A non-destructive prefabricated component internal defect detection method and system

Abstract

The present invention discloses a method and system for non-destructive prefabricated component internal defect detection. The method comprises: acquiring a detection scanning image; wherein the detection scanning image is obtained by a radar host transmitting a broadband narrow pulse signal to the prefabricated component to be detected and synchronously acquiring the echo signal; performing an alignment operation on the detection scanning image based on a standard scanning image of a standard prefabricated component; performing a denoising operation on the detection scanning image after the alignment operation; and performing defect type detection and defect area labeling on the denoised detection scanning image using a trained defect detection network to achieve the effect of reducing power consumption and radiation.

Description

Technical Field

[0001] The present invention relates to the technical field of building quality detection, and in particular to a method and system for detecting internal defects of non-destructive prefabricated components. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] In the field of prefabricated buildings, various types of prefabricated concrete components are used extensively, and their quality directly affects the safety and service life of the final building. However, due to the interference of various factors in the production, transportation, lifting, reinforcement and other links of prefabricated concrete components, cracks may appear in the components. The presence of cracks not only affects the performance of the structure but also reduces the durability of the concrete structure. In order to accurately measure the depth and direction of the cracks, it is necessary to inspect the internal conditions of the prefabricated components. The commonly used measurement methods are:

[0004] Core sampling: This method involves drilling core samples from concrete structures or components to prepare concrete specimens. Testing of these core samples verifies and validates the strength and integrity of the concrete in the area. This method provides a relatively accurate reflection of the concrete condition near the sampling area, but results may vary significantly at greater distances. Furthermore, for lower-strength concrete structures, the core drilling process can easily disrupt the bond between the mortar and coarse aggregate, affecting the accuracy of test results. This method is generally not suitable for structures with high integrity requirements, such as prestressed concrete structures.

[0005] Rebound method: This method uses a spring-driven hammer to strike the concrete surface through a striking rod. The distance the hammer rebounds is measured, and the ratio of the rebound distance to the initial spring length is used as a strength-related indicator to estimate concrete strength. This method is simple in principle and easy to test, but requires the test value to have a good correlation with the concrete strength under certain conditions, and it cannot effectively represent the specific damage within the concrete.

[0006] Ultrasonic method: This method utilizes the proportional relationship between the longitudinal propagation velocity and frequency of ultrasound and the elastic modulus and density of concrete. An ultrasonic transducer transmits ultrasonic pulses, and a receiver at the other end detects the ultrasonic wave velocity, amplitude, and other waveform parameters, thereby inferring the strength and internal structural defects of the structure. This method is non-destructive and has strong penetrating power. However, it requires the concrete to be of essentially the same quality both inside and outside. Furthermore, the test requires a grid drawing and point-by-point detection, resulting in low detection efficiency.

[0007] X-ray method: This method uses medical or industrial CT to scan concrete specimens, obtaining images of the concrete's internal aggregate, mortar, pores, and cracks. This method is also non-destructive and can examine the specimen's subtle internal structure. However, this method is relatively expensive, lacks portability, and emits radiation. It can generally only be used to test relatively small structural components and is primarily used in concrete research, with limited practical application.

[0008] Ground Penetrating Radar (GPR): This method uses a transmitting antenna to emit extremely narrow electromagnetic detection pulses. As these electromagnetic signals propagate through underground media, their electromagnetic field strength, propagation path, and waveform information vary depending on the medium. The receiving system extracts effective information such as the echo's round-trip time, phase, and amplitude, allowing the properties and physical structure of the medium to be inferred. This method not only locates the location and structural information of the detected target, but also uses digital signal processing to image the waveform data, facilitating user decision-making.

[0009] Each of the aforementioned testing methods has its pros and cons, and it's difficult to achieve a balance between non-destructiveness, convenience, adaptability, safety, and affordability. For internal defect detection in precast concrete components, core sampling is destructive, resulting in high drilling costs and poorly representative test results. The rebound method cannot reveal internal structural details. The ultrasonic method requires point-by-point grid testing, resulting in low monitoring efficiency. The X-ray method requires a strict radiation shielding environment and is unsuitable for on-site testing. Ground-penetrating radar offers advantages in non-destructiveness, efficiency, structure, debuggability, performance, and safety.

[0010] However, traditional ground-penetrating radar is typically used for underground inspection or detection of thick concrete structures. The pulse signal intensity emitted by the transmitter is relatively high. Although concrete strongly attenuates electromagnetic signals, the thickness of prefabricated components used in most buildings is less than one meter. Excessively large transmitted pulses not only shorten the battery life of the device but also cause electromagnetic leakage, potentially interfering with other wireless communication or sensor equipment. Data acquisition systems are significantly affected by jitter noise, which causes clock jitter interference. Furthermore, detection resolution is primarily determined by the pulse bandwidth, and pulse amplitude and bandwidth are mutually restrictive.

[0011] The radar receiver collects ultra-wideband narrow pulses through a high-speed analog-to-digital converter. If the device is in multi-channel mode, the amount of data collected per unit time will be greater. If real-time analysis of the measured structure is to be achieved, the data transmission channel from the radar host to the analysis device will become a bottleneck.

[0012] Ground-penetrating radar image analysis involves processing the scanned image and then combining it with the subjective experience of the inspector to determine whether internal defects exist. However, for real-time inspection of prefabricated components, on the one hand, the components often contain interlaced rebar and aggregates. These discontinuous media conditions are reflected in the scanned image, making the determination of internal defects more difficult. On the other hand, radar scanned images are not as intuitive as traditional images, and wiring sometimes requires adjustments to the reflected signal to weaken the pulse trailing edge and reduce the pulse width. The time and accuracy of the inspector's judgment of relatively complex images will fluctuate, affecting the real-time detection. Finally, the avalanche effect of the ground-penetrating radar transistor produces a severe trailing edge of the pulse, requiring the use of high-power step diodes to shape the pulse, reduce the pulse width, and minimize the attenuation of the pulse amplitude. Summary of the Invention

[0013] In order to address the shortcomings of the existing technology, the present invention provides a method and system for non-destructive prefabricated component internal defect detection; combined with the scenario characteristics of prefabricated component detection, adaptive adjustment of the transmitted pulse signal strength is performed, and an adjustment method is given to achieve the effect of reducing power consumption and radiation; different from the data processing of traditional ground penetrating radar, the original point data is used, and the data is compressed and decompressed, alleviating the demand for bandwidth during data transmission; a cloud-edge collaborative defect judgment framework is proposed, image registration is performed within the framework to eliminate detection interference, and a U-net network is used for defect judgment. The data of multiple devices can be collected in the cloud for model training, and each detection terminal is only responsible for running the trained network and performing detection processing.

[0014] In a first aspect, the present invention provides a method for non-destructive detection of internal defects of prefabricated components;

[0015] A non-destructive prefabricated component internal defect detection method comprising:

[0016] Acquire a detection scanning image; wherein the detection scanning image is obtained by the radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously collecting the echo signal;

[0017] Based on the standard scanned image of the standard prefabricated component, the detection scanned image is aligned; and the detection scanned image after the alignment operation is subjected to denoising.

[0018] The trained defect detection network is used to detect defect types and mark defect areas on the denoised detection scan images.

[0019] In a second aspect, the present invention provides a non-destructive prefabricated component internal defect detection system;

[0020] A non-destructive prefabricated component internal defect detection system, comprising:

[0021] An acquisition module is configured to: acquire a detection scanning image; wherein the detection scanning image is obtained by the radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously acquiring the echo signal;

[0022] A pre-processing module is configured to: perform an alignment operation on the detection scan image based on the standard scan image of the standard prefabricated component; and perform a denoising operation on the detection scan image after the alignment operation;

[0023] The defect detection module is configured to: use the trained defect detection network to detect defect types and mark defect areas on the denoised detection scan image.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) Through the MARX circuit with configurable avalanche transistor stages and a set of signal reception feedback adjustment processes, the transmitter pulse intensity can be adaptively adjusted, reducing the system operating power and extending the battery life; at the same time, it reduces electromagnetic radiation and reduces the interference between the nondestructive testing system and other wireless systems.

[0026] (2) By caching the raw data in the radar host, the scanned data is processed by JPEG-LS lossless compression before being sent to the data processing and analysis module, eliminating the bandwidth bottleneck of a large amount of raw data transmission.

[0027] (3) The SIFT algorithm is used for image registration to eliminate the interference of internal standard steel bars and embedded parts on defect detection; the Unet image semantic segmentation network is used for automatic defect detection; the cloud server aggregates sample data from multiple detection terminals for training, achieving continuous improvement in the performance of the defect detection network.

[0028] (4) By capturing the internal defect images of prefabricated components through anisotropic structural tensors, the noise impact between multimodal images is greatly reduced, and the real data is more reliable; it has good robustness to nonlinear intensity differences and enhances the matching function of the algorithm. At the same time, feature fusion is used to sample and stack the feature layers, which is conducive to fusing the effective feature layers of all features and classifying each pixel. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0030] Figure 1 This is a structural diagram of a non-destructive prefabricated component internal defect detection method and system according to Example 1;

[0031] Figure 2 The radar host structure of embodiment 1;

[0032] Figure 3 The narrow pulse generating circuit structure with adjustable amplitude and pulse width of embodiment 1;

[0033] Figure 4 This is the output pulse calibration process of Example 1;

[0034] Figure 5 This is the client structure of embodiment 1. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0037] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0038] All data in this embodiment is obtained in compliance with laws and regulations and based on the consent of the user, and is used legally.

[0039] Embodiment 1 This embodiment provides a non-destructive method for detecting internal defects of prefabricated components;

[0040] like Figure 1 As shown, a non-destructive prefabricated component internal defect detection method includes:

[0041] S101: Acquire a detection scanning image; wherein the detection scanning image is obtained by a radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously collecting the echo signal;

[0042] S102: performing an alignment operation on the detection scan image based on the standard scan image of the standard prefabricated component; and performing a denoising operation on the detection scan image after the alignment operation;

[0043] S103: Using the trained defect detection network to perform defect type detection and defect area labeling on the denoised detection scan image.

[0044] It should be understood that the radar host generates broadband narrow pulses with adjustable amplitude and performs equivalent synchronous acquisition of echoes to obtain the original detection data of the prefabricated component under test, and performs block and lossless compression processing on the combined scanned image.

[0045] Further, if Figure 2 As shown, the radar host includes: a main control module; the main control module is connected to the positioning module, the power management module, the data compression and transmission module, the signal receiving and processing module and the pulse generating and sending module respectively;

[0046] Among them, the main control module is used to coordinate the working timing of each module, analyze the control instructions given by the data compression and transmission module, and control the delay chip to generate a sending trigger signal that meets the synchronous sampling delay requirements;

[0047] The pulse generation and transmission module uses a high-voltage pulse generator circuit to generate high-frequency narrow pulses, and adjusts the amplitude of the output signal by controlling the number of cascade stages. The output pulse signal is transmitted to the transmitting antenna, which converts the signal into an electromagnetic pulse and sends it to the medium;

[0048] The signal receiving and processing module uses an analog-to-digital converter to perform equivalent sampling of the pulse echo according to the sampling trigger signal generated by the main control module. The collected raw data is cached and the resulting scanned 2D grayscale image is compressed according to the JPEG-LS standard. The resulting image is entropy-coded for the prediction error. The height is determined by the number of equivalent sampling points in a single waveform, and the length is the number of scan channels obtained within 1 second. To improve the real-time performance of the compression process, the compression process is performed on the FPGA in the radar host.

[0049] The data compression and transmission module realizes two-way data transmission. On the one hand, it transmits the compressed data generated by the radar host and the original data of the selected target area to the client. On the other hand, it transmits the control instructions issued by the client to the main control module for adjusting the receiving signal gain and setting the working mode.

[0050] The positioning module uses the Beidou positioning module for wide-area positioning, and uses the ranging wheel on the antenna to locate the local position of the object being measured during the detection process;

[0051] The power management module cuts off the power supply to the pulse generation and transmission module, signal reception and processing module and transceiver antenna in a timely manner according to the working status or standby command of the main control module to achieve low-power operation of the system.

[0052] The main control module coordinates the working timing of each module, analyzes the control instructions from the data transmission module, and controls the delay chip to generate a sending trigger signal that meets the synchronous sampling delay requirements.

[0053] The power management module timely cuts off the power supply to the pulse generation and transmission module, the signal reception and processing module and the transceiver antenna according to the working status or standby command of the main control module, so as to achieve low power consumption operation of the system.

[0054] Furthermore, the data compression and transmission module realizes bidirectional data transmission, on the one hand, transmitting the compressed data generated by the radar host and the original data of the selected target area to the client, and on the other hand, transmitting the control instructions issued by the client to the main control module for adjusting the received signal gain and setting the working mode;

[0055] The signal receiving and processing module uses an analog-to-digital converter to perform equivalent sampling of the pulse echo based on the sampling trigger signal generated by the main control module. The collected raw data is cached and the resulting scanned 2D grayscale image is compressed according to the JPEG-LS standard. The height of the compressed image is determined by the number of equivalent sampling points in a single waveform, and the length is the number of scan channels obtained within a single second. To improve the real-time performance of the compression process, the compression process is performed on the FPGA in the radar host.

[0056] JPEG-LS (A New Standard for Lossless Compression of Still Images) uses the three most adjacent pixels of the pixel to be predicted as prediction reference points.

[0057] The pulse generating and sending module includes several stages of narrow pulse generating circuits;

[0058] The first stage narrow pulse generating circuit includes a DC power supply VCC, the negative electrode of the DC power supply VCC is grounded; the positive electrode of the DC power supply VCC is connected to the current limiting resistor R through the on-off switch K0. c1 The first end of the current limiting resistor R c1 The second end is connected to the collector of transistor C0. The three stages of the transistor are base, emitter and collector respectively. The emitter of transistor C0 is grounded.

[0059] The second stage narrow pulse generating circuit includes the positive electrode of the DC power supply VCC connected to the current limiting resistor R through the on-off switch K0.c2 The first end of the current limiting resistor R c2 The second end is connected to the collector of the avalanche transistor, and the emitter of the avalanche transistor is connected to the current limiting resistor R e1 Grounded; the base of the avalanche transistor is connected to the collector of the transistor C0 through the energy storage capacitor C1;

[0060] The i-th stage narrow pulse generating circuit includes the positive electrode of the DC power supply VCC connected to the on-off switch K0 through the on-off switch K ci The first end of the on-off switch K ci The second end of the current limiting resistor R ci The first end of the current limiting resistor R ci The second end of is connected to the collector of the i-th avalanche transistor, and the emitter of the i-th avalanche transistor is connected to the collector of the i-th avalanche transistor through the current limiting resistor R e(i-1) Ground; the base of the i-th avalanche transistor and the on-off switch K bi The first end of the on-off switch K bi The second end of the capacitor C i-1 Connected to the collector of the i-1th avalanche transistor; the collector of the i-th avalanche transistor is connected to the energy storage capacitor C i With the load resistor R l The first end of the load resistor R l The second end is grounded.

[0061] The pulse generation and sending module consists of a charging circuit and a discharging circuit, such as Figure 3 As shown. The charging circuit includes a DC power supply (VCC), a current limiting resistor (R ci +R ei ) and energy storage capacitor (C i ); the discharge circuit includes a storage capacitor (C i ), avalanche transistor and load resistor (R l The avalanche transistors at each level close to the load in the circuit can be connected through the multi-select switch (SW) and the on-off switch (K ci , K bi ) to achieve access or bypass.

[0062] Further, if Figure 4 As shown, the pulse generation and transmission module adaptively adjusts the strength of the transmitted pulse signal before entering the normal measurement mode; the specific adjustment steps include:

[0063] Before the formal measurement, the transceiver antenna is placed on the surface of the prefabricated component to be measured. The radar host receives the calibration command sent by the client through the network interface and enters the output pulse calibration mode;

[0064] The main control module of the radar host sets the gain value of the signal receiving gain and sends the gain value to the signal receiving and processing module. The pulse generation circuit is only connected to the first two levels, the VCC switch is turned on, and the energy storage capacitor is charged. The main control module converts the pulse signal into a discrete digital signal through an analog-to-digital conversion through a delay circuit, and then uses a multi-stage shift register as a delay circuit. Driven by the clock signal, it shifts step by step, and the signal is delayed and a trigger pulse is generated, causing the avalanche transistor to undergo an avalanche effect and generate a high-frequency narrow pulse.

[0065] The signal receiving and processing module, under the default gain, uses the default voltage amplification factor to count the maximum voltage V of the scan signal collected within 1s max Determine the maximum voltage V max Is it greater than the receiving full-scale voltage V full half of

[0066] If this condition is not met, the VCC switch is turned off, the number of avalanche transistors connected to the pulse generation circuit is increased by 1, the VCC switch is turned on, and the signal receiving and processing module counts the maximum value V of the scan signal collected within 1s at the default gain. max ;

[0067] Judge V max Is it greater than the receiving full-scale voltage V full If the conditions are met, the access level is set to the current level and the normal measurement mode is entered.

[0068] Furthermore, the step S102: performing an alignment operation on the detection scan image based on the standard scan image of the standard prefabricated component specifically includes:

[0069] The scale-invariant feature transform (SIFT) algorithm is used to register the detection scan image based on the standard scan image of the standard prefabricated component.

[0070] According to the registration results, the detection scan image is spliced ​​and re-cut. The size of the re-cut image is consistent with the standard scan image of the standard prefabricated component, and the boundaries are aligned;

[0071] The re-cropped scanned image is normalized with the standard scanned image.

[0072] It should be understood that the acquired scanned images of corresponding positions of the prefabricated components are aligned with the scanned images of standard components of the same specifications through the image registration network.

[0073] Furthermore, the denoising process is performed on the detection scan image after the alignment operation, specifically comprising:

[0074] The brightness values ​​at the corresponding positions of the normalized detection scan image and the corresponding positions of the normalized standard scan image are subtracted, and the difference is then normalized to obtain an image with echo interference eliminated; the image with echo interference eliminated removes interference from steel bars and prefabricated accessories in prefabricated components.

[0075] It should be understood that the values ​​of the corresponding coordinate points of the standard component are subtracted from the scanned image to obtain a scanned image without the influence of the internal steel bars and embedded parts.

[0076] Furthermore, the S103: using the trained defect detection network to perform defect type detection and defect area labeling on the denoised detection scan image; wherein the defect detection network is implemented using a U-net network.

[0077] Furthermore, the trained defect detection network is trained in non-real time on a cloud server. After the training is completed, the network parameters after the training are regularly transmitted back to the client. The client updates the parameters of the U-net network stored in its own storage according to the returned network parameters, and uses the network with updated parameters to implement defect detection.

[0078] Furthermore, the training process of the trained defect detection network includes:

[0079] Constructing a training set; the training set includes detection scan images of known defect types and defect locations; the known defect types include: cracks, cavities, insufficient thickness, excessive steel bar spacing, insufficient steel bar diameter, insufficient aggregate density, and other anomalies;

[0080] The training set is input into the U-net network and the U-net network is trained. When the loss function of the U-net network reaches the minimum value or the number of iterations reaches the set number, the training is stopped to obtain the trained defect detection network.

[0081] Furthermore, both S102 and S103 are implemented through the client.

[0082] The client uses a general PC or a tablet computer with an Ethernet interface for template alignment and defect detection, and is composed of the following functional modules (such as Figure 5 As shown): data receiving and display module, template alignment module, defect detection module, data upload and model update module.

[0083] Among them, the data receiving and display module receives and parses the compressed scanned image through the network (UDP or TCP protocol can be used) and decompresses the image; the display part selects the corresponding layer according to the display mode (it can output the original scanned image, the scanned image after echo interference reduction, the semantic segmentation boundary or the stacked combination of multiple layers), and outputs the scanned image in chronological order to achieve seamless and continuous output of adjacent images.

[0084] The template alignment module receives the decompressed RGB image and reads the corresponding scanned image from the standard component scanned image template database according to the selected component specifications. The two sets of images are then input into the SIFT algorithm for image registration. Since the standard component scanned image (number of scan points * number of scan channels within 1s) is not aligned with the boundary of the detected scanned image, the scanned image obtained by detection is spliced ​​and re-cut according to the registration result. The size of the re-cut image is consistent with the standard component scanned image, and the boundary is aligned. The re-cut scanned image and the standard component scanned image are normalized at the same time, and the brightness value of the corresponding position of the normalized scanned image is subtracted from the brightness value of the corresponding position of the normalized standard component scanned image, and the difference is normalized to between 0 and 255 to eliminate echo interference.

[0085] The defect detection module receives echo-free images and uses a U-net image semantic segmentation network to segment various defects (cracks, cavities, insufficient thickness, excessive rebar spacing, insufficient rebar diameter, insufficient aggregate density, and other anomalies). The network model parameter loading module is responsible for setting network parameters.

[0086] The U-net image semantic segmentation network consists of three parts:

[0087] In the first part, feature extraction is performed through convolution and pooling. The backbone feature extraction part is used to obtain the initial effective feature layer. The backbone feature extraction is a stack of convolution and maximum pooling, and the effective feature layer is used for feature fusion. After each pooling layer, a feature map of a scale is obtained (including the original image scale, there are a total of 5 scales);

[0088] In the second part, the feature extraction part is strengthened through upsampling and deconvolution processing of the same level, and the sampling results are stacked and features are fused; the low-resolution image containing high-level abstract features is converted to high-resolution while retaining the high-level abstract features, and is spliced ​​with the channels corresponding to the feature extraction part. After two convolution operations, a feature map is generated, and classification is performed using a convolution with a convolution kernel size of 1*1 and the number of categories to be segmented. A final fusion and effective feature layer of all features is obtained, and the final heat map of the corresponding number is obtained. The values ​​of the corresponding positions of the obtained heat maps are then used as the input of the softmax function, and the defect category corresponding to the softmax value with a relatively large probability is calculated as the defect type to which the pixel belongs.

[0089] The third part is to obtain the last valid feature layer through prediction, classify each feature point, and mark the defect type on the image. The layer packages the original scanned image, the echo-removed image, and the annotation layer, and corresponds the three layers to the RGB channels of the output image respectively.

[0090] The data upload and model update module is responsible for data exchange with the cloud server. It receives Unet model parameters from the server and, based on the system configuration, selects data that meets the requirements (this can be manually reviewed and confirmed by experienced professionals or cross-validated through other detection methods). The module then packages the data and sends it to the cloud server over the network for server-side Unet training.

[0091] The cloud server consists of: model training database, model training and verification module, and model parameter distribution module.

[0092] The model training database consists of samples containing defect annotation data uploaded by all inspection units on the server. Before model training, the existing samples in the database are randomly divided into a training set and a validation set in an 8:2 ratio.

[0093] The model training and validation module uses samples from the database to train a U-net image semantic segmentation network. Five-fold cross-validation is used during training. Data augmentation is performed using random rotation, random scaling, random elastic transformation, gamma correction, and mirroring. The training process uses stochastic gradient descent, using the Adam optimizer with a learning rate of 3e-4. During training, a scanned image is used as an original sample. This original sample is then divided into multiple overlapping image blocks. These blocks are then fed into the training unit as a training batch to optimize the network parameters.

[0094] The loss function selects Dice loss and uses Dice to calculate the similarity of samples:

[0095]

[0096] Where u is the probability output (softmax output), v is the hard-coded (one hot encoding) segmentation label, K is the number of multi-classification categories, and Dice (Dice coefficient) is the set similarity measurement function.

[0097] The learning rate adjustment strategy is to calculate the exponential moving average loss of the training and validation sets. If the exponential moving average loss of the training set does not decrease by 5e-3 within 30 epochs, the learning rate is decayed by a factor of 5. Training is stopped if the exponential moving average loss of the validation set does not decrease by 5e-3 within 60 epochs, or if the learning rate is less than 1e-6.

[0098] The model parameter delivery module triggers the update and delivery of model data based on the following two conditions: the analysis accuracy of the validation set is improved by more than a preset improvement threshold (e.g., 0.5%); new defect label categories are obtained by clustering other anomalies, and network training is completed.

[0099] Example 2

[0100] This embodiment provides a non-destructive prefabricated component internal defect detection system;

[0101] A non-destructive prefabricated component internal defect detection system, comprising:

[0102] An acquisition module is configured to: acquire a detection scanning image; wherein the detection scanning image is obtained by the radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously acquiring the echo signal;

[0103] A pre-processing module is configured to: perform an alignment operation on the detection scan image based on the standard scan image of the standard prefabricated component; and perform a denoising operation on the detection scan image after the alignment operation;

[0104] The defect detection module is configured to: use the trained defect detection network to detect defect types and mark defect areas on the denoised detection scan image.

[0105] It should be noted that the acquisition module, preprocessing module, and defect detection module described above correspond to steps S101 to S103 in Example 1. The examples and application scenarios implemented by these modules and the corresponding steps are the same, but are not limited to the contents disclosed in Example 1. It should be noted that the above modules, as part of a system, can be executed in a computer system, such as a set of computer-executable instructions.

[0106] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0107] The proposed system can be implemented in other ways. For example, the system embodiment described above is merely illustrative. For example, the above module division is only a logical function division. In actual implementation, other division methods may be used. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not implemented.

[0108] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A non-destructive prefabricated component internal defect detection method, characterized in that: include: Acquire a detection scanning image; wherein the detection scanning image is obtained by the radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously collecting the echo signal; The radar host includes: a main control module; the main control module is connected to the positioning module, the power management module, the data compression and transmission module, the signal receiving and processing module and the pulse generating and sending module respectively; The pulse generating and sending module includes several stages of narrow pulse generating circuits; The i-th stage narrow pulse generating circuit includes a positive electrode of a DC power supply VCC connected to an on-off switch K0 and an on-off switch K1. ci The first end of the on-off switch K ci The second end of the current limiting resistor R ci The first end of the current limiting resistor R ci The second end of is connected to the collector of the i-th avalanche transistor, and the emitter of the i-th avalanche transistor is connected to the collector of the i-th avalanche transistor through the current limiting resistor R e（i-1） Ground; the base of the i-th avalanche transistor and the on-off switch K bi The first end of the on-off switch K bi The second end of the capacitor C i-1 Connected to the collector of the i-1th avalanche transistor; the collector of the i-th avalanche transistor is connected to the energy storage capacitor C i With the load resistor R l The first end of the load resistor R l The second end of is grounded; At the default gain, the maximum voltage V of the scan signal collected within 1s is counted by the default voltage magnification factor. max Determine the maximum voltage V max Is it greater than the receiving full-scale voltage V full half of If this condition is not met, the VCC switch is turned off, the number of avalanche transistors connected to the pulse generation circuit is increased by 1, the VCC switch is turned on, and the signal receiving and processing module counts the maximum value V of the scan signal collected within 1s at the default gain. max ; Judge V max Is it greater than the receiving full-scale voltage V full If the conditions are met, the access level is set to the current level and the normal measurement mode is entered; Based on the standard scanned image of the standard prefabricated component, the detection scanned image is aligned; and the detection scanned image after the alignment operation is subjected to denoising. The trained defect detection network is used to detect defect types and mark defect areas on the denoised detection scan images.

2. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: The main control module is used to coordinate the working timing of each module, analyze the control instructions given by the data compression and transmission module, and control the delay chip to generate a sending trigger signal that meets the synchronous sampling delay requirements; The pulse generation and transmission module uses a high-voltage pulse generator circuit to generate high-frequency narrow pulses, and adjusts the amplitude of the output signal by controlling the number of cascade stages. The output pulse signal is transmitted to the transmitting antenna, which converts the signal into an electromagnetic pulse and sends it to the medium; The signal receiving and processing module uses an analog-to-digital converter to perform equivalent sampling on the pulse echo according to the sampling trigger signal generated by the main control module, caches the collected original data, and compresses the collected scanned two-dimensional grayscale image. The compressed image is entropy-coded for the prediction error value.

3. A non-destructive prefabricated component internal defect detection method according to claim 2, characterized in that: The pulse generation and transmission module adaptively adjusts the strength of the transmitted pulse signal before entering the normal measurement mode; The specific adjustment steps include: Before the formal measurement, the transceiver antenna is placed on the surface of the prefabricated component to be measured. The radar host receives the calibration command sent by the client through the network interface and enters the output pulse calibration mode; The main control module of the radar host sets the gain value of the signal receiving gain and sends the gain value to the signal receiving and processing module. The pulse generation circuit is only connected to the first two levels, the VCC switch is turned on, and the energy storage capacitor is charged. The main control module converts the pulse signal into a discrete digital signal through an analog-to-digital conversion through a delay circuit, and then uses a multi-stage shift register as a delay circuit. Driven by the clock signal, it shifts step by step, and the signal is delayed and a trigger pulse is generated, causing the avalanche transistor to undergo an avalanche effect and generate a high-frequency narrow pulse.

4. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: Based on the standard scanned image of the standard prefabricated component, the detection scanned image is aligned; specifically, the following operations are performed: The scale-invariant feature transformation algorithm is used to register the detection scan images based on the standard scan images of standard prefabricated components; According to the registration results, the detection scan image is spliced ​​and re-cut. The size of the re-cut image is consistent with the standard scan image of the standard prefabricated component, and the boundaries are aligned; The re-cropped scanned image is normalized with the standard scanned image.

5. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: The denoising process of the detection scan image after the alignment operation is performed specifically includes: The brightness values ​​at the corresponding positions of the normalized detection scan image and the corresponding positions of the normalized standard scan image are subtracted, and the difference is then normalized to obtain an image with echo interference eliminated; the image with echo interference eliminated removes interference from steel bars and prefabricated accessories in prefabricated components.

6. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: The trained defect detection network is trained in a non-real-time manner on a cloud server. After the training is completed, the trained network parameters are regularly transmitted back to the client. The client updates the parameters of the U-net network stored in its own storage according to the returned network parameters, and uses the network with updated parameters to perform defect detection.

7. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: The training process of the trained defect detection network includes: Constructing a training set; the training set includes detection scan images of known defect types and defect locations; the known defect types include: cracks, cavities, insufficient thickness, excessive steel bar spacing, insufficient steel bar diameter, and insufficient aggregate density; The training set is input into the U-net network and the U-net network is trained. When the loss function of the U-net network reaches the minimum value or the number of iterations reaches the set number, the training is stopped to obtain the trained defect detection network.

8. A non-destructive prefabricated component internal defect detection method according to claim 1, characterized in that: The trained defect detection network is used to detect defect types and mark defect areas on the denoised detection scan images. The defect detection network is implemented using the U-net network.

9. A system based on the non-destructive prefabricated component internal defect detection method according to any one of claims 1 to 8, characterized in that: include: An acquisition module is configured to: acquire a detection scanning image; wherein the detection scanning image is obtained by the radar host transmitting a broadband narrow pulse signal to the prefabricated component under test and synchronously acquiring the echo signal; A pre-processing module is configured to: perform an alignment operation on the detection scan image based on the standard scan image of the standard prefabricated component; and perform a denoising operation on the detection scan image after the alignment operation; The defect detection module is configured to: use the trained defect detection network to detect defect types and mark defect areas on the denoised detection scan image.

Images