Top-hat transformation was used to suppress abrupt grayscale changes in bright water stain areas within the images, and adaptive grayscale equalization was used to adjust the image grayscale distribution to enhance texture details. Perspective transformation was used to correct the viewing angle of the oblique-view images acquired at different rotation angles, converting the obliquely acquired local images into a standard frontal view. Feature points of each frontal view local image were extracted based on an image feature matching algorithm, and feature matching and image alignment fusion were completed to stitch together a complete panoramic image covering the entire surface of the vibrating slab.

10. The method for monitoring damage to a vibrating slab based on a dual-channel attention temporal convolutional network according to claim 1, characterized in that, The method of iteratively optimizing the dual-channel attention temporal convolutional network through semi-supervised incremental learning includes: A teacher-student semi-supervised incremental learning framework was constructed. The currently deployed dual-channel attention-based temporal convolutional network was used as the teacher model to infer from unlabeled multimodal samples collected during self-cleaning cycles. Samples with confidence scores higher than a preset threshold were selected to generate corresponding pseudo-labels. The original labeled dataset and the newly added pseudo-labeled samples were combined to form a mixed training set, which was then input into the student model for incremental fine-tuning training. During training, a knowledge distillation loss was introduced to constrain the consistency of feature distribution between the student and teacher models. An interpretable algorithm was used to analyze the feature contribution of the damage assessment results, statistically analyzing the influence of each dimension of features on damage grading. The damage grading threshold and membership function parameters were corrected based on re-examination data, completing the iterative optimization of the dual-channel attention-based temporal convolutional network.

Images