Quality prediction method and system based on bottle blank visual features

Abstract

The invention discloses a quality prediction method and system based on bottle blank visual features, and belongs to the technical field of quality monitoring, and the method comprises the steps: obtaining a multi-view forming image of a to-be-predicted bottle blank, carrying out the real-time timing sequence tracking through a timing sequence node, constructing a timing sequence image set in combination with a historical bottle blank, and obtaining a visual feature vector, so as to construct an image resume; generating a fusion feature vector based on the image resume; constructing a quality state vector of the to-be-predicted bottle blank, performing defect evolution prediction by using a time sequence prediction model, generating a predicted quality state vector, generating a trajectory unit through a decoding hierarchical architecture, calculating and correcting an initial failure probability, thereby generating a defect evolution trajectory, and marking a risk early warning point; a process optimization suggestion is generated by dividing granularity nodes, constructing weighted edges, generating a cross-scale association graph, constructing an association library in combination with a mapping model, constructing a directional causal association graph and carrying out causal association mining, and updating and feedback optimization are carried out on an image resume.

Description

Technical Field

[0001] This invention relates to a quality prediction method and system based on the visual features of preforms, belonging to the field of quality monitoring technology. Background Technology

[0002] In the field of bottle preform production, quality control is a key link to ensure the reliability of subsequent products. Due to its advantages of high efficiency and non-contact operation, visual inspection technology has been widely used in bottle preform quality inspection. By acquiring images of the bottle preform's appearance through image acquisition equipment, visual features are extracted, and then it is determined whether there are any obvious defects in the bottle preform. Real-time quality screening is carried out on the production site, which effectively reduces the outflow of obviously defective products.

[0003] However, existing technologies do not consider the complete correlation between the initial visual characteristics of the preform and its quality performance throughout the entire product lifecycle. Specifically, they only determine the current quality state of the preform and do not predict the quality change trajectory of the initial visual characteristics in subsequent storage, use and other stages. This leads to the failure to detect defects with delayed manifestation characteristics. At the same time, it is difficult to establish an effective correlation between local visual feature anomalies of the preform and the overall performance degradation of the final product. Furthermore, there is a lack of ability to explain why defects evolve into quality problems and the evolution mechanism. As a result, quality prediction is limited to the current state judgment level and cannot meet the production requirements for the reliability of the product throughout its entire lifecycle. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a quality prediction method and system based on the visual features of preforms. Through image time-series acquisition and image history construction, multi-feature fusion, defect evolution prediction, local-global correlation modeling, and causal tracing, a complete correlation between initial visual features and quality throughout the entire life cycle is established, delayed explicit defects are detected, evolution mechanisms are explained, and the reliability requirements throughout the entire life cycle are met.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Quality prediction methods based on the visual features of preforms include:

[0007] The process involves acquiring multi-view molding images of the preform to be predicted, using time-series nodes for real-time time-series tracking, combining historical preforms to construct a time-series image set, and obtaining visual feature vectors to build an image history.

[0008] Based on the image history, dynamic spatiotemporal features and quality gene features are obtained, and a fused feature vector is generated through normalization, dimension matching, and splicing.

[0009] The quality state vector of the preform to be predicted is constructed, the defect evolution is predicted using a time-series prediction model, the predicted quality state vector is generated, the trajectory unit is generated through a decoding hierarchical architecture, the initial failure probability is calculated and corrected, thereby generating the defect evolution trajectory and marking the risk warning points.

[0010] By dividing granular nodes and constructing weighted edges, a cross-scale association graph is generated. An association library is constructed in conjunction with a mapping model, and a directional causal association graph is constructed. Causal association mining is performed to generate process optimization suggestions and update and optimize the image history.

[0011] Specifically, the steps of real-time time tracing include:

[0012] Acquire multi-view molding images of a single preform and generate an initial image set of the preform to be predicted;

[0013] Assign a unique identification code to the preform to be predicted, establish an identification association library, and associate it with auxiliary information;

[0014] Based on the entire life cycle of the preform to be predicted, multiple time-series nodes are set, including storage stage, simulation stage, and testing stage.

[0015] Based on the multi-view molding images of each time node, the periodic time image set of the preform to be predicted is obtained;

[0016] By using scale-invariant feature transformation, the initial image set of the preform to be predicted is registered with the periodic time-series image set to generate a real-time time-series image sequence of the preform to be predicted.

[0017] Specifically, the steps of real-time time tracing include:

[0018] Collect multi-view molding images of historical preforms throughout their entire life cycle, generate a supplementary time-series image set, and register them with the initial images of historical preforms to generate a historical time-series image sequence.

[0019] A time-series image sequence is constructed, and preprocessed and multi-source features are extracted, including texture features, edge features, and gray-level distribution features, to generate the primary feature vector of the preform to be predicted;

[0020] The preprocessed time-series image sequence and primary feature vector are subjected to quality screening to generate an effective image sequence and visual feature vector of the preform to be predicted.

[0021] Using the identification code of the preform to be predicted as the core index, an image history of the preform to be predicted is constructed using a distributed storage architecture.

[0022] Set multi-level quality labels for the image history, including initial state, process state, and quality state.

[0023] Specifically, the steps for generating the fused feature vector include:

[0024] Based on the effective image sequence, linear interpolation and normalization processing are performed;

[0025] Construct a sequence recognition network and configure dilated convolutional layers to generate intermediate feature maps for multiple temporal nodes;

[0026] Using multi-level quality labels as the basis for attention guidance, initial attention weights are obtained to generate multi-temporal node feature maps;

[0027] Global average pooling is performed on the multi-temporal node feature maps, and the contribution is used to fuse and normalize them to generate dynamic spatiotemporal features.

[0028] Based on the effective image sequence, key regions are divided, and multi-level local feature maps are obtained through local window self-attention and channel attention.

[0029] Quality gene features are generated through adaptive weighted fusion and mutual information entropy screening;

[0030] The dynamic spatiotemporal features and the quality gene features are respectively subjected to layer normalization processing and dimensionality consistency verification.

[0031] By using attention-weighted concatenation and contribution weights, a fused feature vector is generated.

[0032] Specifically, the steps for defect evolution prediction include:

[0033] Based on the image history, process-aided features and decay features are obtained;

[0034] Construct a coding branch architecture, including a main branch, a first auxiliary branch, and a second auxiliary branch, which take the fused feature vector, process-aided features, and decay features as branch inputs, and output visual feature vector, process-related vector, and decay trend vector, respectively.

[0035] A preliminary fusion feature vector is generated through a dynamic gating interaction mechanism;

[0036] The preliminary fused feature vector is reconstructed and optimized to generate a quality state vector;

[0037] A time series prediction model is constructed and multiple prediction branches are set. Based on the number of time series nodes, prediction targets are configured for each prediction branch. At the same time, an evolutionary adaptive layer is configured to assign time steps and attention weights to the prediction branches.

[0038] Acquire stage characteristic factors and basic time step codes, and fuse them to generate time-series stage codes;

[0039] A progressive prediction mechanism is adopted, in which the output of the previous prediction branch is passed to the next prediction branch through residual connection, and cross-branch residual correction is set at the same time.

[0040] A composite loss function is set to iteratively train the time series prediction model and output the predicted quality state vector.

[0041] Specifically, the steps for defect evolution prediction include:

[0042] Based on a three-level structure of features, images, and semantics, a decoding hierarchical architecture is constructed to generate trajectory units;

[0043] The environmental visual features in the image history are obtained and combined with the predicted quality state vector. The initial failure probability of each time node is calculated through multi-level logistic regression.

[0044] Based on historical preform samples and environmental visual features, the initial failure probability is corrected to obtain the failure probability of each time-series node.

[0045] Based on the trajectory units and failure probabilities of each time node, the defect evolution trajectory of the preform to be predicted is generated.

[0046] Set a failure probability threshold, traverse the defect evolution trajectory, and filter risk warning points and risk persistence stages;

[0047] A similar defect verification mechanism is set up to correct the defect evolution trajectory based on similarity.

[0048] Specifically, the steps to build the associated library include:

[0049] The risk area and critical area of ​​the preform to be predicted are obtained, and secondary nodes are divided, including coarse-grained nodes and fine-grained nodes, and a node association table is generated.

[0050] Construct static basic features and dynamic evolution features for each granularity node, and bind them to the granularity node;

[0051] Configure fusion weights for each granularity node, and generate granularity node features based on the fusion weights and layer normalization processing;

[0052] Construct the association attributes of edges, generate weighted edges using weighted summation, and simultaneously construct cross-scale edges and jump edges;

[0053] A cross-scale correlation graph is generated based on the mutual information entropy between the failure probability and the fused feature vector.

[0054] Construct a three-order inference architecture and configure an evolution-oriented message passing mechanism;

[0055] Multi-strategy aggregation and risk weighting are applied to global layer node features to generate graph-level relational feature vectors.

[0056] Construct a mapping model to generate quantitative indicators of performance degradation and the performance impact contribution of each node;

[0057] Calibration is performed using a time-series decay calibration factor and time-series consistency loss.

[0058] Calculate the matching degree between the predicted values ​​of the preform performance and the actual values ​​of historical samples, and construct an association library.

[0059] Specifically, the steps of causal association mining include:

[0060] Based on the cross-scale correlation graph, the nodes are updated with image features and causal candidate factors;

[0061] An initial causal relationship graph is constructed using multi-view shaped images and the edge weights of weighted edges;

[0062] The structure of the initial causal relationship graph is learned by using an image feature-adaptive PC algorithm.

[0063] A graph causal encoder is constructed, which captures the differences in causal transmission of image features at different evolutionary stages through a multi-head self-attention mechanism, and generates a directional causal association graph.

[0064] Based on causal CAM, image mapping is performed on the directional causal relationship graph to generate a causal significance heatmap;

[0065] Adaptive threshold segmentation is used to extract causal salient regions, and causal driving features are screened by combining risk region images to generate a three-dimensional correlation report.

[0066] Specifically, the steps in causal association mining also include:

[0067] A visual mapping library of process parameters and visual features is constructed. Abnormal visual features, corresponding causal links, and associated process parameter ranges of the visual mapping library are matched to the preform to be predicted. The contribution ratio of abnormal visual features to defect evolution is calculated.

[0068] Calculate the visual process calibration factor, generate process optimization suggestions, calculate the confidence level of the process optimization suggestions, and retain only process optimization suggestions whose confidence level is greater than a preset confidence evaluation threshold;

[0069] Based on the directional causal relationship graph, the image history of the preform to be predicted is upgraded causally. Based on the distributed storage architecture, the newly added layer is structurally associated with the original data to form a four-dimensional image history.

[0070] Feedback optimization is performed based on the four-dimensional image history, and a dynamic mapping model is constructed to generate the final quality status label of the preform to be predicted.

[0071] The quality prediction system based on the visual features of bottle preforms includes: an image acquisition module, a feature fusion module, a defect evolution module, a correlation analysis module, and a causal optimization module.

[0072] The feature fusion module is used to extract dynamic spatiotemporal features and quality gene features. After normalizing and matching the dynamic spatiotemporal features and quality gene features, attention-weighted concatenation is performed to generate a fused feature vector.

[0073] The feature fusion module is used to extract dynamic spatiotemporal features and quality gene features. After normalizing and dimension matching the two types of features, attention-weighted concatenation is performed to generate a fused feature vector.

[0074] The defect evolution module is used to extract process-aided features and degradation features, generate a quality state vector through an encoding branch architecture, generate a defect evolution trajectory using a time-series prediction model and a progressive prediction mechanism, correct the failure probability with historical data and environmental features, and mark risk warning points.

[0075] The association analysis module is used to divide granularity nodes, construct cross-scale association graphs, and build an association library through a three-order inference architecture and evolution-oriented message passing;

[0076] The causal optimization module is used to construct a directional causal relationship graph, perform causal relationship mining, generate process optimization suggestions, and update and optimize the image history.

[0077] The beneficial effects of this invention are:

[0078] By generating a complete time-series image set through a ring-shaped acquisition array and time-series tracking, and combining image history with multi-level quality labels, full lifecycle data traceability is achieved, avoiding data distortion and feature loss, laying the foundation for long-term quality correlation. Through the fusion extraction of dynamic spatiotemporal features and quality gene features, early defect signals invisible to the naked eye are captured in advance, solving the problem of missed detection of delayed explicit defects. A time-series prediction model generates defect evolution trajectories with risk labels, clarifying the trend of quality changes and overcoming the limitation of only being able to determine the current state. Through cross-scale correlation modeling, a quantitative correlation between local visual feature anomalies and overall performance degradation is established, filling the gaps in correlation. Through causal correlation mining and process visual matching, the defect evolution mechanism is explained and targeted optimization suggestions are generated, forming a closed loop in the entire process. This realizes the upgrade of preform quality from current state judgment to full lifecycle prediction, ensuring the accuracy of quality assessment, the foresight of prediction, and the pertinence of control, fully meeting the production needs for product reliability throughout the entire lifecycle, and improving the efficiency of quality control and the intelligence of process optimization. Attached Figure Description

[0079] Figure 1 Here is a flowchart of a quality prediction method based on the visual features of preforms;

[0080] Figure 2 This is a flowchart of real-time timing tracking in this invention;

[0081] Figure 3 This is a flowchart of the defect evolution prediction process in this invention;

[0082] Figure 4 This is a structural diagram of a quality prediction system based on the visual features of preforms. Detailed Implementation

[0083] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0084] Example 1:

[0085] refer to Figures 1 to 3 As shown in the figure, this embodiment introduces a quality prediction method based on the visual features of bottle preforms, including the following steps:

[0086] Step S1: Simultaneously acquire multi-view molding images of the preform to be predicted through a ring acquisition array. Utilize the configured time-series nodes to perform real-time time-series tracking of the preform to be predicted, generating a periodic time-series image set. Combine this with historical preforms to construct a supplementary time-series image set, integrate them to generate a time-series image set, and perform feature extraction and quality screening to generate a visual feature vector. Combine this with the identification code of the preform to be predicted, record the interference factors of the injection blow molding process, monitor the status of the controlled object, and structurally associate the visual feature vector and image data to construct an image history and set multi-level quality labels. This enables traceability of data throughout the entire life cycle of a single preform, accurate defect identification, and dynamic tracking, avoiding image distortion, feature loss, and invalid data. It accurately represents the quality status at each stage, supports long-term quality performance prediction, and ensures the accuracy and completeness of preform quality assessment.

[0087] Step S2: Extract intermediate features from multiple temporal nodes using the constructed sequence recognition network, and enhance temporal attention with multi-level quality labels to generate dynamic spatiotemporal features. Delineate key regions using the mechanical properties of the preform structure, and extract multi-level local features through local window self-attention and channel attention. Generate quality gene features through adaptive weighted fusion and mutual information entropy filtering to capture early defect signals. Normalize and match the dynamic spatiotemporal features and quality gene features, and use attention weighted splicing to generate a fusion feature vector of the preform to be predicted. This achieves complementary advantages between spatiotemporal features and local features, eliminates redundant information, reduces computational load, and identifies late-prominent defects that are not visible to the naked eye in advance, ensuring the accuracy and foresight of preform quality prediction.

[0088] Step S3: Obtain process auxiliary features and decay features, generate quality state vectors using the coding branch architecture, predict defect evolution using the time series prediction model, use the quality state vector as input, generate predicted quality state vectors for each time series node through progressive prediction, generate trajectory units with semantic annotations through the decoding hierarchical architecture, calculate the initial failure probability, correct the failure probability through historical data and environmental features, thereby generating the defect evolution trajectory and marking risk warning points, achieving accurate prediction of preform defect evolution and efficient support for production quality control;

[0089] Step S4: By dividing granular nodes and constructing weighted edges, a cross-scale association graph is generated. The defect association is captured by combining a three-order inference architecture and evolution-oriented message passing. After aggregating and refining graph-level features, the performance degradation is mapped through a hybrid model. Time-series calibration and experimental verification ensure the consistency of prediction. The degradation quantification index and regional impact contribution are output. An association library is constructed to achieve accurate prediction, process traceability and quality control support.

[0090] S5: Construct a directional causal relationship graph, mine causal relationships, and combine comparative learning and matching of process parameters and visual features to locate abnormal visual features and corresponding causal links and related process parameters, thereby generating process optimization suggestions. Introduce a causal label layer and an optimization trajectory layer to update the image history and provide feedback to optimize feature extraction, time-series prediction models, cross-scale relationship graphs, and image acquisition parameters. This enables a preform quality control and process optimization system that accurately traces preform defects, intelligently optimizes processes, and efficiently controls and predicts quality.

[0091] In this application, taking injection blow molding as an example, there are interference factors from plastics, equipment, and environment during the injection blow molding process. These interference factors are not generated by the normal operation of the system itself. Among them, plastic disturbances include melt viscosity fluctuations caused by batch changes of raw materials, equipment disturbances include injection pressure fluctuations caused by hydraulic system leaks, and environmental disturbances include sudden changes in workshop temperature and humidity. The status monitoring data of the controlled object includes cavity pressure, melt temperature, mold temperature, real-time product wall thickness, blow molding air pressure, and injection screw speed. These disturbances have a significant impact on the injection blow molding process and product quality. Therefore, by constructing a preform quality control and process optimization system, the quality problems caused by disturbances can be identified and corrected.

[0092] Specifically, the steps of real-time time tracing include:

[0093] For each individual preform after blow molding, multiple industrial cameras configured in the production line are used to initially acquire multi-view molding images of the preform through a synchronous triggering mechanism. This generates an initial image set for the preform to be predicted, avoiding image blurring or feature distortion caused by preform movement and ensuring that the initial images can completely and accurately reflect the original state of the preform after molding. The multiple industrial cameras configured in the preform production line form a circular acquisition array, covering all views of the preform, avoiding feature loss caused by a single view. At the same time, the industrial cameras are synchronously calibrated before acquiring the molding images to eliminate lens distortion and view deviation between different cameras, ensuring the geometric consistency of images from each view.

[0094] A unique identification code is assigned to each preform to be predicted, and laser marking technology is used to mark the identification code on non-critical stress areas of a single preform, such as the edge of the bottle mouth. An identification association library is established to simultaneously record and associate auxiliary information related to the production of the preform to be predicted, such as the production batch of a single preform, blow molding process parameters, and raw material batches, so as to achieve traceability of data throughout the entire life cycle of a single preform.

[0095] Based on the entire lifecycle of the preform to be predicted, multiple time-series nodes are set, including storage, simulation, and testing stages. Each stage has multiple time-series nodes to record quality changes at different stages. Each time-series node uses the same parameters as the initial acquisition to ensure the uniformity of image acquisition conditions at each stage and avoid feature comparison errors caused by differences in equipment or parameters. This allows for the acquisition of multi-view molding images of the preform to be predicted at each time-series node for real-time time-series tracking. The quality status of the preform at each time-series node is recorded synchronously, including whether defects have expanded, whether failure has occurred, and its specific manifestations. These images are then bound to the corresponding multi-view molding images to generate a periodic time-series image set for the current preform to be predicted. The periodic time-series image set only contains the measured data of the time-series nodes that have already occurred and does not include images of time-series nodes that have not occurred. At the same time, the periodic time-series image set is dynamically supplemented as the preform to be predicted progresses through its lifecycle.

[0096] Collect multi-view molding images of the same type of preform produced in the production line throughout its entire life cycle, including initial images after injection molding, time-series images at each stage, and images after failure. Simultaneously record and associate auxiliary information, quality status throughout the entire life cycle, and final quality performance of the historical preforms. This serves as a supplementary time-series image set for supervised learning, filling the gap in the correlation between initial visual features and long-term quality performance.

[0097] Using scale-invariant feature transformation, the initial image set of the preform to be predicted is registered with the periodic time series image set. Local feature points are detected and feature descriptors are generated. Homologous feature point pairs are selected through feature matching. Image spatial alignment is achieved based on homography matrix estimation to eliminate the viewing angle deviation caused by preform placement offset and slight deformation, thereby generating a real-time time series image sequence of the preform to be predicted. Similarly, the initial images of historical preforms are registered with the supplementary time series image set to generate a historical time series image sequence. The historical time series image sequence is then associated with the real-time time series image sequence of the preform to be predicted to generate a time series image sequence.

[0098] To eliminate the impact of the industrial environment on image quality, the time-series image sequence is preprocessed, including: using a nonlocal mean denoising algorithm to remove random noise from the industrial environment and preserve image detail features; decomposing the reflection component and illumination component of the image based on the multi-scale retinal enhancement theory to eliminate residual reflection and brightness deviation caused by uneven illumination and restore the true grayscale information of the image; and using an adaptive histogram equalization algorithm to enhance the grayscale contrast of small defect areas while avoiding overexposure in local areas, making small defects easier to identify, such as microbubbles and subtle texture anomalies.

[0099] Based on the preprocessed time-series image sequence, texture features are extracted using the local binary method to capture local gray-level distribution patterns. Edge features are extracted using the Sobel operator to characterize the bottle preform outline and defect edge information. The gray-level co-occurrence matrix is ​​calculated, and energy, entropy, contrast, and correlation are extracted to generate gray-level distribution features. After normalizing various features, they are spliced ​​together to form the primary feature vector of the bottle preform to be predicted.

[0100] The preprocessed time-series image sequence and primary feature vectors are quality-screened. Invalid data such as blurry, overexposed / underexposed images are removed by the image signal-to-noise ratio. The feature validity is tested by the feature variance. Valid feature vectors are retained, and finally, the valid image sequence and visual feature vector of the preform to be predicted are generated.

[0101] Using the identification code of the preform to be predicted as the core index, the periodic time series image set, effective image sequence, visual feature vector, and auxiliary information are structurally associated. A distributed storage architecture is used to construct the image history of the preform to be predicted. The history of the current preform to be predicted is continuously updated as the life cycle progresses, and the history of historical samples is complete full-cycle data.

[0102] Multi-level quality labels are set for image history, including initial state, process state, and quality state, to avoid the problem that a single label cannot accurately reflect quality changes. Among them, the initial state labels include no defects, microbubbles, uneven texture, minor edge deformation, and local grayscale anomalies; the process state labels include no defect expansion, slow defect expansion, rapid defect expansion, and stable defects; and the final quality state labels include normal service, failure at a specific time, and substandard performance.

[0103] Specifically, the steps for generating the fused feature vector include:

[0104] The corresponding valid image sequence is retrieved from the image history according to the identification code of the preform to be predicted. This includes the measured sequence of the preform to be predicted as the node that has occurred and the complete full-cycle sequence of the same type of historical preform. In order to avoid the difference in the number of time nodes of different preforms, the sequence length is unified by the time-series linear interpolation completion method. The pixel values ​​of the images in the valid image sequence are normalized to eliminate the numerical interference caused by illumination fluctuations and equipment response differences, so as to ensure the stability of feature extraction and generate a standardized valid image sequence.

[0105] 3D ResNet is selected as the backbone network for spatiotemporal feature extraction. A sequence recognition network is constructed by stacking three-dimensional residual blocks to alleviate the gradient vanishing problem in deep network training and ensure the effective transmission of long-term features. Dilated convolutional layers are configured in the sequence recognition network, and different dilation rates are set to adapt to different time spans. This expands the feature receptive field without increasing the computational load, ensuring that the defect evolution trend can be captured from short storage stage to long durability testing stage. With effective image sequences as input, intermediate feature maps of multiple time nodes are output.

[0106] Using multi-level quality labels as the basis for attention guidance, the intermediate feature map is compressed into a single-channel response map, and the matching degree between each time-series node and the key stages of defect evolution (first manifestation of defect, rapid expansion, and stabilization) is calculated. If the node label is rapid expansion of defect, the matching degree is assigned a high value; if the label is no expansion of defect, the matching degree is assigned a low value. The matching degree is converted into an initial attention weight, which is then normalized and multiplied element-wise with the intermediate feature map of the corresponding time-series node to strengthen the features of the key evolution stages and suppress redundant node features without defect changes, thereby generating a multi-time-series node feature map.

[0107] Global average pooling is performed on the feature maps of multiple time-series nodes to compress the high-dimensional feature map of each time-series node into a fixed-dimensional feature vector, retaining the core spatiotemporal features of the node. A fully connected fusion layer is constructed, and the feature vectors of all time-series nodes are input in chronological order. The contribution of different node features to the overall evolution trend is learned through the weight matrix, and a high-dimensional vector with uniform dimension is output and normalized to generate dynamic spatiotemporal features.

[0108] Based on effective image sequences and combined with the mechanical properties of bottle preform structures, key regions prone to failure are identified and cropped to avoid invalid information from non-critical regions consuming computational resources. The images of key regions are divided into non-overlapping local image blocks and converted into embedding vectors. The embedding vectors are then divided into multiple overlapping local windows, and self-attention is calculated only within each local window to reduce computational load. At the same time, cross-window attention connects adjacent windows to ensure the coherence between local and global features and avoid feature fragmentation caused by window segmentation. Channel-dimensional attention weighting is applied to the feature map of each local window to automatically strengthen channel features related to micro-defects and suppress irrelevant channels. Finally, multi-level local feature maps are generated, including low-level detailed features, mid-level abstract features, and high-level global features.

[0109] By using adaptive weighted multi-scale fusion, learnable weight parameters are assigned to each layer of feature map. By supplementing the training with the final quality label of the time series image set, the contribution of each layer of features is adaptively adjusted. Multi-level local feature maps are spliced ​​along the channel dimension into a fusion feature tensor. With the final quality label as the target, the mutual information entropy between each feature dimension and the target variable in the fusion feature tensor is calculated. Feature dimensions with mutual information entropy higher than the adaptive threshold are selected, and redundant and irrelevant feature dimensions are removed. The selected feature dimensions are recombined to form the quality gene features of the current preform to be predicted and the historical preforms, so as to capture the early signal of delayed explicit defects that are invisible to the human eye in advance.

[0110] Dynamic spatiotemporal features and quality gene features are subjected to layer normalization to eliminate numerical scale differences between different feature dimensions, and dimension consistency is checked. If the two types of feature dimensions are inconsistent, the low-dimensional features are mapped to the same dimension as the high-dimensional features through a fully connected layer to ensure dimension matching during fusion.

[0111] By using attention-weighted concatenation, the contribution weights of two types of features to quality prediction are calculated through a single hidden layer neural network. The two types of feature vectors are weighted separately and then concatenated along the feature dimension to generate a fused feature vector, thereby achieving complementary advantages of the two types of features.

[0112] Specifically, the steps for defect evolution prediction include:

[0113] If we rely solely on visual features and ignore the two implicit evolutionary factors of process influence and natural degradation, it is difficult to fully characterize the evolutionary potential of the preform. Based on the image history of the preform to be predicted, we extract visual correlation features from auxiliary information and initial images, such as the visual representation of preform texture and density corresponding to process parameters, to generate process auxiliary features. At the same time, we fit and calculate the quality degradation factor based on the grayscale change rate and defect area growth rate of time-series images as degradation features to reflect the natural degradation trend of the preform.

[0114] The coding branch architecture is constructed, including a main branch, a first auxiliary branch, and a second auxiliary branch. The main branch adopts a deep residual network, takes the fused feature vector as input, and extracts the visual feature vector through nonlinear mapping. The first auxiliary branch adopts a gated recurrent unit, takes the process auxiliary feature as input, strengthens the temporal correlation between process parameters and visual features, and thus obtains the process correlation vector. The second auxiliary branch adopts a fully connected layer, takes the decay feature as input, accurately encodes the natural decay trend of the material, and thus obtains the decay trend vector.

[0115] The cross-correlation coefficient between the main branch vector and the auxiliary branch vector is calculated through a dynamic gating interaction mechanism. The dynamic gating value is generated through the Sigmoid activation function to configure the fusion ratio of each branch feature and perform weighted summation to obtain the preliminary fused feature vector.

[0116] An adversarial autoencoder is used to reconstruct and optimize the initial fused feature vector. The reconstruction loss and adversarial loss are used as optimization targets. Through adversarial training between the generator and the discriminator, redundant information is removed and the core representation is strengthened, thereby generating a quality state vector.

[0117] A multi-scale adaptive time-series prediction model based on Transformer is constructed. Taking the quality state vector as input, multiple prediction branches are set, including short-term, medium-term, and long-term, to adapt to the differences in defect evolution speed at different stages. At the same time, based on the number of time-series nodes, corresponding prediction targets are configured for each prediction branch. For example, the short-term branch outputs the quality state of the next 1-3 time-series nodes, the medium-term branch outputs 4-8 time-series nodes, and the long-term branch outputs more than 9 time-series nodes to ensure full life-cycle coverage. Among them, the short-term corresponds to the storage stage, the medium-term corresponds to the early stage of use, and the long-term corresponds to the later stage of use.

[0118] An evolutionary adaptive layer is configured in the time series prediction model to dynamically match the evolutionary characteristics of different defect types. Based on image history, historical time series image sequences of historical preforms are extracted, evolutionary speed features of different defect types are extracted, and evolutionary speed indicators corresponding to defect types are classified and statistically analyzed, such as defect area growth rate and length expansion rate. A defect type-evolutionary speed mapping library is constructed as the decision basis for the evolutionary adaptive layer. Based on the initial defect type of the preform to be predicted and the evolutionary speed of the nodes that have occurred, the corresponding evolutionary rules are matched from the mapping library. The time step and attention weight are dynamically allocated to the three prediction branches. For defects with slow evolution, the long-term branch is emphasized, and for defects with rapid evolution, the short-term branch is strengthened, which solves the problem of insufficient accuracy caused by the one-size-fits-all approach of traditional multi-branch prediction.

[0119] Based on the environmental and stress characteristics of each stage of the preform to be predicted, such as low stress in the storage stage and high stress in the use stage, stage characteristic factors are obtained, and sine and cosine position coding is adopted. Time progression relationship is injected to generate basic time step coding. Time step information and stage characteristics are fused to generate time-series stage coding, thereby injecting time-series stage coding into each prediction branch.

[0120] A progressive prediction mechanism is adopted, in which the output of the previous prediction branch is passed to the next prediction branch through residual connection. The short-term prediction branch takes the quality state vector as input and outputs the quality state vector at the next moment as input to the medium-term prediction branch. At the same time, cross-branch residual correction is set, and the information interaction between branches is realized through cross-branch attention mechanism. The advantages of different prediction branches are used to correct the prediction bias of a single branch, avoid the accumulation of errors, and generate the predicted quality state vector after correction of each prediction branch.

[0121] A composite loss function is set up, which includes stage alignment, trajectory smoothing, and prediction accuracy. The composite loss function is used as the optimization objective. The time series prediction model is iteratively trained using the quality state vector of historical preforms and the true labels until the loss function converges. Finally, the predicted quality state vector of each time series node is output. Among them, stage alignment is the mean square error between the predicted quality state vector and the true time series quality state vector to ensure prediction accuracy. Trajectory smoothing is to constrain the difference between the predicted quality state vectors of adjacent time nodes to avoid jump distortion. Prediction accuracy is the stage alignment loss based on KL divergence to constrain the consistency between the predicted features and the true feature distribution of the corresponding stage to ensure the balanced prediction accuracy of each stage.

[0122] Based on a three-level structure of features, images, and semantics, a hierarchical decoding architecture is constructed. Using measured images of historical preforms and defect labels as supervision signals, the architecture minimizes the pixel-level loss between the predicted image and the real image, and the classification loss between the semantic annotation and the real label, thereby generating single-temporal-node trajectory units with semantic annotations. Specifically, the first-level decoding uses a transposed convolutional network to restore the predicted quality state vector of each temporal node to the predicted visual image, intuitively presenting the morphological evolution of defects. The second-level decoding embeds a defect semantic annotation layer and trains a semantic classifier based on defect type labels to automatically identify and label the defect type and evolution state in the predicted image. The third-level decoding is used to bind the predicted visual image and semantic annotation of a single temporal node to form an independent trajectory unit.

[0123] The environmental visual features in the image history are obtained, and combined with the predicted quality state vectors of each time series node, the initial failure probability of each time series node is calculated through multi-level logistic regression.

[0124] Based on the mapping relationship between the initial failure probability and the actual failure result of historical preform samples, the original probability is corrected. Based on the environmental visual characteristics, the environmental impact coefficient is calculated by statistically correlating the environmental characteristics and failure rate of historical preform samples to make a secondary adjustment to the failure probability, thereby obtaining the failure probability of each time node to match the failure risk in the actual environment.

[0125] Based on the time sequence, the trajectory units and failure probabilities of all time nodes are concatenated to generate the defect evolution trajectory of the preform to be predicted, and the timestamp and predicted visual image of each time node are marked.

[0126] Based on the statistical analysis of critical failure probabilities of historical preform samples, a failure probability threshold is set. By traversing each time node in the defect evolution trajectory, the time node where the failure probability first exceeds the failure probability threshold is marked as a risk warning point, and the time node where the failure probability continues to exceed the failure probability threshold is marked as a risk persistence stage, thus providing a clear warning signal for production quality control.

[0127] A defect verification mechanism is set up. Based on the real evolution trajectory in historical preform samples, clustering is performed according to defect type. The dynamic time regularization similarity between the predicted trajectory of the preform to be predicted and the corresponding defect type cluster center trajectory is calculated. If the similarity is higher than the set threshold, it indicates that the predicted trajectory conforms to the typical pattern, and the predicted trajectory is directly retained. If the similarity is lower than the set threshold, it indicates that the predicted trajectory has a deviation. The 1-2 real historical trajectories that are most similar to the predicted trajectory in the clustering library are called, their evolutionary features are extracted, and the corresponding nodes of the predicted trajectory are corrected for a second time to ensure that the trajectory conforms to the actual evolution law, thereby generating the verified defect evolution trajectory.

[0128] Specifically, the steps to build the associated library include:

[0129] For the preform to be predicted, the local areas corresponding to the risk warning points are obtained and marked as risk areas. Combined with the divided key areas, secondary nodes are divided, including coarse-grained nodes and fine-grained nodes. Through multi-view image registration results, the inclusion relationship between coarse-grained nodes and fine-grained nodes is clarified, forming a node association table to ensure information exchange between regions of different scales. Specifically, based on the functional modules of the preform to be predicted, such as the bottle mouth, bottle body, bottle bottom, and bottle neck, coarse-grained nodes are divided to cover all physical areas of the preform to be predicted, ensuring the global representation of overall performance. On the basis of coarse-grained nodes, risk areas and key areas are refined, and fine-grained nodes are formed with local image blocks corresponding to quality gene features as the smallest unit to accurately capture the location and morphology of micro-defects.

[0130] Based on the quality gene features and primary feature vectors of the preform to be predicted, static basic features of each granularity node are constructed. Based on the defect evolution trajectory and dynamic spatiotemporal features, dynamic evolution features of each granularity node are constructed and bound to the corresponding granularity node.

[0131] By using a single hidden layer neural network and combining risk warning points, fusion weights are configured for each granularity node. If a certain granularity node of the preform to be predicted contains a risk warning point, it indicates that the current granularity node is a high-risk node. At this time, the proportion of dynamic evolution features is strengthened. Otherwise, the current granularity node is a low-risk node. At this time, static basic features are strengthened. Based on the fusion weights and layer normalization processing, granularity node features are generated to eliminate the numerical scale differences of different node features, ensure that all node features are comparable, and avoid the limitation that a single static feature cannot reflect the changes in the correlation during the evolution process.

[0132] For the set of granular nodes of the preform to be predicted, the association attributes of the edges are constructed, including spatial association, structural association, evolutionary association, and feature association. The association attributes are fused by weighted summation to generate weighted edges. At the same time, cross-scale edges and long-distance jump edges are constructed. Among them, spatial association is based on the physical position relationship of the multi-view molding image, structural association reuses the mechanical transmission path of the preform structure, evolutionary association is based on the expansion path in the defect evolution trajectory, and feature association is obtained by calculating the similarity of the fused feature vectors corresponding to any two granular nodes. Cross-scale edges are the edges between coarse-grained and fine-grained nodes, and long-distance jump edges are the edges between non-adjacent nodes but related in evolutionary transmission.

[0133] Based on the mutual information entropy of failure probability and fused feature vector, invalid nodes with failure probability below the failure probability threshold and extremely low feature information are eliminated, such as non-critical region nodes without defects. Weakly related edges with weights below the weight threshold are also eliminated, such as long-distance edges between non-critical regions without evolutionary transmission. Redundant edges with highly similar weights are merged to generate a cross-scale association graph.

[0134] A three-tiered reasoning architecture encompassing micro, local, and global layers is constructed to adapt to short-term, medium-term, and long-term evolutionary stages. The micro-level layer employs a graph convolutional network to focus on fine-grained nodes and capture the local transmission relationships of micro-defects, adapting to the characteristics of slow short-term evolution. The local layer uses a graph attention network to connect fine-grained and coarse-grained nodes, strengthening the reasoning about the impact of high-risk nodes on adjacent regions through an attention mechanism, adapting to the characteristics of stable medium-term evolution. The global layer uses a graph Transformer, based on coarse-grained nodes and long-distance jump edges, to capture long-distance evolutionary transmission relationships across modules, adapting to the characteristics of accelerated long-term evolution.

[0135] An evolution-oriented message passing mechanism is configured. At the micro level, the neighborhood aggregation weight is adjusted based on the defect expansion rate of short-term evolution. The faster the expansion rate, the higher the contribution of the neighborhood nodes. Cross-scale message passing aggregates micro-layer features to the corresponding coarse-grained nodes through average pooling. After fusing the risk warning point correction features, the local layer node features are updated. Local-global message passing calculates attention weights through GAT to highlight the influence of risk warning point nodes. At the global layer, multi-head self-attention of graph Transformer is used to capture long-distance evolutionary transmission and record the defect influence path graph during the transmission process. The contribution source of each node feature update is clarified, making the correlation transmission process concrete and traceable.

[0136] Multi-strategy aggregation and risk weighting are adopted for global layer node features. Mean pooling captures the global average association, maximum pooling highlights the strong association of high-risk nodes, and attention-weighted pooling allocates aggregation weights based on failure probability. After concatenating the three pooling results, feature refinement is carried out through autoencoder to remove redundant information, strengthen the core association features that are strongly related to the overall performance, and generate graph-level association feature vectors.

[0137] A mapping model between graph-level features and dynamic performance degradation is constructed. An architecture combining multilayer perceptron and temporal convolutional network is adopted. The multilayer perceptron is used to capture the nonlinear static correlation between graph-level associated feature vectors and overall performance degradation, while the temporal convolutional network models the temporal dynamic degradation law, adapts to the temporal evolution characteristics, and avoids the bias caused by static mapping. The graph-level associated feature vectors and defect evolution trajectory are used as inputs to output quantitative indicators of performance degradation, such as remaining service time, degree of performance degradation, and performance impact contribution of each node, quantifying the proportion of the impact of local areas on overall performance.

[0138] Based on the measured quality status and environmental impact coefficient of the preform nodes that have occurred, the decay rate corresponding to the defect type and evolution stage is matched as the time-series decay calibration factor. The time-series consistency loss constrains the continuity of the performance prediction values ​​of adjacent time-series nodes to avoid jump distortion. The time-series decay calibration factor and time-series consistency loss are used for calibration to ensure that the performance prediction and evolution prediction are consistent.

[0139] Based on the measured data of the preform to be predicted and the complete data of historical preform samples, the matching degree between the predicted performance value of the preform to be predicted and the actual value of the historical samples is calculated. The contribution of the node performance of the preform to be predicted is visualized and overlaid with the actual failure location of the historical preform samples and the predicted risk area of ​​the preform to be predicted, so as to intuitively verify the matching degree of the key influence area. Historical cases of similar defects in the image history are retrieved and the consistency between the current association results and historical cases is compared. A defect-path-performance association library is constructed to store and verify the effective association relationship, providing a basis for subsequent quality control.

[0140] Specifically, the steps of causal association mining include:

[0141] Based on the cross-scale correlation graph of the preform to be predicted, the causal correlation graph is initialized and constructed. The node attributes in the causal correlation graph are updated with image features and causal candidate factors. Among them, the causal candidate factors include quality gene features, dynamic spatiotemporal features, defect evolution trajectory and node performance influence contribution.

[0142] By utilizing the physical location relationships of multi-view shaped images and the edge weight matrix of weighted edges, high-association-strength edges that meet the association strength criteria are selected, and an initial causal association graph is constructed to ensure that all nodes and edges in the graph revolve around image processing-derived data.

[0143] Using an image feature-adaptive PC algorithm, structural learning is performed on the initial causal relationship graph, including: using the fused feature vectors of historical bottle blank samples, defect evolution trajectories, and cross-scale relationship graphs as training sets, and through conditional independence testing, mutual information verification based on kernel functions, only inputting image feature vectors, and eliminating pseudo-relational edges, such as edges that are only spatially adjacent but have no causal transmission of image features.

[0144] Based on Transformer, a graph causal encoder is constructed. Taking the image feature sequence and evolution stage labels of historical preform samples as input, a multi-head self-attention mechanism is used to capture the differences in causal transmission of image features in different evolution stages, including short-term, medium-term and long-term stages, thereby generating an optimized directional causal association graph and clarifying the complete directional causal link of image features, defect evolution and performance degradation.

[0145] Based on causal CAM, causal relationships are mapped to intuitive images. Taking the core image features in the directional causal relationship map, such as microscopic features of microbubbles and crack edge features, as input, the activation region of the feature in the original multi-view molding image is calculated through gradient attribution analysis to generate a causal significance heatmap.

[0146] By using adaptive threshold segmentation, causal salient regions, such as the image region corresponding to microbubbles, are extracted and superimposed with the risk region image in the defect evolution trajectory for verification. Causal driving features with an overlap greater than the overlap threshold between causal salient regions and risk regions are retained, thereby generating a three-dimensional correlation report that clarifies the specific physical region in the original image, the corresponding image features, and the defect evolution link it triggers.

[0147] Based on the process-aided features of historical preform samples, a mapping pair between process parameter vectors and visual feature vectors is constructed. A visual mapping model is trained using contrastive learning. The process parameter vectors of historical preform samples and the corresponding multi-view molding images are used as inputs, and similar visual features corresponding to similar process parameters are used as training targets. This generates a visual mapping library of process parameters and visual features. Each record contains the range of process parameters, typical visual feature images, and feature vectors.

[0148] By combining the directional causal relationship graph, process-sensitive causal links are screened, such as injection molding temperature texture features → microbubble features → defect expansion, and incorporated into the visual mapping library to form a four-level mapping relationship including process parameters, visual features, causal links, and defects.

[0149] Based on the multi-view molding images, quality gene features, and defect evolution trajectories of the preform to be predicted, a visual mapping library is matched to locate abnormal visual features, corresponding causal links, and associated process parameter ranges. At the same time, the contribution ratio of abnormal visual features to defect evolution is calculated based on the causal link strength matrix to clarify the core source tracing target.

[0150] Based on the deviation between abnormal visual features and standard visual features in the visual mapping library, and combined with the process adjustment coefficients of historical preform samples, a visual process calibration factor is calculated to obtain the direction and magnitude of process parameter adjustment. At the same time, all optimized images must have clear visual feature indicators to avoid blind process adjustment, thereby generating process optimization suggestions.

[0151] By combining the correlation data of historical preform sample points, and based on the logistic regression model, the confidence level of the process optimization suggestions is calculated using causal link strength, mapping library matching degree, and visual deviation value. Only process optimization suggestions with a confidence level greater than the preset confidence assessment threshold are output.

[0152] Based on the directional causal relationship graph and defect tracing results, the image history of the preform to be predicted is upgraded to a causal model. On the basis of the original history data, a causal label layer and an optimized trajectory layer are added. The causal label layer labels the images and features in the history with causal attributes, including core visual feature labels, causal link labels, and process-related visual labels. The optimized trajectory layer records the execution status of process optimization suggestions, the optimized multi-view forming images, and the comparison results of the fused feature vectors before and after optimization.

[0153] Based on a distributed storage architecture, the identification code of the preform to be predicted is used as the core index to structurally associate the new layer with the original data, forming a four-dimensional image history.

[0154] Feedback optimization is performed based on four-dimensional image history. For the fusion feature vector of the preform to be predicted, the core visual features identified are selected as the key features for extraction. The attention weighted fusion parameters are adjusted, and causal feature enhancement is added to ensure that causal driving features are preferentially retained in the fusion feature vector. For the time-series prediction model, the causal link strength in the directional causal association graph is used as the basis for adjusting the attention weight of multi-scale prediction. The composite loss function is optimized. For the cross-scale association graph, the directional links in the causal association graph are used as the priority edges of the cross-scale association graph and given higher initial weights. The calculation logic of the contribution of node performance is iteratively updated and the causal effect coefficient is incorporated. At the same time, for the image acquisition parameters, the camera calibration parameters and the acquisition density of time-series nodes are adjusted based on the visual feature improvement data after process optimization.

[0155] Based on the four-dimensional image history of historical preform samples, a dynamic mapping model containing image features and quality status is constructed. The core visual feature vector and defect evolution image index are used as inputs, and the output is the final quality status label.

[0156] By combining the evolution trend of time-series image features, an adaptive thresholding algorithm is used to dynamically update all thresholds in quality prediction, such as generating adaptive thresholds in fused feature vectors and failure probability thresholds in defect evolution prediction, and generating threshold update reports.

[0157] Example 2:

[0158] Please see Figure 4 Another embodiment of the present invention provides a quality prediction system based on the visual features of preforms, comprising: an image acquisition module, a feature fusion module, a defect evolution module, a correlation analysis module, and a causal optimization module;

[0159] The image acquisition module is used to synchronously acquire multi-view molding images of the preform to be predicted through a ring acquisition array. After synchronous calibration and preprocessing to eliminate distortion and illumination interference, a time-series image set is generated by tracking the time-series nodes throughout the entire life cycle. Combined with historical preform data to supplement the correlation, visual features are extracted and quality is screened. Based on the identification code, an image history with multi-level quality labels is constructed to achieve traceability of single preform data throughout the entire life cycle, avoid feature loss and invalid data, and accurately characterize the quality status at each stage.

[0160] The feature fusion module is used to extract dynamic spatiotemporal features through sequence recognition network, delineate key areas by combining the mechanical properties of bottle preform structure, extract multi-level local features through local window self-attention and channel attention, generate quality gene features through adaptive weighted fusion and mutual information entropy screening, normalize and dimension-match the two types of features and then attention-weighted splice them to generate fused feature vectors, capture early signals of delayed explicit defects that are not visible to the naked eye in advance, achieve complementary advantages of features and reduce the amount of computation;

[0161] The defect evolution module is used to extract process-aided features and decay features. It generates a quality state vector through an encoding branch architecture, and uses a time-series prediction model and a progressive prediction mechanism. Combined with an evolutionary adaptive layer and time-series stage encoding, it generates a defect evolution trajectory with semantic annotation. The failure probability is corrected by historical data and environmental features, risk warning points are marked, and the module is optimized through verification with similar defects. This breaks through the limitation of only being able to determine the current state and provides accurate early warning for quality control.

[0162] The correlation analysis module is used to divide granular nodes, integrate spatial, structural, evolutionary, and feature correlations to construct a cross-scale correlation graph, capture defect correlations through a three-order inference architecture and evolution-oriented message passing, aggregate and refine graph-level features and then map performance degradation through a hybrid model, temporal calibration and experimental verification to ensure prediction consistency, output quantitative indicators of performance degradation and regional impact contribution, construct a defect-path-performance correlation library, and establish a quantitative correlation between local visual anomalies and overall performance degradation;

[0163] The causal optimization module is used to construct a directional causal relationship graph. It visualizes the image regions corresponding to causal driving features through causal CAM, locates abnormal visual features and associated process parameters by combining a four-level mapping library, generates process optimization suggestions bound to visual verification standards, upgrades the four-dimensional causal image history, provides feedback on the feature extraction, prediction modeling and image acquisition parameters of the optimization preceding modules, dynamically updates the quality control threshold, and realizes accurate defect tracing, intelligent process optimization and closed-loop control of the entire process.

[0164] Working principle and effects:

[0165] By acquiring multi-view time-series images of preforms throughout their entire lifecycle through a circular acquisition array and time-series tracking, and combining historical data to construct a complete image history and multi-level quality labels, the system addresses data distortion and feature loss issues, achieving full lifecycle traceability. By fusing dynamic spatiotemporal features and quality gene features, it proactively captures early signals of delayed-manifesting defects, avoiding potential missed detections. Based on a coding branch architecture and time-series prediction model, it generates defect evolution trajectories with risk annotations. Through cross-scale correlation modeling, it establishes a quantitative correlation between local visual anomalies and overall performance degradation, overcoming the limitation of only determining the current state. By leveraging causal correlation mining and process-visual matching, it explains the defect evolution mechanism and generates targeted optimization suggestions, forming a closed-loop optimization process. This enables accurate prediction of preform quality throughout its entire lifecycle, defect tracing, and intelligent process optimization, ensuring accurate assessment and targeted control, and meeting production requirements for product reliability.

[0166] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for quality prediction based on visual features of preforms, characterized in that, include: The process involves acquiring multi-view molding images of the preform to be predicted, using time-series nodes for real-time time-series tracking, combining historical preforms to construct a time-series image set, and obtaining visual feature vectors to build an image history. Based on the image history, dynamic spatiotemporal features and quality gene features are obtained, and a fused feature vector is generated through normalization, dimension matching, and splicing. The quality state vector of the preform to be predicted is constructed, the defect evolution is predicted using a time-series prediction model, the predicted quality state vector is generated, the trajectory unit is generated through a decoding hierarchical architecture, the initial failure probability is calculated and corrected, thereby generating the defect evolution trajectory and marking the risk warning points. By dividing granular nodes and constructing weighted edges, a cross-scale association graph is generated. An association library is constructed in conjunction with a mapping model, and a directional causal association graph is constructed. Causal association mining is performed to generate process optimization suggestions and update and optimize the image history.

2. The method of quality prediction based on visual features of preforms according to claim 1, characterized in that, The steps of real-time timing tracing include: Acquire multi-view molding images of a single preform and generate an initial image set of the preform to be predicted; Assign a unique identification code to the preform to be predicted, establish an identification association library, and associate it with auxiliary information; Based on the entire life cycle of the preform to be predicted, multiple time-series nodes are set, including the storage stage, simulation stage, and testing stage. Based on the multi-view molding images of each time node, the periodic time image set of the preform to be predicted is obtained; By using scale-invariant feature transformation, the initial image set of the preform to be predicted is registered with the periodic time-series image set to generate a real-time time-series image sequence of the preform to be predicted.

3. The method of quality prediction based on visual features of preforms according to claim 2, characterized in that, The steps of real-time timing tracing include: Collect multi-view molding images of historical preforms throughout their entire life cycle, generate a supplementary time-series image set, and register them with the initial images of historical preforms to generate a historical time-series image sequence. A time-series image sequence is constructed, and preprocessed and multi-source features are extracted, including texture features, edge features, and gray-level distribution features, to generate the primary feature vector of the preform to be predicted; The preprocessed time-series image sequence and primary feature vector are subjected to quality screening to generate an effective image sequence and visual feature vector of the preform to be predicted. Using the identification code of the preform to be predicted as the core index, an image history of the preform to be predicted is constructed using a distributed storage architecture. Set multi-level quality labels for the image history, including initial state, process state, and quality state.

4. The method of claim 3, wherein the quality prediction based on visual features of the preform is characterized by, The steps for generating the fused feature vector include: Based on the effective image sequence, linear interpolation and normalization processing are performed; Construct a sequence recognition network and configure dilated convolutional layers to generate intermediate feature maps for multiple temporal nodes; Using multi-level quality labels as the basis for attention guidance, initial attention weights are obtained to generate multi-temporal node feature maps; Global average pooling is performed on the multi-temporal node feature maps, and the contribution is used to fuse and normalize them to generate dynamic spatiotemporal features. Based on the effective image sequence, key regions are divided, and multi-level local feature maps are obtained through local window self-attention and channel attention. Quality gene features are generated through adaptive weighted fusion and mutual information entropy screening; The dynamic spatiotemporal features and the quality gene features are respectively subjected to layer normalization processing and dimensionality consistency verification. By using attention-weighted concatenation and contribution weights, a fused feature vector is generated.

5. The quality prediction method based on the visual features of the preform according to claim 4, characterized in that, The steps for defect evolution prediction include: Based on the image history, process-aided features and decay features are obtained; Construct a coding branch architecture, including a main branch, a first auxiliary branch, and a second auxiliary branch, which take the fused feature vector, process-aided features, and decay features as branch inputs, and output visual feature vector, process-related vector, and decay trend vector, respectively. A preliminary fusion feature vector is generated through a dynamic gating interaction mechanism; The preliminary fused feature vector is reconstructed and optimized to generate a quality state vector; A time series prediction model is constructed and multiple prediction branches are set. Based on the number of time series nodes, prediction targets are configured for each prediction branch. At the same time, an evolutionary adaptive layer is configured to assign time steps and attention weights to the prediction branches. Acquire stage characteristic factors and basic time step codes, and fuse them to generate time-series stage codes; A progressive prediction mechanism is adopted, in which the output of the previous prediction branch is passed to the next prediction branch through residual connection, and cross-branch residual correction is set at the same time. A composite loss function is set to iteratively train the time series prediction model and output the predicted quality state vector.

6. The quality prediction method based on the visual features of the preform according to claim 5, characterized in that, The steps for defect evolution prediction include: Based on a three-level structure of features, images, and semantics, a decoding hierarchical architecture is constructed to generate trajectory units; The environmental visual features in the image history are obtained and combined with the predicted quality state vector. The initial failure probability of each time node is calculated through multi-level logistic regression. Based on historical preform samples and environmental visual features, the initial failure probability is corrected to obtain the failure probability of each time-series node. Based on the trajectory units and failure probabilities of each time node, the defect evolution trajectory of the preform to be predicted is generated. Set a failure probability threshold, traverse the defect evolution trajectory, and filter risk warning points and risk persistence stages; A similar defect verification mechanism is set up to correct the defect evolution trajectory based on similarity.

7. The quality prediction method based on the visual features of the preform according to claim 6, characterized in that, The steps to build an associated library include: The risk area and critical area of ​​the preform to be predicted are obtained, and secondary nodes are divided, including coarse-grained nodes and fine-grained nodes, and a node association table is generated. Construct static basic features and dynamic evolution features for each granularity node, and bind them to the granularity node; Configure fusion weights for each granularity node, and generate granularity node features based on the fusion weights and layer normalization processing; Construct the association attributes of edges, generate weighted edges using weighted summation, and simultaneously construct cross-scale edges and jump edges; A cross-scale correlation graph is generated based on the mutual information entropy between the failure probability and the fused feature vector. Construct a three-order inference architecture and configure an evolution-oriented message passing mechanism; Multi-strategy aggregation and risk weighting are applied to global layer node features to generate graph-level relational feature vectors. Construct a mapping model to generate quantitative indicators of performance degradation and the performance impact contribution of each node; Calibration is performed using a time-series decay calibration factor and time-series consistency loss. Calculate the matching degree between the predicted values ​​of the preform performance and the actual values ​​of historical samples, and construct an association library.

8. The quality prediction method based on the visual features of the preform according to claim 7, characterized in that, The steps involved in causal association mining include: Based on the cross-scale correlation graph, the nodes are updated with image features and causal candidate factors; An initial causal relationship graph is constructed using multi-view shaped images and the edge weights of weighted edges; The structure of the initial causal relationship graph is learned by using an image feature-adaptive PC algorithm. A graph causal encoder is constructed, which captures the differences in causal transmission of image features at different evolutionary stages through a multi-head self-attention mechanism, and generates a directional causal association graph. Based on causal CAM, image mapping is performed on the directional causal relationship graph to generate a causal significance heatmap; Adaptive threshold segmentation is used to extract causal salient regions, and causal driving features are screened by combining risk region images to generate a three-dimensional correlation report.

9. The quality prediction method based on the visual features of the preform according to claim 8, characterized in that, The steps in causal association mining also include: A visual mapping library of process parameters and visual features is constructed. Abnormal visual features, corresponding causal links, and associated process parameter ranges of the visual mapping library are matched to the preform to be predicted. The contribution ratio of abnormal visual features to defect evolution is calculated. Calculate the visual process calibration factor, generate process optimization suggestions, calculate the confidence level of the process optimization suggestions, and retain only process optimization suggestions whose confidence level is greater than a preset confidence evaluation threshold; Based on the directional causal relationship graph, the image history of the preform to be predicted is upgraded causally. Based on the distributed storage architecture, the newly added layer is structurally associated with the original data to form a four-dimensional image history. Feedback optimization is performed based on the four-dimensional image history, and a dynamic mapping model is constructed to generate the final quality status label of the preform to be predicted.

10. A quality prediction system based on the visual features of preforms, used to implement the quality prediction method based on the visual features of preforms as described in any one of claims 1-9, characterized in that, include: Image acquisition module, feature fusion module, defect evolution module, correlation analysis module, causal optimization module; The image acquisition module is used to acquire multi-view molding images of the preform to be predicted, generate a time-series image set by tracking time-series nodes and combining historical preforms, extract visual feature vectors, and construct an image history based on the identification code. The feature fusion module is used to extract dynamic spatiotemporal features and quality gene features. After normalizing and matching the dynamic spatiotemporal features and quality gene features, attention-weighted concatenation is performed to generate a fused feature vector. The defect evolution module is used to extract process-aided features and degradation features, generate a quality state vector through an encoding branch architecture, generate a defect evolution trajectory using a time-series prediction model and a progressive prediction mechanism, correct the failure probability with historical data and environmental features, and mark risk warning points. The association analysis module is used to divide granularity nodes, construct cross-scale association graphs, and build an association library through a three-order inference architecture and evolution-oriented message passing; The causal optimization module is used to construct a directional causal relationship graph, perform causal relationship mining, generate process optimization suggestions, and update and optimize the image history.

Images