A method for quantifying working processes on manufacturing oors was established that uses a wearable sensor device and an ontology-based stream data processing system. Using this method, the measurement of manufacturing process e ciency from sensor data extracted from such a device worn by workers on the job was con rmed at the Fuji Xerox factory.
A real-time understanding of the situation at production sites is essential for management in manufacturing companies. Their ability to exploit unstructured as well as structured knowledge in the production eld determines their manufacturing competitiveness. Knowledge in the production eld includes three factors: production machinery, productive resources, and humans engaged in production. The machinery and resource factors, which are easy to measure quantitatively and can therefore be projected onto structured data, have been thoroughly utilized in manufacturing management. In contrast, the human factor, which is unstructured and therefore di cult to sense, has been little utilized so far.
The purpose of this study is to establish a method for quantifying the human factor on manufacturing oors. We rst propose to monitor factory job processes with wearable sensor devices. The meanings of signals detected by such devices, however, are not uniquely determined; that is, the same signals detected in di erent contexts can have di erent meanings. To de ne the meaning of each detected signal uniquely in each context, we further propose to exploit ontology-based complex modeling.
Kharlamov et al. [
Real-time data addressing is achieved by a data stream management system
(DSMS) or a complex event processing (CEP) system [
Wearable sensor devices were used and optimized to correspond to human activity speci c to the production eld. In order to reduce the barrier to introduction, a CEP system for semantically ltering stream data was implemented. 3.1
Wearable Device Optimization
Fig 1(a) shows the wearable sensor device used in this study. The device was
developed by LAPIS Semiconductor in the HAPIC project [
The character of motion in a factory work process di ers greatly from that in daily living because factory work requires more e ective motion for mass production. Fig 1(b) shows the acceleration sensor signal for a factory job process at di erent sensing frequencies (10 Hz and 50 Hz). For daily living, 10 Hz is the parameter value, whereas 50 Hz is the parameter value optimized for this work. This result shows that the daily living sensing frequency cannot detect the operations of factory workers. The 50-Hz frequency requires a higher communication bandwidth, but there is a trade-o between high communication bandwidth and battery management. We resolve this con ict to optimize the communication data pro le, device pro le, and system pro le. There are many workers (more than several thousand) and a wide variety of job processes (more than several hundred) on the production oor. State control of such huge and complex events is di cult with a conventional DSMS/CEP module. We propose a method for semantically ltering stream data to reduce such complexity. Fig 2 shows the architecture of our system. Sensed data are collected in the sensor/actuator control hub and transferred to Key-Value stream data. Key data are created from the time stamp and the device address using RFC4122, and Value data are created from the sensed raw data. Stream data corresponding to reactivity and scalability are stored in the Key-Value Database (KVDB). The service module semantically lters the stream data from the KVDB using ontology-based reasoning and query rewriting to key select queries. The reduced quantity of stream data can be accessed by a simple CEP module such as embedded event processing (eep). The CEP module has pre-processing submodules such as a low path lter, feature extraction submodules such as a fast Fourier transform (FFT), classi cation submodules such as a hidden Markov model (HMM), and post-processing submodules. In order to address the working process context for the sensed data, we created a manufacturing ontology from the operation instruction manual (OIM), which describes in detail the assembly motions, procedures, and related parts information. The manufacturing ontology consists of the basic motions, procedures, and product component ontology. Manufacturing context reasoning was produced to match the basic motion ontology to the classi ed sensed data. 4
Many factory workers in Japan have performed an improvement activity called Kaizen. Conventional Kaizen begins with a working process quanti cation that is carried out via job analysis and process time measurements. The working process quanti cation requires much time and a high level of skill because the job analysis requires a deep knowledge of work processes.
We assume a simple scenario for real-time manufacturing process quanti cation as follows: An assembly worker puts the wearable sensor device on his or her wrist and collects job process data with it. The manufacturing engineer select the job to improve. System shows directly selected job process e ciency value (activity e ciency, motion e ciency) and standard value. If job e ency value di er widely from standard value, they go to improvement process.
We examined the above scenario in the Fuji Xerox Printer assembly line. The manufacturing engineer queried the job and collected the data, which were ltered using queries rewritten from it. The ltered data components such as acceleration, gyro, and geomagnetic data were then used to calculate the direction and magnitude of displacement that occurred during the job process. Activity e ciency (derived from integrating the magnitude of displacement) and motion e ciency (derived from the anisotropic ratio of the direction) were able to be viewed in real time and compared to standard one.
Table 1 shows activity e ciency and motion e ciency value at di erent production lot compared to standard one. This result shows big activity e ciency dispersion is exist in the job, and that indicate the job process containes redundant motion and needs to improve it. indicator activity e ciency motion e ciency (%) 58.7 55
We con rmed that our method can be used for quanti cation of a manufacturing process in Fuji Xerox. Our system allows real-time manufacturing process improvement using human factor information, which has not been satisfactorily utilized to date. Our next step will be to apply it in other scenarios such as the creation of links between process quanti cation data and enterprise data.