JP7096708B2 - Inventory management device and inventory management method - Google Patents

Description

The present invention relates to an inventory management technique, and more particularly to an inventory management technique for automatically diagnosing an item inventory using a convolutional neural network.

In the manufacturing, logistics, and retail industries, inventory management is performed for a huge number of products. In inventory management, it is important for companies to detect inventory problems such as overstock and shortages at an early stage, and to adjust supply and demand such as adjusting production volume and reviewing sales plans.
In the field of inventory management, one person in charge manages inventory while dealing with daily exceptions such as emergency orders, delivery date changes, and production delays, and the number of items is hundreds to thousands. Therefore, the inventory transition of the target product over a certain period can be displayed as a graph so that the person in charge can grasp the inventory increase / decrease tendency and retention status at a glance, and the inventory turnover rate, inventory days, inventory moving average, and safety inventory level. Techniques that support early detection of problems by presenting a plurality of indicators such as those for objective inventory evaluation are used (see, for example, Patent Documents 1 and 2).

Detection using information such as inventory transition graphs and indicators is used as initial diagnosis, and in the actual field, the person in charge is in charge of production, sales, and inventory (PSI: Production, Sales and Inventory) of the detected item. Check the detailed transition of, for example, read the situation such as "Inventory has become excessive as a result of mass production even though sales have been on a downward trend during this period", and whether it is a problem inventory or not. Make a final diagnosis.

Patent Document 1 discloses the inventions of the applicants, and describes a technique for displaying thumbnails of inventory transition graphs of a large number of items, a technique for sorting graphs by inventory turnover and crossover ratio, and detailed PSI for each item. Disclosure of technology for displaying detailed transitions in graphs.
Patent Document 2 discloses the inventions of the applicants, so that the ideal safety stock level and the actual safety stock level can be calculated from the PSI data and displayed on a graph so that the deviation can be grasped at a glance. By doing so, we will disclose the technology that supports the optimization of safety stock.

Japanese Unexamined Patent Publication No. 2009-187449 Japanese Patent No. 5457858

However, since inventory diagnosis of items based on detailed changes in PSI and supply / demand adjustment for problem solving are highly specialized tasks, the diagnosis results and adjustment contents may vary depending on the experience and know-how of the person in charge. Many companies, including manufacturers, have problems such as overlooking problems and delays in solving them, and making it difficult to pass on know-how for training new personnel.

The present invention provides a technique that can easily grasp the transition status of PSI, which is a factor of item inventory diagnosis and problem inventory, by using a convolutional neural network that has learned the experience and know-how of the person in charge implemented in the past. The purpose is to do.

According to one aspect of the present invention, the input production / sales / inventory (PSI) data of the item is used to diagnose whether or not the item is problematic inventory, and the item is diagnosed as problematic inventory. In this case, it is an inventory management device that outputs the transition status of PSI data during the period that contributed more than a predetermined contribution to the diagnosis result as the basis of diagnosis or the factor of problem inventory, and changes the supply-demand balance based on the PSI data. , A data learning unit that trains as a feature quantity using a convolutional neural network model, and an inventory diagnosis unit that inputs PSI data of the item to be diagnosed to the trained convolutional neural network model and executes inventory diagnosis for the item. And, an inventory management device characterized by having.

When the result of the diagnosis by the inventory diagnosis unit is problem stock, it propagates back from the output layer of the trained convolutional neural network model to the convolutional layer arranged after the input layer, thereby sequentially PSI in the period of the filter width. Factor estimation that obtains the contribution to the diagnosis result for the filter convolved in the data or the feature map obtained from it, and estimates the PSI data for the period corresponding to the convolution area with the predetermined contribution or more as the basis of diagnosis or the factor of the problem inventory. It is preferable to have a portion.

Furthermore, PSI data by item and day for the item to be diagnosed, teacher data used for learning the convolutional neural network model, and PSI data by item and day for multiple items diagnosed at a certain point in the past. Data consisting of a pair of diagnosis result data, model setting data that defines the filter width and number of filters for at least the convolution layer built after the input layer, learning setting data that defines items related to model learning, inventory diagnosis A data input unit that reads diagnostic setting data that defines items related to and expands it on the memory, an input layer into which PSI data of the item is input according to the model setting data, and a convolution layer are arranged after the input layer. It has a hidden layer having at least one layer and a model building unit for constructing a convolutional neural network model composed of an output layer for outputting the diagnosis result of an item, and the data learning unit has the convolutional neural network. The parameters of the model are initialized with a uniform random number, a normal random number, or a constant value including zero, and PSI data of a plurality of items arbitrarily acquired from the teacher data is input to the model construction unit to perform inventory diagnosis, and the above. It is preferable to sequentially update the parameter values using the gradient so that the error between the diagnosis result output from the model building unit and the diagnosis result of the item group in the teacher data becomes small.

Further, when the factor estimation period is included in the pre-specified adoption period, the item is included in the diagnosis result, the period estimated as a factor, and the character string expressing the PSI transition status of the period for the item to be diagnosed. A screen display unit that displays the screen in the shape of a rectangle, heat map, balloon, etc. on the stock thumbnail list of the group or the PSI detail graph of individual items, and a data output unit that outputs the diagnosis results for the items to be diagnosed to an external file or database. And, it is preferable to have.

According to another aspect of the present invention, the input production / sales / inventory (PSI) data of the item is used to diagnose whether the item is problematic inventory and the item is diagnosed as problematic inventory. In this case, it is an inventory management method using a computer that outputs the transition status of PSI data during the period in which the contribution to the diagnosis result is more than a predetermined contribution as the basis of diagnosis or the factor of the problem inventory, and is based on the PSI data. A data learning step to learn changes in the supply-demand balance as feature quantities using a convolutional neural network model, and input PSI data of the item to be diagnosed into the trained convolutional neural network model to perform inventory diagnosis for the item. An inventory management method is provided, characterized in that it has an inventory diagnostic step to perform.

When the result of the diagnosis by the inventory diagnosis step is a problem stock, it propagates back from the output layer of the trained convolutional neural network model to the convolutional layer arranged after the input layer, thereby sequentially PSI in the period of the filter width. A factor estimation step for estimating the contribution of the filter convoluted to the data or the feature map obtained from the data to the diagnostic result, and estimating the PSI data for the period corresponding to the convolutional region with a predetermined contribution or more as the diagnostic basis or the cause of the problem. It is preferable to have.

According to the present invention, a person in charge of inventory management can easily perform inventory diagnosis of an item and change of PSI, which is a factor of problematic inventory, by using a convolutional neural network that has learned the experience and know-how of the person in charge in the past. I can grasp it. As a result, it is possible to reduce oversight of problem inventory and delay in solving the problem, and even an inexperienced person in charge can perform inventory diagnosis utilizing the know-how of a skilled person.

It is a functional block diagram which shows one configuration example of the stock management apparatus by one Embodiment of this invention. It is an image diagram which shows the outline of the inventory management technique which presents the problem inventory by this embodiment and its factor. It is an image diagram which shows the outline of the inventory management technique which presents the problem inventory by this embodiment and its factor. It is an image diagram which shows the outline of the inventory management technique which presents the problem inventory by this embodiment and its factor. It is a flowchart which shows the flow of model construction. It is a flowchart which shows the outline of the learning process by a model learning part. It is a figure which shows the detailed processing flow of forward propagation. It is a figure which shows the detailed processing flow of back propagation. It is a flowchart which shows an example of the processing flow of a factor estimation part. It is a figure which shows the display example on the inventory thumbnail list screen and the PSI detail screen of the item to be diagnosed in the prior art. It is a figure which shows the display example on the inventory thumbnail list screen and the PSI detail screen of the item to be diagnosed in the prior art. This is an example of visualizing the inventory diagnosis result by the CNN model and the maximum contribution area that is considered to be a factor on the inventory thumbnail list screen. This is an example in which the diagnosis result of a certain target item and the factor due to the maximum contribution area are displayed on the PSI detail screen. It is a figure which shows the example which displayed the diagnosis result of a certain target item and the factor by a heat map on a PSI detail screen.

Therefore, from the PSI data of the item that has already been inventoried by the person in charge and the inventory diagnosis result for the item, the transitional state of PSI in a certain period related to the problem inventory is learned using a convolutional neural network (CNN). By doing so, when the PSI data of the item to be diagnosed is given as the input of the model, the automatic diagnosis of whether the item is in problem stock and the PSI status of the period estimated to be the cause of the problem are output. We propose inventory management technology.

FIG. 1 is a functional block diagram showing a configuration example of an inventory management device according to an embodiment of the present invention. 2A to 2C are image diagrams showing an outline of a problem inventory according to the present embodiment and an inventory management technique for presenting the factors thereof. First, a rough image will be described with reference to FIGS. 2A to 2C.

(Outline of configuration and processing)
As shown in FIG. 2A, the inventory management technique according to the present embodiment first inputs the past inventory diagnosis data 15 into the CNN model 17 to learn the characteristics of the PSI data that correlates with the diagnosis of the problem inventory, and then follows. It is a technique that can obtain the inventory diagnosis result of the item as the output 21 when the PSI data 11 of the item is given to the trained CNN model 17 as an input. For example, the inventory diagnosis results D1 to D3 are graphs showing the inventory status of each item displayed on the display screen, with the horizontal axis representing time and the vertical axis representing inventory quantity. Each graph display shows the item number and the variable sufficiency rate.

Based on such a graph display, as shown in the balloons P1, P2, and P3, the factor of the problem inventory corresponding to the inventory diagnosis results D1 to D3 may be displayed.
For example, with respect to the graph display D1, factor P1 "The sales volume is stable, but the production volume in the last three months has increased, so the inventory is in excess." good.
Regarding the graph display D2, factor P2 "We increased the production volume in preparation for the sales increase trend from half a year ago, but the sales have decreased in the last two months, so it is overstocked." May be displayed.
Further, regarding the graph display D3, the factor P3 “the production amount is insufficient with respect to the sales tendency from one year ago, and the risk of shortage increases” may be displayed. The operator may input the factor based on the judgment of the operator.

FIG. 2B is a diagram showing an outline of basic processing of the CNN model 17. Examples of items diagnosed as problematic inventory by the inventory diagnosis of the person in charge include D11, D12, and D13. Areas R1, R2, and R3 show changes in the supply-demand balance that the person in charge focused on at the time of diagnosis. Finding the basis for diagnosis that "inventory is excessive because we continued production", and in R2, diagnosis that "inventory is excessive because we continued production while holding inventory that can meet demand" Finding the grounds, R3 finds the diagnostic grounds that "production cannot keep up with the increasing trend of sales, inventory runs out, and shortages occur." However, R1, R2, and R3 are tacit-knowledged as the know-how of the person in charge, and are generally rarely recorded as a diagnosis history. Therefore, in the CNN model 17, when a set of PSI data of an item and diagnosis result data as to whether or not the item is problematic is given as past inventory diagnosis data 15, convolution processing is performed for a period length specified in advance. Through, learn the characteristics of PSI data that correlate with the diagnosis of problem inventory. As a result, it is expected that the CNN model 17 will learn the judgment of the change in the supply-demand balance, which has been made into know-how, as shown in R1, R2, and R3.

If the CNN model 17 that learns the characteristics of the PSI data related to the problem inventory is constructed from the diagnosis history data 15, the PSI data 11 of the item to be diagnosed is input to the model, and the convolution processing is performed by the period length specified in advance. It is possible to determine the strength of the characteristics of the input data related to the problem inventory and the probability of being classified as the problem inventory through the full combination process, etc., and to diagnose "problem" or "no problem" for the item at output 31. can.

FIG. 2C is a diagram showing an outline of the processing flow in the CNN model 17 according to the present embodiment, and is a diagram showing the configuration shown in FIG. 2B in more detail.
As shown in FIG. 2C, first, the PSI time series data 11 is input and standardized (normalized) (17-1). Next, Conv processing (convolution processing) 17-2 is performed. The Conv process 17-2 calculates the correlation between the PSI data of the region and the filter while gradually shifting the convolution region having a predetermined period length such as data 11a, 11b, 11c, ... In the time direction. In the normal Conv process 17-2, a filtering process using a plurality of filters such as filter A, filter B, ... Is executed, and various features are learned from the PSI data. In FIG. 2C, the reaction value 33 for the filter A and the reaction value 33 for the filter B in each region are obtained. Next, by performing Tanh treatment 17-3, each reaction value is converted into a value from -1 to +1. A function that converts a reaction value, such as Tanh, is called an activation function. Next, Dropout processing 17-4 is performed.

Next, Affine processing 17-5 is performed. The Affine process 17-5 is a process of aggregating (37) the reaction values 35 of each filter such as filters A and B into one.
Next, by performing the Tanh treatment 17-6, the reaction value of the Affine treatment 17-5 is converted into a value from -1 to +1.
Next, Affine processing 17-7 is performed. In the Afine process 17-7, the reaction values 39 of each filter aggregated in the Tanh process 17-6 are aggregated into the degree of association with each classification label with or without a problem, and then the problem is obtained from the aggregated degree of association. (41) The Softmax process 17-8 is performed to obtain the probability of being classified as a label with a presence and the probability of being classified as a label without a problem.
The Dropout process may be inserted before the Conv process or the Affine process. The Dropout process randomly selects and invalidates a plurality of units of the Conv layer and the Affine layer at the time of learning in order to improve the classification performance of the CNN model for unknown input data. In FIG. 2C, the Dropout process 17-4 is inserted before the first Affine process.

By the above processing, the labeling processing 17-9 based on the classification probability can be executed for the PSI data of the item. As a label, the two values of "problem" and "no problem" are the simplest, but for example, "problem" is subdivided into "excessive safety", "inventory retention", "increase trend", "decrease trend", etc. It may be a multi-valued classification.
Given the teacher data (a data set consisting of a set of {PSI time series, diagnosis result labels}), the CNN model processes 17-1 to 17-9 are executed for the PSI time series, and the model is executed. The error between the label as output and the actual label can be obtained. Therefore, if the error back propagation method is used, the filter value of the Conv layer and the parameter value of the Affine layer can be updated so as to reduce the error. Such parameter update is called CNN model learning.

Factor extraction for the inventory diagnosis result of the item to be diagnosed using the trained CNN model is performed as follows.
That is, for the PSI data of the item diagnosed as the problem inventory in the trained CNN model, the area surrounded by the thick frame that contributes to the diagnosis is specified from the filter. By setting a value other than the maximum output value of Softmax processing (or Affine processing before that) to 0 and propagating back to the Conv layer, the filter that contributes most to the increment of the output value of the corresponding layer. Identify the unit.

(Detailed configuration and processing)
Hereinafter, more detailed and specific processing will be described.
FIG. 1 is a functional block diagram showing a configuration example of an inventory management system according to the present embodiment.
For example, an example will be described in which an inventory management person in charge of a consumer goods manufacturer who is expected to produce thousands of kinds of products executes an inventory diagnosis of the item in charge using the inventory management device 1 shown in FIG. Here, the time when the inventory diagnosis is carried out is, for example, April 1, 2018.

The person in charge of inventory management uses the data input unit 1-1 to determine the PSI data T201 of the item to be diagnosed, the teacher data T202 used for learning the convolutional neural network (CNN) model, and the network configuration of the model. The CNN model configuration setting data T203, the CNN model learning setting data T204 describing the model learning setting items, and the inventory diagnosis setting data T205 describing the inventory diagnosis setting items are read into a memory (not shown) of the inventory management device 1, respectively. The inventory management device 1 is, for example, an information processing device, for example, a server, a PC, a PDA, or the like. The database for storing each data is associated with the inventory management device 1 via, for example, a network.

First, Table 1 shows an example of a format in which the PSI data T201 of the item to be diagnosed is held in the memory. Table 1 shows the item code, date, production quantity, sales quantity, and inventory quantity.

Table 1 shows, for example, the daily PSI actual values for the last one year (April 1, 2017 to March 31, 2018) for 100 items managed by the person in charge of inventory management.

Next, Table 2 shows an example of the format in which the teacher data T202 used for learning the CNN model is held in the memory.

The teacher data according to the present embodiment is composed of PSI data of an item that has been subject to inventory diagnosis in the past and a diagnosis result of the item. In the example of Table 2, for example, the daily PSI actual values and the diagnosis result types for one year (April 1, 2016 to March 31, 2017) for 1,000 items that have been inventoried in the past are shown. It is stored. In this embodiment, since the relationship between the transition status of PSI data for one year and the diagnosis result is learned by the CNN model, the same value is stored in the daily diagnosis result type of each item.

The type of diagnosis result is a numerical value corresponding to "0" for no problem on the memory, "1" for excess inventory, and "2" for understock. It is also possible to subdivide the types of diagnosis results into, for example, "no problem", "excess", "rapid increase", "rapid decrease", and "retention", and assign numerical values corresponding to each. Further, although the PSI data in Tables 1 and 2 are quantities (quantity and number of packages), any numerical value such as an amount of money that records daily results can be used in the same manner as the quantity.

Next, Table 3 shows an example of a format in which the CNN model configuration setting data T203 showing the network architecture of the CNN model is held in the memory.

In the example of Table 3, the layer type is defined for each layer number, and the setting items and the setting values are stored for each layer type. In the present embodiment, the input layer is indispensably designated as the first layer. In the input layer, the shape of the PSI data which is the input value of the CNN model is set. In the present embodiment, for one year's worth of daily data of PSI, the relationship with the diagnosis result is learned while considering the relationship between the production and the inventory transition with respect to the sales tendency in a certain period by the convolution process of the filter. Therefore, the data length is 365 days and the number of channels is 3. For convenience of explanation, the data of 2/29 of the leap year is not considered. In addition, in order to match the data scale for each item, it is set to standardize the data as a pre-conversion.

A convolutional layer is indispensable and designated as the second layer. In this layer, items necessary for extracting features of PSI data that correlate with the diagnosis result are set. The filter width is a period in which the correlation between the PSI data and the diagnosis result is considered in the convolution process. The number of filters is the number of filters to be folded into the PSI data, and the value of each filter is the feature amount of the PSI data extracted through the learning of the CNN model. Stride is the number of days that the convolution process moves each filter from the past to the future. The W ₂ initialization is a method of initializing each element of the matrix W ₂ which is the main parameter of the convolution process (file width × number of channels) × number of filters. Here, the n × m matrix means a matrix having n rows and m columns.

_The b2 initialization is a method of initializing _each element of the vector b2 whose dimension, which is a bias parameter of the convolution process, is equal to the number of filters. The activation function is a function that non-linearly transforms the output after performing a matrix operation on the input of the convolution layer. The value of the activation function of the convolutional layer is also called the feature map. The dropout is a rate at which a part of the units constituting the layer is randomly inactivated during learning in order to improve the generalization performance of the CNN model.
From the third layer onward, a convolutional layer, a pooling layer, a fully connected layer, and the like can be arbitrarily specified.

In Table 3, two fully connected layers are specified. The fully connected layer in the CNN model serves as a classifier for classifying PSI data into diagnostic result types based on the features extracted in the convolutional layer. For the fully connected layer of the third layer, the input dimension is a dimension set based on the output data of the previous layer. The output dimension is the output dimension of the matrix operation performed in the layer. The W3 initialization is a method of initializing _each element of the matrix W3 of the input _dimension × the output dimension, which is the main parameter of the fully connected layer. The _b3 initialization is a method of initializing the element of the output _dimension vector b3 which is a bias parameter of the layer. The activation function and dropout are similar to the convolution layer. The setting items of the fully connected layer of the fourth layer are the same as those of the third layer.

The output layer is required and specified for the final layer. In this layer, items related to the output of the CNN model for obtaining the final inventory diagnosis result from the features extracted by the convolutional process are set for the PSI data. The dimension of the layer is the number of diagnosis result types, and it is necessary to match the output dimension of the previous layer (fully connected layer). The output value calculation is a method of calculating which diagnosis result type is appropriate for the input PSI data.

In this embodiment, the setting items for configuring the CNN model are limited to the above, but additional setting items may be provided in each layer in order to improve the learning speed and generalization performance of the model. For example, in order to avoid large fluctuations in the distribution of the feature map due to changes in the parameter values each time learning is performed, batch normalization processing may be performed before applying the activation function, or inventory diagnosis type. Regularization (load attenuation) may be applied to learn the parameters that do not contribute to the classification of. When a ramp function (ReLU, etc.) is used as the activation function, it is possible to set to execute the local response normalization process in order to prevent the output of the activation function from becoming infinite.

Next, Table 4 shows an example of a format in which the CNN model learning setting data T204, which defines a method for learning the CNN model using the teacher data, is held in the memory.

The number of epochs and the number of iterators are items that define the number of learnings. Learning is executed by the number of epochs x the number of iterators.
The batch size is the number of items arbitrarily extracted from the item set of the entire teacher data. The extracted subitem set becomes the teacher data used for one learning. The loss function is a function for inputting PSI data of an item which is teacher data and obtaining an error between the diagnosis result type output from the CNN model and the actual diagnosis result type of the item. Learning in a neural network is to update the parameter values of the CNN model to minimize the loss function using the teacher data. The parameter update method is a method of updating the parameter value using the gradient information when the loss function is differentiated by each parameter. The learning rate is an item that defines how much the gradient information is reflected when updating the parameters.

Table 5 shows an example of the format in which the inventory diagnosis setting data T205 is held in the memory.

The diagnosis result type defines the options for inventory diagnosis. These options are associated with the diagnostic result type of teacher data and the output of the CNN model. In the diagnostic factor display, the part strongly related to the output of the CNN model when the PSI data of a certain item is input (that is, the inventory diagnosis result by the CNN model of the item) is used with the information of the feature amount extracted by the convolutional layer. Set the method of displaying on the screen as the basis for diagnosis. The maximum contribution period is the period of PSI data corresponding to the feature amount that contributes most to the inventory diagnosis result, and is displayed as the basis for diagnosis. In the heat map, the weighted linear sum of the feature map obtained from the convolution of the input PSI data is obtained by weighting the average contribution to the diagnosis result for each filter, and the linear sum is the relevance of the convolution period. It is regarded as strength, and the strength of the degree of relevance for each convolution period is displayed as the basis for diagnosis. The factor adoption period is a setting for considering the inventory as a problem only when the data period that contributes to the inventory diagnosis result by the CNN model is within the adoption period. This is because, for example, if the period strongly related to the diagnosis result of a certain item is estimated to be 10 months before the diagnosis time, even if it is found that there is a problem with the PSI data of the period, it cannot be dealt with in practice. The purpose is to obtain diagnostic results by focusing on items that have problems with the latest PSI data. In the present embodiment, the data period of the item to be diagnosed is set as it is, but in the above case, the latest period such as 3 months before the diagnosis is set.

Next, the person in charge of inventory management constructs a CNN model set by the CNN model configuration setting data using the model construction unit 1-2 of the inventory management device 1.

FIG. 3 is a flowchart showing the flow of model construction. When the process is started (step S1), the model construction process is repeatedly executed for each layer number of the CNN model configuration setting data (steps S2 to S7).
The layer number is determined in step S3, and when the layer number i is 1, it is an input layer. In step S4, the model building unit 1-2 determines the number of items to be input to the input layer as the number of rows, data length × channel. Generate a matrix X whose number is the number of columns, and initialize all the elements with 0. When the layer number i is 5, it is an output layer, and in step S9, the model building unit 1-2 generates a matrix Y having the number of input items as the number of rows and the dimension as the number of columns, and all the elements are 0. Initialize with. When the layer number i is 2 to 4, it is a hidden layer. In step S5, in the case of the convolutional layer, the model building unit 1-2 first generates a parameter matrix Wi with the filter width × the number of channels as the number of rows and the number of filters as the number of columns, and averages ₀ / standard for each element. Initialize with a normal random value with a deviation of 1. Next, in step S6, a bias parameter vector bi having elements of the number of filters is generated, and each element is initialized to ₀ . In the case of a fully connected layer, the model construction unit 1-2 first generates a parameter matrix Wi with the input dimension as the number of rows and the output dimension as the number of columns, and initially uses a normal random number value with an average of ₀ and a standard deviation of 1. To become. Next, in step S6, a bias parameter vector bi having an output dimension as an element is generated, and each element is initialized to ₀ . For the initialization of the parameters _Wi and bi, another method such as the initialization of He may be applied by changing the setting of the CNN model configuration setting data _T203 .

Next, the person in charge of inventory management executes learning of the CNN model based on the teacher data T202 by using the model learning unit 1-3 of the inventory management device 1. FIG. 4 shows an outline of the learning process by the model learning unit 1-3.

As shown in FIG. 4, when the process is started (step S11), the learning process by the model learning unit 1-3 is repeatedly executed as many times as the number of epochs × the number of iterators (S12 to S20). In one iterator, as shown in step S14, the model learning unit 1-3 first selects 20 batch-sized items from 1000 items in the teacher data T202 according to a uniform random number. Next, in steps S15 and S16, the learning process is executed using the pair of the PSI data for one year and the diagnosis result type for the selected 20 items as the teacher data of the itilate. Learning consists of two processes, forward propagation and back propagation. The forward propagation process U303 in step S15 is a process of propagating the teacher data from the input layer to the output layer based on the current parameter values of the CNN model. The detailed processing flow of forward propagation is shown in FIG. The back propagation process U304 in step S16 is a process of calculating the gradient of the parameter in the loss function by propagating the error obtained from the output value of forward propagation and the diagnosis result type of the teacher data from the output layer to the input layer. .. The detailed processing flow of back propagation is shown in FIG. In addition, N in FIGS. 5 and 6 at the time of learning is 20.

Finally, as shown in step S17, the model learning unit 1-3 updates the parameter value using the gradient information of the parameter obtained by the back propagation. For example, in the AdaGrad method set in the present embodiment, the parameters Wi and bi of the _i -th (2 ≦ _i ≦ 4) layer are updated as follows, respectively.

Here, the initial values of h _w and h _b are 0, and α is the learning rate. An update method other than AdaGrad, such as SGD or Adam, may be set.
In this way, some data is randomly selected from the entire teacher data, and the model output and the actual model output and the actual model output are repeated based on the error between the output of the CNN model using the data and the actual output. It is possible to obtain a parameter value that reduces the error with the output of. In particular, the main parameter W2 of the convolutional layer is based on the one _- year PSI time-series data given as teacher data, "the inventory is excessive as a result of mass production in a certain period of sales decline." In consideration of the mutual relationship of PSI corresponding to the know-how of the person in charge, the characteristics of each time-series data that correlates with the diagnosis result type are learned within the range of the filter width.

In the learning process of the present embodiment, one epoch ends when the iterator is repeated 10 times. Before moving on to the next epoch, I changed the seed of the uniform random number, but this is not a required process. With this setting, the number of epochs x number of iterates x batch size is 1,000, so when the learning process for all epochs is completed, the relationship between the PSI data of all items and the diagnosis result type from the uniformity of random numbers. Learn once.

Next, the person in charge of inventory management uses the inventory diagnosis unit 1-4 to perform inventory diagnosis of, for example, 100 items to be diagnosed. For the diagnosis of the target item, the input layer is a 100 × 1095 matrix X in which the standardized PSI data for 365 days of each item is stored in one row with the number of input items N = 100, and the forward propagation of the CNN model with learned parameters is performed. You just need to find the output. The procedure is the same as the forward propagation during learning shown in FIG. 5, but the dropout that randomly invalidates a certain percentage of units in the hidden layer is not performed at the time of diagnosis, and the input matrix to the layer is (1- Dropout rate) is multiplied and propagated to the next layer as it is. Since the softmax value (classification probability of each diagnosis result type) of the target item is stored in each row of the output matrix Y, the diagnosis result in which the column number with the largest value (maximum classification probability) is predicted from the model. Obtained as a type.

Next, the person in charge of inventory management uses the factor estimation unit 1-5 to obtain a region of PSI data that greatly contributes to the diagnosis result as a basis for inventory diagnosis by the CNN model. Figure 7 shows the flow of factor identification. Although the details of the factor identification method differ depending on the selection items for presenting the diagnostic factors, they are common in that the position of the convolution area and the value of the feature map, which greatly contribute to the classification probability of the diagnosis result type, are obtained from the gradient information in the layer. ing.
As shown in FIG. 5, in the forward propagation process of the CNN model, first, the process is started in step S31, and in step S32, the process of generating the data to be input to the input layer unit is performed. For example, for the matrix X of N × 1095 of the input layer where the number of input items is N, the standardized PSI for 365 days of the item j (j = 1, ..., N) selected as the teacher data or the diagnosis target. The data is stored as the data on the jth line of X.

In step S33, conversion (preprocessing) for executing convolution by matrix operation is performed. For example, the matrix X is converted into a matrix X ₂ of (number of items × number of convolution areas) × (filter width × number of channels). However, the number of convolution areas is calculated as (data length-filter width) / stride + 1, and this time it is (365-90) / 1 + 1 = 276. In _each row of the matrix X2, PSI data of the convolution target period obtained by striding the filter width for each item is stored.
In step S34, a filter convolution process for extracting a feature amount from the data is performed. For example, the convolution operation X'= X ₂ × W ₂ + b ₂ is executed, and the matrix X'of (the number of items × the number of convolution areas) × the number of filters is calculated. However, the sum of the matrix and the vector in X ₂ × W ₂ + b ₂ is a process of adding the value of the k-th element of the vector to each element of the k-th column of the matrix.

In step S35, post-processing for undoing the data conversion of the pre-processing is performed. For example, the matrix X'is converted into a matrix X'' of the number of items × (the number of convolution areas × the number of filters). In each row of the matrix X'' of N × 2760, 276 reaction values obtained by convolving PSI data with the filter width for each item are stored as many as the number of filters.
In step S36, the output value calculation process of each unit of the convolution layer is performed. For example, the matrix X ₃ of N × 2760 is calculated as X ₃ = tanh (X''). However, the application of the activation function to the matrix X'' is the process of obtaining the value of the activation function for each element of the matrix. The matrix X ₃ is a feature map for each item.
In step S37, it is determined whether the forward propagation process is executed by model learning or by diagnosing the target item. If Yes, the process proceeds to step S38, and if No, the process proceeds to step S43.
In step S38, in order to prevent overfitting of the model, a process of arbitrarily selecting and invalidating a plurality of units of the fully connected layer is performed. For example, the elements of the matrix X3 _are randomly set to 0 at the dropout rate.

On the other hand, in step S43, a process of reflecting the overfitting reduction effect at the time of diagnosis is performed. For example, let the matrix X ₃ = (1-dropout rate) × X ₃ . Next, from steps S38 and S43, the process proceeds to step S39, and a full combination process is performed in which the convolution results of the plurality of filters are aggregated into one. For example, the full binding process X ₄ = tanh (X ₃ × W ₃ + b ₃ ) including the application of the activation function is executed, and the matrix X ₄ of N × 276 is calculated. The method of summing a matrix and a vector and the application of the activation function to the matrix are the same as above.
Next, the process proceeds to step S40, and a full binding process is performed in which the aggregated convolution results are associated with the diagnosis type. For example, the full join process X ₅ = X ₄ × W ₄ + b ₄ is executed, and the matrix X ₅ of N × 3 is calculated. The method of summing the matrix and the vector is the same as above.

Next, the process proceeds to step S41, and a process (output layer output) of outputting the degree of association with the diagnosis type with the classification probability is performed. For example, the result of softmax (X ₅ ) is assigned to each element of the matrix Y of N × 3. here. The softmax function is a function for obtaining a softmax value (probability of being classified into each diagnosis result type) for each row of the matrix _X5 . Then, the process is terminated (step S42).

As shown in FIG. 6, in the back propagation process of the CNN model, the process is first started in step S51, and in step S52, the gradient (partial differential) of the classification probability conversion is obtained for the error between the output layer output and the teacher output. Perform processing. For example, the actual diagnosis result type (0 to 2) of each item is converted into the same three-dimensional onehot vector as the output layer to generate an N × 3 matrix T, and the cross entropy of the output matrix Y and the teacher label matrix T is generated. The gradient matrix _dX4 = (YT) / N of the softmax function based on the error is back-propagated to the fully connected layer of the previous layer.

In step S53, a process of obtaining the output of the fully connected layer of the fourth layer and the gradient of the parameter with respect to the error is performed. For example, the output gradient matrix dX ₃ = dX ₄ × W ₄ ^t of N × 276 is back-propagated to the fully connected layer of the previous layer. However, W ₄ ^t means the transposed matrix of W ₄ . Further, the gradient of the parameter W ₄ of the layer is obtained as dW ₄ = X ₄ ^t × dX ₄ , and the gradient of the parameter b ₄ is obtained as db ₄ = sum (dX ₄ ). Here, the sum function is a function that outputs a vector whose element is the sum of each column of the matrix.
In step S54, a process of obtaining the output of the fully connected layer of the third layer and the gradient of the parameter with respect to the error is performed. For example, the gradient matrix dX'of the tanh function is obtained by dX'= dX ₃ * (I-X ₄ * X ₄ ), and the output gradient matrix dX ₂ = dX'x W ₃ ^t of N × 2760 is obtained by the convolution layer of the front layer. It propagates back to. However, I is an N × 2760 matrix in which all the elements are 1, and the operator * is an operation for obtaining the product of each element of the matrix. Further, the gradient of the parameter W ₃ of the layer is obtained as dW ₃ = X ₃ ^t × dX', and the gradient of the parameter b ₃ is obtained as db ₃ = sum (dX').

In step S55, a process of invalidating the back propagation of the unit invalidated by the forward propagation is performed. For example, the element of the matrix _dX2 corresponding to the element of the matrix X3 that was set to ₀ in the forward propagation dropout is set to 0.
In step S56, a process of obtaining the gradient of the feature map with respect to the error is performed. For example, the gradient matrix dX'of the tanh function is obtained by dX'= dX ₂ * (IX ₃ * X ₃ ).
In step S57, preprocessing is performed to obtain the output gradient and the parameter gradient of the convolution layer by matrix operation. For example, the matrix dX'of N × 2760 is converted into the matrix dX'' of (number of items × number of convolution areas) × number of filters. In each row of dX', the gradient of the loss function for 276 reaction values of each item is stored as many as the number of filters. Therefore, each row of dX'is divided into filter units, the gradient of the loss function is sorted by item and convolution area, and stored in each row of dX''.

In step S58, a process of obtaining the output gradient and the parameter gradient of the convolution layer with respect to the error is performed. For example, the following formula, (number of items × number of convolution areas) × (filter width × number of channels) output gradient matrix dX ₁ = dX'' × W ₂ ^t is calculated. Further, the gradient of the parameter W ₂ of the layer is obtained as dW ₂ = X ₂ ^t × dX'', and the gradient of the parameter b ₂ is obtained as db ₂ = sum (dX'') / the number of convolution regions.
In step S59, post-processing for undoing the data conversion of the pre-processing is performed. For example, the gradient matrix dX ₁ of (number of items × number of convolution areas) × (filter width × number of channels) is converted into a matrix dX of N × 1095, back-propagated to the input layer, and the process is terminated (step S60).

FIG. 7 is a flowchart showing an example of the processing flow of the factor estimation unit 1-5. The process is started (step S61), and in step S62, a process for removing the influence other than the diagnosis type having the maximum degree of relevance is performed. For example, for each row of the N × 3 matrix _X5 in which the PSI data of N items to be diagnosed are forward-propagated to the fully connected layer of the fourth layer, the values of the columns other than the column having the maximum value are set to 0. Find dX ₄ .
In step S63, a process of obtaining the output gradient of the fully connected layer with respect to the degree of relevance is performed. For example, the gradient matrix dX ₃ = dX ₄ × W ₄ ^t of N × 276 is back-propagated to the fully connected layer of the previous layer.
In step S64, a process of obtaining the output gradient of the fully connected layer with respect to the degree of relevance is performed. For example, the gradient matrix dX'of the tanh function is obtained by dX'= dX ₃ * (I-X ₄ * X ₄ ), and the gradient matrix dX ₂ = dX'x W ₃ ^t of N × 276 is used as the convolution layer of the front layer. Backpropagate.
In step S65, a process of reflecting the overfitting reduction effect in the gradient calculation is performed. For example, it propagates back to the convolution layer as dX ₂ = (1-dropout rate) × dX ₂ .

In step S66, the presentation method is determined. If the presentation method is the maximum contribution period, the process proceeds to step S67, and if the presentation method is a heat map, the process proceeds to step S70.
In step S67, a process of obtaining the gradient of the feature map with respect to the degree of relevance is performed. For example, the gradient matrix dX'of the tanh function is obtained by dX'= dX ₂ * (IX ₃ * X ₃ ). Next, in step S68, a process of specifying a filter corresponding to the convolutional layer unit having the maximum gradient and a region in which the filter is convoluted is performed. For example, for each row of the matrix dX'of N × 2760, the column number c that takes the maximum value is obtained, and the quotient +1 when c is divided by the number of convolution areas is the filter number f, and the remainder r is the convolution area in the filter f. The N × 2 matrix Z stored as a position is generated, and the process is terminated (step S69).

In step S70, a process of obtaining the weight of the feature map with respect to the degree of relevance is performed for each filter. For example, for _each row of the gradient matrix dX2 of N × 2760, an N × 10 matrix A storing the average value of the gradients of the feature map obtained for each filter is generated.
In step S71, a process of obtaining a weighted output after filter convolution for the input data is performed. For example, for _each row of the feature map matrix X3 of N × 2760 that forward propagates the PSI data of the item to be diagnosed to the convolutional layer of the second layer, the feature map of the filter number f (f = 1, ..., 10) is Considering that it is stored in the (f-1) × 276 + 1st column to the f × 276th column, the value in the (f-1) × 276 + 1st column to the f × 276th column is the fth column of the matrix A. Multiplies the values of to generate the matrix B.

In step S72, a process of obtaining the density of the heat map of each region in which the filter is folded is performed from the weighted output. For example, for each row of the matrix B of N × 2760, the values of the weighted feature maps at the mth (m = 1, ..., 256) are added for each filter f. If the sum of the mth th is greater than 0, the value is generated, and if it is 0 or less, a matrix Z of N × 276 containing 0 is generated.
Then, the processing is terminated from steps S68 and S72 (step S69).

Next, the person in charge of inventory management visualizes the inventory diagnosis result of the item to be diagnosed by the CNN model and the factors that are the basis of the diagnosis on the screen by using the screen display unit 1-6.
8 and 9 are diagrams showing display examples on the inventory thumbnail list screen and the PSI detail screen of the items to be diagnosed in the prior art. On the other hand, FIG. 10 is an example of visualizing the inventory diagnosis result by the CNN model and the maximum contribution area which is considered to be a factor on the inventory thumbnail list screen. The inventory diagnosis result by the CNN model is visualized as a rectangle superimposed on the frame of the inventory thumbnail of each item diagnosed as the problem inventory.
It is preferable that the rectangular line has a different color or line type (thick line, broken line, alternate long and short dash line, etc.) depending on the type of problem. In addition, the maximum contribution area is visualized as a rectangle filled with a constant transmittance on the inventory graph. The display position of the maximum contribution region can be determined by the position r of the convolution region of the matrix Z, which is the output of the factor estimation unit 1-5.

In addition, FIG. 11 is an example in which the diagnosis result of a certain target item and the factor due to the maximum contribution area are displayed on the PSI detail screen. The maximum contribution area is visualized as a filled rectangle on the PSI detail graph. When the mouse cursor is moved on the rectangle, a balloon expressing the production / sales / inventory status of the area may be displayed.
In the present embodiment, by comparing the sales growth rate, the number of production times, the production amount, and the average inventory in the maximum contribution area and the other areas, the PSI status of the period that contributed most to the diagnosis is different from that of other periods. Information on how different they are can be presented.
It is not limited to the display mode according to the present embodiment as long as it is character string information indicating the status of PSI in the maximum contribution area and promotes the understanding of the person in charge regarding the diagnostic basis.

Further, FIG. 12 is a diagram showing an example in which the diagnosis result of a certain target item and the factor by the heat map are displayed on the PSI detailed screen. The heat map is displayed on the PSI detailed graph as a vertical line in which the shade of color is adjusted according to the degree of contribution of each convolution area to the diagnosis result. The degree of contribution of each convolution region to the diagnosis result can be obtained from the matrix Z which is the output of the factor estimation unit 1-5.
In the visualization of the diagnosis results and factors of the items in FIGS. 10, 11, and 12, the most recent PSI situation is limited to the case where the period estimated as the factor of the problem inventory is included in the factor adoption period. It is also possible to display the inventory diagnosis result focusing on the item having a problem with.

The person in charge of inventory management can adjust the supply and demand for solving the problem inventory based on the inventory diagnosis result by the CNN model visualized by the screen display unit 1-6 and its factors. Further, if the diagnosis result of the CNN model is different from the judgment of the person in charge, the diagnosis result can be corrected on the screen.

Further, the person in charge of inventory management uses the data output unit 1-7 to perform the diagnosis result data T206 (
The diagnosis results of each item can be output to an external file or database in the format shown in Table 6). Since the diagnosis result data can be used as it is as the next teacher data, the CNN model can always learn the latest diagnosis result and reflect it in the future inventory diagnosis.

Processing and control can be performed by software processing by CPU (Central Processing Unit) or GPU (Graphics Processing Unit), by ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Hardware).

Further, in the above-described embodiment, the configuration and the like shown in the illustration are not limited to these, and can be appropriately changed within the range in which the effect of the present invention is exhibited. In addition, it can be appropriately modified and implemented as long as it does not deviate from the scope of the object of the present invention.
In addition, each component of the present invention can be arbitrarily selected, and an invention having the selected configuration is also included in the present invention.

The present invention can be used for an inventory management device.

1 Inventory management device 1-1 Data input unit 1-2 Model construction unit 1-3 Model learning unit 1-4 Inventory diagnosis unit 1-5 Factor estimation unit 1-6 Screen display unit 1-7 Data output unit 17 Problem inventory AI model for learning the characteristics of PSI data 17-1 Standardization (normalization) processing 17-2 Conv processing (convolution (correlation) processing)
33 Reaction value 17-3 Tanh processing 17-4 Dropout insertion processing 17-5 Affine processing 17-6 Tanh processing 17-7 Affine processing 17-8 Softmax processing 21 Output

Claims (4)

If the input item's production / sales / inventory (PSI) data is used to diagnose whether the item is problematic inventory and the item is diagnosed as problematic inventory, a predetermined contribution to the diagnosis result. It is an inventory management device that outputs the transition status of PSI data during the period that contributed more than once as the basis of diagnosis or the factor of problem inventory.
A data learning unit that learns to output information indicating whether or not the item is in question stock based on the feature amount, using the change in the supply-demand balance based on the PSI data as a feature amount using a convolutional neural network model. When,
An inventory management device comprising, for the convolutional neural network model that has been trained, an inventory diagnosis unit that inputs PSI data of an item to be diagnosed and executes an inventory diagnosis for the item. When the result of the diagnosis by the inventory diagnosis unit is problem stock, it propagates back from the output layer of the trained convolutional neural network model to the convolutional layer arranged after the input layer, thereby sequentially PSI in the period of the filter width. Factor estimation that obtains the contribution to the diagnosis result for the filter convolved in the data or the feature map obtained from it, and estimates the PSI data for the period corresponding to the convolution area with the predetermined contribution or more as the basis of diagnosis or the factor of the problem inventory. The inventory management device according to claim 1, further comprising a unit. If the input item's production / sales / inventory (PSI) data is used to diagnose whether the item is problematic inventory and the item is diagnosed as problematic inventory, a predetermined contribution to the diagnosis result. It is an inventory management method using a computer that outputs the transition status of PSI data during the period that contributed more than once as the basis of diagnosis or the factor of problem inventory.
A data learning step in which a change in the supply-demand balance based on the PSI data is used as a feature quantity using a convolutional neural network model, and information indicating whether or not the item is in question stock is output based on the feature quantity. When,
An inventory management method comprising:, for the convolutional neural network model that has been trained, an inventory diagnosis step of inputting PSI data of an item to be diagnosed and executing an inventory diagnosis for the item. When the result of the diagnosis by the inventory diagnosis step is a problem stock, it propagates back from the output layer of the trained convolutional neural network model to the convolutional layer arranged after the input layer, thereby sequentially PSI in the period of the filter width. A factor estimation step for estimating the contribution of the filter convoluted to the data or the feature map obtained from the data to the diagnostic result, and estimating the PSI data for the period corresponding to the convolutional region with a predetermined contribution or more as the diagnostic basis or the cause of the problem. The inventory management method according to claim 3, wherein the product has.

Images