Intelligent detection method for forming defects of PET (Polyethylene Terephthalate) sheet based on machine vision

Abstract

The invention belongs to the technical field of artificial intelligence and machine vision, and particularly relates to a PET sheet forming defect intelligent detection method based on machine vision, which comprises the following steps: constructing a system integrating dynamic imaging regulation and control, multi-modal feature fusion, physical and visual coupling modeling and closed-loop feedback optimization; adjusting imaging parameters in real time by using a process context vector, and extracting pixel-level, texture-level and semantic-level features under multiple scales; performing deformation correction on the features in combination with a differentiable physical constraint layer, and completing defect classification through a graph neural network; meanwhile, a closed-loop optimization mechanism driven by online incremental learning and reinforcement learning is introduced, self-adaptive evolution of detection performance is achieved, the method remarkably improves the detection robustness and accuracy of low-contrast and high-variation defects, and it is guaranteed that the takt time of a production line is synchronous.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine vision technology, specifically a machine vision-based intelligent detection method for defects in PET sheet forming. Background Technology

[0002] In the field of plastics processing and packaging manufacturing, polyethylene terephthalate (PET) sheets are widely used in food containers, pharmaceutical packaging and other products due to their excellent transparency, mechanical strength and recyclability. As downstream industries increase their requirements for surface quality, defects such as crystal points, scratches and bubbles generated during the thermoforming process of PET sheets have become key factors restricting the yield and end product performance. In order to meet the online quality control needs of high-speed continuous production lines, intelligent inspection technology based on machine vision has replaced traditional manual visual inspection and become the mainstream solution for high-precision, non-contact defect identification. The current mainstream PET sheet defect detection system in the industry consists of a high-resolution line scan camera, an illumination module, an image acquisition card, and a traditional image processing and analysis unit. Its workflow is as follows: continuous frame images are acquired simultaneously while the sheet is being transported at a constant speed. Suspected defect areas are extracted through grayscale threshold segmentation, edge detection, and other methods. Then, the defect areas are classified and judged according to preset geometric or statistical feature rules. This solution effectively improved detection efficiency in the early stages and alleviated the problem of missed detections and misjudgments by humans. It has advantages in response speed and cost in defect scenarios with simple structure and high contrast, and has long occupied a dominant position in the market.

[0003] As PET sheet production evolves towards ultra-thin and high-speed processes, existing visual inspection paradigms based on static feature matching suffer from the following problems: The dynamic ambiguity of defect characterization and the mismatch between static criteria of detection models; under the coupling of high-speed traction and thermoplastic deformation, the morphology of the same type of defect varies significantly with the working conditions, and the detection logic that relies on fixed templates or thresholds is prone to false alarms and missed detections; moreover, the decoupling of image acquisition and defect judgment makes it impossible to dynamically adjust imaging parameters, which exacerbates the invisibility of low-contrast defects. This problem could lead to the following risks: Relaxing thresholds allows minor defects to be released, and manual adjustments increase maintenance costs and lack a feedback loop.

[0004] Therefore, this invention provides an intelligent detection method for PET sheet forming defects based on machine vision. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by the present invention to solve its technical problem is: the intelligent detection method for PET sheet forming defects based on machine vision, as described in the present invention, includes a multispectral linear array imaging unit deployed downstream of the forming section during the continuous conveying of PET sheet to synchronously trigger the acquisition of original image data at a preset frame rate. The embedded edge computing node acquires multi-source process parameters from the traction roller encoder, infrared thermal imaging sensor and ambient light monitoring module in real time, forming a process context vector that is strictly time-aligned with the current image frame; Based on the process context vector, the dynamic imaging control module generates exposure time, light source intensity and focal length shift instructions for the next frame image acquisition in real time according to the preset imaging quality evaluation function, and directly drives the imaging unit to perform parameter adjustment through the hardware synchronization interface to ensure that the defect area can obtain the best signal-to-noise ratio imaging results under different working conditions. Preferably, the dynamically optimized image frame is fed into a multi-scale feature extraction engine. This engine adopts a cascaded convolutional structure to extract local gradient response, directional frequency domain energy distribution and contextual semantic association features in parallel at three levels: pixel level, texture level and semantic level. The three types of features are then mapped to a unified high-dimensional feature space. The physical and visual coupling discrimination module receives the high-dimensional feature vector and combines it with the temperature gradient, stretching rate and cooling uniformity index in the current process context vector. Through the differentiable physical constraint layer, the feature is deformed and normalized, so that the appearance difference of the same type of defect under different process perturbations is compressed to the minimum discrimination distance. Based on the corrected feature vector, the defect classification decision unit calls a lightweight graph neural network model maintained by an online incremental learning mechanism to complete defect type identification and confidence assessment. The system writes the current test results, imaging parameter adjustment records, and process context information into a distributed log buffer. The closed-loop feedback optimization module periodically analyzes the correlation between historical test performance and parameter configuration, and automatically updates the weight coefficients in the imaging quality evaluation function and the deformation mapping parameters of the physical constraint layer, thereby achieving continuous adaptive evolution of the test strategy.

[0007] Preferably, the multispectral linear array imaging unit consists of at least three independent channels, corresponding to the visible light band, near-infrared band, and polarization-sensitive band, respectively. Each channel is equipped with an independent LED array light source and a bandpass filter, and communicates with the main control unit via an RS485 bus to ensure microsecond-level synchronization between light source switching and image acquisition. The image sensor chip of each channel integrates a global shutter and on-chip noise reduction circuit. Its output signal is transmitted to the image acquisition card via the CameraLink interface, and then sent to the GPU acceleration unit of the edge computing node for preprocessing via the PCIe bus.

[0008] Preferably, the dynamic imaging control module has a built-in imaging quality evaluation function, which is defined as a nonlinear mapping relationship between the process context vector and the local contrast, edge sharpness, and noise power spectral density of the current image frame. This mapping relationship is stored in the non-volatile storage area of ​​the security element chip in the form of trainable parameters and is protected against illegal tampering through hardware encryption. When a new frame of image arrives, the edge computing node first calculates the gradient magnitude histogram entropy value of the preset region of interest in the image. If the entropy value is lower than the preset security threshold, it is determined that the current imaging contrast is too low, and the dynamic control process is triggered. The control command generation process adopts a combination of table lookup and interpolation. Based on the linear velocity and surface temperature in the process context vector, the optimal exposure-light source combination is retrieved from the pre-stored three-dimensional parameter surface, and the light source driving circuit and camera exposure controller are driven by the pulse width modulation signal.

[0009] Preferably, the multi-scale feature extraction engine consists of a three-stage parallel processing pipeline: the first stage is a pixel-level processing unit, which uses a Sobel operator kernel with fixed weights to perform convolution operations on the input image, generate gradient components in the horizontal and vertical directions, and calculate their L2 norm as local contrast features. The second stage is the texture-level processing unit, which deploys four sets of direction-selective Gabor filters, corresponding to the four main directions of 0°, 45°, 90° and 135° respectively. After frequency domain decomposition of the image, the mean and variance of the energy response in each direction are extracted. The third level is the semantic processing unit, which adopts a lightweight MobileNetV3 backbone network. Its input is a low-resolution version of the original image after bicubic interpolation scaling, and its output is a 1280-dimensional semantic feature vector after global average pooling. After the three-level output features are concatenated through channels, they are reduced to 512 dimensions through a 1×1 convolutional layer to form a unified high-dimensional feature vector.

[0010] Preferably, the physical and visual coupling discrimination module includes a differentiable physical constraint layer, which is constructed based on the constitutive relation model of PET material during thermoforming. Specifically, the model takes the longitudinal stretch ratio, transverse stretch ratio, and cooling rate in the process context vector as input variables and outputs an affine transformation matrix for inverse deformation compensation of the spatial position sensitive components in the high-dimensional feature vector. The parameters of the affine transformation matrix are jointly optimized during the training phase through the backpropagation algorithm to ensure that the representation distance of the same defect category under different stretching states in the compensated feature space is less than the preset clustering radius. During the inference phase, the matrix is ​​calculated in real time by the FPGA coprocessor and applied to the feature vector, with the latency controlled within a single frame processing cycle.

[0011] Preferably, the graph neural network model used by the defect classification decision unit constructs a dynamic graph structure with defect candidate regions as nodes and spatial adjacency and feature similarity as edge weights. Each node is initially embedded with its corresponding high-dimensional feature vector, and after two layers of graph convolution operations, neighbor information is aggregated and the node representation is updated. The fully connected classification head outputs the defect category probability distribution based on the updated node embedding. The parameters of the model are stored in a Trusted Execution Environment (TEE) isolation area and can only be updated through incremental learning requests verified by digital signatures. The incremental learning mechanism is triggered by the closed-loop feedback optimization module. When the false alarm rate or false negative rate of multiple consecutive production batches exceeds the preset convergence judgment condition, the system automatically extracts relevant samples from the distributed log buffer, performs small-batch fine-tuning, and ensures that the model's generalization ability does not degrade through cross-validation.

[0012] Preferably, the closed-loop feedback optimization module runs in an independent real-time operating system kernel, the core of which is a policy optimizer based on reinforcement learning. The optimizer takes historical detection accuracy, imaging parameter adjustment frequency, and production line downtime as state inputs and the weight update vector of the imaging quality evaluation function as action outputs. After each optimization cycle, the new policy is written to the safe boot partition and loaded and takes effect during the next system cold boot. To prevent policy oscillation, the optimizer has a built-in sliding window consistency check mechanism, which only submits parameter updates when the optimization direction is consistent for three consecutive times.

[0013] Preferably, the distributed log buffer adopts a circular queue structure and is deployed in a dual-port RAM. The DMA controller automatically writes detection metadata during the idle period of the image processing pipeline. The metadata includes timestamps, image hash values, process context vectors, imaging parameters, defect coordinates, category labels, and confidence levels. The buffer is also periodically read by a remote monitoring terminal through an Ethernet interface for production line quality traceability and process parameter backtracking analysis.

[0014] Preferably, the entire inspection system communicates with the host MES system via the IEC61850 extended specification, supports GOOSE fast message transmission of defect alarm events, and uploads complete inspection reports via MMS service; all communication links are encrypted end-to-end using the TLS1.3 protocol, and the keys are automatically rotated quarterly by the hardware security module.

[0015] The beneficial effects of this invention are as follows: The present invention provides an intelligent detection method for PET sheet forming defects based on machine vision, which avoids the invisibility of low-contrast defects caused by fixed exposure parameters through a process context-driven dynamic imaging control mechanism. By leveraging the synergistic effect of multi-scale feature fusion and physical constraint correction, the apparent variation of defects caused by process disturbances is compressed within the feature space, avoiding the false alarms and missed detections that are prone to occur when the detection logic relies on fixed templates or thresholds; a closed-loop linkage between defect discrimination logic and imaging strategy is realized, enabling the system to continuously optimize detection performance without external intervention. This ensures the safe incremental updates of the graph neural network model within a trusted execution environment, preventing model failure due to imaging condition drift. It also supports end-to-end quality traceability and production line collaborative control through the integration of distributed logs and industrial communication protocols. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the intelligent detection method for PET sheet forming defects based on machine vision, as described in this invention. Detailed Implementation

[0018] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0019] like Figure 1 As shown in the embodiment of the present invention, an intelligent detection method for PET sheet forming defects based on machine vision is deployed in the downstream of the traction section of the PET thermoforming production line, forming a physical collaboration with the traction roller, cooling roller and winding device. The system hardware platform consists of a multispectral linear array imaging unit, an embedded edge computing node, an infrared thermal imaging sensor, an ambient light monitoring module, a traction roller encoder, an FPGA coprocessor, a security element chip, a dual-port RAM log buffer, and an industrial communication interface. The software logic encompasses five major functional modules: dynamic imaging control, multi-scale feature extraction, physical and visual coupling discrimination, graph neural network classification decision-making, and closed-loop feedback optimization. Each module achieves millisecond-level collaboration through data streams with strictly aligned timestamps.

[0020] In practice, the multispectral linear array imaging unit is installed transversely along the direction perpendicular to the PET sheet conveying, at a height of 320mm above the sheet surface, and integrates three independent imaging channels: The first channel is equipped with a global shutter CMOS sensor, and also features a visible light bandpass filter and a white LED array light source; The second channel uses a near-infrared sensor, a narrow-band filter, and an infrared LED light source; The third channel uses a custom polarization-sensitive CMOS chip, whose micro-polarization array is periodically arranged in four directions: 0°, 45°, 90°, and 135°. Combined with a switchable linearly polarized light source, it is used to capture the birefringence effect induced by surface stress. All three channels receive master control commands via RS485 bus and are triggered for image acquisition by the same hardware synchronization signal; Image data is output to the PCIeGen3x4 image acquisition card via the CameraLinkBaseConfiguration interface, and then sent to the GPU memory space of the NVIDIA Jetson AGX Orin edge computing node.

[0021] The traction roller encoder is an incremental photoelectric encoder, installed at the end of the main traction roller shaft. It outputs pulse signals in real time to the GPIO interrupt pin of the edge computing node to calculate the current linear velocity v (unit: m / min) and generate the image acquisition trigger timing. The infrared thermal imaging sensor is an FLIRA655sc uncooled focal plane array with a field of view of 25°×19° and a temperature measurement range of 30-300°C. It acquires the surface temperature distribution of the sheet and transmits it to the edge node via the GigEVision protocol. The ambient light monitoring module consists of a digital light intensity sensor, whose output is used to compensate for the impact of stray ambient light on image contrast. The above three types of sensor data and image frames are aligned at the nanosecond level using the PTP (Precision Time Protocol) to form a process context vector. ; in, The linear velocity is (m / min). The average surface temperature of the sheet (°C). and These represent the transverse and longitudinal temperature gradients (°C / mm), respectively. For ambient illuminance (lux). The cooling rate (°C / s) is estimated from the time derivative of the infrared sequence.

[0022] The dynamic imaging control module runs in the real-time Linux kernel space of the edge computing node, and its core is the imaging quality evaluation function. This function will use the process context vector With the current image frame Mapped to scalar quality score; Specifically, the system pre-defines a region of interest (ROI) with a size of 512×512 pixels located in the central area of ​​the sheet in the image, and calculates the Shannon entropy of the gradient magnitude histogram for this region: ; in, For the gradient magnitude to fall into the first Total intervals (in total) The probability of (a series of equally wide intervals). If If the current image contrast is too low, the imaging parameters need to be adjusted. The control instruction generation process relies on a three-dimensional parametric surface pre-stored in the non-volatile memory area of ​​the Infineon OPTIGA™ TrustM chip. ,in Exposure time (μs) The light source driving current is (mA). This surface was constructed using offline calibration experiments: In m / min Grid sampling was performed within the °C range at a step size of 2 m / min and 5 °C. For each group Combinatorial traversal μs and The combination of mA is used to select the parameter pair that maximizes the signal-to-noise ratio (SNR) within the ROI as the surface points; When running online, the system determines the current... In and ,exist Perform trilinear interpolation to obtain the target and It also controls the camera exposure controller (BasleraceacA2000-50gc) and the LED driver circuit (MeanWellLDD-H series) respectively via PWM signals, with an adjustment delay of less than 1ms.

[0023] Dynamically optimized image frames It is fed into the multi-scale feature extraction engine; The engine consists of a three-stage parallel pipeline, all deployed in the CUDA stream of the GPU to achieve zero-copy processing; The first-level pixel-level processing unit... Applying the fixed-weight Sobel operator: ; Calculate horizontal and vertical gradient components , And synthesize local contrast feature maps. The mean and standard deviation of λ form a 2-dimensional feature vector. The second-level texture processing unit deploys four Gabor filters, and its kernel function is defined as follows: ; in , , , Pixels , , ; right After convolution, calculate the average energy of the response in each direction. With variance This forms an 8-dimensional texture feature vector; The third-level semantic processing unit will The bicubic interpolation scales the image to 224×224 pixels and inputs it into a lightweight MobileNetV3-Small backbone network. This network is pre-trained on ImageNet and the last three layers are fine-tuned on this task to output a 1280-dimensional global semantic feature vector. The third-level features are concatenated through channels to form a 1290-dimensional original feature, which is then reduced to 512 dimensions through a 1×1 convolutional layer, denoted as . .

[0024] Physical and visual coupling discrimination module receives With process context vector First, extract the stretching-related parameters: longitudinal stretch ratio. ( (Extruder exit line speed), transverse draw ratio ( For the current width, (Wide width of the die head exit), cooling rate ; Based on the thermoviscoelastic constitutive model of PET material, an affine transformation matrix is ​​constructed. : ; Among the elements for , , The nonlinear function is parameterized by a three-layer fully connected network, which is jointly optimized with the subsequent classifier during the training phase. because The first 256 dimensions correspond to spatial location-sensitive features (from the intermediate layer activations of MobileNetV3), which are then reshaped into... Feature map Apply reverse deformation compensation: ; Sub-pixel sampling is achieved through bilinear interpolation. The compensated feature map is then subjected to global average pooling and concatenated with the subsequent 256-dimensional non-spatial features to form the corrected feature vector. ; The calculation is performed by the Xilinx Kintex-7XC7K325T FPGA coprocessor, which utilizes its hard-core DSP resources to achieve parallel coordinate transformation, with a single-frame processing latency of 0.8ms.

[0025] Defect classification decision unit As input, firstly, non-maximum suppression (NMS) is used to generate defect candidate regions on the feature map, with each region corresponding to a node. Its initial embedding System construction dynamic diagram edge set Determined by two criteria: if the spatial distance between two nodes Pixels, and cosine similarity Then establish an edge Weight Graph neural networks consist of two layers of graph convolutional neural networks (GCNs): ; in For nodes The neighborhood group, For learnable weight matrix, This is the ReLU activation function. Final node representation. Input a fully connected classification header and output the probability distribution of 6 types of defects (crystal points, scratches, bubbles, black spots, uneven thickness, and edge burrs). and confidence level ; The model parameters are encrypted and stored in the trusted execution environment (TEE) of the ARMTrustZone, and are only accessible to incremental learning requests verified by ECDSA-P256 digital signatures.

[0026] The closed-loop feedback optimization module runs in a standalone FreeRTOS kernel with a cycle time of 10 minutes. Its state vector... ,in This represents the detection accuracy over the past 1000 frames. Adjust the imaging parameters at a frequency (times / minute). The number of production line downtimes caused by false alarms; Action vectors Corresponding image quality evaluation function The increment of each weight in the equation, where NPSD is the noise power spectral density; The policy optimizer employs the Proximal Policy Optimization (PPO) algorithm, whose reward function... ,coefficient , , The new strategy is written to the secure boot partition of the eMMC, only when the optimization direction is consistent three times consecutively within the sliding window (i.e., ... It will only take effect on the next cold start if (this condition is met).

[0027] The distributed log buffer uses a 64MB dual-port SRAM (CypressCY7C028AV) organized as a circular queue. Each record occupies 256 bytes and includes: a 64-bit timestamp (PTP synchronization), a 128-bit image SHA3-256 hash value, a 24-byte process context vector, and 12 bytes of imaging parameters. , , focal length offset), 8-byte defect coordinates (x, y), 1-byte category label, 1-byte confidence score (0-255); The DMA controller automatically writes records during idle periods after the GPU completes feature extraction, and the remote monitoring terminal reads the latest 1,000 records at a frequency of 1Hz via gigabit Ethernet. The system broadcasts a high-confidence defect event within <2ms using the IEC61850-8-1 GOOSE message. The system uploads a complete test report via MMS service. All communications are TLS 1.3 enabled, and session keys are rotated quarterly by the InfineonOPTIGA™ TrustM hardware security module.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based intelligent detection method for defects in PET sheet molding, characterized in that, The method is executed by an intelligent detection system deployed on a continuous PET sheet conveying production line. The system includes a multispectral linear array imaging unit, embedded edge computing nodes, multi-source process sensors, and a closed-loop feedback optimization module. The method includes the following steps: During the PET sheet transport process, the multispectral linear array imaging unit synchronously acquires raw image data at a preset frame rate; The embedded edge computing node acquires multi-source process parameters from the traction roller encoder, infrared thermal imaging sensor and ambient light monitoring module in real time, and constructs a process context vector that is strictly time-aligned with the current image frame. The dynamic imaging control module generates instructions for acquiring the next frame of image in real time based on the process context vector, and drives the multispectral linear array imaging unit to perform parameter adjustments. The dynamically optimized image frame is input into the multi-scale feature extraction engine. The multi-scale feature extraction engine extracts associated features in parallel at three levels: pixel level, texture level and semantic level, and maps the associated features to a unified high-dimensional feature space to form the original high-dimensional feature vector. The original high-dimensional feature vector and the process context vector are input into the physical and visual coupling discrimination module. This module constructs a differentiable physical constraint layer based on the thermoforming constitutive relation model of PET material, performs inverse deformation compensation on the spatial position sensitive component in the original high-dimensional feature vector, and generates a corrected feature vector. Based on the corrected feature vector, the defect classification decision unit calls the graph neural network model to complete the defect type identification and confidence assessment. The detection results, imaging parameter adjustment records, and process context information are written into the distributed log buffer. The closed-loop feedback optimization module periodically analyzes the correlation between historical detection performance and parameter configuration, and automatically updates the weight coefficients of the imaging quality evaluation function and the deformation mapping parameters of the physical constraint layer, thereby realizing the continuous adaptive evolution of the detection strategy.

2. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, The multispectral linear array imaging unit includes at least three independent imaging channels, corresponding to the visible light band, the near-infrared band, and the polarization-sensitive band, respectively. Each of the three independent imaging channels is equipped with an independent LED array light source, a bandpass filter, and a global shutter image sensor, and achieves inter-frame synchronization through a hardware synchronization signal; The image data from the three independent imaging channels is transmitted to the image acquisition card and then sent to the GPU acceleration unit of the embedded edge computing node.

3. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, In the dynamic imaging control module, based on the process context vector, the dynamic imaging control module generates acquisition exposure time, light source intensity and focal length shift instructions according to the imaging quality evaluation function. The imaging quality evaluation function is defined as a nonlinear mapping relationship between the process context vector and the local contrast, edge sharpness and noise power spectral density of the current image frame.

4. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, The multi-scale feature extraction engine consists of a three-stage parallel processing pipeline: The first-level pixel-level processing unit uses a fixed-weight Sobel operator kernel to calculate the horizontal and vertical gradient components and synthesize local contrast features; The second-level texture processing unit deploys filters in four directions, corresponding to the four main directions of 0°, 45°, 90° and 135°, respectively, to extract the mean and variance of the energy of the frequency domain response in each direction; The third-level semantic processing unit uses a lightweight MobileNetV3 backbone network to process the low-resolution image after bicubic interpolation scaling and outputs a global semantic feature vector. The three-level output features are concatenated through channels and then reduced to 512 dimensions through a convolutional layer to form the original high-dimensional feature vector.

5. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, In the physical and visual coupling discrimination module, the differentiable physical constraint layer takes the longitudinal stretching ratio, transverse stretching ratio and cooling rate in the process context vector as input and outputs an affine transformation matrix. This matrix is ​​used to perform inverse deformation compensation on the feature map corresponding to the spatial location-sensitive components in the original high-dimensional feature vector; The parameters of the affine transformation matrix are jointly optimized through backpropagation during the model training phase and are calculated and applied in real time by the FPGA coprocessor during the inference phase.

6. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, The graph neural network model used by the defect classification decision unit constructs a dynamic graph structure with defect candidate regions as nodes and spatial adjacency and feature similarity as edge weights. The initial embedding of the node is its corresponding corrected feature vector; After aggregating neighbor information through two layers of graph convolution operations, the fully connected classification head outputs the probability distribution and confidence level of the defect category. The parameters of the graph neural network model are stored in a trusted execution environment isolation zone, and can only be updated by incremental learning requests verified by digital signatures.

7. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 6, characterized in that, The incremental learning mechanism is triggered by the closed-loop feedback optimization module: When the false alarm rate or false negative rate of multiple consecutive production batches exceeds the preset convergence criteria, the system automatically extracts relevant samples from the distributed log buffer.

8. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, The closed-loop feedback optimization module runs on an independent real-time operating system kernel. Its status inputs include historical detection accuracy, imaging parameter adjustment frequency, and production line downtime. Its action output is the weight update vector of the imaging quality evaluation function. The module employs a policy optimizer based on reinforcement learning, whose reward function comprehensively considers detection performance, regulation overhead, and production line stability.

9. The intelligent detection method for PET sheet molding defects based on machine vision according to claim 1, characterized in that, The distributed log buffer is deployed in a dual-port RAM using a circular queue structure, and the detection metadata is automatically written by the DMA controller during the idle period of the image processing pipeline. The metadata includes timestamps, image hash values, process context vectors, imaging parameters, defect coordinates, category labels, and confidence levels. The buffer allows remote monitoring terminals to periodically read it via Ethernet for quality traceability and process retrospective analysis.

Images