Machine Learning for Appearance Grading of Sawn Timber using Cameras and X-ray Computed Tomography

↓ Direkt till sidans innehåll
↓ Direkt till sidans sekundära innehåll (sidomenyn)

Sökning: WFRF:(Olofsson Linus) > Machine Learning fo...

Machine Learning for Appearance Grading of Sawn Timber using Cameras and X-ray Computed Tomography

Olofsson, Linus (författare): Luleå tekniska universitet,Träteknik

Sandberg, Dick, 1967- (preses): Luleå tekniska universitet,Träteknik

Broman, Olof (preses): Luleå tekniska universitet,Träteknik

visa fler...

Fredriksson, Magnus, 1984- (preses): Luleå tekniska universitet,Träteknik

Skog, Johan (preses): Luleå tekniska universitet,Träteknik

Verkasalo, Erkki (opponent): Bio-based Business and Industry, Natural Resources Institute Finland, Joensuu, Finland

visa färre...

(creator_code:org_t)

ISBN 9789177908746
Luleå University of Technology, 2021
Engelska.
Serie: Doctoral thesis / Luleå University of Technology 1 jan 1997 → …, 1402-1544

Relaterad länk:: https://ltu.diva-por... (primary) (Raw object); visa fler...; https://urn.kb.se/re...; visa färre...

Doktorsavhandling (övrigt vetenskapligt/konstnärligt)

Abstract Ämnesord

Stäng

This doctoral thesis deals with a new approach for the appearance grading of sawn timber adapted to the requirements of modern sawmilling industries and timber market situations. Appearance grading of sawn timber allows wood products to be made with a speciﬁc visual style due to wood features such as knots. Identifying and grading sawn timber by its visual style is a holistic-subjective task that is inherently suitable for humans. However, with the ever-increasing demand for a faster and more consistent grading operation, humans have been replaced by automatic systems during the past few decades. However, the human perception of the appearance of sawn timber is not something easily deﬁned coherently and concisely for use in automatic systems, resulting in automatic systems struggling to perform appearance grading using conventional rule-based grading. As shown in this thesis, machine-learning methods can be used to teach an automatic system to perform holistic-subjective grading in a way that emulates manual grading while still performing the fast and consistent grading associated with automatic systems. This thesis introduced machine learning for product-adapted appearance grading of sawn timber and studied the use of machine learning to appearance grade sawn timber according to standardised quality grades, using an X-ray computed tomography (CT) scanner and a camera-based board scanner.In the studies presented in this thesis, measurement data from the CT scan-ner and the board scanner was used to create a set of variables only regarding knots. The variable sets and the grades of the sawn timber were modelled by projection to latent structures (PLS) models. The grade of the sawn timber was determined in three ways; ﬁrstly, manual grading according to standard-ised quality grades; secondly, called the product grade, the sawn timber was delivered to a wall-panelling customer, and the grade of the sawn timber was determined by the quality yield at the customer; and thirdly, called the image grade, images were extracted from the board scanner and used to estimate the quality yield of the wall-panelling customer manually. The grading in each scanning system was performed using a machine-learning method and a conventional rule-based approach, and their performances were compared.Seven data sets were collected in the studies presented in this thesis, each with a combination of variable sets from the scanners and quality grades as described above. In each study, one or more PLS models were trained to model the relationship between a variable set and a quality grade and used to predict the quality of the sawn timber. A PLS model predicts a score for each piece of sawn timber, and if that score passes a classiﬁcation threshold, the model assigns a quality grade. This classiﬁcation threshold could be tuned manually to introduce a bias in the model and thereby change the sorting outcome.When performing standardised appearance grading of dried sawn timber, both a PLS model and rule-based grading achieved about 80% grading accuracy, while a manual grader agreed to 95% with the PLS model and to 81% with the rule-based grading in a veriﬁcation test. Furthermore, when performing customer-adapted grading of the standardised grades, a PLS model managed an 84% grading accuracy compared to 64% of the rule-based approach. These results show how a conventional rule-based ap-proach struggled with performing customer-adapted grading compared to a PLS model. When performing standardised grading, however, both meth-ods achieved similar grading accuracy, but only the grading performed by the PLS model could not be signiﬁcantly distinguished from the targeted standardised grades.Using a PLS model to perform product-adapted grading of dried sawn tim-ber resulted in a grading accuracy of about 70%–80% for di˙erent scenarios. These gradings resulted in a quality yield, pass or fail, of about 80% for the wall-panelling customer. According to the customer, rule-based grad-ing did not yield impressive product-adapted results, and no metric was given. Furthermore, this thesis showed that the image grade was as useful as the product grade for training the PLS models, which greatly simpliﬁes the logistical process of creating a data set for training a product-adapted machine-learning model. Had a traceability method been used to collect the data from the scanners automatically, the image grade would allow for completely software-based data collection, which is very much in line with the industry 4.0 concept.A CT scanner enables the appearance grading of virtual sawn timber in the 3D images of the scanned logs, which allows the logs to be sawn for maxi-mum value or quality yield. The CT scanner was made to perform a primary product-adapted grading using either a PLS model or a rule-based approach. In addition to this primary grading, the CT scanner and board scanner were programmed to perform a small secondary grading by limiting a small set of measurements that the CT scanner could not suÿciently account for. For example, large pith deviations were limited in the CT scanner, and rotten knots were forbidden by the board scanner, as these measurements were associated with a high risk of resulting in poor quality wall panels for the customer. With this setup, a dataset of 300 pieces of virtual sawn timber was studied. Using rule-based primary grading, the sawmill delivered about 200 pieces of sawn timber with a product yield of 77% for the customer, after the board scanner rejected 28 pieces (12%). Then, by controlling the classiﬁcation threshold of a PLS model to make the primary grading very strict, meaning that the log was sawn to only yield very likely high-quality pieces of sawn timber, the sawmill could deliver 114 pieces of sawn timber with a product yield of 90%, after the board scanner rejected 9 pieces (7%). These results show that a PLS model achieved higher grading accuracy and higher quality yield than a rule-based approach. Furthermore, the classiﬁca-tion threshold of the PLS model allows for easy and intuitive control over the sorting outcome, something that the rule-based approach does not support.This thesis showed that a PLS-based machine-learning model could be used to perform holistic-subjective appearance grading by both a CT scanner and a board scanner, where a rule-based approach struggled in all but the most familiar case of standardised grading. Once a framework for a machine-learning method such as PLS has been implemented, this thesis showed the ease of customising and ﬁne-tuning the grading performance to be in line with customers needs. A customer or product adaptation could conceivably be initiated and ﬁnalised completely in software by automatically collecting the data using a traceability method, collecting the reference grades needed for training by grading images of sawn timber, and using the intuitive clas-siﬁcation threshold to ﬁne-tune the sorting outcome.

Ämnesord

LANTBRUKSVETENSKAPER -- Lantbruksvetenskap, skogsbruk och fiske -- Trävetenskap (hsv//swe)
AGRICULTURAL SCIENCES -- Agriculture, Forestry and Fisheries -- Wood Science (hsv//eng)

Nyckelord

Träteknik
Wood Science and Engineering

Publikations- och innehållstyp

vet (ämneskategori)
dok (ämneskategori)

Hitta via bibliotek

Machine Learning for Appearance Grading of Sawn Timber using Cameras and X-ray C... (Sök publikationen i LIBRIS)

Till lärosätets databas

Hitta mer i SwePub

Av författaren/redakt...: Olofsson, Linus; Sandberg, Dick, ...; Broman, Olof; Fredriksson, Mag ...; Skog, Johan; Verkasalo, Erkki

Om ämnet

LANTBRUKSVETENSKAPER: LANTBRUKSVETENSK ...; och Lantbruksvetensk ...; och Trävetenskap

Delar i serien: Doctoral thesis ...

Av lärosätet: Luleå tekniska universitet

Sök utanför SwePub

Sök vidare i:: Google; Google Book Search; Google Scholar

Kungliga biblioteket hanterar dina personuppgifter i enlighet med EU:s dataskyddsförordning (2018), GDPR. Läs mer om hur det funkar här.
Så här hanterar KB dina uppgifter vid användning av denna tjänst.

LIBRIS.kb.se