An automatic knocking detection and identification system for building wall hollow

Abstract

The application discloses a kind of building wall hollow automatic knocking detection and identification system, including mobile platform module, standardization knocking execution module, multimodal signal acquisition module, hollow feature intelligent identification and self-learning module and data management and visualization module;Wherein, mobile platform module is connected with standardization knocking execution module, data management and visualization module respectively;Standardization knocking execution module is connected with multimodal signal acquisition module, hollow feature intelligent identification and self-learning module respectively;Hollow feature intelligent identification and self-learning module are internally provided with small sample meta-learning submodule, multimodal attention fusion submodule and online self-learning submodule.This application system is deployed to different construction projects, and has quick adaptive capacity when facing different wall materials, without re-collecting mass data to train model, saving preparation time and manpower cost in early stage, avoid the performance of traditional model to appear sudden drop in cross-scene application.

Description

Technical Field

[0001] This invention relates to the field of building wall hollow detection technology, specifically an automated tapping detection and identification system for building wall hollowness. Background Technology

[0002] Building wall hollowness detection and identification refers to a non-destructive testing process that utilizes professional physical detection and signal analysis technologies to accurately locate and quantitatively assess detachment and voids between the wall finish layer and the base layer in building construction projects, caused by poor adhesion, shrinkage cracking, or improper construction. Traditional techniques primarily involve manual tapping of the wall with handheld testing tools, relying on the inspector's auditory experience to determine the location and extent of hollowness. This method is inefficient and highly susceptible to subjective influences. With the development of sensing and intelligent technologies, modern hollowness identification has gradually transitioned to instrument-based detection, primarily using infrared thermal imaging, ultrasonic arrays, and impact echo methods. Infrared thermal imaging quickly pinpoints abnormal areas by capturing temperature differences on the wall surface, while intelligent detectors based on sound or vibration combine artificial intelligence and machine learning algorithms to automatically extract and analyze the time-frequency domain characteristics of the tapping signal, effectively filtering out environmental noise interference and achieving accurate quantitative calculation of the hollow area. This technology is widely used in the completion acceptance of new construction projects, renovation of old residential areas, and regular safety inspections of existing buildings. It is a key technical means to prevent the risk of injury from falling exterior wall finishes, ensure public safety, and improve the level of project quality supervision.

[0003] However, existing intelligent detectors only collect acoustic or vibration signals of a single modality and identify them using pre-trained traditional machine learning models. These models rely on a large number of labeled samples for training, requiring retraining with new data for different wall materials. This results in long adaptation cycles, high costs, and significant performance fluctuations across different application scenarios. Furthermore, single-modal signal acquisition is susceptible to environmental noise interference, has limited feature dimensions, and struggles to comprehensively cover the signal characteristics of different types of hollow areas, leading to high false positive and false negative rates. This fails to meet the high-precision detection requirements of complex engineering scenarios and lacks a self-learning optimization mechanism, preventing the model's performance from continuously improving with the accumulation of detection data. Summary of the Invention

[0004] The purpose of this invention is to provide an automated tapping detection and identification system for hollow walls in buildings, in order to overcome the shortcomings of existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An automated tapping detection and identification system for hollow walls in buildings includes a mobile platform module, a standardized tapping execution module, a multimodal signal acquisition module, an intelligent identification and self-learning module for hollow features, and a data management and visualization module. The motion control output of the mobile platform module is connected to the position calibration input of the standardized tapping execution module, and the status feedback output of the mobile platform module is connected to the device status acquisition input of the data management and visualization module. The action trigger output of the standardized striking execution module is connected to the synchronous acquisition input of the multimodal signal acquisition module, and the force parameter output of the standardized striking execution module is connected to the prior parameter input of the hollow drum feature intelligent recognition and self-learning module. The feature encoding output terminal of the multimodal signal acquisition module is connected to the multimodal feature input terminal of the hollow drum feature intelligent recognition and self-learning module, and the raw data output terminal of the multimodal signal acquisition module is connected to the raw data storage input terminal of the data management and visualization module. The detection result output terminal of the hollow drum feature intelligent recognition and self-learning module is connected to the result input terminal of the data management and visualization module. The model update parameter output terminal of the hollow drum feature intelligent recognition and self-learning module is self-connected to form a self-learning closed loop. The material adaptation instruction output terminal of the hollow drum feature intelligent recognition and self-learning module is back-connected to the force adjustment input terminal of the standardized tapping execution module. The path planning instruction output of the data management and visualization module is connected to the path execution input of the mobile platform module. The model optimization trigger output of the data management and visualization module is connected to the online learning input of the hollow drum feature intelligent recognition and self-learning module. The user interaction instruction output of the data management and visualization module is connected to the emergency control input of the mobile platform module and the start / stop control input of the standardized tapping execution module, respectively.

[0006] Furthermore, the mobile platform module internally includes a motion control unit, a position calibration component, a status feedback module, and a path execution engine. The motion control unit comprises a DC servo driver, a stepper motor control circuit, and motion trajectory generation logic. The DC servo driver receives pulse signals and outputs speed-regulating current. The stepper motor control circuit is used for forward and reverse rotation and speed adjustment of the motor. The motion trajectory generation logic generates a continuous motion path based on preset coordinate points. The position calibration component is equipped with a laser displacement sensor, a mechanical limit switch, and a coordinate calibration algorithm. The laser displacement sensor collects the relative distance between the platform and the wall in real time. The mechanical limit switch sets the boundary range of the platform's movement. The coordinate calibration algorithm corrects the platform's current coordinates using sensor data. The status feedback module consists of a speed encoder, an attitude sensor, and a data transmission interface. The speed encoder collects motor speed data in real time. The attitude sensor detects the platform's tilt angle and vibration amplitude. The data transmission interface converts the status data into a standard communication protocol format. The path execution engine includes a path parser, a motion planner, and an emergency stop controller. The path parser breaks down the input path instructions into individual coordinate points. The motion planner calculates the motion speed and acceleration between adjacent coordinate points. The emergency stop controller receives external commands to trigger a motor power-off mechanism.

[0007] Furthermore, the standardized striking execution module internally includes a striking drive assembly, a force adjustment unit, a position alignment module, and an action trigger controller. The striking drive assembly comprises an electromagnetic striking device, a striking head replacement mechanism, and a power supply circuit. The electromagnetic striking device drives the striking head to reciprocate via electromagnetic force. The striking head replacement mechanism uses a snap-fit ​​structure to achieve quick disassembly and installation of the striking head. The power supply circuit provides stable DC voltage and current. The force adjustment unit is equipped with a pressure sensor, an electromagnetic force adjustment circuit, and force calibration logic. The pressure sensor collects data on the reaction force when the striking head contacts the wall. The electromagnetic force adjustment circuit adjusts the force by changing the electromagnetic wire... The striking force is adjusted by the magnitude of the coil current, and the force calibration logic corrects the force output value based on sensor data. The position alignment module includes a visual positioning sensor, a position adjustment motor, and an alignment algorithm. The visual sensor identifies marked points on the wall surface, the position adjustment motor drives the striking drive component to move horizontally and vertically, and the alignment algorithm calculates the position adjustment amount based on sensor data. The action trigger controller includes a synchronization signal receiver, a trigger logic circuit, and an action timer. The synchronization signal receiver receives external trigger commands, the trigger logic circuit controls the start, stop, and movement cycle of the striker, and the action timer records the duration and interval of a single strike.

[0008] Furthermore, the multimodal signal acquisition module internally includes an acoustic signal acquisition unit, a vibration signal acquisition component, a synchronization control module, and a feature encoding engine. The acoustic signal acquisition unit comprises a MEMS microphone array, a preamplifier, and a filtering circuit. The MEMS microphone array consists of four microphones arranged in a rectangular layout. The preamplifier amplifies the weak signals acquired by the microphones, and the filtering circuit filters out low-frequency and high-frequency noise from the environment. The vibration signal acquisition component is equipped with a piezoelectric accelerometer, a signal conditioning circuit, and an AD converter. The piezoelectric accelerometer is fixed to the impact drive component to acquire vibration data, and the signal conditioning circuit amplifies and... The filtering process involves an AD converter that converts analog signals into digital signals. The synchronization control module includes a high-precision clock generator, a signal synchronizer, and a trigger signal interface. The high-precision clock generator provides a clock signal with microsecond-level accuracy, the signal synchronizer aligns the tap trigger signal with the acquisition signal in time, and the trigger signal interface receives external synchronization commands to control the acquisition start time. The feature encoding engine includes a feature extractor, a data compressor, and an encoding format converter. The feature extractor extracts characteristic parameters such as formants and decay times from the original signal, the data compressor uses a lossless compression algorithm to reduce data storage space, and the encoding format converter converts the feature data into a standard JSON format.

[0009] Furthermore, the intelligent hollow feature recognition and self-learning module internally includes a few-shot meta-learning submodule, a multimodal attention fusion submodule, and an online self-learning submodule. Specifically, the few-shot meta-learning submodule, during the meta-training phase, constructs a general hollow feature recognition pre-model across wall materials based on the MAML framework. It first builds a task set including common wall materials such as concrete, ceramic tile, and stone. Each task For a single wall material, only 10 labeled samples are included: 5 sets of hollow walls and 5 sets of normal walls; an initial model parameter is trained using MAML. This allows the parameter to converge quickly on any new task with only 1-2 gradient updates; during meta-training, for each task... First, calculate the loss using the sample from that task. And based on this, the parameters after a single update are obtained. ; then in the mission The updated loss is calculated on the test samples, and finally the initial parameters are optimized by backpropagation through the sum of the test losses of all tasks. .

[0010] Furthermore, the data management and visualization module internally includes a raw data storage unit, a test result processing module, a device status monitoring component, and a user interface. The raw data storage unit comprises a solid-state storage array, a data index manager, and a backup controller. The solid-state storage array uses RAID5 mode for redundant data storage. The data index manager establishes timestamps and device ID indexes for the stored raw data. The backup controller periodically synchronizes data to external storage devices. The test result processing module includes a result classifier, a labeling tool, and a data statistics component. The result classifier categorizes test results according to the location and severity of voids. The labeling tool supports manual annotation of the test results. The results are then labeled and corrected. The data statistics component counts the number and distribution ratio of hollow areas in different regions. The equipment status monitoring component includes a data stream collector, a status indicator, and an anomaly alarm. The data stream collector receives status data from each module in real time. The status indicator displays the operating, standby, and abnormal status of the equipment in the form of indicator lights. The anomaly alarm triggers an audible and visual alarm when the equipment malfunctions. The user interface includes a path planning panel, a test result display area, and equipment control buttons. The path planning panel supports manual drawing and importing of test paths. The test result display area shows the distribution of hollow locations in the form of a heat map. The equipment control buttons include start, stop, and pause operation commands.

[0011] Compared with the prior art, the beneficial effects of the present invention are: 1. The automated tapping detection and recognition system for hollow walls in this invention, through a small-sample meta-learning submodule, enables the system to quickly adapt to different wall materials without relying on a large number of labeled samples. Through parameter initialization and rapid fine-tuning of the meta-learning algorithm, it can extract core features from a small number of new material samples and directly transfer them to new scenarios. When deploying the system to different building projects, there is no need to re-collect massive amounts of data to train the model, saving preparation time and manpower costs. Simultaneously, the self-learning mechanism continuously optimizes the model; as detection data accumulates, the recognition accuracy for niche wall materials gradually improves, avoiding the performance drop issues that traditional models experience in cross-scenario applications.

[0012] 2. The automated tapping detection and identification system for hollow walls in this invention, through a multimodal attention fusion submodule, weightedly fuses acoustic and vibration signal features, accurately capturing the unique signal characteristics of hollow areas and filtering out background noise and interference signals in complex environments. This allows the system to automatically focus on key signal dimensions related to hollowness, reducing false positives and false negatives. Furthermore, the fused feature dimensions are more comprehensive, covering the signal characteristics of different types of hollowness. Whether it's shallow or deep hollowness, accurate identification can be achieved through the complementarity of multimodal features, improving the system's reliability and stability in practical engineering and ensuring that the detection results better meet the stringent standards of engineering acceptance. Attached Figure Description

[0013] Figure 1 This is a frame connection diagram of the automated tapping detection and identification system for hollow walls in buildings according to the present invention; Figure 2 This is a schematic diagram showing the sub-module connection of the intelligent recognition and self-learning module for hollow drum features of the present 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 This invention provides an automated tapping detection and identification system for hollow walls in building structures, comprising a mobile platform module, a standardized tapping execution module, a multimodal signal acquisition module, a hollow feature intelligent recognition and self-learning module, and a data management and visualization module. The motion control output of the mobile platform module is connected to the position calibration input of the standardized tapping execution module, and the status feedback output of the mobile platform module is connected to the device status acquisition input of the data management and visualization module. The action trigger output of the standardized tapping execution module is connected to the synchronous acquisition input of the multimodal signal acquisition module, and the force parameter output of the standardized tapping execution module is connected to the prior parameter input of the hollow feature intelligent recognition and self-learning module. The feature encoding output of the multimodal signal acquisition module is connected to the multimodal feature input of the hollow feature intelligent recognition and self-learning module. The original data output of the block is connected to the original data storage input of the data management and visualization module; the detection result output of the hollow drum feature intelligent recognition and self-learning module is connected to the result input of the data management and visualization module; the model update parameter output of the hollow drum feature intelligent recognition and self-learning module is self-connected to form a self-learning closed loop; the material adaptation instruction output of the hollow drum feature intelligent recognition and self-learning module is back-connected to the force adjustment input of the standardized tapping execution module; the path planning instruction output of the data management and visualization module is connected to the path execution input of the mobile platform module; the model optimization trigger output of the data management and visualization module is connected to the online learning input of the hollow drum feature intelligent recognition and self-learning module; and the user interaction instruction output of the data management and visualization module is connected to the emergency control input of the mobile platform module and the start / stop control input of the standardized tapping execution module, respectively.

[0016] In the above embodiments, the mobile platform module internally includes a motion control unit, a position calibration component, a status feedback module, and a path execution engine. The motion control unit comprises a DC servo driver, a stepper motor control circuit, and motion trajectory generation logic. The DC servo driver receives pulse signals and outputs speed-regulating current. The stepper motor control circuit is used for forward and reverse rotation and speed adjustment of the motor. The motion trajectory generation logic generates a continuous motion path based on preset coordinate points. The position calibration component is equipped with a laser displacement sensor, a mechanical limit switch, and a coordinate calibration algorithm. The laser displacement sensor collects the relative distance between the platform and the wall in real time. The mechanical limit switch sets the boundary range of the platform's movement. The coordinate calibration algorithm corrects the platform's current coordinates using sensor data. The status feedback module consists of a speed encoder, an attitude sensor, and a data transmission interface. The speed encoder collects motor speed data in real time. The attitude sensor detects the platform's tilt angle and vibration amplitude. The data transmission interface converts the status data into a standard communication protocol format. The path execution engine includes a path parser, a motion planner, and an emergency stop controller. The path parser breaks down the input path instructions into individual coordinate points. The motion planner calculates the motion speed and acceleration between adjacent coordinate points. The emergency stop controller receives external commands to trigger a motor power-off mechanism.

[0017] In the above embodiments, the standardized striking execution module internally includes a striking drive assembly, a force adjustment unit, a position alignment module, and an action trigger controller. The striking drive assembly comprises an electromagnetic striking device, a striking head replacement mechanism, and a power supply circuit. The electromagnetic striking device drives the striking head to reciprocate via electromagnetic force. The striking head replacement mechanism uses a snap-fit ​​structure to achieve quick disassembly and installation of the striking head. The power supply circuit provides stable DC voltage and current. The force adjustment unit is equipped with a pressure sensor, an electromagnetic force adjustment circuit, and force calibration logic. The pressure sensor collects data on the reaction force when the striking head contacts the wall. The electromagnetic force adjustment circuit adjusts the force by changing the electrical current... The magnetic coil current adjusts the striking force, and the force calibration logic corrects the force output value based on sensor data. The position alignment module includes a visual positioning sensor, a position adjustment motor, and an alignment algorithm. The visual sensor identifies marked points on the wall surface, the position adjustment motor drives the striking drive component to move horizontally and vertically, and the alignment algorithm calculates the position adjustment amount based on sensor data. The action trigger controller includes a synchronization signal receiver, a trigger logic circuit, and an action timer. The synchronization signal receiver receives external trigger commands, the trigger logic circuit controls the start, stop, and movement cycle of the striker, and the action timer records the duration and interval of each strike.

[0018] In the above embodiments, the multimodal signal acquisition module internally includes an acoustic signal acquisition unit, a vibration signal acquisition component, a synchronization control module, and a feature encoding engine. The acoustic signal acquisition unit comprises a MEMS microphone array, a preamplifier, and a filtering circuit. The MEMS microphone array consists of four microphones arranged in a rectangular layout. The preamplifier amplifies the weak signals acquired by the microphones, and the filtering circuit filters out low-frequency and high-frequency noise from the environment. The vibration signal acquisition component is equipped with a piezoelectric accelerometer, a signal conditioning circuit, and an AD converter. The piezoelectric accelerometer is fixed to the impact drive component to acquire vibration data, and the signal conditioning circuit amplifies the vibration signals. The analog-to-digital converter (ADC) converts analog signals into digital signals through filtering and processing. The synchronization control module includes a high-precision clock generator, a signal synchronizer, and a trigger signal interface. The high-precision clock generator provides a clock signal with microsecond-level accuracy, the signal synchronizer aligns the trigger signal with the acquired signal in time, and the trigger signal interface receives external synchronization commands to control the acquisition start time. The feature encoding engine includes a feature extractor, a data compressor, and an encoding format converter. The feature extractor extracts characteristic parameters such as formants and decay times from the original signal, the data compressor uses a lossless compression algorithm to reduce data storage space, and the encoding format converter converts the feature data into a standard JSON format.

[0019] In the above embodiments, such as Figure 2 As shown, the intelligent recognition and self-learning module for hollow features in this embodiment includes a few-shot meta-learning submodule, a multimodal attention fusion submodule, and an online self-learning submodule. The few-shot meta-learning submodule, during the meta-training phase, constructs a general hollow feature recognition pre-model across wall materials based on the MAML framework. The specific method is as follows: S1: Construct a task set that includes common wall materials such as concrete, tile, and stone. Each task It corresponds to one type of wall material and includes only 10 sets of labeled samples (5 sets of hollow walls and 5 sets of normal walls). S2: The core of MAML is training an initial model parameter set. This allows the parameter to converge quickly on any new task with only 1-2 gradient updates; S3: During meta-training, for each task First, calculate the loss using the sample from that task. And based on this, the parameters after a single update are obtained. ; then in the mission The updated loss is calculated on the test samples, and finally the initial parameters are optimized by backpropagation through the sum of the test losses of all tasks. ; The formula for the parameters of a single gradient update within the task is as follows: In the formula: This represents the initial model parameters during the meta-training phase, corresponding to the weights of the convolutional and fully connected layers in the hollow drum recognition model. This represents the model parameters after a single gradient update for the i-th task; This represents the in-task learning rate during the meta-training phase, ranging from 0.01 to 0.1, and is adaptively adjusted based on sample characteristics. Indicates based on initial parameters In the Task Cross-entropy loss on training samples; This represents the loss function with respect to the initial parameters. The gradient; The formula for calculating the total loss function of meta-training is: In the formula: This represents the total loss during the meta-training phase, used to optimize the initial parameters. ; This represents the total number of tasks in the meta-training task set, with a value ranging from 8 to 12. Indicates the first The test sample set for each task includes 5 sets of labeled samples that did not participate in the task update; For updated parameters In the mission Cross-entropy loss on test samples; After entering the rapid adaptation phase, the aforementioned general hollow wall recognition pre-model only needs to collect 10-20 sets of labeled samples (50% hollow and 50% normal samples) when the system encounters unfamiliar wall materials (such as aerated concrete blocks and hollow bricks), based on the initial parameters obtained from meta-training. Performing one gradient update is sufficient to obtain the parameters of the hollow recognition model adapted to this material. The specific formula is as follows: In the formula: The final model parameters after adapting to the new wall material; This represents the learning rate during the rapid adaptation phase, ranging from 0.05 to 0.2, and is typically greater than the original training learning rate. ; A set of labeled samples representing the new wall material, containing 10-20 sets of samples of hollow and normal walls; Indicates initial parameters New mission Loss gradient on the sample.

[0020] This "meta-training-rapid adaptation" mechanism completely solves the pain point of traditional machine learning models requiring a large number of labeled samples in new material scenarios, while ensuring consistency and accuracy in cross-material recognition. It is especially suitable for engineering scenarios with diverse building wall materials and high sample collection costs.

[0021] like Figure 2 As shown, in the above-mentioned intelligent recognition and self-learning module for hollow drum features, its internal multimodal attention fusion submodule constructs a cross-modal feature interaction framework based on the Transformer multi-head attention mechanism. The specific method is as follows: S1: First, the formant spectrum features extracted from the acoustic signal through short-time Fourier transform and the decay time constant features obtained from the vibration signal through wavelet decomposition are encoded into sequence features of the same dimension. and ; S2: Introducing an intermodal attention interaction layer, which automatically identifies complementary information between two modes by calculating the correlation scores of frame-level features: When a 2-5kHz wideband resonance peak unique to hollow drum appears in the acoustic signal, the attention score of the corresponding spectral frame will be higher than that of the normal wall frequency band; at the same time, abnormal segments in the vibration signal whose decay time constant exceeds twice the normal range will be accurately located by the model and given higher interaction weights, thus achieving spatiotemporal alignment of acoustic and vibration features; S3: In the dynamic weight allocation phase, global pooling and softmax normalization are performed on the attention score to generate a modality-adaptive weight vector. and Unlike traditional fixed-weight fusion methods, this module can dynamically adjust the weight ratio based on the signal characteristics of each tap: for example, when the acoustic resonance characteristics of hollow tiles are more pronounced... The weighting percentage can be increased to over 65%; and when the vibration attenuation anomalies of concrete walls are more easily distinguished, It will automatically take the dominant position. Finally, fused features are generated through weighted fusion and residual connections. This provides more robust feature support for hollow drum recognition, improving accuracy by 15%-20% compared to single-modality recognition.

[0022] The frame-level modal attention score calculation formula in this embodiment is as follows: In the formula: The attention score for acoustic features in frame m and vibration features in frame n; Let be the encoded feature vector of the m-th frame of the acoustic signal; This is the encoded feature vector of the nth frame of the vibration signal; To query the feature mapping weight matrix; d is the weight matrix for the key feature mapping; d is the dimension of the feature vector. For row dimension normalization function; The formula for calculating modal global adaptive weights is: In the formula: For the global adaptive weights of the acoustic modes; For the global adaptive weights of the vibration modes; The total number of frames for acoustic features; The total number of frames representing vibration characteristics; This is an operation to maximize the frame-level attention score; The formula for multimodal feature fusion calculation is as follows: ; In the formula: This is the final feature vector after fusion; This is an element-wise multiplication operation; This is a feature vector concatenation operation; This is the mapping weight matrix for the fused features.

[0023] like Figure 2As shown, in the aforementioned intelligent recognition and self-learning module for hollow drum features, its internal online self-learning submodule consists of a data buffer unit, a feature filter, an incremental trainer, a model snapshot manager, and a parameter synchronizer connected in series. The data buffer unit comprises a fixed-length sliding window and real-time data cleaning logic. The sliding window stores the latest acquired multimodal feature data in timestamp-ordered order. The data cleaning logic includes rules for duplicate data removal and outlier filtering, with rules set based on the statistical distribution threshold of the feature dimensions. The feature filter consists of a rule engine and a feature matcher. The rule engine has built-in modal feature association rules, and the feature matcher is responsible for dimensional alignment of newly acquired features with samples in the historical feature library. The alignment process is based on hash mapping of feature encoding. The rule engine filters feature subsets that meet the incremental training conditions based on the matching results. The incremental trainer includes a mini-batch sampler, a gradient calculation unit, and a gradient... The trimmer and mini-batch sampler extract training samples from the filtered feature subset according to the principle of class balance. The gradient calculation unit uses the stochastic gradient descent algorithm to calculate the gradient of the model parameters. The gradient trimmer limits the gradient update magnitude by setting a gradient norm threshold. The model snapshot manager consists of a version control module and a snapshot storage unit. The version control module generates a unique version identifier for each training batch. The snapshot storage unit stores the model parameters and training metadata of the corresponding version in binary format. The metadata includes the training batch, sample quantity, and training timestamp. The version control module supports snapshot backtracking and comparison based on the version identifier. The parameter synchronizer consists of a real-time synchronization protocol and a conflict resolution mechanism. The real-time synchronization protocol uses a publish-subscribe model to synchronize the training parameters with the recognition model. The conflict resolution mechanism determines the priority based on the parameter update timestamp and only retains the parameter update instruction with the latest timestamp. The synchronization process uses an incremental transmission method.

[0024] In the above embodiments, the data management and visualization module of the present invention internally includes a raw data storage unit, a detection result processing module, a device status monitoring component, and a user interface. The raw data storage unit comprises a solid-state storage array, a data index manager, and a backup controller. The solid-state storage array uses RAID5 mode to achieve redundant data storage. The data index manager establishes timestamps and device ID indexes for the stored raw data. The backup controller periodically synchronizes the data to external storage devices. The detection result processing module includes a result classifier, a labeling tool, and a data statistics component. The result classifier categorizes the detection results according to the location and severity of voids. The labeling tool supports manual annotation. The test results are then annotated and corrected. The data statistics component calculates the number and distribution ratio of hollow areas in different regions. The equipment status monitoring component includes a data stream collector, status indicators, and anomaly alarms. The data stream collector receives status data from each module in real time. The status indicators display the equipment's running, standby, and abnormal states in the form of indicator lights. The anomaly alarms trigger audible and visual alarms when the equipment malfunctions. The user interface includes a path planning panel, a test result display area, and equipment control buttons. The path planning panel supports manual drawing and importing of test paths. The test result display area shows the distribution of hollow locations in the form of a heat map. The equipment control buttons include equipment operation commands such as start, stop, and pause.

[0025] This invention provides an automated tapping detection and recognition system for hollow walls. When applied to a building wall hollow detection model, its internal small-sample meta-learning submodule enables rapid adaptation to different wall materials without relying on a large number of labeled samples. Through parameter initialization and rapid fine-tuning of the meta-learning algorithm, it can extract core features from a small number of new material samples and directly transfer them to new scenarios. When deploying the system to different building projects, it eliminates the need to re-collect massive amounts of data to train the model, saving preparation time and manpower costs. Simultaneously, the self-learning mechanism continuously optimizes the model; as detection data accumulates, the recognition accuracy for niche wall materials gradually improves, avoiding the performance drop issues that traditional models experience in cross-scenario applications. Furthermore, its internal multimodal attention fusion submodule accurately captures the unique signal features of hollow areas through weighted fusion of acoustic and vibration signal features, filtering out background noise and interference signals in complex environments. This allows the system to automatically focus on key signal dimensions related to hollowness, reducing false positives and false negatives. At the same time, the fused feature dimensions are more comprehensive, covering the signal manifestations of different types of hollowness. Whether it is shallow or deep hollowness, it can achieve accurate identification through the complementarity of multimodal features, improving the reliability and stability of the system in actual engineering and making the test results more in line with the strict standards of engineering acceptance.

[0026] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automated tapping detection and identification system for hollow areas in building walls, characterized in that, It includes a mobile platform module, a standardized tapping execution module, a multimodal signal acquisition module, an intelligent recognition and self-learning module for hollow drum features, and a data management and visualization module; The motion control output of the mobile platform module is connected to the position calibration input of the standardized tapping execution module, and the status feedback output of the mobile platform module is connected to the device status acquisition input of the data management and visualization module. The action trigger output of the standardized striking execution module is connected to the synchronous acquisition input of the multimodal signal acquisition module, and the force parameter output of the standardized striking execution module is connected to the prior parameter input of the hollow drum feature intelligent recognition and self-learning module. The feature encoding output terminal of the multimodal signal acquisition module is connected to the multimodal feature input terminal of the hollow drum feature intelligent recognition and self-learning module, and the raw data output terminal of the multimodal signal acquisition module is connected to the raw data storage input terminal of the data management and visualization module. The detection result output terminal of the hollow drum feature intelligent recognition and self-learning module is connected to the result input terminal of the data management and visualization module. The model update parameter output terminal of the hollow drum feature intelligent recognition and self-learning module is self-connected to form a self-learning closed loop. The material adaptation instruction output terminal of the hollow drum feature intelligent recognition and self-learning module is back-connected to the force adjustment input terminal of the standardized tapping execution module. The path planning instruction output of the data management and visualization module is connected to the path execution input of the mobile platform module. The model optimization trigger output of the data management and visualization module is connected to the online learning input of the hollow drum feature intelligent recognition and self-learning module. The user interaction instruction output of the data management and visualization module is connected to the emergency control input of the mobile platform module and the start / stop control input of the standardized tapping execution module, respectively.

2. The automated tapping detection and identification system for hollow building walls as described in claim 1, characterized in that: The mobile platform module is internally equipped with a motion control unit, a position calibration component, a status feedback module, and a path execution engine; The motion control unit includes a DC servo driver, a stepper motor control circuit, and motion trajectory generation logic. The DC servo driver receives pulse signals and outputs speed regulation current. The stepper motor control circuit is used for forward and reverse rotation and speed adjustment of the motor. The motion trajectory generation logic generates a continuous motion path based on preset coordinate points. The position calibration component is equipped with a laser displacement sensor, a mechanical limit switch, and a coordinate calibration algorithm. The laser displacement sensor collects the relative distance between the platform and the wall in real time, the mechanical limit switch sets the boundary range of the platform's movement, and the coordinate calibration algorithm corrects the platform's current coordinates based on the sensor data. The status feedback module consists of a speed encoder, an attitude sensor, and a data transmission interface. The speed encoder collects motor speed data in real time, the attitude sensor detects the tilt angle and vibration amplitude of the platform, and the data transmission interface converts the status data into a standard communication protocol format. The path execution engine includes a path parser, a motion planner, and an emergency stop controller. The path parser breaks down the input path instructions into individual coordinate points, the motion planner calculates the motion speed and acceleration between adjacent coordinate points, and the emergency stop controller receives external instructions to trigger the motor power-off mechanism.

3. The automated tapping detection and identification system for hollow building walls as described in claim 1, characterized in that: The standardized tapping execution module is internally equipped with a tapping drive component, a force adjustment unit, a position alignment module, and an action trigger controller. The striking drive assembly includes an electromagnetic striking device, a striking head replacement mechanism, and a power supply circuit. The electromagnetic striking device drives the striking head to reciprocate through electromagnetic force. The striking head replacement mechanism adopts a snap-fit ​​structure to achieve quick disassembly and installation of the striking head. The power supply circuit provides a stable DC voltage and current. The force adjustment unit is equipped with a pressure sensor, an electromagnetic force adjustment circuit and a force calibration logic. The pressure sensor collects the reaction force data when the striking head contacts the wall. The electromagnetic force adjustment circuit adjusts the striking force by changing the magnitude of the electromagnetic coil current. The force calibration logic corrects the force output value based on the sensor data. The position alignment module includes a visual positioning sensor, a position adjustment motor, and an alignment algorithm. The visual sensor identifies the marked points on the wall surface, the position adjustment motor drives the tapping drive component to move horizontally and vertically, and the alignment algorithm calculates the position adjustment amount based on the sensor data. The action trigger controller includes a synchronization signal receiver, a trigger logic circuit, and an action timer. The synchronization signal receiver receives external trigger commands, the trigger logic circuit controls the start and stop of the tapper and its movement cycle, and the action timer records the duration and interval of a single tap.

4. The automated tapping detection and identification system for hollow building walls as described in claim 1, characterized in that: The multimodal signal acquisition module is internally equipped with an acoustic signal acquisition unit, a vibration signal acquisition component, a synchronization control module, and a feature encoding engine; The acoustic signal acquisition unit includes a MEMS microphone array, a preamplifier, and a filtering circuit. The MEMS microphone array consists of four microphones arranged in a rectangular layout. The preamplifier amplifies the weak signals acquired by the microphones, and the filtering circuit filters out low-frequency and high-frequency noise in the environment. The vibration signal acquisition component is equipped with a piezoelectric accelerometer, a signal conditioning circuit, and an AD converter. The piezoelectric accelerometer is fixed on the impact drive component to collect vibration data. The signal conditioning circuit amplifies and filters the vibration signal, and the AD converter converts the analog signal into a digital signal. The synchronization control module includes a high-precision clock generator, a signal synchronizer, and a trigger signal interface. The high-precision clock generator provides a clock signal with microsecond-level precision. The signal synchronizer aligns the tap trigger signal with the acquisition signal in time. The trigger signal interface receives external synchronization commands to control the acquisition start time. The feature encoding engine includes a feature extractor, a data compressor, and an encoding format converter. The feature extractor extracts feature parameters such as formants and decay times from the original signal. The data compressor uses a lossless compression algorithm to reduce data storage space. The encoding format converter converts the feature data into a standard JSON format.

5. The automated tapping detection and identification system for hollow building walls as described in claim 1, characterized in that: The intelligent recognition and self-learning module for hollow drum features is internally configured with a small sample meta-learning sub-module, a multimodal attention fusion sub-module, and an online self-learning sub-module; In the meta-training phase, the small-sample meta-learning submodule constructs a general hollow-dampness recognition pre-model based on the MAML framework, as follows: S1: A task set for constructing common wall materials including concrete, tile, and stone. Each task It corresponds to one type of wall material and includes only 10 sets of labeled samples: 5 sets of hollow walls and 5 sets of normal walls; S2: Train an initial model parameter using MAML This allows the parameter to converge quickly on any new task with only 1-2 gradient updates; S3: During meta-training, for each task First, calculate the loss using the sample from that task. And based on this, the parameters after a single update are obtained. ; then in the mission The updated loss is calculated on the test samples, and finally the initial parameters are optimized by backpropagation through the sum of the test losses of all tasks. ; The formula for the parameters of a single gradient update within the task is as follows: In the formula: This represents the initial model parameters during the meta-training phase, corresponding to the weights of the convolutional and fully connected layers in the hollow drum recognition model. This represents the model parameters after a single gradient update for the i-th task; This represents the in-task learning rate during the meta-training phase, ranging from 0.01 to 0.1, and is adaptively adjusted based on sample characteristics. Indicates based on initial parameters In the Task Cross-entropy loss on training samples; This represents the loss function with respect to the initial parameters. The gradient; The formula for calculating the total loss function of meta-training is: In the formula: This represents the total loss during the meta-training phase, used to optimize the initial parameters. ; This represents the total number of tasks in the meta-training task set, with a value ranging from 8 to 12. Indicates the first The test sample set for each task includes 5 sets of labeled samples that did not participate in the task update; For updated parameters In the mission Cross-entropy loss on test samples; After the aforementioned general hollow wall recognition pre-model enters the rapid adaptation phase, when the system encounters an unfamiliar wall material, it needs to collect 10-20 sets of labeled samples, including 50% hollow samples and 50% normal samples, based on the initial parameters obtained from meta-training. Performing one gradient update is sufficient to obtain the parameters of the hollow recognition model adapted to this material. The specific formula is as follows: In the formula: The final model parameters after adapting to the new wall material; This represents the learning rate during the rapid adaptation phase, ranging from 0.05 to 0.2, and is typically greater than the original training learning rate. ; A set of labeled samples representing the new wall material, containing 10-20 sets of samples of hollow and normal walls; Indicates initial parameters New mission Loss gradient on the sample.

6. The automated tapping detection and identification system for hollow building walls as described in claim 5, characterized in that: The multimodal attention fusion submodule constructs a cross-modal feature interaction framework based on the Transformer multi-head attention mechanism, and the specific method is as follows: S1: First, the formant spectrum features extracted from the acoustic signal through short-time Fourier transform and the decay time constant features obtained from the vibration signal through wavelet decomposition are encoded into sequence features of the same dimension. and ; S2: Introducing an intermodal attention interaction layer, which automatically identifies complementary information between two modes by calculating the correlation scores of frame-level features: When a 2-5kHz wideband resonance peak unique to hollow drum appears in the acoustic signal, the attention score of the corresponding spectral frame will be higher than that of the normal wall frequency band; at the same time, abnormal segments in the vibration signal whose decay time constant exceeds twice the normal range will be accurately located by the model and given higher interaction weights, thus achieving spatiotemporal alignment of acoustic and vibration features; S3: In the dynamic weight allocation phase, global pooling and softmax normalization are performed on the attention score to generate a modality-adaptive weight vector. and Then, fusion features are generated through weighted fusion and residual connection. This provides more robust feature support for hollow drum identification; The formula for calculating the frame-level modal attention score is as follows: In the formula: The attention score for acoustic features in frame m and vibration features in frame n; Let be the encoded feature vector of the m-th frame of the acoustic signal; This is the encoded feature vector of the nth frame of the vibration signal; To query the feature mapping weight matrix; d is the weight matrix for the key feature mapping; d is the dimension of the feature vector. For row dimension normalization function; The formula for calculating modal global adaptive weights is: In the formula: For the global adaptive weights of the acoustic modes; For the global adaptive weights of the vibration modes; The total number of frames for acoustic features; The total number of frames representing vibration characteristics; This is an operation to maximize the frame-level attention score; The formula for multimodal feature fusion calculation is as follows: ； In the formula: This is the final feature vector after fusion; This is an element-wise multiplication operation; This is a feature vector concatenation operation; This is the mapping weight matrix for the fused features.

7. The automated tapping detection and identification system for hollow building walls as described in claim 5, characterized in that: The online self-learning submodule consists of a data buffer unit, a feature filter, an incremental trainer, a model snapshot manager, and a parameter synchronizer connected in series. The data buffer unit consists of a fixed-length sliding window and real-time data cleaning logic. The sliding window stores the latest collected multimodal feature data in order of timestamp. The data cleaning logic includes rules for removing duplicate data and filtering outliers. The rules are set based on the statistical distribution threshold of the feature dimension. The feature filter consists of a rule engine and a feature matcher. The rule engine has built-in modal feature association rules, and the feature matcher is responsible for dimensional alignment of newly collected features with samples in the historical feature library. The alignment process is based on the hash mapping of feature encoding. The rule engine filters feature subsets that meet the incremental training conditions based on the matching results. The incremental trainer includes a mini-batch sampler, a gradient calculation unit, and a gradient clipper. The mini-batch sampler extracts training samples from the filtered feature subset according to the class balance principle. The gradient calculation unit uses the stochastic gradient descent algorithm to calculate the gradient of the model parameters. The gradient clipper limits the gradient update magnitude by setting a gradient norm threshold. The model snapshot manager consists of a version control module and a snapshot storage unit. The version control module generates a unique version identifier according to the training batch. The snapshot storage unit stores the model parameters and training metadata of the corresponding version in binary format. The metadata includes the training batch, the number of samples, and the training timestamp. The version control module supports snapshot backtracking and comparison based on the version identifier. The parameter synchronizer consists of a real-time synchronization protocol and a conflict resolution mechanism. The real-time synchronization protocol uses a publish-subscribe model to synchronize the training parameters with the recognition model. The conflict resolution mechanism determines the priority based on the parameter update timestamp and only retains the parameter update instruction with the latest timestamp. The synchronization process uses an incremental transmission method.

8. The automated tapping detection and identification system for hollow building walls as described in claim 1, characterized in that: The data management and visualization module is internally equipped with a raw data storage unit, a detection result processing module, a device status monitoring component, and a user interface. The original data storage unit includes a solid-state storage array, a data index manager, and a backup controller. The solid-state storage array uses RAID5 mode to achieve redundant data storage. The data index manager establishes timestamps and device ID indexes for the stored original data. The backup controller periodically synchronizes the data to external storage devices. The detection result processing module includes a result classifier, a labeling tool, and a data statistics component. The result classifier classifies the detection results according to the location and severity of the voids. The labeling tool supports manual secondary labeling and correction of the detection results. The data statistics component counts the number and distribution ratio of voids in different regions. The device status monitoring component includes a data stream collector, a status indicator, and an anomaly alarm. The data stream collector receives status data from each module in real time. The status indicator displays the device's running, standby, and abnormal status in the form of indicator lights. The anomaly alarm triggers an audible and visual alarm when the device malfunctions. The user interface includes a path planning panel, a test result display area, and device control buttons. The path planning panel supports manual drawing and importing of test paths. The test result display area shows the distribution of void locations in the form of a heat map. The device control buttons include start, stop, and pause operation commands.

Images