Production business process node vector representation method fusing structure and behavior features

Abstract

The application discloses a production business process node vector representation method fusing structure and behavior characteristics. The method comprises the following steps: firstly, structure characteristics of a production business process are extracted by using a random walk algorithm and a Skip-gram model; then, behavior characteristics of the production business process are extracted by using the Skip-gram model; finally, a deep Boltzmann machine is introduced to fuse the structure characteristics and the behavior characteristics of the production business process, thereby automatically performing vector representation on nodes of the production business process, extracting and fusing various characteristics to improve the accuracy of the node vector representation, and further improving the accuracy of subsequent production business process recommendation, optimization and the like. The application has the advantages that the graph structure characteristics of the production business process itself and the behavior characteristics based on historical execution logs are fused, deep learning technology is used to strengthen the production business process model, the node vector representation which simply depends on the historical execution logs is supplemented and improved, and the accuracy of the node vector representation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of business process management, in particular to a production business process node vector representation method fusing structure and behavior characteristics. BACKGROUND

[0002] With the progress of science and technology and the development of manufacturing industry, manufacturing enterprises are facing the challenges of product cost compression, product change speed acceleration, product quality improvement and service quality improvement, which puts forward higher requirements for the execution of production business process, such as lower cost, shorter time, more optimal resource utilization, etc. Therefore, optimizing production business process to help enterprises continuously increase profits, improve core competitiveness and promote the intelligent upgrading of traditional manufacturing enterprises is an important problem to be solved for manufacturing enterprises. In order to optimize the production business process, the first step is to represent the nodes of the production business process, and the goal is to learn to generate low-dimensional dense vectors to represent the process nodes, so that the production business process optimization task is directly carried out in the learned low-dimensional vector space.

[0003] The traditional production business process node representation method mainly focuses on manually designing or extracting process features, which is low in efficiency and the extraction difficulty increases rapidly with the complexity of the process. In addition, although the traditional process feature extraction is targeted, the extracted feature dimension is small, and mostly only involves single aspect features, which has strong application scene dependency. Therefore, there is an urgent need for an automatic production business process node feature representation method, which considers multiple aspects of features such as structure and behavior of production business process, and fuses and extracts different features to better represent the nodes of production business process. SUMMARY

[0004] In order to automatically represent the nodes in the production business process, the present application proposes a production business process node vector representation method fusing structure and behavior characteristics. First, random walk is performed on the graph structure of the production business process to obtain a node sequence, and on this basis, the structure vector of the node is learned using the Skip-gram model; then the historical execution log of the production business process is input into the Skip-gram model to learn the behavior vector of the node; finally, the structure and behavior vectors of the node are fused using a deep Boltzmann machine to obtain the production business process node feature vector fusing structure and behavior characteristics.

[0005] A production business process node vector representation method fusing structure and behavior characteristics, comprising the following steps:

[0006] (1) input the graph structure G={P,T,E} of the production business process modeled by Petri net, wherein P represents the place node, T is the task node, E is the edge, and the place is the state of the task node.

[0007] (2) Obtain the structure vector of the node in the production business process.

[0008] (3) Input the historical execution log of the production business process L={δ1,δ2,...,δ |L|} Consists of multiple execution trajectories δ i Composition, |L| represents the number of execution traces contained in the historical execution log, and an execution trace δ=<v1,v2,...,v |δ| > It consists of multiple task nodes, where |δ| represents the length of the execution trajectory.

[0009] (4) Use the Skip-gram model to generate behavioral vectors for production business process nodes.

[0010] (5) Using a deep Boltzmann machine, the structural feature vector s and behavioral feature vector b of the low-level production business process nodes are extracted by two Gaussian restricted Boltzmann machines to extract the hidden unit features of the intermediate layer. Then, a layer of binary hidden units is superimposed to fuse v. s and v b For the combination of structural and behavioral characteristics of production business process nodes (v s ,v b The probability of the conditional activation of the visible and hidden units is shown in Equations 5 to 10:

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] Where θ represents the optimal model parameters. Indicates v s v b Let the vector space be N-dimensional and each component of the vector be a real number. K is the number of task nodes in the production process, D is the size of the node table, and h represents the hidden unit. (2) For the second hidden unit, W ij The weights representing the connection between visible unit i and hidden unit j are sigm, which is the sigmoid function.

[0018] (6) Train a deep Boltzmann machine to make the activation probability of the hidden units converge, thereby obtaining the node feature vector that integrates structure and behavior.

[0019] Furthermore, step (2) consists of the following steps:

[0020] (2.1) Construct a random walk sequence generator. Given the current node v in graph G, visit its next node v. next The probability is shown in Formula 1:

[0021]

[0022] Where E represents the edge set of graph G, (v,v) next )∈E represents node v and node v next There are edges between them. Representing node v and node v next There is no edge between them. It is node v and node v next The unnormalized transition probability, Z is the normalization constant. A series of node sequences are obtained by performing random walks on graph G.

[0023] (2.2) The Skip-gram model is introduced to probabilistically model each node in the obtained random walk sequence. The stochastic gradient descent algorithm is used to obtain the structure vector of the node. The objective function of the algorithm is shown in Equation 2:

[0024]

[0025] Where V represents the set of nodes in graph G, Φ(v) is the structure vector of the current node v, and N(v) are the neighbor nodes of v obtained through random walk sampling. If we assume that each sampled neighbor node is independent of the others, then Equation 2 can be simplified to:

[0026]

[0027] Among them, c i Let p(c) represent the i-th node in the neighboring nodes N(v), ∏ represent the product, and p(c) represent the ith node in the neighboring nodes N(v). i |Φ(v)) is obtained from the softmax function defined by formula (4), Pr(c i |Φ(v)) is obtained from the softmax function defined in Equation 4:

[0028]

[0029] Where e represents the natural base, Φ(c i ) and Φ(u) represent node c respectively i The structure vector of node u.

[0030] Further, the step (4) consists of the following steps:

[0031] (4.1) Prepare training data, get the task node set V of the historical execution log, generate a one-hot vector for each node, that is, each node is a Vx1 vector w, and the node table composed of all task nodes constitutes a VxV matrix M.

[0032] (4.2) Initialize two matrices W and W' with a size of Vxd, where d is the space size after encoding the node, W is the node vector matrix when the task node is the center node, and W' is the same. Each column in W and W' corresponds to the vector information of the encoded node in each column of matrix M, which will be updated in the subsequent optimization process, and it is both a weight matrix and a final output node behavior vector.

[0033] (4.3) Use the center node v c Do inner product for each node u o in its context, calculate u o T v c .

[0034] (4.4) Map the node table using the softmax formula to calculate the probability of each node The result gets the probability of a certain node, and then it is converted into one-hot form, and the output result is w.

[0035] (4.5) Maximize P(w o |w c ), that is, maximize the probability of output w c when the one-hot vector of the input node in a certain context is w o .

[0036] (4.6) Perform back propagation to update the two weight matrices W and W', and take the center node vector matrix W as the behavior vector of the node after optimization.

[0037] The technical concept of the application is: first, use the random walk algorithm and the Skip-gram model to extract the structural features of the production business process, then use the Skip-gram model to extract the behavior features of the production business process, and finally introduce the deep Boltzmann machine to fuse the structural features and behavior features of the production business process. Thus, the nodes of the production business process are automatically vector represented, avoiding single-dimensional feature node vector representation, extracting and fusing multiple features to improve the accuracy of node vector representation, and further improving the accuracy of subsequent production business process recommendation, optimization and other tasks.

[0038] The advantage of the application is:

[0039] 1) The structure features of the fusion production business process itself and the behavior features based on the historical execution log are fused, the production business process model is strengthened by using deep learning technology, the node vector representation which simply depends on the historical execution log is supplemented and improved, and the accuracy of the node vector representation is improved;

[0040] 2) The Skip-gram is used to extract the structure and behavior features of the production business process node respectively, the one-sidedness caused by only considering the structure or behavior single feature can be avoided, and the accuracy of the production business process node vector representation is improved;

[0041] 3) The deep Boltzmann machine is introduced to fuse the structure and behavior features of the production business process node, not only simply splicing the structure feature vector and the behavior feature vector, but also finding the complementary relationship between different features, fusing and extracting different features by a certain information fusion method, better representing the production business process node, and improving the accuracy of subsequent business process node classification, recommendation, optimization and the like. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the overall flowchart of the present application;

[0043] Figure 2 is a product processing process example diagram;

[0044] Figure 3 is a random walk example diagram;

[0045] Figure 4 is a historical execution log example diagram of Figure 2 ;

[0046] Figure 5 is a Skip-gram model diagram;

[0047] Figure 6 is a deep Boltzmann machine diagram fusing the structure and behavior of the production business process. DETAILED DESCRIPTION

[0048] The present application is further described below in conjunction with the drawings of the specification.

[0049] As shown in Figure 1 , a production business process node vector representation method fusing structure and behavior features, the method comprises:

[0050] (1) input the graph structure G={P, T, E} of the production business process modeled by the Petri net, wherein P represents the place node, T is the task node, E is the edge, and the place is the state of the task node.

[0051] (2) Obtain the structure vector of the node in the production business process, specifically including the following steps:

[0052] (2.1) Construct a random walk sequence generator, given the current node v in the graph G, access its next node v next The probability is shown in formula 1:

[0053]

[0054] Wherein, is the node v and the node v next The transition probability is not normalized, and Z is the normalization constant. A series of node sequences are obtained by random walk on the graph G.

[0055] (2.2) Introduce the Skip-gram model to model the probability of each node in the obtained random walk sequence, and use the stochastic gradient descent algorithm to obtain the structure vector of the node. The objective function of the algorithm is shown in formula 2:

[0056]

[0057] Wherein, Φ(v) is the structure vector of the current node v, and N(v) is the adjacent node of v obtained by random walk sampling. If it is assumed that each adjacent node obtained by sampling is independent, formula 2 can be simplified as:

[0058]

[0059] Wherein, Pr(c i |Φ(v)) is obtained by the softmax function defined in formula 4:

[0060]

[0061] (3) Input the historical execution log L = {δ1, δ2,..., δ |L|} of the production business process, which is composed of multiple execution traces δ i , |L| represents the number of execution traces contained in the historical execution log, and an execution trace δ = <v1, v2,..., v |δ| > is composed of multiple task nodes, and |δ| represents the length of the execution trace.

[0062] (4) Generate the behavior vector of the node of the production business process using the Skip-gram model, specifically including the following steps:

[0063] (4.1) Prepare the training data, obtain the task node set V of the historical execution log, and generate a one-hot vector for each node, that is, each node is a Vx1 vector w, and the node table composed of all task nodes constitutes a VxV matrix M.

[0064] (4.2) Initialize two matrices W and W' with size Vxd, where d is the size of the space after encoding the nodes, W is the node vector matrix when the task node is the center node, and W' is the same. Each column in W and W' corresponds to the vector information of the encoded node in each column of matrix M, which will be updated in the subsequent optimization process, and it is both the weight matrix and the final output node behavior vector.

[0065] (4.3) Use the center node v c Do the inner product for each node u o in its context, calculate u o T v c .

[0066] (4.4) Map the node table using the softmax formula to calculate the probability of each node The result gets the probability of a certain node and then converts it to one-hot form, and the output result is w.

[0067] (4.5) Maximize P(w o |w c ), that is, maximize the probability of output w c when the one-hot vector of the input node is w o in a certain context.

[0068] (4.6) Backpropagation, update two weight matrices W and W', and take the center node vector matrix W as the behavior vector of the node after optimization.

[0069] (5) Use deep Boltzmann machine to extract intermediate layer hidden unit features by respectively passing the production business process node low-level structure feature vector s and behavior feature vector b through two Gaussian restricted Boltzmann machines, and then superimpose a layer of binary hidden units to fuse v s and v b , the probability of the structure and behavior feature combination (v s , v b ) of the production business process node, the conditional activation probability of the visible unit and the hidden unit are shown in formulas 5-10:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] where, K is the number of task nodes in the production business process, D is the node table size, h represents the hidden unit, h (2) is the second layer hidden unit, W ij represents the weight of the connection between the visible unit i and the hidden unit j, and sigm is the Sigmoid function, i.e.

[0077] (6) Train the deep Boltzmann machine so that the activation probability of the hidden unit converges, and thus obtain the node feature vector that fuses the structure and behavior.

[0078] Figure 2 The product processing process instance graph G = {P, T, E} modeled by the Petri net, where P = {P1, P2, P3, P4, P5, P6, P7, P8, P9, P 10 , P 11 , P 12}, T = {material preparation, splicing, wire welding, quartz tube winding, wire joint, fastening screw, sealing, hot melt glue assembly, assembly, packaging}, E = {P1→material preparation, material preparation→P2, material preparation→P3, P2→splicing, splicing→P3, splicing→P4, P3→wire welding, P4→quartz tube winding, wire welding→P5, quartz tube winding→P6, P5→wire joint, P6→wire joint, wire joint→P7, P7→assembly, P3→fastening screw, P3→hot melt glue assembly, fastening screw→P9, P9→sealing, sealing→P 10 , hot melt glue assembly→P 10 , P 10 →assembly, assembly→P 11 , P 11 →packaging, packaging→P 12}. The process represents a product processing process starting with material preparation, which includes metal and plastic materials: when processing metal materials, the metal materials are first spliced, then wire welding and quartz tube winding are performed simultaneously, and finally wire joint is performed; when processing plastic materials, materials that need to be fastened are fastened with screws, then sealed, and materials that do not need to be fastened are directly glued with hot melt glue. Then, the metal materials and plastic materials are assembled, and the assembled product is packaged.

[0079] Figure 3 represents the extraction of the feature vector of the product processing process using the random walk method Figure 2The node sequence of the product processing flowchart structure. The random walk method randomly selects a task node in the product processing flowchart, performs random breadth-first traversal or depth-first traversal along the edges in the flowchart, and then generates a series of task node sequences, such as "preparing materials → splicing → wire welding → quartz tube winding", "preparing materials → splicing → wire welding → wire joint → assembly → packaging", "preparing materials → fastening screws → sealing → assembly → packaging", "fastening screws → sealing → hot melt glue assembly", "splicing → wire welding → wire joint → assembly → packaging", and the like.

[0080] Figure 4 represents Figure 2 The historical execution log of the product processing flow, wherein each row represents a case, i.e., a real execution sequence, the first column indicates the number of the case, and the second column is the specific execution sequence. For example, the case numbered 1 has the execution sequence "preparing materials → fastening screws → sealing → assembly", and the case numbered 2 has the execution sequence "preparing materials → hot melt glue assembly → assembly".

[0081] Figure 5 represents the Skip-gram model. The process of a training instance of the Skip-gram is as follows: if the structural features of the production business process nodes are to be extracted, the input is the one-hot encoding vector x s of the nodes in the node sequence extracted by the random walk method, if the behavioral features of the production business process are to be extracted, the input is the one-hot encoding vector x b of the nodes in each execution sequence in the historical execution log, and the output is a probability vector y of the output node. The node embedding matrix WV×N is a weight vector between the input layer and the hidden layer, and each row thereof is a node vector of size N. The hidden layer h is a matrix of size 1×N, which can be calculated as follows:

[0082] h = x T W = W (i,·) (Equation 11)

[0083] The context matrix W' N×V is a weight matrix between the hidden layer and the input layer. The output layer is a 1×V matrix, which is calculated as follows:

[0084] y j = hW' (·,j) (Equation 12)

[0085] In order to obtain the probability distribution of the context nodes, each component of the output layer is calculated using a softmax classifier:

[0086]

[0087] In fact, the Skip-gram model will train a large number of instances, and the objective is to minimize the objective function by learning the matrix W and W' through the back propagation algorithm and the stochastic gradient descent algorithm. Various techniques, such as negative sampling, can be introduced to train the Skip-gram model. After training, the node embedding matrix W can be obtained, and the structural or behavioral feature vectors of the nodes can be extracted from W.

[0088] Figure 6 The deep Boltzmann machine structure represents the fusion of the production business process structure and behavior, wherein v s The production business process node structure representation is represented by v b The production business process node behavior representation is represented by v The two features extract the features of the intermediate layer hidden units through two Gaussian restricted Boltzmann machines, and then fuse the two feature vectors through a superimposed binary hidden unit. In the training of the deep Boltzmann machine, first, the two Gaussian restricted Boltzmann machines are trained through the contrastive divergence algorithm. For each Gaussian restricted Boltzmann machine, the data of the hidden layer is learned according to the data of the visible layer. For the second Gaussian restricted Boltzmann machine, the first hidden layer obtained can be regarded as the visible layer to obtain the data of the second hidden layer. Then, the two Gaussian restricted Boltzmann machines are fine-tuned, the data is updated through the CD (Contrastive Divergence) algorithm, and finally a binary hidden unit is superimposed to realize the deep Boltzmann machine, so as to fuse the structural and behavioral feature vectors of the production business process nodes.

[0089] The content described in the embodiments of the present specification is only a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms stated in the embodiments, and the protection scope of the present application also includes equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A method for representing production business process nodes using vectors that integrate structural and behavioral characteristics, characterized in that, The method comprises the following steps: 1) inputting a graph structure G={P, T, E} of a production business process modeled by a Petri net, wherein P represents a place node, T represents a task node, E represents an edge, and the place is a state of the task node; 2) obtaining a structure vector of a node in the production business process; 3) Input the historical execution log of the production business process L={δ1,δ2,...,δ |L| } Consists of multiple execution trajectories δ i Composition, |L| represents the number of execution traces contained in the historical execution log, and an execution trace δ=<v1,v2,...,v |δ| > It consists of multiple task nodes, where |δ| represents the length of the execution trajectory; 4) generating a behavior vector of the node of the production business process by using a Skip-gram model; 5) Using deep Boltzmann machine, by the production business process node low layer structure feature vector s and behavior feature vector b respectively through two Gaussian limited Boltzmann machine to extract the middle layer hidden unit features, and then through the superposition of a layer of binary hidden unit to fuse v s and v b , the probability of the combination of the structure and behavior features of the production business process node (v s ,v b ) is shown in formulas (5)-(10): where θ represents the optimal model parameters, represents v s , v b is an N-dimensional vector space and each component in the vector is a real number, K is the number of task nodes in the production business process, D is the node table size, h represents the hidden unit, h (2) is the second layer hidden unit, W ij represents the weight of the connection between the visible unit i and the hidden unit j, and sigm is a Sigmoid function, that is 6) training a deep Boltzmann machine so that the activation probability of a hidden unit converges, and then obtaining a node feature vector fused with the structure and the behavior.

2. The fusion of structure and behavior feature production business process node vector representation method of claim 1, wherein: Step 2) specifically comprises the following steps: 2.1) Construct a random walk sequence generator, given the current node v in graph G, visit its next node v next The probability of v is given by equation (1): Where E represents the edge set of graph G, (v,v) next )∈E represents node v and node v next There are edges between them. Representing node v and node v next There is no edge between them. It is node v and node v next The unnormalized transition probability, Z is a normalization constant, obtained by performing a random walk on graph G to obtain a series of node sequences; 2.2) introducing a Skip-gram model to probabilistically model each node in the obtained random walk sequence, and obtaining a structure vector of the node by using a stochastic gradient descent algorithm, wherein an objective function of the algorithm is shown in formula (2): wherein V represents a node set of the graph G, Φ(v) is a structure vector of a current node v, and N(v) is a neighboring node of v obtained by random walk sampling; if it is assumed that each neighboring node obtained by sampling is independent of each other, then formula (2) is simplified as: where c i denotes the i-th node in the neighboring nodes N(v), denotes the product, p(c i |Φ(v)) is obtained by the softmax function defined by formula (4): where e denotes the natural base, Φ(c i ) and Φ(u) denote the structure vectors of node c i and node u, respectively.

3. The fusion of structure and behavior feature production business process node vector representation method of claim 1, wherein: Step 4) specifically comprises the following steps: 4.1) preparing training data, obtaining a task node set V of a historical execution log, generating a one-hot vector for each node, that is, each node is a V×1 vector w, and a node table composed of all the task nodes constitutes a V×V matrix M; 4.2) initializing two V×d matrices W and W', wherein d is a space size after coding the node, W is a node vector matrix when the task node is taken as a center node, and W' is the same; each column in W and W' corresponds to vector information of a coded node in each column of the matrix M, and will be updated in a subsequent optimization process, and is both a weight matrix and a finally output node behavior vector; 4.3) Use the center node v c For each node u in its context o Do the inner product, compute u o T v c ; 4.4) Using the softmax formula to map the node table, calculate the probability of each node The result gets the probability of a certain node and then converts it into one-hot form, and the output result is w; 4.5) Maximize P(w o |w c ), i.e. in some context, the one-hot vector of the input node is w c , maximize the probability that the output is w o ; 4.6) performing back propagation, updating the two weight matrices W and W', and taking the center node vector matrix W as the behavior vector of the node after optimization.

Images