Product life cycle quality tracing method and device based on extensible ontology

Abstract

The present invention relates to a method and device for product quality traceability throughout its entire life cycle based on an extensible ontology. The method specifically includes the following steps: S1. Acquiring product information of a quality traceability target, constructing an extensible ontology model of the quality traceability target based on the product information, and fusing relevant data; S2. Constructing multiple views of the extensible ontology model based on the fused relevant data, and performing interpretability evaluation on the obtained views; S3. Extracting features from each view based on a graph convolutional neural network to obtain a feature graph for each view; S4. Fusion of the feature graphs of each view to obtain an organic knowledge graph of the quality traceability target and expand the extensible ontology model to obtain a quality traceability ontology model for quality traceability of the quality traceability target. Compared with the prior art, the present invention has the advantages of fusing multi-source data and knowledge, improving the efficiency and accuracy of quality traceability throughout the product life cycle, and achieving the goal of reducing costs and increasing efficiency.

Description

Technical Field

[0001] The present invention relates to the fields of quality tracing throughout a product's entire life cycle, knowledge management, and artificial intelligence, and in particular to a method and device for quality tracing throughout a product's entire life cycle based on an extensible ontology. Background Art

[0002] In the quality tracing process of a product throughout its life cycle, due to the large span of time and space, complex data types, numerous involved links, and non-unique traceability identifiers, work efficiency is low, costs are inflated, and few people conduct knowledge fusion for the entire tracing process. Based on the collection of big data from all links in the entire life cycle, this paper proposes a multi-view knowledge fusion model and method based on extended ontology and GCN. This model fuses multi-source data and knowledge across domains, layers, and time dimensions, effectively improving the efficiency of quality tracing throughout the product life cycle. Summary of the Invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a product full life cycle quality traceability method and device based on extensible ontology, which integrates multi-source data and knowledge across domains, layers, and time dimensions, and effectively improves the efficiency of quality traceability throughout the product life cycle.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] A product life cycle quality traceability method based on an extensible ontology includes the following steps:

[0006] S1. Obtain product information of the quality traceability target, build an extensible ontology model of the quality traceability target based on the product information, and integrate relevant data;

[0007] S2. Build multiple views of the scalable ontology model based on the fused relevant data and evaluate the interpretability of the obtained views;

[0008] S3. Extract features of each view based on the graph convolutional neural network (GCN) to obtain the feature map of each view;

[0009] S4. The feature graphs of each view are further fused to obtain an organic knowledge graph of the quality traceability target and the extensible ontology model is expanded to obtain a quality traceability ontology model to perform quality traceability on the quality traceability target.

[0010] The extensible ontology model is specifically a data fusion model that spans a long period and is multi-source and heterogeneous.

[0011] The extensible ontology model includes a continuously extensible domain ontology and an ontology extracted from valuable events.

[0012] Furthermore, the continuously extensible domain ontology in the extensible ontology model extends the ontology by introducing external knowledge.

[0013] The step S2 further includes extracting a new associated subgraph from the view after the view has been evaluated for interpretability.

[0014] In step S3, feature extraction is performed on each view to obtain a feature map of each view, which specifically includes map construction, frequency domain map convolution, feature extraction, classification, Adam optimization, L2 weight decay and visualization processing.

[0015] Furthermore, in step S3, feature extraction is performed by a feature extractor, which includes two graph convolution layers and an output layer. The main function of the feature extractor is to extract transferable features from the source domain and target domain data.

[0016] Furthermore, in step S3, classification is performed by a classifier, and the classifier includes a softmax layer. The number of neurons in the softmax layer is the number of categories of the faults of the quality traceability target.

[0017] Furthermore, the purpose of L2 weight decay in step S3 is to reduce the weight to a smaller value and to reduce the problem of overweight model to a certain extent.

[0018] Furthermore, the visualization processing in step S3 includes principal component analysis dimensionality reduction.

[0019] The process of fusing the feature maps of each view in step S4 specifically refers to splitting the data according to different views to obtain multiple analysis views with strong interpretability.

[0020] The organic knowledge graph in step S4 is specifically an organic knowledge graph that retains the hierarchical structure of the domain knowledge graph while also taking into account the knowledge structure oriented to business goals.

[0021] A product life cycle quality traceability device based on an extensible ontology, comprising:

[0022] Heterogeneous data preprocessing module integrates the information systems, statistical tables, and quality reports corresponding to quality traceability targets based on an extensible ontology model;

[0023] Build an interpretable multi-view module, construct multiple views of the scalable ontology model based on expert experience, and obtain multiple interpretability evaluation views;

[0024] Based on the multi-view knowledge fusion module, feature extraction based on graph convolutional neural network is performed on multiple views, and goal-oriented knowledge fusion is performed to obtain an organic knowledge graph of quality traceability goals;

[0025] The quality traceability module expands the extensible ontology model based on the organic knowledge graph to obtain the quality traceability ontology model to perform quality traceability on the quality traceability target.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] Based on the collection of big data from all aspects of the entire life cycle, the present invention uses a multi-view knowledge fusion model based on an extensible ontology model and a graph convolutional neural network to fuse multi-source data and knowledge across domains, layers, and time dimensions, and obtain an organic knowledge graph that is more in line with the quality traceability goal for quality traceability, effectively improving the efficiency and accuracy of quality traceability throughout the product life cycle, and ultimately achieving the goal of reducing costs and increasing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a structural schematic diagram of the present invention;

[0029] Figure 2 It is a structural schematic diagram of the device of the present invention;

[0030] Figure 3 Schematic diagram of multiple views output by a graph convolutional neural network in an embodiment of the present invention, wherein Figure 3 (a), (b), (c), and (d) are views of the output of the graph convolutional neural network;

[0031] Figure 4 Schematic diagram of an organic knowledge graph according to an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram of a module with the largest proportion of failure losses in an embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of an area with large fault losses in an embodiment of the present invention;

[0034] Figure 7 This is a schematic diagram of a region with a relatively long usage time in an embodiment of the present invention;

[0035] Figure 8 This is a graph showing the accuracy of the initial after-sales judgment in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0037] Example

[0038] like Figure 1 As shown in the figure, a product life cycle quality traceability method based on extensible ontology includes the following steps:

[0039] S1. Obtain product information of the quality traceability target, build an extensible ontology model of the quality traceability target based on the product information, and integrate relevant data;

[0040] S2. Build multiple views of the scalable ontology model based on the fused relevant data and evaluate the interpretability of the obtained views;

[0041] S3. Extract features of each view based on the graph convolutional neural network (GCN) to obtain the feature map of each view;

[0042] S4. The feature graphs of each view are further fused to obtain an organic knowledge graph of the quality traceability target and the extensible ontology model is expanded to obtain a quality traceability ontology model to perform quality traceability on the quality traceability target.

[0043] The extensible ontology model is specifically a data fusion model that spans a long period and has multiple sources and heterogeneity.

[0044] The extensible ontology model includes the continuously extensible domain ontology and the ontology extracted from valuable events.

[0045] The continuously extensible domain ontology in the extensible ontology model extends the ontology by introducing external knowledge.

[0046] Step S2 also includes extracting new subgraphs of the associated type after the view has been evaluated for interpretability.

[0047] In step S3, feature extraction is performed on each view to obtain the feature map of each view. The process specifically includes map construction, frequency domain map convolution, feature extraction, classification, Adam optimization, L2 weight decay and visualization processing.

[0048] In step S3, feature extraction is performed using a feature extractor. The feature extractor consists of two graph convolutional layers and an output layer. The graph convolutional layers then serve as activation functions, mapping the outputs of neurons. The main function of the feature extractor is to extract transferable features from the source and target domain data. After feature extraction, the data is normalized. The goal of the model in the feature extractor is to find characteristic regions, learn feature mappings between the source and target domain data in the feature space, and reduce the data distribution differences between the source and target domains. Therefore, the goal of the model in the feature extractor is to learn domain-invariant features.

[0049] In step S3, classification is performed by a classifier. The main function of the classifier is to classify the features extracted by the feature extractor during the model training and prediction process.

[0050] The classifier maps the features learned by the feature extractor to the sample space, where Dropout is applied to randomly discard some neurons. The classifier includes a softmax layer. The number of neurons in the softmax layer is the number of fault categories of the quality traceability target, and the corresponding result represents the fault label.

[0051] In this embodiment, the scalable ontology model is to identify the health status of the detection equipment based on the features learned by the feature extractor. Therefore, the goal of the model on the classifier is to reduce the fault classification error on the source domain data, which is expressed as minimizing the nll_loss loss function on the source domain data set to minimize the classification loss. The calculation method of the Softmax activation function is to first transform each input x into i beg Then divide each value x by i The corresponding exponential sum with base e is calculated as follows:

[0052]

[0053] LogSoftmax is the output of Softmax, and then calculates the logarithm. Its calculation formula is as follows:

[0054]

[0055] As can be seen from the above analysis, the value range of the output of Softmax is [0, 1], and the meaning of the value represents the probability of the corresponding category. However, in machine learning, it is usually to minimize a certain function value, so the probability value is further subjected to a log transformation. After the log transformation, the maximum probability value is the value closest to 0. After the log transformation of other probability values, they are all real values less than 0. In order to meet the minimization condition, the loss function usually takes a negative sign. At this time, if the value at the corresponding position of the actual label tensor in the model output result logsfm(output) is closer to 0, the loss value is smaller; otherwise, the loss value is larger.

[0056] The purpose of L2 weight decay in step S3 is to reduce the weight to a smaller value and to a certain extent reduce the problem of the model being too heavy.

[0057] In this embodiment, the Adam optimization algorithm and the L2 weight decay process are as follows:

[0058] Adam is essentially RMSprop with a momentum term. It uses the first-order moment estimate and second-order moment estimate of the gradient to dynamically adjust the learning rate of each parameter. Its main advantage is that after bias correction, the learning rate has a definite range for each iteration, making the parameters relatively stable. Its formula is as follows:

[0059] m t = β1 * m t-1 + (1 - β1) * g t

[0060] v<00000​​​​​​​​​​​​​​​​​​​​​​​The corrections to the first- and second-order moment estimates in the third and fourth formulas can be approximated as unbiased estimates of the expectation. It can be seen that direct moment estimates of the gradient do not require additional memory and can be dynamically adjusted based on the gradient. The preceding part of the last term forms a dynamic constraint on the learning rate η, with a clear range. α and ∈ are process parameters.

[0065] The purpose of L2 weight decay is to make the weight decay to a smaller value, reducing the problem of model overfitting to a certain extent, so weight decay is also called L2 regularization.

[0066] L2 regularization is to add a regularization term after the cost function:

[0067]

[0068] Among them, C0 represents the original cost function, and the following term is the L2 regularization term: the sum of the squares of all parameters w, divided by the sample size n of the training set, and λ is the regularization term coefficient (i.e., the weight decay coefficient), which weighs the proportion of the regularization term and the C0 term.

[0069] The visualization processing in step S3 includes principal component analysis dimensionality reduction.

[0070] In this embodiment, the visualization process is as follows:

[0071] PCA reduces n-dimensional input data to R dimensions, where r < n. Simply put, PCA is essentially a basis transformation that maximizes the variance of the transformed data. This involves rotating the coordinate axes and translating the origin to minimize the variance between one axis (the principal axis) and the data points. After the coordinate transformation, the orthogonal axes with high variance are removed, resulting in a reduced-dimensional dataset.

[0072] In this embodiment, the SVD method is used to perform PCA dimensionality reduction. Assuming that there are p×n dimensional data samples X, there are p samples in total, and each row is n-dimensional, the p×n real matrix can be decomposed into:

[0073] X=U∑V T

[0074] Y r =U∑ r

[0075] Among them, the dimension of the orthogonal matrix U is p×n, and the dimension of the orthogonal matrix V is n×n (the orthogonal matrix satisfies: UU T =V T V=1), ∑ is an n×n diagonal matrix; divide E into r columns, denoted as E r ; Using U and V, we can get the dimension-reduced data point Y r .

[0076] In this embodiment, the graph construction and frequency domain graph convolution process in step S3 are as follows:

[0077] A graph can be represented as G = (V, E, W), where V represents the nodes in the graph, E represents the edge between two nodes, and W represents the weight of the edge between two vertices. Furthermore, a graph can be represented as a Laplacian matrix, L = DA, where D and A represent the degree matrix and adjacency matrix, respectively. The Laplacian matrix can be eigen-decomposed as follows:

[0078]

[0079] in, is a matrix composed of unit eigenvectors, Λ is a diagonal matrix composed of the eigenvalues ​​of the Laplace matrix, λ1~λ n is the eigenvalue. Because U is an orthogonal matrix (i.e. UU T =E), so the eigendecomposition of the Laplace matrix can also be written as follows:

[0080]

[0081] Because the Laplacian matrix has the above two properties, graph convolution operations can be quickly implemented in the frequency domain.

[0082] The adaptive graph convolutional neural network is mainly based on the frequency domain convolution theory. Its core idea is to perform a graph Fourier transform on the graph signal, perform convolution in the spectrum domain, and then perform an inverse graph Fourier transform on the signal. The graph Fourier transform operation is as follows:

[0083]

[0084] Among them, f is the N-dimensional component of the graph, f(i) corresponds one-to-one to the nodes in the graph; u l (i) represents the i-th component of the l-th eigenvector; u l * (i) means u l (i) is the conjugate vector. The Fourier transform matrix of the figure is The corresponding inverse Fourier transform of the graph is:

[0085]

[0086] So the matrix of the inverse Fourier transform on the graph is According to the convolution theorem (that is, the convolution of functions f(t) and g(t) is the inverse transform of the product of their Fourier transforms), the graph convolution is as follows:

[0087]

[0088] Among them, f*h represents the convolution of functions f(t) and g(t), U T f represents the Fourier transform of f, and g represents the convolution kernel.

[0089] According to the characteristics of the Laplace matrix itself, (f*h) G =U(U T h⊙U T f), where ⊙ is the Hadamard product, which represents the element-wise product operation of two vectors of the same dimension at corresponding positions. Therefore, the output of the graph convolutional network is expressed as follows:

[0090] y output =σ(Ug θ (λ)U T x)

[0091] Among them, x is the input feature, σ(*) is the activation function, g θ (λ) is the convolution kernel.

[0092] The process of fusing the feature maps of each view in step S4 specifically refers to splitting the data according to different views to obtain multiple analysis views with strong interpretability.

[0093] The organic knowledge graph in step S4 is specifically an organic knowledge graph that retains the hierarchical structure of the domain knowledge graph while also taking into account the knowledge structure oriented towards business goals.

[0094] like Figure 2 As shown, a product life cycle quality traceability device based on an extensible ontology includes:

[0095] Heterogeneous data preprocessing module integrates the information systems, statistical tables, and quality reports corresponding to quality traceability targets based on an extensible ontology model;

[0096] Build an interpretable multi-view module, construct multiple views of the scalable ontology model based on expert experience, and obtain multiple interpretability evaluation views;

[0097] Based on the multi-view knowledge fusion module, feature extraction based on graph convolutional neural network is performed on multiple views, and goal-oriented knowledge fusion is performed to obtain an organic knowledge graph of quality traceability goals;

[0098] The quality traceability module expands the extensible ontology model based on the organic knowledge graph to obtain the quality traceability ontology model to perform quality traceability on the quality traceability target.

[0099] In the specific implementation, an extensible ontology for Haier water heater quality traceability is constructed, as shown in Table 1:

[0100] Table 1 Haier water heater expandable body information table

[0101]

[0102]

[0103]

[0104]

[0105] The pseudo code flow of the GCN algorithm is as follows:

[0106] Input: node number node, category class, feature feature and adjacency matrix adj of source data;

[0107] Output: clustering effect and classification accuracy of the model.

[0108] Step 1: Initialize hyperparameters, such as the number of iterations epochs, learning rate lr, weight decay weight_decay, and randomly initialize network parameters, such as weight parameter W, number of hidden layers hidden, dropout, etc.

[0109] Step 2: Calculate the feature representation features of the source data after passing through the feature extractor;

[0110] Step 3: Calculate the nll_loss of the classification and update the classifier parameters through model backpropagation;

[0111] Step 4: Calculate the model loss according to formula (14) and the Adam single-shot optimization algorithm, and update the feature extractor parameters through model backpropagation;

[0112] Step 5: Repeat step 2 until the network converges or reaches the pre-set number of iterations.

[0113] By extending the ontology construction, we can establish rich data relationships and links between multi-source heterogeneous data and evaluate the weights between data, so as to achieve flexible and efficient data fusion. Based on GCN feature extraction, after interpretability evaluation, we can get the following: Figure 3 The multiple views shown above are split according to different views to obtain multiple analytical views with strong interpretability in the quality traceability process, and then fused. Figures 4 to 8 As shown, the final result is an organic knowledge graph, which not only retains the hierarchical structure of the domain knowledge graph, but also takes into account the knowledge structure oriented to business goals.

[0114] Specifically, refer to Figure 5As shown, the module with the largest loss ratio in the water heater is the inner tank. The main problems are water pipe welding problems, flange problems, inner tank girth welding problems, inner tank straight seam welding problems, inner tank quality problems, and inner tank T-shaped port problems. The average service life of the water heater in the embodiment is 5.4 years. Figure 6 As shown, the region with longer water heater usage time is referenced Figure 7 As shown, the accuracy rate of after-sales initial judgment is based on the standard Figure 8 shown.

Claims

1. A product life cycle quality traceability method based on extensible ontology, characterized by: The specific steps include: S1. Obtain product information of the quality traceability target, build an extensible ontology model of the quality traceability target based on the product information, and integrate relevant data; S2. Build multiple views of the scalable ontology model based on the fused relevant data and evaluate the interpretability of the obtained views; S3. Extract features from each view based on a graph convolutional neural network to obtain a feature map of each view. S4. The feature graphs of each view are further fused to obtain an organic knowledge graph of the quality traceability target and the extensible ontology model is expanded to obtain a quality traceability ontology model to perform quality traceability on the quality traceability target; The extensible ontology model includes a continuously extensible domain ontology and an ontology extracted from valuable events; The product information and related data include: Traceability work order, including traceability work order number, process step number, process step sequence, process step name, entry date, and removal date. Product design, including product number, product name, barcode, serial number, warranty period, cost, selling price, gross profit, Inventory inventory, including material / product number, location, quantity, status, first-in-first-out date, batch number, barcode, Production work order, including work order number, work order date, product number, production quantity, status, equipment number, process number, Whole machine quality inspection, including number, faulty module, faulty submodule, fault details, final judgment, final judgment date, and responsible person. Logistics, including number, shipping date, shipping place, receiving place, waybill number, status, User purchase: including purchase number, purchase date, location, dealer point, invoice number, User data, including number, user name, contact information, purchase date, repair date, user level, After-sales service outlets: including number, date of receiving repair request, date of repair request, location, dealer point, initial judgment, Scrap: including product number, product name, barcode, serial number, loss, scrap / return date, scrap / return judgment, Bill of Materials: including parent number, child number, quantity, sequence number, unit, update date, version number, and status; The organic knowledge graph in step S4 is specifically an organic knowledge graph that retains the hierarchical structure of the domain knowledge graph while also taking into account the knowledge structure oriented to business goals.

2. The product life cycle quality traceability method based on extensible ontology according to claim 1 is characterized in that: The extensible ontology model is specifically a data fusion model that spans a long period and is multi-source and heterogeneous.

3. The product life cycle quality traceability method based on extensible ontology according to claim 1 is characterized in that: The continuously extensible domain ontology in the extensible ontology model extends the ontology by introducing external knowledge.

4. The product life cycle quality traceability method based on extensible ontology according to claim 1 is characterized in that: In step S3, feature extraction is performed on each view to obtain a feature map of each view, which specifically includes map construction, frequency domain map convolution, feature extraction, classification, Adam optimization, L2 weight decay and visualization processing.

5. The product life cycle quality traceability method based on extensible ontology according to claim 4 is characterized in that: In step S3, feature extraction is performed by a feature extractor, and the feature extractor includes two graph convolution layers and an output layer.

6. The product life cycle quality traceability method based on extensible ontology according to claim 4 is characterized in that: In step S3, classification is performed by a classifier, and the classifier includes a softmax layer. The number of neurons in the softmax layer is the number of categories of the faults of the quality tracing target.

7. The product life cycle quality traceability method based on extensible ontology according to claim 4 is characterized in that: The visualization processing in step S3 includes principal component analysis dimensionality reduction.

8. A device for implementing the product lifecycle quality traceability method based on an extensible ontology as described in any one of claims 1 to 7, characterized in that: include: Heterogeneous data preprocessing module integrates the information systems, statistical tables, and quality reports corresponding to quality traceability targets based on an extensible ontology model; Build an interpretable multi-view module, construct multiple views of the scalable ontology model based on expert experience, and obtain multiple interpretability evaluation views; Based on the multi-view knowledge fusion module, feature extraction based on graph convolutional neural network is performed on multiple views, and goal-oriented knowledge fusion is performed to obtain an organic knowledge graph of quality traceability goals; The quality traceability module expands the extensible ontology model based on the organic knowledge graph to obtain the quality traceability ontology model to perform quality traceability on the quality traceability target.

Images