Insole product quality traceability method and system based on internet of things

Abstract

This invention discloses a method and system for tracing the quality of insoles based on the Internet of Things (IoT), belonging to the field of IoT technology. The method includes: acquiring RFID tag data and spectral image data of insoles at each process node on the production line; performing timestamp alignment on the RFID tag data and spectral image data to construct a bimodal feature vector of the product state for each process node; inputting the bimodal feature vector of the product state for each process node into a graph attention network to learn the state transit dependencies between adjacent nodes, generating a traceability feature map containing the quality transmission path; and locating the initial process node and propagation path of the abnormal quality event based on the feature differences of each node in the traceability feature map. The system includes a memory and a processor to execute the method steps. Through bimodal data fusion and graph attention network modeling, accurate traceability of the source and propagation path of the anomaly is achieved.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, specifically to a method and system for tracing the quality of insoles based on IoT. Background Technology

[0002] The production line for shoe insoles involves multiple continuous process nodes such as cutting, pressing, printing, and molding. Deviations in any of these stages can propagate downstream along the transmission path, causing mass quality defects. Existing shoe insole quality traceability technologies typically deploy UHF RFID readers on the production line. These readers obtain unique identifiers and timestamps from the passive RFID tags embedded in the insoles, creating discrete logistical records. When a quality incident occurs, the process nodes traversed by the product are traced based on the tag reading history. This is then combined with manual sampling or offline testing equipment to analyze retained samples and infer potential problem areas. However, RFID tag data only reflects when and at which node the product passed; it does not carry information about the material state of the product after processing at that node. Relying solely on this identification and time record, traceability results can only locate a list of processes the product passed through, but cannot confirm at which node the anomaly actually occurred, nor can it reconstruct the transmission and evolution of quality defects between different processes.

[0003] Some production lines are equipped with industrial cameras or spectral devices to capture images of the insole surface. However, image acquisition and RFID reading / writing operate independently, resulting in inconsistent time bases. This makes it difficult to accurately match high-dimensional spectral material information with product identity information over time. Even with manual alignment afterward, a unique correspondence between product status and identity at each node cannot be guaranteed. Alignment deviations can easily lead to misassociations, making it impossible to accurately link material anomalies to specific products and processes. Furthermore, existing anomaly identification methods often use fixed thresholds or single-node statistical analysis, treating each process node as an isolated detection point and ignoring the continuous impact of processing residues, deformation, or chemical migration from previous processes on subsequent processes. When an anomaly is detected at a node, it is impossible to determine whether the anomaly originated from the current node or was transmitted from an upstream node. This leads to blind tracing of the root cause, unclear propagation paths, high investigation costs, and an inability to reconstruct the complete chain of anomalies accumulating or transforming with material flow.

[0004] Therefore, it is necessary to solve how to strictly correlate RFID tag data and spectral image data in time in a multi-node production line environment to form a bimodal representation that can simultaneously characterize product identity and material status, and how to construct a dynamic dependency model for the transmission of quality status between process nodes, so as to accurately lock the initial process node when an anomaly is detected and restore the propagation path of the anomaly along the production process. Summary of the Invention

[0005] This invention provides a method and system for tracing the quality of insole products based on the Internet of Things. It aims to locate the initial process node and reconstruct the propagation path of quality anomalies in the insole product production line by aligning the timestamps of RFID tag data and spectral image data with dual-modal feature fusion, and combining graph attention network to model the state propagation dependency between process nodes.

[0006] To achieve the above objectives, this invention provides the following technical solution: This invention provides a method for tracing the quality of insoles based on the Internet of Things (IoT), comprising: acquiring RFID tag data and spectral image data of insoles at each process node on the production line; performing timestamp alignment on the RFID tag data and spectral image data to construct a bimodal feature vector of product state for each process node; inputting the bimodal feature vector of product state of each process node into a graph attention network to learn the state transmission dependency between adjacent nodes and generate a traceability feature map containing the quality transmission path; and locating the initial process node and propagation path of the abnormal quality event based on the feature difference degree of each node in the traceability feature map. This method integrates RFID spatiotemporal information and hyperspectral material information into a unified feature representation after temporal alignment, and automatically learns the transmission law of quality state between process nodes using a graph attention network. It can capture weak anomalies from multimodal data and trace the source and propagation chain of anomalies through a graph structure, significantly improving the accuracy and efficiency of locating insole quality anomalies.

[0007] As a preferred technical solution of the present invention, the process of acquiring RFID tag data and spectral image data of insole products at various process nodes on the production line includes: deploying UHF RFID readers and hyperspectral imagers at each process node of the production line; reading the passive RFID tags built into the insole products using UHF RFID readers to obtain RFID tag data including the tag's unique identifier, reading timestamp, and signal strength value; and acquiring the spectral reflectance cube of the insole products in the visible and near-infrared bands using the hyperspectral imager to extract spectral image data of specific wavelengths on the surface of the insole products. By utilizing the simultaneous deployment of UHF RFID and hyperspectral imaging, it is ensured that the same product simultaneously obtains both identification and material spectral information at each process node, providing a precise data foundation for subsequent multimodal alignment and fusion.

[0008] Preferably, the process of aligning RFID tag data and spectral image data with timestamps to construct a bimodal feature vector for the product status at each process node includes: using the reading timestamp of the UHF RFID reader as a benchmark, selecting the spectral image data with the smallest absolute difference between the acquisition time and the reading timestamp of the hyperspectral imager as the matching data; expanding the matched spectral image data into a one-dimensional spectral sequence along the wavelength dimension; and concatenating the tag's unique identifier with the signal strength value and the one-dimensional spectral sequence to generate a bimodal feature vector for the product status. By using timestamp nearest neighbor matching, the time deviation between RFID reading and spectral acquisition is eliminated, ensuring a strict correspondence between spatiotemporal information and material information in the feature vector, enabling the features to accurately reflect the true product status at each process node.

[0009] Furthermore, after normalizing the one-dimensional spectral sequence, it is then vector-concatenated with the unique identifier of the tag after one-hot encoding and the signal intensity value to eliminate the impact of the difference in spectral data dimensions on subsequent network training and improve the robustness of feature representation.

[0010] As another preferred embodiment of the present invention, the process of inputting the bimodal feature vectors of the product state of each process node into a graph attention network to learn the state transit dependencies between adjacent nodes and generate a traceability feature map containing the quality transit path includes: defining each process node of the production line as a vertex in a graph structure, and defining the material flow relationship between adjacent process nodes as an edge in the graph structure; for each process node, using its bimodal feature vector of the product state as the initial feature of that vertex; in the graph attention network, performing weighted aggregation on the features of the neighboring vertices of each vertex, with the weight determined by the similarity between the vertex features, and outputting the updated feature vector of each vertex after multi-layer aggregation; and organizing the updated feature vectors of all vertices into a traceability feature map according to the edge connection relationship. This scheme uses the graph attention mechanism to adaptively focus on the features of upstream nodes that have a greater impact on the quality state of the current node, so that the generated traceability feature map not only retains the independent state of each node, but also explicitly encodes the information on the transmission and evolution of quality along the process route.

[0011] Preferably, in the graph attention network, the features of the neighboring vertices of each vertex are weighted and aggregated, with the weight determined by the similarity between the vertex features. The specific process of outputting the updated feature vector of each vertex after multi-layer aggregation includes: using the bimodal feature vector of the product state corresponding to each process node as the input feature of the corresponding vertex in the graph attention network, and performing a linear transformation on the input features of all vertices, recording the transformed feature vector as the intermediate representation vector of that vertex; for each edge in the graph structure, calculating the attention coefficient between the two vertices it connects, by concatenating the intermediate representation vectors of the two vertices, and inputting the concatenated vector into a single-layer feedforward layer. The neural network, after being mapped by a nonlinear activation function, obtains a scalar value, which serves as the original attention value for that edge. A flexible maximum normalization operation is applied to the original attention values ​​of all adjacent edges of each vertex, ensuring the sum of the normalized attention coefficients is 1. These normalized attention coefficients represent the similarity weights between vertex features. The intermediate representation vectors of all adjacent vertices of each vertex are weighted and summed according to their corresponding normalized attention coefficients to obtain the aggregated feature vector of that vertex. The aggregated feature vector is then residually concatenated with the original intermediate representation vector of the current vertex. After processing through a multi-layer graph attention network with layer-by-layer aggregation, the updated feature vector of each vertex is output. Through residual concatenation and multi-layer aggregation mechanisms, both the node's own features are preserved, and neighborhood-transferred features are fully integrated, effectively amplifying subtle quality shifts between multi-hop nodes without smoothing them out.

[0012] Furthermore, the single-layer feedforward neural network uses a linear unit with leakage correction as a nonlinear activation function to alleviate gradient vanishing and retain weak feature responses in the negative value range, thereby improving the sensitivity to boundary anomalies.

[0013] As a preferred technical solution of the present invention, the process of locating the initial process node and propagation path of an abnormal quality event based on the feature difference degree of each node in the traceability feature graph includes: calculating the cosine distance between the updated feature vector and the standard feature vector of each vertex in the traceability feature graph to obtain the feature difference degree of each vertex; marking vertices with feature difference degrees exceeding a preset difference threshold as abnormal nodes; selecting the process node with the earliest occurrence time from the abnormal nodes as the initial abnormal node; and traversing adjacent abnormal nodes along the edge direction in the traceability feature graph starting from the initial abnormal node, recording the abnormal node sequence as the quality anomaly propagation path. Using cosine distance to measure feature difference degree is insensitive to feature scale and can stably identify abnormal nodes that deviate from the normal distribution pattern. Then, the source of the anomaly and the propagation chain are jointly determined by the temporal sequence and graph structure constraints, avoiding misjudgment of isolated abnormal points.

[0014] Preferably, when traversing adjacent abnormal nodes, only the sequence of abnormal nodes whose feature difference does not decrease monotonically along the propagation direction is retained as the final quality abnormality propagation path, so as to ensure that the propagation path reflects the real physical process of abnormal accumulation or amplification and eliminates erroneous connections caused by accidental fluctuations.

[0015] As a further improvement of the present invention, before calculating the cosine distance between the updated feature vector and the standard feature vector of each vertex in the traceability feature map, the method further includes: acquiring RFID tag data and spectral image data of multiple consecutive batches of the same insole product model under normal production conditions; constructing bimodal feature vectors of normal product states for multiple batches according to the method described above; inputting the bimodal feature vectors of normal product states for multiple batches into a graph attention network to obtain updated feature vectors of multiple normal nodes; calculating the mean of the updated feature vectors of multiple normal nodes by dimension to obtain the standard feature vector. The standard feature vector obtained from historical normal samples is used as a comparison benchmark, making the anomaly judgment statistically based, adaptable to normal production fluctuations, and able to accurately detect minute shifts.

[0016] As another preferred embodiment of the present invention, after marking vertices with feature differences exceeding a preset difference threshold as anomalous nodes, the method further includes: obtaining the difference vector between the updated feature vector marked as anomalous node and the standard feature vector; sorting the absolute values ​​of each dimension in the difference vector from largest to smallest, and selecting the preset quantity dimension with the highest sorting value as the anomalous feature dimension; and back-mapping the anomalous feature dimension to the original spectral band to determine the key spectral feature wavelength range that caused the anomalous quality event. Through the difference vector sorting and band back-mapping, the specific spectral band causing the anomaly can be directly identified, providing process engineers with clues to the causes of the anomaly at the chemical composition or physical structure level.

[0017] Preferably, after selecting the earliest occurring process node from the abnormal nodes as the initial abnormal node, the method further includes: obtaining the signal strength value in the product state dual-modal feature vector of the initial abnormal node; estimating the spatial position offset of the insole product at the initial abnormal node based on the inverse relationship between the signal strength value and the RFID tag reading / writing distance; and marking the initial abnormal node whose spatial position offset exceeds a preset offset threshold as a conveyor belt jamming associated abnormal node. Utilizing the distance information implicit in the RFID signal strength, it is possible to simultaneously determine whether the abnormality is caused by conveyor belt jamming or positional deviation, thus helping to distinguish between process abnormalities and transmission abnormalities.

[0018] As a further optimization of this invention, after recording the abnormal node sequence as the quality anomaly propagation path, the method further includes: obtaining the time interval between adjacent abnormal nodes in the quality anomaly propagation path; obtaining the running speed of the conveyor belt between adjacent abnormal nodes and calculating the theoretical transmission time between adjacent abnormal nodes; marking the edges between adjacent abnormal nodes where the ratio of the time interval to the theoretical transmission time exceeds a preset range as propagation delay abnormal edges; and outputting the upstream and downstream process nodes connected to the propagation delay abnormal edges as quality traceability checkpoints. By comparing the actual time interval with the theoretical transmission time, delay anomalies in material flow can be detected, assisting in troubleshooting production cycle anomalies or buffer waiting problems.

[0019] Preferably, after outputting the upstream and downstream process nodes connected by the propagation delay anomaly edge as quality traceability check points, the method further includes: obtaining the anomaly feature dimensions of the starting node and the ending node of the propagation delay anomaly edge; calculating the intersection dimension of the anomaly feature dimensions of the starting node and the ending node; and outputting the spectral bands corresponding to the intersection dimension as cross-node common anomaly features to the quality analysis terminal for cause clustering display. Through cross-node anomaly feature intersection analysis, common spectral anomaly features spanning multiple process nodes are extracted, facilitating quality engineers to perform fault mode clustering and root cause classification.

[0020] This invention also includes an IoT-based insole product quality traceability system. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned IoT-based insole product quality traceability method. This system deploys the above-mentioned method as a computer program on the processor, achieving automated, end-to-end traceability analysis and anomaly alarms for insole product quality on the production line.

[0021] The technical effects and advantages provided by the present invention in the above technical solution are as follows: At each process node of the production line, UHF RFID tag data and spectral image data collected by a hyperspectral imager are simultaneously acquired from the insole products. Using the RFID reading timestamp as a benchmark, the spectral image data with the smallest time difference is selected for matching. The unique tag identifier is then thermally encoded and vectorized with the signal strength value and a wavelength-dimension-expanded and normalized one-dimensional spectral sequence to form a dual-modal feature vector of the product status. This vector simultaneously carries the product's physical identity, spatial location perception, and the spectral reflectance characteristics of the surface material in the visible to near-infrared bands. This ensures that the product status at each process node is not merely a marker of process arrival, but a comprehensive description of the status with high-dimensional material information. Compared to using only RFID tags or relying solely on image detection, this dual-modal feature can more sensitively reflect subtle process fluctuations such as changes in material density, differences in coating uniformity, and the degree of adhesive curing. Furthermore, the time synchronization between the tag and the spectrum avoids mismatches that could lead to incorrect identity-status binding.

[0022] Using each process node on the production line as a graph vertices and the material flow relationships between adjacent nodes as edges, the bimodal feature vector of the product state of each node is input into a graph attention network. The network performs attention-based weighted aggregation on the features of neighboring vertices of each vertex. The attention coefficient is learned from the similarity between vertex features, giving higher weights to upstream nodes that are similar to the current node's state features or have strong transitive relationships. Through multi-layer aggregation and residual connections, the transitive dependencies of quality states between adjacent nodes are learned, outputting a traceability feature map containing the quality transit path. This mechanism transforms each process node on the production line from an isolated detection point into an interconnected state transit network. Abnormal states can be naturally characterized in a forward propagation through the edge weights in the graph structure, not only preserving the behavior of anomalies at a single node but also capturing their trajectory of accumulation, transformation, or amplification between processes. After calculating the cosine distance between the updated feature vector of each vertex and the standard feature vector in the traceability feature map to obtain the difference, abnormal nodes with difference exceeding the threshold are marked. The process node with the earliest occurrence time is selected as the initial abnormal node, and adjacent abnormal nodes are traversed along the edge direction, recording the abnormal node sequence as the quality anomaly propagation path. Therefore, the root cause of abnormal events can be accurately pinpointed to the earliest process node that deviates from the normal state. At the same time, the complete range of impact of the abnormality as it spreads along the material flow to the subsequent process can be restored. This avoids the limitation of existing methods that lack upstream and downstream dependency models, which makes it impossible to distinguish between the source and the transmission effect. It transforms quality inspection from blind inspection of the entire line segment by segment to targeted path restoration. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0024] Figure 1 This is a flowchart of a method for tracing the quality of insole products based on the Internet of Things. Figure 2 This is a flowchart for generating dual-modal feature vectors of product status based on radio frequency identification and spectral images; Figure 3 This is a schematic diagram of the product state traceability feature map generation process based on graph attention network; Figure 4 This is a flowchart for obtaining the propagation path of quality anomalies based on feature difference. Figure 5 This is a flowchart for identifying propagation delay anomalies and extracting common anomaly features across nodes based on the ratio of time interval to theoretical transmission duration. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] See Figure 1 This invention provides a method for tracing the quality of insole products based on the Internet of Things (IoT), comprising: acquiring RFID tag data and spectral image data of insole products at each process node on the production line; performing timestamp alignment on the RFID tag data and spectral image data to construct a bimodal feature vector of product state for each process node; inputting the bimodal feature vector of product state for each process node into a graph attention network to learn the state transit dependency between adjacent nodes and generate a traceability feature map containing the quality transit path; and locating the initial process node and propagation path of the abnormal quality event based on the feature difference degree of each node in the traceability feature map.

[0027] In specific implementation, please refer to Figure 2Ultra-high frequency (UHF) RFID readers and hyperspectral imagers are deployed at various process nodes on the production line. The UHF RFID readers read the passive RFID tags embedded in the insoles to obtain RFID tag data, including the tag's unique identifier, reading timestamp, and signal strength value. The hyperspectral imager collects spectral reflectance cubes of the insoles in the visible and near-infrared bands, extracting spectral image data of specific wavelengths on the insole surface.

[0028] Timestamp alignment is performed on RFID tag data and spectral image data. Using the read timestamp of the UHF RFID reader as a reference, the absolute difference between each acquisition time of the hyperspectral imager and the read timestamp of the UHF RFID reader is calculated. From all the spectral image data acquired by the hyperspectral imager, the spectral image data that minimizes the absolute difference is selected as the matching data to be matched with the read timestamp of the UHF RFID reader.

[0029] The matched spectral image data is expanded along the wavelength dimension to generate a one-dimensional spectral sequence. Each element of the one-dimensional spectral sequence corresponds to the spectral reflectance value at a specific wavelength. The one-dimensional spectral sequence is then normalized using the minimum-maximum normalization formula:

[0030] in, Represents the first spectral element in a one-dimensional spectral sequence. Spectral reflectance values ​​corresponding to each wavelength This represents the minimum spectral reflectance value among all wavelengths in a one-dimensional spectral sequence. This represents the maximum value of the spectral reflectance values ​​corresponding to all wavelengths in a one-dimensional spectral sequence. Indicates the normalized i-th The reflectance value corresponding to each wavelength.

[0031] The unique identifier of the label is uniquely encoded. The unique identifier is a category variable with a predetermined number of categories. After one-hot encoding, a binary vector with a dimension equal to the number of categories is obtained. Only the dimension corresponding to the category to which the unique identifier belongs has a value of 1, while the other dimensions have values ​​of 0. The binary vector obtained after one-hot encoding, the signal strength value, and the normalized one-dimensional spectral sequence are sequentially concatenated to generate a bimodal feature vector of the product state. The dimension of the bimodal feature vector of the product state is determined by the unique encoding dimension, the single dimension of the signal strength value, and the length of the one-dimensional spectral sequence.

[0032] In specific implementation, please refer to Figure 3Each process node in the production line is defined as a vertex in a graph data structure, and the material flow relationships between adjacent process nodes are defined as edges in the graph data structure. Graph data structures are used... express, This represents a set of vertices, with each vertex corresponding to a process node. Represents a set of edges, where each edge corresponds to a material flow path. If the process node... The output materials flow directly into the process node. Then there exists a path from point to The directed edge.

[0033] For each process node, obtain the corresponding bimodal feature vector of the product state, and use this bimodal feature vector as the initial feature vector of the corresponding vertex in the graph data structure. The initial eigenvector is denoted as , The dimension is , The value is determined by adding the one-hot encoding dimension of the tag's unique identifier, the dimension 1 of the signal strength value, and the length of the one-dimensional spectral sequence.

[0034] The initial feature vectors of all vertices in the graph data structure are input into the graph attention network. The graph attention network contains multiple graph attention layers, each performing a weighted aggregation and update of the vertex features. The number of graph attention layers is set to 2. The input of the first graph attention layer is the initial feature vector of each vertex, and the input of the second graph attention layer is the output of the first graph attention layer. The internal processing method of each graph attention layer is the same.

[0035] In a graph attention layer of a graph attention network, a linear transformation is first performed on the input feature vectors of all vertices. A shared linear transformation matrix is ​​introduced. , The shape is ,in, Let represent the dimension of the intermediate representation vector after the linear transformation. Set to 128. For vertices Its input feature vector is (in the first layer) In subsequent layers (The output of the previous layer) is used to obtain the vertex after linear transformation. intermediate representation vector .

[0036] For each edge in the graph data structure Calculate the edges The attention coefficient between the two connected vertices. The attention coefficient is calculated using a single-layer feedforward neural network. The single-layer feedforward neural network uses a linear unit with leakage correction as the non-linear activation function. The vertices are then... intermediate representation vector With vertex intermediate representation vector Concatenate the vectors to obtain the concatenated vector. The dimension of the concatenated vector is The concatenated vector is input into a single-layer feedforward neural network, which consists of a weight vector. Parameterization The dimension is First calculate the weight vector. The inner product of the concatenated vectors is then applied to the inner product result using a linear unit function with leakage correction to obtain the edge. The original attention value. The linear unit function with leakage correction outputs the product of the input value and the leakage slope of 0.2 when the input is negative, and outputs the original value when the input is non-negative. The formula for calculating the original attention value is:

[0037] in, Representing an edge The original attention value; This represents a linear unit function with leakage correction; Represents the weight vector transpose; Represents vertices The intermediate representation vector; Represents vertices The intermediate representation vector; This indicates a vector concatenation operation.

[0038] For each vertex Collect all vertices The original attention values ​​corresponding to the edges starting from a given point are then normalized using a flexible maximum normalization operation to obtain normalized attention coefficients. The normalization operation is calculated by taking each original attention value... The exponent divided by the vertex The sum of the original attention values ​​of all neighboring edges, expressed by the formula: ,in Represents vertices The set of neighboring vertices. Normalized attention coefficients. satisfy The normalized attention coefficient is the similarity weight between vertex features.

[0039] Vertex The intermediate representation vectors of all neighboring vertices are weighted and summed according to their corresponding normalized attention coefficients to obtain the vertex. Aggregated feature vectors .

[0040] Aggregate feature vectors With vertex The original intermediate representation vector input to the current graph attention layer. Perform a residual join, which is done by adding elements one by one to obtain the vertices. Initial update vector The vector will be initially updated. The output of this graph attention layer is passed to the next graph attention layer. After being aggregated layer by layer through two graph attention layers, the updated feature vector of each vertex is output. The updated feature vector is obtained based on the output of the second graph attention layer, i.e., the vertex's... The updated feature vector is the output of the second graph attention layer. .

[0041] The updated feature vectors of all vertices are organized into a source feature graph according to the edge connections in the graph data structure. In the source feature graph, each vertex carries an updated feature vector, and each edge retains its original direction and connection relationship.

[0042] In specific implementation, please refer to Figure 4 Obtain the updated feature vector corresponding to each vertex in the traceability feature map. Obtain the standard feature vector, which is calculated by averaging the updated feature vectors of multiple batches of products of the same insole product model under normal production conditions.

[0043] The standard feature vector is obtained as follows: Radio frequency identification (RFID) tag data and spectral image data of multiple consecutive batches of the same insole product model under normal production conditions are acquired, with the number of batches set to 50. For each batch, using the reading timestamp of the UHF RFID reader as a benchmark, the spectral image data with the smallest absolute difference between the acquisition time and the reading timestamp of the hyperspectral imager is selected as the matching data. The matched spectral image data is expanded into a one-dimensional spectral sequence along the wavelength dimension, and the one-dimensional spectral sequence is subjected to min-max normalization. The unique tag identifier is then one-hot encoded and concatenated with the signal strength value and the normalized one-dimensional spectral sequence to generate a bimodal feature vector for the normal product state. After a batch of products passes through all process nodes, a bimodal feature vector for the normal product state is generated for each process node.

[0044] The bimodal feature vectors of 50 batches of normal product states are input into a graph attention network according to the process nodes. The structure and parameters of the graph attention network are fixed, and the graph attention network contains two graph attention layers, with a shared linear transformation matrix set in each graph attention layer. Linear transformation matrix The shape is , The dimension of the bimodal feature vector for the normal product state is defined; each graph attention layer also has a weight vector with a dimension of 256. Used to calculate the original attention values ​​and weight vectors on the edges. Corresponding to the single-layer feedforward neural network applied to concatenated vectors, the nonlinear activation function of the single-layer feedforward neural network is a linear unit with leakage correction, and the leakage slope is set to 0.2; the output of each layer uses residual connections. The input of the graph attention network is the bimodal feature vector of the normal product state of each process node in each batch as the initial feature vector of the corresponding vertex, and the output is the updated feature vector of each vertex. For the updated feature vectors output by vertices of the same process node in 50 batches, the arithmetic mean is calculated dimension by dimension, and the arithmetic mean of all dimensions is combined into a vector as the standard feature vector of that process node.

[0045] Calculate the cosine distance between the updated feature vector of each vertex in the source feature map and the standard feature vector of the corresponding process node, and use the cosine distance as the feature dissimilarity of that vertex. The formula for calculating the cosine distance is:

[0046] in, Represents the cosine distance, with a value range of 1. ; This represents the updated feature vector of a vertex; This represents the standard feature vector of the process node corresponding to this vertex; Representing vectors with vector The inner product; Representing vectors The L2 norm; Representing vectors The L2 norm.

[0047] The preset difference threshold is set to 0.15. The preset difference threshold is determined by calculating the average of all normal characteristic differences at each process node under normal production conditions for 50 batches. and standard deviation Set the preset difference threshold to The calculated value is 0.15. The feature difference degree... Vertices with a value greater than 0.15 are marked as abnormal nodes.

[0048] From the set of vertices marked as abnormal nodes, obtain the reading timestamp of the UHF RFID reader corresponding to each abnormal node, compare the reading timestamps of all abnormal nodes, and select the process node with the earliest reading timestamp as the initial abnormal node.

[0049] Starting from the initial abnormal node in the source feature graph, traverse adjacent abnormal nodes along the edges of the graph data structure. For the currently traversed abnormal node, check if its neighboring vertices are marked as abnormal nodes. If a neighboring vertex is an abnormal node, compare the feature difference between the neighboring abnormal node and the current abnormal node. If the feature difference between the neighboring abnormal node and the current abnormal node is not less, add the neighboring abnormal node to the abnormal node sequence and continue traversing with that neighboring abnormal node as the new current node; if the feature difference between the neighboring abnormal node and the current abnormal node is less, terminate the traversal along that neighborhood direction. Repeat the above process, and the final recorded abnormal node sequence is used as the quality abnormality propagation path. The node order in the quality abnormality propagation path follows the sequential direction of material flow on the production line, and the feature difference between adjacent nodes in the abnormal node sequence satisfies a monotonically non-decreasing relationship.

[0050] In specific implementation, after vertices whose feature differences exceed a preset difference threshold are marked as abnormal nodes, the updated feature vectors of the marked abnormal nodes are obtained. The standard feature vectors of the corresponding process nodes for these abnormal nodes are obtained. The standard feature vectors are obtained as follows: RFID tag data and spectral image data from 50 consecutive batches of the same insole product model under normal production conditions are obtained. For each batch, using the reading timestamp of the UHF RFID reader as a benchmark, the spectral image data with the smallest absolute difference between the hyperspectral imager acquisition time and the reading timestamp is selected as the matching data. The matched spectral image data is expanded into a one-dimensional spectral sequence according to the wavelength dimension, and the one-dimensional spectral sequence is subjected to minimum-maximum normalization. The unique tag identifier is one-hot encoded and then concatenated with the signal strength value and the normalized one-dimensional spectral sequence to generate a normal product state dual-modal feature vector. The normal product state dual-modal feature vectors from the 50 batches are input into a graph attention network according to the process nodes. The graph attention network contains two graph attention layers, each with a shared linear transformation matrix. The shape of the linear transformation matrix is... , To define the dimensions of the bimodal feature vector for normal product status, each graph attention layer also includes a weight vector with a dimension of 256. This weight vector corresponds to a single-layer feedforward neural network applied to the concatenated vector. The nonlinear activation function of the single-layer feedforward neural network is a linear unit with leakage correction, and the leakage slope is set to 0.2. Each layer output uses residual connections. For the updated feature vectors of the vertex outputs of the same process node in 50 batches, the arithmetic mean is calculated dimension by dimension. The arithmetic mean of all dimensions is combined into a single vector, which serves as the standard feature vector for that process node.

[0051] The difference vector is obtained by subtracting the updated feature vector of the abnormal node from the standard feature vector dimension by dimension. The dimension of the difference vector is the same as that of the updated and standard feature vectors. The value of each dimension in the difference vector represents the direction and amount of deviation of the abnormal node from the standard state in that dimension. The absolute value of each dimension in the difference vector is calculated. The calculated absolute values ​​of each dimension are sorted from largest to smallest, resulting in a sequence from largest to smallest absolute value. Starting from the first dimension in the sequence, a preset number of dimensions are selected as abnormal feature dimensions. The preset number is set to 10. The basis for this setting is that in 50 batches of products under normal production conditions, the absolute values ​​of each dimension of the difference vector between the updated and standard feature vectors of each batch of products are calculated and sorted. The upper limit of the numerical fluctuation range of the top 10 dimensions is extended upward by 20% as a reference threshold. When the absolute value of a certain dimension of the difference vector of an abnormal node exceeds the reference threshold of that dimension, that dimension is included in the abnormal feature dimension. Statistically, in abnormal quality events, there are usually no more than 10 dimensions that exceed their respective reference thresholds simultaneously.

[0052] Based on the selected anomaly feature dimension, the index number of the anomaly feature dimension is back-mapped to the original spectral band. There is a correspondence between the index number of the anomaly feature dimension and the wavelength position in the one-dimensional spectral sequence. This correspondence is determined when the one-dimensional spectral sequence is generated; that is, when the spectral image data is expanded along the wavelength dimension, the position of the first wavelength in the one-dimensional spectral sequence is determined. Each element corresponds to a wavelength in the spectral image data. The spectral reflectance value. For each anomaly feature dimension index number, query the wavelength position in the corresponding one-dimensional spectral sequence, based on the wavelength. The numerical value determines the corresponding spectral band interval, and this spectral band interval is marked as the key spectral characteristic wavelength interval.

[0053] In some implementations, after selecting the earliest occurring process node from the abnormal nodes as the initial abnormal node, the product state bimodal feature vector of the initial abnormal node is obtained. The product state bimodal feature vector is composed of a binary vector obtained after one-hot encoding, a signal strength value, and a normalized one-dimensional spectral sequence, concatenated sequentially. The value corresponding to the dimension containing the signal strength value is extracted from the product state bimodal feature vector; this value is the signal strength value recorded when the UHF RFID reader reads the passive RFID tag.

[0054] Based on the inverse relationship between signal strength and RFID tag read / write distance, the spatial offset of the insole product at the initial anomaly node is estimated. This inverse relationship is expressed using a logarithmic distance path loss model, and the calculation formula is as follows:

[0055] in, This indicates that the UHF RFID reader has a reading distance of [missing information]. The received signal strength value, in dBm; Indicates the reference distance The reference signal strength value received by the UHF RFID reader. By pre-setting a reference distance The average value was obtained by taking multiple measurements at a distance of 1 meter for the same type of passive RFID tag. The value is -30dBm; This indicates the actual spatial distance between the built-in passive RFID tag and the UHF RFID reader antenna in the insole product, in meters. Indicates the reference distance. Set to 1 meter; The path loss index, ranging from 2.0 to 4.0, is determined in a production line environment by measuring signal strength values ​​at different known distances using pre-deployed reference tags and fitting the data using the least squares method. The value is 2.8; It represents a logarithmic operation with base 10.

[0056] Rearranging the above formula yields the estimated spatial distance. The calculation formula is Signal strength value Reference signal strength value Reference distance and path loss index Substituting into the calculation formula, we obtain the estimated spatial distance. Estimating spatial distance The spatial position offset is obtained by subtracting the nominal working distance of the UHF RFID reader antenna at the initial abnormal node. The nominal working distance of the UHF RFID reader antenna is the expected standard distance value determined during the production line design when the insole product passes the UHF RFID reader antenna on the conveyor belt.

[0057] The preset offset threshold is set to 0.15 meters. This setting is based on the mechanical tolerance range of the production line conveyor belt and the normal positional fluctuation range of the insole product on the conveyor belt, which is within ±0.1 meters. After considering the measurement error margin, the preset offset threshold is determined to be 0.15 meters. The spatial position offset is compared with 0.15 meters. If the spatial position offset exceeds 0.15 meters, the initial abnormal node is marked as a conveyor belt jamming-related abnormal node. A conveyor belt jamming-related abnormal node indicates that the actual spatial position of the insole product at this process node deviates from the expected standard position beyond the normal range, and is associated with material retention or displacement caused by conveyor belt jamming.

[0058] In specific implementation, please refer to Figure 5 After recording the sequence of anomalous nodes as the propagation path of quality anomalies, the time information of each pair of adjacent anomalous nodes in the propagation path is obtained. For adjacent anomalous nodes in the propagation path... and abnormal nodes Abnormal nodes Corresponding to a process node, an abnormal node This corresponds to another process node immediately downstream in the corresponding process sequence. (Identify the abnormal node.) The corresponding UHF RFID reader reads the timestamp. Get abnormal nodes The corresponding UHF RFID reader reads the timestamp. Calculate abnormal nodes. with abnormal nodes The time interval between Time interval This indicates that the insole product has an abnormal node. Transfer to abnormal node The actual time spent.

[0059] Acquire the conveyor belt speed between adjacent process nodes. The conveyor belt speed is output in real time by the production line's programmable logic controller (PLC) via an industrial Ethernet interface, with the speed data sampled once per second. Abnormal nodes... Corresponding read timestamp with abnormal nodes Corresponding read timestamp The average conveyor belt speed is obtained by arithmetically averaging all sampled conveyor belt speeds within the specified time period. .

[0060] Obtain the physical distance between conveyor belts between adjacent process nodes. The physical distances between conveyor belts at each process node of the production line are pre-stored as fixed parameters in the traceability system's configuration database. Indicates abnormal nodes From the current process node to the abnormal node The actual length of the conveyor belt laid between the process nodes. The values ​​are read from the configuration database.

[0061] Calculate the theoretical transmission time between adjacent abnormal nodes. The formula for calculating the theoretical transmission time is:

[0062] in, Indicates abnormal nodes To the abnormal node The theoretical transmission time, in seconds; Indicates abnormal nodes From the current process node to the abnormal node The physical distance between conveyor belts at the respective process nodes, in meters; Indicates in to The average operating speed of the conveyor belt during the time period, expressed in meters per second.

[0063] Calculate time interval Compared with theoretical transmission time ratio The preset ratio range is set to... The preset ratio range is set as follows: Under normal production conditions, time data for 100 consecutive batches of products is collected. The ratio of the time interval between adjacent process nodes to the theoretical transmission time is calculated, and the 99% confidence interval for the normal ratio distribution is obtained. Based on this range, an extension of 0.05 is added to both ends as a tolerance, ultimately determining the preset ratio range as follows: The calculated ratio Compare with the preset ratio range, if the ratio Less than 0.85 or ratio If the value is greater than 1.20, then the abnormal node will be considered. with abnormal nodes The edges between them in the source feature map are marked as propagation delay anomaly edges.

[0064] Connect the abnormal nodes to the abnormal edges with propagation delay. and abnormal nodes The corresponding process nodes are output as quality traceability checkpoints. These checkpoints are highlighted in the quality monitoring interface using both the process node number and the process node name.

[0065] In some implementations, after outputting the upstream and downstream process nodes connected to the propagation delay anomaly edge as quality traceability check points, the starting node of the propagation delay anomaly edge is obtained. Anomaly feature dimensions and endpoint nodes The dimension of abnormal features. Starting node. The method for obtaining the abnormal feature dimension is as follows: obtain the starting node. The difference vector between the updated feature vector and the standard feature vector of the corresponding process node is used to sort the absolute values ​​of each dimension of the difference vector from largest to smallest, and then the top 10 dimensions are selected as the starting node. The dimension of abnormal features. Endpoint node. The method for obtaining the anomaly feature dimension is as follows: obtain the endpoint node. The difference vector between the updated feature vector and the standard feature vector of the corresponding process node is used to sort the absolute values ​​of each dimension of the difference vector from largest to smallest, and then the top 10 dimensions are selected as the endpoint nodes. The dimension of abnormal features.

[0066] Calculate the starting node Anomaly feature dimensions and endpoint nodes The intersection dimension of the abnormal feature dimensions. The intersection dimension belongs to both the starting node and the starting node. The set of abnormal feature dimensions belongs to the endpoint node. The dimensional index numbers of the set of anomalous feature dimensions are used. Each dimensional index number in the intersection dimension is then mapped inversely to the original spectral band. The mapping relationship is determined by the correspondence between the dimensional index number and the wavelength position when the one-dimensional spectral sequence is generated. The spectral band corresponding to each dimensional index number in the intersection dimension is marked as a cross-node common anomalous feature.

[0067] Spectral band information corresponding to common anomaly features across nodes, and starting node and the endpoint The process node number, process node name, and propagation delay anomaly marker information are output to the quality analysis terminal via the data interface. The quality analysis terminal groups one or more cross-node common anomaly features received according to spectral band range. Cross-node common anomaly features that occur more frequently within the same band range are grouped into one cause category. The quality analysis terminal performs clustering display on all cause categories, and the clustering display content includes band range, involved process nodes, and anomaly frequency.

[0068] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for tracing the quality of insoles based on the Internet of Things, characterized in that, Includes the following steps: Acquire RFID tag data and spectral image data of insole products at each process node on the production line; Perform timestamp alignment on RFID tag data and spectral image data to construct a dual-modal feature vector of product status for each process node; The product state bimodal feature vectors of each process node are input into the graph attention network to learn the state transit dependency between adjacent nodes and generate a traceability feature map containing the quality transit path. Based on the feature differences of each node in the traceability feature map, the initial process node and propagation path of the abnormal quality event can be located.

2. The method for quality traceability of insoles based on the Internet of Things according to claim 1, characterized in that, The process of acquiring RFID tag data and spectral image data of insole products at each process node on the production line includes the following steps: UHF radio frequency identification readers and hyperspectral imagers are deployed at each process node of the production line. The passive RFID tag built into the insole product is read by an ultra-high frequency RFID reader to obtain RFID tag data, which includes the tag's unique identifier, reading timestamp, and signal strength value. The spectral reflectance cubes of the insole product in the visible and near-infrared bands are collected by a hyperspectral imager, and spectral image data of specific wavelengths on the surface of the insole product are extracted.

3. The method for quality traceability of insoles based on the Internet of Things according to claim 2, characterized in that, The process of aligning RFID tag data and spectral image data with timestamps to construct a bimodal feature vector of product status for each process node includes the following steps: Using the reading timestamp of the UHF RFID reader as a benchmark, the spectral image data with the smallest absolute difference between the acquisition time and the reading timestamp of the hyperspectral imager is selected as the matching data. The matched spectral image data is expanded into a one-dimensional spectral sequence according to the wavelength dimension; The unique identifier of the label is uniquely encoded and then concatenated with the signal strength value and one-dimensional spectral sequence to generate a dual-modal feature vector of the product status.

4. The method for quality traceability of insoles based on the Internet of Things according to claim 3, characterized in that, After normalizing the one-dimensional spectral sequence, it is then vector-concatenated with the unique identifier of the tag after one-heat encoding and the signal intensity value.

5. The method for quality traceability of insoles based on the Internet of Things according to claim 3, characterized in that, The process of inputting the bimodal feature vectors of product state at each process node into a graph attention network to learn the state transit dependencies between adjacent nodes and generate a traceability feature map containing the quality transit path includes the following steps: Each process node in the production line is defined as a vertex in the graph structure, and the material flow relationship between adjacent process nodes is defined as an edge in the graph structure. For each process node, its product state bimodal feature vector is used as the initial feature of that vertex; In a graph attention network, the features of the neighboring vertices of each vertex are weighted and aggregated. The weights are determined by the similarity between the vertex features. After multiple layers of aggregation, the updated feature vector of each vertex is output. The updated feature vectors of all vertices are organized into a source feature map according to the edge connectivity.

6. The method for quality traceability of insoles based on the Internet of Things according to claim 5, characterized in that, In graph attention networks, the features of neighboring vertices of each vertex are weighted and aggregated. The weights are determined by the similarity between the vertex features. The specific process of outputting the updated feature vector of each vertex after multiple aggregations includes the following steps: The product state bimodal feature vector corresponding to each process node is used as the input feature of the corresponding vertex in the graph attention network, and a linear transformation is performed on the input features of all vertices. The transformed feature vector is recorded as the intermediate representation vector of that vertex. For each edge in the graph structure, the attention coefficient between the two vertices it connects is calculated. The attention coefficient is calculated by concatenating the intermediate representation vectors of the two vertices, inputting the concatenated vector into a single-layer feedforward neural network, and obtaining a scalar value after mapping by a nonlinear activation function, which is used as the original attention value of the edge. Apply a flexible maximum normalization operation to the original attention values ​​of all adjacent edges of each vertex, so that the sum of the normalized attention coefficients is 1. The normalized attention coefficients are the similarity weights between vertex features. The aggregated feature vector of each vertex is obtained by weighting and summing the intermediate representation vectors of all its neighboring vertices according to the corresponding normalized attention coefficients. The aggregated feature vector is residually concatenated with the original intermediate representation vector of the current vertex, and then processed by a multi-layer graph attention network that aggregates layer by layer to output the updated feature vector of each vertex.

7. The method for quality traceability of insoles based on the Internet of Things according to claim 6, characterized in that, The single-layer feedforward neural network uses a linear unit with leakage correction as a nonlinear activation function.

8. The method for quality traceability of insoles based on the Internet of Things according to claim 5, characterized in that, Based on the feature differences of each node in the source tracing feature map, the process of locating the initial process node and propagation path of the abnormal quality event includes the following steps: Calculate the cosine distance between the updated feature vector and the standard feature vector of each vertex in the source feature map to obtain the feature difference degree of each vertex; Vertices whose feature difference exceeds a preset difference threshold are marked as abnormal nodes; Select the process node that appears earliest from the abnormal nodes as the initial abnormal node; Starting from the initial abnormal node in the source feature map, the adjacent abnormal nodes are traversed along the edge direction, and the abnormal node sequence is recorded as the quality abnormality propagation path.

9. The method for quality traceability of insoles based on the Internet of Things according to claim 8, characterized in that, When traversing adjacent abnormal nodes, only the sequence of abnormal nodes whose feature difference does not decrease monotonically along the propagation direction is retained as the final quality abnormality propagation path.

10. A quality traceability system for insoles based on the Internet of Things, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the IoT-based insole product quality traceability method as described in any one of claims 1 to 9.

Images