Big data-based hardware fitting production and processing quality intelligent management system

Abstract

The present application relates to the technical field of intelligent manufacturing control, in particular to a hardware fitting production and processing quality intelligent management system based on big data, comprising: data fusion, quality prediction, regulation and decision and execution control modules. Multi-source heterogeneous production data flow of hardware production line is collected, improved gradient boosting decision tree algorithm is used to carry out feature fusion and defect mode mining, and quality feature hologram covering multiple indexes is generated. Through the attention mechanism dynamic prediction network, the finished product quality grade and defect probability distribution are accurately output, and the personalized regulation instruction is generated by matching the quality standard. The defect data is mapped to the production line space coordinates, and the online control of the defective workpiece is completed. The scheme can realize dynamic prediction of production quality, accurate adjustment of process and pre-disposal of defects, and comprehensively optimize the quality control level of the whole process of hardware fitting processing.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing control technology, and in particular to an intelligent management system for the production and processing quality of hardware accessories based on big data. Background Technology

[0002] In the large-scale production and processing of hardware accessories, the industry generally uses conventional data statistical algorithms for production quality control. Data collection often relies on a single dimension, with multi-source, heterogeneous data such as equipment operating conditions, process parameters, and raw material properties stored independently and analyzed in a decentralized manner. Production quality inspections typically follow the traditional model of manual sampling and final inspection of finished products. Data mining methods are simplistic and crude, only able to identify superficial quality anomalies and failing to uncover the underlying logic and evolutionary patterns of defects within the production process.

[0003] Conventional quality analysis algorithms have weak data fusion adaptability, failing to deeply couple and analyze diverse production data. They cannot simultaneously cover multiple key information such as material batch quality, process operation status, and defect distribution location, making it difficult to form integrated quality data visualization content. Quality prediction models are mostly static architectures with fixed parameters, unable to adapt to the dynamic fluctuations of hardware processing conditions. Defect risk assessment has a single dimension, and quality prediction is not sufficiently relevant.

[0004] Hardware processing sites often employ standardized process adjustment schemes, resulting in homogenized processing parameter adjustment modes that fail to match the processing losses and operating conditions of different workpieces. The screening and handling of defective workpieces is concentrated at the finished product output stage, lacking real-time intervention methods during production. High-risk defective workpieces continue to undergo all processing steps, indicating insufficient proactive quality control in the production process and an inability to promptly stop processing anomalies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a big data-based intelligent management system for the production and processing quality of hardware accessories.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a big data-based intelligent management system for the production and processing quality of hardware accessories, comprising: The data fusion module collects multi-source heterogeneous data streams generated during the operation of the hardware parts production line, calls the improved gradient boosting decision tree algorithm to perform feature fusion and defect pattern mining on the multi-source heterogeneous data streams, and generates a quality feature hologram. The quality feature hologram includes the material batch quality spectrum, process capability index and potential defect spatial distribution. The quality prediction module inputs the quality feature hologram into a dynamic prediction network based on an attention mechanism to generate a quality grade prediction result and defect probability distribution for the current batch of finished hardware accessories. The control decision module generates a set of personalized quality control instructions for a single workpiece based on the comparison between the quality level prediction results and the preset quality standards. The set of personalized quality control instructions includes tool compensation adjustment instructions, process parameter correction instructions, and sorting path planning instructions. The execution control module maps the probability distribution of the defects to the spatial coordinates of the physical production line, and drives the actuator to identify, intercept and reprocess high-probability defective workpieces online.

[0007] As a further aspect of the present invention, the step of invoking the improved gradient boosting decision tree algorithm to perform feature fusion and defect pattern mining on the multi-source heterogeneous data stream to generate a quality feature hologram includes: The multi-source heterogeneous data stream includes raw material spectral analysis data, machine tool processing status time-series data, online visual inspection image sequences, and real-time workpiece size measurement data. Feature extraction is performed on the spectral analysis data of the raw materials to obtain the feature vectors of raw material components and impurities; The machining state time-series data of the machine tool are analyzed to extract machining dynamic feature vectors that characterize spindle vibration, feed speed stability and cutting force fluctuation; The online visual inspection image sequence is processed to extract surface visual feature vectors that characterize the surface texture, scratch depth, and gloss uniformity of the workpiece. The real-time measurement data of the workpiece dimensions are analyzed to calculate the dimensional deviation feature vector and the geometric tolerance feature vector. The raw material composition feature vector, the impurity feature vector, the processing dynamic feature vector, the surface visual feature vector, the dimensional deviation feature vector, and the geometric tolerance feature vector are merged to form a high-dimensional original feature set; The high-dimensional original feature set is input into the improved gradient boosting decision tree algorithm, which outputs a low-dimensional fusion feature that can distinguish quality states. The low-dimensional fusion feature is the quality feature hologram.

[0008] As a further aspect of the present invention, the working principle of the improved gradient boosting decision tree algorithm is as follows: The improved gradient boosting decision tree algorithm constructs a set of decision trees based on the principle of multi-scale feature interaction, which is achieved by introducing a feature cross-attention mechanism. When constructing each decision tree, the feature cross-attention mechanism dynamically evaluates the interaction importance between different feature dimensions and adjusts the feature split point selection strategy according to the interaction importance. Based on the distribution density of historical quality defect samples, an adaptive adjustment strategy is assigned to the sample weights in the decision tree set. The adaptive adjustment strategy makes the algorithm pay more attention to rare defect patterns during training. In each iteration of gradient boosting, not only is a new decision tree fitted based on the prediction residual, but the complexity of the tree is also regularized based on the sparsity constraint of the quality feature hologram. All decision trees generated iteratively are integrated. During the integration process, the recognition accuracy of each decision tree on the validation set for a specific defect category is dynamically weighted, and the low-dimensional fusion feature is finally output.

[0009] As a further aspect of the present invention, the improved gradient boosting decision tree algorithm constructs a set of decision trees based on the multi-scale feature interaction principle, which is implemented by introducing a feature cross-attention mechanism, including: In the initial stage of constructing the decision tree set, an initial importance score is calculated for each feature dimension in the high-dimensional original feature set; Initialize an attention parameter matrix, which is used to quantify the potential interaction strength between any two different feature dimensions; When constructing each decision tree in the decision tree set, for the current node to be split, the cross-attention score between each pair of feature dimensions is dynamically calculated based on the attention parameter matrix and the feature values ​​of the samples contained in the current node. Based on the cross-attention score, the initial importance score is weighted and corrected to generate a dynamic feature importance score; Based on the dynamic feature importance score, a set of candidate feature dimensions for splitting the current node is selected from high to low. Based on the Gini impurity or information gain criterion, the best splitting feature and its splitting threshold are selected from the candidate feature dimension set to complete the splitting of the current node; After the entire decision tree is constructed, the attention parameter matrix is ​​updated using gradient backpropagation based on the prediction error of the decision tree on the training set to optimize the quantification of the feature interaction strength. The execution steps are iterated until the construction of all decision trees in the decision tree set is completed.

[0010] As a further aspect of the present invention, the quality feature hologram is input into a dynamic prediction network based on an attention mechanism to generate a quality grade prediction result and defect occurrence probability distribution for the current batch of finished hardware accessories, including: The attention-based dynamic prediction network includes an encoder and decoder structure; The encoder performs feature transformation on the input quality feature hologram to generate a set of hidden layer feature representations containing contextual information; The decoder integrates a self-attention computation layer and a cross-attention computation layer. The self-attention computation layer is used to capture the long-range dependencies between different feature dimensions within the hidden layer feature representation, and generate an enhanced hidden layer feature representation. The cross-attention computation layer is used to establish the association mapping between the enhanced hidden layer feature representation and the predefined quality level template vector; Based on the strength of the association mapping, the quality level prediction result is output. The quality level prediction result is a multi-dimensional vector, and each dimension of the vector represents the probability of being assigned to a specific quality level. Simultaneously, the decoder outputs a defect attention weight matrix, and performs a dot product operation between the defect attention weight matrix and the quality feature hologram to obtain the defect occurrence probability distribution reflecting the contribution of different features to various types of defects.

[0011] As a further aspect of the present invention, the encoder performs feature transformation on the input quality feature hologram to generate a set of hidden layer feature representations containing contextual information, including: The quality feature hologram is input into the first linear transform layer of the encoder to map the dimension of the input feature to the preset hidden layer dimension, thereby obtaining the initial hidden layer feature representation; The initial hidden layer feature representation is input into the multi-head self-attention layer of the encoder, and the multi-head self-attention layer projects the initial hidden layer feature representation into multiple different feature subspaces to generate multiple projected feature representations. Within each feature subspace, the attention weights between each feature position and all other feature positions in the projected feature representation are calculated to obtain the attention-weighted feature representation corresponding to each feature subspace; The attention-weighted feature representations corresponding to multiple feature subspaces are concatenated and fused through a second linear transformation layer to generate a fused attention feature representation. The fused attention feature representation is residually connected to the initial hidden layer feature representation, and the connection result is subjected to layer normalization to generate the first intermediate hidden layer feature representation. The first intermediate hidden layer feature representation is input into the feedforward neural network layer of the encoder, and the feedforward neural network layer performs nonlinear transformation and feature enhancement on the first intermediate hidden layer feature representation. The output of the feedforward neural network layer is again residually connected to the first intermediate hidden layer feature representation, and the connection result is again processed by layer normalization to generate a hidden layer feature representation containing context information.

[0012] As a further aspect of the present invention, based on the comparison between the quality grade prediction result and the preset quality standard, a set of personalized quality control instructions for a single workpiece is generated, including: Extract the probability value of the current workpiece being assigned to a non-optimal quality level from the quality level prediction results; When the probability value exceeds a preset control threshold, the key feature dimension that causes the probability value to increase is located from the quality feature hologram; Based on the mapping relationship library between the key feature dimensions and the control parameters of the processing equipment, the amount of control parameter adjustment required to reduce the probability value is derived in reverse. The control parameter adjustment amount is converted into specific equipment executable instructions, forming the tool compensation amount adjustment instruction and the process parameter correction instruction; Simultaneously, based on the real-time logistics status of the production line and the current quality level prediction results of the workpiece, an optimal sorting or return path is planned, and the sorting path planning instruction is generated.

[0013] As a further aspect of the present invention, locating the key feature dimension that leads to the increase in the probability value from the quality feature hologram includes: Extract the attention weight vector related to the current workpiece from the defect attention weight matrix of the dynamic prediction network based on the attention mechanism; The attention weight vector is sorted, and several feature dimensions with weight values ​​exceeding a set threshold are selected. These feature dimensions are then marked as candidate key feature dimensions. By tracing back the generation process of the quality feature hologram, the original data source corresponding to the candidate key feature dimension is queried from the improved gradient boosting decision tree algorithm; Further analysis of the original data sources reveals whether the anomalies are in the raw material data, machine tool processing status data, visual data, or dimensional measurement data. The original data source ultimately identified as the root cause, along with its specific abnormal feature values, are collectively defined as the key feature dimension that leads to an increase in the predicted probability value of the quality level.

[0014] As a further aspect of the present invention, the defect occurrence probability distribution is mapped to the spatial coordinates of the physical production line, and the actuator is driven to perform online identification, interception, and reprocessing of high-probability defective workpieces, including: The probability distribution of the defect occurrence is associated and bound with the unique identifier and real-time location coordinates of each workpiece on the production line; Establish a spatial probability heatmap, with the physical location of the production line as the horizontal axis and the probability of defect occurrence as the vertical axis; In the spatial probability heatmap, different probability response threshold intervals are set, and each probability response threshold interval corresponds to an actuator response strategy. When the workpiece moves to the critical station equipped with the actuator, the system queries the probability value of the defect occurrence corresponding to the workpiece in the spatial probability heat map in real time. Based on the probability response threshold range in which the defect occurrence probability value falls, a corresponding response strategy is triggered. The response strategy includes inkjet marking, pneumatic interception push rod action, or robotic arm grabbing to the reprocessing station.

[0015] As a further aspect of the present invention, a corresponding response strategy is triggered based on the probability response threshold range in which the defect occurrence probability value falls, including: When the probability value of the defect occurrence is in the low probability range, the inkjet marking strategy is triggered, and the inkjet printer is controlled to spray an identification code containing the probability value and the defect type on the non-critical surface of the workpiece. When the probability value of the defect occurrence is in the medium probability range, the diversion and interception strategy is triggered, and the pneumatic push rod is controlled to push the workpiece from the main conveyor belt to the parallel verification and inspection branch line. When the probability value of the defect occurrence is in the high probability range, the immediate reprocessing strategy is triggered, and the robotic arm is controlled to directly grab the workpiece to the reprocessing station. At the same time, a reprocessing instruction containing the inferred defect type and suggested process parameters is sent to the control system of the reprocessing station. After all actions are completed, the workpiece identification, triggering strategy, execution time, and post-execution workstation information are recorded in the quality traceability log.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Leveraging an improved gradient boosting decision tree algorithm, this system processes multi-source heterogeneous data streams from the production line. Optimized algorithm logic facilitates feature fusion across various data types, deeply analyzing the relationships between different data points in the production process and uncovering patterns of hidden defects. By integrating material batch quality data, process capability indicators, and defect location distribution information, a multi-dimensional, holographic display of quality characteristics is constructed. This enriches the analytical dimensions of production quality data, improves the three-dimensional presentation of the entire production process's quality status, and refines the analytical levels for tracing the origins of production defects, allowing for a more intuitive presentation of hidden quality issues in the production process.

[0017] A dynamic prediction network based on an attention mechanism is constructed. Leveraging the network's dynamic weight adjustment characteristics, it adapts to real-time changes in hardware production conditions, capturing subtle quality fluctuations during processing and completing finished product quality stratification and defect distribution state prediction. A corresponding relationship is established between defect probability distribution data and the physical spatial coordinates of the production line. Combined with differences in quality thresholds, differentiated process adjustment and material sorting instructions are generated, refining the precision of single-workpiece processing control. The system collaborates with production line actuators to identify and handle abnormal workpieces on-site, shortening the flow path of defective workpieces within the production line, reducing the spread of processing anomalies, and optimizing the operational logic of quality control throughout the entire production process. Attached Figure Description

[0018] Figure 1 This is a sequence diagram of the intelligent management system for the production and processing quality of hardware accessories based on big data as described in this invention. Figure 2 A flowchart illustrating the process of generating a quality feature hologram using the improved gradient boosting decision tree algorithm; Figure 3 A flowchart illustrating the working principle of the improved gradient boosting decision tree algorithm. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1The overall implementation scheme of the intelligent management system for the production and processing quality of hardware accessories based on big data includes: a data fusion module, a quality prediction module, a control and decision-making module, and an execution control module. The data fusion module collects multi-source heterogeneous data streams generated during the operation of the hardware accessories production line, covering the entire process from raw material input to finished product output. This module uses an improved gradient boosting decision tree algorithm to process the multi-source heterogeneous data streams. The algorithm performs feature fusion and deep analysis on the data, mining deep correlations related to defect patterns and generating a comprehensive quality feature hologram. This hologram integrates multi-dimensional quality information such as material batch quality spectrum, process capability index, and potential defect spatial distribution. The quality prediction module receives the quality feature hologram as input and feeds it into a built-in attention-based dynamic prediction network. The network analyzes and deduces from the hologram, generating a quality grade prediction result for the current batch of finished hardware accessories and simultaneously outputting the defect occurrence probability distribution. The control and decision-making module receives the output from the quality prediction module, compares the predicted quality level with the preset quality standards in the system in real time, and generates a set of personalized quality control instructions for individual workpieces based on the comparison results. This set specifically includes tool compensation adjustment instructions, process parameter correction instructions, and sorting path planning instructions. The execution control module is responsible for mapping the defect occurrence probability distribution generated by the prediction module to the actual spatial coordinates of the physical production line. Through this mapping relationship, it drives various actuators on the production line to perform a series of control actions on the identified high-probability defective workpieces, such as online identification, physical interception, or guidance to reprocessing stations.

[0022] In one embodiment of the present invention, the multi-source heterogeneous data stream acquired by the system includes raw material spectral analysis data, machine tool processing status time-series data, online visual inspection image sequences, and real-time workpiece dimension measurement data. (See also...) Figure 2The raw material spectral analysis data is processed to extract raw material component feature vectors and impurity feature vectors reflecting material composition and purity. Time-domain and frequency-domain analysis is performed on machine tool processing time-series data to extract processing dynamic feature vectors characterizing equipment operational stability, including information such as spindle vibration, feed rate stability, and cutting force fluctuations. Online visual inspection image sequences are processed to extract surface visual feature vectors characterizing workpiece appearance quality, including measures such as surface texture, scratch depth, and gloss uniformity. Real-time workpiece dimension measurement data is analyzed to calculate dimensional deviation feature vectors and geometric tolerance feature vectors. The extracted raw material component feature vectors, impurity feature vectors, processing dynamic feature vectors, surface visual feature vectors, dimensional deviation feature vectors, and geometric tolerance feature vectors are merged to form a high-dimensional original feature set. This high-dimensional original feature set is input into an improved gradient boosting decision tree algorithm. The algorithm filters, combines, and compresses the high-dimensional features, outputting a low-dimensional fused feature with stronger discriminative ability for quality status. This low-dimensional fused feature is defined as the system's quality feature hologram.

[0023] In practical implementation, the data fusion module of the intelligent management system for the production and processing quality of hardware accessories based on big data collects multi-source heterogeneous data streams generated during the operation of the hardware accessories production line. These multi-source heterogeneous data streams include raw material spectral analysis data, machine tool processing status time-series data, online visual inspection image sequences, and real-time workpiece dimension measurement data. Feature extraction is performed on the raw material spectral analysis data to obtain raw material composition feature vectors and impurity feature vectors. The raw material spectral analysis data is acquired by a spectrometer and converted into a numerical sequence. Feature selection methods are used to extract the raw material composition feature vector representing the material's elemental composition and the impurity feature vector indicating the level of contaminants from the numerical sequence. The machine tool processing status time-series data is analyzed to extract processing dynamic feature vectors characterizing spindle vibration, feed speed stability, and cutting force fluctuations. The machine tool processing status time-series data comes from sensors, and signal processing techniques are used to calculate vibration amplitude, velocity variation coefficient, and force signal spectrum characteristics to construct the processing dynamic feature vectors. The online visual inspection image sequence is processed to extract surface visual feature vectors characterizing the workpiece surface texture, scratch depth, and gloss uniformity. The online visual inspection image sequence is captured by an industrial camera, and image analysis algorithms are applied to quantify texture roughness, scratch pixel depth, and gloss variance to generate surface visual feature vectors. Real-time workpiece dimension measurement data is analyzed to calculate dimensional deviation feature vectors and geometric tolerance feature vectors. The real-time workpiece dimension measurement data comes from measuring instruments; dimensional errors and geometric tolerance indices are calculated by comparing them with design standard values ​​to form dimensional deviation and geometric tolerance feature vectors. In some embodiments, the above feature extraction process employs standardization methods to ensure consistent vector dimensions. For example, raw material composition feature vectors and impurity feature vectors are dimensionality-reduced using principal component analysis; processing dynamic feature vectors are generated through time-domain statistics and frequency-domain transformation; surface visual feature vectors are derived based on gray-level co-occurrence matrix and edge detection algorithms; and dimensional deviation and geometric tolerance feature vectors are obtained through geometric calculations. Optionally, the acquisition frequency of multi-source heterogeneous data streams is synchronized with the production line cycle time to ensure that the feature vectors are aligned with the workpiece processing sequence.

[0024] In practical implementation, the feature vectors of raw material composition, impurities, processing dynamics, surface visual characteristics, dimensional deviations, and geometric tolerances are merged to form a high-dimensional original feature set. This high-dimensional original feature set is achieved through vector concatenation, where all feature vectors are sequentially connected into a single composite vector. The high-dimensional original feature set is represented as follows: in: Represents the high-dimensional original feature set. Represents the feature vector of raw material composition. Represents the impurity feature vector. Represents the dynamic feature vector of the process. Represents the surface visual feature vector. Represents the eigenvector of dimensional deviation. This represents the geometrical tolerance feature vector. It can be understood that the high-dimensional original feature set serves as input to the improved gradient boosting decision tree algorithm for subsequent feature fusion and defect pattern mining. In some embodiments, feature merging is performed before feature merging to eliminate the influence of dimensions, for example, by using min-max scaling to map each feature vector to a uniform numerical range. The high-dimensional original feature set is input into the improved gradient boosting decision tree algorithm, which performs feature selection and compression on the high-dimensional original feature set, outputting low-dimensional fused features that are discriminative of quality states; these low-dimensional fused features constitute the full quality feature map. Optionally, the training process of the improved gradient boosting decision tree algorithm uses historical production data, with quality labels as supervision signals to optimize model parameters.

[0025] In one embodiment of the present invention, the algorithm constructs a decision tree set based on the multi-scale feature interaction principle, which is implemented by introducing a feature cross-attention mechanism. See also... Figure 3 In the initial stage of constructing the decision tree ensemble, an initial importance score is calculated for each feature dimension in the high-dimensional original feature set. An attention parameter matrix is ​​initialized to quantify the potential interaction strength between any two different feature dimensions. When constructing each decision tree, for the current node to be split, the cross-attention score between each pair of feature dimensions is dynamically calculated based on the attention parameter matrix and the feature values ​​of the samples contained in the current node. Based on the cross-attention score, the initial importance scores of the features are weighted and corrected to generate dynamic feature importance scores. Based on the dynamic feature importance scores, a set of candidate feature dimensions for splitting the current node is selected from high to low. Based on the Gini impurity or information gain criterion, the optimal splitting feature and its splitting threshold are selected from the set of candidate feature dimensions to complete the node split. After the entire decision tree is constructed, the attention parameter matrix is ​​updated using gradient backpropagation based on the prediction error of the decision tree on the training set to optimize the quantification of feature interaction strength. The above steps are iteratively executed until all decision trees in the decision tree ensemble are constructed. Furthermore, the algorithm employs an adaptive adjustment strategy to assign sample weights to the decision tree set based on the distribution density of historical quality defect samples, enabling the algorithm to focus more on rare defect patterns during training. In each iteration of gradient boosting, a new decision tree is fitted not only based on the prediction residuals but also regularized according to the sparsity constraint of the quality feature hologram. All decision trees generated iteratively are then integrated, with dynamic weighting based on the recognition accuracy of each decision tree for specific defect categories on the validation set during the integration process, and low-dimensional fused features are output.

[0026] In its implementation, the improved gradient boosting decision tree algorithm constructs a decision tree set based on the principle of multi-scale feature interaction. This principle is achieved by introducing a feature cross-attention mechanism. In the initial stage of constructing the decision tree set, an initial importance score is calculated for each feature dimension in the high-dimensional original feature set. This initial importance score is obtained through statistical correlation analysis between features and quality labels. An attention parameter matrix is ​​initialized, which quantifies the potential interaction strength between any two different feature dimensions. The dimension of the attention parameter matrix corresponds to the number of features in the high-dimensional original feature set, and its elements are randomly initialized at the start of training. When constructing each decision tree in the decision tree set, for the node to be split, the cross-attention score between each pair of feature dimensions is dynamically calculated based on the attention parameter matrix and the feature values ​​of the samples contained in the current node. This dynamic calculation process involves linear transformation of the feature vectors and Softmax normalization of the attention weights. It can be understood that the cross-attention score reflects the strength of the mutual influence between different feature dimensions under a specific node sample distribution. Based on the cross-attention score, the initial importance score is weighted and corrected to generate a dynamic feature importance score. This weighting correction is achieved by weighting and summing the cross-attention score with the initial importance score. Based on the dynamic feature importance score, a set of candidate feature dimensions for splitting the current node is selected from high to low. This set contains the top K feature dimensions by dynamic feature importance score. Based on the Gini impurity or information gain criterion, the optimal splitting feature and its splitting threshold are selected from the candidate feature dimension set to complete the splitting of the current node. After constructing the entire decision tree, the attention parameter matrix is ​​updated using gradient backpropagation based on the prediction error of the decision tree on the training set to optimize the quantification of feature interaction strength. The aforementioned steps are iteratively executed until all decision trees in the decision tree set are constructed. In some embodiments, the calculation of the dynamic feature importance score can be expressed as: in: Indicates the importance score of dynamic features. The initial importance score represents the feature. This represents the elements in the attention parameter matrix corresponding to features i and j. Represents the feature value based on the current node sample. and The calculated interaction function value. Optional, interaction function. It can be a product of feature values ​​or a distance-based metric.

[0027] In its implementation, the improved gradient boosting decision tree algorithm assigns an adaptive adjustment strategy to the sample weights in the decision tree set based on the distribution density of historical quality defect samples. This adaptive adjustment strategy makes the algorithm focus more on rare defect patterns during training. The distribution density of historical quality defect samples is obtained by statistically analyzing the frequency of different defect categories in the training data. In each iteration of gradient boosting, a new decision tree is fitted not only based on the prediction residuals but also regularized to the tree's complexity according to the sparsity constraint of the entire quality feature graph. The sparsity constraint is achieved by adding an L1 regularization term to the loss function of the tree structure. All decision trees generated iteratively are integrated. During the integration process, each decision tree is dynamically weighted based on its recognition accuracy for a specific defect category on the validation set, outputting low-dimensional fusion features. This dynamic weighting is achieved by assigning a weight coefficient positively correlated with recognition accuracy to each decision tree. In some embodiments, the adaptive adjustment strategy for sample weights can be implemented based on cost-sensitive learning, assigning higher weights to samples of rare defect categories. Optionally, the hyperparameters of the sparsity constraint regularization term are determined through cross-validation. It is understandable that dynamic weighting based on recognition accuracy helps improve the ensemble model's ability to identify minority defects.

[0028] In one embodiment of the present invention, a quality feature hologram is input into a dynamic prediction network based on an attention mechanism, the network comprising an encoder and a decoder structure. The encoder performs feature transformation on the input quality feature hologram to generate a set of hidden layer feature representations containing contextual information. The encoder's processing involves inputting the quality feature hologram into its first linear transformation layer, mapping the dimension of the input features to a preset hidden layer dimension, obtaining an initial hidden layer feature representation. The initial hidden layer feature representation is then input into the encoder's multi-head self-attention layer, which projects the initial hidden layer feature representation into multiple different feature subspaces, generating multiple projected feature representations. Within each feature subspace, the attention weights between each feature position and all other feature positions in the projected feature representation are calculated, obtaining an attention-weighted feature representation corresponding to each feature subspace. The attention-weighted feature representations corresponding to multiple feature subspaces are concatenated and fused through a second linear transformation layer to generate a fused attention feature representation. This fused attention feature representation is then residually connected to the initial hidden layer feature representation, and the connection result is subjected to layer normalization to generate a first intermediate hidden layer feature representation. The first intermediate hidden layer feature representation is input into the encoder's feedforward neural network layer, which performs nonlinear transformations and feature enhancements. The output of the feedforward neural network layer is then residually connected to the first intermediate hidden layer feature representation, and the connection result is again processed by layer normalization to generate a hidden layer feature representation containing contextual information. The decoder integrates a self-attention computation layer and a cross-attention computation layer. The self-attention computation layer captures long-range dependencies between different feature dimensions within the hidden layer feature representation, generating an enhanced hidden layer feature representation. The cross-attention computation layer establishes an association mapping between the enhanced hidden layer feature representation and a predefined quality grade template vector. Based on the strength of the association mapping, a quality grade prediction result is output, which is a multi-dimensional vector where each dimension represents the probability of a workpiece belonging to a specific quality grade. Simultaneously, the decoder outputs a defect attention weight matrix, which is multiplied by the quality feature hologram to obtain a defect occurrence probability distribution reflecting the contribution of different features to various defects.

[0029] In practical implementation, the full quality feature map is input into a dynamic prediction network based on an attention mechanism to generate a prediction result of the quality grade and the probability distribution of defects in the current batch of finished hardware accessories. The dynamic prediction network based on the attention mechanism includes an encoder and a decoder structure. The encoder performs feature transformation on the input full quality feature map to generate a set of hidden layer feature representations containing contextual information. The encoder's processing includes inputting the full quality feature map into the encoder's first linear transformation layer to map the dimension of the input features to a preset hidden layer dimension to obtain the initial hidden layer feature representation. The initial hidden layer feature representation is input into the encoder's multi-head self-attention layer, which projects the initial hidden layer feature representation into multiple different feature subspaces to generate multiple projected feature representations. In each feature subspace, the attention weight between each feature position and all other feature positions in the projected feature representation is calculated to obtain the attention-weighted feature representation corresponding to each feature subspace. It can be understood that the process of calculating attention weights enables features to aggregate information based on the global context. The attention-weighted feature representations corresponding to multiple feature subspaces are concatenated and fused through a second linear transformation layer to generate a fused attention feature representation. The fused attention feature representation is residually connected to the initial hidden layer feature representation, and the connection result is layer-normalized to generate a first intermediate hidden layer feature representation. This first intermediate hidden layer feature representation is then input into the encoder's feedforward neural network layer. The feedforward neural network layer performs nonlinear transformations and feature enhancements on the first intermediate hidden layer feature representation; it typically contains two linear transformation layers and an activation function. The output of the feedforward neural network layer is again residually connected to the first intermediate hidden layer feature representation, and the connection result is again layer-normalized to finally generate a hidden layer feature representation containing contextual information. In some embodiments, the calculation of attention weights in the multi-head self-attention layer can be expressed as: in: The query matrix is ​​obtained by linear projection from the initial hidden layer feature representation. The key matrix is ​​obtained by linear projection from the initial hidden layer feature representation. The value matrix is ​​obtained by linear projection from the initial hidden layer feature representation. The dimension of the key vector. This represents the normalization exponential function, used to transform the calculated score into a probability distribution.

[0030] In specific implementation, the decoder integrates a self-attention computation layer and a cross-attention computation layer. The self-attention computation layer is used to capture the long-range dependencies between different feature dimensions within the hidden layer feature representation to generate an enhanced hidden layer feature representation. The cross-attention computation layer is used to establish an association mapping between the enhanced hidden layer feature representation and a predefined quality grade template vector. The quality grade template vector is a predefined feature vector representing different quality grades (such as superior, qualified, and substandard). It can be understood that the strength of the association mapping reflects the degree of matching between the input quality feature map and each quality grade template. Based on the strength of the association mapping, the quality grade prediction result is output. The quality grade prediction result is a multi-dimensional vector, where each dimension represents the probability of belonging to a specific quality grade. Simultaneously, the decoder outputs a defect attention weight matrix. The dimension of the defect attention weight matrix is ​​"number of defect categories × number of feature dimensions," and its element values ​​represent the contribution weight of each feature dimension to predicting a specific category of defect. Performing a dot product operation between the defect attention weight matrix and the quality feature map yields a defect occurrence probability distribution reflecting the contribution of different features to various types of defects. In some embodiments, the format of the quality grade prediction result is shown in Table 1. Table 1: Quality Grade Prediction Results Optionally, the predefined quality grade template vector can be obtained by calculating the centroids of the full quality feature maps of historical high-quality samples, qualified samples, and defective samples. The cross-attention calculation layer establishes an association mapping by calculating the similarity between the enhanced hidden layer feature representation and the quality grade template vector. Optionally, the calculation of the defect occurrence probability distribution involves performing row-by-row Softmax normalization on the dot product results to ensure that the sum of the probabilities of each type of defect is 1.

[0031] In one embodiment of the present invention, a set of personalized quality control instructions for a single workpiece is generated based on a comparison between the quality grade prediction result and a preset quality standard. The probability value of the current workpiece belonging to a non-optimal quality grade is extracted from the quality grade prediction result. When this probability value exceeds a preset control threshold, the key feature dimension causing the increase in this probability value is located from the quality feature hologram. Specifically, the key feature dimension is located by extracting the attention weight vector related to the current workpiece from the defect attention weight matrix of the dynamic prediction network based on an attention mechanism. This attention weight vector is sorted, and several feature dimensions with weight values ​​exceeding a set threshold are selected and marked as candidate key feature dimensions. The generation process of the quality feature hologram is traced back, and the original data source corresponding to the candidate key feature dimensions is queried from the improved gradient boosting decision tree algorithm. Further analysis is conducted on the original data source to determine whether the anomaly is in the raw material data, machine tool processing status data, visual data, or dimensional measurement data. The original data source ultimately identified as the root cause and its specific abnormal feature value are collectively defined as the key feature dimension causing the increase in the quality grade prediction probability value. After identifying the key feature dimensions, the required adjustment amount of the control parameters to reduce the probability value is derived by using the mapping relationship library between the dimension and the control parameters of the processing equipment. This adjustment amount is then converted into specific executable instructions for the equipment, forming tool compensation adjustment instructions and process parameter correction instructions. Simultaneously, based on the real-time logistics status of the production line and the predicted current quality level of the workpiece, an optimal sorting or return path is planned, generating a sorting path planning instruction.

[0032] In practical implementation, a set of personalized quality control instructions for individual workpieces is generated based on the comparison between the quality grade prediction results and preset quality standards. The control decision module extracts the probability value of the current workpiece belonging to a non-optimal quality grade from the quality grade prediction results. The quality grade prediction results are output by the quality prediction module, which includes the probability of the workpiece belonging to multiple grades such as "superior," "qualified," and "defective." When the extracted probability value of belonging to a non-optimal quality grade exceeds a preset control threshold, the system initiates the key feature dimension localization process. The preset control threshold is a configurable numerical parameter. The key feature dimension causing the increased probability value is located from the full quality feature map. The localization process is achieved by analyzing the defect attention weight matrix generated within the dynamic prediction network based on the attention mechanism. The attention weight vector related to the current workpiece is extracted from the defect attention weight matrix of the dynamic prediction network based on the attention mechanism. The attention weight vector is a one-dimensional array with the same length as the number of feature dimensions in the full quality feature map. Each element value represents the influence weight of the corresponding feature dimension on the current prediction result. The attention weight vectors are sorted, and several feature dimensions with weight values ​​exceeding the set threshold are selected. These feature dimensions are marked as candidate key feature dimensions. Tracing back the generation process of the full quality feature map, the original data sources corresponding to the candidate key feature dimensions are queried from the improved gradient boosting decision tree algorithm, i.e., determining which original data were fused to generate the candidate key feature dimensions. Further analysis is performed on the original data sources to determine whether the feature anomalies are in raw material data, machine tool processing status data, visual data, or dimensional measurement data. This analysis is completed by comparing the deviations of the candidate key feature dimension values ​​with the historical normal data distribution. This step aims to trace back from the fused high-level features to the original production data level. The original data source ultimately identified as the root cause and its specific anomalous feature values ​​are collectively defined as the key feature dimensions that lead to an increase in the predicted probability value of the quality grade. In some embodiments, the location results of the key feature dimensions and the generation of subsequent instructions are shown in Table 2. Table 2: Personalized Quality Control Instruction Generation Table In practical implementation, the required adjustment amount of control parameters to reduce the probability value is derived by reverse derivation based on the mapping relationship library between key feature dimensions and processing equipment control parameters. The mapping relationship library is a pre-generated lookup table or regression model that stores the correspondence between outliers of key feature dimensions and correction amounts of equipment control parameters. It can be understood that the mapping relationship library is constructed through historical data mining or a process knowledge base. The control parameter adjustment amount is converted into specific executable instructions for the equipment, forming tool compensation adjustment instructions and process parameter correction instructions. For example, "tool compensation +0.02mm" is converted into a G-code instruction recognizable by the CNC system, and "spindle speed reduced by 5%" is converted into a setting parameter modification instruction for the equipment controller. Simultaneously, combined with the real-time logistics status of the production line, an optimal sorting or return path is planned based on the predicted current quality level of the workpiece, generating a sorting path planning instruction. The real-time logistics status includes information such as the equipment status of each station, conveyor belt speed, and buffer occupancy. In some embodiments, the reverse derivation of the control parameter adjustment amount can be achieved through a linear mapping function: in: This represents the vector of control parameters that need to be adjusted (such as speed, feed, and compensation). This represents the weight matrix in the mapping relation library. This represents the vector of key feature dimensions that have been located. Optionally, optimal path planning employs a dynamic path search algorithm based on real-time status, aiming to minimize processing latency or maximize throughput. Optionally, all generated personalized quality control instruction sets are sent to the corresponding controllers on the production line for execution via industrial communication protocols.

[0033] In one embodiment of the present invention, the defect occurrence probability distribution is mapped to the spatial coordinates of the physical production line, driving the actuator to perform online identification, interception, and reprocessing of high-probability defective workpieces. The defect occurrence probability distribution is associated with the unique identifier and real-time position coordinates of each workpiece on the production line. A spatial probability heatmap is established, with the physical location of the production line as the horizontal axis and the defect occurrence probability as the vertical axis. Different probability response threshold intervals are set in the spatial probability heatmap, and each probability response threshold interval corresponds to an actuator response strategy. When a workpiece moves to a critical station equipped with an actuator, the system queries the corresponding defect occurrence probability value of the workpiece in the spatial probability heatmap in real time. Based on the probability response threshold interval where the defect occurrence probability value falls, the corresponding response strategy is triggered. Specific response strategies include: when the defect occurrence probability value is in the low probability interval, triggering a coding identification strategy, controlling the coding machine to spray an identification code containing the probability value and defect type on the non-critical surface of the workpiece; when the defect occurrence probability value is in the medium probability interval, triggering a diversion interception strategy, controlling a pneumatic pusher to push the workpiece from the main conveyor belt to a parallel verification and inspection branch line. When the probability of a defect occurring is in the high-probability range, an immediate reprocessing strategy is triggered. The robotic arm is then controlled to directly grab the workpiece and move it to the reprocessing station. Simultaneously, a reprocessing instruction containing the predicted defect type and suggested process parameters is sent to the control system of the reprocessing station. After all actions are completed, the workpiece identification, triggering strategy, execution time, and post-execution station information are recorded in the quality traceability log.

[0034] In practical implementation, the defect occurrence probability distribution is mapped to the spatial coordinates of the physical production line, driving the actuators to identify, intercept, and reprocess high-probability defective workpieces online. The execution control module associates the defect occurrence probability distribution with the unique identifier and real-time position coordinates of each workpiece on the production line. The unique identifier of the workpiece is obtained by scanning the QR code or RFID tag attached to the workpiece, and the real-time position coordinates are provided by position sensors or a vision positioning system deployed on the production line. A spatial probability heatmap is established, with the physical location of the production line as the horizontal axis and the defect occurrence probability as the vertical axis. The physical location of the production line is usually measured by the distance from the start of the production line. The defect occurrence probability value comes from the defect occurrence probability distribution output by the quality prediction module. In the spatial probability heatmap, different probability response threshold ranges are set, each corresponding to an actuator response strategy. For example, the low probability range is set to [0, 0.3), the medium probability range to [0.3, 0.7), and the high probability range to [0.7, 1.0]. When the workpiece moves to a critical station equipped with an actuator, the system queries the corresponding defect occurrence probability value in the spatial probability heatmap in real time. This query is achieved by matching the probability record stored in the spatial probability heatmap with the workpiece's unique identifier. Based on the probability response threshold range where the defect occurrence probability value falls, a corresponding response strategy is triggered. Response strategies include inkjet marking, pneumatic interception pusher action, or robotic arm grasping and moving the workpiece to a reprocessing station. In some embodiments, the process of mapping discrete workpiece positions and probability values ​​to a continuous heatmap can employ a kernel density estimation method, the formula of which can be expressed as: in: Represents the coordinates of the production line position. The estimated probability density value at that location. This indicates the total number of workpieces in the current batch. Indicates bandwidth parameter, Represents the kernel function. Indicates the first Real-time position coordinates of each workpiece Indicates the first The probability value of defects occurring in each workpiece.

[0035] In practical implementation, the response strategy triggered based on the probability response threshold range of the defect occurrence probability value includes the following specific operations: When the defect occurrence probability value is in the low probability range, a coding identification strategy is triggered, controlling the coding machine to spray an identification code containing the probability value and defect type on the non-critical surfaces of the workpiece. When the defect occurrence probability value is in the medium probability range, a diversion and interception strategy is triggered, controlling the pneumatic pusher to push the workpiece from the main conveyor belt to the parallel verification and inspection branch line. When the defect occurrence probability value is in the high probability range, an immediate reprocessing strategy is triggered, controlling the robotic arm to directly grab the workpiece to the reprocessing station, and simultaneously sending a reprocessing instruction containing the inferred defect type and suggested process parameters to the control system of the reprocessing station. It can be understood that the inferred defect type is obtained from the defect category with the highest probability in the defect occurrence probability distribution, and the suggested process parameters are obtained from the process parameter correction instructions generated by the control decision module. After all actions are completed, the workpiece identification, triggering strategy, execution time, and post-execution station information are recorded in the quality traceability log. The quality traceability log is stored in the form of a database table for production process traceability and analysis. In some embodiments, the timing of the pneumatic actuator and the robotic arm's movements is synchronized with the conveyor belt speed and precisely coordinated by a programmable logic controller. Optionally, the marking codes sprayed in the marking strategy adopt a machine-readable data matrix code format, facilitating rapid scanning and identification in subsequent stages. Optionally, sending reprocessing instructions to the reprocessing station is achieved through a manufacturing execution system interface or the OPCUA protocol.

[0036] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A big data-based intelligent management system for the production and processing quality of hardware accessories, characterized in that: The system includes: The data fusion module collects multi-source heterogeneous data streams generated during the operation of the hardware parts production line, calls the improved gradient boosting decision tree algorithm to perform feature fusion and defect pattern mining on the multi-source heterogeneous data streams, and generates a quality feature hologram. The quality feature hologram includes the material batch quality spectrum, process capability index and potential defect spatial distribution. The quality prediction module inputs the quality feature hologram into a dynamic prediction network based on an attention mechanism to generate a quality grade prediction result and defect probability distribution for the current batch of finished hardware accessories. The control decision module generates a set of personalized quality control instructions for a single workpiece based on the comparison between the quality level prediction results and the preset quality standards. The set of personalized quality control instructions includes tool compensation adjustment instructions, process parameter correction instructions, and sorting path planning instructions. The execution control module maps the probability distribution of the defects to the spatial coordinates of the physical production line, and drives the actuator to identify, intercept and reprocess high-probability defective workpieces online.

2. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 1, characterized in that, The improved gradient boosting decision tree algorithm is invoked to perform feature fusion and defect pattern mining on the multi-source heterogeneous data stream, generating a quality feature hologram, including: The multi-source heterogeneous data stream includes raw material spectral analysis data, machine tool processing status time-series data, online visual inspection image sequences, and real-time workpiece size measurement data. Feature extraction is performed on the spectral analysis data of the raw materials to obtain the feature vectors of raw material components and impurities; The machining state time-series data of the machine tool are analyzed to extract machining dynamic feature vectors that characterize spindle vibration, feed speed stability and cutting force fluctuation; The online visual inspection image sequence is processed to extract surface visual feature vectors that characterize the surface texture, scratch depth, and gloss uniformity of the workpiece. The real-time measurement data of the workpiece dimensions are analyzed to calculate the dimensional deviation feature vector and the geometric tolerance feature vector. The raw material composition feature vector, the impurity feature vector, the processing dynamic feature vector, the surface visual feature vector, the dimensional deviation feature vector, and the geometric tolerance feature vector are merged to form a high-dimensional original feature set; The high-dimensional original feature set is input into the improved gradient boosting decision tree algorithm, which outputs a low-dimensional fusion feature that can distinguish quality states. The low-dimensional fusion feature is the quality feature hologram.

3. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 2, characterized in that, The improved gradient boosting decision tree algorithm works as follows: The improved gradient boosting decision tree algorithm constructs a set of decision trees based on the principle of multi-scale feature interaction, which is achieved by introducing a feature cross-attention mechanism. When constructing each decision tree, the feature cross-attention mechanism dynamically evaluates the interaction importance between different feature dimensions and adjusts the feature split point selection strategy according to the interaction importance. Based on the distribution density of historical quality defect samples, an adaptive adjustment strategy is assigned to the sample weights in the decision tree set. The adaptive adjustment strategy makes the algorithm pay more attention to rare defect patterns during training. In each iteration of gradient boosting, not only is a new decision tree fitted based on the prediction residual, but the complexity of the tree is also regularized based on the sparsity constraint of the quality feature hologram. All decision trees generated iteratively are integrated. During the integration process, the recognition accuracy of each decision tree on the validation set for a specific defect category is dynamically weighted, and the low-dimensional fusion feature is finally output.

4. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 3, characterized in that, The improved gradient boosting decision tree algorithm constructs a set of decision trees based on the multi-scale feature interaction principle, which is implemented by introducing a feature cross-attention mechanism, including: In the initial stage of constructing the decision tree set, an initial importance score is calculated for each feature dimension in the high-dimensional original feature set; Initialize an attention parameter matrix, which is used to quantify the potential interaction strength between any two different feature dimensions; When constructing each decision tree in the decision tree set, for the current node to be split, the cross-attention score between each pair of feature dimensions is dynamically calculated based on the attention parameter matrix and the feature values ​​of the samples contained in the current node. Based on the cross-attention score, the initial importance score is weighted and corrected to generate a dynamic feature importance score; Based on the dynamic feature importance score, a set of candidate feature dimensions for splitting the current node is selected from high to low. Based on the Gini impurity or information gain criterion, the best splitting feature and its splitting threshold are selected from the candidate feature dimension set to complete the splitting of the current node; After the entire decision tree is constructed, the attention parameter matrix is ​​updated using gradient backpropagation based on the prediction error of the decision tree on the training set to optimize the quantification of the feature interaction strength. The execution steps are iterated until the construction of all decision trees in the decision tree set is completed.

5. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 1, characterized in that, The quality feature hologram is input into a dynamic prediction network based on an attention mechanism to generate a quality grade prediction result and defect probability distribution for the current batch of finished hardware accessories, including: The attention-based dynamic prediction network includes an encoder and decoder structure; The encoder performs feature transformation on the input quality feature hologram to generate a set of hidden layer feature representations containing contextual information; The decoder integrates a self-attention computation layer and a cross-attention computation layer. The self-attention computation layer is used to capture the long-range dependencies between different feature dimensions within the hidden layer feature representation, and generate an enhanced hidden layer feature representation. The cross-attention computation layer is used to establish the association mapping between the enhanced hidden layer feature representation and the predefined quality level template vector; Based on the strength of the association mapping, the quality level prediction result is output. The quality level prediction result is a multi-dimensional vector, and each dimension of the vector represents the probability of being assigned to a specific quality level. Simultaneously, the decoder outputs a defect attention weight matrix, and performs a dot product operation between the defect attention weight matrix and the quality feature hologram to obtain the defect occurrence probability distribution reflecting the contribution of different features to various types of defects.

6. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 5, characterized in that, The encoder performs feature transformation on the input quality feature hologram to generate a set of hidden layer feature representations containing contextual information, including: The quality feature hologram is input into the first linear transform layer of the encoder to map the dimension of the input feature to the preset hidden layer dimension, thereby obtaining the initial hidden layer feature representation; The initial hidden layer feature representation is input into the multi-head self-attention layer of the encoder, and the multi-head self-attention layer projects the initial hidden layer feature representation into multiple different feature subspaces to generate multiple projected feature representations. Within each feature subspace, the attention weights between each feature position and all other feature positions in the projected feature representation are calculated to obtain the attention-weighted feature representation corresponding to each feature subspace; The attention-weighted feature representations corresponding to multiple feature subspaces are concatenated and fused through a second linear transformation layer to generate a fused attention feature representation. The fused attention feature representation is residually connected to the initial hidden layer feature representation, and the connection result is subjected to layer normalization to generate the first intermediate hidden layer feature representation. The first intermediate hidden layer feature representation is input into the feedforward neural network layer of the encoder, and the feedforward neural network layer performs nonlinear transformation and feature enhancement on the first intermediate hidden layer feature representation. The output of the feedforward neural network layer is again residually connected to the first intermediate hidden layer feature representation, and the connection result is again processed by layer normalization to generate a hidden layer feature representation containing context information.

7. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 5, characterized in that, Based on the comparison between the predicted quality level and the preset quality standard, a set of personalized quality control instructions for individual workpieces is generated, including: Extract the probability value of the current workpiece being assigned to a non-optimal quality level from the quality level prediction results; When the probability value exceeds a preset control threshold, the key feature dimension that causes the probability value to increase is located from the quality feature hologram; Based on the mapping relationship library between the key feature dimensions and the control parameters of the processing equipment, the amount of control parameter adjustment required to reduce the probability value is derived in reverse. The control parameter adjustment amount is converted into specific equipment executable instructions, forming the tool compensation amount adjustment instruction and the process parameter correction instruction; Simultaneously, based on the real-time logistics status of the production line and the current quality level prediction results of the workpiece, an optimal sorting or return path is planned, and the sorting path planning instruction is generated.

8. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 7, characterized in that, Locating the key feature dimensions that cause the increased probability value from the quality feature hologram includes: Extract the attention weight vector related to the current workpiece from the defect attention weight matrix of the dynamic prediction network based on the attention mechanism; The attention weight vector is sorted, and several feature dimensions with weight values ​​exceeding a set threshold are selected. These feature dimensions are then marked as candidate key feature dimensions. By tracing back the generation process of the quality feature hologram, the original data source corresponding to the candidate key feature dimension is queried from the improved gradient boosting decision tree algorithm; Further analysis of the original data sources reveals whether the anomalies are in the raw material data, machine tool processing status data, visual data, or dimensional measurement data. The original data source ultimately identified as the root cause, along with its specific abnormal feature values, are collectively defined as the key feature dimension that leads to an increase in the predicted probability value of the quality level.

9. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 1, characterized in that, Mapping the probability distribution of the defects to the spatial coordinates of the physical production line, and driving the actuator to identify, intercept, and reprocess high-probability defective workpieces online, including: The probability distribution of the defect occurrence is associated and bound with the unique identifier and real-time location coordinates of each workpiece on the production line; Establish a spatial probability heatmap, with the physical location of the production line as the horizontal axis and the probability of defect occurrence as the vertical axis; In the spatial probability heatmap, different probability response threshold intervals are set, and each probability response threshold interval corresponds to an actuator response strategy. When the workpiece moves to the critical station equipped with the actuator, the system queries the probability value of the defect occurrence corresponding to the workpiece in the spatial probability heat map in real time. Based on the probability response threshold range in which the defect occurrence probability value falls, a corresponding response strategy is triggered. The response strategy includes inkjet marking, pneumatic interception push rod action, or robotic arm grabbing to the reprocessing station.

10. The intelligent management system for hardware accessories production and processing quality based on big data as described in claim 9, characterized in that, Based on the probability response threshold range in which the defect occurrence probability value falls, a corresponding response strategy is triggered, including: When the probability value of the defect occurrence is in the low probability range, the inkjet marking strategy is triggered, and the inkjet printer is controlled to spray an identification code containing the probability value and the defect type on the non-critical surface of the workpiece. When the probability value of the defect occurrence is in the medium probability range, the diversion and interception strategy is triggered, and the pneumatic push rod is controlled to push the workpiece from the main conveyor belt to the parallel verification and inspection branch line. When the probability value of the defect occurrence is in the high probability range, the immediate reprocessing strategy is triggered, and the robotic arm is controlled to directly grab the workpiece to the reprocessing station. At the same time, a reprocessing instruction containing the inferred defect type and suggested process parameters is sent to the control system of the reprocessing station. After all actions are completed, the workpiece identification, triggering strategy, execution time, and post-execution workstation information are recorded in the quality traceability log.

Images