Kazuya ITO*
Toshiyuki UMEZAWA*
Akiko YOKOYAMA*
*Ebara Environmental Plant Co., Ltd.
In February 2019, an automatic crane system outfitted with waste identification AI started operation at the Funabashi Hokubu Incineration Plant. Previously, waste crane operation was dependent on the experience and judgment of operators, but this system automates it by utilizing AI while maintaining stable incinerator operation. In a six-day demonstration experiment, we achieved an automatic operation rate of about 90 % without adversely affecting waste combustion or requiring constant monitoring of the pit by an operator. Based on these results, the system was put into practical operation. It has maintained an automatic operation rate equivalent to the value achieved in the demonstration experiment for more than half a year.
Keywords: Incineration Plant, Automation, Crane Operation, Deep Learning, AI
In Japan, it is required that the public projects are streamlined through the utilization of private companies due to chronic shortages of manpower resulting from the reduction of the workforce and stringent local government finances. This is also the case with the waste processing business, and recently, comprehensive contracts, such asDBO (Design Build Operate) contracts, have been on the increase. As we face shortages of skilled workers, we must establish schemes that envision long-term operation and are intended for further advanced facility control.
The operation of a waste incineration facility requires many personnel, but crane operation, such as stirring for homogenizing the properties of the waste inside a waste bunker (hereinafter referred to as a “bunker”) to stabilize waste combustion, is still not completely automatized. In this operation, the operator recognizes the properties of the waste visually and loads the homogenized waste selectively into the incinerator by manual or semi-automatic operation. Thus, we first developed AI which is capable of grasping the properties of the waste using deep learning, as like the skilled operator’s visual sense. We also established an automatic crane system by integrating AI output into crane control. As a result, we have been able to automatize a series of operations shown in Figure 1, making it possible to make stable and advanced decisions that do not rely on the proficiency of personnel.
Fig.1 Concept of the system
In this report, we introduce the automatic crane system with waste identification AI (hereinafter referred to as the “system”).
First of all, we explain the core of the system: waste identification AI (hereinafter referred to as “AI”). This AI learns the characteristics of waste through deep learning that uses images of actual waste bunker as training data, and it can identify the following.
(1) waste that affects combustion or devices if it is loaded into the incinerator in a mass, such as pruning branches or sludge (hereinafter referred to as “special waste”)
(2) waste suitable for loading into the incinerator
(3) waste in unbroken bags
Figure 2 shows an example of waste identification by AI. By displaying the types of waste identified by AI with using different colors and comparing that with the original image captured at the bunker, we can prove that AI can accurately identify types of waste. In addition, Table 1 shows the results of the identification accuracy of the AI-estimated types of waste assessed using F1 scores (Note) through comparison with the results of judgement of respective types of waste by a skilled operator. We quantitatively confirmed that these values were good enough for the system.
Fig.2 Results of waste identification by AI
Moreover, Figure 3 shows the results of comparison between the loadability of waste at the location assigned to each area inside the bunker, which operators judged based on their experience, and the degree of broken / unbroken garbage bags estimated by AI (hereinafter referred to as the “ratio of broken bag”) when implementing the developed AI in this system. It is clarified by these results that there is a threshold value of the ratio of broken bag that can serve as the boundary between loadable and unloadable waste. Integrating the threshold value into the system will make it possible to load waste in an equivalent state to the state of waste that an operator loads.
Fig.3 Relationship between loading criteria by skilled operator and ratio of broken bag
(Note) F1 scores: Indicators considering two ratios: recall (for example, the ratio at which AI can estimate actual pruning branches as pruning branches) and precision (the ratio at which the pruning branches estimated by AI are actually pruning branches).
This section describes the configuration of the system. As shown in Figure 4, the system is divided into four main components.
(1) Camera for bunker photographing
(2) AI that identifies the waste inside the bunker
(3) Advanced control that judges appropriate crane operation
(4) Crane control that promptly executes commands from the advanced control
Fig.4 Configuration of the automatic crane system with waste identification AI
This system can undergo remote support if it is designed to be remote-connectable using a closed VPN connection, for example. The sections below describe each of the abovementioned components.
Industrial cameras take photos of the waste inside the bunker at fixed time intervals. Determine the number of cameras and installation locations considering not only the size and structure of the bunker but also an arrangement that does not cause any blind spots by the waste level in the bunker.
The images taken by the cameras are imported, and the ratio of each type of waste existing at each location in the bunker is inferred by AI as the output. The results of inference by AI can be reviewed on the screen as needed. Thus, the accuracy of AI can be easily checked by, for example, directly viewing the bunker fcom the crane operator’s room and comparing it with the results of inference by AI.
In the advanced control, the aforementioned inference results by AI, waste-level data transmitted from the crane control, waste loading request signals, present situation at receiving area (whether garbage trucks are permitted to approach), the number of trucks inside the facility and the carry-in status, whether the bunker gates are open or closed, and various other conditions related to facility operation are comprehensively judged, and crane operation commands, e.g. its grasping and releasing locations, and the direction about the incinerator into which the waste will be loaded are sent to the crane control. The operator can check through the screen of the advanced control system whether the system is operating safely, whether there is special waste inside the bunker, and more. It is also possible to change the settings of the system (time limit for reducing the waste level in front of the bunker gates in preparation for receipt on the following day, bunker gates to be used, etc.) according to the operating state of the facility.
In crane control, commands are received from advanced control to execute crane operation. In addition, the level of waste inside the bunker is constantly monitored by installing a level sensor on the girder of the crane or using a stereo camera or ToF camera.
Note that the system can control the crane fully automatically, and that it can be introduced to and operated with an existing crane control system if the speeds of individual motions are sufficiently high.
The main functions of the system equipped with waste identification AI are as follows:
(1) Finding special waste and withdrawing it to a specific place (can be stored in memory and withdrawn even if covered by general waste)
(2) Loading waste suitable for loading in response to a loading request from the incinerator
(3) Finding unbroken garbage bags, breaking them, and stirring the waste to be suitable for loading into the incinerator
(4) Moving the waste in front of the bunker gates without causing congestion of garbage trucks attributable to access by the crane to the bunker gates
With the aid of these functions, the system can be operated automatically not only in the nighttime but also in the daytime without requiring constant monitoring by the crane operator.
To confirm that using the system will allow crane operations to be automatized without any problems, we conducted a demonstration experiment for six consecutive days at the Funabashi Hokubu Incineration Plant (127 t/d × three grate-type incinerators). In the experiment, we verified whether the functions of above (1) to (4) were executed appropriately and, at the same time, whether there were any differences between the combustion stability during operation by the automatic crane system and during manual operation.
Table 2 shows the results of comparison between the combustion state during operation by the automatic crane system and during manual operation by a crane operator in the demonstration experiment period.
We found little difference in operation performance between the two cases, and confirmed that automatic crane operation by the system was capable of securing functions for the stable operation of the waste incineration facility in automatic combustion control (ACC).
In addition, waste was smoothly received in the daytime without congestion of garbage trucks, and the amount of waste processed was maintained at the rated value.
The duration of automatic operation was about 90% of the entire period, and the manual operation in the rest of the period was not urgent and was mainly for the mixing of special waste that was automatically withdrawn. From these results, we confirmed that automatic operation was feasible in both daytime and nighttime without constant monitoring by the crane operator.
The above-mentioned results demonstrated that the system was practical for automatic operation without any problems and also helped realize manpower saving. The system started operation in February 2019 at the Funabashi Hokubu Incineration Plant.
We investigated the status of automatization after the start of system operation. Figure 5 shows the results of the investigation of the status of automatization for about a month. In the investigation, we assessed the ratio of automatization in the waste receiving period in the daytime and the waste non-receiving period separately in addition to the ratio of automatic operation throughout the daytime and the nighttime.
Fig.5 Results of operation of the automatic crane system with waste identification AI
These results prove that a ratio of automatic operation as high as that during the demonstration period was realized after the start of operation. In the waste receiving period in the daytime in particular, it had been difficult to realize automatization with a conventional crane because preference to be given for the smooth receiving of waste. The system did, however, realize both smooth receiving of waste and automatization. In the waste receiving period in the daytime, an automatization ratio of 94% was achieved, excluding manual operation for training and seasonal operation.
As explained thus far, we confirmed that the system operated without any problems and could contribute to manpower saving in crane operations and also stable operation of the incinerator.
We will introduce the system to other facilities and to realize advanced automatic operation of waste incineration facilities by combining the system with an automatic combustion control system or another control system.
Finally, we would like to extend our deep gratitude to the people of Funabashi City and of Ridge-i Inc. for their great help and support in the development and demonstration experiment of the system and to everyone else who cooperated.
Under the Scenes of our Lives High-pressure pump - Role and Application -
50% capacity boiler feed pump(BFP）playing an active role in a 1 000 MW thermal power plant
Large-capacity, Ultrahigh-efficiency, High-pressure Pumps for Seawater RO Desalination Delivered to Carlsbad Desalination Plant in the U.S.
Streamlines in crossover passage and velocity distributions at inlet of the second-stage impeller (Left:original,Right:optimized)
Discussion Meeting Symposium Ebara research system - Cooperation between research and business to create a new future -
Discussion Meeting (Mr. HIYAMA, Mr. SOBUKAWA, Mr. GOTO）
Under the Scenes of our Lives Standard Pumps - Essential Part of our Everyday Lives -
Examples of standard pumps
Inquiry about Ebara Engineering Review
Inquiry Form