A method and system for detecting production quality of engineering plastics

Abstract

The present invention relates to the field of computer vision technology, and discloses a method and system for detecting the production quality of engineering plastics. The method uses an industrial camera to collect a sequence of images of the surface of engineering plastics on a production line in real time, and simultaneously obtains the process parameters of an injection molding machine. A ViT‑Adapter visual feature extraction network based on a multi-scale sliding window is constructed to fuse surface texture features and process parameters. A defect classification model is trained using a joint loss function of an improved Focal Loss and contrastive learning, and zero-sample anomaly segmentation is implemented based on WinCLIP to achieve pixel-level positioning of surface bubbles and scratches. Finally, a defect heat map is output and the production line parameters are linked to self-adjustment to visualize the results. The method improves the accuracy and efficiency of defect detection, realizes automated adjustment, reduces manual intervention, and improves production quality and intelligence.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a method and system for detecting the production quality of engineering plastics. Background Art

[0002] During the production of engineering plastics, surface defects such as bubbles and scratches seriously affect product quality. Traditional manual inspection methods are inefficient and easily affected by subjective factors. In recent years, although defect detection methods based on machine vision have been developed, most of them rely solely on image features and ignore the impact of process parameters on defect formation, resulting in low detection accuracy. At the same time, existing methods still have difficulties in handling complex defect types and achieving pixel-level positioning, and lack a linkage adjustment mechanism with the production line. Therefore, there is an urgent need for a method that can integrate image features and process parameters to achieve efficient and accurate defect detection and automatically adjust production line parameters. Summary of the Invention

[0003] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a method and system for detecting the production quality of engineering plastics, so as to solve the problems of low detection efficiency and accuracy and lack of linkage adjustment mechanism in the prior art.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present invention provides a method for detecting the production quality of engineering plastics, the method comprising:

[0006] Step S100: Using an industrial camera to collect real-time images of the surface of engineering plastics on the production line and synchronously obtain process parameters of the injection molding machine;

[0007] Step S200: constructing a multi-scale sliding window ViT-Adapter visual feature extraction network, and cross-modally fusing the surface image features of engineering plastics with the real-time process parameters of the injection molding machine through an adaptive attention mechanism to generate a fused feature vector;

[0008] Step S300: constructing a dual-branch training network based on the fused feature vector, and using a joint loss function of improved Focal Loss and contrastive learning to train a defect classification model;

[0009] Step S400: Using the defect classification model to extract deep semantic features, combined with the zero-sample anomaly segmentation algorithm of the WinCLIP framework, the surface bubbles and scratches are located at the pixel level to generate a defect mask map;

[0010] Step S500: Generate a defect heat map in real time based on the defect location results and visualize the results. Based on the defect area ratio and spatial distribution of the heat map, adjust the injection molding machine process parameters through the PID control algorithm.

[0011] Preferably, in a possible implementation of the first aspect, the image acquisition frame rate of the industrial camera is The relationship with the injection molding cycle is:

[0012]

[0013] in is the injection molding cycle time, As a redundancy factor, the ejection signal of the injection molding machine is obtained in real time through the encoder to trigger the acquisition.

[0014] Preferably, in a possible implementation of the first aspect, the feature extraction network includes:

[0015] Global features extracted by the ViT-L / 16 backbone network and sliding window local features Perform spatial alignment, is a real number matrix, and feature fusion is achieved through deformable convolution:

[0016]

[0017] in is the fused feature tensor, For the sliding window weight coefficients, generated by the gated attention mechanism, is the variable convolution, and s is the number of sliding windows.

[0018] Preferably, in a possible implementation of the first aspect, the cross-modal fusion feature adopts a pyramid structure to perform cross-scale feature interaction:

[0019] Constructing a 3-level feature pyramid , where the P2 layer resolution is 1 / 4 and the channel is 512; the P3 layer resolution is 1 / 8 and the channel is 256; the P4 layer resolution is 1 / 16 and the channel is 128; cross-scale feature interaction is performed through a bidirectional feature pyramid network:

[0020]

[0021] in For the Layer output feature map, For the Layer input feature map, is the feature weight of the current layer, is the feature weight of the adjacent layer, For the Layer input feature map, and is a learnable weight that satisfies , is an upsampling or downsampling operation.

[0022] Preferably, in a possible implementation manner of the first aspect, the defect classification model training adopts an improved loss function:

[0023]

[0024] in and is the loss weight coefficient, is Focal Loss, is the contrast loss term:

[0025]

[0026] in is the similarity score of the positive sample pair, is the similarity score of the negative sample pair, n is the index of the negative sample, and the temperature coefficient , N is the number of negative samples.

[0027] Preferably, in a possible implementation of the first aspect, the defect classification model training adopts multimodal data enhancement:

[0028] Image generation conditioned on process parameters, generating synthetic data with defect labels through VAE-GAN:

[0029]

[0030] in For expected value calculation, is the discriminator output, For the generator with latent variables and process conditions The generated image, is the latent variable distribution output by the encoder, is the prior distribution, is the KL divergence calculation, is the KL divergence weight, is a conditional vector containing process parameters of preset dimensions.

[0031] Preferably, in a possible implementation of the first aspect, the anomaly segmentation algorithm includes:

[0032] Abnormal semantic vectors generated by CLIP text encoder and visual features Calculate pixel-level similarity:

[0033]

[0034] in is a real matrix, Pixel The anomaly score, For the image block at position The visual feature vector of Defect semantic vector generated for the text encoder, is the scaling function, the scaling factor , is the number of training cycles.

[0035] Preferably, in a possible implementation manner of the first aspect, the result visualization includes:

[0036] Map the defect heat map to the CAD model to generate a 3D quality distribution cloud map. The defect area coloring follows:

[0037]

[0038] Also mark the maximum defect size ;

[0039] Among them is is the coordinate of the centroid of the defect area, is the set of all defective pixels.

[0040] Preferably, in a possible implementation manner of the first aspect, the adjustment of the process parameters of the injection molding machine is implemented as follows:

[0041] According to the defect area ratio and location distribution , calculate the correction amount of injection molding machine:

[0042]

[0043] The PID coefficient , , , is the proportion of defect area in the current detection cycle, is the integral of the historical defect area, is the defect area change rate.

[0044] In a second aspect, the present invention provides an engineering plastic production quality inspection system, the system comprising:

[0045] Image acquisition module, used to collect real-time images of the surface of engineering plastics on the production line through industrial cameras and simultaneously obtain the process parameters of the injection molding machine;

[0046] The feature extraction and fusion module builds a multi-scale sliding window ViT-Adapter visual feature extraction network, which uses an adaptive attention mechanism to perform cross-modal fusion of engineering plastic surface image features and injection molding machine real-time process parameters to generate a fused feature vector.

[0047] The model training module builds a dual-branch training network based on the fused feature vector and uses the joint loss function of improved Focal Loss and contrastive learning to train the defect classification model;

[0048] The anomaly segmentation module uses the defect classification model to extract deep semantic features and combines it with the WinCLIP framework's zero-shot anomaly segmentation algorithm to perform pixel-level positioning of surface bubbles and scratches and generate a defect mask.

[0049] The result output module generates a defect heat map in real time based on the defect location results and visualizes the results. Based on the defect area ratio and spatial distribution in the heat map, the injection molding machine process parameters are adjusted through the PID control algorithm.

[0050] The beneficial effects of the present invention are as follows: the present invention uses intelligent sensors to collect images and process parameters in real time, constructs a multi-scale sliding window ViT-Adapter network to extract fusion features, adopts an improved loss function to train a defect classification model, realizes zero-sample anomaly segmentation based on WinCLIP, outputs a defect heat map and links the production line parameters for self-adjustment.

[0051] This method uses visual inspection to improve the accuracy and efficiency of defect detection, achieves pixel-level positioning, and can automatically adjust production line parameters, reducing manual intervention and improving production quality and intelligence.

[0052] In addition, the application of technologies such as multimodal data enhancement and cross-scale feature interaction of pyramid structures further enhances the generalization ability and robustness of the model, providing a new solution for quality inspection of engineering plastic production. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 A flow chart of a method for testing the production quality of engineering plastics is provided for this application.

[0055] Figure 2A structural diagram of an engineering plastic production quality inspection system is provided for this application.

[0056] Explanation of the accompanying symbols: 1-image acquisition module, 2-feature extraction and fusion module, 3-model training module, 4-abnormal segmentation module, 5-result output module. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0058] Example 1: Figure 1 As shown, the present invention provides a method for detecting the production quality of engineering plastics, comprising:

[0059] Step S100: Using an industrial camera to collect real-time images of the surface of engineering plastics on the production line, and synchronously obtain process parameters of the injection molding machine.

[0060] In this example, the image acquisition system uses a Basler ace2 16k industrial camera equipped with a SchneiderKreuznach Cinegon 1.4 / 12 mm fixed-focus lens. It is mounted 1.2 meters above the injection molding machine's mold opening station, at a 45° angle to the product surface to maximize imaging coverage. The lighting system comprises 12 ring-shaped LED cold light sources, using a PWM dimming controller to achieve stepless brightness adjustment from 0 to 1000 lux, eliminating reflections from the metal mold.

[0061] The image acquisition frame rate is dynamically adjusted according to the injection molding cycle: when the cycle time tcycle = 15 seconds, the formula (where ξ=1.8), the frame rate is 8.64Hz, ensuring that at least 12 frames of images are collected in each cycle, and the ejection signal of the injection molding machine is obtained in real time through the encoder to trigger the acquisition.

[0062] The synchronous acquisition system acquires process parameters through Kistler 5074A pressure sensor and FLIR A655sc infrared thermal imager. The data is transmitted to the industrial computer via OPC UA protocol, and the timestamp alignment accuracy is achieved. Level. The image preprocessing process includes:

[0063] (1) Non-uniform illumination correction: CLAHE algorithm is used with a grid size of 64×64 and a contrast limit threshold of 2.0.

[0064] (2) Motion blur removal, based on Wiener filtering, point spread function PSF size 5×5.

[0065] (3) Multi-frame super-resolution reconstruction uses the VSRNet model to fuse four consecutive frames of images into a resolution of 8192×8196. Data storage adopts a layered architecture. The original images are compressed using H.265 encoding and stored in the NAS storage array. The process parameters are stored in the time series database InfluxDB, enabling millisecond-level correlation queries.

[0066] Step S200: Construct a multi-scale sliding window ViT-Adapter visual feature extraction network, and perform cross-modal fusion of engineering plastic surface image features and injection molding machine real-time process parameters through an adaptive attention mechanism to generate a fusion feature vector.

[0067] In this embodiment, the backbone network adopts the ViT-Large model, with a patch size of 16×16, 24 layers, and a hidden layer dimension of 1024. The input image is processed into blocks to generate a 14×14 global feature map. , with a dimension of 14×14×1024. In view of the subtle defects on the surface of injection molded parts, five sliding windows of different scales are designed, namely 3×3, 5×5, 7×7, 9×9, and 11×11. Local features are extracted from each window through RoI Align. , with a dimension of 5×5×256, and spatially aligned with global features through deformable convolution DeformConv (offset learning rate 0.1, convolution kernel 3×3) to solve the problem of mold texture deformation. The process parameters are mapped into a 128-dimensional conditional vector through the MLP encoder (3-layer fully connected, hidden layer dimension 256), and input into the gated attention network together with the visual features to dynamically generate weights. .

[0068] In the specific implementation, the feature fusion formula , where n=36 is dynamically calculated based on the image size of 2048×2048. The attention weight generation network adopts a dual-path architecture: the global path acquires contextual information through average pooling, while the local path uses the SENet channel attention mechanism. Finally, the fused weights are output through sigmoid activation. Training uses a curriculum learning strategy, initially enabling only 3×3 and 5×5 windows, and gradually introducing larger window sizes to prevent the model from prematurely falling into local optima.

[0069] At the same time, feature fusion adopts a multi-level pyramid structure to achieve deep interaction of cross-scale features. The pyramid structure consists of three levels, and the feature fusion module adopts a three-level feature pyramid structure, including P2, P3, and P4 levels, which correspond to 1 / 4 (P2 layer, 512 channels), 1 / 8 (P3 layer, 256 channels), and 1 / 16 (P4 layer, 128 channels) of the original image resolution respectively. The P2 layer upsamples the 14×14 feature map output from the ViT backbone to 28×28 resolution through depthwise separable convolution (kernel size 3×3, stride 2), and adopts hybrid void convolution (void rate ) enhances the perception of large-area flow marks; the P3 layer is obtained by downsampling the P2 layer through maximum pooling (kernel size 2×2, stride 2), and integrates the cross-region texture association extracted by the void convolution (dilation=2); the P4 layer is directly connected to the global feature Fg of the ViT backbone, compresses the channel to 128 dimensions through 1×1 convolution, and conditionally splices it with the real-time process parameters to form a process perception feature vector. The process parameters include melt temperature, mold temperature, injection pressure, holding pressure, clamping force, injection speed, holding time, cooling time, back pressure, screw speed, melt viscosity and material moisture content.

[0070] Cross-scale fusion is achieved between pyramid levels through a bidirectional feature interaction path:

[0071]

[0072] in For the Layer output feature map, For the Layer input feature map, is the feature weight of the current layer, is the feature weight of the adjacent layer, For the Layer input feature map, and is a learnable weight that satisfies , It is an upsampling or downsampling operation. The top-down path upsamples the P4 layer features to the P3 layer resolution through bilinear upsampling, and combines them with the P3 layer features through gated addition. The gated weights are generated by two layers of fully connected networks. The bottom-up path performs dilated spatial pyramid pooling on the P2 layer features, downsamples the output features to the P3 layer resolution through 3×3 convolution, and then splices them with the P3 layer features. The channel weights are then recalibrated through the SE attention module. The core of cross-layer interaction lies in the dynamic multi-scale attention mechanism: for each sliding window position , extract the multi-resolution feature blocks of the corresponding areas from the P2-P4 layers, and calculate the cross-layer feature correlation matrix by cosine similarity , after Softmax normalization, the hierarchical fusion weights are generated.

[0073] The pyramid structure is deeply coupled with the sliding window: for each sliding window, the P2 layer covers an area 4 times the original size of the window (e.g., a 7×7 window corresponds to 28×28 pixels), and the P2 layer is transformed into a continuous image by the deformable convolution DeformConv (offset range The P3 layer maps to an area twice the size of the window (14×14 pixels) and uses deformable RoI Align to extract local features. The P4 layer directly maps to the center of the window (7×7 pixels), incorporating global context through a cross-attention mechanism. A cross-layer feature compensation module is designed to address resolution mismatch: P4 layer features are upscaled to the P3 layer resolution through sub-pixel convolution, then concatenated with P3 layer features in the channel dimension and compressed to 256 dimensions through 3×3 convolution. Similarly, P3 layer features are compensated to the P2 layer through the same operation, forming a closed-loop feature enhancement pathway.

[0074] Step S300: constructing a dual-branch training network based on the fused feature vector, and using the joint loss function of improved Focal Loss and contrastive learning to train the defect classification model.

[0075] In this embodiment, a joint loss function is used to solve the imbalance problem of injection molding defect samples. .in and is the loss weight coefficient, and this embodiment adopts , , is Focal Loss, the category balance coefficient in Focal Loss , difficult sample focusing coefficient The contrastive learning part builds a Memory Bank containing 1024 negative samples, and the temperature coefficient Softening similarity distribution, loss function ,in is the cosine similarity between the anchor point and the positive sample, is the negative sample similarity, n is the index of the negative sample, and N is the number of negative samples.

[0076] The training strategy uses a two-stage optimization strategy: Focal Loss is used to stabilize the feature space in the first 50 epochs, and contrastive loss is introduced in the last 30 epochs. Hard sample mining is also enabled, with 256 samples sampled per batch. The optimizer used is Rectified Adam, with a weight decay coefficient of 0.01 and a gradient clipping threshold of 5.0.

[0077] Step S400: Utilize the defect classification model to extract deep semantic features, combine with the zero-sample anomaly segmentation algorithm of the WinCLIP framework, perform pixel-level positioning of surface bubbles and scratches, and generate a defect mask map.

[0078] In this embodiment, the text encoder of the WinCLIP framework is extended to support process parameter condition input. After the text prompt (such as "surface bubble, diameter 2mm") is spliced ​​with the real-time process parameters, a 512-dimensional defect semantic vector is generated through the BERT model. The visual branch uses 3-layer strided convolution to extract multi-scale features, and the calculation steps are middle, Pixel The anomaly score, For the image block at position The visual feature vector of Defect semantic vector generated for the text encoder, is the scaling function, the scaling factor Dynamically adjusted during training, T=0.12 in the early stage promotes feature learning, T=0.09 in the middle stage balances accuracy, and T=0.08 in the late stage improves positioning details.

[0079] The training data uses synthetic defect generation technology to add realistic defects to normal samples through a VAE-GAN model (latent space dimension 64, generator contains 5 residual blocks): ,in For expected value calculation, is the discriminator output, For the generator with latent variables and process conditions The generated image, is the latent variable distribution output by the encoder, is the prior distribution, is the KL divergence calculation, is the KL divergence weight, It is a conditional vector containing preset dimensional process parameters. This embodiment uses 12-dimensional process parameters, including melt temperature, mold temperature, injection pressure, holding pressure, clamping force, injection speed, holding time, cooling time, back pressure, screw speed, melt viscosity and material moisture content. The test phase uses a double threshold method: abnormal score Determined to be a confirmed defect, Mark it as a suspected area and calculate the area of ​​the connected domain.

[0080] Step S500: Generate a defect heat map in real time based on the defect location results and visualize the results. Based on the defect area ratio and spatial distribution of the heat map, adjust the injection molding machine process parameters through the PID control algorithm.

[0081] In this embodiment, the defect heat map is calculated according to the formula Rendering, the red channel intensity is positively correlated with the anomaly score. The defect centroid is also calculated. and maximum extension distance ,when PID control module triggers parameter adjustment. Formula to calculate adjustment amount, PID coefficient , , , where the integral term time window is set to 30 minutes, and the differential term uses a sliding average filter to eliminate noise.

[0082] The visualization system is built on the WebGL engine and supports interactive viewing of 3D injection molded parts models. It can be rotated and scaled at will and overlaid with heat maps. The data dashboard integrates trend analysis functions to display the defect area change curve for the last 200 cycles. When there are three consecutive cycles, The adjustment command is sent to the injection molding machine controller via the Modbus / TCP protocol, and the parameter update delay is less than 50ms.

[0083] Example 2: Figure 2 As shown, the present invention provides an engineering plastic production quality detection system, comprising:

[0084] Image acquisition module 1 is used to collect real-time images of the surface of engineering plastics on the production line through an industrial camera and simultaneously obtain the process parameters of the injection molding machine.

[0085] Feature extraction and fusion module 2 constructs a multi-scale sliding window ViT-Adapter visual feature extraction network, and uses an adaptive attention mechanism to cross-modally fuse the surface image features of engineering plastics with the real-time process parameters of the injection molding machine to generate a fused feature vector.

[0086] Model training module 3 builds a dual-branch training network based on the fused feature vector and uses the joint loss function of improved Focal Loss and contrastive learning to train the defect classification model.

[0087] Anomaly segmentation module 4 uses the defect classification model to extract deep semantic features and combines it with the zero-sample anomaly segmentation algorithm of the WinCLIP framework to perform pixel-level positioning of surface bubbles and scratches and generate a defect mask map.

[0088] The result output module 5 generates a defect heat map in real time according to the defect location results and visualizes the results. Based on the defect area ratio and spatial distribution of the heat map, the injection molding machine process parameters are adjusted through the PID control algorithm.

[0089] 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 present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for detecting the production quality of engineering plastics, characterized in that: The method comprises: Step S100: Using an industrial camera to collect real-time images of the surface of engineering plastics on the production line and synchronously obtain process parameters of the injection molding machine; Step S200: constructing a multi-scale sliding window ViT-Adapter visual feature extraction network, and cross-modally fusing the surface image features of engineering plastics with the real-time process parameters of the injection molding machine through an adaptive attention mechanism to generate a fused feature vector; Step S300: constructing a dual-branch training network based on the fused feature vector, and using a joint loss function of improved Focal Loss and contrastive learning to train a defect classification model; Step S400: Using the defect classification model to extract deep semantic features, combined with the zero-sample anomaly segmentation algorithm of the WinCLIP framework, the surface bubbles and scratches are located at the pixel level to generate a defect mask map; The anomaly segmentation algorithm includes: Abnormal semantic vectors generated by CLIP text encoder and visual features Calculate pixel-level similarity: in is a real matrix, Pixel The anomaly score, For the image block at position The visual feature vector of Defect semantic vector generated for the text encoder, is the scaling function, the scaling factor , is the number of training cycles; Step S500: Generate a defect heat map in real time based on the defect location results and visualize the results. Based on the defect area ratio and spatial distribution of the heat map, adjust the injection molding machine process parameters through the PID control algorithm.

2. The method for testing the production quality of engineering plastics according to claim 1, wherein: The image acquisition frame rate of the industrial camera The relationship with the injection molding cycle is: in is the injection molding cycle time, As a redundancy factor, the ejection signal of the injection molding machine is obtained in real time through the encoder to trigger the acquisition.

3. The method for testing the production quality of engineering plastics according to claim 1, wherein: The feature extraction network comprises: Global features extracted by the ViT-L / 16 backbone network and sliding window local features Perform spatial alignment, is a real matrix, and feature fusion is achieved through deformable convolution: in is the fused feature tensor, For the sliding window weight coefficients, generated by the gated attention mechanism, is the variable convolution, and s is the number of sliding windows.

4. The method for testing the production quality of engineering plastics according to claim 3, wherein: The cross-modal fusion feature adopts a pyramid structure to perform cross-scale feature interaction: Constructing a 3-level feature pyramid , where the P2 layer resolution is 1 / 4 and the channel is 512; the P3 layer resolution is 1 / 8 and the channel is 256; the P4 layer resolution is 1 / 16 and the channel is 128; cross-scale feature interaction is performed through a bidirectional feature pyramid network: in For the Layer output feature map, For the Layer input feature map, is the feature weight of the current layer, is the feature weight of the adjacent layer, For the Layer input feature map, satisfying , is an upsampling or downsampling operation.

5. The method for testing the production quality of engineering plastics according to claim 1, wherein: The defect classification model training adopts an improved loss function: in and is the loss weight coefficient, is Focal Loss, is the contrast loss term: in is the similarity score of the positive sample pair, is the similarity score of the negative sample pair, n is the index of the negative sample, and the temperature coefficient , N is the number of negative samples.

6. The method for testing the production quality of engineering plastics according to claim 1, wherein: Defect classification model training uses multimodal data enhancement: Image generation conditioned on process parameters, generating synthetic data with defect labels through VAE-GAN: in For expected value calculation, is the discriminator output, For the generator with latent variables and process conditions The generated image, is the latent variable distribution output by the encoder, is the prior distribution, is the KL divergence calculation, is the KL divergence weight, is a conditional vector containing process parameters of preset dimensions.

7. The method for testing the production quality of engineering plastics according to claim 1, wherein: The result visualization includes: Map the defect heat map to the CAD model to generate a 3D quality distribution cloud map. The defect area coloring follows: Also mark the maximum defect size ; Among them is is the coordinate of the centroid of the defect area, is the set of all defective pixels.

8. The method for testing the production quality of engineering plastics according to claim 1, wherein: The injection molding machine process parameter adjustment is achieved as follows: According to the defect area ratio , calculate the correction amount of injection molding machine: The PID coefficient , , , is the proportion of defect area in the current detection cycle, is the integral of the historical defect area, is the defect area change rate.

9. An engineering plastic production quality inspection system, characterized in that: The system is used to implement the method for detecting the production quality of engineering plastics according to any one of claims 1 to 8, comprising: Image acquisition module, used to collect real-time images of the surface of engineering plastics on the production line through industrial cameras and simultaneously obtain the process parameters of the injection molding machine; The feature extraction and fusion module builds a multi-scale sliding window ViT-Adapter visual feature extraction network, which uses an adaptive attention mechanism to perform cross-modal fusion of engineering plastic surface image features and injection molding machine real-time process parameters to generate a fused feature vector. The model training module builds a dual-branch training network based on the fused feature vector and uses the joint loss function of improved Focal Loss and contrastive learning to train the defect classification model; The anomaly segmentation module uses the defect classification model to extract deep semantic features and combines it with the WinCLIP framework's zero-shot anomaly segmentation algorithm to perform pixel-level positioning of surface bubbles and scratches and generate a defect mask. The result output module generates a defect heat map in real time based on the defect location results and visualizes the results. Based on the defect area ratio and spatial distribution in the heat map, the injection molding machine process parameters are adjusted through the PID control algorithm.

Images