Method and system for detecting defects of polyethylene plastic hollow plate

Abstract

The invention discloses a polyethylene plastic hollow plate defect detection method and system, and relates to the technical field of nondestructive testing, the system is composed of a plurality of functional modules, and the system comprises: a multi-mode sensing module for collecting multi-source data including a hollow plate surface image, three-dimensional morphology data and an ultrasonic reflection signal; the historical database construction module is used for performing space-time alignment and feature level fusion on the multi-source data to generate a composite defect feature matrix; collecting ultrasonic signal waveforms of known defects, extracting characteristic parameters, and establishing a historical waveform library; calculating a characteristic parameter classification range of each type of defects based on a historical waveform library; judging whether the detection waveform parameter falls into a classification range corresponding to the processing noise or not, and if so, rejecting the signal; otherwise, keeping; and the defect area identification module is used for dividing the surface of the hollow plate into uniform grids based on the surface image of the hollow plate, and counting the density of the residual ultrasonic detection sources in each grid.

Description

Technical Field

[0001] The invention relates to the technical field of nondestructive testing, and in particular to a method and system for detecting defects in a polyethylene plastic hollow plate. Background Art

[0002] With the expansion of material application scenarios, such as building load-bearing components and electronic industrial packaging, the detection dimension extends from appearance to mechanical properties; quantitative evaluation of material reliability is carried out through drop ball impact test, bending strength test, etc., for example, through the mechanical index threshold specified in "QB / T1651-1992 Polyethylene Plastic Hollow Board"; standardized testing procedures began to be established at this stage; the mechanical properties of polyethylene hollow board directly affect its reliability as packaging boxes and building partitions; internal support rib fracture or surface cracks may cause structural instability, and detection can eliminate safety hazards in advance; polyethylene plastic hollow board is widely used in logistics and transportation, building partitions, automotive interiors and other fields due to its light weight, corrosion resistance, high impact resistance and other characteristics; however, its production process is prone to material defects. Unevenness, mold wear, process parameter fluctuations and other factors introduce a variety of defects, including surface scratches, dents, bulges (appearance defects) and internal cracks, bubbles, and delamination (structural defects); these defects not only affect the appearance of the product, but also significantly reduce its mechanical properties, and even cause structural failure during use; in the electronics industry, if there is uneven distribution of conductive fillers in anti-static hollow boards, it may cause short circuits in components; at the same time, in food and medicine packaging, surface micropores or contamination will destroy the sterile environment, and toxicity testing and environmental performance testing are required to ensure compliance; polyethylene hollow boards can reduce costs by adding recycled materials, but too high a proportion of old materials will lead to a decrease in mechanical properties; testing the purity of raw materials and process defects, such as seam defects and uneven polishing, can balance cost and quality.

[0003] Traditional defect identification of polyethylene plastic hollow plates usually uses single-modal detection technology, and it is difficult to detect internal structural defects (such as stratification and bubbles), and is easily affected by environmental interference such as surface reflection and dust, with a high false alarm rate. At the same time, traditional systems usually use fixed thresholds or predefined noise templates to filter interference signals, but in actual production, the time-frequency characteristics of processing noise will dynamically shift due to equipment aging and changes in ambient temperature and humidity, resulting in incomplete noise filtering or true signals being mistakenly deleted. In addition, offline detection and manual adjustment are often combined, and the detection results cannot be fed back to the production line in real time, resulting in the expansion of the scale of defective batches and increased production costs. Summary of the invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: A polyethylene plastic hollow board defect detection system, comprising: Multimodal sensing module collects multi-source data, including hollow board surface image, three-dimensional shape data and ultrasonic reflection signal; The historical database construction module performs spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collects ultrasonic signal waveforms of known defects, extracts feature parameters, and establishes a historical waveform library; based on the historical waveform library, calculates the classification range of feature parameters for each type of defect; determines whether the detection waveform parameters fall into the classification range corresponding to the processing noise, and if so, removes the signal; otherwise, retains it; The defect area recognition module divides the surface of the hollow plate into uniform grids based on the surface image of the hollow plate, and counts the density of the remaining ultrasonic detection sources in each grid; it uses the region growing algorithm to expand the search of adjacent areas with the high-density grid as the center and merge continuous abnormal areas; it calculates the signal density ratio of the abnormal area, and if it exceeds the preset threshold, it is judged as a structural defect; otherwise, it uses a hybrid model constructed by a convolutional neural network and a support vector machine to perform appearance detection on the input three-dimensional morphology data and output the location of the appearance defect.

[0005] Furthermore, the surface image, three-dimensional morphology data and ultrasonic reflection signal of the hollow board are collected: The optical camera is used to collect RGB images of the hollow board surface for illumination correction and distortion correction. The laser scanner is used to obtain point cloud data of the hollow board surface and convert it into three-dimensional morphology data. Ultrasonic probes arranged at preset intervals emit ultrasonic waves into the hollow board and receive reflected signals.

[0006] Furthermore, the composite defect feature matrix: The multi-source data are time synchronized and space registered, and parameter features are extracted, including at least texture features, height distribution features of three-dimensional morphology and time-frequency domain features of ultrasonic signals. The extracted parameter features are fused to generate a composite defect feature matrix.

[0007] Furthermore, the process of parameter feature fusion is as follows: S201: Use the ResNet-18 neural network to process the input hollow plate surface image, generate an initial feature map through the convolution layer of the network, dynamically calculate the attention weight matrix based on the regional significance in the initial feature map, and multiply the attention weight matrix with the original feature map element by element for weighted fusion; after processing the weighted fused feature map, generate a composite feature representation vector of the image modality; S202: Taking each point in the point cloud data as the center, selecting adjacent points within a certain neighborhood around it, and calculating the curvature value of the center; taking each point as the center, counting the vertical heights of all points in the area within the same neighborhood; obtaining the variance of the height value, combining the local curvature value and the height variance value corresponding to each point to form a two-dimensional feature vector; S203: Decompose the preprocessed ultrasonic signal through multi-scale to obtain sub-bands of different frequency bands, each sub-band corresponds to a set of wavelet packet decomposition coefficients, square the wavelet packet decomposition coefficients of each sub-band to obtain the energy value of the corresponding frequency band; divide each energy value by the total energy to obtain the energy proportion of the frequency band; calculate the natural logarithm of each energy proportion value, and then multiply it with the corresponding energy proportion to obtain the energy-logarithm product of each frequency band; sum the products of all frequency bands and take the negative value to finally obtain the energy entropy value; S204: Based on the three-dimensional morphology data and the ultrasonic detection data, the three-dimensional features and the ultrasonic features are weighted respectively by preset weight coefficients, and then spliced ​​with the image features to finally obtain a composite feature matrix.

[0008] Furthermore, a historical waveform library is established: Based on the ultrasonic signal waveform in the defective hollow plate, the time domain, frequency domain and time-frequency domain characteristics of the ultrasonic signal are extracted, and a multi-dimensional feature vector is constructed. The wavelet packet is decomposed through time-frequency joint analysis, and the ultrasonic signal is refined into sub-bands. Based on the energy entropy quantification and the complexity of local energy distribution, the defect disturbance is detected; the defects are classified and the noise is removed through the synergistic effect of multiple features; among which, the synergistic effect of multiple features includes: signal strength, energy dynamics, frequency response and multi-scale energy distribution.

[0009] Furthermore, the characteristic parameter classification range of defects is: Under normal production conditions without defects, ultrasonic signals are continuously collected; the mean and standard deviation are calculated based on the time domain and frequency domain characteristics of the noise signal; K-means clustering is used to divide the clusters according to the defect type based on the characteristic parameters in the historical defect waveform library; the minimum, maximum and median of each characteristic parameter of the cluster for each type of defect is calculated through range definition; and combined with process experience, the abnormal clustering boundaries are corrected to determine whether the detection waveform parameters fall within the classification range corresponding to the processing noise, and noise is removed, and the combined parameters are screened through cross-validation.

[0010] Furthermore, the density of remaining ultrasonic detection sources in each grid is counted: S301: Divide the detection area into a plurality of grid units according to the acoustic field characteristics of the ultrasonic probe and the material thickness; and dynamically adjust according to the acoustic path difference threshold; S302: performing signal acquisition and feature extraction by identifying physical features of the ultrasonic detection source in the grid; S303: Count the valid signal points in the grid, normalize the sound pressure in the grid based on the sound field characteristics, calculate the relative density, adjust the density threshold according to the material thickness, map the multi-grid density data into a two-dimensional dot matrix in the C-scan image, and generate a comprehensive density distribution.

[0011] Furthermore, continuous abnormal regions are merged: By traversing all grids, the grids that meet the high-density condition are marked as seed points; an independent queue to be expanded is created for each seed point to store the position of the adjacent grids currently to be detected; based on the density value and judgment threshold of each grid, it is determined whether to merge the adjacent grids.

[0012] Furthermore, the signal density ratio of the abnormal area is calculated, the hybrid model is constructed, and the appearance defect position is output: The abnormal area boundary is extended outward by a number of grid widths, and the average density of the annular area formed after the expansion is taken as the reference value; the number of signals in each grid in the reference area is counted, the number of signals is added up to obtain the total number of signals, and the total number of signals is divided by the total number of grids in the reference area to obtain the reference area density; the abnormal area density is divided by the reference area density to obtain the signal density ratio; The ratio of the area and the square of the perimeter of the hollow board surface image recognition is calculated. When the result is close to 1, it is an ideal circle. The ratio of the maximum extension length of the defect in the x and y directions is taken, and the ratio of the long side to the short side of the minimum circumscribed rectangle is used to determine whether it is a linear defect based on the set threshold. The feature vector generated by CNN is received by SVM, and the defect type marked based on historical training data is classified, and the radial basis kernel function is used to handle nonlinear classification problems; the defect type and confidence are output; adjacent defect slices are merged to generate a continuous defect area bounding box; according to the grid division rules during preprocessing, the slice coordinates are mapped back to the actual physical position of the hollow plate.

[0013] A method for detecting defects in a polyethylene plastic hollow plate comprises the following steps: Step 1: Collect multi-source data, including the surface image of the hollow board, three-dimensional shape data and ultrasonic reflection signal; Step 2: Perform spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collect ultrasonic signal waveforms of known defects, extract feature parameters, and establish a historical waveform library; based on the historical waveform library, calculate the classification range of feature parameters for each type of defect; determine whether the detection waveform parameters fall into the classification range corresponding to the processing noise, if they match, remove the signal; otherwise, retain it; Step 3: Based on the surface image of the hollow board, the surface of the hollow board is divided into uniform grids, and the density of the remaining ultrasonic detection sources in each grid is counted; the region growing algorithm is used to expand the search of adjacent areas with the high-density grid as the center, and merge continuous abnormal areas; the signal density ratio of the abnormal area is calculated, and if it exceeds the preset threshold, it is judged as a structural defect; otherwise, the hybrid model constructed by the convolutional neural network and the support vector machine is used to perform appearance detection on the input three-dimensional morphology data and output the location of the appearance defect.

[0014] The present invention provides a method and system for detecting defects in a polyethylene plastic hollow plate, which has the following beneficial effects: (1) The present invention achieves high-precision and comprehensive defect detection of polyethylene plastic hollow plates through multimodal data fusion technology. It also integrates optical cameras, laser scanners and ultrasonic probes to collect surface images, three-dimensional morphology data and internal sound wave reflection signals respectively, and generates a composite defect feature matrix through spatiotemporal alignment and feature-level fusion. It avoids the risk of misjudgment caused by feature ambiguity in traditional methods through quantifiable features. It uses a dynamic weight mechanism to allocate the differentiated contribution of sensor data based on mutual information, for example, the ultrasonic feature weight accounts for 40%-50%, and the three-dimensional morphology and image features each account for 25%-30%, which significantly improves the discriminability of the composite feature matrix.

[0015] (2) The PCA dimensionality reduction technology eliminates redundant dimensions, significantly reducing computational complexity while retaining 95% of key information. This enables the system to efficiently process high-dimensional data, ensuring the accuracy of the detection results. It also provides a reliable data basis for subsequent defect classification and noise removal, and solves the problem that single-modal detection is susceptible to environmental interference or one-sided information.

[0016] (3) The present invention effectively distinguishes real defects from processing noise through the region growing algorithm and the dynamic threshold judgment of the signal density ratio. For example, the adjacent areas are expanded and searched with the high-density grid as the center, and the continuous abnormal areas are merged. The ratio of the density of the abnormal area to the density of the reference area is calculated, which significantly reduces the false alarm rate and improves the recognition sensitivity of small defects such as cracks and delamination. Secondly, the system uses a hybrid model to detect appearance defects on three-dimensional morphological data, and combines geometric features such as the minimum circumscribed rectangle and aspect ratio analysis to achieve accurate positioning and classification of appearance defects such as dents and scratches, with a confidence level of more than 95%.

[0017] (4) The early warning and feedback module synchronizes the defect location, type and process parameter adjustment suggestions, such as extruder temperature and mold pressure, to the production line in real time, forming a full closed-loop management of "detection-analysis-adjustment". For example, when a bubble defect is detected, the system automatically triggers the mold pressure to increase by 0.2-0.5MPa, significantly reducing the defect rate. This not only improves the production quality and yield rate of hollow boards, but also enhances the robustness of the system through a dynamic threshold adaptive mechanism, making it suitable for long-term stable operation under different working conditions, and showing significant practicality and engineering value in actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the system flow of the present invention; Figure 2 It is a schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Example 1 See also Figure 1 This embodiment provides a polyethylene plastic hollow plate defect detection system, the detection system comprising: The multi-modal sensing module collects the surface image, three-dimensional shape data and ultrasonic reflection signal of the hollow board respectively; The process of collecting the surface image, three-dimensional shape data and ultrasonic reflection signal of the hollow board is as follows: For the convenience of explanation, the hollow board in the whole article refers to the polyethylene plastic hollow board; Optical camera surface imaging steps: When the polyethylene plastic hollow board enters the inspection area, the optical camera is triggered to capture the surface image; through image preprocessing, grayscale correction is performed to eliminate uneven lighting and enhance edges, thereby highlighting defects such as cracks and scratches, outputting a high-contrast image, and marking the coordinates of the suspected defective area; Optical camera: A high-resolution linear array CCD camera with a resolution of ≥0.1mm is used, equipped with a ring-shaped LED light source to eliminate surface reflection interference; the camera is installed along the conveyor belt direction of the hollow plate and continuously captures surface images at a fixed frame rate of more than 200Hz; Laser scanning 3D shape reconstruction steps: The laser beam scans the surface of the hollow plate, and the CMOS sensor captures the displacement of the reflected light spot in real time; the three-dimensional coordinates of each point are calculated through the triangulation formula; a three-dimensional point cloud model with millimeter-level accuracy is generated to detect uneven thickness, concave or convex defects; Laser scanner: Use a laser scanner based on the principle of triangulation, with a scanning frequency of ≥500Hz; the laser emits a linear laser beam, receives the reflected light spot through a high-speed CMOS sensor, and generates three-dimensional point cloud data in real time with an accuracy of ±0.05mm; Ultrasonic internal defect detection steps: The ultrasonic probe emits a pulse signal to penetrate the inside of the hollow plate; the reflected wave signal is received to extract the time domain waveform and frequency domain features, such as echo amplitude and flight time; the threshold method is used to identify abnormal reflections, such as sudden changes in acoustic impedance caused by bubbles, and locate the depth of defects; Ultrasonic probe: Equipped with a high-frequency pulse ultrasonic probe with a center frequency of 5MHz. It adopts an integrated transceiver design. The probe array is arranged equidistantly along the width of the hollow plate, and the sampling frequency is ≥1kHz. The coupling agent is used to ensure that the sound waves effectively penetrate the material and receive the internal reflection signal. Send synchronization signals through the PLC controller to ensure that the optical camera, laser scanner and ultrasonic probe start data acquisition in the same time window; The historical database construction module performs spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collects ultrasonic signal waveforms of known defects, extracts feature parameters, and establishes a historical waveform library; based on the historical waveform library, calculates the classification range of feature parameters for each type of defect; determines whether the detection waveform parameters fall into the classification range corresponding to the processing noise, and if they match, the signal is removed; otherwise, it is retained; The process of generating the composite defect feature matrix is: Through multimodal data preprocessing, quantifiable feature extraction such as LBP histogram and wavelet packet entropy, dynamic weight fusion optimization based on mutual information, and PCA dimension reduction, a highly discriminative composite defect feature matrix is ​​constructed to achieve a full-process technical closed loop from data collection to feature fusion; Multi-source data: By preprocessing multi-source data, graying and Gaussian filtering are performed on the surface image to eliminate light noise; outlier points are eliminated and smoothed on the three-dimensional morphology data to correct measurement errors; and the ultrasonic reflection signal is band-pass filtered and baseline corrected to suppress environmental interference; Feature extraction: Extract the texture features of the surface image of the hollow board, the height distribution features of the three-dimensional morphology and the time-frequency domain features of the ultrasonic signal; wherein the texture features of the surface image at least include grayscale gradient, edge sharpness and local binary pattern histogram; the height distribution features of the three-dimensional morphology at least include surface curvature, regional concavity and normal vector deviation; the time-frequency domain features of the ultrasonic signal at least include peak amplitude, decay time, spectrum main frequency energy proportion and wavelet packet decomposition entropy; Feature normalization and weight allocation: Through timestamp synchronization and spatial registration, the surface image, three-dimensional morphology data and ultrasonic signal are aligned to find the optimal space-time transformation matrix, so that the difference between the surface image and the three-dimensional morphology data after the matrix transformation is minimized. At the same time, the time axis of the ultrasonic signal is converted to a time coordinate system consistent with the spatial data and multiplied by a weight coefficient. The Z-score of each modal feature is normalized to eliminate the dimensional difference. Based on the feature information of the historical data, the weight is dynamically allocated, of which the ultrasonic feature weight accounts for 40%-50%, and the image and morphology features account for 25%-30% each. It is used to balance the contribution of different modal data. Finally, the surface image and the transformed three-dimensional morphology data are fused through convolution operation, and the weighted ultrasonic signal is added. The overall difference value is calculated and minimized, so that the data of different modalities are accurately matched in time and space. Composite defect feature matrix generation: S201: Use the improved ResNet-18 neural network to process the input hollow board surface image, and generate the initial feature map through the convolution layer of the network, where the feature map contains the basic visual information of the image such as local texture and edge; dynamically calculate the attention weight matrix based on the regional saliency in the initial feature map, analyze the importance of different areas in the image, such as areas with drastic texture changes or high contrast, and assign higher weight values ​​to key areas, i.e. high-weight areas, to highlight the characteristics of the potential defect areas; multiply the generated attention weight matrix by the original feature map element by element, enhance the feature expression of the high-weight area, and suppress the redundant information of the non-key area; after processing the weighted fused feature map, finally generate a 512-dimensional vector as the composite feature representation of the image modality; S202: Taking each point in the point cloud data as the center, select the adjacent points within a certain neighborhood around it, and the surrounding 1 unit or 2 unit area; calculate the curvature value at the point by fitting the local surface or plane; wherein the curvature reflects the curvature of the surface, and the curvature value of the convex or concave area will be significantly higher than that of the flat area; repeat this process for all points to generate the local curvature distribution value of the polyethylene plastic hollow board surface; taking each point as the center, count the vertical heights of all points in the area within the same neighborhood; obtain the variance of the height value, the larger the variance value, the more drastic the height fluctuation of the area, which may correspond to the surface unevenness or defect area, generate the height variance distribution value of the entire surface, combine the local curvature value C and the height variance value H corresponding to each point, and form a two-dimensional feature vector, which integrates the bending characteristics of the surface shape of the polyethylene plastic hollow board and the irregularity of the height change, for subsequent defect recognition and feature fusion; S203: Input the preprocessed ultrasonic signal into the wavelet packet transform model, perform multi-scale decomposition on the signal, decompose the ultrasonic signal into sub-bands of different frequency bands, each sub-band corresponds to a set of wavelet packet decomposition coefficients, and the wavelet packet decomposition coefficients reflect the energy distribution characteristics of the signal in different frequency bands; square the wavelet packet decomposition coefficients of each sub-band to obtain the energy value of the corresponding frequency band; divide each energy value by the total energy to obtain the energy proportion of the frequency band; calculate the natural logarithm of each energy proportion value, and then multiply it with the corresponding energy proportion to obtain the energy-logarithm product of each frequency band; sum the products of all frequency bands and take the negative value to finally obtain the energy entropy value; this value quantifies the uniformity of the energy distribution of the signal in different frequency bands. The higher the entropy value, the more dispersed the energy distribution, which may correspond to complex defect characteristics; traverse and scan the absolute value of the original ultrasonic signal, and record the maximum value of the signal amplitude; the maximum value is the peak value of the reflected wave, which reflects the maximum energy intensity reflected back when the ultrasonic wave encounters a defect inside the material; S204: extract the three-dimensional feature F3D, multiply it by the preset weight coefficient 0.6, and enhance the contribution weight of the three-dimensional morphology data; combine the energy entropy and peak power of the ultrasound, and multiply it by the preset weight coefficient 0.4 to balance the influence of the ultrasonic feature; splice the weighted three-dimensional features, weighted ultrasonic features, and unweighted image features Fimg in sequence to form a final composite feature matrix; the final generated composite matrix Mdefect is a 516-dimensional real vector, which integrates the features of three types of modes: image, three-dimensional morphology, and ultrasound; it should be noted that the weight coefficients 0.6 and 0.4 are optimized and determined by the grid search method in the experiment to ensure the optimal contribution ratio of different modal data and improve the accuracy of defect detection; splice the weighted multimodal features into the initial matrix according to the spatial position after time-space alignment; use principal component analysis to reduce the dimension of the initial matrix, retain the principal components with a cumulative contribution rate ≥ 95%, and generate a composite defect feature matrix; Create a historical waveform library: Through experiments, hollow board samples with typical defects are prepared, or hollow board waste products that are rejected due to unqualified quality during the production process are collected, such as natural defects such as bubbles and cracks. The defect characteristics are based on the defects in the actual production process; and through the defect data of plastic products in open data sets such as PHM Challenge and MVTec AD, the hollow board detection scene is adapted through transfer learning, and its ultrasonic signal waveform is collected; In the extraction of characteristic parameters for defect detection, the system constructs a multi-dimensional feature vector by comprehensively analyzing the time domain, frequency domain and time-frequency domain characteristics of ultrasonic signals; at the time domain level, the peak amplitude reflects the strength of the defect reflection signal, the signal energy measures the overall energy distribution, and the rise time captures the dynamic changes of the signal to distinguish the defect type; at the frequency domain level, the concentrated area of ​​the signal energy within the frequency range is calculated by the centroid of the spectrum to identify the frequency offset caused by material anomalies; and through joint time-frequency analysis, the wavelet packet is decomposed to refine the ultrasonic signal into sub-bands, and the complexity of the local energy distribution is quantified based on energy entropy to accurately detect tiny defect disturbances; from the synergistic effect of multiple features such as signal strength, energy dynamics, frequency response and multi-scale energy distribution, a highly discriminative quantitative basis is provided for defect classification and noise removal; The superimposed waveforms are separated by using a deep learning model, and the time-frequency parameters are dynamically adjusted through an optimization algorithm to match the characteristics of Gaussian modulated pulses to achieve accurate separation of noise and effective signals. The final classification waveform library is constructed, through the digital storage and interpolation reconstruction mechanism of feature vectors (such as peak and trough positions, contour intersection amplitudes), to achieve high-fidelity waveform restoration and rapid retrieval and comparison under noise background. Calculate the classification range of characteristic parameters for each type of defect: The classification range corresponding to processing noise: Under normal production conditions without defects, ultrasonic signals (≥500 groups) are continuously collected to ensure coverage of scenes such as normal vibration of the production line and environmental interference; the mean and standard deviation of the time domain and frequency domain characteristics of the noise signal are calculated respectively; based on the K-sigma criterion, K=3 is usually used, and K is the proportion of the data set falling within a specific range; the noise parameter confidence interval is defined as [μ-3σ, μ+3σ], where μ and σ are the mean and standard deviation of the frequency domain characteristics respectively; it usually covers 99.7% of the normal noise distribution; Defect characteristic parameter classification range calculation: Based on the characteristic parameters in the historical defect waveform library, Gaussian mixture model or K-means clustering is used to divide clusters according to defect types; through range definition, the minimum, maximum and median of each characteristic parameter of each type of defect cluster are calculated; and combined with process experience, abnormal cluster boundaries are manually corrected, such as removing outliers, and finally a characteristic parameter classification range table for each type of defect is formed, as shown below: Table 1: Classification range of characteristic parameters of crack defects:

[0021] Table 2: Classification range of characteristic parameters of bubble defects:

[0022] Table 3: Classification range of characteristic parameters of delamination defects:

[0023] Table 4: Classification range of characteristic parameters of defects caused by processing noise:

[0024] It should be noted that: The defect signal is based on the cluster analysis of the historical waveform library, combined with the manual correction boundary; the noise signal is calculated through the K-sigma criterion, the confidence interval is calculated, and it covers 99.7% of the normal noise distribution; among the characteristic parameters, the peak amplitude reflects the defect reflection intensity, and the crack / delamination has a higher amplitude due to structural mutation; the main frequency position is that the bubble is biased towards low frequency due to cavity resonance, and the stratification is biased towards high frequency due to interface reflection; the proportion of harmonic components is that the processing noise has very few harmonics, and the real defect generates harmonics due to nonlinear reflection; the wavelet packet energy entropy is used to quantify the signal complexity, and the defect signal entropy value is higher; Determine whether the detection waveform parameters fall within the classification range corresponding to the processing noise, and remove the noise. Through cross-validation, select the combination parameters with strong discrimination power, which significantly improves the category separation degree of the feature space. At the same time, introduce a dynamic threshold mechanism to automatically calibrate the classification boundary according to real-time production variables, such as material thickness changes and process parameter fluctuations. For example, in the detection scenario of polyethylene plastic hollow plates, the ratio of the major diameter to the threshold is reduced to capture tiny cracks, while the threshold is relaxed to avoid false alarms when detecting thick plates, ensuring that the classification rules always maintain adaptive matching with the actual working conditions of the production line. The synergy of these two methods takes into account both theoretical rigor and engineering practicality, forming a flexible and adjustable defect judgment system. For the ultrasonic signal currently being tested, extract the time domain, frequency domain and time-frequency joint characteristic parameters that are the same as those in the historical waveform library; compare the parameters with the noise confidence interval and defect classification range table item by item, and determine: If all characteristic parameters fall within the noise confidence interval, it is determined to be processing noise and the signal is removed; If at least one characteristic parameter exceeds the noise interval and falls into the classification range of a certain type of defect, it is determined to be a real defect signal, and the defect type is retained and associated; If the parameter matches neither the noise nor the known defect, it is marked as a "signal to be verified", triggering the manual re-inspection process; During the manual re-inspection process, the unknown defect signals confirmed have their characteristic parameters added to the historical waveform library, cluster analysis is re-performed and the classification range is updated; and the latest noise signals are collected regularly, and the confidence intervals are recalculated to adapt to the aging of production line equipment or environmental changes; The defect area recognition module divides the surface of the hollow plate into uniform grids based on the surface image of the hollow plate, and counts the density of the remaining ultrasonic detection sources in each grid; the regional growing algorithm is used to expand the search of adjacent areas with the high-density grid as the center, and merge continuous abnormal areas; the signal density ratio of the abnormal area is calculated, and if it exceeds the preset threshold, it is determined to be a structural defect; conversely, a hybrid model constructed by a convolutional neural network and a support vector machine is used to perform appearance inspection on the input three-dimensional morphology data and output the location of the appearance defect; Divide into a uniform grid: S301: Divide the detection area into multiple grid units according to the acoustic field characteristics of the ultrasonic probe and the material thickness; dynamically adjust according to the acoustic path difference threshold, such as dividing it into independent imaging areas when the acoustic path difference exceeds the set value, and the grid size is usually related to the wavelength of the sound wave; for example, in the acoustic field simulation, the maximum grid unit size is set to 1 / 10 of the wavelength of the sound wave to balance the accuracy and the amount of calculation; for peak detection, a 5×5 pixel search interval is often used to count the signal density; S302: performing signal collection and feature extraction by identifying physical features of the ultrasonic detection source in the grid; For example: count the number of bright spots (such as pixel grayscale values ​​exceeding the threshold) in a 5×5 pixel area to determine the presence of a wave peak; in a microcrack model, identify the sound source density in the crack area through the sideband components of the mixing signal (such as sum frequency and difference frequency) to perform a nonlinear response signal; use a multivariate Gaussian beam model to calculate the square integral of the sound pressure amplitude in the grid as the sound source energy density to obtain the sound pressure integral; If the sound field is uniform along the thickness direction of the polyethylene plastic hollow plate, the sound pressure integral in the grid is: Where D is the integral of the sound pressure in the grid; A is the grid area; is the sound pressure amplitude at the spatial coordinate (x, y); is the medium density; c is the speed of sound; dA is the integral over the grid area A; S303: Count the valid signal points in the grid, normalize the sound pressure or energy in the grid in combination with the acoustic field characteristics, calculate the relative density, adjust the density threshold according to the material thickness or process parameter fluctuations, map the multi-grid density data into a two-dimensional dot matrix in the C-scan image, and generate a comprehensive density distribution by overlaying, superposition or extreme value extraction; for example, the peak density of each frame in the sequence image is counted after suppressing false alarms through morphological operations; C Scanning: Observe and analyze the defects at a specific depth inside the hollow board; Merge consecutive anomaly regions: In the ultrasonic signal density distribution after the surface of the polyethylene plastic hollow board is evenly meshed, the density value of each mesh is the number of effective ultrasonic signals detected in the area; the density value ≥ 2 times the average density of adjacent meshes is used as the high-density mesh judgment threshold, and the adjacent meshes are merged when the density difference is ≤ 15%; Traverse all grids and mark the grids that meet the high-density condition as seed points; create an independent queue to be expanded for each seed point to store the current adjacent grid positions to be detected; Step 1: Filter out all high-density grids from the density distribution map; For example: the density of grid A is 25 (the average density of adjacent grids is 10), which meets the 2.5 times threshold and is marked as a seed point; Step 2: Perform the following operations on the queue to be expanded for each seed point: a1: Take out the first grid in the queue (such as seed point A); a2: Search its adjacent grids (four or eight neighborhoods, usually four neighborhoods are selected to reduce the amount of calculation): up, down, left, and right four-directional grids; a3: Determine whether to merge: If the adjacent grid is not marked as an abnormal area, and the difference between its density value and the current grid is ≤15% (for example, the current density is 25, and the adjacent grid densities 21-29 all meet the conditions); at the same time, the density of the adjacent grid must be higher than the global minimum abnormal density threshold (for example, density ≥5, to avoid merging low-noise areas); Step 3: Region merging and queue updating: If the merging conditions are met: mark the adjacent grids as the same abnormal area; add them to the queue to be expanded of the current seed point, waiting for subsequent expansion detection; if the conditions are not met: abandon the adjacent grid and do not merge; Step 4: Iteration termination: When the queue to be expanded of a seed point is empty, the growth of the region is stopped; after all seed points are processed, the merged continuous abnormal region set is output; It should be noted that: When merging abnormal areas, during the cross-seed area merging process: if two independently grown abnormal areas meet during the expansion process, that is, adjacent grids are detected by both queues at the same time, and the density difference is ≤15%, they are merged into the same defective area; when performing edge processing, when the grid is located at the edge of the hollow plate, only the existing adjacent grids are detected, such as the upper left corner grid is only detected in the right and lower directions, and the virtual grids beyond the boundary are ignored; For example: The density of the central grid (1, 1) is 25 (the adjacent average density = (5+8+6+20) / 4≈9.75), which meets the 2.5 times threshold (25≥9.75×2); The grid (2, 1) density is 20 (the adjacent average density = (25+9+7) / 3≈13.67), which does not meet the 2-fold threshold (20<13.67×2) and is not marked as a seed point; Region growing process: Starting from the seed point (1, 1), detect its four neighbors: Above (0, 1) density 8 (difference = |25-8| / 25=68%>15%) → do not merge; Below (2, 1) density 20 (difference = |25-20| / 25=20%>15%) → not merged; Left side (1, 0) density 6 (difference = |25-6| / 25=76%>15%) → not merged; Right side (1, 2) density 22 (difference = |25-22| / 25=12%≤15%) → merge and join the queue; Continue expanding from the newly merged (1, 2) grid (density 22): Above (0, 2) density 12 (difference = |22-12| / 22≈45%>15%) → not merged; The density at the bottom (2, 2) is 18 (difference = |22-18| / 22≈18%>15%) → not merged; The left side (1, 1) has been merged; Right side (1, 3) density 10 (difference = |22-10| / 22≈55%>15%) → not merged; The queue is empty and the growth stops; Output: The merged abnormal areas are grids (1, 1) and (1, 2), and the signal density ratio is (25+22) / 2=23.5; Calculate the signal density ratio of the abnormal area: Expand 2-3 grid widths outward from the abnormal area boundary, exclude other abnormal areas, and take the average density of this annular area as the reference value; For example: The abnormal area covers the grid range: to ; Reference area range: to , remove the internal abnormal area; when the adjacent area is affected by other defects, the average density of the entire hollow board is used as the reference value; Calculate the density of the abnormal area: Count the number of signals in each grid in the reference area, add up these signal numbers to get the total signal number, divide the total signal number by the total number of grids in the reference area to get the reference area density; divide the abnormal area density by the reference area density to get the signal density ratio; For example: Assume that the reference area contains 12 grids, and the number of signals in each grid is 8; calculate the total number of signals: 8*12=96; calculate the reference area density: 96 / 12=8.0; Using the above calculation results: the density of the abnormal area is 20.0; the density of the reference area is 8.0; the signal density ratio is calculated as: 20.0 / 8.0=2.5; The preset threshold determines: Basis for setting the preset threshold: Based on historical data statistics, the maximum SDR under normal production conditions is 1.8 (μ+3σ); If SDR ≥ 2.0, it is determined to be a structural defect (such as cracks, delamination); in this case, SDR = 2.5> 2.0, which is determined to be a structural defect; If SDR<2.0, it is judged as a structural defect (such as cracks, delamination), and the input 3D shape data is inspected for appearance; It should be noted that: during high-speed production, the signal acquisition time is shortened, the density value may decrease, and the threshold is adjusted down proportionally. For example, if the speed increases by 20%, the threshold is adjusted to 1.8; the background noise level is regularly tested. If the noise density increases by 10%, the threshold is simultaneously increased by 10% to avoid misjudgment; The hybrid model constructed: The ratio of the area to the square of the perimeter is calculated by the surface image recognition of the hollow board. When the result is close to 1, it is an ideal circle, and the range is usually set to 0.85-1.0; the ratio of the maximum extension length of the defect in the x and y directions is taken. The aspect ratio of the elliptical defect is generally in the range of 1.2-2.0. The ratio of the long side to the short side of the minimum circumscribed rectangle is usually set to a threshold of >3.0, and it is judged as a linear defect if it exceeds; for example, the ratio of the long side to the short side of the minimum circumscribed rectangle of knife wires and cracks is usually set to a threshold of >3.0; judged by the standard deviation of the main direction angle of the minimum circumscribed rectangle, if the standard deviation is <10°, it is a regular linear defect; Support vector machine classification and positioning: Classification Model: The feature vector generated by CNN is received by SVM, and the defect types marked, such as scratches, dents, and bulges, are classified based on historical training data: Use radial basis kernel function to handle nonlinear classification problems; output defect type and confidence level, such as "scratch, confidence level 92%"; Positioning logic: If a slice is judged as a defect, its original grid position coordinates are recorded, such as the coordinates of the slice center point (x, y); adjacent defective slices are merged to generate a continuous defect area boundary box; otherwise, no processing is performed; Defect position calibration and mapping: According to the grid division rules during preprocessing, the slice coordinates are mapped back to the actual physical position of the hollow board: if the grid resolution is 1cm×1cm, the slice center coordinates (x, y) = (5, 5) correspond to the actual position of the hollow board (5cm, 5cm); Bounding box correction: Use the non-maximum suppression algorithm to merge adjacent bounding boxes with high overlap rates to avoid repeated calibration; Output format: defect location is marked with physical coordinate system (X, Y) or grid index (i, j), and associated with defect type and confidence level; For example: Input data: The three-dimensional morphology slice showed that the height value of a certain area dropped abnormally, which was suspected to be a depression in the polyethylene plastic hollow board; CNN was used to extract feature vectors, and SVM was determined to be "depression" with a confidence level of 95%; the center coordinates of the slice were grid (3, 7), mapped to the actual position (30 mm, 70 mm); Merge three adjacent defect slices and output the defect area as a rectangular frame: Upper left corner (25mm, 65mm), lower right corner (35mm, 75mm); Output: {type: concave, position: [(25, 65), (35, 75)], confidence: 95%}; Early warning and feedback module, marking defect location and type, and synchronously feeding back to the production line for adjustment; In the digital twin model of the hollow board or the real-time monitoring interface, defects are marked with color codes (such as red for structural defects and yellow for appearance defects); bounding boxes and defect type labels (such as "cracks - high confidence") are superimposed; defect information (type, location, timestamp) is stored in the database to support batch traceability and quality analysis, generate feedback instructions and match adjustment strategies, and match the corresponding table of production line adjustment parameters according to the defect type and process knowledge base: Table 5: Classification of defect type adjustment parameters and directions:

[0025] Through the deep integration of intelligent decision-making and industrial control, a full closed-loop management from defect detection to production line optimization is achieved, which significantly improves the production quality and efficiency of polyethylene hollow boards.

[0026] Example 2 See also Figure 2 Based on Example 1, this embodiment also provides a method for detecting defects in a polyethylene plastic hollow plate, comprising the following specific steps: Step 1: Collect multi-source data, including the surface image of the hollow board, three-dimensional shape data and ultrasonic reflection signal; Step 2: Perform spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collect ultrasonic signal waveforms of known defects, extract feature parameters, and establish a historical waveform library; based on the historical waveform library, calculate the classification range of feature parameters for each type of defect; determine whether the detection waveform parameters fall into the classification range corresponding to the processing noise, if they match, remove the signal; otherwise, retain it; Step 3: Based on the surface image of the hollow board, the surface of the hollow board is divided into uniform grids, and the density of the remaining ultrasonic detection sources in each grid is counted; the region growing algorithm is used to expand the search of adjacent areas with the high-density grid as the center, and merge continuous abnormal areas; the signal density ratio of the abnormal area is calculated, and if it exceeds the preset threshold, it is judged as a structural defect; otherwise, the hybrid model constructed by the convolutional neural network and the support vector machine is used to perform appearance detection on the input three-dimensional morphology data and output the location of the appearance defect.

[0027] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0028] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0029] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A polyethylene plastic hollow board defect detection system, characterized in that: The system includes: Multimodal sensing module collects multi-source data, including hollow board surface image, three-dimensional shape data and ultrasonic reflection signal; The historical database construction module performs spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collects ultrasonic signal waveforms of known defects, extracts feature parameters, and establishes a historical waveform library; based on the historical waveform library, calculates the classification range of feature parameters for each type of defect; determines whether the detection waveform parameters fall into the classification range corresponding to the processing noise, and if so, removes the signal; otherwise, retains it; The defect area recognition module divides the surface of the hollow plate into uniform grids based on the surface image of the hollow plate, and counts the density of the remaining ultrasonic detection sources in each grid; it uses the region growing algorithm to expand the search of adjacent areas with the high-density grid as the center and merge continuous abnormal areas; it calculates the signal density ratio of the abnormal area, and if it exceeds the preset threshold, it is judged as a structural defect; otherwise, it uses a hybrid model constructed by a convolutional neural network and a support vector machine to perform appearance detection on the input three-dimensional morphology data and output the location of the appearance defect.

2. A polyethylene plastic hollow board defect detection system according to claim 1, characterized in that: The process of collecting the surface image, three-dimensional shape data and ultrasonic reflection signal of the hollow board is as follows: The optical camera is used to collect RGB images of the hollow board surface for illumination correction and distortion correction. The laser scanner is used to obtain point cloud data of the hollow board surface and convert it into three-dimensional morphology data. Ultrasonic probes arranged at preset intervals emit ultrasonic waves into the hollow board and receive reflected signals.

3. A polyethylene plastic hollow board defect detection system according to claim 1, characterized in that: The process of generating the composite defect feature matrix is: The multi-source data are time synchronized and space registered, and parameter features are extracted, including at least texture features, height distribution features of three-dimensional morphology and time-frequency domain features of ultrasonic signals. The extracted parameter features are fused to generate a composite defect feature matrix.

4. A polyethylene plastic hollow board defect detection system according to claim 3, characterized in that: The process of fusing the parameter features is as follows: S201: Use the ResNet-18 neural network to process the input hollow plate surface image, generate an initial feature map through the convolution layer of the network, dynamically calculate the attention weight matrix based on the regional significance in the initial feature map, and multiply the attention weight matrix with the original feature map element by element for weighted fusion; after processing the weighted fused feature map, generate a composite feature representation vector of the image modality; S202: Taking each point in the point cloud data as the center, selecting adjacent points within a certain neighborhood around it, and calculating the curvature value of the center; taking each point as the center, counting the vertical heights of all points in the area within the same neighborhood; obtaining the variance of the height value, combining the local curvature value and the height variance value corresponding to each point to form a two-dimensional feature vector; S203: Decompose the preprocessed ultrasonic signal through multi-scale to obtain sub-bands of different frequency bands, each sub-band corresponds to a set of wavelet packet decomposition coefficients, square the wavelet packet decomposition coefficients of each sub-band to obtain the energy value of the corresponding frequency band; divide each energy value by the total energy to obtain the energy proportion of the frequency band; calculate the natural logarithm of each energy proportion value, and then multiply it with the corresponding energy proportion to obtain the energy-logarithm product of each frequency band; sum the products of all frequency bands and take the negative value to finally obtain the energy entropy value; S204: Based on the three-dimensional morphology data and the ultrasonic detection data, the three-dimensional features and the ultrasonic features are weighted respectively by preset weight coefficients, and then spliced ​​with the image features to finally obtain a composite feature matrix.

5. A polyethylene plastic hollow board defect detection system according to claim 1, characterized in that: The process of establishing the historical waveform library is as follows: Based on the ultrasonic signal waveform in the defective hollow plate, the time domain, frequency domain and time-frequency domain characteristics of the ultrasonic signal are extracted, and a multi-dimensional feature vector is constructed. Through the joint analysis of time and frequency, the wavelet packet is decomposed, and the ultrasonic signal is refined into sub-bands. Based on the energy entropy quantification and the complexity of the local energy distribution, the defect disturbance is detected; Defect classification and noise removal are achieved through the synergy of multiple features, including signal strength, energy dynamics, frequency response and multi-scale energy distribution.

6. A polyethylene plastic hollow board defect detection system according to claim 5, characterized in that: The process of calculating the classification range of the characteristic parameters of each type of defect is as follows: Under normal production conditions without defects, ultrasonic signals are continuously collected; based on the time domain and frequency domain characteristics of the noise signal, the mean and standard deviation are calculated respectively; based on the characteristic parameters in the historical defect waveform library, K-means clustering is used to divide clusters according to defect types; through range definition, the minimum, maximum and median of each characteristic parameter of the cluster of each type of defect is calculated; Combined with process experience, the abnormal clustering boundaries are corrected, and it is determined whether the detection waveform parameters fall within the classification range corresponding to the processing noise. The noise is then removed and the combined parameters are screened through cross-validation.

7. A polyethylene plastic hollow board defect detection system according to claim 1, characterized in that: The process of counting the remaining ultrasonic detection source density in each grid is as follows: S301: Divide the detection area into a plurality of grid units according to the acoustic field characteristics of the ultrasonic probe and the material thickness; and dynamically adjust according to the acoustic path difference threshold; S302: performing signal acquisition and feature extraction by identifying physical features of the ultrasonic detection source in the grid; S303: Count the valid signal points in the grid, normalize the sound pressure in the grid based on the sound field characteristics, calculate the relative density, adjust the density threshold according to the material thickness, map the multi-grid density data into a two-dimensional dot matrix in the C-scan image, and generate a comprehensive density distribution.

8. A polyethylene plastic hollow board defect detection system according to claim 7, characterized in that: The process of merging the continuous abnormal regions is: By traversing all grids, the grids that meet the high-density condition are marked as seed points; Create an independent queue to be expanded for each seed point to store the adjacent grid positions to be detected; based on the density value and judgment threshold of each grid, determine whether to merge the adjacent grids.

9. A polyethylene plastic hollow board defect detection system according to claim 8, characterized in that: The process of calculating the signal density ratio of the abnormal area, constructing the hybrid model, and outputting the appearance defect position is as follows: Expand the abnormal area boundary outward by a certain number of grid widths, and take the average density of the annular area formed after the expansion as the reference value; Count the number of signals in each grid in the reference area, add up the number of signals to get the total number of signals, divide the total number of signals by the total number of grids in the reference area to get the reference area density; divide the abnormal area density by the reference area density to get the signal density ratio; The ratio of the area and the square of the perimeter of the hollow board surface image recognition is calculated. When the result is close to 1, it is an ideal circle. The ratio of the maximum extension length of the defect in the x and y directions is taken, and the ratio of the long side to the short side of the minimum circumscribed rectangle is used to determine whether it is a linear defect based on the set threshold. The feature vector generated by CNN is received by SVM, and the defect type marked based on historical training data is classified, and the radial basis kernel function is used to handle nonlinear classification problems; the defect type and confidence are output; adjacent defect slices are merged to generate a continuous defect area bounding box; according to the grid division rules during preprocessing, the slice coordinates are mapped back to the actual physical position of the hollow plate.

10. A method for detecting defects in polyethylene plastic hollow plates, characterized in that: The steps include: Step 1: Collect multi-source data, including the surface image of the hollow board, three-dimensional shape data and ultrasonic reflection signal; Step 2: Perform spatiotemporal alignment and feature-level fusion of multi-source data to generate a composite defect feature matrix; collect ultrasonic signal waveforms of known defects, extract feature parameters, and establish a historical waveform library; based on the historical waveform library, calculate the classification range of feature parameters for each type of defect; determine whether the detection waveform parameters fall into the classification range corresponding to the processing noise, and if they match, remove the signal; Otherwise, keep it; Step 3: Based on the surface image of the hollow board, the surface of the hollow board is divided into uniform grids, and the density of the remaining ultrasonic detection sources in each grid is counted; the region growing algorithm is used to expand the search of adjacent areas with the high-density grid as the center, and merge continuous abnormal areas; the signal density ratio of the abnormal area is calculated, and if it exceeds the preset threshold, it is judged as a structural defect; otherwise, the hybrid model constructed by the convolutional neural network and the support vector machine is used to perform appearance detection on the input three-dimensional morphology data and output the location of the appearance defect.

Images