Concrete strength nondestructive testing system and method based on adaptive excitation and machine learning

Abstract

The invention belongs to the technical field of machine learning and concrete detection, and provides a concrete strength nondestructive detection system and detection method based on adaptive excitation and machine learning, and the main scheme is as follows: an active excitation unit executes pre-scanning operation before formal detection, and emits preset pulse waves to concrete according to preset excitation parameters; capturing a first feedback sound wave signal through a sound wave sensing unit; the data processing unit extracts frequency spectrum information and adaptively determines an optimal excitation parameter; the active excitation unit emits stress waves to the concrete according to the optimal excitation parameters for formal detection, and a second feedback sound wave signal is captured through the sound wave sensing unit; and the data processing unit extracts the characteristic parameters and inputs the characteristic parameters to the trained learning model to obtain a preliminary strength presumption value, and judges whether the confidence coefficient of the detection meets the requirement, if so, the confidence coefficient is taken as a final strength presumption value, and otherwise, the optimal excitation parameter is adjusted and the re-detection is prompted.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and concrete testing technology, and in particular to a non-destructive testing system and method for concrete strength based on adaptive excitation and machine learning. Background Technology

[0002] Existing non-destructive testing technologies for concrete strength (such as the rebound method) heavily rely on mechanical impact rebound signals. Furthermore, the traditional rebound method, based on the principle of kinetic energy conservation, has inherent limitations in its physical model for ultra-high performance concrete with high hardness and high elastic modulus. When using the rebound method, most of the impact energy is dissipated, resulting in a weak effective rebound signal. In addition, the sensor's dynamic range is insufficient, leading to inaccurate measurements, especially with ultra-high strength concrete of C100 and above, where low energy transfer efficiency and poor signal-to-noise ratio cause detection failures. Moreover, instruments based on a single testing principle cannot cover a wide range of materials from ordinary to ultra-high strength, requiring users to configure multiple devices, resulting in high costs and cumbersome operation. Furthermore, the impact energy and sensor response required for different concrete strengths vary significantly, making it impossible for a single hardware structure to meet the full range of testing needs. Summary of the Invention

[0003] The purpose of this invention is to provide a non-destructive testing system and method for concrete strength based on adaptive excitation and machine learning. It can combine acoustic sensors and machine learning models to adaptively determine the optimal excitation parameters for different concrete types under different testing scenarios. It can also determine whether the estimated concrete strength value meets the requirements by the confidence level of the test. This avoids the problems of high hardware investment costs and large measurement errors caused by the use of multiple devices in traditional rebound method measurement.

[0004] The technical solution adopted by this invention to solve its technical problem is as follows: On one hand, the present invention provides a non-destructive testing system for concrete strength based on adaptive excitation and machine learning, comprising: The active excitation unit is used to receive preset excitation parameters, perform a pre-scan operation after power-on and before formal testing, and emit a preset pulse wave to the concrete with the preset excitation parameters. When the optimal excitation parameters are received, it emits a stress wave to the concrete with the optimal excitation parameters for formal testing. The acoustic wave sensing unit is used to capture the first feedback acoustic wave signal of the concrete surface in response to a preset pulse wave, and to capture the second feedback acoustic wave signal of the concrete surface in response to a stress wave, and send them to the data processing unit. The data processing unit stores preset excitation parameters, which are used to send the preset excitation parameters to the active excitation unit. When the first feedback acoustic signal is received, its spectrum information is extracted, and the optimal excitation parameters for the current concrete type under the current detection scenario are adaptively determined based on the spectrum information and sent to the active excitation unit. It also stores a pre-built and trained machine learning model based on acoustic features and intensity mapping. When the second feedback acoustic wave signal is received, its feature parameters are extracted and input into the pre-trained learning model. The pre-trained machine learning model is used to obtain a preliminary intensity estimate and to determine whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate. Otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The pre-trained machine learning model is then updated and optimized.

[0005] In some embodiments, the data processing unit also stores information on concrete specimens of different texture types and information on different testing scenario types; The preset excitation parameters corresponding to the preset pulse wave have a frequency range of 20kHz-150kHz. When the data processing unit receives the first feedback acoustic signal, it extracts its spectral information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectral information. This means that the data processing unit identifies the frequency band with the highest signal-to-noise ratio and the most significant response based on the spectral information extracted from the first feedback acoustic signal, and extracts the center frequency and bandwidth parameters of the frequency band with the most significant response as the optimal excitation parameters.

[0006] In some embodiments, the energy levels of the preset pulse wave are adjustable, and its excitation voltage amplitude is controlled by the driving circuit to be between 100V and 1000V, and its pulse width is between 1μs and 50μs. The waveform of the preset pulse wave is a square wave pulse and / or a narrow pulse and / or a frequency-modulated pulse and / or a window function modulated pulse.

[0007] In some embodiments, when the data processing unit receives the first feedback acoustic signal, it extracts its spectral information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectral information, including the following steps: The active excitation unit performs the first excitation according to the preset excitation parameters, and the data processing unit receives the first feedback acoustic signal; The data processing unit extracts the spectral information from the first feedback acoustic signal, calculates the key quality indicators of the first feedback acoustic signal based on the spectral information, and determines whether the key quality indicators have all reached the preset threshold. If they have not reached the threshold, the excitation parameters are adjusted and the adjusted excitation parameters are sent to the active excitation unit. The active excitation unit performs cyclic excitation according to the adjusted excitation parameters, and the data processing unit continues to calculate the key quality indicators of the first feedback acoustic signal and determine whether all key quality indicators have reached the preset threshold. The process is repeated until all key quality indicators have reached the preset threshold, and the center frequency and bandwidth parameters of the frequency band with the most significant response in the spectrum information corresponding to all reaching the preset threshold are taken as the optimal excitation parameters.

[0008] In some embodiments, the active excitation unit, the acoustic wave sensing unit, and the data processing unit are all integrated into the same device; The active excitation unit is an active excitation probe with a built-in piezoelectric ceramic stack; The acoustic wave sensing unit is a miniature microphone array or an optical sensor, located on the same side of the housing of the active excitation probe.

[0009] In some embodiments, a reference microphone is further provided on the housing of the device, and a filter is provided inside the housing; The reference microphone is used to collect ambient background noise; The filter is a minimum mean square adaptive filter, used to receive ambient background noise and a first feedback acoustic signal or a second feedback acoustic signal, and to use the ambient background noise as a reference input to dynamically remove the noise components associated with the first feedback acoustic signal or the second feedback acoustic signal, and then transmit the dynamically removed noise components to the data processing unit.

[0010] In some embodiments, the device housing is further provided with a variety of digital bandpass filter banks; The various digital bandpass filter banks are used to dynamically remove noise components related to the environmental background noise from the first or second feedback acoustic signal, and filter out frequency components unrelated to the detection scene and concrete type.

[0011] In some embodiments, when the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, the extracted feature parameters include spectral centroid shift, energy attenuation rate, resonant frequency change, and wavelet entropy.

[0012] In some embodiments, the data processing unit obtains a preliminary strength estimate through a trained machine learning model and determines whether the confidence level of the current detection meets the requirements. If the requirements are not met, it means: The trained machine learning model pre-stores datasets covering different types of concrete in different detection scenarios; When the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, it extracts the spectral centroid shift, energy attenuation rate, resonant frequency change and wavelet entropy, and forms a feature vector. Calculate the Mahalanobis distance or k-nearest neighbor distance between the feature vector and the features of various samples in the dataset. If the distance exceeds the preset distance, it means that the confidence level of this detection does not meet the requirements.

[0013] On the other hand, the present invention also provides a non-destructive testing method for concrete strength based on adaptive excitation and machine learning, applied to the aforementioned non-destructive testing system for concrete strength based on adaptive excitation and machine learning, comprising the following steps: The data processing unit stores preset excitation parameters and a machine learning model based on acoustic features and intensity mapping that has been built and trained. The data processing unit sends the preset excitation parameters to the active excitation unit, and the active excitation unit receives the preset excitation parameters. After power-on and before formal testing, a pre-scan operation is performed, and a preset pulse wave is emitted to the concrete with the preset excitation parameters. The first feedback acoustic signal of the concrete surface in response to the preset pulse wave is captured by the acoustic wave sensing unit and sent to the data processing unit. When the data processing unit receives the first feedback acoustic signal, it extracts its spectrum information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectrum information, and sends them to the active excitation unit. When the active excitation unit receives the optimal excitation parameters, it emits stress waves to the concrete with the optimal excitation parameters for formal detection, and captures the second feedback acoustic wave signal of the concrete surface in response to the stress wave through the acoustic wave sensing unit, and sends it to the data processing unit. When the data processing unit receives the second feedback acoustic signal, it extracts its feature parameters and inputs them into the trained learning model. The trained machine learning model obtains a preliminary intensity estimate and determines whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate; otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The trained machine learning model is then updated and optimized.

[0014] The beneficial effects of this invention are as follows: First, this invention eliminates the need for contact testing using the traditional rebound method, instead employing non-contact acoustic sensing. This avoids sensor wear and potential damage to the test surface caused by contact measurements, thus improving the durability and applicability of the equipment. Second, this invention enables high-precision testing of concrete with a full strength range from C15 to C150+, particularly addressing the detection blind spot for ultra-high-strength concrete. Third, the machine learning model in this invention can adaptively control the optimal excitation parameters for different concrete types under different testing environments, resulting in more accurate test results. Furthermore, each test result can be used to provide real-time samples for the machine learning model, further enhancing its scalability and enabling faster concrete testing under new testing environments. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the composition structure of a non-destructive testing system for concrete strength based on adaptive excitation and machine learning, according to Embodiment 1 of the present invention. Figure 2 This is a flowchart of a non-destructive testing method for concrete strength based on adaptive excitation and machine learning, according to Embodiment 2 of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0017] Example 1

[0018] This embodiment provides a non-destructive testing system for concrete strength based on adaptive excitation and machine learning. A schematic diagram of its structure can be found in [reference needed]. Figure 1 The system mainly consists of the following parts: The active excitation unit is used to receive preset excitation parameters, perform a pre-scan operation after power-on and before formal testing, and emit a preset pulse wave to the concrete with the preset excitation parameters. When the optimal excitation parameters are received, it emits a stress wave to the concrete with the optimal excitation parameters for formal testing. The acoustic wave sensing unit is used to capture the first feedback acoustic wave signal of the concrete surface in response to a preset pulse wave, and to capture the second feedback acoustic wave signal of the concrete surface in response to a stress wave, and send them to the data processing unit. The data processing unit stores preset excitation parameters, which are used to send the preset excitation parameters to the active excitation unit. When the first feedback acoustic signal is received, its spectrum information is extracted, and the optimal excitation parameters for the current concrete type under the current detection scenario are adaptively determined based on the spectrum information and sent to the active excitation unit. It also stores a pre-built and trained machine learning model based on acoustic features and intensity mapping. When the second feedback acoustic wave signal is received, its feature parameters are extracted and input into the pre-trained learning model. The pre-trained machine learning model is used to obtain a preliminary intensity estimate and to determine whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate. Otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The pre-trained machine learning model is then updated and optimized.

[0019] It should be noted that, because the active excitation unit in this embodiment possesses adjustable parameters, scene awareness, and closed-loop optimization capabilities, the system can dynamically optimize the excitation signal through a "perception-decision-adjustment" closed loop, targeting concrete specimens of different textures (e.g., from porous C30 to dense C150) and different testing scenarios (e.g., smooth laboratory test blocks versus rough field structural surfaces). This maximizes the acquisition of effective information and minimizes signal loss and error. The adaptive control of the excitation signal can be described in detail from two aspects: adaptive selection of the frequency domain and precise control of energy and pulse waveform.

[0020] First, adaptive selection for the frequency domain: Since stress waves of different frequencies have different propagation characteristics in concrete, high-frequency waves have high resolution but are sensitive to attenuation, and are suitable for surface or near-surface detection of dense, high-strength concrete. Low-frequency waves have strong penetration but low resolution, and are suitable for deep detection of concrete with lower strength, loose internal structure or large volume.

[0021] Therefore, before the formal measurement, the system automatically performs a rapid frequency scan, i.e., a pre-scan mode. During this mode, the excitation unit emits a set of short pulses covering a preset wide bandwidth (e.g., 20kHz - 150kHz), and the acoustic sensing unit simultaneously receives the feedback. The signal processing unit then analyzes the spectrum of this scan feedback in real time, identifying the frequency band with the highest signal-to-noise ratio and the most significant response. For example, for high-strength, dense concrete, which has a higher resonant frequency, the system will select 80kHz-120kHz as the optimal excitation frequency band; while for porous, low-strength concrete, the optimal response may be 30kHz-60kHz.

[0022] Finally, the system will apply the optimized center frequency and bandwidth parameters to the excitation signal in the subsequent formal testing.

[0023] Secondly, precise control of energy and pulse waveform is crucial. Insufficient excitation energy results in a weak signal, especially for high-strength concrete, while excessive energy can cause nonlinear effects or signal saturation, and may even cause micro-damage to low-strength concrete. The pulse waveform determines the spectral content and temporal resolution of the excitation signal.

[0024] Therefore, to achieve graded adjustable energy (voltage / pulse width), the high-voltage drive circuit of the excitation unit needs to support precise digital control of voltage amplitude (e.g., 100V-1000V) and pulse width (e.g., 1μs-50μs). In this case, the system can have a built-in preset excitation energy that roughly corresponds to the concrete strength grade. Users can select "high-strength mode" (high voltage, narrow pulse width) or "medium-strength mode" (medium voltage, medium pulse width), or the system can automatically determine the appropriate mode based on the pre-scan results.

[0025] Furthermore, regarding the waveform library and intelligent selection, the system firmware stores a variety of optimized excitation waveforms, including not only simple square wave pulses, but also: narrow pulses (with extremely wide spectrums for high-resolution detection and analysis of fine structures); frequency-modulated pulses (with specific frequency-time relationships, beneficial for separating effective signals from noise); and window function-modulated pulses (such as sinusoidal pulses modulated by the Hanning window, which reduce spectral leakage). The system can intelligently select or combine these waveforms according to the detection target (emphasizing intensity accuracy or internal defect detection).

[0026] Therefore, based on the above-mentioned adaptive excitation control, in this embodiment, the data processing unit also stores information on concrete specimens of different texture types and information on different detection scenario types; the preset excitation parameters corresponding to the preset pulse wave have a frequency band range of 20kHz-150kHz; when the data processing unit receives the first feedback acoustic wave signal, it extracts its spectrum information and adaptively determines the optimal excitation parameters for the current concrete type under the current detection scenario based on the spectrum information. This means that the data processing unit identifies the frequency band with the highest signal-to-noise ratio and the most significant response based on the spectrum information extracted from the first feedback acoustic wave signal, and extracts the center frequency and bandwidth parameters of the frequency band with the most significant response as the optimal excitation parameters.

[0027] The energy levels of the preset pulse wave are adjustable, and its excitation voltage amplitude can be controlled by the driving circuit to be between 100V and 1000V, and its pulse width is between 1μs and 50μs; the waveform of the preset pulse wave is a square wave pulse and / or a narrow pulse and / or a frequency-modulated pulse and / or a window function modulated pulse.

[0028] It should be noted that, regarding excitation adaptive control, in addition to the two types mentioned above (adaptive selection of frequency domain and precise control of energy and pulse waveform), feedback-based closed-loop optimization of excitation can also be included. In this case, the system treats each excitation-feedback cycle as a learning opportunity and uses it to optimize the next excitation. Feedback-based closed-loop optimization of excitation can include the following steps: Initial stimulation: The system performs the first stimulation with default or preset parameters and collects feedback signal S1.

[0029] Signal quality assessment: The processing unit calculates key quality indicators of S1 in real time, such as peak amplitude, signal-to-noise ratio, and significance of the main frequency component.

[0030] Parameter iteration: If the quality indicators are not met (e.g., peak amplitude is below the threshold, or signal-to-noise ratio is insufficient), the system will automatically fine-tune the excitation parameters. If the signal is too weak, the excitation voltage will be increased in steps; if the signal is saturated, the voltage will be reduced or the pulse width will be decreased; if the spectrum shows a blurry response at key frequencies, the center frequency will be adjusted or the excitation waveform will be replaced.

[0031] Final measurement: After 1-3 rapid iterations, when the feedback signal quality reaches the optimal standard, the system locks this optimal excitation parameter and performs the final high-precision measurement under this parameter.

[0032] Therefore, in this embodiment, the data processing unit extracts the spectrum information of the first feedback acoustic signal upon receiving it, and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectrum information. This may include the following steps: The active excitation unit performs the first excitation according to the preset excitation parameters, and the data processing unit receives the first feedback acoustic signal; The data processing unit extracts the spectral information from the first feedback acoustic signal, calculates the key quality indicators of the first feedback acoustic signal based on the spectral information, and determines whether the key quality indicators have all reached the preset threshold. If they have not reached the threshold, the excitation parameters are adjusted and the adjusted excitation parameters are sent to the active excitation unit. The active excitation unit performs cyclic excitation according to the adjusted excitation parameters, and the data processing unit continues to calculate the key quality indicators of the first feedback acoustic signal and determine whether all key quality indicators have reached the preset threshold. The process is repeated until all key quality indicators have reached the preset threshold, and the center frequency and bandwidth parameters of the frequency band with the most significant response in the spectrum information corresponding to all reaching the preset threshold are taken as the optimal excitation parameters.

[0033] At this point, for different concrete conditions, this embodiment can employ different adaptive excitation strategies to achieve the detection objective. For example, for ultra-high strength concrete (C100+), in order to excite high-frequency resonance of the material, avoid energy dissipation, and obtain a high signal-to-noise ratio signal, a high-frequency band (e.g., 100kHz) can be pre-scanned, using high voltage and narrow pulses. For porous / low strength concrete (C30), in order to enhance wave penetration, avoid excessive signal attenuation, and obtain deep information, a low-frequency band (e.g., 35kHz) can be pre-scanned, using medium voltage and longer pulse widths. For concrete with rough surfaces, in order to reduce surface scattering loss and ensure effective signal strength, a lower frequency can be used to bypass the influence of surface roughness, and the energy may be increased to ensure sufficient feedback. For thin-walled concrete components, in order to avoid waveform overlap and accurately analyze wave velocity and attenuation, a wide-band narrow pulse can be selected to clearly separate the front-end surface wave and the rear-end reflected wave.

[0034] Therefore, the adaptive adjustment mechanism of the active excitation unit in this embodiment can solve the influence of different concrete textures (such as aggregate type, particle size, and moisture content) and test scenarios (such as surface roughness and environmental noise) on signal propagation. The system in this embodiment can be set with multiple adjustable active excitation strategies: Frequency adaptive selection: The system has multiple built-in excitation frequency bands (e.g., 20kHz–100kHz) and automatically selects the optimal excitation frequency band based on the preset concrete type (e.g., ordinary C30, high-strength C80, ultra-high-strength C120). For example, for high-strength concrete, a higher frequency is used to improve resolution; for porous or rough-surface concrete, a lower frequency is used to enhance penetration.

[0035] Adjustable energy and waveform: The excitation unit supports pulse width modulation (PWM) and voltage amplitude adjustment to adapt to concrete of different depths and densities. The system can dynamically adjust the excitation energy based on the strength of the initial feedback signal to avoid signal saturation or excessive weakness.

[0036] Frequency sweep excitation mode: When the type of concrete is uncertain, the system can start the frequency sweep mode, emit a series of pulses with gradually changing frequencies, and select the optimal excitation parameters by analyzing the response at each frequency.

[0037] In practical applications, in order to achieve centralized and lightweight measurement, the active excitation unit, acoustic wave sensing unit and data processing unit in this embodiment are all integrated into the same device, and a display screen can be set on the outer shell of the device to display the estimated value of concrete strength. In this embodiment, the active excitation unit is an active excitation probe with a built-in piezoelectric ceramic stack; the acoustic wave sensing unit is a miniature microphone array or optical sensor, located on the same side of the outer shell of the device where the active excitation probe is located.

[0038] Here, in order to effectively capture the feedback acoustic wave signal radiated from the concrete surface and to minimize signal loss and measurement error caused by different specimen textures (such as smooth and rough surfaces) and different detection scenarios (such as environmental noise, temperature and humidity changes), the system in this embodiment adopts a front-end signal enhancement and acquisition strategy, an adaptive signal preprocessing and noise reduction strategy, and feature signal extraction and scene adaptive identification at the signal sensing and preprocessing level.

[0039] Specifically, for front-end signal enhancement and acquisition strategies, multi-sensor arrays and beamforming technology can be employed. The acoustic wave sensing unit does not use a single microphone, but rather a miniature microphone array (e.g., a 2x2 or linear array), with these sensors arranged in a specific geometric shape in space. At this point, through digital beamforming algorithms, the system can spatially amplify the acoustic wave signal from the normal direction of the concrete test surface while suppressing environmental noise from other directions (such as wind noise and mechanical vibration noise). This is equivalent to creating an "auditory spotlight," precisely "focusing" on the measured point and significantly improving the signal-to-noise ratio.

[0040] In addition, this embodiment can also use optical sensing as an alternative for high-noise scenarios. In this case, in situations with extremely high ambient noise or where absolute non-contact is required, the system can use a laser Doppler vibrometer as the sound wave sensing unit. In this case, the laser Doppler vibrometer indirectly "hears" the sound waves by detecting the tiny vibration velocity generated on the concrete surface due to the propagation of stress waves. It is completely unaffected by air noise and can provide extremely high displacement resolution, making it particularly suitable for capturing the weak vibration signals of ultra-high strength concrete.

[0041] For adaptive signal preprocessing and noise reduction strategies, to achieve adaptive noise cancellation, the system can integrate a reference microphone on the probe housing. This microphone is specifically used to collect ambient background noise and does not receive feedback signals from the concrete. In this case, a minimum mean square adaptive filter is used, taking the noise collected by the reference microphone as a reference input, and dynamically subtracting related noise components from the main sensing signal in real time. Therefore, steady-state or slowly varying background noise (such as continuous rumbling in a factory workshop) can be effectively eliminated in the data processing unit. To achieve synchronous triggering and signal averaging, the system uses the transmission time of the active excitation pulse as the precise trigger point and performs multiple (e.g., 32 or 64) acquisitions of the feedback signal. Since the excitation signals are synchronous, the true acoustic response of the concrete is reflected in each acquisition. Since the time is fixed and random noise is uncorrelated, the time-domain signal averaging method can be used to superimpose and average the signals collected multiple times. This will enhance the real signal and reduce random noise by canceling each other out. In addition, in order to achieve time-varying gain control, in this embodiment, a programmable amplifier can be used before the signal enters the analog-to-digital converter. Its gain changes with time (i.e. with signal depth) according to a preset curve. At this time, since the energy of the stress wave attenuates when it propagates in the concrete, the echo signal in the later stage is very weak. Therefore, a low gain can be set in the early stage of the signal to prevent saturation. Then, the gain is increased exponentially with time (corresponding to the depth of wave propagation) to "flatten" the amplitude of the entire time-domain signal, ensuring that even the weak deep reflection signal can be fully digitized by the analog-to-digital converter and avoid information loss.

[0042] Finally, regarding feature signal extraction and scene adaptive identification, in this embodiment, to achieve intelligent separation of direct and reflected waves, the surface wave (earliest arrival), volume wave (longitudinal and transverse waves), and possible bottom reflected waves are first separated from the time-domain signal according to the propagation time window. For different test scenarios (such as thin plate components or large-volume foundations), the system will intelligently select to analyze different wave components. For example, for shallow detection, the focus is on analyzing high-frequency surface waves, while for strength assessment, more attention is paid to the volume wave characteristics carrying internal structural information. To achieve filtering optimization based on material properties, the system can incorporate multiple digital bandpass filter banks, whose passband range corresponds to the typical acoustic response frequency band of concrete. Therefore, before the detection begins, the system will perform a rapid frequency sweep measurement and automatically select an optimal analysis frequency band based on the spectral energy distribution of the feedback signal. This can filter out unimportant frequency components in the scenario (such as low-frequency vibration interference or high-frequency electronic noise) and focus on the main frequency band that best reflects the strength information, thereby reducing errors introduced by irrelevant information.

[0043] Therefore, through the comprehensive processing of space (sensor array), time (signal averaging), frequency (adaptive filtering), and multiple physical quantities (acoustic / optical sensing) described above, the acoustic wave sensing unit of this embodiment can stably capture high-quality concrete feedback acoustic wave signals under various complex working conditions, laying a reliable data foundation for subsequent feature extraction and strength inversion.

[0044] It should be noted that the core task of the data processing unit in this embodiment is to convert the received raw time-domain signal into characteristic parameters that are highly sensitive to and robust to concrete strength, and finally output the compressive strength value through a mapping model with self-learning capability.

[0045] Traditional ultrasonic testing often only uses time-domain (wave velocity) or frequency-domain (spectrum) analysis, which does not fully utilize the information. However, this embodiment uses joint time-frequency analysis to capture the dynamic characteristics of the frequency components of stress waves changing over time during propagation. Therefore, in order to reveal the internal structure of concrete, it is necessary to process the preprocessed signal using short-time Fourier transform and wavelet transform. In this embodiment, short-time Fourier transform is used to perform preliminary and rapid time-frequency observation of the signal. Through a sliding time window, the local spectrum within each time period is calculated to generate the time spectrum. For wavelet transform: in order to provide good time resolution at high frequencies and good frequency resolution at low frequencies, this embodiment prefers Morlet wavelet or Daubechies series wavelets as the basis wavelets because they are similar to the attenuation oscillation characteristics of concrete stress waves.

[0046] In practical applications, the preprocessed signal can first be represented in time and frequency, that is, the preprocessed echo signal s(t) can be subjected to continuous wavelet transform to obtain its wavelet coefficient matrix CWT(a, b), where a is the scale (corresponding to frequency) and b is the translation (corresponding to time). This matrix completely describes the distribution of the signal energy in the two-dimensional plane of time and frequency. Then feature extraction is performed, that is, more robust feature parameters are directly extracted from the time and frequency representation. Therefore, in this embodiment, when the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, the extracted feature parameters include spectral centroid shift, energy attenuation rate, resonant frequency change and wavelet entropy.

[0047] After the above feature parameters are determined, the data processing unit obtains a preliminary intensity estimate using the trained machine learning model and determines whether the confidence level of this detection meets the requirements. If it meets the requirements, the preliminary intensity estimate can be directly used as the final intensity estimate. If it does not meet the requirements, a confidence level assessment is needed. Therefore, in this embodiment, the data processing unit obtains a preliminary intensity estimate using the trained machine learning model and determines whether the confidence level of this detection meets the requirements. When it does not meet the requirements, it means: The trained machine learning model pre-stores datasets covering different types of concrete in different detection scenarios; When the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, it extracts the spectral centroid shift, energy attenuation rate, resonant frequency change and wavelet entropy, and forms a feature vector. Calculate the Mahalanobis distance or k-nearest neighbor distance between the feature vector and the features of various samples in the dataset. If the distance exceeds the preset distance, it means that the confidence level of this detection does not meet the requirements.

[0048] In practical applications, a machine learning model based on acoustic features and strength mapping can be pre-built offline. During model building, a large number of standard concrete specimens covering C15 to C150+ need to be prepared or collected in the laboratory, covering key variables such as different coarse aggregate types (limestone, granite), maximum particle size, and water content. Then, a feature library is established. When the feature library is established, each specimen is tested on different surfaces (polished, template surface, rough surface), and all the above feature parameters are extracted to form a high-dimensional feature vector (containing the above feature parameters). After the dataset is prepared, the model can be selected and trained. In this embodiment, a gradient boosting decision tree, such as XGBoost or LightGBM, can be selected. During model training, the feature vector corresponding to each specimen and its measured compressive strength (obtained through a pressure machine destructive test) can be used as labels and input into the GBDT model for supervised learning. The model learns the complex mapping relationship f(feature)->strength from multi-dimensional acoustic features to a single strength value.

[0049] After the above model is established, model adaptation and online calibration can be performed. At this point, scene recognition and model fine-tuning can be performed first, which is achieved through the following steps: Step 1: During on-site testing, the system first extracts the feature vector of the current signal.

[0050] Step 2: Input the vector into the trained master model to perform preliminary strength estimation.

[0051] Step 3: Simultaneously, the system calculates the Mahalanobis distance or k-nearest neighbor distance between the feature vector and the features of various samples in the model training set to assess uncertainty.

[0052] Step 4: If the system determines that the current measurement point differs significantly from the known data distribution of the model (for example, encountering special aggregates that have not been found in the training set), it will give a "low confidence" prompt when displaying the results.

[0053] Secondly, incremental learning and user calibration can be performed, which can be achieved through the following steps: Step 1: Users can drill core samples on-site or find areas with known design strength on the same structure as reference points.

[0054] Step 2: The system performs testing at the reference point, and the user inputs the reference strength value (measured value or design value of the core sample) into the system.

[0055] Step 3: The system uses this new (feature vector, true strength) data pair to perform incremental learning on the original GBDT model, that is, without retraining the entire model, it uses a small amount of new data to locally fine-tune the model's decision boundary.

[0056] Step 4: After several such calibrations, the model will gradually "learn" the characteristics of local materials and environment, and the detection accuracy in the area will be significantly improved.

[0057] Example 2

[0058] Based on Example 1, this example provides a non-destructive testing method for concrete strength based on adaptive excitation and machine learning. See the flowchart below. Figure 2 The method may include the following steps: S1. Store the preset excitation parameters in the data processing unit, and store the established and trained machine learning model based on acoustic features and intensity mapping. S2. The preset excitation parameters are sent to the active excitation unit through the data processing unit, and the preset excitation parameters are received through the active excitation unit. After power-on and before formal testing, a pre-scan operation is performed, and a preset pulse wave is emitted to the concrete with the preset excitation parameters. S3. The first feedback acoustic signal of the concrete surface responding to the preset pulse wave is captured by the acoustic wave sensing unit and sent to the data processing unit. S4. When the data processing unit receives the first feedback acoustic signal, it extracts its spectrum information and adaptively determines the best excitation parameters for the current concrete type in the current detection scenario based on the spectrum information, and sends them to the active excitation unit. S5. When the active excitation unit receives the optimal excitation parameters, it emits stress waves to the concrete with the optimal excitation parameters for formal detection, and captures the second feedback acoustic wave signal of the concrete surface in response to the stress waves through the acoustic wave sensing unit, and sends it to the data processing unit. S6. When the data processing unit receives the second feedback acoustic signal, it extracts its feature parameters and inputs them into the trained learning model. The trained machine learning model obtains a preliminary intensity estimate and determines whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate; otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The trained machine learning model is then updated and optimized.

[0059] As can be seen from the description of Embodiment 1, the application scenario and implementation principle of this embodiment are the same as those of Embodiment 1, so they will not be repeated here.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-destructive testing system for concrete strength based on adaptive excitation and machine learning, characterized in that, include: The active excitation unit is used to receive preset excitation parameters, perform a pre-scan operation after power-on and before formal testing, and emit a preset pulse wave to the concrete with the preset excitation parameters. When the optimal excitation parameters are received, it emits a stress wave to the concrete with the optimal excitation parameters for formal testing. The acoustic wave sensing unit is used to capture the first feedback acoustic wave signal of the concrete surface in response to a preset pulse wave, and to capture the second feedback acoustic wave signal of the concrete surface in response to a stress wave, and send them to the data processing unit. The data processing unit stores preset excitation parameters, which are used to send the preset excitation parameters to the active excitation unit. When the first feedback acoustic signal is received, its spectrum information is extracted, and the optimal excitation parameters for the current concrete type under the current detection scenario are adaptively determined based on the spectrum information and sent to the active excitation unit. It also stores a pre-built and trained machine learning model based on acoustic features and intensity mapping. When the second feedback acoustic wave signal is received, its feature parameters are extracted and input into the pre-trained learning model. The pre-trained machine learning model is used to obtain a preliminary intensity estimate and to determine whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate. Otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The pre-trained machine learning model is then updated and optimized.

2. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 1, characterized in that, The data processing unit also stores information on concrete specimens of different textures and information on different testing scenarios; The preset excitation parameters corresponding to the preset pulse wave have a frequency range of 20kHz-150kHz. When the data processing unit receives the first feedback acoustic signal, it extracts its spectral information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectral information. This means that the data processing unit identifies the frequency band with the highest signal-to-noise ratio and the most significant response based on the spectral information extracted from the first feedback acoustic signal, and extracts the center frequency and bandwidth parameters of the frequency band with the most significant response as the optimal excitation parameters.

3. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 2, characterized in that, The energy levels of the preset pulse wave are adjustable, and its excitation voltage amplitude is controlled by the driving circuit to be between 100V and 1000V, and its pulse width is between 1μs and 50μs. The waveform of the preset pulse wave is a square wave pulse and / or a narrow pulse and / or a frequency-modulated pulse and / or a window function modulated pulse.

4. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 3, characterized in that, When the data processing unit receives the first feedback acoustic signal, it extracts its spectral information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectral information, including the following steps: The active excitation unit performs the first excitation according to the preset excitation parameters, and the data processing unit receives the first feedback acoustic signal; The data processing unit extracts the spectral information from the first feedback acoustic signal, calculates the key quality indicators of the first feedback acoustic signal based on the spectral information, and determines whether the key quality indicators have all reached the preset threshold. If they have not reached the threshold, the excitation parameters are adjusted and the adjusted excitation parameters are sent to the active excitation unit. The active excitation unit performs cyclic excitation according to the adjusted excitation parameters, and the data processing unit continues to calculate the key quality indicators of the first feedback acoustic signal and determine whether all key quality indicators have reached the preset threshold. The process is repeated until all key quality indicators have reached the preset threshold, and the center frequency and bandwidth parameters of the frequency band with the most significant response in the spectrum information corresponding to all reaching the preset threshold are taken as the optimal excitation parameters.

5. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 1, characterized in that, The active excitation unit, acoustic wave sensing unit, and data processing unit are all integrated into the same device; The active excitation unit is an active excitation probe with a built-in piezoelectric ceramic stack; The acoustic wave sensing unit is a miniature microphone array or an optical sensor, located on the same side of the housing of the active excitation probe.

6. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 5, characterized in that, The device housing is also equipped with a reference microphone, and a filter is installed inside the housing; The reference microphone is used to collect ambient background noise; The filter is a minimum mean square adaptive filter, used to receive ambient background noise and a first feedback acoustic signal or a second feedback acoustic signal, and to use the ambient background noise as a reference input to dynamically remove the noise components associated with the first feedback acoustic signal or the second feedback acoustic signal, and then transmit the dynamically removed noise components to the data processing unit.

7. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 6, characterized in that, The device housing is also equipped with a variety of digital bandpass filter groups; The various digital bandpass filter banks are used to dynamically remove noise components related to the environmental background noise from the first or second feedback acoustic signal, and filter out frequency components unrelated to the detection scene and concrete type.

8. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 1, characterized in that, When the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, the extracted feature parameters include spectral centroid shift, energy attenuation rate, resonant frequency change, and wavelet entropy.

9. The non-destructive testing system for concrete strength based on adaptive excitation and machine learning according to claim 8, characterized in that, The data processing unit obtains a preliminary strength estimate through a trained machine learning model and determines whether the confidence level of this detection meets the requirements. If it does not meet the requirements, it means: The trained machine learning model pre-stores datasets covering different types of concrete in different detection scenarios; When the trained machine learning model based on acoustic features and intensity mapping receives the second feedback acoustic signal, it extracts the spectral centroid shift, energy attenuation rate, resonant frequency change and wavelet entropy, and forms a feature vector. Calculate the Mahalanobis distance or k-nearest neighbor distance between the feature vector and the features of various samples in the dataset. If the distance exceeds the preset distance, it means that the confidence level of this detection does not meet the requirements.

10. A nondestructive testing method for concrete strength based on adaptive excitation and machine learning, applied to a nondestructive testing system for concrete strength based on adaptive excitation and machine learning as described in any one of claims 1-9, characterized in that, Includes the following steps: The data processing unit stores preset excitation parameters and a machine learning model based on acoustic features and intensity mapping that has been built and trained. The data processing unit sends the preset excitation parameters to the active excitation unit, and performs a pre-scan operation after the active excitation unit is powered on and before the formal test, and emits a preset pulse wave to the concrete with the preset excitation parameters. The first feedback acoustic signal of the concrete surface in response to the preset pulse wave is captured by the acoustic wave sensing unit and sent to the data processing unit. When the data processing unit receives the first feedback acoustic signal, it extracts its spectrum information and adaptively determines the optimal excitation parameters for the current concrete type in the current detection scenario based on the spectrum information, and sends them to the active excitation unit. When the active excitation unit receives the optimal excitation parameters, it emits stress waves to the concrete with the optimal excitation parameters for formal detection, and captures the second feedback acoustic wave signal of the concrete surface in response to the stress wave through the acoustic wave sensing unit, and sends it to the data processing unit. When the data processing unit receives the second feedback acoustic signal, it extracts its feature parameters and inputs them into the trained learning model. The trained machine learning model obtains a preliminary intensity estimate and determines whether the confidence level of this detection meets the requirements. If it does, it is used as the final intensity estimate; otherwise, the optimal excitation parameters are adjusted and a re-detection is prompted. The trained machine learning model is then updated and optimized.

Images