A visual inspection method and system for PTC starter production

Abstract

This invention discloses a visual inspection method and system for PTC starter production, relating to the field of automated visual inspection. It involves acquiring original images of the PTC starter to be inspected, enhancing them to generate standardized inspection images, extracting geometric structure and surface texture features based on a multi-scale decomposition algorithm to construct multi-dimensional feature vectors, inputting these vectors into a deep belief network model, and outputting defect classification results and location information. Based on the defect classification results, it generates control commands for qualified release, unqualified rejection, or adaptive adjustment of process parameters. This invention improves the accuracy and intelligence of PTC starter defect detection, standardizes defect judgment, promotes the linkage between inspection and production processes, and adapts to the online quality control needs of large-scale automated PTC starter production.

Description

Technical Field

[0001] This invention relates to the field of PTC starter manufacturing technology, specifically to a visual inspection method and system for PTC starter manufacturing. Background Technology

[0002] PTC starters are core starting components for refrigeration equipment compressors. Their core structure includes a PTC thermistor chip, metal electrode plates, conductive pins, and an insulating shell. Even minor defects during the manufacturing process can lead to compressor starting failure, leakage, or even burnout. Therefore, online high-precision defect detection is a core aspect of PTC starter production quality control.

[0003] Currently, the production and testing technology for PTC starters still faces many technical challenges and cannot meet the demands of large-scale automated production. Specifically:

[0004] Traditional manual inspection is inefficient and has a high rate of false positives and false negatives: it relies on the operator's experience for visual judgment, which is poor at identifying minute defects such as chip microcracks and electrode micro-soldering, with a false positive and false negative rate as high as 8%-12%, and the efficiency of a single manual inspection is only 50 pieces / minute, which is far lower than the production line speed.

[0005] Conventional visual inspection features are limited: Existing visual inspection often uses simple geometric contour matching, which only extracts the geometric structure features of the workpiece without considering surface texture features. It cannot distinguish between poor welding of electrode sheets and surface stains, and its accuracy in identifying subtle defects is insufficient.

[0006] Intelligent detection models lack quantitative judgment and self-updating capabilities: Some existing technologies use simple neural networks for defect identification, but there is no unified standard for quantitative scoring of defects. They only make qualitative judgments, resulting in poor consistency of results. Moreover, the models cannot be updated online after training, and their generalization ability to new defects that occur in the production process is weak. They need to be retrained offline, resulting in poor adaptability.

[0007] The testing and production processes are disconnected and lack closed-loop linkage: the existing testing technology only realizes the "release of qualified and rejection of unqualified" in the back-end, without feeding back the test results to the front-end production process. It is impossible to adjust the welding, assembly and other process parameters for repairable defects, resulting in the repeated occurrence of the same defects and making it difficult to improve the production yield.

[0008] Poor image preprocessing results are highly susceptible to environmental interference: uneven lighting, dust, equipment vibration, and other factors in the production line can cause noise, reflection, and brightness deviation in the acquired images. Existing preprocessing methods only use single filtering or lighting correction, which cannot effectively eliminate interference and affect the accuracy of subsequent feature extraction and defect identification.

[0009] To address the aforementioned technical challenges, there is an urgent need in this field for a visual inspection method and system for PTC starter production that integrates multi-dimensional feature extraction, quantitative defect assessment, online model updates, and closed-loop linkage between inspection and process, in order to achieve high-precision, high-efficiency, and intelligent online quality control. Summary of the Invention

[0010] This invention proposes a visual inspection method and system for PTC starter production to solve the technical problems mentioned in the background art.

[0011] To address the aforementioned technical problems, in a first aspect, this invention proposes a visual inspection method for PTC starter production, the specific technical solution of which is as follows:

[0012] A visual inspection method for PTC starter manufacturing includes:

[0013] Acquire the original image of the PTC starter to be tested;

[0014] The original image is enhanced to generate a standardized image to be inspected;

[0015] Based on a preset multi-scale decomposition algorithm, geometric structure features and surface texture features in the image to be inspected are extracted, and a multi-dimensional feature vector is constructed.

[0016] The multidimensional feature vector is input into a trained deep belief network model, which outputs defect classification results and defect location information.

[0017] Based on the defect classification results, corresponding control instructions are generated. The control instructions include at least a qualified release instruction, a non-qualified rejection instruction, or a process parameter adaptive adjustment instruction.

[0018] As a further improvement to the technical solution of the present invention, the step of generating the image to be inspected specifically includes:

[0019] The original image is subjected to illumination uniformity correction based on a two-dimensional gamma function;

[0020] A hybrid filtering algorithm combining adaptive median filtering and bilateral filtering is used to remove noise from the corrected image; and

[0021] Based on the physical contour template of the PTC starter, the target area containing the electrode sheet, PTC thermistor chip, and pin soldering surface is segmented by the template matching algorithm.

[0022] As a further improvement to the technical solution of the present invention, the step of extracting the multidimensional feature vector specifically includes:

[0023] The image to be inspected is decomposed into low-frequency approximation components and multiple high-frequency detail components based on wavelet transform.

[0024] Extract the gray-level co-occurrence matrix eigenvalues ​​of the low-frequency approximation components, wherein the gray-level co-occurrence matrix eigenvalues ​​include contrast, energy, and correlation;

[0025] Extract the local binary pattern texture spectral features of the high-frequency detail components; and

[0026] The gray-level co-occurrence matrix eigenvalues ​​are fused with the local binary pattern texture spectrum features to construct the multidimensional feature vector.

[0027] As a further improvement to the technical solution of this invention, the step of outputting the defect classification result specifically includes:

[0028] The multidimensional feature vector is input into a deep belief network, which includes multiple restricted Boltzmann machines stacked sequentially and a backpropagation classification layer;

[0029] The defect confidence score is calculated based on the multidimensional feature vector, and the defect classification result is confirmed based on the comparison result of the defect confidence score and the preset threshold.

[0030] As a further improvement to the technical solution of the present invention, the defect confidence score S is calculated using the following formula:

[0031] ;

[0032] in, For the multidimensional feature vector, and The weight matrix and bias terms of the backpropagation classification layer are, Output weight coefficients for deep networks. These are the geometric constraint weighting coefficients. Let be the prior weight of the i-th geometric feature. The radial basis function deviation value is constructed based on the aforementioned geometric structural features. This represents the total number of geometric features involved in the calculation.

[0033] As a further improvement to the technical solution of the present invention, the step of generating corresponding control instructions based on the defect classification results specifically includes:

[0034] When the defect classification result is a qualified product, the qualified release instruction is generated;

[0035] When the defect classification result is an unrepairable defect, the non-compliance rejection instruction is generated and sent to the rejection device;

[0036] When the defect classification result belongs to the preset repairable category, the process parameter adaptive adjustment instruction is generated and sent to the production execution system. The process parameter adaptive adjustment instruction is used to adjust the front-end welding process parameters or assembly pressure parameters.

[0037] As a further improvement to the technical solution of the present invention, it also includes:

[0038] The false detection samples and newly added defect samples collected during the production process are used as incremental data to perform online incremental learning and updates on the deep belief network model.

[0039] Secondly, this invention proposes a visual inspection system for PTC starter production, comprising:

[0040] The image acquisition module is used to acquire raw images of the PTC starter under test.

[0041] An image preprocessing module is used to enhance the original image to generate a standardized image to be inspected;

[0042] The feature extraction module is used to extract geometric structure features and surface texture features from the image to be inspected based on a preset multi-scale decomposition algorithm, and construct a multi-dimensional feature vector;

[0043] The defect identification module is used to input the multidimensional feature vector into a trained deep belief network model and output defect classification results and defect location information; and

[0044] The hierarchical control module is used to generate corresponding control instructions based on the defect classification results. The control instructions include at least a qualified release instruction, a non-qualified rejection instruction, or a process parameter adaptive adjustment instruction.

[0045] As a further improvement to the technical solution of the present invention, the image preprocessing module includes:

[0046] An illumination compensation unit is used to perform illumination uniformity correction on the original image based on a two-dimensional gamma function.

[0047] A filtering and noise reduction unit is used to remove noise from the corrected image using a hybrid filtering algorithm combining adaptive median filtering and bilateral filtering; and

[0048] The region of interest extraction unit is used to segment the target region containing the electrode sheet, PTC thermistor chip, and pin bonding surface based on the physical contour template of the PTC starter using a template matching algorithm.

[0049] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the visual inspection method for PTC starter production as described in any one of claims 1 to 7.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] This invention achieves image standardization by enhancing the original image of the PTC starter. It relies on a multi-scale decomposition algorithm to accurately extract geometric structures and surface texture features from the image and construct a multi-dimensional feature vector. Combined with a deep belief network model, it can efficiently output accurate defect classification results and defect location information. Furthermore, it can generate corresponding pass / fail release, fail / reject, or parameter adaptive adjustment control commands based on the defect judgment results. Overall, it realizes the intelligent and standardized production inspection of PTC starters, effectively improving the accuracy of defect detection and the scientific nature of judgment. Simultaneously, it achieves linkage between inspection results and production execution, solving the problems of strong subjectivity, inaccurate defect identification, and disconnect between inspection and production processes in traditional inspection methods. It is suitable for the online quality control needs of automated PTC starter production. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the visual inspection method for PTC starter production according to the present invention.

[0054] Figure 2 This is a schematic diagram of the PTC starter production visual inspection system of the present invention. Detailed Implementation

[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] like Figure 1-2As shown, this invention proposes a visual inspection method for PTC starter production. The method involves image acquisition → standardized preprocessing → multi-scale feature fusion extraction → deep belief network quantization recognition → hierarchical control command generation → online incremental model learning, forming a closed-loop detection process. The specific steps are as follows:

[0057] Original image acquisition: The PTC starter to be tested on the production line conveyor belt is captured by an industrial camera. The relative position of the camera and the workpiece is kept fixed to ensure that the acquired image can clearly capture the features of the core detection areas such as the electrode plate, PTC thermistor chip, and pin soldering surface.

[0058] Image enhancement processing: The original image undergoes multi-step standardization preprocessing to generate the image to be inspected. Specifically, this includes: performing illumination homogenization correction on the original image based on a two-dimensional gamma function to eliminate uneven illumination and reflection interference; using a hybrid filtering algorithm of adaptive median filtering and bilateral filtering to remove salt-and-pepper noise and Gaussian noise from the corrected image while preserving image edge features; and using a template matching algorithm based on the physical contour template of the PTC starter to segment the target area containing the electrode sheet, PTC thermistor chip, and pin soldering surface, eliminating interference from irrelevant areas such as the background and conveyor belt.

[0059] Multidimensional feature vector construction: Based on a preset multi-scale decomposition algorithm, geometric structure features and surface texture features of the image to be inspected are extracted and fused to construct a multidimensional feature vector. Specifically, this includes: decomposing the image to be inspected into low-frequency approximate components and multiple high-frequency detail components based on wavelet transform; extracting the gray-level co-occurrence matrix feature values ​​(contrast, energy, correlation) of the low-frequency approximate components as geometric structure features; extracting the local binary pattern texture spectrum features of the high-frequency detail components as surface texture features; and fusing the two types of features after normalization to construct a multidimensional feature vector.

[0060] Defect Quantification and Identification: A multi-dimensional feature vector is input into a trained deep belief network model, outputting defect classification results and defect location information. Specifically, the deep belief network consists of multiple stacked restricted Boltzmann machines and a backpropagation classification layer, achieving layer-by-layer dimensionality reduction and non-linear mapping of features; a defect confidence score is calculated using a preset defect confidence score formula; based on the comparison between the score and a preset threshold, the defect classification result (qualified product / specific defect type) is confirmed; and the specific location information of the defect is output through the mapping of pixel coordinates to physical coordinates.

[0061] Hierarchical control instruction generation: Based on the defect classification results, corresponding control instructions are generated. The control instructions include at least qualified release instructions, unqualified rejection instructions, and process parameter adaptive adjustment instructions, realizing closed-loop linkage with the production line: qualified products generate qualified release instructions to control the flow of workpieces to subsequent processes; unrepairable defects generate unqualified rejection instructions, which are sent to the rejection device to achieve accurate rejection of unqualified products; repairable defects generate process parameter adaptive adjustment instructions, which are sent to the production execution system to adjust the process parameters of front-end welding, assembly, etc.

[0062] Online incremental learning of the model: The false detection samples and newly added defect samples collected during the production process are used as incremental data to update the deep belief network model online. Only the weights of the backpropagation classification layer of the model are fine-tuned, without changing the weights of the restricted Boltzmann machine, so as to realize the model's ability to identify newly added defects without offline retraining.

[0063] To implement the above detection method, this invention proposes a PTC starter production video inspection system, including an image acquisition module, an image preprocessing module, a feature extraction module, a defect identification module, and a hierarchical control module. The functions of each module correspond one-to-one with the steps of the detection method. The specific structure and functions are as follows:

[0064] Image acquisition module: This is the system's image input unit, used to acquire raw images of the PTC starter to be tested. The core hardware includes an industrial camera, an industrial lens, and an adjustable fill light. The brightness of the fill light can be adjusted according to the lighting environment of the production line to ensure the clarity of the raw image acquisition. The acquired image data is transmitted to the image preprocessing module.

[0065] Image preprocessing module: Electrically connected to the image acquisition module, it is used to enhance the original image to generate a standardized image to be inspected. This module further includes an illumination compensation unit, a filtering and noise reduction unit, and a region of interest extraction unit connected in sequence, which respectively realize illumination uniformity correction, hybrid filtering and noise reduction, and target region segmentation in the detection method. The processed image to be inspected is transmitted to the feature extraction module.

[0066] Feature extraction module: Electrically connected to the image preprocessing module, it is used to extract the geometric structure features and surface texture features of the image to be inspected based on a preset multi-scale decomposition algorithm, construct a multi-dimensional feature vector, and use a hardware acceleration chip to realize high-speed operation of the feature extraction algorithm. The constructed multi-dimensional feature vector is transmitted to the defect recognition module.

[0067] Defect recognition module: Electrically connected to the feature extraction module, it is the core intelligent recognition unit of the system. It has a built-in trained deep confidence network model, which is used to input multi-dimensional feature vectors into the model and output defect classification results and defect location information. This module is equipped with an embedded AI computing platform to realize the rapid calculation of defect confidence scores and accurate defect recognition. The recognition results are transmitted to the hierarchical control module.

[0068] Hierarchical control module: Electrically connected to the defect identification module, it is the system's control output unit. It is used to generate and send corresponding control commands based on the defect classification results. The core hardware is an industrial PLC controller with multiple digital / analog output interfaces. It can be directly connected to the production line's conveyor belt control system, rejection device, and production execution system. At the same time, it records all detection data to achieve production quality control traceability.

[0069] The modules of this system adopt an industrial-grade communication protocol, with image data transmission latency ≤0.05s and control command response time ≤0.1s, ensuring the real-time performance and stability of the system and adapting to the continuous production needs of the production line.

[0070] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the visual inspection method for PTC starter production. This computer-readable storage medium includes, but is not limited to, ROM, RAM, CD-ROM, magnetic disk, and floppy disk.

[0071] The present invention will be further described below with reference to embodiments:

[0072] Example 1:

[0073] This embodiment, combining core algorithm steps and key formula calculation examples, elaborates on the entire execution process of the detection method, adapting to the online detection of standard PTC starters (pin center distance 10mm, electrode area 15mm²), specifically including steps S1~S6:

[0074] Step S1: Acquisition of raw images

[0075] Employing a 12-megapixel area array industrial camera, paired with a 16mm fixed-focus industrial lens and a ring LED fill light, the camera lens is perpendicular to the PTC starter detection surface at a height of 200mm. The fill light brightness is adjusted to 800lm, the acquisition frame rate is 30 frames / second, the image resolution is 4096×3072, and the acquisition format is 8-bit grayscale. Continuous acquisition of images of PTC starters on the production line conveyor belt ensures clear capture of the subtle features of the electrode plates, PTC thermistor chips, and pin soldering surfaces. The acquired raw images are transmitted to subsequent modules via a USB 3.0 interface.

[0076] Step S2: Image enhancement processing to generate a standardized image to be inspected.

[0077] Illumination homogenization correction: Illumination compensation of the original image is performed based on a two-dimensional gamma function. The expression for the two-dimensional gamma function is: ,in The corrected grayscale value is γ, which is the correction coefficient. Depending on the lighting environment of the production line, the value of γ ranges from 0.8 to 1.2. When the image is locally too bright, γ < 1 is used, and when the image is locally too dark, γ > 1 is used. This can achieve uniform image lighting and eliminate reflections and brightness deviations.

[0078] Hybrid filtering noise reduction: First, adaptive median filtering is used to eliminate salt-and-pepper noise caused by dust. The filtering window is adaptively adjusted from 3×3 to 7×7 according to the degree of pixel noise. Then, bilateral filtering is used to eliminate Gaussian noise caused by equipment vibration. The spatial kernel standard deviation is set to 2, the grayscale kernel standard deviation is set to 50, and the image edge features are preserved.

[0079] Target region segmentation: Construct a standard physical contour template for a PTC starter, calculate the correlation coefficient between the image to be processed and the template using a normalized cross-correlation template matching algorithm, and determine the region with a correlation coefficient ≥ 0.95 as the target region. Segment the region containing the electrode sheet, PTC thermistor chip, and pin soldering surface, and exclude irrelevant regions such as the conveyor belt and background to generate a standardized image to be inspected.

[0080] Step S3: Multi-scale feature extraction, constructing multi-dimensional feature vectors

[0081] Wavelet transform decomposition: The image to be examined is decomposed into three levels at multiple scales using the db4 wavelet basis function, resulting in a low-frequency approximate component. and 6 high-frequency detail components ( );

[0082] Geometric structure feature extraction: for Generate a gray-level co-occurrence matrix (step size 1, orientation 0° / 45° / 90° / 135°, gray level 64), extract three feature values: contrast, energy, and correlation, and take the average value of the four orientations to obtain the 3D geometric structure features;

[0083] Surface texture feature extraction: The texture spectrum features are extracted from the six high-frequency detail components using the 8-neighbor circular LBP operator (radius 1). Each component generates a 256-dimensional feature vector, which is then concatenated to obtain a 1536-dimensional surface texture feature.

[0084] Feature fusion: The two types of features are mapped to the [0,1] interval through min-max normalization and then concatenated and fused into a 1539-dimensional multidimensional feature vector.

[0085] Step S4: The model consists of three Restricted Boltzmann Machines (RBMs) stacked sequentially and one Backpropagation (BP) classification layer. The number of neurons in the three RBMs are 1539, 1024, and 512 respectively, achieving layer-by-layer dimensionality reduction and nonlinear mapping of features. The Backpropagation classification layer is a fully connected layer with 8 neurons, corresponding to 7 types of defects in PTC starters (chip cracking, electrode soldering failure, pin misalignment, solder surface contamination, shell deformation, pin breakage, and chip corner missing) and 1 type of qualified product.

[0086] 2. Model Pre-training and Fine-tuning: First, unsupervised learning was used to pre-train the three restricted Boltzmann machines layer by layer. The training samples were 100,000 labeled PTC starter image feature vectors, with a learning rate of 0.01, a batch size of 128, and 500 iterations. Then, supervised learning was used to fine-tune the entire deep belief network, using cross-entropy as the loss function, Adam as the optimizer, a learning rate of 0.001, and 300 iterations, so that the model's recognition accuracy converged to over 99.9%.

[0087] 3. Defect confidence score: The score S is calculated using the following formula:

[0088] ;

[0089] in, For the 1539-dimensional feature vector, This is the weight vector (dimension 512×8) for the backpropagation classification layer. This is the weight vector (dimension 512×8) for the backpropagation classification layer. The weight coefficients are output for the deep network (0.7 in this example). (The geometric constraint weighting coefficient is set to 0.3 in this embodiment). The prior weights (contrast) of the i-th geometric feature =0.4, energy =0.3, Correlation =0.3), The deviation value of the radial basis function constructed based on the geometric structural features (in this embodiment, it is the absolute difference between the measured geometric features and the standard geometric features). The total number of geometric features involved in the calculation (n=3). It is a Sigmoid activation function with an output value of [0,1].

[0090] Formula calculation example: Taking the electrode sheet cold solder joint defect as an example, the measured multi-dimensional feature vector is mapped by a deep network. =0.92, and the geometric feature deviation measures are respectively =0.08、 =0.12、 =0.09, substituting into the formula, we get:

[0091] S=0.7×0.92+0.3×(0.4×0.08+0.3×0.12+0.3×0.09)=0.644+0.3×0.095=0.6725.

[0092] 4. Defect Classification and Location Output: The preset defect confidence score threshold is 0.5. When S≥0.5, it is judged as the corresponding defect type, and when S<0.5, it is judged as a qualified product. At the same time, through pixel coordinate mapping, the pixel coordinates of the defect area in the image are converted into the actual physical coordinates of the PTC starter, and the defect location information is output. For example, the score S=0.6725≥0.5 for the above electrode cold solder joint is judged as an electrode sheet cold solder joint defect, and the physical coordinates of the cold solder joint area (X:5.2mm, Y:3.8mm) are output.

[0093] Step S5: Generate hierarchical control instructions based on defect classification results.

[0094] Based on the defect classification results, this step generates three types of hierarchical control commands: qualified release, unqualified rejection, and adaptive adjustment of process parameters. This achieves closed-loop linkage with the production line rejection device and the production execution system, as detailed below:

[0095] Qualified release instruction: When the defect confidence score S < 0.5, the product is judged to be qualified. A qualified release instruction is generated and sent to the production line conveyor belt control system to control the conveyor belt to continue running and transfer the qualified product to the subsequent packaging process.

[0096] Non-conforming rejection instruction: When the defect classification result is an unrepairable defect (chip crack, shell deformation, pin breakage, chip corner missing), a non-conforming rejection instruction is generated and sent to the production line rejection device (pneumatic pusher). The rejection device pushes the non-conforming product into the collection box at a preset position. The push response time is ≤0.1s, which does not affect the continuous operation of the production line.

[0097] Adaptive process parameter adjustment command: When the defect classification result is a repairable defect (electrode sheet cold solder joint, pin misalignment, solder surface contamination), an adaptive process parameter adjustment command is generated and sent to the production execution system. The front-end production process parameters are adjusted according to the defect type and severity. Specific adjustment example:

[0098] Electrode sheet cold solder joint: Based on the area of ​​the cold solder joint and the confidence score, the welding current is adjusted from 180A to 190~200A, and the welding time is adjusted from 20ms to 25~30ms;

[0099] Pin misalignment: Adjust the pressure of the pin positioning cylinder in the assembly process from 0.4MPa to 0.45~0.5MPa to correct the pin positioning deviation;

[0100] Contamination of welding surface: Adjust the air volume of the dust removal air knife before welding from 5m³ / min to 6~7m³ / min to eliminate dust contamination on the welding surface.

[0101] Step S6: Online incremental learning and updating of the deep belief network model

[0102] To address the model's weak generalization ability to new defects, this step uses false positive samples (samples incorrectly identified by the model) and new defect samples (samples of new defects appearing after adjustments to the production process) collected during production as incremental data to perform online incremental learning updates on the deep belief network model, eliminating the need for offline retraining. The specific implementation is as follows:

[0103] Incremental data collection and annotation: False detection samples of the model are collected through manual verification stations on the production line, and newly added defect samples appearing in the production process are manually annotated. The annotation content includes defect type, feature vector, and defect confidence score. When the incremental data volume reaches 1,000 images, a model update is triggered.

[0104] Model weight fine-tuning: Only the weights of the backpropagation classification layer of the deep belief network are fine-tuned, while keeping the weights of the restricted Boltzmann machine unchanged to avoid damaging the model's feature extraction capabilities; fine-tuning uses mini-batch training with a batch size of 32, a learning rate of 0.0001, and 50 iterations to quickly enable the model to identify new defects;

[0105] Model update verification: After fine-tuning, the model is verified using a test set. When the model's accuracy in identifying new defects is ≥99%, the model update is confirmed to be complete. The updated model is then deployed to the detection system to achieve self-updating of the model.

[0106] Practical application effect test

[0107] The detection method and system of this invention were deployed on the automated production line of a PTC starter manufacturer for a 30-day practical application effect test. The production line speed was 50 units / minute, with an average daily production of 72,000 PTC starters. The test objects were standard specification PTC starters (pin center distance 10mm, electrode area 15mm²). The test results are as follows:

[0108] Inspection accuracy: A total of 2.16 million PTC starters have been inspected, with 2,158 defective products identified, 2 false positives, and 0 missed positives, resulting in a false positive / false negative rate of 0.0009%, far lower than the 8%-12% of existing inspection technologies; the identification rate for minute defects such as chip microcracks and electrode micro-soldering reaches 99.95%;

[0109] Inspection efficiency: The single system can inspect an average of 72,000 pieces per day, with an inspection efficiency of 3,000 pieces / hour. It is fully adapted to the production line speed, and the efficiency is more than 10 times higher than that of manual inspection. It saves 8 manual inspection personnel and reduces labor costs by 80%.

[0110] Production yield: Through adaptive adjustment of process parameters, the recurrence rate of repairable defects such as electrode soldering defects and pin misalignment was reduced from 12% to 1.5%, and the overall production yield of PTC starters was increased from 92% to 99.2%.

[0111] Model adaptability: When a new type of defect (chip surface scratches) appears during the production process, the model's recognition accuracy for this new defect increases to 99.5% within 24 hours after online incremental learning, without the need for offline retraining, and the model update time is ≤30 minutes;

[0112] System stability: During the 30-day testing period, the system operated without failure for 710 hours, with a failure rate of ≤0.01%, a control command response time of ≤0.08s, and no delay in detection and rejection actions, fully meeting the requirements of continuous operation of industrial production lines 24 / 7.

[0113] Test results show that the detection method and system of the present invention can effectively solve the pain points of existing PTC starter production detection technology, realize high-precision, high-efficiency and intelligent online quality control, and have significant industrial application value.

[0114] The visual inspection method and system for PTC starter production of this invention integrates deep learning and machine vision technologies, significantly improving the overall accuracy of PTC starter production inspection. It can accurately identify various minor defects, achieving standardization and quantification of defect judgment, and effectively reducing false negatives and missed detections. By establishing a closed-loop linkage mechanism between inspection and front-end production processes, it can adaptively adjust process parameters for repairable defects, reducing the recurrence of similar defects and significantly improving product yield. The online incremental learning mechanism on the model can quickly adapt to new defect types, requiring only minor weight adjustments to complete updates without offline retraining, greatly reducing model maintenance costs. A dedicated multi-step image preprocessing scheme can effectively eliminate various interferences caused by uneven lighting, dust, and equipment vibration on the production line, ensuring the feature clarity of the acquired images and laying a solid foundation for accurate inspection. At the same time, the inspection system adopts a modular architecture design, which can be directly connected to existing production lines, making deployment convenient, operation stable, and command response timely. It can fully adapt to the full-process quality control needs of large-scale automated production of PTC starters, comprehensively solving industry pain points such as low inspection accuracy, strong subjective judgment, disconnect from production processes, and poor model adaptability in traditional inspection.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0116] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A visual inspection method for PTC starter manufacturing, characterized in that, include: Acquire the original image of the PTC starter to be tested; The original image is enhanced to generate a standardized image to be inspected; Based on a preset multi-scale decomposition algorithm, geometric structure features and surface texture features in the image to be inspected are extracted, and a multi-dimensional feature vector is constructed. The multidimensional feature vector is input into a trained deep belief network model, which outputs defect classification results and defect location information. Based on the defect classification results, corresponding control instructions are generated. The control instructions include at least a qualified release instruction, a non-qualified rejection instruction, or a process parameter adaptive adjustment instruction.

2. The visual inspection method for PTC starter production according to claim 1, characterized in that, The steps for generating the image to be inspected specifically include: The original image is subjected to illumination uniformity correction based on a two-dimensional gamma function; A hybrid filtering algorithm combining adaptive median filtering and bilateral filtering is used to remove noise from the corrected image; and Based on the physical contour template of the PTC starter, the target area containing the electrode sheet, PTC thermistor chip, and pin soldering surface is segmented by the template matching algorithm.

3. The visual inspection method for PTC starter production according to claim 1, characterized in that, The steps for extracting the multidimensional feature vector specifically include: The image to be inspected is decomposed into low-frequency approximation components and multiple high-frequency detail components based on wavelet transform. Extract the gray-level co-occurrence matrix eigenvalues ​​of the low-frequency approximation components, wherein the gray-level co-occurrence matrix eigenvalues ​​include contrast, energy, and correlation; Extract the local binary pattern texture spectral features of the high-frequency detail components; and The gray-level co-occurrence matrix eigenvalues ​​are fused with the local binary pattern texture spectrum features to construct the multidimensional feature vector.

4. The visual inspection method for PTC starter production according to claim 1, characterized in that, The specific steps for outputting defect classification results include: The multidimensional feature vector is input into a deep belief network, which includes multiple restricted Boltzmann machines stacked sequentially and a backpropagation classification layer; The defect confidence score is calculated based on the multidimensional feature vector, and the defect classification result is confirmed based on the comparison result of the defect confidence score and the preset threshold.

5. The visual inspection method for PTC starter production according to claim 4, characterized in that, The defect confidence score S is calculated using the following formula: ； in, For the multidimensional feature vector, and The weight matrix and bias terms of the backpropagation classification layer are, Output weight coefficients for deep networks. These are the geometric constraint weighting coefficients. Let be the prior weight of the i-th geometric feature. The radial basis function deviation value is constructed based on the aforementioned geometric structural features. This represents the total number of geometric features involved in the calculation.

6. The visual inspection method for PTC starter production according to claim 1, characterized in that, The steps for generating corresponding control instructions based on the defect classification results specifically include: When the defect classification result is a qualified product, the qualified release instruction is generated; When the defect classification result is an unrepairable defect, the non-compliance rejection instruction is generated and sent to the rejection device; When the defect classification result belongs to the preset repairable category, the process parameter adaptive adjustment instruction is generated and sent to the production execution system. The process parameter adaptive adjustment instruction is used to adjust the front-end welding process parameters or assembly pressure parameters.

7. The visual inspection method for PTC starter production according to claim 1, characterized in that, Also includes: The false detection samples and newly added defect samples collected during the production process are used as incremental data to perform online incremental learning and updates on the deep belief network model.

8. A visual inspection system for PTC starter production, characterized in that, include: The image acquisition module is used to acquire raw images of the PTC starter under test. An image preprocessing module is used to enhance the original image to generate a standardized image to be inspected; The feature extraction module is used to extract geometric structure features and surface texture features from the image to be inspected based on a preset multi-scale decomposition algorithm, and construct a multi-dimensional feature vector; The defect identification module is used to input the multidimensional feature vector into a trained deep belief network model and output the defect classification result and defect location information. as well as The hierarchical control module is used to generate corresponding control instructions based on the defect classification results. The control instructions include at least a qualified release instruction, a non-qualified rejection instruction, or a process parameter adaptive adjustment instruction.

9. The PTC starter production visual inspection system according to claim 8, characterized in that, The image preprocessing module includes: An illumination compensation unit is used to perform illumination uniformity correction on the original image based on a two-dimensional gamma function. A filtering and noise reduction unit is used to remove noise from the corrected image using a hybrid filtering algorithm combining adaptive median filtering and bilateral filtering; and The region of interest extraction unit is used to segment the target region containing the electrode sheet, PTC thermistor chip, and pin bonding surface based on the physical contour template of the PTC starter using a template matching algorithm.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the visual inspection method for PTC starter production as described in any one of claims 1 to 7.

Images