A production logistics resource allocation method based on dynamic spatiotemporal knowledge graph

Abstract

The present invention discloses a production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph, comprising the following steps: S01, investigating and analyzing data generated by large-scale Internet of Things devices by a neural network method; S02, establishing a dynamic spatiotemporal knowledge graph model according to the relationship between tasks in production logistics in the data, considering the Euclidean distance and time window between each point, and establishing a weighted directed graph; S03, allocating resources in the weighted directed graph by Dijkstra's graph algorithm. The production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph provided by the present invention can reduce the average duration of each task in production logistics and the average driving distance of vehicles, and solve the problems of spatial disorder and temporal asynchrony of resources in production logistics.

Description

Technical Field

[0001] The present invention relates to a production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph, and belongs to the technical field of modern logistics. Background Art

[0002] At present, the research and application scope of Industry 4.0 is being upgraded from "automatic execution of some production stages" to "intelligent operation of the entire production process". In some factories, no matter how intelligent they are, when materials enter the production line, the efficiency improvement is already very limited because the operation plan and control parameters are basically fixed. Production logistics refers to the logistics activities that occur in production. It is an important process that connects multiple production activities before and after. It usually occupies 95% of the production time and determines the efficiency of the entire production process. Because the current customer demand for products has gradually shifted from "meeting basic functional needs" to "pursuing personalized needs", production logistics has become complicated, and there are more and more uncertain factors in the production process. The resource allocation of production logistics is also greatly affected by the large working area, frequent resource interaction, long operation cycle and high staff participation.

[0003] Production logistics resources exhibit the following two characteristics during the operation and execution of production logistics: (1) Spatial disorder: Production logistics resources frequently enter and exit different production storage buffers according to the task requirements of production logistics. Due to high mobility and randomness, production logistics resources are usually difficult to locate. (2) Temporal asynchrony: Before production, the logistics buffer needs to wait for all materials and personnel to meet the requirements before production can begin. Excessive waiting time will not only reduce throughput, but also increase energy consumption. After production, due to the lack of timely and feasible allocation of production logistics resources, many work-in-process buffers have to be established to hedge time by increasing space, but this will lead to chaos in the allocation of production logistics resources and frequent accidents.

[0004] In order to effectively allocate production logistics resources, existing research on production logistics resource allocation has made some breakthroughs, but there are still some problems. Most studies consider the changes in time demand caused by production fluctuations, establish mathematical scheduling models for resource allocation, and solve the minimum distribution cost. Such methods can reduce costs to a certain extent, but they only focus on resource management in the production stage. The uncertainty of production logistics resources in the geographical and temporal dimensions should also be considered in the logistics stage.

[0005] For the manufacturing industry, most production logistics operations are completed in indoor and outdoor environments. In recent years, researchers have used the Internet of Things to sense and transmit data in physical space as virtual copies and model manufacturing resources into knowledge graphs. Some point out that the digital world and the real world are coupled through indoor positioning systems and real-time positioning systems are integrated into digital twins to predict and monitor production status, but most of these studies discuss specific objects and work tasks, and few studies consider the spatiotemporal consistency of global physical and virtual entities. For the problem of demand allocation in production logistics, most studies have focused on the construction method of knowledge graphs and have made great progress, but there is limited research on the use of graph algorithms for decision-making in reproduction logistics operations. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph, which can reduce the average duration of each task in production logistics and the average driving distance of vehicles, and solve the problems of spatial disorder and temporal asynchrony of resources in production logistics.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] A production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph comprises the following steps:

[0009] S01, investigate and analyze data generated by large-scale IoT devices through neural network methods;

[0010] S02, establish a dynamic spatiotemporal knowledge graph model based on the relationship between tasks in production logistics in the data, consider the Euclidean distance and time window between each point, and establish a weighted directed graph;

[0011] S03, allocating resources in the weighted directed graph by using Dijkstra's graph algorithm.

[0012] S01 specifically includes the following steps:

[0013] S011, configure Bluetooth tags on movable resources of production logistics, which include vehicles, materials and personnel, and deploy microcomputers embedded with Bluetooth communication modules as gateways at fixed locations;

[0014] S012, training deep neural networks using the concept of position fingerprinting;

[0015] S013, update the parameters in the gradient descent algorithm using the back propagation algorithm.

[0016] In S012, in the offline phase, RSSI values ​​of different reference points are collected to create a network map; in the online phase, the real-time measurement value is matched with the most similar signal pattern in the network map according to the Euclidean distance.

[0017] S012 The specific steps are:

[0018] use represents the RSSI value collected from the kth gateway of the i-th location unit. [l] Represents the neurons in the lth hidden layer, and establishes a triplet (r i , S i , T i ) as the training model, where S i is the semantic space information of the i-th location unit, T i is the timestamp of the RSSI value collected in the ith location unit, r i represents the set of RSSI values ​​of the gateway of the i-th location unit;

[0019] Use formula (1) to vectorize the neurons in the hidden layer.

[0020] A [l] =g [l] (W [l] A [l-1] )+b [l] (1)

[0021] Where W [l] is the weight matrix, b [l] is the bias vector of the lth layer, g [l] represents the sigmoid or ReLU activation function of the lth layer, A [l-1] Represents the output of the previous layer;

[0022] The softmax function is used as the output layer, and vectorization is performed as shown in formula (2). The calculation results of the output layer are shown in formulas (3), (4) and (5).

[0023]

[0024] Z [L] =W [L] A [L-1] +b [L] (3)

[0025]

[0026]

[0027] Where Z [L]represents the linear value of the L layer, and the softmax activation function is recorded as T, y i represents the position estimation probability in location unit i, C represents the number of location units, Represents the set of all output values;

[0028] The first two elements in all triplets, namely the RSSI value and the corresponding location unit index, are used as the total training set (X, Y) for supervised learning, and the classification cross entropy loss function of softmax is defined as formula (6),

[0029]

[0030] where h i Represents the predicted value The i-th true label of , s represents the number of triplets, and C represents the number of position units.

[0031] In S013, the back propagation algorithm updates the gradient descent algorithm in the following specific steps: using training parameters W and training parameters b, α represents the learning rate, and performing β iterations, forward propagation from layer 1 to L: A [l] =g [l] (W [l] A [l-1] )+b [l] ;

[0032] Calculate the cross entropy loss function:

[0033] Calculate the gradient of the loss function:

[0034] Update parameters: W = W-α·dW, b = b-α·db;

[0035] After the iteration, the updated W and b are returned.

[0036] S02 The specific steps are as follows: Construct a dynamic spatiotemporal knowledge graph DSTKG (V, E) to represent the relationship between production logistics resources, where V represents the set of vertices, E represents the set of weights, and a directed edge can represent the distance from the site to the last execution unit, with v i represents the i-th vertex, i<=k, k represents the number of vertices, d mn Indicates point v m With v n The distance, e mn Indicates point v m To point v n The weight of the edge, x mn Indicates point v m to v n The edge of is used, if v m to vn If the edge is used, it is represented as 1, otherwise it is 0. The objective function of transportation cost is constructed as shown in formula (7):

[0037]

[0038] Where m and N represent the sequence numbers of the points.

[0039] Beneficial effects of the present invention: The present invention provides a production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph. The production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph helps to find the shortest path for production logistics problems, so that the average execution time of each task and the driving distance of the vehicle are significantly reduced, thereby greatly reducing the cost of resource allocation in production logistics and solving the problems of spatial disorder and temporal asynchrony of resources in production logistics. DETAILED DESCRIPTION

[0040] The present invention is further described below. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are not intended to limit the protection scope of the present invention.

[0041] In order to solve the practical problems encountered in resource allocation of production logistics, the present invention provides a production logistics resource allocation method based on dynamic spatiotemporal knowledge graph, which mainly includes the following steps:

[0042] Step 1: Investigate and analyze the data generated by large-scale IoT devices through the method of neural network. The device in the present invention includes a low-power IoT Bluetooth tag, a gateway and a cloud server, wherein the gateway is a microcomputer embedded with a Bluetooth communication module. The low-power IoT Bluetooth tag is set on the movable resources of the production logistics, which include vehicles, materials and personnel, and the gateway is deployed in a fixed position. The gateway receives the signal broadcast by the BLE tag. In order to make full use of the edge computing capability, the gateway filters and compresses the noise data through the signal processing method, and then sends it to the cloud server.

[0043] Then the concept of location fingerprinting is used to train the deep neural network. In the offline stage, the RSSI values ​​of different reference points are collected to create a network map; in the online stage, the real-time measurement value is matched with the most similar signal pattern in the network map according to the Euclidean distance. RSSI is an indication of the strength of the received signal, and its implementation is performed after the reverse channel baseband receiving filter. Compared with TOA, TDOA and AOA, the use of RSSI is low-cost and easy to implement, and no other devices need to be deployed. Therefore, the present invention uses RSSI value as the ranging standard.

[0044] use represents the RSSI value collected from the kth gateway of the i-th location unit. [l]Represents the neurons in the lth hidden layer, and establishes a triplet (r i , S i , T i ) as the training model, where S i is the semantic space information of the i-th location unit, T i is the timestamp of the RSSI value collected in the ith location unit, r i represents the set of RSSI values ​​of the gateway of the i-th location unit;

[0045] Use formula (1) to vectorize the neurons in the hidden layer.

[0046] A [l] =g [l] (W [l] A [l-1] )+b [l] (1)

[0047] Where W [l] is the weight matrix, b [l] is the bias vector of the lth layer, g [l] represents the sigmoid or ReLU activation function of the lth layer, A [l-1] Represents the output of the previous layer;

[0048] The softmax function is used as the output layer, and vectorization is performed as shown in formula (2). The calculation results of the output layer are shown in formulas (3), (4) and (5).

[0049]

[0050] Z [L] =W [L] A [L-1] +b [L] (3)

[0051]

[0052]

[0053] Where Z [L] represents the linear value of the L layer, and the softmax activation function is recorded as T, y i represents the position estimation probability in location unit i, C represents the number of location units, Represents the set of all output values;

[0054] The first two elements in all triplets, namely the RSSI value and the corresponding location unit index, are used as the total training set (X, Y) for supervised learning, and the classification cross entropy loss function of softmax is defined as formula (6),

[0055]

[0056] where h i Represents the predicted value The i-th true label of , s represents the number of triplets, and C represents the number of position units.

[0057] Use the back propagation algorithm to update the parameters in the gradient descent algorithm. The specific steps of the back propagation algorithm to update the gradient descent algorithm are as follows: Use training parameters W and training parameters b, α represents the learning rate, perform β iterations, and forward propagate from layer 1 to L: A [l] =g [l] (W [l] A [l-1] )+ b [l] ;

[0058] Calculate the cross entropy loss function:

[0059] Calculate the gradient of the loss function:

[0060] Update parameters: W = W-α·dW, b = b-α·db;

[0061] After the iteration, the updated W and b are returned.

[0062] Step 2: Based on the relationship between tasks in production logistics in the data, a dynamic spatiotemporal knowledge graph model is established, and a weighted directed graph is established by considering the Euclidean distance and time window between each point. The specific steps are as follows: Use the parameters trained in the above steps to predict the classification problem, and construct a dynamic spatiotemporal knowledge graph DSTKG (V, E) to represent the relationship between production logistics resources, where V represents the set of vertices, E represents the set of weights, and the directed edge can represent the distance from the site to the last execution unit, with v i represents the i-th vertex, i<=k, k represents the number of vertices, d mn Indicates point v m With v n The distance, e mn Indicates point v m To point v n The weight of the edge, x mn Indicates point v m to v n The edge of is used, if v m to v n If the edge is used, it is represented as 1, otherwise it is 0. The objective function of the transportation cost is constructed as shown in formula (7). From formula (7), we can see that if we want to find the minimum transportation cost, we only need to find the weighted shortest path of the weighted directed graph.

[0063]

[0064] Where m and N represent the sequence numbers of the points.

[0065] Step 3: Allocate resources in the weighted directed graph using Dijkstra's graph algorithm. For solving the weighted shortest path, since Dijkstra's graph algorithm can traverse all points, the present invention uses Dijkstra to calculate the single-source shortest path. The path sought represents the solution with the lowest cost, which is the optimal resource allocation method.

[0066] The present invention adopts a production logistics resource allocation method based on a dynamic spatiotemporal knowledge graph to help find the shortest path for production logistics problems, significantly reducing the average execution time of each task and the driving distance of the vehicle, thereby greatly reducing the cost of resource allocation in production logistics.

[0067] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A production logistics resource allocation method based on dynamic spatiotemporal knowledge graph, characterized by: The following steps are involved: S01, investigate and analyze the data generated by large-scale IoT devices through neural network methods, specifically including the following steps: S011, configure Bluetooth tags on movable resources of production logistics, which include vehicles, materials and personnel, and deploy microcomputers embedded with Bluetooth communication modules as gateways at fixed locations; S012 uses the concept of location fingerprinting to train a deep neural network. In the offline stage, RSSI values ​​of different reference points are collected to create a network map. In the online stage, the real-time measurement values ​​are matched with the most similar signal pattern in the network map according to the Euclidean distance. The specific steps are as follows: i k Denotes the RSSI value collected from the kth gateway of the i-th location unit, with a [l] Represents the neurons in the lth hidden layer, and establishes a triplet (r i ,S i ,T i ) as the training model, where S i is the semantic space information of the i-th location unit, T i is the timestamp of the RSSI value collected in the ith location unit, r i represents the set of RSSI values ​​of the gateway of the i-th location unit; Use formula (1) to vectorize the neurons in the hidden layer. A [l] =g [l] (W [l] A [l-1] )+b [l] (1) Where W [l] is the weight matrix, b [l] is the bias vector of the lth layer, g [l] represents the sigmoid or ReLU activation function of the lth layer, A [l-1] Represents the output of the previous layer; The softmax function is used as the output layer, and vectorization is performed as shown in formula (2). The calculation results of the output layer are shown in formula (3), formula (4) and formula (5). Z [L] =W [L] A [L-1] +b [L] (3) Where Z [L] represents the linear value of the L layer, and the softmax activation function is recorded as T, y i represents the position estimation probability in location unit i, C represents the number of location units, Represents the set of all output values; The first two elements in all triplets, namely the RSSI value and the corresponding location unit index, are used as the total training set (X, Y) for supervised learning, and the classification cross entropy loss function of softmax is defined as formula (6), where h i Represents the predicted value The i-th true label of , s represents the number of triplets, and C represents the number of position units; S013, update the parameters in the gradient descent algorithm using the back propagation algorithm; S02, based on the relationship between tasks in production logistics in the data, a dynamic spatiotemporal knowledge graph model is established. Considering the Euclidean distance and time window between each point, a weighted directed graph is established to construct a dynamic spatiotemporal knowledge graph DSTKG (V, E) to represent the relationship between production logistics resources, where V represents the set of vertices, E represents the set of weights, and the directed edge represents the distance from the site to the last execution unit, with v as the i represents the i-th vertex, i<=k, k represents the number of vertices, d mn Indicates point v m With v n The distance, e mn Indicates point v m To point v n Edge weight, x mn Indicates point v m to v n The edge of is used, if v m to v n If the edge is used, it is represented as 1, otherwise it is 0. The objective function of transportation cost is constructed as shown in formula (7): Where m and n represent the serial numbers of the points; S03, allocating resources in the weighted directed graph by using Dijkstra's graph algorithm.

2. According to claim 1, a production logistics resource allocation method based on dynamic spatiotemporal knowledge graph is characterized by: In S013, the back propagation algorithm updates the gradient descent algorithm in the following specific steps: using training parameters W and training parameters b, α represents the learning rate, and performing β iterations, forward propagation from layer 1 to L: A [l] =g [l] (W [l] A [l-1] )+b [l] ; Calculate the cross entropy loss function: Calculate the gradient of the loss function: Update parameters: W = W-α·dW, b = b-α·db; After the iteration, the updated W and b are returned.

Images