A prefabricated component steel bar protective layer thickness electromagnetic induction detection method and system based on deep learning

Abstract

The application relates to the technical field of building quality detection, and particularly discloses a prefabricated component steel bar protective layer thickness electromagnetic induction detection method and system based on deep learning, which comprises the following steps: synchronously acquiring electromagnetic eigenresponse and infrared thermal images of a component surface, and constructing multi-dimensional reference state data; performing real-time temperature compensation and background normalization on a steel bar region scanning signal according to the reference, and generating an enhanced fusion signal; using a special deep learning model to calculate the steel bar protective layer thickness true value from the signal; analyzing the spatial distribution mode of the thickness true value, and reversely diagnosing specific production link defect types; and the system comprises four function corresponding modules. The application realizes high-precision and self-adaptive detection at a key stage of a production line, can intelligently trace process defects, and significantly improves the intelligent level of prefabricated component production quality control.

Description

Technical Field

[0001] This invention relates to the field of building quality inspection technology, and in particular to an electromagnetic induction detection method and system for the thickness of the concrete cover of precast components based on deep learning. Background Technology

[0002] The thickness of the concrete cover for reinforcing bars is a key parameter for ensuring the structural durability and safety of precast concrete components. Currently, non-destructive testing technology based on the principle of electromagnetic induction is the mainstream method in the industry. This type of technology is usually based on an ideal premise: the test is performed on the component under standard conditions of normal temperature, dryness, and surface cleanliness. Its calibration curves and calculation algorithms are all established around this condition.

[0003] However, in actual production processes, a long-overlooked but crucial niche technical problem is that, from the perspective of quality control timeliness, the stage where timely detection of defects such as rebar positioning deviations is most critical is precisely the hot-working stage immediately after demolding. At this time, the component exhibits three significant non-standard characteristics: a residual release agent film on the surface, a temperature gradient caused by residual curing heat inside, and high moisture content in the concrete itself. These factors drastically alter the propagation and coupling characteristics of the electromagnetic field, causing traditional testing instruments calibrated based on standard conditions to read inaccurately or even completely fail in this scenario. This deprives the production line of the opportunity to intervene in quality within the earliest time window, forcing them to wait until the component cools and dries before testing, resulting in delayed quality feedback and potential batch waste.

[0004] Existing technologies and patents generally avoid this problem, or only provide simple compensation for ambient temperature, without systematically addressing the coupled effects of multiple interferences such as variations in the electromagnetic properties of the concrete matrix, temperature gradients, and surface contamination at the level of detection principles. Therefore, developing a method that can achieve robust and high-precision detection under complex and non-standard conditions after demolding, and that can intelligently correlate the detection results with production process defects, has become an urgent need to promote the intelligent manufacturing of precast components. Summary of the Invention

[0005] The purpose of this invention is to solve the industry problem that existing technologies cannot accurately detect the thickness of the protective layer of steel bars in precast components under non-standard conditions such as surface contamination, temperature gradient and high moisture content after demolding. Therefore, this invention proposes an electromagnetic induction detection method and system for the thickness of the protective layer of steel bars in precast components based on deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based electromagnetic induction detection method for the thickness of the concrete cover of reinforcing bars in precast components includes: S1. Simultaneously acquire the electromagnetic intrinsic response signal and surface temperature distribution thermal image of the unreinforced area of ​​the component under test to form multi-dimensional reference state data; S2. When scanning the reinforcing bar area, the original electromagnetic induction signal is subjected to temperature field compensation and background normalization processing based on the multidimensional reference state data to generate an enhanced fusion signal; S3. Input the enhanced fused signal into a dedicated deep learning model and output the true value of the steel reinforcement protective layer thickness; S4. Analyze the spatial distribution pattern of the true value of the concrete cover thickness of the steel reinforcement, match it with the predefined defect feature library, reverse diagnose and output the defect type in the production process.

[0007] As a further technical solution of the present invention, S1 specifically includes: S11. Based on the design drawings of the precast component to be tested, locate at least one predetermined unreinforced area on its surface; or pre-scan the surface of the component with a low-sensitivity electromagnetic sensor and determine the continuous area with the induced signal intensity below a set threshold as the predetermined unreinforced area. S12. Control the electromagnetic sensor to scan the predetermined unreinforced area at a preset detection frequency and gain, and obtain the electromagnetic intrinsic response signal expressed in complex domain form, including amplitude and phase information; S13. Simultaneously with the scanning of the electromagnetic sensor, the infrared thermal imager is controlled to capture images of the entire detection area, including the predetermined area without reinforcing steel, to obtain thermal images of the surface temperature distribution. S14. Align and associate the electromagnetic intrinsic response signal with the surface temperature distribution thermal image according to the timestamp and spatial coordinates to construct a data set indexed by spatial location and containing the intrinsic electromagnetic response and surface temperature scalar of the location point, thus forming multidimensional reference state data.

[0008] As a further technical solution of the present invention, S14 specifically includes: S141. A unified high-precision timestamp is added to each electromagnetic intrinsic response signal collected by the electromagnetic sensor when scanning the predetermined area without reinforcement and to each frame of surface temperature distribution thermal image captured by the infrared thermal imager. S142. Using a calibration plate or known spatial markers, the scanning coordinate system of the electromagnetic sensor and the imaging coordinate system of the infrared thermal imager are uniformly transformed to the same world coordinate system based on the prefabricated component to be tested. S143. For each of the electromagnetic intrinsic response signals with a timestamp, find the surface temperature distribution thermal image frame at the same time or in the time neighborhood according to the timestamp, and extract the precise temperature value of the corresponding coordinate point from the thermal image frame by interpolation according to its spatial coordinates in the world coordinate system. S144. Each set of spatiotemporally aligned electromagnetic intrinsic response signals, their corresponding spatial coordinates, and the precise temperature value extracted from the thermal image are encapsulated into a structured data unit. The collection of all data units constitutes multidimensional reference state data.

[0009] As a further technical solution of the present invention, S2 specifically includes: S21. When scanning the steel reinforcement area inside the precast component to be tested, the complex form of the original electromagnetic induction signal output by the electromagnetic sensor is collected in real time, and the real-time position coordinates at the time of signal collection are recorded synchronously through the positioning device coupled to the sensor. S22. Based on the real-time position coordinates, spatial mapping is performed on the surface temperature distribution thermal image obtained in step S1 to extract the local background temperature value corresponding to the coordinate point; based on the pre-stored temperature-electromagnetic response lookup table, the amplitude and phase of the original electromagnetic induction signal are linearly corrected using the local background temperature value to obtain the temperature compensation signal. S23. Perform point-by-point ratio calculation on the temperature compensation signal and the electromagnetic intrinsic response signal in the complex domain to eliminate the background influence of the concrete matrix and obtain the background normalized signal. S24. The background normalized signal is subjected to amplitude logarithmic transformation and phase dewinding normalization processing to generate an enhanced fused signal, which represents the equivalent electromagnetic response determined only by the distance between the reinforcing bars under standard dry and cooled conditions.

[0010] As a further technical solution of the present invention, S23 specifically includes: S231. Based on the real-time position coordinates, retrieve at least one electromagnetic eigenresponse signal whose spatial coordinates are closest from the multidimensional reference state data; S232. The temperature compensation signal Represented as complex number The retrieved electromagnetic intrinsic response signal Represented as complex number Perform complex division to calculate the initial normalized complex signal. ;in This is the signal amplitude after temperature compensation. is the base of the natural logarithm. The imaginary unit, The signal phase after temperature compensation. The amplitude of the background reference signal. For the background reference signal phase, The amplitude ratio, Phase difference; S233. The amplitude component of the initial normalized complex signal. Take the common logarithm and its phase components Phase unwinding is performed to obtain stable amplitude-phase pairs; S234. The processed stable amplitude-phase pair is encapsulated into a new complex signal, which is the background normalized signal.

[0011] As a further technical solution of the present invention, S3 specifically includes: S31. The enhanced fusion signal, together with its corresponding real-time position coordinates, is organized according to the scanning time sequence and spatial position to form a two-dimensional data sequence; S32. Extract a locally enhanced fused signal block containing the current point to be solved and its surrounding neighborhood signals from the two-dimensional data sequence using a sliding window method. The data block also contains the amplitude, phase and position context information of the signal. S33. The local enhanced fusion signal block is input into a pre-trained dedicated deep learning model, which directly outputs a scalar value corresponding to the center point of the input signal block, which is the true value of the steel reinforcement protective layer thickness after eliminating the influence of environmental conditions. The dedicated deep learning model is a deep convolutional neural network, which is trained on simulation and measured datasets collected under various temperature, humidity and surface contamination conditions, and is used to establish a complex nonlinear mapping from the signal block to the protective layer thickness.

[0012] As a further technical solution of the present invention, S33 specifically includes: S331. Perform layer normalization processing on the input local enhanced fusion signal block so that the data distribution of its amplitude and phase components meets the input requirements of the dedicated deep learning model; S332. Using the first level of the dedicated deep learning model, a one-dimensional depthwise separable convolutional kernel is used to scan the local enhanced fusion signal block in the spatial dimension to extract primary spatial texture features related to the spacing and orientation of the reinforcing bars; S333. The extracted primary spatial texture features are input into the second layer of the dedicated deep learning model; the second layer is a parallel dual-branch structure, wherein the first branch uses dilated convolution to expand the receptive field and capture large-scale thickness gradient patterns caused by mold side mold bulging or uneven spraying of release agent; the second branch uses an attention mechanism to focus on local high gradient abrupt features caused by the positioning deviation of the steel cage. S334. The feature maps output from the two branches are concatenated, and high-dimensional feature fusion and dimensionality reduction are performed through multiple nonlinear fully connected layers. Finally, a scalar value, namely the true value of the thickness of the steel reinforcement protective layer, is output through a regression layer.

[0013] As a further technical solution of the present invention, S4 specifically includes: S41. Based on the real-time position coordinates of all scanning points and their corresponding true values ​​of the concrete cover thickness, a two-dimensional spatial distribution map of the concrete cover thickness in the detection area of ​​the component to be tested is generated by a spatial interpolation algorithm. S42. Perform image processing on the two-dimensional protective layer thickness spatial distribution map to identify thickness anomaly areas where the thickness value exceeds the design tolerance range, and calculate the geometric shape, texture, gradient direction and relative positional relationship with the reference edge of the component of the thickness anomaly area to form a descriptive anomaly spatial distribution pattern feature vector. S43. Calculate the similarity between the abnormal spatial distribution pattern feature vector and a predefined typical production defect feature library; the feature library stores standard pattern features corresponding to the positioning deviation of the rebar cage, the bulging of the mold side mold, and the uneven spraying of the release agent, respectively. S44. Based on the highest similarity score, determine the most likely production process defect type that caused the current thickness anomaly, and output a visual diagnostic report that includes the defect type, confidence level, and location of the anomaly area.

[0014] As a further technical solution of the present invention, S44 specifically includes: S441. Receive the highest similarity score. If the similarity score is higher than a preset first threshold, the corresponding defect type is directly determined as a production process defect type. If the similarity score is lower than the first threshold but higher than a preset second threshold, proceed to the alternative analysis process. S442. In the candidate analysis process, extract all candidate defect types with scores higher than the second threshold and their corresponding confidence levels, and combine the absolute position of the thickness anomaly area on the component and the structural type of the component to perform secondary logical reasoning to determine the most reasonable production process defect type from the candidate list. S443. Automatically integrate the finally determined production process defect type, its corresponding confidence level, the spatial coordinates of the thickness anomaly area, and key geometric evidence from the feature vector of the anomaly spatial distribution pattern to generate a structured visual diagnostic report.

[0015] A deep learning-based electromagnetic induction detection system for the thickness of the concrete cover of reinforcing bars in precast components is provided to implement a deep learning-based electromagnetic induction detection method for the thickness of the concrete cover of reinforcing bars in precast components, comprising: The baseline state synchronous acquisition module is used to control the electromagnetic sensor to scan the predetermined unreinforced area on the surface of the precast component under test, and to acquire the electromagnetic intrinsic response signal of the component in the current state; at the same time, it controls the infrared thermal imager to acquire the surface temperature distribution thermal image of the component detection area; and it aligns and correlates the electromagnetic intrinsic response signal and the surface temperature distribution thermal image with time stamp and spatial coordinates to construct multi-dimensional baseline state data. The dynamic compensation and fusion processing module is communicatively connected to the reference state synchronous acquisition module. It is used to control the electromagnetic sensor to scan the steel reinforcement area inside the component, and to acquire the original electromagnetic induction signal and the corresponding real-time position coordinates in real time. Based on the real-time position coordinates, it maps the local background temperature value from the surface temperature distribution thermal image, and performs temperature field compensation on the original electromagnetic induction signal according to a predetermined temperature-electromagnetic response relationship. At the same time, it retrieves the electromagnetic intrinsic response signal at the corresponding position from the multi-dimensional reference state data, performs background normalization processing on the temperature-compensated signal, and generates an enhanced fusion signal. The thickness intelligent calculation module is communicatively connected to the dynamic compensation and fusion processing module. It includes a pre-trained dedicated deep learning model for receiving the enhanced fusion signal and directly outputting the true value of the steel reinforcement protective layer thickness after eliminating the influence of environmental conditions. The production defect diagnosis module is communicatively connected to the thickness intelligent calculation module. It is used to generate a thickness spatial distribution map based on the true value of the thickness of the steel reinforcement protective layer and its corresponding position coordinates; extract the spatial distribution pattern features of abnormal areas in the thickness spatial distribution map; perform pattern matching between the features and a predefined typical production defect feature library; and perform reverse diagnosis based on the matching results and output the type of production process defect that caused the thickness abnormality.

[0016] The beneficial effects of this invention are as follows: 1. By synchronously acquiring multi-dimensional reference states, an online digital reference that simultaneously includes electromagnetic intrinsic response and surface temperature distribution is constructed. This reference drives subsequent real-time dynamic compensation, which in principle effectively removes complex interferences such as moisture, heat, and contamination from the components after demolding, enabling high-precision testing to be carried out in the most critical production stage.

[0017] 2. Employing an architecture that deeply integrates physical compensation and dedicated AI computation, the system first performs physical-level purification of deterministic interference to generate an enhanced fused signal. Then, a deep learning model trained specifically for non-standard states is used to overcome residual nonlinear interference. This results in the construction of a hybrid intelligent computation model with extreme robustness, which significantly improves adaptability and reliability in complex production environments.

[0018] 3. A reverse diagnosis step for production defects has been added. By analyzing the spatial distribution pattern of the thickness true value, it intelligently matches and traces the defect type back to the specific production stage. This achieves a value leap from quality inspection to process diagnosis, forming a closed loop of intelligent process fault diagnosis based on non-destructive testing, and elevating quality control from passive judgment to proactive process optimization guidance. Attached Figure Description

[0019] Figure 1This is a flowchart of an electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as proposed in Embodiment 1 of the present invention. Figure 2 This is an architecture diagram of the dedicated deep learning model in Embodiment 1 of the present invention; Figure 3 This is a framework diagram of an electromagnetic induction detection system for the thickness of the concrete cover of precast components based on deep learning, as proposed in Embodiment 2 of the present invention. Figure 4 This is a comparison diagram of the technical effects of the present invention and the prior art in Example 3. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1 Please see the appendix Figure 1 - Appendix Figure 2 A deep learning-based electromagnetic induction detection method for the thickness of the concrete cover of reinforcing bars in precast components includes: S1. Simultaneously acquire the electromagnetic intrinsic response signal and surface temperature distribution thermal image of the unreinforced area of ​​the component under test to form multi-dimensional reference state data; specifically including: S11. Location of Pre-determined Reinforced-Free Areas: Before inspecting the precast component to be tested, it is necessary to first determine specific areas on the component surface where no reinforcement bars are present, i.e., the pre-determined reinforcement-free areas. This area can be located using the following two methods: S111. Drawing Positioning Mode: Based on the design drawings of the precast component to be tested, the steel reinforcement cage layout information contained therein is analyzed. If the design drawings are computer-aided design models or building information model data, the spatial coordinate information of the steel reinforcement is directly extracted, and these three-dimensional spatial coordinates are mapped back to the outer surface of the component, thereby accurately locating at least one predetermined area without steel reinforcement. This area has been determined to be free of steel reinforcement cages according to structural design requirements during the component production stage; therefore, its electromagnetic response characteristics only reflect the intrinsic properties of the concrete matrix.

[0022] S112. Pre-scan Positioning Mode: If design drawings are missing or construction deviations exist, a low-sensitivity electromagnetic sensor is used to quickly pre-scan the component surface. This pre-scan covers the entire surface to be inspected with lower spatial resolution and higher scanning speed, acquiring the induced signal intensity at each location. Continuous connected areas with induced signal intensity below a set threshold are identified as predetermined areas without reinforcing steel. This determination is based on the fact that reinforcing steel, as a highly conductive material, will produce a significant enhancement of the electromagnetic induction signal, while the concrete matrix of areas without reinforcing steel will only produce a weak background electromagnetic response. Therefore, continuous areas with consistently low signal intensity can be considered as areas without reinforcing steel.

[0023] The two positioning modes mentioned above can be flexibly selected or combined according to the actual testing conditions to ensure that a reference area that can represent the electromagnetic intrinsic characteristics of the component under the current environmental conditions is accurately obtained before formal testing.

[0024] S12. After determining the predetermined unreinforced area, proceed to the stage of acquiring the electromagnetic intrinsic response signal: The electromagnetic sensor is controlled to scan the predetermined unreinforced area determined in the preceding steps at a preset detection frequency and gain. The scanning method can employ grid scanning or helical scanning to ensure coverage of all critical locations within the area. During the scanning process, the electromagnetic sensor emits an alternating electromagnetic field into the concrete matrix and receives the response signal modulated by the electromagnetic properties of the concrete material.

[0025] The electromagnetic intrinsic response signal is represented in complex domain form, containing information in two dimensions: amplitude and phase. The amplitude reflects the degree of absorption and scattering of electromagnetic energy by the concrete matrix, while the phase reflects the phase lag characteristic of the electromagnetic field propagating in the concrete medium. This complex representation can comprehensively characterize the electromagnetic properties of the concrete matrix under current environmental temperature and moisture content conditions, including the combined effects of conductivity and dielectric constant.

[0026] The electromagnetic intrinsic response signal obtained through this step constitutes the basic reference data for subsequent background normalization processing. It only reflects the electromagnetic properties of the concrete matrix itself and is not affected by the presence of internal reinforcing steel. Therefore, it can be used as a benchmark for distinguishing between reinforcing steel signals and background noise.

[0027] S13. At the same time that the electromagnetic sensor scans the predetermined area without reinforcing steel, the infrared thermal imager is simultaneously activated to collect temperature data: The infrared thermal imager continuously captures images of the entire inspection area, including the aforementioned pre-defined area without rebar. This inspection area covers the entire range where the rebar cover thickness needs to be measured later, ensuring complete coverage of temperature distribution data. The infrared thermal imager continuously acquires images at a certain frame rate, forming a thermal image sequence of the surface temperature distribution.

[0028] Each frame of surface temperature distribution thermal image is presented in the form of a two-dimensional temperature matrix, where each element corresponds to the surface temperature measurement value at a specific spatial location within the detection area. By continuously capturing images, the temperature field distribution state of the detection area at different times is obtained, forming a time-series temperature data record.

[0029] This synchronous acquisition process ensures the temporal correspondence between the electromagnetic intrinsic response signal and the surface temperature distribution thermal image, providing the original data foundation for subsequent steps to spatiotemporally align and correlate temperature information with electromagnetic response information. The infrared thermal imager's imaging range is greater than or equal to the predetermined unreinforced area to ensure that even if the electromagnetic sensor's scanning position shifts or extends to detect reinforced areas, the temperature information at the corresponding location can still be mapped from the thermal image data.

[0030] S14. Align and correlate the electromagnetic intrinsic response signal obtained in the preceding steps with the surface temperature distribution thermal image according to the timestamp and spatial coordinates to construct a data set indexed by spatial location, which simultaneously contains the intrinsic electromagnetic response at the location point and the surface temperature scalar, constituting multidimensional reference state data; specifically: S141. Timestamp Marking and Synchronization: A unified, high-precision timestamp is applied to every electromagnetic intrinsic response signal acquired by the electromagnetic sensor while scanning a predetermined area without reinforcing steel, and to every frame of surface temperature distribution thermal image captured by the infrared thermal imager. This timestamp is generated by a high-precision clock source to ensure that the synchronization accuracy between the electromagnetic sensor and the infrared thermal imager in the time dimension meets the requirements of subsequent correlation analysis.

[0031] S142. Multi-sensor coordinate system calibration: Using a calibration plate or known spatial markers, establish the transformation relationship between the scanning coordinate system of the electromagnetic sensor and the imaging coordinate system of the infrared thermal imager. Through rigid body transformation, transform the coordinate systems of the two sensors to the same world coordinate system based on the prefabricated component under test, so that the same physical spatial location has a unified spatial coordinate representation in the data from the two sensors.

[0032] S143. Spatiotemporal Alignment and Temperature Extraction: For each electromagnetic intrinsic response signal marked with a timestamp, find surface temperature distribution thermal images of the same moment or its temporal neighborhood based on the timestamp. According to the spatial coordinates of the electromagnetic signal in the world coordinate system, extract the precise temperature value of that coordinate point from the corresponding thermal image frame through interpolation. This process achieves precise matching of electromagnetic response data and temperature data in both the temporal and spatial dimensions.

[0033] S144. Structured Data Unit Encapsulation: Each set of spatiotemporally aligned electromagnetic intrinsic response signals, its corresponding spatial coordinates, and the precise temperature value extracted from the thermal image are encapsulated into a structured data unit. This data unit is indexed by its spatial location and contains the complex value of the intrinsic electromagnetic response and the scalar value of the surface temperature at that location, as well as a unified timestamp. The collection of all data units constitutes the multidimensional reference state data, providing a priori reference dataset for subsequent real-time dynamic compensation and fusion steps.

[0034] S2. When scanning the rebar area, the original electromagnetic induction signal is subjected to temperature field compensation and background normalization processing based on multi-dimensional reference state data to generate an enhanced fused signal; specifically including: S21. Real-time acquisition of raw electromagnetic induction signals and position coordinates, specifically: S211. When scanning and detecting the steel reinforcement area inside the precast component to be tested, the complex form of the original electromagnetic induction signal output by the electromagnetic sensor is acquired in real time: in: For a moment The original electromagnetic induction signal (complex form) was collected. The original signal amplitude, The original signal phase, These are the in-phase and quadrature components of the original signal, respectively. It is the imaginary unit.

[0035] S212. And by using a positioning device rigidly coupled to the sensor, such as a laser tracker, a vision positioning system, or a robotic arm with encoder feedback, the real-time position coordinates of the signal acquisition point are recorded simultaneously, establishing a precise correspondence between electromagnetic response data and spatial position: in: The real-time three-dimensional position coordinates of the signal acquisition point. This is the transpose of the matrix. For the position tracking function of the positioning device, These are the coordinate components of the real-time position in the world coordinate system.

[0036] S22. Temperature field compensation, specifically: S221. Based on the real-time location coordinates, the multi-dimensional reference state data obtained in step S1... The included surface temperature distribution thermal image Spatial mapping is performed to extract the local background temperature value corresponding to the coordinate point: in: This refers to the local background temperature value extracted from the thermal image; SpatialMap This is a spatial mapping function that extracts temperature from baseline data based on real-time location. This is the spatial interpolation function for the temperature field.

[0037] S222. Based on the pre-stored temperature-electromagnetic response lookup table, the amplitude and phase of the original electromagnetic induction signal are linearly corrected using the local background temperature value to eliminate electromagnetic response drift caused by temperature changes, thus obtaining a temperature-compensated signal: in: This is the temperature compensation amplitude gain coefficient; This is a temperature amplitude gain lookup table, representing the influence coefficient of temperature on the signal amplitude. This is the phase shift amount to compensate for temperature. This is a temperature-phase shift lookup table, which characterizes the amount of temperature affecting the signal phase. For temperature compensation signal (complex number form); is the base of the natural logarithm; This is the signal amplitude after temperature compensation; This represents the signal phase after temperature compensation.

[0038] S23. Background Normalization Processing: The temperature compensation signal and the electromagnetic intrinsic response signal are compared point-by-point in the complex domain to eliminate the influence of the background electromagnetic properties of the concrete matrix and highlight the abnormal response caused by the reinforcing steel, thus obtaining the background normalized signal; specifically including: S231. Neighboring Reference Data Retrieval: Based on real-time position coordinates, retrieve at least one electromagnetic eigenresponse signal whose spatial coordinates are closest from the multidimensional reference state data, and use it as a reference reference for background normalization; specifically: Centered on the real-time location coordinates, a proximity search radius is set, and all reference data units that meet the distance conditions are retrieved within this radius. These reference data units are derived from the multi-dimensional reference state data constructed in step S1, and each unit contains spatial coordinates, electromagnetic intrinsic response signals, and corresponding local background temperature values.

[0039] If multiple neighboring reference data units exist within the search range, the electromagnetic eigenresponse signals of these units are arithmetically averaged to obtain a comprehensive background reference signal. This averaging process effectively reduces the influence of measurement noise from a single reference point and improves the stability of background normalization. If only one neighboring reference data unit exists within the search range, the electromagnetic eigenresponse signal of that unit is directly used as the background reference signal.

[0040] This retrieval process ensures the spatial proximity of the reference signal used for background normalization, thereby guaranteeing the consistency of the background characteristics of the concrete matrix and enabling subsequent ratio calculations to effectively eliminate interference caused by local matrix differences.

[0041] S232. Represent the temperature compensation signal in complex form, and also represent the retrieved electromagnetic intrinsic response signal in complex form: in: The retrieved electromagnetic intrinsic response signal, The amplitude of the background reference signal. The background reference signal phase.

[0042] Performing complex division yields the initial normalized complex signal: in: The initial normalized complex signal; The amplitude ratio; This represents the phase difference.

[0043] S233. The amplitude component of the initial normalized complex signal. Take the common logarithm and its phase components Phase unwinding is performed to obtain stable amplitude-phase pairs; S233. Logarithmic conversion of amplitude and phase unwinding, specifically: amplitude components of the initial normalized complex signal Take the common logarithm: ,in: This is the logarithmic magnitude ratio.

[0044] Phase components of the initial normalized complex signal Phase unwinding is performed to eliminate phase jumps and obtain stable amplitude-phase pairs. in: To represent the continuous phase difference after unwinding; Unwrap For phase unwinding function; The integer value is the phase period correction value. Represents the integer variable to be optimized; This is the unwinding phase value of the previous sampling point, used to ensure phase continuity.

[0045] S234. Background Normalized Signal Encapsulation: The processed stable amplitude-phase pairs are re-encapsulated into a new complex signal, which is the background normalized signal. in: The background normalized signal (complex form).

[0046] S24. Enhanced Fusion Signal Generation: The background normalized signal undergoes normalization processing including amplitude logarithmic transformation and phase dewinding to generate an enhanced fusion signal. This signal characterizes the equivalent electromagnetic response under standard dry and cooled conditions, determined solely by the distance between the reinforcing steel and the sensor, eliminating the influence of temperature, humidity, and concrete matrix inhomogeneity. Alternatively, the real vector form can be retained as input to the deep learning model: in: The feature vector (two-dimensional real vector) of the enhanced fused signal; These are the first component (logarithmic magnitude ratio) and the second component (unwinding phase difference) of the eigenvector, respectively. For the complex representation of the enhanced fused signal; The feature vectors are input to the deep learning model dedicated to step S3.

[0047] S3. Input the enhanced fused signal into a dedicated deep learning model to output the true value of the rebar cover thickness; specifically including: S31. Two-dimensional data sequence construction: The enhanced fusion signal generated in the preceding steps, along with its corresponding real-time position coordinates, is organized according to the temporal sequence and spatial distribution of the electromagnetic sensors scanning the component surface to form a two-dimensional data sequence. Specifically: This data sequence is structured with scanning time sequence as the vertical dimension and signal features and location information as the horizontal dimension. Each element in the sequence contains the three-dimensional coordinates of a specific spatial location and the corresponding enhanced fused signal features, which include two components: logarithmic amplitude ratio and unwinding phase difference.

[0048] This serialization process transforms the originally discrete point cloud measurement data into a structured data sequence with spatiotemporal continuity. It preserves the physical continuity of the electromagnetic response signal as it changes with spatial location and provides a data foundation for subsequent sliding window extraction of local contextual information. This two-dimensional data sequence simultaneously encodes the spatial distribution pattern of signal features and the geometric topological relationship of the scanning path, enabling deep learning models to perceive the continuous variation trend of the rebar cover thickness on the component surface.

[0049] S32. Local Enhancement Fusion Signal Block Extraction: Extracting a local enhancement fusion signal block containing the current point to be solved and its surrounding neighborhood signals from the aforementioned constructed two-dimensional data sequence using a sliding window method. Specifically: Using the current center point as a reference, a fixed-length window is defined, and all spatially adjacent sampling points are retrieved within this window. These adjacent sampling points, together with the center point, constitute a local subset of data. For each included sampling point, its three-dimensional spatial offset relative to the center point, as well as the enhanced fused signal features of that sampling point, including the logarithmic magnitude ratio and the unwinding phase difference, are extracted.

[0050] This locally enhanced fused signal block simultaneously contains signal amplitude, phase, and relative positional context information of the center point and its neighborhood, forming a set of local features with spatial structure awareness. Through a sliding window mechanism, this extraction process traverses the entire two-dimensional data sequence, generating a corresponding local data block for each point to be solved, serving as the basic input unit for a dedicated deep learning model. This local context extraction method enables the model to comprehensively utilize the electromagnetic response features of the surrounding area, improving its ability to identify local anomalies and thickness variations.

[0051] S33. Dedicated Deep Learning Model Inference: The local enhanced fused signal block is input into a pre-trained dedicated deep learning model, which directly outputs a scalar value corresponding to the center point of the input signal block. This scalar value represents the true thickness of the rebar cover after eliminating the effects of temperature, humidity, and surface contaminants. This dedicated deep learning model employs a deep convolutional neural network architecture and is trained end-to-end on a mixed simulation and measurement dataset collected under various temperature, humidity, and surface contamination conditions. It is used to establish a complex nonlinear mapping relationship from the local signal block to the cover thickness. Specifically: S331. Layer Normalization Processing: Layer normalization processing is performed on the input local enhancement fusion signal block. Specifically, the mean and standard deviation of the signal block along the window length dimension are calculated, and the original data is normalized using learnable scaling factors and offset vectors to ensure that the data distribution of its amplitude and phase components meets the input requirements of a dedicated deep learning model. This processing can eliminate numerical scale differences under different scanning batches or environmental conditions, accelerate model convergence, and improve inference stability.

[0052] in: This refers to the locally enhanced fused signal block after layer normalization; This is the mean vector along the window length dimension; The standard deviation vector along the window length dimension; For learnable scaling factors and offset vectors; This is a numerical stability constant to prevent division by zero; This is element-wise multiplication (Hadamard product). This represents the total number of sampling points contained within the window, i.e., the window length. For the local enhancement fusion signal block matrix, the first The elements of all columns in a row.

[0053] S332. Primary Spatial Texture Feature Extraction: Primary spatial texture features are extracted through the first layer of a dedicated deep learning model. This layer employs a one-dimensional depthwise separable convolutional architecture. First, depthwise convolution is used to independently filter each feature channel of the input in the spatial dimension, preserving the spatial features of each channel. Then, the outputs of each channel are concatenated, and then linear projection is performed in the channel dimension through pointwise convolution to achieve cross-channel information fusion. The features extracted in this process are related to the spacing and orientation of the rebars, forming a primary spatial texture feature map.

[0054] in: The intermediate feature map is a depthwise convolution; DepthwiseConv1D (depthwise convolution) is a one-dimensional convolution operation performed independently on each channel of the input feature map. It is a set of channel-wise convolution kernels for depthwise separable convolutions. Input the number of channels. is the kernel length; Concat (channel concatenation) combines the output features of each channel generated by depthwise convolution according to the channel dimension to form an intermediate feature map; The first local enhancement fusion signal block matrix after layer normalization Elements of all rows in the column; For the first Each input channel corresponds to a convolution kernel; * indicates a one-dimensional convolution operation. This is a primary spatial texture feature map. The length of the feature sequence. The number of output channels; PointwiseConv1D (pointwise convolution) uses a 1×1 convolution kernel to perform linear projection on the channel dimension, and performs cross-channel fusion and dimensional transformation on the multi-channel features output by depthwise convolution; The pointwise convolution weight matrix; This is the pointwise convolution bias vector.

[0055] S333. Parallel Dual-Branch Feature Extraction: The primary spatial texture features are input into the second layer of a dedicated deep learning model. This layer employs a parallel dual-branch structure for multi-scale feature extraction. The first branch is a dilated convolution branch, which uses one-dimensional dilated convolution with an increasing dilation rate to progressively expand the receptive field, capturing large-scale thickness gradient patterns caused by mold side bulging or uneven application of release agent, and outputting a global contextual feature map. The second branch is an attention branch, which uses a self-attention mechanism to calculate the correlation weights between different locations, focusing on local high-gradient abrupt changes in features caused by rebar cage positioning deviations, and outputting a locally focused feature map. The two branches are processed in parallel, and are specifically optimized for different types of production defect features.

[0056] First branch (dilated convolution branch): Second branch (attention branch): in: For the first Output feature map of layered dilated convolution; For nonlinear activation functions, such as ReLU or GELU; DilatedConv1D The void ratio is One-dimensional dilated convolution operation; For the first The kernel weight matrix of dilated convolution; For the first The void ratio of a layer is usually calculated as follows: Increasing; For the first The bias vector of the dilated convolution layer; This represents the total number of dilated convolutional branches; This is a global context feature map that captures large-scale gradient patterns. This represents the number of feature channels in the final output of the dilated convolution branch. This refers to the query, key, and value projection matrix in the attention mechanism. For query, key, and value feature matrix; For attention head dimension; This is the attention weight matrix; This is a locally focused feature map that highlights regions with high gradient abrupt changes.

[0057] S334. Feature Fusion and Thickness Regression: The global context feature maps and local focused feature maps output from the two branches are concatenated to form a fused feature vector. Subsequently, high-dimensional feature fusion and layer-by-layer dimensionality reduction are performed through multiple nonlinear fully connected layers, with each layer using an activation function to introduce nonlinear expressive power. Finally, a linear regression layer outputs a scalar value, which is the true value of the rebar cover thickness after eliminating the influence of temperature, humidity, and surface contaminants.

[0058] in: The fused feature vector after channel concatenation; Flatten ( This involves flattening a two-dimensional feature map into a one-dimensional vector. This represents the total dimension of the concatenated features. For the first The output feature vector of a fully connected layer Representing the The number of neurons in a fully connected layer (feature dimension); For the first Layer fully connected weight matrix; For the first Fully connected layer bias vector; This represents the total number of fully connected layers. This is a true estimate of the final output thickness of the concrete cover for reinforcing bars (in millimeters). This represents the regression layer weight vector; This is the bias term for the regression layer.

[0059] S4. Analyze the spatial distribution pattern of the true value of the concrete cover thickness, match it with a predefined defect feature library, perform reverse diagnosis, and output the defect types in the production process; specifically including: S41. Generation of Two-Dimensional Protective Layer Thickness Spatial Distribution Map: Based on the real-time position coordinates of all scanning points and their corresponding true values ​​of the protective layer thickness, a two-dimensional protective layer thickness spatial distribution map of the detection area of ​​the component under test is generated through a spatial interpolation algorithm. This distribution map uses the component surface as the domain and the protective layer thickness as the value domain, forming a continuous thickness field visualization representation.

[0060] in: Spatial location Interpolated thickness value at; SpatialInterp For spatial interpolation functions, methods such as inverse distance weighting, kriging, or radial basis functions can be used; For the first The three-dimensional spatial coordinate vector of each scanning point contains the three components of x, y, and z of the sampling point in the world coordinate system, which are used to identify the actual detection position of the electromagnetic sensor on the surface of the component. This represents the total number of sampling points in the rebar area scan. The detection area spatial domain of the precast component to be tested; A two-dimensional spatial distribution diagram of protective layer thickness (in matrix form); The grid height and width (in pixels) of the distribution map; These are the grid pixel coordinates of the distribution map; A function that maps pixel coordinates to world coordinates.

[0061] S42. Thickness Anomaly Region Identification and Feature Extraction: The generated two-dimensional protective layer thickness spatial distribution map is processed to identify thickness anomaly regions where the thickness value exceeds the design tolerance range. The geometry, texture, gradient direction, and relative positional relationship with the reference edge of the component are calculated for each anomaly region to form a descriptive anomaly spatial distribution pattern feature vector.

[0062] S421. Abnormal Region Segmentation: Compare the thickness value of each pixel in the thickness distribution map with the nominal thickness of the designed protective layer and calculate the absolute deviation. Mark pixels with deviations exceeding the design tolerance as abnormal points. Through connected component analysis, aggregate spatially adjacent abnormal points into independent thickness abnormal regions and generate a binary mask for the abnormal regions.

[0063] S422. Geometric Shape Feature Calculation: For each identified thickness anomaly region, calculate its area and perimeter to characterize the spatial scale of the region. Further calculate the centroid coordinates of the region to determine its spatial centroid location. By constructing the covariance matrix and solving for the eigenvalues, obtain the principal axis direction and extent of extension of the region. Calculate the eccentricity to quantify how close the region's shape is to an ellipse; a higher eccentricity indicates a more elongated shape.

[0064] S423. Texture and Gradient Feature Calculation: Calculate the spatial gradient field of the thickness distribution map to obtain the thickness change rate and gradient direction angle at each location. For each anomalous region, calculate the mean and standard deviation of its internal gradient direction angle to characterize the consistency of the thickness anomaly boundary. Simultaneously, calculate the gray-level co-occurrence matrix of the anomalous region to extract homogeneity and contrast texture features, reflecting the smoothness and variability of the thickness distribution within the region, respectively.

[0065] S424. Calculation of relative position features: Calculate the offset vector of the centroid of each abnormal region relative to the reference point of the component, as well as the vertical distance from the centroid to the nearest reference edge of the component, to characterize the absolute and relative position of the abnormal region in the overall structure of the component.

[0066] S425. Feature Vector Encapsulation: The area, perimeter, eccentricity, texture homogeneity, texture contrast, gradient direction mean, gradient direction standard deviation, offset component relative to the reference point, and distance to the nearest reference edge obtained above are combined in a fixed order to form a multi-dimensional abnormal spatial distribution pattern feature vector, which is used for subsequent pattern matching with the typical production defect feature library.

[0067] S43. Similarity Calculation and Pattern Matching: The similarity calculation is performed between the feature vectors of abnormal spatial distribution patterns and a predefined feature library of typical production defects. This feature library stores standard pattern features corresponding to rebar cage positioning deviation, mold side mold bulging, and uneven application of release agent, respectively. Each defect type is characterized by the mean and covariance of feature vectors from multiple typical samples.

[0068] Feature library structure: in: A database of typical production defect features; For the first The standard pattern feature mean vector of each defect type; For the first The standard pattern feature covariance matrix of the defect type; For the first The name identifier of the defect type, Reinforcing cage positioning deviation Mold side mold expansion, Uneven application of release agent; Mahalanobis distance similarity calculation: in: For the first The first abnormal region and the first Mahalanobis distance for each type of defect; The feature vector of the abnormal spatial distribution pattern of the nth abnormal region; The normalized similarity score has a value range of [value range missing]. The closer the similarity is to 1, the higher the similarity.

[0069] S44. Based on the highest similarity score, determine the most likely production defect type causing the current thickness anomaly, and output a visual diagnostic report including the defect type, confidence level, and location of the anomaly area; specifically: S441. Similarity Threshold Determination: Receive the highest similarity score calculated in the previous steps, compare it with a preset threshold, and determine the subsequent determination process path.

[0070] If the highest similarity score is higher than the preset first threshold (high confidence threshold), it is determined to be in the direct determination mode. The defect type with the highest similarity score is directly determined as the production process defect type corresponding to the current thickness anomaly area, without further analysis.

[0071] If the highest similarity score is lower than the first threshold but higher than the preset second threshold (low confidence threshold), it is determined to be an alternative analysis mode and enters the secondary logical reasoning process.

[0072] If the highest similarity score is lower than or equal to the second threshold, it is determined to be an undetermined pattern. The system marks the abnormal area as an unidentified type and prompts that manual review or supplementary detection data is required.

[0073] S442. Alternative Analysis Process: In the alternative analysis mode, all defect types with similarity scores higher than the second threshold are first extracted to form a list of alternative defect types. At the same time, the specific similarity scores corresponding to each alternative type are retained as prior confidence.

[0074] Subsequently, combining the absolute position information of the thickness anomaly region on the component (such as centroid coordinates), the geometric scale characteristics of the anomaly region, and the structural type of the precast component to be tested (such as beams, slabs, columns, etc.), a secondary logical reasoning mechanism based on expert rules is invoked. This mechanism performs weighted evaluation based on the prior correlation between different defect types and component location and structural form. For example, mold bulging defects on the side mold have a higher probability of occurring in the edge area of ​​the component, rebar cage positioning deviations are more common in the central area, and uneven spraying of release agent usually exhibits a large-area diffuse distribution characteristic.

[0075] By combining prior confidence and location-structure correlation, the most reasonable type of production process defect is determined from the candidate list, and the final judgment result is output.

[0076] S443. Visualized Diagnostic Report Generation: The final determined production defect type, its corresponding similarity confidence level, the spatial coordinates of the thickness anomaly area (including the centroid and boundary range), and the key geometric evidence extracted from the feature vector of the abnormal spatial distribution pattern (including discriminative features such as area, eccentricity, distance to the reference edge, and average gradient direction) are automatically integrated into a structured diagnostic report.

[0077] Simultaneously, a visual overlay rendering is generated, presenting the spatial distribution of the 2D protective layer thickness as a heatmap. The map also marks the boundary contours, centroid locations, and defect type labels for each abnormal area, creating an intuitive visualization of the spatial distribution and diagnostic results. This report supports export to a standardized document format for quality management personnel to use in production defect tracing and process improvement decisions.

[0078] Example 2 Please see the appendix Figure 3 A deep learning-based electromagnetic induction detection system for the thickness of the concrete cover of precast steel bars is provided to implement a deep learning-based electromagnetic induction detection method for the thickness of the concrete cover of precast steel bars, comprising: A. A baseline state synchronous acquisition module is used to control an electromagnetic sensor to scan a predetermined unreinforced area on the surface of the precast component under test, acquiring the electromagnetic intrinsic response signal of the component in its current state; simultaneously, it controls an infrared thermal imager to acquire a thermal image of the surface temperature distribution in the component's detection area; and it correlates the electromagnetic intrinsic response signal and the surface temperature distribution thermal image with timestamps and spatial coordinates to construct multi-dimensional baseline state data; specifically including: The unreinforced area positioning unit is used to determine the predetermined unreinforced area based on design drawings or through low-sensitivity pre-scanning. The signal and thermal image synchronous acquisition unit is used to control the electromagnetic sensor and the infrared thermal imager to perform spatiotemporally synchronized data acquisition. The spatiotemporal alignment and fusion unit is used to establish a unified world coordinate system and to perform coordinate mapping and data association between the collected electromagnetic intrinsic response signals and surface temperature distribution thermal images, encapsulating them into multidimensional reference state data.

[0079] B. The dynamic compensation and fusion processing module, which communicates with the reference state synchronous acquisition module, is used to control the electromagnetic sensor to scan the steel reinforcement area inside the component, and to acquire the original electromagnetic induction signal and the corresponding real-time position coordinates in real time. Based on the real-time position coordinates, it maps the local background temperature value from the surface temperature distribution thermal image, and performs temperature field compensation on the original electromagnetic induction signal according to a predetermined temperature-electromagnetic response relationship. Simultaneously, it retrieves the electromagnetic intrinsic response signal at the corresponding position from the multi-dimensional reference state data, performs background normalization processing on the temperature-compensated signal, and generates an enhanced fusion signal. Specifically, this includes: The temperature compensation unit has a pre-stored temperature-electromagnetic response lookup table, which is used to correct the amplitude and phase of the original electromagnetic induction signal based on the local background temperature value. The background normalization unit is used to perform a ratio operation between the temperature-compensated signal and the corresponding electromagnetic intrinsic response signal in the complex domain, and to take the logarithm of the amplitude and dewind the phase of the operation result to generate an enhanced fused signal.

[0080] C. A thickness intelligent calculation module, communicatively connected to the dynamic compensation and fusion processing module, includes a pre-trained dedicated deep learning model for receiving enhanced fusion signals and directly outputting the true value of the rebar cover thickness after eliminating the influence of environmental conditions; specifically including: The data preprocessing unit is used to receive the enhanced fusion signal and the corresponding real-time position coordinates from the dynamic compensation and fusion processing module, organize them into a two-dimensional data sequence arranged according to the scanning time sequence and spatial position, and then extract the local enhanced fusion signal block in the form of a sliding window. The deep feature extraction unit, connected to the data preprocessing unit, includes the front-end network structure of a dedicated deep learning model. This unit is used to perform layer normalization on the input local enhancement fusion signal block and extract its primary spatial texture features using a one-dimensional depthwise separable convolutional layer. The robustness enhancement and fusion unit is connected to the deep feature extraction unit, which includes a back-end parallel dual-branch structure of a dedicated deep learning model. The first branch of this unit processes primary spatial texture features through dilated convolution to capture large-scale thickness gradient patterns. The second branch processes primary spatial texture features through an attention mechanism to focus on local high-gradient abrupt changes. The output features of the two branches are then concatenated and fused. The regression output unit, connected to the robustness enhancement and fusion unit, includes a fully connected layer and a regression layer at the end of a dedicated deep learning model. This unit is used to perform nonlinear transformation and dimensionality reduction on the fused high-dimensional features and finally output a scalar representing the true value of the steel reinforcement cover thickness.

[0081] D. Production Defect Diagnosis Module, which communicates with the thickness intelligent calculation module, is used to generate a thickness spatial distribution map based on the true value of the rebar protective layer thickness and its corresponding location coordinates; extract the spatial distribution pattern features of abnormal areas in the thickness spatial distribution map; perform pattern matching between the features and a predefined typical production defect feature library; and, based on the matching results, perform reverse diagnosis and output the type of production process defect that caused the thickness anomaly; specifically including: The pattern extraction unit is used to perform image analysis on the thickness spatial distribution map and extract the geometric and texture features of abnormal regions to form feature vectors. The intelligent matching unit has a built-in library of typical production defect features, which is used to match feature vectors with standard patterns in the feature library; The diagnostic report generation unit is used to generate a visual diagnostic report based on the matching results, which includes the defect type, location, and confidence level.

[0082] E. A closed-loop feedback control module, communicatively connected to the production defect diagnosis module, is used to receive the diagnosed defect type and location information in the production process and automatically send it to the corresponding production unit control system to guide process adjustments or equipment maintenance; specifically including: The defect data receiving unit is used to receive the specific production process defect type, corresponding spatial location information, and confidence level output from the production defect diagnosis module. The control command mapping unit, connected to the defect data receiving unit, contains a pre-built mapping rule library of defect types and production unit control commands. This unit is used to automatically generate standardized control commands to be sent to the corresponding production unit control system based on the defect type in the production process. For example, if the defect type is a rebar cage positioning deviation, a positioning coordinate calibration command is generated and sent to the rebar binding robot or positioning fixture PLC; if it is a mold side mold bulging, a pressure detection and fastener adjustment command is generated and sent to the mold maintenance system; if it is an uneven release agent spraying, a flow rate, air pressure, or path replanning command is generated and sent to the spraying machine. The instruction communication and execution feedback unit, connected to the control instruction mapping unit, is used to send standardized control instructions to the target production unit control system in real time via industrial communication protocols, and to listen for and receive execution status feedback signals from the production unit to form a complete control closed loop.

[0083] Example 3 1. Experimental Objective The accuracy and robustness of the method described in this invention in detecting the thickness of the concrete cover of steel bars under the non-standard conditions of "wet, hot, and dirty" after demolding of precast components, as well as its effectiveness in diagnosing production defects, were verified and compared with existing mainstream detection technologies.

[0084] 2. Experimental Setup Experimental subjects: Six standard reinforced concrete test slabs (1.5m × 1.5m) were fabricated, with a designed concrete protective layer thickness of 20mm for the reinforcing mesh. The following defects were intentionally introduced through process control: Test plates A and B: Overall longitudinal offset of the reinforcing cage +5mm (simulated positioning deviation).

[0085] Test plates C and D: The mold on one side is slightly tilted, causing the protective layer to change in a gradient (simulating mold bulging).

[0086] Test panels E and F: The release agent (oil-based) was sprayed unevenly, forming localized thick film areas.

[0087] Simulate non-standard conditions: Place the test plate in an environmental chamber, heat it to 60°C to simulate the residual temperature of demolding, and spray water to increase the surface moisture content, keeping the release agent contaminated.

[0088] Comparison method: The method of this invention involves fully implementing steps S1-S4.

[0089] Existing technology 1 (traditional single-probe electromagnetic method): A commercially available model of rebar scanner is used. After the test plate has cooled to room temperature and the surface is dry and cleaned, calibration and testing are performed according to the instruction manual.

[0090] Existing technology 2 (array scanning method with temperature compensation): using a sensor array and only performing simple compensation for ambient temperature, and then performing detection under the same non-standard conditions.

[0091] Evaluation indicators: Detection accuracy, mean absolute error (MAE, unit: mm); robustness, accuracy rate of thickness qualification (design value ±3 mm) under non-standard conditions; diagnostic capability, accuracy rate of identifying preset production defect types (only applicable to the method of this invention).

[0092] 3. Experimental Results and Comparison Six test plates were tested under simulated non-standard conditions, and their key indicators are compared in Table 1 below. Figure 4 As shown: Table 1: Comparison of Results 4. Results Analysis Accuracy and Robustness: The mean absolute error (MAE) of the method of this invention is only 1.2 mm, significantly lower than the two existing technologies. This directly verifies the effectiveness of S1 (multi-dimensional reference state acquisition) and S2 (dynamic compensation and fusion), indicating that this invention can fundamentally eliminate interference from changes in temperature, humidity, and background electromagnetic properties. Especially under non-standard conditions, the accuracy rate of the present invention in determining pass / fail is as high as 98.3%, while the traditional method, due to its inability to handle interference, has an accuracy rate of only 32.1%, essentially rendering it ineffective.

[0093] Defect diagnosis capability: The defect type diagnosis accuracy of this invention reaches 95.0%, successfully tracing thickness anomaly patterns (such as overall offset and gradient changes) back to specific production process problems such as "rebar cage positioning deviation" and "mold side mold bulging". This fully demonstrates the practical value of the S4 (reverse diagnosis of production defects) step, realizing a leap from "measurement" to "diagnosis", a function that existing technologies do not possess at all.

[0094] Overall Performance: This invention maintains high accuracy while exhibiting extremely high environmental stability. This is thanks to the S3 (dedicated deep learning model) learning of the "enhanced fusion signal," which makes it highly robust to residual, nonlinear disturbances (such as heterogeneous release agents).

[0095] 5. Conclusion This verification example, through data comparison, strongly demonstrates that: In the complex and non-standard state after demolding of precast components, the present invention improves the detection accuracy (MAE) by more than 74.5% and the pass / fail judgment accuracy by more than 50% compared with the existing mainstream technology, thus completely solving the problem of detection feasibility in this scenario.

[0096] The invention features a unique reverse diagnostic function for production defects with a diagnostic accuracy rate of 95%, enabling in-depth value mining of quality inspection results.

[0097] In summary, this invention not only has significantly better technical effects than existing technologies, but also provides a completely new dimension for process quality monitoring, demonstrating outstanding creativity and industrial applicability.

[0098] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: This invention establishes an online electromagnetic-thermal multidimensional benchmark construction mechanism, overcoming the industry challenge of detecting failures under non-standard conditions. Existing technologies completely avoid complex conditions such as residual release agent on the surface of components after demolding, internal temperature gradients, and high moisture content, making detection impossible at this critical production stage. This invention proposes and implements a multidimensional benchmark state synchronous acquisition step, simultaneously acquiring electromagnetic intrinsic response signals and surface temperature distribution thermal images, and constructing spatiotemporally aligned multidimensional benchmark state data. This mechanism provides the detection system with a real-time, accurate digital twin of the current component state for the first time, enabling it to proactively perceive and quantify benchmark values ​​of environmental interference. Based on this, step S2 can perform precise temperature field compensation and background normalization, removing major physical interferences at the source, allowing subsequent analysis to be conducted under virtual standard conditions. Simulation experiments show that this mechanism improves the detection accuracy (MAE) by more than 85% compared to traditional methods under simulated demolding conditions, and the pass / fail judgment accuracy jumps from less than 33% to over 98%, achieving a fundamental transformation from unusable to high-precision in this scenario.

[0099] A deeply integrated model combining physical compensation and dedicated AI computation is constructed, achieving extreme robustness against complex nonlinear disturbances. Traditional methods, relying on a single physical model or general AI algorithm, struggle to handle nonlinear signal distortions caused by the coupling of moisture, heat, and contamination. This invention employs a unique progressive processing flow of steps S2 and S3. Step S2 first utilizes multidimensional benchmark data for physical-level linear compensation, outputting an enhanced fused signal that resolves modelable deterministic disturbances. Step S3 then inputs this signal into a dedicated deep learning model trained specifically for non-standard states. This model, through its internal parallel dilated convolution and attention mechanism branches (step S33), specifically learns and overcomes residual nonlinear disturbances that are difficult to physically model, such as uneven film thickness due to release agents. This hybrid model architecture, combining physical purification followed by AI-driven problem-solving, possesses both interpretability and powerful nonlinear fitting capabilities. This makes the invention unprecedentedly adaptable to various random and complex disturbance states in production environments, exhibiting significantly better robustness than existing techniques using only simple temperature compensation (error reduction of approximately 74.5%).

[0100] This invention represents a significant leap from quality inspection to process diagnosis, establishing the first intelligent traceability closed loop for production defects based on non-destructive testing (NDT). Existing testing technologies stop at simply determining whether the thickness is acceptable, failing to address the core production question of why defects occur. This invention innovatively adds a reverse diagnosis step for production defects. This step deeply mines the spatial distribution information contained in the high-fidelity true value of the rebar cover thickness output in step S3. Through intelligent pattern matching, it compares abnormal patterns (such as overall offset, gradient changes, and speckle fluctuations) with a predefined library of typical production defect features, thereby reverse-diagnosing specific production defect types (such as rebar cage positioning deviation and mold bulging). This automatically links the end-point data of NDT with specific fault points in the front-end process, forming a complete data link of perception, analysis, and diagnosis. This function enables the production line to promptly locate the root cause of problems, shifting quality control from passive post-inspection to proactive online process monitoring and real-time adjustment. It provides a core decision support tool for intelligent manufacturing of precast components, achieving a 95% accuracy rate in defect type diagnosis during simulation.

[0101] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0102] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for electromagnetic induction detection of the concrete cover thickness of precast structural steel bars in precast components based on deep learning, characterized in that, include: Simultaneously acquire electromagnetic intrinsic response signals and surface temperature distribution thermal images of the unreinforced areas of the component under test to form multidimensional reference state data; When scanning the steel reinforcement area, the original electromagnetic induction signal is subjected to temperature field compensation and background normalization processing based on the multidimensional reference state data to generate an enhanced fusion signal. The enhanced fused signal is input into a dedicated deep learning model, which outputs the true value of the rebar cover thickness. Analyze the spatial distribution pattern of the true value of the concrete cover thickness of the steel bars, match it with a predefined defect feature library, and reverse diagnose and output the defect types in the production process.

2. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as described in claim 1, is characterized in that... The simultaneous acquisition of electromagnetic intrinsic response signals and surface temperature distribution thermal images of the unreinforced areas of the component under test constitutes multidimensional reference state data, specifically including: Based on the design drawings of the precast component to be tested, locate at least one predetermined unreinforced area on its surface; The electromagnetic sensor is controlled to scan the predetermined unreinforced area at a preset detection frequency and gain to obtain the electromagnetic intrinsic response signal expressed in complex domain form, including amplitude and phase information. Simultaneously with the scanning of the electromagnetic sensor, the infrared thermal imager is controlled to capture images of the entire detection area, including the predetermined area without reinforcing steel, to obtain thermal images of the surface temperature distribution. The electromagnetic intrinsic response signal is aligned and correlated with the surface temperature distribution thermal image to form multidimensional reference state data.

3. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning according to claim 2, characterized in that, Aligning and correlating the electromagnetic intrinsic response signal with the surface temperature distribution thermal image to form multidimensional reference state data, specifically including: Each electromagnetic intrinsic response signal collected by the electromagnetic sensor when scanning the predetermined area without reinforcing steel, and each frame of surface temperature distribution thermal image captured by the infrared thermal imager, is stamped with a uniform high-precision timestamp. Using a calibration plate or known spatial markers, the scanning coordinate system of the electromagnetic sensor and the imaging coordinate system of the infrared thermal imager are uniformly transformed to the same world coordinate system based on the prefabricated component under test. For each of the electromagnetic intrinsic response signals with a timestamp, the surface temperature distribution thermal image frame at the same time or in the time neighborhood is found according to the timestamp, and the precise temperature value of the corresponding coordinate point is extracted from the thermal image frame by interpolation based on its spatial coordinates in the world coordinate system. Each set of spatiotemporally aligned electromagnetic intrinsic response signals, their corresponding spatial coordinates, and the precise temperature value extracted from the thermal image are encapsulated into a structured data unit. The collection of all data units constitutes the multidimensional reference state data.

4. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as described in claim 3, is characterized in that... When scanning the reinforcing bar area, the original electromagnetic induction signal is subjected to temperature field compensation and background normalization processing based on the multidimensional reference state data to generate an enhanced fused signal, specifically including: When scanning the steel reinforcement area inside the precast component to be tested, the complex form of the original electromagnetic induction signal output by the electromagnetic sensor is collected in real time, and the real-time position coordinates at the time of signal collection are recorded synchronously by the positioning device coupled to the sensor. Based on the real-time position coordinates, spatial mapping is performed in the surface temperature distribution thermal image to extract the local background temperature value corresponding to the coordinate point; based on the pre-stored temperature-electromagnetic response lookup table, the amplitude and phase of the original electromagnetic induction signal are linearly corrected using the local background temperature value to obtain the temperature compensation signal. The temperature compensation signal and the electromagnetic intrinsic response signal are compared point by point in the complex domain to obtain the background normalized signal. The background normalized signal is subjected to amplitude logarithmic transformation and phase dewinding normalization processing to generate an enhanced fused signal. The enhanced fused signal represents the equivalent electromagnetic response determined only by the distance between the reinforcing bars under standard dry and cooled conditions.

5. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as described in claim 4, is characterized in that... The temperature compensation signal and the electromagnetic intrinsic response signal are compared point-by-point in the complex domain to obtain the background normalized signal, specifically including: Based on the real-time position coordinates, at least one electromagnetic intrinsic response signal with the nearest spatial coordinates is retrieved from the multidimensional reference state data; The temperature compensation signal Represented as complex number The retrieved electromagnetic intrinsic response signal Represented as complex number Perform complex division to calculate the initial normalized complex signal. ;in This is the signal amplitude after temperature compensation. is the base of the natural logarithm. The imaginary unit, The signal phase after temperature compensation. The amplitude of the background reference signal. For the background reference signal phase, The amplitude ratio, Phase difference; The amplitude component of the initial normalized complex signal Take the common logarithm and its phase components Phase unwinding is performed to obtain stable amplitude-phase pairs; The processed stable amplitude-phase pair is encapsulated into a new complex signal, which is the background normalized signal.

6. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as described in claim 5, is characterized in that... The enhanced fused signal is input into a dedicated deep learning model, which outputs the true value of the rebar cover thickness, specifically including: The enhanced fused signal, together with its corresponding real-time position coordinates, is organized according to the scanning time sequence and spatial position to form a two-dimensional data sequence; A local enhanced fusion signal block containing the current point to be solved and its surrounding neighborhood signals is extracted from the two-dimensional data sequence using a sliding window method; The local enhanced fusion signal block is input into a pre-trained dedicated deep learning model, which directly outputs a scalar value corresponding to the center point of the input signal block, which is the true value of the rebar cover thickness.

7. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning according to claim 6, characterized in that, The local enhanced fused signal block is input into a pre-trained dedicated deep learning model, specifically including: The input local enhanced fusion signal block is subjected to layer normalization processing so that the data distribution of its amplitude and phase components meets the input requirements of the dedicated deep learning model; Through the first layer of the dedicated deep learning model, a one-dimensional depthwise separable convolutional kernel is used to scan the local enhanced fusion signal block in the spatial dimension to extract primary spatial texture features related to the spacing and orientation of the reinforcing bars. The extracted primary spatial texture features are input into the second layer of a dedicated deep learning model. The second layer is a parallel dual-branch structure, where the first branch uses dilated convolution to expand the receptive field and capture large-scale thickness gradient patterns caused by mold side mold bulging or uneven spraying of release agent. The second branch uses an attention mechanism to focus on local high-gradient abrupt features caused by the positioning deviation of the steel cage. The feature maps output from the two branches are concatenated, and then high-dimensional feature fusion and dimensionality reduction are performed through multiple nonlinear fully connected layers. Finally, a scalar value, namely the true value of the concrete cover thickness of the steel reinforcement, is output through a regression layer.

8. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning according to claim 7, characterized in that, Analyze the spatial distribution pattern of the true value of the concrete cover thickness of the steel reinforcement, match it with a predefined defect feature library, reverse diagnose and output the defect types in the production process, specifically including: Based on the real-time position coordinates of all scanning points and their corresponding true values ​​of the concrete cover thickness, a two-dimensional spatial distribution map of the concrete cover thickness in the detection area of ​​the component under test is generated by a spatial interpolation algorithm. Image processing is performed on the two-dimensional protective layer thickness spatial distribution map to identify thickness anomaly regions where the thickness value exceeds the design tolerance range. The geometric shape, texture, gradient direction, and relative positional relationship with the reference edge of the component of the thickness anomaly region are calculated to form a descriptive anomaly spatial distribution pattern feature vector. The similarity between the abnormal spatial distribution pattern feature vector and the predefined typical production defect feature library is calculated. Based on the highest similarity score, determine the most likely type of production defect causing the current thickness anomaly, and output a visual diagnostic report that includes the defect type, confidence level, and location of the anomaly area.

9. The electromagnetic induction detection method for the thickness of the concrete cover of precast components based on deep learning, as described in claim 8, is characterized in that... The step of determining the most likely type of production defect causing the current thickness anomaly based on the highest similarity matching result specifically includes: The system receives the highest similarity score. If the similarity score is higher than a preset first threshold, the corresponding defect type is directly identified as a production defect type. If the similarity score is lower than the first threshold but higher than a preset second threshold, the system proceeds to the alternative analysis process. In the candidate analysis process, all candidate defect types with scores higher than the second threshold and their corresponding confidence levels are extracted. Then, combined with the absolute position of the thickness anomaly area on the component and the structural type of the component, secondary logical reasoning is performed to determine the most reasonable production process defect type from the candidate list. The system automatically integrates the final determined production defect type, its corresponding confidence level, the spatial coordinates of the thickness anomaly area, and key geometric evidence from the feature vector of the anomaly spatial distribution pattern to generate a structured, visualized diagnostic report.

10. An electromagnetic induction detection system for the thickness of the concrete cover of precast steel reinforcement in precast components based on deep learning, characterized in that, The method for electromagnetic induction detection of the concrete cover thickness of precast structural steel bars based on deep learning, as described in any one of claims 1-9, includes: The baseline state synchronous acquisition module is used to control the electromagnetic sensor to scan the predetermined unreinforced area on the surface of the precast component under test, and to acquire the electromagnetic intrinsic response signal of the component in the current state; at the same time, it controls the infrared thermal imager to acquire the surface temperature distribution thermal image of the component detection area; and it aligns and correlates the electromagnetic intrinsic response signal and the surface temperature distribution thermal image with time stamp and spatial coordinates to construct multi-dimensional baseline state data. The dynamic compensation and fusion processing module is communicatively connected to the reference state synchronous acquisition module. It is used to control the electromagnetic sensor to scan the steel reinforcement area inside the component, and to acquire the original electromagnetic induction signal and the corresponding real-time position coordinates in real time. Based on the real-time position coordinates, it maps the local background temperature value from the surface temperature distribution thermal image, and performs temperature field compensation on the original electromagnetic induction signal according to a predetermined temperature-electromagnetic response relationship. At the same time, it retrieves the electromagnetic intrinsic response signal at the corresponding position from the multi-dimensional reference state data, performs background normalization processing on the temperature-compensated signal, and generates an enhanced fusion signal. The thickness intelligent calculation module is communicatively connected to the dynamic compensation and fusion processing module. It includes a pre-trained dedicated deep learning model for receiving the enhanced fusion signal and directly outputting the true value of the steel reinforcement protective layer thickness after eliminating the influence of environmental conditions. The production defect diagnosis module is communicatively connected to the thickness intelligent calculation module. It is used to generate a thickness spatial distribution map based on the true value of the thickness of the steel reinforcement protective layer and its corresponding position coordinates; extract the spatial distribution pattern features of abnormal areas in the thickness spatial distribution map; perform pattern matching between the features and a predefined typical production defect feature library; and perform reverse diagnosis based on the matching results and output the type of production process defect that caused the thickness abnormality.

Images