Logistics System

Abstract

To provide a system, a device and a method capable of executing failure detection and prediction of a device with high accuracy using machine learning.MEANS FOR SOLVING THE PROBLEM: In a complicated device including at least one of a multi-tier automatic warehouse, a car, an airplane, a factory manufacturing facility and a plant, failure detection and / or failure prediction is executed using machine learning. A device state from a normal state to a failure state is set as at least one state between the normal state and the failure state and a process leading to the failure is grasped as a state transition between them. Sensor data including at least one of acoustic, vibration, acceleration, load, velocity, image, temperature and humidity may be obtained and the data through a step of processing the obtained sensor data and extracting a feature amount may be regarded as an input of the machine learning.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This application relates to warehouse control, which is related to technology for controlling the movement of goods within a logistics center, regardless of the warehouse method and structure, including the structure, size, loading / unloading method, and goods movement method of the logistics center, and in particular to an apparatus, system, program, and recording medium related to a method for analyzing logistics and optimizing logistics, and detecting and predicting failures using a 3D simulator and AI. [Background technology]

[0002] Logistics centers in the distribution industry and the like are equipped with automated multi-story warehouses that include a storage warehouse section for storing cases containing products (storage cases) and a picking station that picks the required number of products from cases removed from the storage warehouse section and sends them to a shipping line. In addition, downstream of the picking station, there may also be a packaging line that packages the picked products using an automatic box-sealing device or the like.

[0003] An automated multi-story warehouse is a warehouse management system (WMS) equipped with a host computer (or server, cloud, etc.) that centrally manages warehouse operations. The WCS (Warehouse Control System) is a system that contains a computer connected to the WMS host computer. It is managed by the System.

[0004] Based on instructions from the WMS, the WCS issues work instructions such as receiving, transporting, retrieving, picking, and automatic sealing to equipment within the automated warehouse (transportation devices, lifting devices, buffer conveyors, etc.), picking stations, and packaging lines via wireless remote control and communication lines.

[0005] In recent years, the proliferation of online sales (e-commerce) has led to the expansion of logistics centers, while the variety and number of goods handled are steadily increasing. In this competitive environment, the need for shorter delivery times has led to an ever-increasing demand for efficient logistics center operation. However, the size, configuration, and operation methods of a logistics center vary depending on the type, number, and size of the goods handled. In addition, depending on the type of goods handled, there can be significant seasonal fluctuations, as well as fluctuations due to business factors such as campaigns. Logistics center hardware is individually designed to take these factors into account. As a result, the software that controls the inbound and outbound shipments to and from warehouses and the transfer of goods within a logistics center must also be individually designed. Furthermore, the performance of this control software directly determines the performance of a logistics center, including its outbound efficiency.

[0006] As such, control software for distribution centers not only must be individually designed, but because it is a critical factor in determining performance, it requires a huge investment of time and resources to develop.One way to improve this resource-intensive situation in which the development of control software for distribution centers is to build general-purpose software in advance, then customize it by setting the attributes of each warehouse (number of storage shelves, layout, number of shelves, number of openings, type of transfer means, performance, number of units, etc.) as hardware, and then automatically generate control software that matches each warehouse.

[0007] However, it is extremely difficult to build such general-purpose control software in advance. Generally, a logistics center's control software grasps the aforementioned warehouse attributes and then inputs the current, ever-changing status of the logistics center, including the number of occupied and available space slots and their addresses, the number and addresses of available transfer vehicles, and the current warehouse status, including transfer commands indicating which items, how many, and to which exit gate. After calculating these, the software outputs control commands indicating the path along which items should be moved. However, the number of combinations of warehouse attributes and status required to control the movement of items within a warehouse is enormous. Therefore, it is virtually impossible to build general-purpose software that takes all situations into account. As a result, it has been necessary to create individual control software for each individual automated warehouse hardware, each with its own limited operating conditions.

[0008] It is also common to monitor the status of complex equipment. For this purpose, various sensors (such as acoustics, vibration, acceleration, load, speed, images, temperature, and humidity) are attached to necessary locations inside and outside the equipment, and the signals obtained from these sensors are monitored. When the equipment breaks down, an abnormal signal is sent from the sensor, which is used to detect the failure. Traditionally, failures have been detected by building an algorithm to detect abnormal signals from each sensor and implementing it as software. However, there is also a method that uses machine learning technology to learn the abnormal signals sent from the sensors and use this to detect abnormalities.

[0009] While these technologies make it possible to detect faults, they are still not sufficient. For example, algorithm-based fault detection can miss faults that should be detected because it is difficult to incorporate all failure modes into the software. On the other hand, fault detection using machine learning cannot achieve satisfactory fault detection because the training data is insufficient due to the rarity of faults.

[0010] In addition to failure detection, there is also a technical demand for failure prediction. This demand arises from the original objective of keeping equipment running without interruption. If failure prediction becomes possible, it will be possible to avoid unexpected equipment downtime by replacing parts and components (units that make up equipment made up of multiple parts) that are likely to fail. Moreover, there is a technical demand for prediction as early as possible. This is because some of the parts and components used in equipment can be expensive and have long delivery times.

[0011] Furthermore, the recent record-breaking expansion of e-commerce has led to a serious labor shortage in the logistics industry, with logistics centers actively recruiting staff. However, neighboring centers are competing for staff positions, exacerbating the labor shortage. This has led to unprecedented demand for labor-saving measures at logistics centers. To address this issue, automation and labor-saving measures for material handling equipment are being implemented, such as combining automated warehouses with Goods-to-Person (GTP) stations to remove products from the automated warehouse in the order of delivery, and then picking them at the GTP station with the minimum number of staff and transporting them to the packaging line. Meanwhile, for infrequently picked items, the traditional method of having people walk to the shelves where they are stored remains in use. However, systems that analyze worker movement patterns are being considered to standardize and streamline these tasks.

[0012] Furthermore, methods are being considered for taking photographs of the condition of products during the packing process before shipping them from the logistics center, linking this to WMS data, and leaving evidence that the product was shipped properly within the center. These systems can be said to be part of the recent trend of promoting the digitalization of logistics centers through the use of IoT. [Prior art documents] [Patent documents]

[0013] [Patent Document 1] Japanese Patent Application Publication No. 02-255412 [Patent Document 2] Japanese Patent Application Publication No. 05-204891 [Patent Document 3] Japanese Patent Publication No. 2020-007060 Summary of the Invention [Problem to be solved by the invention]

[0014] Therefore, the applicant decided to make it his goal to provide a system, program, and recording medium that can be applied universally for optimization, including control software for individually designed material handling equipment at logistics centers and the movement of workers.

[0015] In addition, for systems that are already in operation, system operation information (flow rate and speed of cases, etc. moving on the conveyor, and case / product information, etc.) ) Furthermore, adding the above-mentioned system operation information collection function to an existing system would require not only the installation of physical sensors but also system modifications, and modifying a system while it is in operation would pose a high risk of affecting the operation of the existing system, and there is also a lack of time to actually carry out the modification work, so the reality is that modifications are not possible.

[0016] However, under these circumstances, it is necessary to collect operational information, visualize the data using simulations, and utilize it in data analysis systems using AI, etc., to improve the efficiency of system operation.

[0017] Furthermore, when the purpose is to measure the flow rate of goods from image data and its analysis, it is necessary to determine where and how to install the device in the logistics system (line or moving part of an automated warehouse), what data to collect, and process the resulting data into visualized data to show the actual state of logistics in chronological order, thereby improving logistics control, including proposals for responding to fluctuations in logistics due to seasons, busy seasons, slow seasons, etc. In addition to quantity, image analysis technology can also be used to analyze the movement lines of workers who perform picking work on foot, which is useful for improving the efficiency and leveling of the work of these workers.

[0018] Furthermore, it is natural that logistics operations deteriorate over time or through frequent use, and once a breakdown occurs in the logistics system, from a safety standpoint, the line must be stopped for a certain period of time and over a certain area to carry out maintenance work, which causes a certain delay in logistics operations.

[0019] The goal is to achieve "uninterrupted logistics" through predictive maintenance services that support stable operation of logistics centers, and to provide a constant monitoring system for mobile objects for predictive maintenance services that guarantee this.

[0020] Complex equipment such as automated warehouses, automobiles, aircraft, factory manufacturing equipment, and plants are made up of a huge number of parts and components. As a result, not only are the failure modes complex and difficult to predict, but in some cases, overlooking a minor failure can lead to a major breakdown. While it goes without saying that it is ideal to keep equipment operating normally without stopping it, failures are inevitable, and as a practical response, there is a technical demand for minimizing equipment downtime and even predicting failures in advance.

[0021] Therefore, the present application focuses on the above-mentioned problems and aims to provide a generalized method, device, system, program, and recording medium for controlling a logistics center that can be generally applied to improve the operation of logistics center equipment, detect and predict equipment failures and other conditions using a 3D simulator and AI with high accuracy, and improve the efficiency of work content, thereby enabling visualization of logistics without the need to create individual control software for the logistics center. [Means for solving the problem]

[0022] In order to solve the above problems, the logistics system of the present invention is characterized in that, throughout the entire process from arrival at the logistics center to shipment, identification information of the transported goods, workers, and material handling equipment obtained by sensing means and / or image acquisition means is obtained via a communication line, the logistics flow rate of the transported goods and the movement lines of the workers and / or material handling equipment are optimized using a 3D simulator on a computer, and the logistics flow rate information and failure detection and / or failure prediction information are displayed.

[0023] By digitally recreating an actual operating logistics center and improving the operational efficiency of equipment and machinery beyond that of conventional control logic, and combining it with AI to further improve logistics efficiency, we aim to optimize logistics volume by maximizing the capabilities of equipment and machinery while responding to customer demands, and to optimize the movement of AGVs and / or workers, as well as predicting and maintaining logistics equipment failures.

[0024] A 3D simulator is used to digitally recreate a logistics center, and the actual equipment and facilities in the physical space are reproduced exactly in the virtual space. This allows for a perfect reproduction of actual movements, allowing the movement of products and equipment in an actual logistics center to be accurately reproduced and verified digitally.

[0025] By using a 3D simulator, it is possible to conduct all kinds of verification before building a new logistics center, expanding an existing one, or changing the layout, to see how much efficiency will actually be achieved.

[0026] It features 3D simulation as software that accurately simulates hardware, and solves the above-mentioned issues through image processing, tagging, and AI technology.

[0027] Furthermore, by processing data from an operating logistics center into visualized data and viewing the actual logistics volume over time, it may be possible to improve logistics volume control in response to fluctuations in logistics volume due to seasons, busy periods, slow periods, etc.

[0028] The visualization data may include data on the movement of AGVs and / or workers.

[0029] Here, accurate simulation of the hardware means that if the same control signals as those in the WMS or WCS are input, the hardware will operate in the same way as in an actual logistics center.

[0030] A feature of this system is that it can digitally recreate an actual logistics center in operation, introduce AI to improve logistics efficiency, and increase the efficiency of equipment and machinery in and out of storage compared to conventional control logic, thereby increasing the amount of logistics that can be handled.

[0031] By using reinforcement learning AI, it can learn with less data than conventional AI, analyze the results of its learning itself, and generate new data.By introducing reinforcement learning AI into logistics control, it can be characterized as automatically generating optimal solutions.

[0032] The system may be characterized by comprising a sensor installed on an elevator in a logistics center and / or a structure that structurally supports the elevator and / or a part related to the structure, and acquiring sensing information relating to physical quantities including vibrations in real time or at regular time intervals; a wireless transmitting unit that transmits the sensing information acquired by the sensor via wireless; a receiving unit that receives the sensing information transmitted from the wireless transmitting unit; and a 3D simulator that uses artificial intelligence (AI) technology to predict when to replace parts of the elevator equipment from the sensing information received by the receiving unit.

[0033] In order to reliably detect machine abnormalities, the system may be characterized by adopting a method of detecting abnormal data by comparing it with normal data that has been highly modeled using AI machine learning, rather than simply determining abnormalities based on differences in collected data.

[0034] Regarding the comparison of modeled normal data, the model is created based on past data as well as the latest data collected in real time, and a method may be characterized in that equipment with the same specifications as the actually installed machinery and equipment is constructed as a 3D simulator on a PC screen based on the model, and AI performs simulations using the 3D simulator, which is used as a means for making optimal judgments about whether or not there is an abnormality.

[0035] With the aim of improving facility operation, the system may be characterized by the quantity management of transported items, the movement line management of AGVs and / or workers (picking, assorting), and the safety management of transported items.

[0036] The operational status of the customer's WCS, PDR, images, and other digital information such as log information can be collected in real time from the customer's equipment to grasp the situation, and this information can then be used in a 3D simulator to reproduce it in a digital space on a PC.

[0037] Based on this reproduced situation, AI can use machine learning to determine whether the customer's current equipment condition is optimal, and if it is not, it can be used as a management system to propose measures to make it optimal.

[0038] Furthermore, by using high-performance, highly reliable wireless communication, low-cost, high-performance IoT sensors for predictive maintenance can be used to visualize the operating status of equipment and predict abnormalities, thereby optimizing the timing of part replacement, etc., and the ``predictive maintenance service'' can be characterized as making it possible to achieve ``uninterrupted logistics.''

[0039] This device is equipped with sensors, cameras, batteries, and communication functions and is installed near the logistics line to collect the above system operation information and transmit it to a server via wireless communication.By using this device, it can be introduced without disrupting the operation of an active logistics center.

[0040] The above equipment is equipped with a camera and an image recognition function that recognizes the barcodes, product codes, and product names of cases passing in front of the camera. It also has an image recognition function that counts the number of cases passing in front of it. The above device can also be equipped with an RFID reader, which collects RFID tag information from individual items to identify them. It is also possible to measure the size of passing cases using image recognition.

[0041] As a system using the above-mentioned device, analytical tags are automatically attached to the acquired data in order to analyze the data.

[0042] Using a camera and a server-based photography system, videos, images, reservation links, etc. can be placed anywhere within the generated 3D model. For example, links can be embedded into 3D models to access necessary logistics data.

[0043] It also has a management console that allows you to set the orientation of the sensor and the type of information data to be collected from the device.

[0044] The means may be characterized as a means for detecting the status of equipment that has already been installed or is in operation by adding a retrofit sensor without stopping the equipment, and capable of knowing the status of the equipment with sufficient spatiotemporal resolution from information including the position and movement of the equipment.

[0045] By obtaining data as information from the device, it may be possible to efficiently transfer and consolidate items, taking into account time, energy such as electricity used, physical noise, etc., and to prevent or minimize hardware breakdowns through predictive maintenance.

[0046] It features data collection, platform, and tagging, and may also build a management platform with visualized data.

[0047] Data from automated warehouses, which hold a particularly important position in logistics control, includes information on the location and movement of shuttles, lifters, carts, conveyors, and items, as well as information on the storage status of racks. Visualization is achieved by collecting this information with tags and displaying the current status on a platform screen within a 3D image of a physical rack reconstructed from a blueprint. This means collecting data to display the actual movement of shuttles and conveyors, i.e., the exact current situation as it can be seen, on a computer screen. This visualized data can also be distributed to the necessary departments via a communications network.

[0048] In the case of automated warehouses, which require large amounts of operational data to build a management platform, various sensors must be installed on the equipment while it is in operation. However, installing sensors requires the automated warehouse to be temporarily stopped, which is impossible. Therefore, we have developed a system that collects data on shuttles, lifters, carts, conveyors, the position and movement of items, and rack storage status information without stopping the automated warehouse in operation.

[0049] As one example, multiple cameras could be installed along the perimeter of the rack on each floor and each shelf, and the shuttle's position could be determined from the images obtained from them. However, if there are multiple shuttles or rack contents on the same floor or shelf, the shuttle or contents in front may shade the camera and prevent the shuttle from being captured. To reduce this possibility, the probability of shadowing can be calculated, and if it is within an acceptable range, it is acceptable to temporarily be unable to obtain the shuttle's position, or the entire rack can be treated as a hexahedron and cameras installed on all sides, thereby minimizing blind spots.

[0050] As an example, ultrasonic waves are used. If there are no obstacles, reflected waves from an ultrasonic transmitter installed at the end of the rack are detected, and the deviation in the reflected frequency due to the time and Doppler effect is detected to determine the shuttle's position, speed, and acceleration. The acceleration measurement can also be used to determine the mass of the item placed on it.

[0051] In one embodiment, a camera, RFID, BT (Bluetooth: one of the standards for short-range wireless communication), a beacon, etc. are used in combination.

[0052] In one example, a camera and ultrasonic waves are used in combination. This allows the system to determine information that cannot be obtained using the Doppler effect of ultrasonic waves, such as which shelf contains what items, or whether the shelf is empty. When both the camera and the ultrasonic transmitter / receiver are installed at the end of the aisle, 1) Ultrasonic sensors detect the shuttle's movement (speed and acceleration) and position. 2) The camera determines whether the shuttle is carrying an item. 3) The shuttle carrying the item moves and stops, and when the item moves left and right and the shuttle becomes empty, it is determined that the item has been stored on the shelf. The left and right movement of the item is captured by a camera. 4) If an item is placed on a previously empty shuttle, it is determined that the item has been taken from the shelf. 5) By repeatedly obtaining the information from 1) to 4) above, it is possible to determine which shelves are empty and which shelves are filled. 6) Furthermore, by reading the ID of each item with a camera, it becomes possible to manage each item individually.

[0053] In addition, cameras and various sensors can be installed not only in fixed locations such as the perimeter of racks, walls, pillars, ceilings, and other automated warehouses in logistics centers, but also in portable predictive maintenance systems, such as a box-shaped container containing cameras and sensors and a PC that collects data and communicates it with systems in the management department that manages the logistics center's WMS, WCS, and other systems. This system can be transported and moved within conveyor lines such as conveyors and sorters or automated warehouses to detect the transport status of material handling equipment such as conveyors, sorters, and automated warehouses, and inspect the wear and tear of these components. In this case, real-time data collection utilizes log data from sensors pre-installed in the conveyor lines and automated warehouses, and decisions are made based on the content of this log data. In such a system, only the cameras and sensors required for the portable system are required, making it possible to build a compact and efficient predictive maintenance system.

[0054] When the input to the WCS that controls the automated warehouse is, for example, to consolidate i item numbers X, j item numbers Y, and k item numbers Z at location O, and this is also input to the 3D simulator, a signal that controls the automated warehouse in the same way as the output from the WCS is output from the 3D simulator, and the virtual automated warehouse operates based on this control signal to move the items to the desired location, and a simulation is performed in which the time required for this is equivalent to that of an actual automated warehouse.

[0055] For example, tag information such as product name, manufacturer name, quantity, and item category (stationery, electrical appliances, books, clothing, etc.) is managed by ID.

[0056] By running the 3D simulator in parallel with the actual movement of the logistics center, the 3D simulator is verified by comparing it with real data.

[0057] By using a 3D simulator to collect learning data for the control AI of an actual logistics center and to train the AI, it is possible to collect a large amount of learning data that will enable the AI, especially deep learning, to perform control efficiently.

[0058] This allows for efficient control of real logistics centers and can also be used for predictive maintenance.

[0059] For the predictive maintenance service, the system is configured to include sensors installed on elevators and / or structures that structurally support the elevators and / or components related to the structures in a logistics center, which acquire sensing information related to physical quantities including vibrations in real time or at regular time intervals; a wireless transmission unit that transmits the sensing information acquired by the sensors via wireless; a receiving unit that receives the sensing information transmitted from the wireless transmission unit; and a 3D simulator that predicts the replacement time for components of the elevators and / or the structures based on the sensing information received by the receiving unit.

[0060] The 3D simulator may predict the replacement time of the lifting equipment parts from the sensing information using artificial intelligence (AI) technology.

[0061] According to this configuration, the main challenge in developing a sensor was to create one that could be easily installed in existing equipment. In this application, to reliably detect machine anomalies, a method is adopted in which abnormal data is detected by comparing it with normal data, rather than simply determining anomalies based on differences in collected data. Alternatively, a method may be adopted in which abnormal data is detected by comparing it with normal data that has been highly modeled using AI machine learning. Furthermore, by using high-performance, highly reliable wireless communication, a low-cost, high-performance IoT sensor for predictive maintenance can be realized. Using the newly developed IoT sensor, a "predictive maintenance service" can visualize equipment operating status and predict anomalies, optimizing part replacement timing, making "uninterrupted logistics" possible.

[0062] The equipment parts replacement prediction system has a configuration that predicts the replacement time of the equipment parts by artificial intelligence (AI) technology using the output from sensors installed in the equipment. More specifically, the system can be embodied as an equipment parts replacement prediction system that predicts the replacement time of the parts by installing vibration sensors on the equipment of an automated warehouse and using artificial intelligence (AI) technology to create a trained model of the relationship between the amount of change over time in the frequency of vibrations generated from the parts that make up the equipment and the abnormality level of the parts.

[0063] In this application, the term "equipment" refers to equipment that assists in transportation and loading and unloading operations, and in particular, is a concept that includes not only the above-mentioned transport devices, lifting devices (lifters), and buffer conveyors, but also forklifts, pallets, and general conveyors as work machines used to improve the efficiency of logistics operations.

[0064] The facility may be embodied as an elevator device installed in the storage warehouse section of the automated warehouse.

[0065] The facility may be realized as a stacker crane type transport device installed in the storage warehouse section of the automated warehouse.

[0066] The sensor may be a vibration sensor that detects vibrations of the equipment, and the artificial intelligence technology may be configured to create a learned model of the relationship between the amount of change over time in the frequency of the vibrations of the equipment and the degree of abnormality of the component.

[0067] The sensor may be configured to analyze vibrations generated from a plurality of components constituting the facility using an FFT analyzer, thereby decomposing the vibrations into frequency components of the components constituting the facility.

[0068] The vibration sensor may be embodied to be attached to a drive train of the equipment.

[0069] The drive system may be embodied as one including a configuration including active rotation means and passive rotation means and fixing means for fixing them, for example, a motor and pulleys, and bolts for fixing them.

[0070] The sensor may be configured to be retrofitted to existing equipment that does not have the sensor installed.

[0071] The method for predicting replacement of equipment parts includes the steps of: detecting vibrations generated from the equipment using a vibration sensor installed in the equipment; analyzing the vibrations using an FFT analyzer to resolve the vibrations into frequencies; and predicting the replacement time of the parts using artificial intelligence (AI) technology that uses a trained model of the relationship between the amount of change in the frequency over time and the degree of abnormality of the equipment.

[0072] We discovered that this was due to the extremely large total number of components that make up a logistics center, resulting in a so-called combinatorial explosion. Therefore, instead of counting all possible combinations, we came up with a method in which we treat the features describing logistics center attributes and logistics center status, as well as the entire transfer instructions, as patterns, input them into a machine learning method such as an artificial neural network (ANN), and have it train to perform a simulation tailored to the objective. This machine learning method learns the features representing logistics center attributes and status as patterns and outputs a simulation signal tailored to the objective. A new transfer command is then input into the machine learning method, which then outputs a transfer simulation signal. In this way, by learning patterns tailored to the target logistics center, it becomes possible to use the same machine learning method with different attributes to simulate logistics centers with different attributes.

[0073] This is possible because a logistics center has a large number of components, but only a small number of component types, and the status of each component can be obtained automatically. First, because there are only a few component types, it is possible to describe all attributes for each component type. Meanwhile, the status of a logistics center during operation—that is, the status of the components within the center (such as the availability of each entrance and the current location and availability of loading and unloading vehicles)—changes constantly, and the number of combinations is enormous. However, if all of this information can be collected, combined, and processed as necessary to input it into a machine learning algorithm as features, it becomes possible to simulate the logistics center. Considering that the status of a logistics center can be automatically obtained from sensors installed within the center, it is clear that the features to be input into the machine learning algorithm can be automatically obtained. In other words, a vast number of features obtained from attributes and status can be automatically obtained and used as patterns to input the machine learning algorithm. One aspect of the present invention was conceived as a result of focusing on the following characteristics of machine learning means including ANNs, which can avoid the combinatorial explosion that occurs when counting everything by treating them as patterns; the characteristics that although the total number of components is large, the types are few and therefore describable; the characteristics that the situation can be obtained automatically and therefore the features can be obtained automatically; and the characteristics of automated multi-story warehouses, which can treat these as patterns.

[0074] However, when using an ANN as a machine learning method, it is still difficult to simulate all types of distribution centers with the exact same ANN. For example, if the total number of components is different, the number of input nodes in the ANN must be changed. If the number of input nodes is different, the number of hidden layers and the number of nodes in each hidden layer must also be changed. Meanwhile, the number of output nodes differs depending on the type and number of simulation targets. Therefore, the number of input nodes, hidden layers, the number of nodes in each hidden layer, and the number of output nodes must be changed to match the attributes of each distribution center. To meet this requirement, it is necessary to set the above ANN variables to match the attributes of the distribution center being simulated. However, since many of these settings require know-how, adjustments must be made while referring to the learning results. Therefore, these settings must be made via a GUI screen, allowing the user to set them while referring to the learning results.

[0075] On the other hand, learning requires a large amount of operational data for the logistics center being simulated. However, this is not always possible. It is particularly difficult to prepare data for newly designed logistics centers because there is no operational history. To solve this problem, one embodiment of the present invention builds a 3D simulator of the logistics center, performs virtual operations, and uses the resulting data as learning data for the machine learning means. To achieve this, one embodiment of the present invention can also include an interface for inputting data from the 3D simulator into the machine learning means. Learning can be performed by inputting attribute and status data for the logistics center being simulated, as well as various transfer command data, from the 3D simulator through this interface.

[0076] However, when using ANN as a machine learning method, what is input is generally an important parameter that determines performance, and selecting it requires advanced know-how. Therefore, it is possible to obtain better results by processing data from a logistics center or its 3D simulator rather than inputting it as is. However, know-how plays a major role in determining which logistics center data to input and how to process it. Therefore, one aspect of the present invention provides a GUI interface for selecting and processing input parameters, and a function for selecting or processing input parameters while referring to the learning results.

[0077] The role of the 3D simulator is not limited to preparing learning data for the machine learning means. In the above learning, the data on logistics center attributes and logistics center status output from the 3D simulator is input into the machine learning means, which then outputs a simulation signal for the logistics center, and the logistics center can be virtually operated based on this output. The machine learning means outputs inefficient simulation signals in the early stages of learning, but as learning progresses, it learns to output efficient simulation signals. In either case, forming a loop between the 3D simulator and the machine learning means enables the machine learning means to quickly learn efficient simulations.

[0078] The method uses machine learning, including an artificial neural network (ANN), which is characterized by performing general-purpose simulations of logistics centers independent of logistics center attributes by treating the entire information, including as patterns, at least one of the following: the number of storage shelves and their layout, the number of tiers per storage shelf, the number of openings which are the unit of item storage on each tier, the type and performance of the transfer means for moving items, and the number and layout of those; and transfer commands, including at least one of the following: the current number of occupied openings, the number of available openings and their addresses, the number of available transfer means, the current addresses of each transfer means, and which items, how many, and to which destination.

[0079] The machine learning may be performed by at least one of ANN including deep learning, reinforcement learning, Monte Carlo tree search, and Rapidly Exploring Random Tree (RRT).

[0080] When an artificial neural network is used as a means for performing the machine learning, attributes of the artificial neural network including at least one of the number of input nodes, the number of hidden layers, the number of nodes in each hidden layer, the number of output nodes, and the connection topology between layers are set; or when a method including reinforcement learning, Monte Carlo tree search, or RRT is used, attributes of the logistics center including at least one of the number of shelves, tiers, entrance layout, and number of transfer means are set, and the structure of the machine learning means is set on a GUI screen before learning, and learning is performed on it, and a simulation is performed using the trained machine learning means.

[0081] A simulator may be used to simulate the operation of a target logistics center related to the logistics center, and the output from the simulator may be used as learning data and / or teacher data or reward for the machine learning means.

[0082] During the learning related to the machine learning, a simulation may be performed using an output from a machine learning means, and the resulting output of the simulator may be input to the artificial neural network to perform learning.

[0083] All or part of the parameters describing the distribution center attributes, distribution center status, and transfer instructions may be input to the machine learning means as is and / or after processing.

[0084] The current attributes and status of the logistics center and transfer instructions may be input to a machine learning means that has been trained to perform the desired simulation, and the logistics center may be simulated using the output of the resulting artificial neural network.

[0085] The status data of the logistics center to be input into the trained model may include at least one of the following: item occupancy information and availability information on shelves in the logistics center, stored item information, and the current position of the transfer means, and the transfer command may include at least one of spatial information of the source and destination and information on the items being transferred.

[0086] From the input information, a simulation parameter including at least one of the transfer order and time, and the transfer route address may be output so that the logistics center operates in accordance with the purpose.

[0087] The system may further include at least one of a GUI (graphical user interface, etc.) for setting the structure of a machine learning means including an artificial neural network, reinforcement learning, Monte Carlo tree search, and RRT.

[0088] The device may further include a GUI for switching whether or not the output from the machine learning means is used in the simulation of the 3D simulator.

[0089] The 3D simulator is also configured to include a first input section into which logistics center attributes are input, including at least one of the number of storage shelves and their layout, the number of tiers per storage shelf, the number of openings which are the unit of item storage on each tier, the type and performance of the transfer means for moving items, and the number and layout of those means; a second input section into which transfer commands are input, including at least one of the current number of occupied openings, the number of vacant openings and their addresses, the number of available transfer means, the current addresses of each transfer means, and which items, how many, and to which destination; and a learning section that performs machine learning, including an artificial neural network (ANN), by treating all of the information obtained from the first and second input sections as a pattern.

[0090] The machine learning may be performed by at least one of ANN including deep learning, reinforcement learning, Monte Carlo tree search, and Rapidly Exploring Random Tree (RRT).

[0091] When an artificial neural network is used as a means for performing the machine learning, attributes of the artificial neural network including at least one of the number of input nodes, the number of hidden layers, the number of nodes in each hidden layer, the number of output nodes, and the connection topology between layers are set; or when a method including reinforcement learning, Monte Carlo tree search, or RRT is used, attributes of the logistics center including at least one of the number of shelves, tiers, entrance layout, and number of transfer means are set, and the structure of the machine learning means is set on a GUI screen before learning, and learning is performed on it, and a simulation is performed using the trained machine learning means.

[0092] A 3D simulator may be used to simulate the logistics center, and the output from the 3D simulator may be used as learning data and / or teaching data or reward for the machine learning means.

[0093] During the learning related to the machine learning, a 3D simulator may be simulated using an output from a machine learning means, and the resulting output of the simulator may be input to the artificial neural network to perform learning.

[0094] All or part of the parameters describing the distribution center attributes, distribution center status, and transfer instructions may be input to the machine learning means as is and / or after processing.

[0095] The current attributes and status of the logistics center and transfer instructions may be input to a machine learning means that has been trained to perform the desired simulation, and the logistics center may be simulated using the output of the resulting artificial neural network.

[0096] The status data of the logistics center to be input into the trained model may include at least one of the following: item occupancy information and availability information on shelves in the logistics center, stored item information, and the current position of the transfer means, and the transfer command may include at least one of spatial information of the source and destination and information on the items being transferred.

[0097] From the input information, a simulation parameter including at least one of the transfer order and time, and the transfer route address may be output so that the logistics center operates in accordance with the purpose.

[0098] The system may further include at least one of a GUI (graphical user interface, etc.) for setting the structure of a machine learning means including an artificial neural network, reinforcement learning, Monte Carlo tree search, and RRT.

[0099] The device may further include a GUI for switching whether or not the output from the machine learning means is used in the simulation of the 3D simulator.

[0100] A typical application of reinforcement learning is route discovery. In reinforcement learning, a lifelike agent is imagined within the software, and the agent searches for a route. For example, a starting point and a goal point are given, and various routes exist between them. Some of these routes connect the two points, but some do not reach the goal. Among the routes to the goal, some are quick and others are long routes. Each route connecting the starting point and the goal point has multiple milestones (coordinates), and the agent transitions from its current coordinates to the next coordinate, receiving a reward as a result. If it reaches the goal in this way, it receives a large reward for the route it has taken up to that point. By repeating this learning process, the agent will eventually be able to proceed from the starting point to the goal without getting lost. In this way, the agent finds the route that should lead to the goal.

[0101] Although different from the state transitions in route planning, failures also involve state transitions. Failures do not occur suddenly; rather, they are thought to progress from a normal state through multiple states before finally failing. For example, before a motor stops rotating, there may be warning signs such as poor rotation or instability. Or, in the case of a component within a device made up of multiple parts, the malfunctions of the multiple parts may gradually become more severe, eventually leading to the failure of the component itself. In this way, parts and components will likely go through various malfunction states from a normal state until they finally fail. Therefore, if we consider the normal state as the starting point and the malfunction state as the goal, and associate the various states of each part between them with milestones in route planning, reinforcement learning can be used for failure detection and prediction. [Effects of the Invention]

[0102] According to the aspects of the present invention, instead of the conventional "reactive maintenance" that responds after a problem occurs, or the "preventive maintenance" that monitors the situation during regular inspections and replaces parts, it is now possible to perform "predictive maintenance" that uses the latest AI (artificial intelligence) and IoT (Internet of Things) to predict signs of failure, thereby realizing "uninterrupted logistics." [Brief explanation of the drawings]

[0103] [Figure 1] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 2A] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 2B] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 2C] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 2D] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 3] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 4A] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 4B]FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 4C] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 4D] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 5] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 6] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 7] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 8] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 9] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 10] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 11] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 12] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 13] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 14] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 15] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 16] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 17] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 18] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 19] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 20] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 21] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 22] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 23] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 24]FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 25] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 26] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 27] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 28] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 29] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 30] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 31] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 32] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 33] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 34] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 35] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 36] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 37] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 38] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 39] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 40] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 41] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 42] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 43] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 44] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 45] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 46]FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 47] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 48] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 49] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 50] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 51] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 52] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. [Figure 53] FIG. 1 is an explanatory diagram according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0104] Hereinafter, an embodiment of the present invention will be described with reference to FIG.

[0105] In the following, the scope necessary for the explanation to achieve the object of the present invention will be shown schematically, and the scope necessary for the explanation of the relevant parts of the present invention will be mainly explained, and the parts where explanation is omitted will be based on publicly known technology.

[0106] As mentioned above, in logistics centers where further evolution is required even in stable operation, the present application aims to systematically realize "predictive maintenance" that predicts signs of failure, rather than "reactive maintenance" that responds after a problem occurs, or "preventive maintenance" that monitors the situation during regular inspections and replaces parts, as has been commonly done in the past for the maintenance of machinery and equipment. As one means for achieving this, the inventor has worked on developing "predictive maintenance" technology that predicts signs of failure using the latest AI (artificial intelligence) and IoT (Internet of Things).

[0107] The main challenge in the development was to adopt a sensor that is highly accurate and can be easily installed on existing equipment. In order to reliably detect abnormalities in machinery, a method was adopted in which abnormal data was detected by comparing it with normal data that had been highly modeled using AI machine learning, rather than simply determining anomalies based on differences in collected data. Furthermore, by leveraging the applicant's extensive technology and experience in logistics equipment, a low-cost, high-performance IoT sensor for predictive maintenance was developed that uses high-performance, highly reliable wireless communication. Using the newly developed IoT sensor, a "predictive maintenance service" can visualize the operating status of equipment and predict abnormalities, thereby optimizing the timing of part replacements, making "unstoppable logistics" possible.

[0108] In one embodiment of the present application, a 3D simulator is used to bring further advances to logistics centers. This point will be described in detail.

[0109] As seen in the development of predictive maintenance services, the applicant has been actively working on 3D simulators for logistics centers. As we have entered the COVID-19 era, the applicant recognizes that its efforts in 3D simulators are more significant than ever before. One such initiative is the development of GTR (Goods to Robot: Robot Pick), which aims to improve the efficiency of the entire logistics center and further expand contactless work by incorporating actual work data, such as inbound and outbound warehousing data, into the layout and human traffic of the logistics center and deriving the optimal layout. Depending on the products handled, this may even enable full automation of the logistics center.

[0110] Furthermore, 3D simulators are becoming increasingly important in achieving the Sustainable Development Goals, and through the development of new technologies, they can contribute to solving problems that the applicant defines as the most important social issues, such as reducing environmental impact and addressing the drastic decline in the working population. One example of this effort is the development of a control platform that provides integrated control of AGVs (Automatic Guided Vehicles: unmanned transport robots), whose operation methods vary from manufacturer to manufacturer. The transition to a conveyorless system will make it easier to respond to environmental changes.

[0111] Next, the significance of introducing wireless communication in one embodiment of the present application will be considered, and the concept of a wireless system will be explained in itemized form.

[0112] We will consider the significance of introducing wireless communication and explain the concepts of wireless communication methods in bullet points.

[0113] · Predictive maintenance requires the installation of a large number of sensors (100 to 10,000 or more). A transmission line is required to transmit the signal from the sensor to the server. Even if the transmission lines are consolidated along the way, multiple receiving devices are required on the receiving side, such as the server. (1) Large warehouses and factories with numerous pieces of equipment require long-distance wiring using transmission lines. However, in warehouses and factories, there are many noise sources such as power equipment, which can interfere with long-distance sensor signal transmission, especially when the sensor is located close to power lines. (2) On the other hand, there are situations where sensors must be installed after the equipment is in operation. In this case, wiring the sensor signal transmission line is not easy. (3) On the other hand, it is also necessary to transmit sensor signals from moving objects within the factory space. · When the movement of a moving object is irregular or cannot be determined at the time of design, transmission via wires is difficult. ···(4) Even if the range of movement is limited, there are cases where the distance traveled is virtually infinite, such as in rotational motion, and in this case too, transmission via wires is difficult. (5) Particularly for reasons (1) and (2), transmitting sensor signals via wires is costly. In particular, for reasons (1) to (5), wired sensor signal transmission is nearly impossible. To solve the above problems with wired sensor signal transmission, a sensor device including a power source is installed at the target for signal acquisition. The use of wireless transmission allows for flexible support for a large number of sensors, retrofit sensors, an increased number of sensors, and any type of moving object. An increase in the number of sensors can be accommodated simply by changing the software settings. To achieve wireless transmission that is resistant to noise, noise-resistant signal processing is applied to the wireless transmission. · It gives you more freedom in installing sensors.

[0114] The key points of one aspect of the present invention are as follows. Point 1: This relates to a highly flexible means of signal transmission between objects that move only limitedly within short distance spaces such as warehouses and factories. Point 2: In the case of the object described in Point 1, which has a limited movement such as rotation but a movement distance that can be considered infinite, communication is performed wirelessly with the object. Point 3: A means of communicating with the outside world will be adopted in a situation where the sensor installation location can only be determined after the equipment has been installed. Point 4: Including a vibration sensor for fault detection, wireless communication with the target is superimposed on the means of point 3 above for signal detection that depends on the equipment after installation, which differs for each design. Point 5: In addition to the measures mentioned in points 1 and 3, in industrial facilities, including factories and warehouses where high-power equipment is installed, noise-resistant transmission methods, including PCM, FSK, PSK, and frequency spreading, should be used.

[0115] The following is a bulleted explanation of the forms using the 3D simulator.

[0116] (1) The 3D simulator constructs and displays a 3D layout of an actual logistics center on a computer. (2) In (1), moving objects such as shuttles and conveyor belts are moved. (3) 3D Display what is stored and what is not stored on which shelves in the logistics center. (4) Navigate inside a 3D logistics center. (5) A histogram is displayed showing the time from when the transfer command is issued to when the transfer is completed (time efficiency check). (6) A histogram of the power consumed from the time the transfer command is issued until the transfer is completed is displayed (power efficiency check). (7) Display of cumulative data on the operating time and travel distance of each component (e.g., each shuttle, each lifter, etc.) (to identify correlations with failures) (8) Based on the image data displayed on the 3D logistics center, we will intervene in the operations of a real logistics center and try to improve efficiency through human intuition. (9) Change various parameters of the 3D logistics center to test the changes in efficiency. 10) Change the hardware parameters of the 3D logistics center and observe changes in failure frequency, etc. 11) Observe the changes in 10) when various parameters are changed and past transfers are recreated. → Identify the key points for improving efficiency. 12) Show operational performance such as number of items processed per hour and processing time per item.

[0117] Next, the concept of the shape using the 3D simulator will be explained in bullet points.

[0118] 1) To create a highly accurate 3D simulator, we first model the components that make up the actual logistics center, and then combine them to create the 3D simulator. At this time, the units used vary depending on the characteristic component units of the logistics center (for example, focusing on failures, procurement units, or factory production units). 2) When modeling the subcomponents that make up a component, the more granular the sub-subcomponents are, the more accurate the simulation becomes. However, this requires a huge amount of time and effort to build. This granularity does not need to be consistent across the entire logistics center (for example, some components may model bearings, while others may not). 3) Two modes are available: real-time mode and high-speed mode. The real-time mode is used for comparison and verification with a real warehouse, and the high-speed mode is used for A1 learning to speed up learning.

[0119] Actual logistics center to The occurrence of a fault in the simulator is correlated with the operation of the 3D simulator, and the operation history of the 3D simulator is used as raw data for predicting the occurrence of a fault.

[0120] The raw data may be acquired in component units.

[0121] This article will explain the concepts of AI learning and data collection for control AI in a real logistics center. In particular, it will explain Deep Learning (DL).

[0122] A large amount of learning data is required for DL ​​to be able to perform control efficiently. By issuing various transfer commands to the 3D simulator, it is possible to know the time and power required to move an item. If this is used as learning data for DL, DL can be operated at high speed. efficiency It is expected that the system will learn good control of the

[0123] Furthermore, it is possible to introduce other AI methods in addition to DL. In addition, because the 3D simulator models are a combination of components, when designing a new logistics center, a customized 3D simulator can be constructed simply by combining the necessary components, allowing for quick response.

[0124] Furthermore, it is possible to perform predictive maintenance while operating the 3D simulator. To do this, the 3D simulator is used to compare basic data such as failure probability and MTBF (mean time between failures) with actual operating time and travel distance. Instead of displaying actual operating time, it can also be displayed as relative time or on any scale.

[0125] The acquired failure data is used as training data for predictive maintenance for DL ​​learning, allowing the AI ​​to learn the correlation between the operating time and distance of each component and failure.

[0126] 2A to 2D show examples of a logistics center according to an embodiment of the present invention. The warehouse has the functions of receiving, storing, sorting, and shipping. FIG. 2A is a perspective view of an automated multi-story warehouse as seen from the shipping side. In this example, there are eight rows of shelves 1, each with seven levels, each with several dozen openings, each of which holds trays of items. Received items 2 are placed into an empty opening on shelf 1. These received items may remain at the receiving opening until a shipping command is received, or they may be moved between shelves, levels, or openings depending on the situation in the warehouse. When a shipping command is received, the items are moved to the shipping end of each shelf by an item transfer means (not shown), and in some cases are placed on a standby conveyor 3, where they wait until an elevator (not shown) arrives to transport them to a lower floor where they will be transferred to conveyor 4. When the elevator arrives, it moves to a lower floor and transfers the items 5 onto the conveyor 4, where they arrive at the GTP (Goods To Person) station 8, which removes the items from the tray and transfers them onto a shipping tray 7 on an outgoing conveyor 6. At the GTP station, a worker 9 picks the required number of items carried by the conveyor and places them into the shipping tray 7. This work may be performed automatically by a robot arm 10. Once all the items have been picked, the shipping tray is placed on the outgoing conveyor 6, transported to the outgoing station, and transferred onto a truck. Note that instead of using a belt conveyor to transport items, a roller conveyor can also be used, and this applies to all of the cases described below.

[0127] As another example of an application of the technical concept of the present application, a digital picking system is shown in Figure 2B. In this system, items to be shipped are placed on shipping trays 11, which move horizontally on conveyor 12. Each picker 13 picks the required number of items to be shipped from the shelf 14 for which he or she is responsible, and places the items on the shipping tray 11 moving on the conveyor in front of them. To ensure this is done correctly, shipping instruction displays 15 may be installed in front of each entrance of shelf 14.

[0128] Another example of an application of the technical concept of the present application is shown in Figure 2C. In this example, numerous small shelves 16 are placed on the warehouse floor, and items are stored on them. The bottom of the shelves 16 is raised above the floor, allowing a thin AGV (Automatic Guided Vehicle) 17 to enter. A close-up of the shelves 16 and AGV 17 is shown within the arc. The AGV 17 is equipped with a lifter, which rises when it enters the center of the shelf 16, lifting the shelf 16 and transporting it to its destination. This is a different method of transferring items from Figures 1A and 1B. Since items are stored on several low-height shelves, the items are stored more flatly. While AGVs require markers (typically installed on the floor) to indicate their location, the use of an Autonomous Moving Robot (AMR) can eliminate the need for markers on the floor or elsewhere.

[0129] Another example of an object to which the technical idea of ​​the present application is applied is shown in Figure 2D. In this example, a stacker crane 19 is used as a transfer means to access items stored on shelf 18. The stacker crane can access the entire front of the shelf by carrying the item and moving it horizontally and vertically, making it possible to take the item in and out. The items placed on the stacker crane are then transferred to a conveyor or the like for storage, transfer, or removal.

[0130] As such, logistics centers have a variety of storage means and layouts, and in addition, goods transfer means include belt conveyors, roller conveyors, lifting devices, self-propelled electric mobile carts, stacker cranes, AGVs / AMRs, etc., and various picking means are used, such as manual methods and robotic arm methods.One embodiment of the present invention is directed to a logistics center control technology that uses machine learning means to obtain the desired control for a target logistics center, instead of designing and building control software for a wide variety of logistics centers individually for each target logistics center.

[0131] ANNs, such as deep learning, treat data to be processed as patterns instead of logically describing them as input and produce the desired output. In one embodiment of the present invention, the data to be processed includes warehouse attributes (the number and layout of storage shelves, the type, number, and performance of transfer vehicles, etc.), warehouse status (the type and number of stored items and their storage locations, which shelves are available, which transfer vehicles are available, the current location of the transfer vehicles, etc.), and transfer instructions (which items to transfer, how many, and to which destinations). On the other hand, the output information may include, but is not limited to, transfer time or transfer route when controlling to minimize transfer time or power consumption. For example, when controlling to shorten delivery time, learning can be performed to output a shorter delivery time. Alternatively, when controlling to reduce the operational power of a logistics center, learning can be performed to reduce power consumption. In addition to these, it is possible to perform targeted control by performing learning according to each purpose, such as leveling the transfer load at the time of retrieval, minimizing noise, shortening the time it takes to enter the warehouse, etc. The type of learning to be performed can be determined by selecting the input signal and / or teacher signal / reward according to the control purpose.

[0132] In this way, machine learning can learn in accordance with the distribution center being controlled. This characteristic of machine learning makes it possible to perform general-purpose control without relying on the attributes of the distribution center being controlled. Machine learning can use ANNs such as deep learning, as well as, but is not limited to, reinforcement learning, Monte Carlo tree search, RRT (Rapidly exploring Random Tree), or a combination of these.

[0133] As described above, learning data is necessary to control a logistics center using machine learning techniques. However, there is insufficient data for training the ANN in a newly designed or pre-delivered logistics center. Therefore, a simulator for the target logistics center is constructed. The simulator simulates the behavior of the target logistics center. It can represent various logistics center conditions on a computer, such as the number and layout of storage shelves, the type, number, and performance of transfer means, the type and number of stored items, and their associated behavior, as well as which shelves are available, which transfer means are available, and the current location of the transfer means. By providing various commands (e.g., transfer commands including, but not limited to, inbound and outbound commands), the simulator can mimic the behavior of an actual logistics center. Such a simulator can be constructed, and a large amount of learning data can be generated using it. Furthermore, the simulator can be used to control the target logistics center constructed using the output of the machine learning technique during learning. By providing rewards or training signals to, for example, shorten outbound times, the machine learning technique can be trained. One embodiment is shown in Figure 3. Note that from Figure 2 onwards, the explanation will be mainly focused on a warehouse with the structure of Figure 2A. However, since the transfer of goods involves moving them from source coordinates to destination coordinates, and this principle applies to logistics centers in general, those skilled in the art will understand that the generality of the present invention will not be lost even if the explanation is mainly based on the structure of Figure 2A.

[0134] In Figure 3, the warehouse simulator 21 (3D simulator) represents, on a computer, storage location attributes such as warehouse storage type (e.g., shelf storage, flat storage), number of shelves, number of shelf tiers, number of entrances, shelf coordinates, tier coordinates, and entrance coordinates, as well as transfer method attributes such as transfer method type (e.g., electric mobile cart, stacker crane, conveyor, AGV / ARR, etc.) and number of vehicles, travel speed, maximum load weight, and current coordinates. Alternatively, a state where no items are stored can be set as the initial state 23 and items are placed there, or a state where a certain number of items are stored can be set as the initial state. A transfer command 26 is input from an external device to the machine learning means 22, and warehouse attributes 24 and warehouse status 25 are input from the warehouse simulator 21, and a control output 27 is output. Initially, since no learning has been performed, the system is unable to output a control signal consistent with the objective. However, as learning progresses by providing rewards and teacher signals 28 consistent with the control objective, the system begins to output appropriate control signals.

[0135] Figure 3 shows an example in which the control output of the machine learning means is input to a warehouse simulator and used to control the warehouse. This allows the machine learning means to progress in learning by integrating it with the warehouse simulator. However, as an alternative, simple pre-built control software can be used to perform minimum control such as warehousing, transfer, and retrieval, while the machine learning means learns in parallel. In this case, once the learning of the machine learning means has progressed, it can take over control of the warehouse simulator from the pre-built control software.

[0136] 4A to 4D show an example of machine learning means according to one embodiment of the present invention. FIG. 4A shows an example using an ANN 30 incorporating deep learning. Warehouse attributes 24 (e.g., number of shelves, layout, number of shelves, number of bays, type, performance, and number of transfer means), warehouse status 25 (e.g., number of occupied bays, number of vacant bays, their addresses, and number and addresses of available transfer means), and transfer instructions 26 (e.g., which items, how many, and to which exit gates) are input to the input layer 31. The connections between the input layer 31, hidden layer 32, and output layer 33 form generalized patterns that learn various warehouse statuses and output results in line with the control objective. The output layer 33 outputs predicted parameter values, such as transfer time and transfer power, according to the control objective, as well as control signals 27, such as transfer paths. For example, but not limited to, when training to shorten transfer times, the transfer times output from deep learning are stored, and learning is performed to further shorten the transfer times in the next output, and the transfer paths at that time are used to control the warehouse. In FIG. 4A, the input layer, hidden layer, and output layer are arranged one-dimensionally, but the arrangement is not limited to this and may be two or more dimensions.

[0137] In Figure 4A, warehouse attributes and warehouse conditions are patterned and input into the ANN, but Figure 4B shows one embodiment in which a successful example 34 of a transfer path is input as a pattern to the ANN for learning. In this example, an item to be transferred is present in the opening indicated by start 35, and is transferred to the opening indicated by end 36, and this is a successful path that is considered to have the shortest time as a transfer path. This method patterns examples of successful paths for warehouse conditions represented by various available openings and available transfer means, as well as various starts and ends, and has the ANN shown in Figure 4A learn many of these patterns.

[0138] 4C shows an example using reinforcement learning according to yet another embodiment of the present invention. This shows a case where there are five shelves 37, A to E, and the grid 38 of each shelf is the entrance, and its coordinates are represented by (Ax1, Ay1) to (Axn, Ayn), (Ex1, Ey1) to (Exn, Eyn). Consider the case where an item stored in the light gray entrance (Ax2, Ay2) of shelf A is transferred to the dark gray entrance (Ex3, Ey3) of shelf E. Figure 4CIn the figure, thin lines indicate paths from all openings to the next open opening on the shelf, and a route cost (e.g., the time required for travel) is set or defined for each path. This allows the time required for various routes from (Ax2, Ay2) to (Ex3, Ey3) to be evaluated. This time evaluation may use the unit path travel time, taking into account the current position of the transfer vehicle. Since some openings are actually blocked, this information is registered in advance as warehouse status for all openings on shelves A to B, and learning is performed to avoid blocked openings. Given the nature of item storage locations as discrete values, the reinforcement learning algorithm is not limited to this. Other methods that are used when both the state space and the action space are discretized include Q-learning, Deep Q-Networks (DQN), Cross Entropy Method (CEM), State-action-reward-state-action (SARSA), Deep SARSA, and Rapidly Exploring Random Tree (RRT). The path with the shortest time is selected from the paths evaluated in this way.

[0139] 4D shows an example of an RRT according to yet another embodiment of the present invention. The RRT tentatively determines the target address of the next step. For example, Figure 4DConsider the transfer of trays stored in the gray opening (A22) of shelf A to the dark gray opening (E33) of shelf E. Shelves B through D exist between them, and the first step is to transfer the tray to the opening (B31) of shelf B marked with an *. There are various possible paths, and a path is randomly selected from them to virtually transfer the tray. If an obstacle is encountered during the virtual transfer (e.g., the opening is already blocked by a tray and cannot be used), the path search is stopped and the search for the next path begins. The same search is repeated from the opening * of shelf B to the opening ● of shelf C, then from the opening ● of shelf C to the opening ▲ of shelf D, and finally from the opening ▲ of shelf D to the dark gray opening of shelf E. In this way, a path from the gray opening of shelf A to the dark gray opening of shelf E is searched for. The desired transfer route can be obtained by selecting the shortest path from the searched paths. Although FIGS. 4A to 4D show examples in which deep learning, reinforcement learning, and RRT are used independently, they may also be used in combination.

[0140] Figure 5 shows an example of a GUI screen 40 for setting the configuration of machine learning means for an individual logistics center according to one embodiment of the present invention. When using a deep learning type ANN 30 like the one shown in Figure 3A, it is necessary to set the number of input nodes, the number of layers in the middle layer and the number of nodes in each layer, and the number of nodes in the output layer according to the attributes of the warehouse to be controlled, and also determine the inputs and outputs to the ANN and the teacher signals according to the control objective, and set and adjust them to obtain optimal results. The GUI for doing this is shown in Figure 5.

[0141] Here we will show an example using deep learning. First, set the number of input nodes, output nodes, hidden layers, and nodes in each hidden layer of the deep learning ANN in the setting boxes for the number of nodes 41 and the number of layers 42 in Figure 4.

[0142] Next, set the feature quantities 43 to be input to the ANN. To do this, select the attributes, conditions, or transfer orders required to form each feature quantity from the column 44 of all attributes and conditions of the target warehouse displayed on the left side of the screen, and the contents of the transfer orders 45. To do this, use the "Sel / Math" button 46 to select at least one item from the warehouse attributes, warehouse conditions, or transfer orders. If you want to obtain feature quantities by performing a calculation from multiple warehouse attributes and transfer orders, you can set the necessary calculation after selecting the warehouse attributes. This calculation is not limited to mathematical calculations that can be described with equations, but can also be some kind of logical calculation represented by and / or.

[0143] On the other hand, the "Teacher Signal Setting" button 47 is used to set the teacher signal. By pressing this button, the teacher signal is selected and various settings are made. Furthermore, the output setting button 48 is used to assign meaning to the output node and make various settings. For example, when learning to shorten the departure time, at least one of the output nodes is set as an exit time node, and furthermore, the teacher signal is set to perform learning operation so that the value of this output setting node becomes smaller. Note that these settings can be operated at fixed values ​​once learning is complete, so they can be made in the early stages of ANN learning.

[0144] FIG. 6 shows an example of a GUI screen 50 when reinforcement learning is used as the machine learning method in yet another embodiment of the present invention. Since reinforcement learning searches for a path from the starting point to the goal, the topology of the search target space exists as a prerequisite. In this embodiment, the topology is the structure of the target warehouse, which is provided in advance as warehouse attributes 51 and stored on the computer as a warehouse representation 52 in the reinforcement learning method. Meanwhile, information on available space within the warehouse and the current location of the transfer means are input as warehouse status 53. A transfer command 54 is issued based on these attributes, and a transfer path is searched for based on these attributes. To shorten the outgoing time, the path with the shortest time is selected from the paths that reach the transfer shelf as a result of the movement, and this path is also output 55. In this case, a function 56 for storing and calculating the output of the machine learning method may be provided as a means for selecting the shortest path. Note that reinforcement learning may employ a learning method that rewards users retroactively from the point in time when a path search was successful during the learning process. While FIG. 5 illustrates an example of reinforcement learning, other methods, such as Monte Carlo tree search and RRT, can be used. A GUI appropriate for such methods can also be used.

[0145] FIG. 7 shows an example of a GUI 60 for setting an interface between a simulator and machine learning means according to yet another embodiment of the present invention. The GUI screen first graphically displays the warehouse configuration (in this example, there are seven shelves, each with three levels) in a warehouse simulator section 61. An overview 62 of the machine learning means is also displayed, along with a graphical display of an interface 63 through which warehouse attributes and warehouse status output from the warehouse simulator are input to the machine learning means. In addition to the above, the machine learning means also receives transfer commands 64 and reward / teacher signals 65, which output control signals 66 to the warehouse simulator, thereby controlling the warehouse. This control output may be turned on or off, and can be set using an ON / OFF button 67.

[0146] For example, when a delivery command is issued to deliver an item stored on the first level of shelf 1 to the third level of shelf 7 and input to the machine learning means 62, the machine learning means outputs a corresponding control signal. If this control signal is input to the warehouse simulator, the warehouse simulator starts virtual operation in response to the control signal 66, and graphically displays the delivery of an item stored in the gray opening on the first level of shelf 1 via the gray openings on shelves 2, 3, 4, 5, 6, and 7. In FIG. 7, the display may be such that the openings occupied by items and the open openings can be graphically distinguished by using different display colors, for example. This makes it possible to realistically display how items are transferred in response to the control signal from the machine learning means.

[0147] This display is important for the following reasons. As mentioned above, the configuration of the machine learning means (for example, in the case of an ANN, the number of nodes in the input layer, the number of hidden layers and the number of nodes in each hidden layer, and the number of output nodes), as well as the features input to the machine learning means and the method for calculating them, are largely dependent on know-how. Therefore, by observing the real movement of goods as described above, humans can intuitively grasp what features should be input to the machine learning means. This makes it possible to determine more appropriate features from warehouse attributes and warehouse conditions.

[0148] Figure 7 shows the warehouse configuration of Figure 2A, with a deep learning type ANN used as the machine learning method, but it goes without saying that these should be changed depending on the target warehouse configuration and the learning method used. Also, although a button for calculating feature values ​​is displayed on the same screen, this is optional. For example, clicking on the machine learning method switches to the phase shown in Figure 4 or Figure 5, where detailed feature value settings can be made. Meanwhile, while Figure 7 does not display a button for setting the output from the machine learning method, this may be included on the screen of Figure 7.

[0149] The automated warehouse 1 is composed of a WMS (warehouse management system) 400 equipped with a host computer (or server, cloud, etc.) that collectively manages warehouse operations, and a WCS (warehouse control system) that incorporates a computer connected to the WMS 400 host computer. It is managed by the National Instruments Control System (NIST) 500.

[0150] As shown in FIG. 9, the WCS 500 includes a computer equipped with a CPU, main memory, and external interface, and connected to the host computer of the WMS 400, and storage connected to this computer via the external interface.

[0151] Based on instructions from the WMS 400, the WCS 500 issues work instructions such as warehousing, transport, retrieval, picking, and automatic sealing to the material handling equipment (transporting device 30, lifting device 50, buffer conveyor 60, etc.) in the automated warehouse 1, the picking station 200, and the packaging means 300, etc., via wireless remote control or communication lines.

[0152] As shown in Fig. 10, each storage case 40 stored at any opening of the shelf 10 in the storage warehouse section 100 contains one or more product items 70 each having a first identifier 71 made up of a barcode. In addition, the exterior of the storage case 40 is provided with a second identifier 41 made up of a barcode that can be associated with the first identifier 71 of the product item 70.

[0153] The first identifier 71 is used to identify the product item 70 and is different for each product. The second identifier 41 is used to identify the storage case 40 and is different for each storage case 40. The first identifier 71 and the second identifier 41 may be configured as an identifier other than a barcode, such as a two-dimensional code or RFID.

[0154] The attribute data (quantity, date, destination, storage location, weight, etc.) of all product items 70 stored in the storage case 40 is stored in correspondence with the first identifier of the product item and the second identifier of the storage case 40 storing the product item, and is managed by the above-mentioned WMS400 and WCS500.

[0155] Although not shown in the figures, the logistics center equipped with the automated multi-story warehouse 1 is equipped with a truck berth where product items 70 from each manufacturer arrive. When the product items 70 from each manufacturer arrive at the truck berth, each product item 70 is unpacked from a packaging container such as a cardboard box and placed on a tray for storage in the storage warehouse section 100. The traying process is a process of linking (associating as described above) the first identifier 71 attached to each product item 70 with the second identifier 41 attached to the storage case 40.

[0156] The second identifier 41 attached to the storage case 40 is read when the storage case 40 passes in front of a barcode reader installed in the traying work area. The second identifier 41 is read using a known technique such as irradiating an LED light source and receiving the reflected light with a photodiode. Meanwhile, the first identifier 71 attached to the product item 70 is read by a handy scanner installed at the product input location of the truck berth every time the product item 70 is placed into the storage case 40.

[0157] The computer of the WCS 500 then creates a traying table based on the read first identifiers 71, their quantities, and the second identifiers 41, and stores this in a traying table memory area in the storage. This allows the WCS 500 to know the names and quantities of the product items 70 stored in each storage case 40.

[0158] The storage cases 40 for which the traying work described above has been completed are transported to the storage warehouse section 100 by the receiving conveyor 600 (see FIG. 8). One end of the receiving conveyor 600, which connects the traying work area with the storage warehouse section 100, is located above the picking station 200.

[0159] When storing the storage case 40 in the storage warehouse section 100, first, a storage command issued by the host computer of the WMS 400 is transmitted to the computer of the WCS 500.

[0160] The CPU of the WCS 500 grasps the names and quantities of the product items 70 to be stored in the storage warehouse section 100 by the above-mentioned storing command. The CPU of the WCS 500 compares the storage data table with the storing table based on a predetermined work instruction program stored in the hard disk drive, creates optimal storing order information, and issues work instructions to the material handling equipment (transport device 30, buffer conveyor 60, lifting device 50, etc.) of the storage warehouse section 100. Here, the storing order information is: Each storage case 40 This refers to the route and sequence that determines the order in which items will be stored and at which entrances.

[0161] At this time, the CPU of the WCS 500 refers to the inventory table and the traying table to confirm in which storage case 40 each product item 70 is stored, and then compares the storage information table with the traying table to determine on which shelf 10 and at which level each storage case 40 should be stored.

[0162] When determining the address (indicating the address of the entrance of the shelf 10 in the aisle 20) for storing the storage case 40, the address information of the storage case 40 already stored is read from the storage data table, and the storage case 40 is stored in the shelf 10 in the aisle 20 where the number of storage cases 40 is small.

[0163] Each storage case 40 for which a storage command has been issued moves to the shelf 10 in the aisle 20 in sequence in accordance with the operation of the material handling equipment, and is stored there. Then, the WCS 500 updates the storage data table in accordance with this storage data (address information of the storage case 40, time of storage, etc.).

[0164] When an order is placed by a customer of a store or the like for a product item 70 stored in any opening of the shelf 10 in this manner, the storage case 40 containing the product item 70 is removed from the opening of the shelf 10 by the conveying device 30 and transported to the picking station 200 via the buffer conveyor 60 and the lifting device 50.

[0165] When a product item 70 is to be shipped from the storage warehouse unit 100, a shipping command for the product item 70 is first issued from the WMS 400, and this command is transmitted to the picking station 200 via the computer of the WCS 500. The shipping command contains shipping information that associates the product item 70 with its quantity, shipping time, and delivery destination.

[0166] The CPU of the WCS 500 compares the storage information with the shipping information according to a predetermined work instruction program stored in the storage, creates optimal shipping order information, and operates the material handling equipment (transport device 30, buffer conveyor 60, lifting device 50, etc.).

[0167] Specifically, the WCS 500 selects a storage case 40 containing a product item 70 corresponding to a retrieval command according to a predetermined work instruction program stored in the storage, and moves the conveying device 30 to retrieve the storage case 40 from the shelf 10. When the conveying device 30 arrives at the storage location of the storage case 40 to be retrieved, it grasps the storage case 40 and transfers it to the buffer conveyor 60.

[0168] When a storage case 40 released from the storage warehouse section 100 arrives at the picking station 200, a picking section 90 such as a worker or robot picks the product item 70 according to instructions on a display screen that shows the product items 70 and their quantities in the storage case 40, and places the product items 70 into a shipping case (shipping medium) 80 (or packaging material such as cardboard) for the shipping destination that flows along the shipping conveyor 91.

[0169] When the product items 70 have been placed in the shipping case 80, the picking unit 90 touches the picking complete button displayed on the display screen. This causes the WCS 500 to transport the shipping case 80 toward the packing unit 300 in accordance with the predetermined work instruction program stored in the storage.

[0170] Of the storage cases 40 from which product items 70 have been picked at the picking station 200, those with product items 70 remaining inside are transported back to the storage warehouse section 100 by the buffer conveyor 60 and the lifting device 50, and are returned to the shelf 10 by the transport device 30. In this case, the location to which the storage case 40 is returned does not have to be the original location, and any available storage space on any shelf 10 can be used (free location method).

[0171] Meanwhile, once the picking of the product items 70 has been completed and the storage cases 40 are emptied, they are transported to a traying work area upstream of the storage warehouse section 100, and after the newly arrived product items 70 are stored therein, they are transported to the storage warehouse section 100 via the receiving conveyor 600.

[0172] In the packing section 300, the merchandise items 70 are removed from the shipping cases 80 sent from the picking station 200, packed in packing materials such as cardboard boxes, and then sent for shipping and delivery processing.

[0173] The shipping case 80 sent to the packing department 300 is checked for discrepancies between the shipping information read from the storage of the WCS 500 and information such as the product items 70 and their quantity inside the shipping case 80, customer information and store information for the shipping destination, and shipping date and time.The product items 70 inside the shipping case 80 are then removed and packed using an automatic sealing device or the like.

[0174] In the automated multi-story warehouse 1 where the above-described operations are carried out, if a malfunction occurs in the material handling equipment (conveying device 30, lifting device 50, buffer conveyor 60, etc.) in the storage warehouse section 100 and the transportation and storage / retrieval of product items 70 stops, the arrival time to customers may be delayed, resulting in significant losses.

[0175] In particular, among the material handling equipment in the storage warehouse section 100, the lifting device 50 is installed in a narrow space within the storage warehouse section 100. Therefore, when the material handling equipment operates 24 hours a day for an extended period, such as in the automated warehouse 1 at a logistics center, it is difficult for maintenance workers to enter the storage warehouse section 100 and perform regular inspections of the lifting device 50. Furthermore, the lifting device 50 is equipped with pulleys and belts that serve as a power transmission mechanism for vertically moving the platform 55 on which the storage cases 40 are placed up and down, but because the bearings built into the pulleys cannot be visually inspected from the outside, it is extremely difficult to manually detect the extent of any abnormality in the bearings or the need for replacement.

[0176] Therefore, a parts replacement prediction system for the lifting device 50 according to this embodiment will be described below with reference to the drawings.

[0177] FIG. 11 is a schematic diagram showing the main parts of the lifting device 50 installed in the storage warehouse section 100, and FIG. 12 is a schematic diagram showing an enlarged view of a part (near the upper pulley) of the lifting device 50 shown in FIG.

[0178] The lifting device 50 includes a drive system (driving components) such as a lower pulley 51 and an upper pulley 52 which are passive rotation means, a belt 53 suspended between them, a motor 54 with a reducer which is active rotation means that drives the rotation of the lower pulley 51, and a base 55 which moves up and down in the vertical direction in conjunction with the rotation of the upper pulley 52. ​​The lower pulley 51 and the motor 54 with a reducer are fixed onto a stage 57 by bolts 56a and 56b (fixing means), and the upper pulley 53 is fixed to a frame 59 by bolt 58 (fixing means).

[0179] Vibration sensors 2 that detect vibrations generated by each component are attached near the above-mentioned components (lower pulley 51, upper pulley 52, motor 54 with reducer, bolts 56a, 56b, 58) that make up the drive system of lifting device 50. A wireless vibration sensor that can automatically measure vibration levels periodically and transmit the measured values ​​to the 3D simulator via wireless communication is suitable as vibration sensor 2.

[0180] Each of the components constituting the lifting device 50 has its own resonance frequency (natural frequency). The vibration sensor 2 simultaneously measures vibrations (acceleration) in three axes (X, Y, and Z directions) generated from each component, and converts the complex vibration waveform resulting from the combination of all these vibrations into FFT (Fast Fourier Transform). By analyzing it with a Fast Fourier Transformation (FFT) analyzer, the vibration source is broken down into simple frequency components for each part.

[0181] Meanwhile, as each component constituting the lifting device 50 deteriorates over time (loosening in the case of bolts), the magnitude and frequency of the vibration generated by that component change. Therefore, by using the vibration sensor 2 to continuously measure the vibration generated by each component over a long period of time and observing the change in frequency over time for each component (the amount of deviation from the natural frequency), it becomes possible to estimate the component in which an abnormality occurred and the time when the component failed.

[0182] Figures 13 and 14 show examples of the frequency of each component, obtained by analyzing the vibration waveform measured by vibration sensor 2 with an FFT analyzer. In each figure, the horizontal axis represents frequency, which increases from left to right. The vertical axis represents the number of measurements (total number of measurements taken for 2 seconds each time), which increases from top to bottom. These figures show excerpts of specified ranges for frequency and number of measurements.

[0183] Figure 13 shows the frequency generated by a part in its initial state immediately after replacement, and Figure 14 shows the frequency generated by the same part after a predetermined time has passed since replacement. It can be seen that the frequency generated by the part after a predetermined time has passed (Figure 14) has deviated from the frequency generated by the same part in its initial state (Figure 13).

[0184] 15 is a graph illustrating the concept of the part replacement prediction system according to this embodiment. The vertical axis of the graph represents the amount of frequency fluctuation generated by the part as an abnormality level, and the horizontal axis represents time.

[0185] As the deterioration of each component constituting the drive system of the lifting device 50 progresses over time, the amount of frequency fluctuation (amount of fluctuation from the natural frequency) generated by that component changes, such as increasing. Therefore, the relationship between the amount of frequency fluctuation generated by each component and the degree of deterioration (degree of abnormality) of the component is learned in advance, and a threshold value is set for the amount of frequency fluctuation based on the learned model.

[0186] When the amount of frequency fluctuation of a certain part exceeds the threshold, the 3D simulator receives the measurement value of the vibration sensor 2 and automatically sends a warning email like the one shown in Table 1 to the terminal device of the logistics center operator via an information and communication network such as the Internet.

[0187] [Table 1] 16, the host device compares the amount of fluctuation in the frequency of the part (input value) with a threshold set based on the learned model described above to predict the time to replace the part (output value), and notifies the operator of the logistics center of the replacement time along with the warning email.The operator of the logistics center, having received the warning email and the notification of the replacement time, then instructs a maintenance worker to replace the part that needs to be replaced.

[0188] Figure 17 shows an example of a signal from the sensor. Figure 17A shows the amplitude signal detected by the vibration sensor, and Figure 17B shows an example of its Fourier transform. As shown in Figure 17A, in a normal state the amplitude is small and the period is constant, but when a slight malfunction occurs the amplitude increases and the period becomes slightly disrupted. If the condition worsens further, the vibration amplitude increases further and the period becomes more disrupted. An example of the period disruption in Figure 17 is shown in Figure 17B. When a malfunction eventually occurs, operation stops, the vibration disappears, and the amplitude becomes zero.

[0189] In this example, we show four milestones leading from a normal state to a fault state. However, if we increase the type of sensor and the resolution of data acquisition, we will notice that even more milestones exist. In this case, as shown in Figure 18, there are many milestones and routes between the normal start (S) and the fault goal (G). For example, the example in Figure 1 shows only data detected by a vibration sensor, but if acoustic sensors, temperature sensors, etc. are also present, many different data will be sent from them. It is possible to combine data from multiple sensors to create a sensor signal state space, that is, an n-dimensional space created by n features extracted from the analysis of sensor signals. In this case, a certain feature cluster is defined as the state of the equipment, and this state starts from the normal state (S) and passes through various routes and milestones to reach the fault state (G) ("1" to "20" in Figure 2).

[0190] In this way, by extracting features from data from each sensor, defining clusters in an n-dimensional feature space, and associating state transitions between clusters with milestones leading to failure, reinforcement learning techniques can be used for failure detection and prediction. Once the diagram in Figure 18 is obtained, it is easy to identify which milestone the equipment is currently in, and therefore, it is possible to estimate how close it is to failure. By obtaining the time required for the transition between each milestone, it is also possible to estimate the time required for failure from the current state.

[0191] Although the above example uses reinforcement learning, it is not limited to this. Among various machine learning methods, any method that can construct multi-step state transitions is within the scope of this patent. [Industrial Applicability]

[0192] As described above, according to the present invention, rather than "reactive maintenance" that responds after a problem occurs, or "preventive maintenance" that examines the situation during regular inspections and replaces parts, "predictive maintenance" that uses the latest AI (artificial intelligence) and IoT (Internet of Things) to predict signs of failure becomes possible, thereby enabling "uninterrupted logistics" to be realized.

Claims

[Claim 1] A means for acquiring, via a communication line, identification information of transported goods, workers, and material handling equipment acquired by a sensing means and / or an image acquisition means in all processes from receiving goods to shipping goods at a logistics center; A 3D simulator for creating a large amount of learning data by constructing and displaying a 3D layout of the logistics center on a computer and modeling it, and selecting input signals and / or teacher signals and rewards according to the control purpose. A logistics system comprising: The 3D simulator is a means for supporting optimization. a means for moving the movable material handling equipment; a means for displaying what is stored and what is not stored on which shelf of the logistics center on the 3D layout; a means for displaying a histogram of time and / or power consumption from an order to transfer an item to completion of the transfer in the logistics center on the 3D layout; a means for displaying integrated data of the operating time and travel distance of each component constituting the material handling equipment in the logistics center on the 3D layout; a means for accepting changes to various parameters including parameters related to the material handling equipment of the logistics center on the 3D layout; a display means for visually displaying a change in the failure frequency, etc., by comparing the failure frequency of the logistics center on the 3D layout related to the changed various parameters with the failure frequency of the logistics center on the 3D layout related to the various parameters before the change; A logistics system equipped with a sensor installed on an elevator body, which is at least one of an elevator in a logistics center in real space, a structure that structurally supports the elevator, and a component related to the structure, and which acquires sensing information related to physical quantities including vibration in real time or at regular time intervals; a wireless transmission unit that wirelessly transmits sensing information acquired by the sensor; a receiving unit that receives the sensing information transmitted from the wireless transmitting unit; Equipped with A continuous monitoring system for moving objects for predictive maintenance services, characterized in that it is equipped with a means for obtaining a trained model using artificial intelligence technology based on the sensing information received by the receiving unit and the obtained learning data, using the amount of fluctuation in the frequency of the part as an input value, comparing it with a threshold value set based on the trained model to predict the time to replace the part as an output value, and notifying the operator of the logistics center of the time to replace the lifting equipment body along with a warning email.

Images