JP7383262B2 - Box selection device, box selection method, and box selection program - Google Patents

Description

The present invention relates to a box selection device, a box selection method, and a box selection program.

BACKGROUND ART Conventionally, a box selection device as described in, for example, Patent Document 1 is known. The box selection device selects a combination of cargo and container based on data such as the volume of the cargo. This box selection device selects all combinations of containers in which the total capacity of the containers is greater than or equal to the total volume of cargo.

Japanese Patent Application Publication No. 2016-88630

The box selection device described above selects all combinations of containers (boxes) that satisfy the conditions. However, when such a method is adopted, as the number of types of boxes increases, the number of combinations increases rapidly, and it becomes necessary to calculate whether all items can be loaded for each combination. In this case, there is a problem in that it becomes difficult to employ this method from the viewpoint of calculation costs and the like. Furthermore, such calculation costs also increase as the number of types of articles increases. That is, there is a problem in that it is difficult to apply the above-described box selection device to a site where there are many types of articles or many options for boxes. On the other hand, if an attempt is made to select boxes while reducing calculation costs, a problem arises in that the accuracy of selection decreases.

An object of the present invention is to provide a box selection device, a box selection method, and a box selection program that can accurately select a suitable box when selecting a box for packaging an article.

A box selection device according to one aspect of the present invention is a box selection device that selects a box for packaging an article based on learning information learned in advance, and a control unit of the box selection device controls the learning information and the box information. a storage unit for storing information, an input information reception unit for creating input information regarding the article, and a neural network calculation unit for inputting the input information into the input layer and calculating an evaluation value regarding the box by neural network calculation based on the learning information. , a selection unit that selects a box based on the evaluation value output from the output layer of the neural network calculation unit, and the neural network calculation unit includes an embedding layer that acquires feature quantities regarding the article from input information, and an intermediate layer that evaluates the feature amount acquired in the layer.

In the box selection device, the neural network calculation section inputs input information regarding the article into the input layer, and calculates an evaluation value regarding the box by neural network calculation based on the learning information. Then, the selection section can select a box suitable for the article to be packed, based on the evaluation value output from the output layer of the neural network calculation section. Here, when there are many types of items to be packed, if the neural network calculation unit tries to output evaluation values by treating each item as independent, learning will be insufficient and the accuracy of the evaluation values will decrease. There is a possibility that it will decrease. On the other hand, even if the products are of different types, they may have a high degree of similarity when focusing on a certain feature. That is, if the neural network calculation unit treats highly similar items equally as similar items, it is possible to suppress a decrease in the evaluation value even when the number of types of items increases. Therefore, the neural network calculation unit includes an embedding layer that acquires feature amounts related to the article from input information, and an intermediate layer that evaluates the feature amounts acquired by the embedding layer. In this way, by having the embedded layer, the neural network calculation unit can output an evaluation value based on the feature amount of the article. In other words, based on the feature values of the items, the neural network calculation unit can treat items with high similarity equally as the same items, substantially reduce the number of types of items, and output evaluation values. can. Therefore, when selecting a box for packaging an article, the box selection device can select a suitable box with high accuracy.

The neural network calculation unit may have embedding vector information corresponding to information included in the input information, and the embedding layer may convert the input information into an embedding vector based on the input information and the embedding vector information. Thereby, the neural network calculation unit can appropriately output the evaluation value based on the feature amount as described above.

The output layer may output an evaluation value for each of the plurality of boxes in the box information stored in the storage section, and the selection section may select the box with the maximum evaluation value. In this case, the selection section can quickly select any of the boxes stored in the storage section based on the evaluation value.

The neural network calculation unit may perform learning using training data based on packing results and obtain embedding vectors in the embedding layer. This makes it possible to obtain an embedding vector that can output highly accurate evaluation values based on actual packaging results.

The neural network calculation unit calculates the discrepancy between the evaluation value output from the output layer and the correct value of the teaching data using a loss function, and uses the result obtained by the calculation to generate an embedding vector by error backpropagation. You may obtain it. Thereby, the neural network calculation unit can appropriately learn the teacher data.

A box selection method according to one aspect of the present invention is a box selection method for selecting a box for packaging an article based on learning information learned in advance, and includes a storage step of storing learning information and box information; an input information reception step that creates input information; a neural network calculation step that inputs the input information into the input layer and calculates an evaluation value for the box by neural network calculation based on the learning information; and an output layer in the neural network calculation step. a selection step of selecting a box based on the evaluation value output from the neural network calculation step, a calculation by an embedding layer that calculates a feature amount regarding the article from input information, and a calculation result from the calculation result by the embedding layer. A computation is performed by an intermediate layer that evaluates the feature amount.

A box selection program according to one aspect of the present invention is a box selection program that selects a box for packaging an article based on learning information learned in advance, and includes a storage step of storing learning information and box information, and a storage step related to the article. an input information reception step that creates input information; a neural network calculation step that inputs the input information into the input layer and calculates an evaluation value for the box by neural network calculation based on the learning information; and an output layer in the neural network calculation step. The computer system executes the selection step of selecting a box based on the evaluation value output from the computer system, and the neural network calculation step includes calculations by an embedding layer that calculates feature quantities related to the article from input information, and calculations by the embedding layer. Based on the results, calculations are performed by an intermediate layer that evaluates the feature quantities.

According to these box selection methods and box selection programs, it is possible to obtain the same functions and effects as those of the box selection device described above.

According to the present invention, there are provided a box selection device, a box selection method, and a box selection program that can accurately select a suitable box when selecting a box for packaging an article.

1 is a schematic configuration diagram showing a box selection device according to an embodiment of the present invention. FIG. 2 is a schematic diagram showing a processing image when the box selection device selects a transport box using a neural network. FIG. 7 is a schematic diagram showing a processing image when the neural network calculation unit of the box selection device performs learning using past packing results as training data. It is a schematic diagram which shows the specific example of the processing content of a box selection apparatus. FIG. 3 is a schematic diagram showing a specific example of vector transformation using an embedding layer. FIG. 3 is a diagram for explaining embedding vectors. FIG. 7 is a schematic diagram showing a specific example of processing contents during learning of the box selection device. It is a flowchart which shows the content of the control process by the control part of a box selection apparatus. FIG. 6 is a diagram illustrating an example of matching processing by an input information receiving unit.

Embodiments of the present invention will be described in detail below with reference to the drawings. In addition, in the drawings, the same or equivalent elements are given the same reference numerals, and overlapping explanations will be omitted.

FIG. 1 is a schematic configuration diagram showing a box selection device 100 according to an embodiment of the present invention. The box selection device 100 is a device that selects in which box a plurality of items should be packed in a box at a distribution site such as a warehouse. Note that after a box is selected, the position and order in which the articles are to be packed in the box is automatically determined by another method. In this embodiment, parts are exemplified as articles to be packed, and a transportation box is exemplified as a box. However, the types of articles and boxes are not particularly limited.

As shown in FIG. 1, the box selection device 100 includes an information input section 2, an information output section 3, and a control section 10. The box selection device 100 is a device that selects a transportation box for packing parts based on learning information learned in advance. Note that FIGS. 2 and 3 may be referred to in order to explain each component of the box selection device 100. FIG. 2 is a schematic diagram showing a processing image when the box selection device 100 selects a transport box using a neural network. FIG. 2 is a schematic diagram showing a processing image when the box selection device 100 selects a transport box using information learned by a neural network. FIG. 3 is a schematic diagram showing specific details when the neural network calculation unit 12 of the box selection device 100 performs learning using past packing results as training data.

The information input unit 2 is an interface for inputting various information. For example, the information input unit 2 includes a keyboard, a mouse, a touch panel, and the like. Alternatively, information may be input to the information input unit 2 via a storage medium or communication. The information output unit 3 is an interface that outputs various information. For example, the information output unit 3 may output information using a monitor, a speaker, or the like. Alternatively, the information output unit 3 may output information via a storage medium or communication.

The control unit 10 is composed of a CPU, RAM, ROM, etc. The control unit 10 includes an input information reception unit 11, a neural network calculation unit 12, a selection unit 13, and a storage unit 14. The control unit 10 may be configured as one or more processors that operate according to a program, one or more dedicated hardware circuits such as ASIC, or a circuit that includes a combination thereof. The processor has a CPU and memories such as RAM and ROM. The memory stores various programs for processing information, and the CPU reads out the various programs and performs calculations.

The input information receiving unit 11 creates input information regarding parts. The input information receiving unit 11 acquires the information input by the information input unit 2 and creates input information using the information. The input information receiving unit 11 acquires identification information associated with each of a plurality of parts, information regarding the number of parts, and the like. The input information is information about parts that has been adjusted to facilitate neural network calculations. As shown in FIG. 2, the input information receiving unit 11 obtains a parts list of parts to be packed, that is, a list indicating which parts and how many parts are to be packed, in a pre-packing stage. Further, as shown in FIG. 3, during learning, the input information receiving unit 11 acquires a parts list of actually packed parts and information about which shipping box was used for packaging. Note that the specific content of the identification information linked to the component and the specific content of the input information will be described later.

The neural network calculation unit 12 calculates an evaluation value regarding the transport box by neural network calculation based on learning information. The neural network calculation unit 12 acquires learning information by performing learning using teacher data based on past packing results (see FIG. 3). In addition, the neural network calculation unit 12 uses the learning information to calculate evaluation values for selecting which shipping box is suitable for the list of parts to be packed at the actual work site (Figure 2 reference).

The neural network calculation unit 12 includes a learning unit 16. The learning unit 16 generates a machine learning model as learning information. In the learning stage (see FIG. 3), the learning unit 16 generates the best neural network that is estimated to have the highest prediction accuracy as a model. The learning unit 16 reads a teacher data set (samples of past packing results) from the storage unit 14 or the like, and executes machine learning while sequentially inputting each sample to a neural network for learning. Note that such a neural network can be configured by, for example, an arithmetic unit, a storage unit, etc. modeled after a neuron model. For example, the learning unit 16 may generate the model by performing deep learning using a multilayer neural network. The type of machine learning is not limited to deep learning, and the learning unit 16 may generate the model using other methods. Note that the specific configuration and processing contents of the neural network calculation unit 12 will be described later.

The selection unit 13 selects a transport box based on the evaluation value output from the output layer of the neural network calculation unit 12. The selection unit 13 acquires the evaluation value and uses the evaluation value to select the most suitable transport box for packaging the parts. The selection unit 13 selects the most suitable transport box from the transport box list (box information) stored in the storage unit 14 in advance. Note that the specific processing contents of the selection unit 13 will be described later.

The storage unit 14 is a part of the box selection device 100 that stores various information. The storage unit 14 stores at least learning information (neural network model) and box information. The box information is information regarding a plurality of transport boxes used at the work site, and is stored in advance in the storage unit 14 before the box selection work. The storage unit 14 stores box information in the form of a shipping box list (see FIG. 2) that includes information such as the size and shape of each of a plurality of shipping boxes.

Next, a specific example of the processing of the neural network calculation unit 12 will be described with reference to FIGS. 2 and 3. First, with reference to FIG. 3, a case will be described in which the neural network calculation unit 12 of the box selection device 100 performs learning using past packing results as training data. Here, as a packing record, "transport box Z" is used when the parts list is "2 pieces of part a," "1 piece of part b," "2 pieces of part c," and "1 piece of part d." This data is used. Note that the parts list in this case may be referred to as "parts list PL1." In addition, the shipping boxes that can be used for packaging are selected from a shipping box list in which six types of shipping boxes (see Figure 2), ``Transporting boxes U, V, W, X, Y, and Z'' are registered. shall be carried out. The transport box list may be referred to as "transport box list BL1."

As shown in FIG. 3, the neural network calculation unit 12 extracts a packed parts list PL1 from the packing record data, and confirms that "transport box Z" has been selected as the packed transport box for the parts list. Extract. The learning unit 16 (see FIG. 1) of the neural network calculation unit 12 operates so that when the parts list PL1 is input to the model M of the neural network, an evaluation value that selects the transport box Z is output. Update model M as appropriate. The neural network calculation unit 12 performs learning using training data based on a plurality of types of parts lists and a large number of packing records related to transport boxes selected for the parts list.

Next, with reference to FIG. 2, a case will be described in which the box selection device 100 selects a transport box. As shown in FIG. 2, the box selection device 100 acquires a transport box list BL1 in advance. Then, when selecting a transport box, the box selection device 100 acquires the above-mentioned parts list PL1 as information indicating the parts to be packed. The neural network calculation unit 12 calculates an evaluation value regarding the transport box using the neural network model M updated through learning shown in FIG. As a result of learning, the neural network calculation unit 12 outputs an evaluation value that makes it easier for transport box Z to be selected from transport box list BL1. Therefore, the selection unit 13 selects the transport box Z based on the evaluation value. The box selection device 100 outputs that transport box Z has been selected.

Next, with reference to FIG. 4, the detailed configuration of the neural network calculation unit 12 and its processing contents will be described. FIG. 4 is a schematic diagram showing a specific example of processing contents of the box selection device 100. First, the process when the input information receiving unit 11 creates input information will be described. In this embodiment, the input information is configured by a combination of product number IDs. That is, as part identification information, a unique serial number is assigned according to the type of part. The input information receiving unit 11 creates input information by arranging product number IDs corresponding to the parts in the parts list. Specifically, part number ID "1" is set for part a, part number ID "2" is set for part b, and part number ID "4" is set for part d. Therefore, the input information receiving unit 11 arranges product number IDs such as "1, 1, 2...4" in order from the first line. The input information reception unit 11 inputs the created input information to the neural network calculation unit 12.

Here, the input information receiving unit 11 may fix the number of parts handled as input information. Note that in this case, the number of nodes ND in the input layer 21 is fixed to a predetermined number. If the number of parts is insufficient, the input information receiving unit 11 may fill in the product number IDs corresponding to the missing parts with predetermined temporary IDs as a matching process. Furthermore, if the number of parts exceeds the number, the input information receiving unit 11 may discard the product number IDs corresponding to the excess parts as a matching process. Specifically, as shown in FIG. 9A, it is assumed that the number of parts (N) handled as input information is fixed at 6 (N=6). Here, the list of parts to be packed includes five parts, which is less than six. In this case, the input information receiving unit 11 fills in the product number ID corresponding to the missing part (sixth part) with a temporary ID of 0 as a matching process. Further, as shown in FIG. 9(b), it is assumed that the number of parts (N) handled as input information is fixed at 4 (N=4). Here, the list of parts to be packed includes five parts, which are more than four. In this case, the input information receiving unit 11 discards the product number ID of the excess component (fifth component h) as a matching process.

Note that the input information receiving unit 11 may organize the input order of the identification information of each component to the node ND of the input layer 21 according to a predetermined rule. For example, the input order of product numbers IDs, etc. of parts to be packed may be arranged in ascending order or descending order of product number IDs, etc. For example, the input information receiving unit 11 in FIG. 4 is arranged in the order of product number ID, "1 (part a), 1 (part a), 2 (part b)...4 (part d)". For example, if the list of parts to be packed is in the order "part b, part d, part a...part a", the order of product number IDs is "2 (part b), 4 (part d), 1 (part a)...1 (part a)". In such a case, the input information receiving unit 11 can arrange product number IDs as shown in FIG. 4 in descending order of number and input them to each node ND of the input layer 21.

The neural network calculation unit 12 has an input layer 21, an embedding layer 22, an intermediate layer 30, and an output layer 23 as a neural network model. The input layer 21 is a layer for inputting parameters for performing neural network calculations. The embedded layer 22 and the intermediate layer 30 are layers that perform predetermined calculations on input parameters. The output layer 23 is a layer that outputs evaluation values. Each layer 21, 22, 30, 23 has nodes ND corresponding to the number of input parameters. For example, the input layer 21 outputs an input vector x=(x ₁ , x ₂ , . . . x _m ) having m parameters as components to the embedding layer 22 as is. The embedding layer 22 and the intermediate layer 30 convert the total input into an output using the weight and activation function of each layer, and pass the output to the output layer 23. The output layer 23 also converts the total input into an output using the weight and activation function of each layer. This output is an output vector y=(y ₁ , y ₂ , . . . , y _n ) of the neural network whose components are n parameters. This output vector y indicates the evaluation value. In _addition , during learning ₍ see Figure 3), if _the correct answer of the output vector for one input vector The weight w of each layer is updated. This is information processing called learning. Note that the details of the processing during learning will be described later.

In this embodiment, a case will be described in which the number of nodes ND in the input layer 21 is set to be the same as the number of parts included in the input information. A vector listing product number IDs is input to the input layer 21 .

The embedding layer 22 is a layer that acquires feature amounts related to parts from input information. The neural network calculation unit 12 has embedded vector information corresponding to information included in the input information (one-hot vector described later). Then, the embedding layer 22 converts the input information into an embedding vector based on the input information and embedding vector information.

The processing contents of the embedding layer 22 will be explained in detail with reference to FIG. FIG. 5 is a schematic diagram showing a specific example of vector conversion by the embedding layer 22. As shown in FIG. 5, the input layer 21 inputs to the embedding layer 22 an input vector having a parameter (for example, a product number ID as described above) for identifying a component in the component list. Here, the input vector is expressed as a one-dimensional vector having the number of rows (hereinafter sometimes referred to as N rows) corresponding to the number of parts in the parts list.

The embedding layer 22 converts the input vector from the input layer 21 into a one-hot vector as preprocessing for embedding vector conversion. The one-hot vector is a vector in which one component is set as "1" and all remaining components are simply shown as "0" in order to identify the type of part. The One-hot vector is a vector with N rows corresponding to the number of parts in the parts list. Note that, for example, the first and second rows are vectors corresponding to part a. The vectors in each row correspond to each part of the input information. On the other hand, the number of dimensions (number of columns) of the One-hot vector is the same as the number of parts types of all parts handled at the work site. Therefore, when thousands to hundreds of thousands of types of parts are handled at a work site, the number of dimensions of the One-hot vector is from several thousand to several hundred thousand.

The embedding layer 22 converts the one-hot vector into an embedding vector by extracting embedding vector information of each component using the one-hot vector. Note that the detailed processing method will be described later.

Here, the embedding vector will be explained with reference to FIG. 6. FIG. 6 is a diagram for explaining embedding vectors. As shown in FIG. 6(a), parts having various shapes and shapes are handled at the packaging site. In this case, even if each part exists independently, the characteristics of each part are unknown. For example, parts c and h are close to rectangular in shape, but their sizes are different, and it is unclear whether they can be considered similar parts. Furthermore, when each part is handled independently, it is impossible to determine whether they are similar without additional information such as the shape of each part. On the other hand, as shown in FIG. 6(b), the embedding vector can indicate the similarity between parts based on the relationship between the values of the feature amounts. For example, it is assumed that the feature amount 1 of the first dimension of the embedding vector (the first column of the embedding vector shown in FIG. 5) indicates the size. The feature quantity 2 of the second dimension of the embedding vector (the second column of the embedding vector shown in FIG. 5) indicates the degree of angularity. FIG. 6(b) is a graph in which the values of each component are plotted with the feature amount 1 as the horizontal axis and the feature amount 2 as the vertical axis. Since parts a and e are both rounded as a whole, the value of feature amount 1 is "0.1". Part h has a rectangular shape as a whole, but has rounded corners, so the value of feature quantity 1 is "0.5". Part c is entirely rectangular and has no rounded corners, so the value of feature quantity 1 is "0.9". Since part a is small, the value of feature quantity 2 is "0.2". Since parts e and h have an intermediate size, the value of feature quantity 2 is "0.5". Since component c is large, the value of feature quantity 2 is "0.9". From this relationship, it can be said that parts a and e are close to each other and are similar parts. On the other hand, parts a and c are far apart from each other and can be said to be different parts. In other words, part a can be handled as a part having a higher degree of similarity to part e than to part c.

In the example shown in FIG. 5, the embedding vector output by the embedding layer 22 is shown in the form of a matrix written on the rightmost side. Information on the feature amount of part a is shown in the first and second lines. The feature amount information in each row corresponds to each part of the input information. Hereinafter, when it is simply referred to as an "embedding vector", it means a matrix that the embedding layer 22 outputs to the intermediate layer 30. Note that, as shown in the matrix of embedding vectors in FIG. 5, the embedding vector has items indicating features of the component in addition to the above-mentioned feature quantities 1 and 2. The number of dimensions of the embedding vector, that is, the number of columns, is fixed to the number of feature items. When the types of parts are different, the combinations of feature values of each item are expressed in different forms while the number of dimensions is fixed. Therefore, even when the number of types of parts becomes enormous, the embedding vectors can represent each part using about several hundred fixed-dimensional vectors.

The neural network calculation unit 12 has embedding vector information for converting input information into an embedding vector as described above. The embedded vector information is a set of values for each feature quantity item in the parts handled at the work site. Here, in FIG. 5, the matrix indicated as "embedding vector information of all types of parts" has the embedding vector information of component a in the first column, the embedding vector information of component b in the second column, and the embedding vector information of component b in the third column. It has embedding vector information for component c, embedding vector information for component d in the fourth column, and embedding information for component e in the fifth column. In this way, the values for each feature quantity item in part X are arranged (here, in a line) and are defined as "embedded vector information of part X." Then, by arranging the "embedded vector information of part Such a matrix is defined as "embedded vector information for all types of parts."

The embedding layer 22 multiplies the One-hot vector by "embedding vector information for all types of parts." Thereby, the embedding layer 22 converts the one-hot vector into an embedding vector by extracting "component embedding vector information" for each component included in the input information. For example, the first and second lines of the input vector indicate "part a", and the first line of the one-hot vector indicates "part a". Therefore, the embedding layer 22 extracts a column of values corresponding to "embedding vector information of component a" from among "embedding vector information of all types of components", and inserts the values into the first and second rows of the embedding vector. Convert it as a row with the value of the row. The third line of the input vector indicates "part c", and the third line of the one-hot vector indicates "part c". Therefore, the embedding layer 22 extracts "embedding vector information of component c" from "embedding vector information of all types of components" and converts the value into the third row of the embedding vector. As described above, the One-hot vector is converted into an embedded vector.

Returning to FIG. 4, the intermediate layer 30 will be explained. The intermediate layer 30 includes a convolution layer 31, a pooling layer 32, and a fully connected layer 33. The convolution layer 31 is a layer that evaluates the feature amount acquired by the embedding layer 22. That is, the embedding vectors include many feature quantities that do not need to be taken into consideration when selecting a transport box. Therefore, the convolution layer 31 applies feature quantities to the embedding vector and extracts the degree of response of each feature quantity as a score. The pooling layer 32 is a layer that converts the feature amount extracted by the convolution layer 31 into a one-dimensional vector and expresses it. That is, since the output layer 23 ultimately outputs the evaluation value as a one-dimensional vector, the pooling layer 32 uses the convolution layer 31 to convert the feature amount represented by a multidimensional vector into one dimension. Only the maximum value is extracted from the scores of each extracted feature quantity. Thereby, the pooling layer 32 extracts only the information that has a large influence on box selection. The fully connected layer 33 is a layer that aggregates feature amounts represented by one-dimensional vectors and calculates evaluation values for each transport box.

The number of nodes ND in the output layer 23 is set to be the same as the number of selectable shipping boxes, that is, the number of shipping boxes in the shipping box list BL1 (see FIG. 2). Each transport box is assigned to each node ND of the output layer 23, and the output value of the corresponding node ND is used as the evaluation value of each transport box.

The selection unit 13 uses the output value of the neural network calculation unit 12 as the evaluation value of each transport box. For example, the selection unit 13 acquires the output value for the transport box U as the evaluation value y1, the output value for the transport box V as the evaluation value y2, and the output value for the transport box Z as the evaluation value yn. Then, the selection unit 13 selects the transport box with the highest evaluation value.

Next, with reference to FIG. 7, learning by the neural network calculation unit 12 will be specifically described. The neural network calculation unit 12 performs learning using training data based on past packaging results, and acquires (acquires or updates) embedding vector information for each component in the embedding layer 22. The teacher data includes information on a list of parts packed in past packing operations and information on selected shipping boxes (see, for example, FIG. 3). Further, the teacher data includes evaluation values for each transport box in the transport box list as correct values. The correct value of the actually selected transport box is the maximum value, and the lower the similarity of the transport box to the selected transport box, the lower the correct value.

As shown in FIG. 7, during learning, a parts list included in the teacher data is input, and the output layer 23 outputs an evaluation value corresponding to the parts list. The neural network calculation unit 12 compares the evaluation value and the correct value for each transport box, and updates the embedding vector information of each component depending on the degree of deviation. Specifically, the neural network calculation unit 12 calculates the discrepancy between the evaluation value output from the output layer 23 and the correct value of the teaching data using a loss function, and uses the calculation result to perform embedding by error backpropagation. Get a vector. More specifically, the neural network calculation unit 12 uses the calculation results to rewrite the embedding vector information of each component forming the embedding vector. As a result, part or all of the "embedded vector information for all types of parts" is rewritten. Then, the neural network calculation unit 12 obtains an embedding vector by converting the One-hot vector using the rewritten "embedding vector information for all types of parts." As the loss function, for example, a cross entropy function shown in equation (1) below is used. Note that when the embedding vector is updated, the updated content is also reflected in the embedding vector information. As a result, the next time the neural network calculation unit 12 outputs an evaluation value, the rewritten embedding vector information is used to output the evaluation value.

Next, with reference to FIG. 8, a box selection method according to this embodiment will be described. FIG. 6 is a flowchart showing the details of the control process by the control unit 10 of the box selection device 100. The box selection method is a method of selecting a transport box for packing parts based on learning information learned in advance. Note that the detailed processing contents in each step are the same as those described in the explanation of each component of the box selection device 100, so the explanation will be omitted.

As shown in FIG. 8, a storage step S10 is executed in which learning information and transport box information are stored in the storage unit 14. Note that the storage step S10 may be executed in advance before the process of FIG. 8 is executed. Here, as described with reference to FIG. 3, a neural network model is generated or updated as learning information.

Next, the input information receiving unit 11 executes an input information receiving step S20 of obtaining information regarding the component and creating input information based on the obtained information. Next, the neural network calculation unit 12 inputs the input information created in the input information reception step S20 to the input layer 21, and calculates the evaluation value regarding the transport box by neural network calculation based on the learning information stored in the storage unit 14. A neural network calculation step S30 is executed to calculate . In the neural network calculation step S30, calculations are performed by the embedding layer 22 that calculates feature quantities related to the part from input information, and calculations are performed by the convolution layer 31 of the intermediate layer 30 that evaluates the feature quantities from the calculation results of the embedding layer 22. Next, the selection unit 13 executes a selection step S40 in which a transport box is selected based on the evaluation value output from the output layer 23 in the neural network calculation step S30. With the above, the control process shown in FIG. 8 is completed. Note that when selecting a transport box for another parts list, the process may be repeated from the input information receiving step S20.

Next, the functions and effects of the box selection device 100 according to this embodiment will be explained.

The box selection device 100 according to the present embodiment is a box selection device 100 that selects a box for packaging an article based on learning information learned in advance. The control unit 10 of the box selection device 100 includes a storage unit 14 that stores learning information and box information, an input information receiving unit 11 that creates input information regarding articles, and inputs input information into an input layer and performs processing based on the learning information. It has a neural network calculation unit 12 that calculates evaluation values for boxes by neural network calculation, and a selection unit 13 that selects boxes based on the evaluation values output from the output layer 23 of the neural network calculation unit 12. The neural network calculation unit 12 includes an embedding layer 22 that acquires feature quantities related to articles from input information, and an intermediate layer 30 that evaluates the feature quantities acquired by the embedding layer 22.

In the box selection device 100, the neural network calculation unit 12 inputs input information regarding the article to the input layer 21, and calculates the evaluation value regarding the box by neural network calculation based on the learning information. Then, the selection unit 13 can select a box suitable for the article to be packed, based on the evaluation value output from the output layer 23 of the neural network calculation unit 12.

Here, a box selection device according to a comparative example will be described. The box selection device according to the comparative example does not have the embedded layer 22 like the present embodiment. The box selection device according to the comparative example updates a neural network model using past packaging results as training data. The teacher data includes information on combinations of a plurality of articles and information on boxes selected for the combinations. Thereby, when selecting a box, the box selection device can select an appropriate box if the same combination of articles has been packed in the past.

The box selection device according to the comparative example treats each article as an independent item and outputs an evaluation value, reducing calculation costs and making it possible to easily select a suitable box. Here, we will explain the problem when the neural network calculation unit 12 attempts to treat each item as an independent item and output an evaluation value when there are many types of items to be packed, as in the comparative example. . For example, in a site where many types of items (tens of thousands of items) are handled, the number of combinations of items to be packed becomes enormous. In this case, even if an attempt is made to obtain training data based on past packaging results, appropriate learning may not be possible with the amount of data that can actually be collected. As a result, learning may become insufficient and the accuracy of the evaluation value may decrease. Furthermore, when the number of types of articles increases, a method can be considered in which similar articles are grouped together by collecting additional information such as shape and purpose and adding additional information to each article. However, as the number of types of articles increases, the process of collecting additional information becomes enormous, making it difficult to implement. That is, in the comparative example, shape information of parts, etc. is required for calculation, so it is necessary to collect shape information, etc. of all parts to be loaded.

On the other hand, even if the products are of different types, they may have a high degree of similarity when focusing on a certain feature. In other words, the neural network calculation unit 12 does not need to handle all items independently, and if items with high similarity are treated equally as the same item, the evaluation value will be improved even when the number of types of items increases. It is possible to suppress the decrease in Therefore, in the present embodiment, the neural network calculation unit 12 includes an embedding layer 22 that acquires feature amounts related to the article from input information, and an intermediate layer 30 that evaluates the feature amounts acquired by the embedding layer 22. In this manner, the neural network calculation unit 12 having the embedded layer 22 can output an evaluation value based on the feature amount of the article. That is, the neural network calculation unit 12 treats highly similar items equally as the same item based on the feature values of the items, substantially reduces the number of types of items, and outputs evaluation values. Can be done. Further, when using the embedding layer 22, it is possible to extract the similarity of each article from the packing record without the need to separately extract additional information. That is, in the box selection device 100 according to the present embodiment, a box can be selected by learning the packing history of combinations of parts, so it is possible to eliminate the need to acquire shape information etc. of each part. . Therefore, when selecting a box for packaging an article, the box selection device 100 can select a suitable box with high accuracy.

The neural network calculation unit 12 has embedding vector information corresponding to information included in the input information (in the embodiment, it has embedding vector information for all types of parts), and the embedding layer 22 has embedding vector information corresponding to information included in the input information. Based on the vector information, the input information may be converted into an embedding vector. Thereby, the neural network calculation unit 12 can appropriately output the evaluation value based on the feature amount as described above.

The output layer 23 may output the evaluation value for each of the plurality of boxes in the box information stored in the storage section 14, and the selection section 13 may select the box with the maximum evaluation value. In this case, the selection unit 13 can quickly select any of the boxes stored in the storage unit 14 based on the evaluation value.

The neural network calculation unit 12 may perform learning using training data based on packing results and obtain embedding vectors in the embedding layer 22. This makes it possible to obtain an embedding vector that can output highly accurate evaluation values based on actual packaging results.

The neural network calculation unit 12 calculates the discrepancy between the evaluation value output from the output layer 23 and the correct value of the teaching data using a loss function, and performs embedding by error backpropagation using the result obtained by the calculation. You can get a vector. Thereby, the neural network calculation unit 12 can appropriately learn the teacher data.

The box selection method according to the present embodiment is a box selection method for selecting a box for packing an article based on learning information learned in advance, and includes a storage step S10 for storing learning information and box information, and input regarding the article. An input information reception step S20 for creating information; a neural network calculation step S30 for inputting input information into the input layer 21 and calculating an evaluation value for the box by neural network calculation based on learning information; and a neural network calculation step S30. , a selection step of selecting a box based on the evaluation value output from the output layer 23, and a neural network calculation step S30 includes a calculation by the embedding layer 22 that calculates a feature amount regarding the article from the input information, and a selection step of selecting a box based on the evaluation value output from the output layer 23. Based on the calculation results by the layer 22, calculations are performed by the intermediate layer 30 that evaluates feature amounts.

The box selection program according to the present embodiment is a box selection program that selects a box for packaging an article based on learning information learned in advance, and includes a storage step S10 for storing learning information and box information, and input regarding the article. An input information reception step S20 for creating information; a neural network calculation step S30 for inputting input information into the input layer 21 and calculating an evaluation value for the box by neural network calculation based on learning information; and a neural network calculation step S30. In step S40, the computer system selects a box based on the evaluation value outputted from the output layer 23, and in the neural network calculation step S30, the embedding layer 22 calculates the feature amount regarding the article from the input information. Based on the calculation and the calculation result by the embedding layer 22, calculation is performed by the intermediate layer 30 that evaluates the feature amount.

According to these box selection methods and box selection programs, it is possible to obtain the same functions and effects as those of the box selection device 100 described above.

The invention is not limited to the embodiments described above.

For example, in the embodiments described above, the embedding layer 22 converted the input information into a one-hot vector and then into an embedding vector. However, the embedding layer 22 does not need to convert the embedding vector into a one-hot vector before converting it, or may convert it into another vector.

Further, in the above-described embodiment, the input information reception unit 11 created the input information using the product number ID of each component, but the form of the input information input to the neural network calculation unit 12 is not particularly limited. .

Note that if there is an "unknown part" such as a new part whose details cannot be ascertained, the input information receiving unit 11 may perform the following preprocessing. For example, the input information receiving unit 11 may add a row indicating "unknown parts" to the column of product number ID of each part arranged as input information, and input the number of all unknown parts in the row. Then, the input information receiving unit 11 may discard the line corresponding to "unknown part".

DESCRIPTION OF SYMBOLS 10... Control part, 11... Input information reception part, 12... Neural network calculation part, 13... Selection part, 14... Storage part, 21... Input layer, 22... Embedding layer, 23... Output layer, 30... Intermediate layer, 100 ...Box selection device.

Claims (7)

A box selection device that selects a box for packing an article based on learning information learned in advance,
The control unit of the box selection device includes:
a storage unit that stores the learning information and box information;
an input information reception unit that creates input information regarding the article;
a neural network calculation unit that inputs the input information to an input layer and calculates an evaluation value regarding the box by neural network calculation based on the learning information;
a selection unit that selects the box based on the evaluation value output from the output layer of the neural network calculation unit;
The neural network calculation unit includes:
an embedding layer that acquires feature amounts related to the article from the input information;
A box selection device, comprising: an intermediate layer that evaluates the feature amount acquired by the embedding layer. The neural network calculation unit has embedded vector information corresponding to information included in the input information,
The box selection device according to claim 1, wherein the embedding layer converts the input information into an embedding vector based on the input information and the embedding vector information. The output layer outputs the evaluation value for each of the plurality of boxes in the box information stored in the storage unit,
The box selection device according to claim 1 or 2, wherein the selection unit selects a box with the maximum evaluation value. The box selection device according to any one of claims 1 to 3, wherein the neural network calculation unit performs learning using training data based on packing records and obtains an embedding vector in the embedding layer. The neural network calculation unit calculates a discrepancy between the evaluation value output from the output layer and the correct value of the teacher data using a loss function, and performs error backpropagation using the result obtained by the calculation. The box selection device according to claim 4, wherein the embedding vector is obtained by:. A box selection method for selecting a box for packaging an article based on learning information learned in advance, the method comprising:
a storing step of storing the learning information and box information;
an input information receiving step of creating input information regarding the article;
a neural network calculation step of inputting the input information to an input layer and calculating an evaluation value regarding the box by neural network calculation based on the learning information;
a selection step of selecting the box based on the evaluation value output from the output layer in the neural network calculation step,
In the neural network calculation step,
calculation by an embedding layer that calculates a feature amount regarding the article from the input information;
A box selection method, wherein a calculation is performed by an intermediate layer that evaluates the feature quantity based on a calculation result by the embedding layer. A box selection program that selects a box for packaging an article based on learning information learned in advance,
a storing step of storing the learning information and box information;
an input information receiving step of creating input information regarding the article;
a neural network calculation step of inputting the input information to an input layer and calculating an evaluation value regarding the box by neural network calculation based on the learning information;
causing a computer system to execute a selection step of selecting the box based on the evaluation value output from the output layer in the neural network calculation step;
In the neural network calculation step,
calculation by an embedding layer that calculates a feature amount regarding the article from the input information;
A box selection program in which a calculation is performed by an intermediate layer that evaluates the feature quantity based on a calculation result by the embedding layer.

Images