Aluminum foil strip defect detection method and system

Abstract

The invention discloses an aluminum foil strip defect detection method and system. The method comprises the following steps: acquiring multi-modal data acquired by an optical sensor and an ultrasonic sensor in a high-speed production line; performing mirror reflection preprocessing on the optical image, and eliminating ambient light interference by using an adaptive filter to obtain a first feature set; performing weighted fusion on the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; judging whether the second feature set contains defect features or not, and if the second feature set contains the defect features, classifying the fusion features through a convolutional neural network; generalizing the defect classification result by adopting a support vector machine to obtain a generalization model; the generalization model is in butt joint with a production line control system, a control instruction is generated according to the defect classification result, and when defects are detected, production line speed parameters are adjusted. According to the invention, the accuracy and robustness of defect detection are improved, intelligent defect detection and control of a high-speed production line are realized, and the product quality and the production efficiency are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of aluminum foil strip detection, and in particular discloses an aluminum foil strip defect detection method and system. Background Art

[0002] Aluminum foil defect detection is a crucial technology in modern manufacturing, particularly in areas such as pharmaceutical packaging, food packaging, and new energy batteries, where surface quality is directly linked to product safety and performance. High-quality aluminum foil must be free of tiny pinholes, scratches, or foreign matter, placing extremely high demands on detection technology. Traditional detection methods, which rely heavily on single-source optical imaging or manual spot checks, struggle to address the real-time recognition of complex surface features on high-speed production lines and are susceptible to interference from reflective materials, leading to missed detections and frequent false alarms. These limitations make existing solutions inadequate in terms of accuracy, efficiency, and adaptability to meet the demands of intelligent production in the context of Industry 4.0.

[0003] In the field of aluminum foil defect detection, the core challenge stems from the complexity of multimodal data acquisition and processing. In high-speed production lines, the specular reflection characteristics of the aluminum foil surface make optical imaging susceptible to interference from ambient light, making it difficult to stably capture the characteristic information of tiny defects. This imaging instability further makes it difficult for intelligent analysis algorithms to accurately distinguish new or invisible defects, such as fine cracks or embedded foreign matter, when classifying defects. The lack of generalization ability of the algorithm requires the system to frequently update training data when faced with aluminum foils of different materials or processes, increasing computational complexity and system maintenance costs. The interaction of the above factors ultimately manifests itself in the lack of synergy between the detection system and existing production line equipment, especially in high-speed dynamic environments, making it difficult to achieve low-latency closed-loop control.

[0004] Therefore, how to overcome mirror reflection interference, improve algorithm generalization capabilities, and achieve real-time collaboration with production line equipment through multimodal data fusion and intelligent algorithm optimization in high-speed production lines has become a key issue in breaking through aluminum foil defect detection technology. Summary of the Invention

[0005] The present invention provides an aluminum foil strip defect detection method and system, aiming to solve at least one defect of the above-mentioned prior art.

[0006] One aspect of the present invention relates to a method for detecting defects in aluminum foil strips, comprising the following steps: Acquire multimodal data collected by optical sensors and ultrasonic sensors in high-speed production lines and generate optical images; Perform specular reflection preprocessing on the optical image and use an adaptive filter to eliminate ambient light interference to obtain a first feature set; Performing weighted fusion on the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; Determine whether the second feature set contains defect features. If the second feature set contains defect features, classify the fused features using a convolutional neural network and output the defect classification results. Use support vector machine to generalize the defect classification results and obtain a generalized model; The generalized model is connected to the production line control system, control instructions are generated based on the defect classification results, and production line speed parameters are adjusted when defects are detected.

[0007] Furthermore, the steps of obtaining multimodal data collected by the optical sensor and the ultrasonic sensor in the high-speed production line and generating an optical image include: Acquire optical data from an optical sensor, acquire ultrasonic data from an ultrasonic sensor, synchronize the data with the ultrasonic data using a timestamp alignment technique, and obtain synchronized multimodal data; For the synchronized multimodal data, Gaussian filtering is used to perform denoising on the synchronized multimodal data to generate an optical image.

[0008] Furthermore, the steps of performing specular reflection preprocessing on the optical image and using an adaptive filter to eliminate ambient light interference to obtain a first feature set include: Acquire an optical image, use a reflection detection algorithm to determine whether there is a specular reflection area in the optical image, and if so, correct the specular reflection area through geometric transformation to obtain a corrected first image; For the first image, an adaptive filter is used to detect the pixel brightness of the first image. If the pixel brightness exceeds a preset threshold, the pixel value is adjusted through local mean filtering to obtain a second image with ambient light interference eliminated; According to the second image, a grayscale histogram of the second image is obtained, and a distribution of the grayscale histogram is adjusted by a histogram equalization technique to obtain a third image with enhanced contrast; For the third image, a convolutional neural network algorithm is used to extract spatial features and texture features of the third image to generate a first feature set.

[0009] Furthermore, the step of performing weighted fusion on the first feature set and the ultrasonic signal according to a preset weight matrix to obtain the second feature set includes: Acquire an ultrasonic signal, detect a noise level of the ultrasonic signal using an adaptive filter, and if the noise level exceeds a preset threshold, adjust the signal value using a local mean filter to obtain a corrected first ultrasonic signal; Extracting frequency features of the first ultrasonic signal using Fourier transform to generate a first frequency feature set; A preset weight matrix is used to perform weighted fusion on the first frequency feature set and the first feature set to obtain a second feature set.

[0010] Furthermore, determining whether the second feature set includes defect features, and if the second feature set includes defect features, classifying the fused features using a convolutional neural network, and outputting defect classification results includes: Obtaining a second feature set, using a preset threshold comparison method to determine whether the second feature set contains defect features, and obtaining a defect feature determination result; Determine whether the defect feature determination result contains the defect feature. If the defect feature determination result contains the defect feature, extract the fusion feature from the second feature set to generate a fusion feature subset. For the fused feature subset, a convolutional neural network is used to classify the fused feature subset to obtain the defect classification result; According to the defect classification results, the defect classification results are converted into structured data through preset mapping rules, and the final classification results are output.

[0011] Furthermore, a support vector machine is used to generalize the defect classification results. The steps of obtaining the generalization model include: Obtaining defect classification results, normalizing the defect classification results using a data preprocessing method, extracting a first feature subset from the defect classification results, and obtaining a normalized feature set; Determining whether the normalized feature set satisfies a preset threshold judgment condition; if the normalized feature set satisfies the preset threshold judgment condition, obtaining a second feature subset from the normalized feature set through a feature extraction algorithm to determine the second feature subset; For the second feature subset, a support vector machine is used to perform classification training on the second feature subset to generate a generalized classification model, thereby obtaining a generalized model.

[0012] Furthermore, the generalized model is connected to the production line control system, and control instructions are generated based on the defect classification results. When a defect is detected, the production line speed parameters are adjusted. The steps include: Acquire real-time detection data from the production line control system, normalize the real-time detection data using a data preprocessing method, extract a first data subset containing defect characteristics, and obtain a normalized data set; Determining whether the normalized data set satisfies a preset threshold judgment condition; if the normalized data set satisfies the preset threshold judgment condition, obtaining a second data subset from the normalized data set using a feature extraction algorithm to determine the second data subset; For the second data subset, a logistic regression algorithm is used to generate a defect classification result to obtain a defect classification result; According to the defect classification results, control instructions are generated through a pre-established instruction mapping table, and the production line speed parameters are adjusted to obtain the adjusted speed parameters.

[0013] Another aspect of the present invention relates to an aluminum foil defect detection system for implementing the above-mentioned aluminum foil defect detection method. The aluminum foil defect detection system comprises: The first acquisition module is used to acquire multimodal data collected by optical sensors and ultrasonic sensors in the high-speed production line and generate optical images; A second acquisition module is used to perform specular reflection preprocessing on the optical image and use an adaptive filter to eliminate ambient light interference to obtain a first feature set; A third acquisition module is used to perform weighted fusion of the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; A judgment module is used to judge whether the second feature set contains defect features. If the second feature set contains defect features, the fused features are classified by a convolutional neural network and a defect classification result is output; The fourth acquisition module is used to generalize the defect classification results using a support vector machine to obtain a generalization model; The adjustment module is used to connect the generalized model with the production line control system, generate control instructions based on the defect classification results, and adjust the production line speed parameters when defects are detected.

[0014] Furthermore, the first acquisition module includes: a first acquisition unit, configured to acquire optical data from the optical sensor, acquire ultrasonic data from the ultrasonic sensor, and synchronize the optical data with the ultrasonic data using a timestamp alignment technology to obtain synchronized multimodal data; The first generating unit is configured to perform denoising processing on the synchronized multimodal data by using Gaussian filtering to generate an optical image.

[0015] Furthermore, the second acquisition module includes: a second acquisition unit, configured to acquire an optical image, determine whether a specular reflection area exists in the optical image using a reflection detection algorithm, and if so, correct the specular reflection area through geometric transformation to obtain a corrected first image; a third acquisition unit, configured to detect pixel brightness of the first image using an adaptive filter, and if the pixel brightness exceeds a preset threshold, adjust the pixel value using a local mean filter to obtain a second image with ambient light interference eliminated; a fourth acquiring unit, configured to acquire a grayscale histogram of the second image based on the second image, and adjust the distribution of the grayscale histogram using a histogram equalization technique to obtain a third image with enhanced contrast; The second generating unit is used to extract the spatial features and texture features of the third image by using a convolutional neural network algorithm to generate a first feature set.

[0016] The beneficial effects achieved by the present invention are: The present invention provides a method and system for detecting defects in aluminum foil strips. The method obtains multimodal data from optical sensors and ultrasonic sensors, performs mirror reflection preprocessing and adaptive filtering on the optical image to eliminate ambient light interference, and then performs weighted fusion with the ultrasonic signal to obtain a fusion feature set. A convolutional neural network is used to classify the fusion features containing defect features, and a support vector machine is used to perform generalization processing to obtain a generalized model. Finally, the generalized model is connected to the production line control system, and control instructions are generated according to the defect classification results to achieve automatic adjustment of the production line speed. The present invention improves the accuracy and robustness of defect detection through multi-sensor data fusion and deep learning algorithms, realizes intelligent defect detection and control of high-speed production lines, and effectively improves product quality and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of an embodiment of a method for detecting defects in aluminum foil according to the present invention. DETAILED DESCRIPTION

[0018] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0019] like Figure 1 As shown, the first embodiment of the present invention provides a method for detecting defects in aluminum foil, comprising the following steps: Step S100: Acquire multimodal data collected by optical sensors and ultrasonic sensors in a high-speed production line to generate an optical image.

[0020] Multimodal data refers to complementary or interrelated datasets acquired in specific scenarios by a variety of heterogeneous sensors or acquisition methods. These datasets can include text, images, time series signals (such as mechanical parameters), and structured metrics (such as density and thickness). The core feature of multimodal data lies in its ability to achieve a comprehensive description and in-depth analysis of the target object or process through cross-modal fusion (such as the collaboration between optical imaging and mechanical testing) and multidimensional representation (spatial distribution + time series). In this embodiment, the multimodal data incorporates optical data detected by an optical sensor and ultrasonic data detected by an ultrasonic sensor.

[0021] Step S200: perform specular reflection preprocessing on the optical image, and use an adaptive filter to eliminate ambient light interference to obtain a first feature set.

[0022] Specular preprocessing refers to the technical process used in optical inspection or computer graphics to optimize specular effects in advance by adjusting light source configuration, material properties, or algorithm parameters to enhance subsequent analysis or rendering. The core goals of specular preprocessing include suppressing interfering reflections (e.g., highlight masking real defects), enhancing effective signals (e.g., defect feature extraction), and optimizing computational efficiency (e.g., precalculating reflection models).

[0023] An adaptive filter is an intelligent filtering system that dynamically adjusts its parameters based on input signal characteristics and preset optimization criteria (such as minimizing mean squared error). Its core goal is to adapt to signal or environmental changes by modifying filter coefficients in real time, thereby achieving tasks such as noise suppression, signal enhancement, and system identification. Unlike traditional fixed-parameter filters, its core features lie in its autonomous learning and dynamic response mechanisms, enabling optimal filtering without relying on prior statistical knowledge.

[0024] The "Primary Feature Set" (PFS) refers to the basic feature set extracted from raw data during the data modeling process. It typically serves as the initial input for subsequent feature engineering or model training. Its core goal is to capture key patterns in the data that are directly relevant to the task objectives and provide underlying support for higher-order feature construction (such as cross-features and statistical aggregation features). The construction of this PFS requires a balance between information content and computational efficiency to avoid redundant noise that interferes with model generalization performance.

[0025] Step S300: weighted fusion of the first feature set and the ultrasonic signal is performed according to a preset weight matrix to obtain a second feature set.

[0026] A weight matrix is a set of learnable parameters used in machine learning and deep learning models to represent the nonlinear mapping relationship between input features and output results. It is typically organized in matrix form, and its element values are dynamically adjusted through a backpropagation algorithm to minimize prediction error. Its mathematical essence is to achieve high-dimensional representation of data and extract complex patterns by combining linear transformations (matrix multiplication) with nonlinear activation functions. The Secondary Feature Set (SFS) is an enhanced feature set generated from the First Feature Set through methods such as cross-modal fusion, high-level combination, or domain knowledge infusion. It aims to improve the model's representation of complex patterns and its adaptability to tasks. Its core purpose is to overcome the limitations of a single modality or basic features, strengthening the model's reasoning and generalization performance through the integration of multi-granular information and semantic association.

[0027] Step S400: determine whether the second feature set contains defect features. If the second feature set contains defect features, classify the fused features through a convolutional neural network and output a defect classification result.

[0028] Default features are key parameters extracted from industrial products or image data to quantitatively describe appearance flaws or abnormal areas. Their core goal is to convert the physical properties of defects (such as shape, texture, and grayscale distribution) into differentiated metrics that can be processed by mathematical models, providing data support for defect classification, location, and severity assessment.

[0029] A convolutional neural network (CNN) is a deep learning architecture specifically designed for processing grid-like data (such as images, video, and audio). Its core concept is to automatically learn the spatial or temporal patterns of input data through local perception, weight sharing, and hierarchical feature extraction. Compared to fully connected neural networks, CNNs significantly reduce the number of parameters through convolution operations while enhancing robustness to geometric changes such as translation and rotation, making them a cornerstone model in the field of computer vision.

[0030] Defect classification results refer to standardized defect category labels and their corresponding confidence scores, ultimately generated through feature extraction and pattern recognition of abnormal regions within industrial products or image data. These results must clearly define defect type, severity, and spatial distribution, providing a quantitative basis for subsequent process optimization and quality decision-making.

[0031] Step S500: generalize the defect classification results using a support vector machine to obtain a generalized model.

[0032] The Support Vector Machine (SVM) is a supervised learning model based on statistical learning theory. Its core goal is to achieve classification or regression of data in high-dimensional space by constructing a maximum margin hyperplane. Its core concept is structural risk minimization, which aims to maximize the distance between the decision boundary and the nearest sample point (support vector) while ensuring classification accuracy, thereby improving the model's generalization ability.

[0033] Generalization processing refers to the technical process of ensuring that machine learning models maintain stable predictive performance on unknown data outside the training set through algorithm design or data optimization. Its core goal is to reduce the risk of overfitting the model to the training data and ensure that the learned patterns are universal, rather than simply memorizing noise or irrelevant details from specific samples.

[0034] A generalizable model is a machine learning model that, after specific design or optimization, maintains stable predictive performance on unseen data or data from a new distribution. Its core characteristic is that it suppresses overfitting of training data, capturing the underlying patterns in the data rather than noise, thereby achieving reliable reasoning across scenarios and tasks.

[0035] Step S600: Connect the generalized model to the production line control system, generate control instructions based on the defect classification results, and adjust the production line speed parameters when defects are detected.

[0036] Production line speed is a core metric for measuring production line efficiency. It refers to the number of products completed or the distance materials traveled per unit time, typically expressed in units such as "pieces / hour," "meters / minute," or "kilograms / hour." Its core function is to balance efficiency and quality by quantifying production pace, and to provide data support for production planning, equipment scheduling, and process optimization.

[0037] Furthermore, in the aluminum foil defect detection method provided in this embodiment, step S100 includes: Step S110: Acquire optical data from the optical sensor, acquire ultrasonic data from the ultrasonic sensor, synchronize the data with the ultrasonic data using a timestamp alignment technique, and obtain synchronized multimodal data; In a possible implementation, the scenario for acquiring optical data and ultrasonic data is set as aluminum foil defect detection.

[0038] The optical sensor captures optical data on the surface of the aluminum foil through a high-resolution camera, recording the surface texture, scratches or corrosion marks of the aluminum foil; the ultrasonic sensor obtains ultrasonic data inside the aluminum foil by emitting and receiving ultrasonic waves, reflecting internal cracks or voids.

[0039] Timestamp alignment technology ensures the synchronization of the two types of data. For example, by unifying the clock signal, the timestamp error between optical data and ultrasonic data can be controlled to the millisecond level, generating synchronized multimodal data.

[0040] This synchronization ensures the temporal and spatial consistency of subsequent analysis and improves the accuracy of defect detection.

[0041] Step S120 : Using Gaussian filtering to perform denoising processing on the synchronized multimodal data to generate an optical image.

[0042] Gaussian filtering is a linear smoothing filter based on the Gaussian function, widely used in image processing, signal denoising, and other fields. Its core concept is to suppress high-frequency noise in data through weighted averaging while preserving the structural characteristics of the original signal as much as possible.

[0043] In this embodiment, the technical chain of "precise synchronization → sub-modal denoising → cross-modal fusion → deep optimization" is used to solve the problems of multimodal data noise coupling and information loss. The generated optical image still has high fidelity in low-light and high-noise scenarios, and can meet the stringent visual requirements of industries such as industry, medicine, and autonomous driving.

[0044] Furthermore, in the aluminum foil defect detection method provided in this embodiment, step S200 includes: Step S210: Acquire an optical image, and use a reflection detection algorithm to determine whether there is a specular reflection area in the optical image. If so, correct the specular reflection area through geometric transformation to obtain a corrected first image.

[0045] For example, when processing optical images in a high-speed production line, a reflection detection algorithm is used to identify specular reflection areas.

[0046] Specular reflection is usually caused by the high glossiness of the aluminum foil surface, which causes the image to be locally too bright, affecting subsequent analysis.

[0047] The reflection detection algorithm analyzes pixel brightness gradients and reflective properties to determine whether there are areas of highlight. For example, when inspecting aluminum foil, a brightness threshold of 200 is set, and areas exceeding this threshold are marked as specular reflections. Geometric transformation correction adjusts the image perspective through rotation or perspective transformation to eliminate the effects of highlights and generate a corrected first image. This method ensures uniform image brightness distribution, providing a reliable foundation for subsequent processing.

[0048] Step S220: For the first image, an adaptive filter is used to detect the pixel brightness of the first image. If the pixel brightness exceeds a preset threshold, the pixel value is adjusted through local mean filtering to obtain a second image with ambient light interference eliminated.

[0049] In a possible implementation, the adaptive filter is used to detect pixel brightness of the first image.

[0050] Adaptive filters dynamically adjust filtering parameters based on local pixel characteristics to adapt to varying lighting conditions. For example, when inspecting the surface of aluminum foil, if the pixel brightness in a certain area exceeds a preset threshold of 180, a local mean filter calculates the average brightness of a 3×3 neighborhood, replaces the excessively high pixel values with the mean, and generates a second image. This step effectively eliminates ambient light interference, ensuring clear image details and laying the foundation for subsequent feature extraction.

[0051] Step S230: Obtain a grayscale histogram of the second image according to the second image, and adjust the distribution of the grayscale histogram using a histogram equalization technique to obtain a third image with enhanced contrast.

[0052] It should be noted that the histogram equalization technology is used to enhance the contrast of the second image.

[0053] The grayscale histogram reflects the distribution of pixel grayscale values. If the distribution is concentrated, the contrast is low.

[0054] Histogram equalization redistributes grayscale values to expand the dynamic range. For example, in an image of aluminum foil, the original grayscale values are concentrated between 50 and 100. Equalization expands the grayscale values to 0 to 255, generating a third image. This process improves the visibility of image textures, facilitating subsequent feature extraction.

[0055] Step S240: For the third image, use a convolutional neural network algorithm to extract spatial features and texture features of the third image to generate a first feature set.

[0056] Specifically, a convolutional neural network algorithm is used to extract spatial features and texture features of the third image.

[0057] Convolutional neural networks capture image information such as edges and textures through multiple layers of convolution and pooling. For example, in aluminum foil defect detection, the first layer uses 32 3×3 convolution kernels to extract low-level features, such as scratch edges. Subsequent layers generate high-level features through pooling operations, forming the first feature set. This feature set contains spatial location and texture information, providing data support for defect classification.

[0058] In one embodiment, the combination of reflectance detection and geometric transformation ensures the integrity of image preprocessing. For example, for highlights on the surface of aluminum foil, the reflectance detection algorithm first locates the highlights, then uses geometric transformation to adjust the viewing angle to produce an image with uniform illumination. Adaptive filtering and histogram equalization further optimize image quality, while a convolutional neural network extracts key features. This multi-step process, from image correction to feature extraction, forms a complete processing chain.

[0059] Advantageously, these techniques are highly adaptable in practical applications. For example, adaptive filters dynamically adjust parameters to address varying lighting conditions, while histogram equalization adapts to varying aluminum foil materials, ensuring robust feature extraction. Together, these methods enhance the robustness of image processing and provide high-quality data for subsequent analysis.

[0060] It's understandable that each technical step in aluminum foil defect detection is closely interdependent. For example, reflection correction provides a clear input for filtering, the filtered image provides uniform brightness for equalization, and the equalized image provides a high-contrast input for the neural network. This interconnected logical relationship ensures an efficient and reliable process from raw image to feature set, ultimately generating a first feature set that accurately reflects the spatial and textural characteristics of the aluminum foil surface.

[0061] Furthermore, in the aluminum foil defect detection method provided in this embodiment, step S300 includes: Step S310: Acquire an ultrasonic signal, use an adaptive filter to detect the noise level of the ultrasonic signal, and if the noise level exceeds a preset threshold, adjust the signal value through local mean filtering to obtain a corrected first ultrasonic signal.

[0062] For example, in aluminum foil defect detection, acquiring ultrasonic signals is a key step. Ultrasonic signals are emitted by an ultrasonic probe and reflected waves are received to detect defects inside or on the surface of the aluminum foil.

[0063] An ultrasonic probe transmits pulses at a fixed frequency, such as 5 MHz. The received signal contains information about the material's properties, but is often accompanied by noise. A signal acquisition device records the raw ultrasonic waveform, generating an initial data stream that forms the basis for subsequent processing.

[0064] In one possible implementation, an adaptive filter is used to detect the noise level of an ultrasound signal. The filter dynamically adjusts parameters based on the local characteristics of the signal to identify the noise intensity. For example, a noise threshold can be set at 20% of the signal amplitude. If the noise amplitude of a particular signal segment exceeds this threshold, the filter will initiate processing.

[0065] Local mean filtering calculates the average amplitude of a five-point neighborhood and replaces signal points with excessive noise to generate a corrected first ultrasonic signal. This process preserves the original characteristics of the signal while reducing noise interference, providing a clear input for frequency analysis.

[0066] Step S320: According to the first ultrasonic signal, frequency features of the first ultrasonic signal are extracted using Fourier transform to generate a first frequency feature set.

[0067] It should be noted that the Fourier transform is used to extract the frequency characteristics of the first ultrasonic signal. The Fourier transform converts the time-domain signal into the frequency domain, revealing the signal's frequency distribution. For example, in aluminum foil inspection, the Fourier transform analyzes the signal's spectrum, extracting the primary frequency component (e.g., 3 MHz) and the secondary frequency component (e.g., 6 MHz), generating a first frequency feature set. This feature set reflects the acoustic properties of the aluminum foil's internal structure and facilitates subsequent feature fusion.

[0068] Step S330: For the first frequency feature set and the first feature set, a preset weight matrix is used to perform weighted fusion on the first frequency feature set and the first feature set to obtain a second feature set.

[0069] Specifically, for the weighted fusion of the first frequency feature set and the first feature set, the preset weight matrix plays a key role.

[0070] The weight matrix allocates proportions according to the importance of the features, for example, the weight of frequency features is set to 0.6 and the weight of texture features is set to 0.4.

[0071] The fusion process integrates the two sets of features through matrix operations to generate a second feature set. For example, in aluminum foil defect detection, frequency features reflect the acoustic response of internal cracks, while texture features reflect the visual information of surface scratches. The second feature set generated by weighted fusion combines the advantages of both, improving the comprehensiveness of defect identification.

[0072] In one embodiment, weighted fusion is implemented to accommodate different inspection scenarios. For example, in a noisy environment, the weight of the frequency feature can be dynamically adjusted to 0.7 to enhance the detection of internal defects. This flexibility ensures the robustness of the fused feature set.

[0073] Ideally, each step of the above method is tightly linked. Ultrasonic signal correction provides high-quality input for frequency feature extraction, while the fusion of frequency and texture features forms a more comprehensive feature description. This interconnected logical relationship ensures an efficient and reliable process from signal acquisition to feature fusion.

[0074] It's no surprise that the combination of ultrasonic signals and optical image features is highly adaptable for aluminum foil inspection. For example, frequency features can reveal tiny internal cracks, while texture features capture surface scratches. The resulting fusion of these two creates a second feature set that provides rich data support for subsequent defect classification. This multi-dimensional feature integration strategy significantly improves inspection accuracy and reliability.

[0075] Furthermore, in the aluminum foil defect detection method provided in this embodiment, step S400 includes: Step S410: Obtain a second feature set, use a preset threshold comparison method to determine whether the second feature set contains defect features, and obtain a defect feature determination result.

[0076] For example, in aluminum foil defect detection, after obtaining the second feature set, it is necessary to determine whether it contains defect features.

[0077] The preset threshold comparison method is based on the distribution characteristics of feature values. For example, if the second feature set includes frequency and texture features, the frequency threshold is set to 3.5MHz, and the texture entropy threshold is set to 0.8. If a feature value exceeds the corresponding threshold, it is determined to contain a defect feature. This method ensures objectivity by quantifying the degree of feature abnormality.

[0078] Step S420 , determining whether the defect feature determination result includes defect features; if the defect feature determination result includes defect features, extracting fusion features from the second feature set to generate a fusion feature subset.

[0079] In one possible implementation, fused features are extracted from the second feature set to generate a fused feature subset. This fused feature subset focuses on core information related to the defect. For example, a subset is generated by extracting the amplitude of the main frequency of 3 MHz in the frequency features and regions with high entropy in the texture features. This subset reduces the complexity of subsequent processing by screening for highly relevant features while retaining the key information about the defect.

[0080] Step 430: For the fused feature subset, a convolutional neural network is used to classify the fused feature subset to obtain a defect classification result.

[0081] It should be noted that the convolutional neural network is used to classify the fused feature subset.

[0082] Convolutional neural networks extract characteristic patterns through multiple layers of convolution and pooling. For example, the network inputs a fused feature subset. The first convolution layer uses a 3x3 kernel to extract local features. The pooling layer compresses the data, and the final fully connected layer outputs a classification result, such as "crack," "scratch," or "no defect." This layered processing effectively captures the spatial and frequency domain characteristics of the feature subset.

[0083] Step S440: According to the defect classification result, the defect classification result is converted into structured data through a preset mapping rule, and the final classification result is output.

[0084] Specifically, the defect classification results need to be converted into structured data for output.

[0085] Pre-set mapping rules associate classification results with information such as defect type and location. For example, if the classification result is "crack," the mapping rules generate structured data containing the defect type, center coordinates, and confidence level, such as "crack, (x:50, y:30), 0.95." This format facilitates subsequent system integration and decision support.

[0086] In one embodiment, the threshold comparison method can be adaptively adjusted based on the detection environment. For example, in a noisy environment, the frequency feature threshold can be dynamically increased to 4 MHz to reduce false positives. This flexibility enhances the robustness of the method.

[0087] The convolutional neural network training process preferably takes into account the diversity of aluminum foil defects. For example, the training dataset includes a variety of crack and scratch samples to ensure the network's generalization ability to different defect patterns. This improves the accuracy of the classification results.

[0088] Understandably, the output of structured data is designed with practical application needs in mind. For example, the output coordinate information can be directly connected to automated equipment to guide precise repairs. This seamless process, from feature identification to result output, ensures efficient and practical testing.

[0089] Furthermore, the aluminum foil defect detection method provided in this embodiment includes step S500: Step S510: Obtain defect classification results, normalize the defect classification results using a data preprocessing method, extract a first feature subset from the defect classification results, and obtain a normalized feature set.

[0090] For example, in an aluminum foil defect detection scenario, after obtaining the classification results, they need to be normalized to ensure data consistency. For example, the classification results may include the defect type and confidence level, such as "scratch, 0.85."

[0091] Data preprocessing methods can remove outliers, such as results with a confidence score below 0.1, by normalizing the confidence score to a range of 0 to 1. This normalization can unify the data format and facilitate subsequent analysis.

[0092] In one possible implementation, a first feature subset is extracted from the classification results to generate a normalized feature set. For example, the defect type, confidence level, and grayscale value of the defect area are extracted from the classification results to form the first feature subset. Grayscale values may range from 0 to 255, and these values are normalized to 0 to 1 through a linear transformation to generate the normalized feature set. This approach focuses on key information and reduces the impact of redundant data.

[0093] Step S520: determine whether the normalized feature set meets the preset threshold judgment condition. If the normalized feature set meets the preset threshold judgment condition, obtain a second feature subset from the normalized feature set through a feature extraction algorithm to determine the second feature subset.

[0094] It should be noted that the normalized feature set must meet the preset threshold judgment conditions to determine whether to proceed further. For example, the confidence threshold is set to 0.7 and the grayscale feature mean threshold is set to 0.5.

[0095] If the normalized feature set meets certain criteria, such as a confidence level of 0.85 and a grayscale mean of 0.6, a feature extraction algorithm is used to obtain a second feature subset. This feature extraction algorithm can use principal component analysis to extract high-variance features, such as the edge strength of the defect area, to form the second feature subset. This screening method can highlight defect-related characteristics.

[0096] Step S530: For the second feature subset, use a support vector machine to perform classification training on the second feature subset to generate a generalized classification model to obtain a generalized model.

[0097] Specifically, a support vector machine (SVM) is used to train the second feature subset to generate a generalized classification model. For example, the second feature subset includes edge strength and regional entropy. The SVM constructs a hyperplane to classify the features into categories such as "cracks" and "scratches."

[0098] During training, labeled samples, such as 1,000 cracks and 800 scratches, are used to ensure the model learns diverse patterns. This classification training produces a model with generalization capabilities.

[0099] In one embodiment, the generation of a generalized classification model takes into account the complexity of aluminum foil defects. For example, by including samples under different lighting conditions in the training data, the model can adapt to the lighting variations in the production environment. This diverse training improves the robustness of the model.

[0100] Preferably, the feature extraction algorithm for the second feature subset can be dynamically adjusted based on the scenario. For example, in a high-noise environment, texture features are prioritized over grayscale features to reduce noise interference. This adaptive adjustment enhances the targeted nature of feature extraction.

[0101] It's easy to understand that the combination of normalization and feature extraction ensures logical coherence from classification results to generalized models. For example, the normalized feature set uses threshold filtering to reduce invalid data; the second feature subset focuses on highly relevant features to reduce training complexity; and the classification training of the support vector machine ensures model accuracy. This interconnected process ensures efficient and reliable defect detection.

[0102] Furthermore, in the aluminum foil defect detection method provided in this embodiment, step S600 includes: Step S610: Acquire real-time detection data from the production line control system, normalize the real-time detection data using a data preprocessing method, extract a first data subset containing defect features, and obtain a normalized data set.

[0103] For example, in the aluminum foil defect detection scenario, acquiring real-time inspection data from the production line control system involves raw data streams captured by sensors and cameras. For example, the camera generates 100 frames per second, each containing pixel grayscale values and defect location information. Data preprocessing methods must ensure the consistency of this data format.

[0104] In one possible implementation, preprocessing involves normalizing grayscale values, linearly transforming them from a range of 0 to 255 to a range of 0 to 1, while also removing noisy data, such as abnormal frames with large grayscale fluctuations. This normalization process unifies the data and facilitates subsequent feature extraction.

[0105] Specifically, when extracting the first data subset containing defect features, key information can be filtered from the normalized data. For example, the texture features and area ratio of the defect area can be extracted from each frame of the image to form the first data subset.

[0106] Texture features are generated using a local binary pattern algorithm, while area fraction is calculated as the ratio of defective pixels to total pixels. Assuming the defect area fraction in an image is 0.2 and the texture feature value is 0.75, these data constitute the first data subset. This approach focuses on the core attributes of the defect and reduces interference from irrelevant information.

[0107] Step S620: Determine whether the normalized data set meets the preset threshold judgment condition. If the normalized data set meets the preset threshold judgment condition, obtain a second data subset from the normalized data set through a feature extraction algorithm to determine the second data subset.

[0108] In one embodiment, the normalized data set must meet a preset threshold judgment condition. For example, the area ratio threshold is set to 0.1 and the texture feature threshold is set to 0.5. If the area ratio of a data set in the data set is 0.2 and the texture feature value is 0.75, the conditions are met and the next step of processing is performed.

[0109] It should be noted that the threshold setting needs to be adjusted according to the actual environment of the production line. For example, under high precision requirements, the threshold can be increased to screen out more significant defect data.

[0110] Preferably, when obtaining the second data subset through a feature extraction algorithm, a dimensionality reduction method, such as linear discriminant analysis, can be used to extract highly discriminative features from the first data subset. For example, the shape factor and contrast characteristics of the defect area can be extracted to generate the second data subset. Assuming a shape factor of 0.8 and a contrast of 0.6, these features can better characterize the defect type. This method improves classification efficiency by focusing on highly relevant features.

[0111] Step S630: For the second data subset, a logistic regression algorithm is used to generate a defect classification result to obtain a defect classification result.

[0112] For example, for the second data subset, a logistic regression algorithm is used to generate defect classification results.

[0113] Logistic regression classifies data into categories such as "bubble" and "wrinkle" by learning the mapping between features and defect categories. Labeled samples, such as 500 bubble samples and 400 wrinkle samples, are used during training to ensure the model learns diverse features. The defect classification result might be "bubble, confidence level 0.9," providing a basis for subsequent instruction generation.

[0114] Step S640: Based on the defect classification result, a control instruction is generated through a pre-established instruction mapping table, and the production line speed parameter is adjusted to obtain the adjusted speed parameter.

[0115] In one possible implementation, a control instruction is generated based on the defect classification results using a pre-established instruction mapping table. For example, the mapping table specifies that if a bubble defect is detected with a confidence level higher than 0.8, the production line speed is reduced by 10%.

[0116] If the classification result is "bubble, 0.9," an instruction is generated to adjust the speed from 1 m / s to 0.9 m / s. This mapping mechanism enables rapid response to defect situations.

[0117] Understandably, the adjusted speed parameters must be fed back to the production line control system in real time. For example, after the speed is adjusted, the system uses sensors to monitor the subsequent defect rate. If the defect rate decreases, the effectiveness of the instruction is verified. This closed-loop mechanism ensures stable and efficient production line operation.

[0118] The present invention relates to an aluminum foil defect detection system, which is used to implement the above-mentioned aluminum foil defect detection method. The aluminum foil defect detection system includes a first acquisition module, a second acquisition module, a third acquisition module, a judgment module, a fourth acquisition module and an adjustment module, wherein the first acquisition module is used to acquire multimodal data collected by an optical sensor and an ultrasonic sensor in a high-speed production line to generate an optical image; the second acquisition module is used to perform mirror reflection preprocessing on the optical image, and adopt an adaptive filter to eliminate ambient light interference to obtain a first feature set; the third acquisition module is used to perform weighted fusion of the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; the judgment module is used to determine whether the second feature set contains defect features. If the second feature set contains defect features, the fused features are classified by a convolutional neural network and a defect classification result is output; the fourth acquisition module is used to generalize the defect classification result by adopting a support vector machine to obtain a generalized model; the adjustment module is used to connect the generalized model with the production line control system, generate control instructions according to the defect classification result, and adjust the production line speed parameters when a defect is detected.

[0119] Furthermore, in the aluminum foil defect detection system provided in this embodiment, the first acquisition module includes a first acquisition unit and a first generation unit, wherein the first acquisition unit is used to acquire optical data from the optical sensor and ultrasonic data from the ultrasonic sensor, and synchronize the data with the ultrasonic data through timestamp alignment technology to obtain synchronized multimodal data; the first generation unit is used to use Gaussian filtering to denoise the synchronized multimodal data to generate an optical image.

[0120] Preferably, in the aluminum foil defect detection system provided in this embodiment, the second acquisition module includes a second acquisition unit, a third acquisition unit, a fourth acquisition unit and a second generation unit, wherein the second acquisition unit is used to acquire an optical image, and use a reflection detection algorithm to determine whether there is a mirror reflection area in the optical image. If so, the mirror reflection area is corrected by geometric transformation to obtain a corrected first image; the third acquisition unit is used to detect the pixel brightness of the first image using an adaptive filter, and if the pixel brightness exceeds a preset threshold, the pixel value is adjusted by local mean filtering to obtain a second image with ambient light interference eliminated; the fourth acquisition unit is used to acquire a grayscale histogram of the second image based on the second image, and adjust the distribution of the grayscale histogram by histogram equalization technology to obtain a third image with enhanced contrast; the second generation unit is used to extract spatial features and texture features of the third image using a convolutional neural network algorithm to generate a first feature set.

[0121] The present embodiment provides an aluminum foil strip defect detection method and system. Compared with the existing technology, by acquiring multimodal data from optical sensors and ultrasonic sensors, the optical image is subjected to mirror reflection preprocessing and adaptive filtering to eliminate ambient light interference, and then weighted fusion is performed with the ultrasonic signal to obtain a fusion feature set. A convolutional neural network is used to classify the fusion features containing defect features, and a support vector machine is used for generalization processing to obtain a generalized model. Finally, the generalized model is connected to the production line control system, and control instructions are generated according to the defect classification results to achieve automatic adjustment of the production line speed. This embodiment improves the accuracy and robustness of defect detection through multi-sensor data fusion and deep learning algorithms, realizes intelligent defect detection and control of high-speed production lines, and effectively improves product quality and production efficiency.

[0122] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.

Claims

1. A method for detecting defects in aluminum foil, characterized in that: The following steps are involved: Acquire multimodal data collected by optical sensors and ultrasonic sensors in high-speed production lines and generate optical images; Performing specular reflection preprocessing on the optical image, and using an adaptive filter to eliminate ambient light interference to obtain a first feature set; Performing weighted fusion on the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; Determining whether the second feature set includes defect features, and if so, classifying the fused features using a convolutional neural network and outputting a defect classification result; Using a support vector machine to generalize the defect classification results to obtain a generalized model; The generalized model is connected to the production line control system, and control instructions are generated according to the defect classification results. When defects are detected, the production line speed parameters are adjusted.

2. The aluminum foil defect detection method according to claim 1, wherein: The step of acquiring multimodal data collected by the optical sensor and the ultrasonic sensor in the high-speed production line and generating an optical image includes: Acquire optical data from an optical sensor, acquire ultrasonic data from an ultrasonic sensor, and synchronize the optical data with the ultrasonic data using a timestamp alignment technology to obtain synchronized multimodal data; For the synchronized multimodal data, Gaussian filtering is used to perform denoising on the synchronized multimodal data to generate an optical image.

3. The aluminum foil defect detection method according to claim 1, wherein: The steps of performing specular reflection preprocessing on the optical image and using an adaptive filter to eliminate ambient light interference to obtain a first feature set include: Acquire an optical image, use a reflection detection algorithm to determine whether there is a specular reflection area in the optical image, and if so, correct the specular reflection area through geometric transformation to obtain a corrected first image; For the first image, an adaptive filter is used to detect the pixel brightness of the first image. If the pixel brightness exceeds a preset threshold, the pixel value is adjusted through local mean filtering to obtain a second image with ambient light interference eliminated. Obtaining a grayscale histogram of the second image based on the second image, and adjusting the distribution of the grayscale histogram using a histogram equalization technique to obtain a third image with enhanced contrast; For the third image, a convolutional neural network algorithm is used to extract spatial features and texture features of the third image to generate a first feature set.

4. The aluminum foil defect detection method according to claim 1, wherein: The step of performing weighted fusion of the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set includes: Acquire an ultrasonic signal, detect a noise level of the ultrasonic signal using an adaptive filter, and if the noise level exceeds a preset threshold, adjust the signal value using a local mean filter to obtain a corrected first ultrasonic signal; Extracting frequency features of the first ultrasonic signal using Fourier transform according to the first ultrasonic signal to generate a first frequency feature set; A preset weight matrix is used to perform weighted fusion on the first frequency feature set and the first feature set to obtain a second feature set.

5. The aluminum foil defect detection method according to claim 1, wherein: The step of determining whether the second feature set includes defect features, and if the second feature set includes defect features, classifying the fused features using a convolutional neural network and outputting defect classification results includes: Obtaining the second feature set, and using a preset threshold comparison method to determine whether the second feature set contains defect features, thereby obtaining a defect feature determination result; determining whether the defect feature determination result includes a defect feature; if the defect feature determination result includes a defect feature, extracting a fusion feature from the second feature set to generate a fusion feature subset; For the fused feature subset, a convolutional neural network is used to classify the fused feature subset to obtain a defect classification result; According to the defect classification result, the defect classification result is converted into structured data through a preset mapping rule, and a final classification result is output.

6. The aluminum foil defect detection method according to claim 1, wherein: The step of generalizing the defect classification result using a support vector machine to obtain a generalized model includes: Obtaining the defect classification result, normalizing the defect classification result using a data preprocessing method, extracting a first feature subset from the defect classification result to obtain a normalized feature set; Determining whether the normalized feature set satisfies a preset threshold judgment condition; if the normalized feature set satisfies the preset threshold judgment condition, obtaining a second feature subset from the normalized feature set by a feature extraction algorithm to determine the second feature subset; For the second feature subset, a support vector machine is used to perform classification training on the second feature subset to generate a generalized classification model to obtain a generalized model.

7. The aluminum foil defect detection method according to claim 1, wherein: The steps of interfacing the generalized model with the production line control system, generating control instructions based on the defect classification results, and adjusting the production line speed parameters when a defect is detected include: Acquiring real-time detection data from a production line control system, normalizing the real-time detection data using a data preprocessing method, extracting a first data subset containing defect characteristics, and obtaining a normalized data set; Determining whether the normalized data set satisfies a preset threshold judgment condition; if the normalized data set satisfies the preset threshold judgment condition, obtaining a second data subset from the normalized data set by using a feature extraction algorithm to determine the second data subset; For the second data subset, a logistic regression algorithm is used to generate a defect classification result to obtain a defect classification result; According to the defect classification result, a control instruction is generated through a pre-established instruction mapping table, and the production line speed parameter is adjusted to obtain the adjusted speed parameter.

8. An aluminum foil defect detection system, used to implement the aluminum foil defect detection method according to any one of claims 1 to 7, characterized in that: The aluminum foil defect detection system includes: The first acquisition module is used to acquire multimodal data collected by optical sensors and ultrasonic sensors in the high-speed production line and generate optical images; a second acquisition module, configured to perform specular reflection preprocessing on the optical image, and adopt an adaptive filter to eliminate ambient light interference to obtain a first feature set; a third acquisition module, configured to perform weighted fusion of the first feature set and the ultrasonic signal according to a preset weight matrix to obtain a second feature set; a judgment module, configured to judge whether the second feature set includes defect features, and if so, classify the fused features using a convolutional neural network and output a defect classification result; A fourth acquisition module is used to generalize the defect classification result using a support vector machine to obtain a generalization model; An adjustment module is used to connect the generalized model to the production line control system, generate control instructions based on the defect classification results, and adjust the production line speed parameters when a defect is detected.

9. The aluminum foil defect detection system according to claim 8, characterized in that: The first acquisition module includes: a first acquisition unit, configured to acquire optical data from the optical sensor, acquire ultrasonic data from the ultrasonic sensor, and synchronize the optical data with the ultrasonic data using a timestamp alignment technology to obtain synchronized multimodal data; The first generating unit is configured to perform denoising processing on the synchronized multimodal data by using Gaussian filtering to generate an optical image.

10. The aluminum foil defect detection system according to claim 8, characterized in that: The second acquisition module includes: a second acquisition unit, configured to acquire an optical image, determine whether a specular reflection area exists in the optical image using a reflection detection algorithm, and if so, correct the specular reflection area through geometric transformation to obtain a corrected first image; a third acquisition unit, configured to detect pixel brightness of the first image using an adaptive filter, and if the pixel brightness exceeds a preset threshold, adjust the pixel value using a local mean filter to obtain a second image with ambient light interference eliminated; a fourth acquiring unit, configured to acquire a grayscale histogram of the second image based on the second image, and adjust the distribution of the grayscale histogram using a histogram equalization technique to obtain a third image with enhanced contrast; The second generating unit is configured to extract spatial features and texture features of the third image by using a convolutional neural network algorithm to generate a first feature set.

Images