Benchmarking moisture prediction in kiln-dried Pacific Coast hemlock wood

Abstract

The uniformity of final moisture content within a drying timber batch is crucial. Lack of such uniformity leads to producing large percentages of over-dried and under-dried timber, resulting in significant quality degradation and value downgrade. This study aims to predict kiln-dried timber moisture content using its initial moisture value, timber weight, and density. The distribution of wood properties in different drying runs was analyzed, and the difference in their means was statistically assessed. Various machine learning models were used for moisture prediction. The performance of the group method of data handling network was compared with the adaptive neuro-fuzzy inference system, support vector regression, decision tree, and random forest method. The best performance was achieved using random forest with the initial moisture content and weight of the wood as input parameters. Finally, the models’ performances were compared and practical recommendations for employing the adopted methodology in industrial settings were provided.

Keywords

Pacific coast hemlock moisture prediction machine learning kiln drying random forest neural networks

Introduction

Kiln drying plays a pivotal role in wood manufacturing and improving the physicomechanical properties of dried timbers (Perre 2007). Kiln drying aims to reduce the initial moisture (M_i) to target moisture (M_t), which is acceptable for the ultimate usage and extended service life (Cabardo 2007; Salas and Moya 2014). Variation in the final moisture (M_f) between kiln-dried timbers results in over-dried and under-dried timbers and consequently downgraded final products. Thus, grasping the parameters involved in predicting the M_f and reducing its variation after kiln drying is paramount (Elustondo and Oliveira 2009). Minimizing M_f variation has long been a critical challenge, especially for the species with very high green moisture (M_g) variation, such as Pacific Coast hemlock (PCH), spruce-pine-fir, and poplar (Cai and Oliveira 2004; Rahimi et al. 2011; Shahverdi et al. 2017).

The M_f significantly impacts the quality of the timber and the engineered wood products made from it. The moisture variation affects the elastic mechanical properties of wood, but it also influences the viscoelastic properties (Fathi et al. 2020a). Other than its impact on the mechanical properties, moisture content is an important parameter when evaluating the wood durability (Ayanleye et al. 2022). Considering the significance of the M_f prediction in the quality of dried timbers, researchers have attempted to estimate and model M_f at different M_t. For this purpose, researchers used probabilistic analysis and numerical stochastic models to simulate M dispersion during and after kiln drying (Cronin et al. 2002, 2003; Elustondo and Avramidis 2002). Elustondo and Avramidis (2005) used three timber drying simulation approaches, namely Cronin's, Monte Carlo, and a continuous probability density (CPD). They claimed that CPD applies to iterative types of calculations such as process optimization or data fitting. Moreover, Watanabe et al. (2016) used the bootstrap method for probabilistic evaluation of M_f in kiln drying timbers. They concluded that the M data did not conform to any specific classical probability distribution such as Normal, Log-Normal, Weibull, or Gamma, which emphasized the requirement for nonparametric statistics. More recently, Rahimi, Avramidis et al. (2021) applied polynomial models to predict the M_f variation of the kiln-dried timbers and proposed a closed formula based on the M_i indices.

Classic methods are incapable of predicting and characterizing M and its variation in wood drying, mainly due to the natural complexity of wood and the variation in the data collected during the wood drying process. However, artificial neural networks (ANNs) are powerful and robust tools for self-organizing adjustment, adaptive learning, and fault-tolerant anti-noise dealing with nonlinear and complex processes (Sablani et al. 2005; Fu et al. 2017). For these reasons, researchers used ANN methods for characterizing and categorizing M and other drying parameters (Wu and Avramidis 2006; Ceylan 2008). Watanabe et al. (2013, 2014) used ANN to predict M_f and surface stresses of kiln-dried sugi (Cryptomeria japonica). Researchers applied ANNs to sort green PCH timbers with near-infrared spectroscopy (NIRS) and utilized networks for online species identification of PCH based on NIRS and chemometrics (Zhou et al. 2020). In addition, machine learning (ML) methods have been applied to kiln-dried PCH to predict M_f and categorize dried timbers accordingly (Rahimi, Nasir et al. 2021; Rahimi and Avramidis 2022).

The most commonly employed ANN in the literature is the multilayer perceptron (MLP) model, widely used for different regression and classification tasks. While it has been called a universal approximator (Hornik et al. 1989), it has certain drawbacks. The black-box nature of the MLP ANN makes it challenging to comprehend the causality of events in the model (Park and Lek 2016). Also, the convergence of the system and its performance may highly depend on the network architecture and the choice of training algorithm (Nasir and Cool 2020a). Thus, training an optimized model may require a trial and error task. Finally, the model's performance is influenced by the complexity of the dataset.

Rahimi and Avramidis (2022) applied different ANN methods, namely MLP, radial basis function (RBF), and group method of data handling (GMDH), to estimate the M_f of the kiln-dried PCH. Predictive models consisted of four inputs: M_t, M_i, weight (w_i), and density (ρ_b). Research findings revealed that GMDH yielded the best performance. However, other ML approaches have not yet been applied to predicting M_f. This study is a continuation of Rahimi and Avramidis (2022), further optimizing M_f predictive models in timber kiln drying. It compares the results of the best ANN model (i.e. GMDH) to those of other ML predictive models, namely adaptive neuro-fuzzy inference system (ANFIS), support vector regression (SVR), decision tree, and random forests, to find the optimal ML model for accurate prediction of moisture in kiln-dried PCH.

Materials and methods

Specimen preparation and drying experiment

A total of 1152 freshly cut green PCH (also known as hem-fir) timbers (50 mm × 100 mm × 2.44 m) were procured, and 2304 kiln specimens (50 mm × 100 mm × 1.25 m) and cookies (50 mm × 100 mm × 30 mm) were cut out of the timbers according to the cutting protocol represented in the former study (Rahimi and Avramidis 2022). The oven-dry weight method measured the two cookies’ average M_g and ρ_b. In the next step, all green specimens were sorted and equally divided into six groups. These groups were formed by combining two M_g variations (high and low) and three setpoints (high, medium, and low). Three setpoints of 21% (group 1 and 2), 16% (group 3 and 4), and 11% (group 5 and 6) were considered as the high, medium, and low M_t levels. The odd groups (1, 3, and 5) had high M_i variation, while even groups (2, 4, and 6) had low M_i variation. The clustering and categorization of the kiln specimens were elaborated on and represented in the previous research (Rahimi, Nasir et al. 2021).

Each drying run contained 384 kiln specimens divided into two timber stacks. Drying runs were carried out in the FPInnovations’ research kiln. All six batches were dried under an identical schedule slightly different from the modified typical industrial schedule for PCH (Shahverdi et al. 2017). The drying schedule was tabulated, and drying trends were represented in former research (Rahimi, Avramidis et al. 2021; Rahimi, Nasir et al. 2021; Rahimi and Avramidis 2022).

Statistical analysis

Statistical analysis was accomplished on wood attributes, namely ρ_b, w_i, M_i, and M_f, to evaluate the measures of central tendency (mean and median) and variability (standard deviation, coefficient of variation, and range). Consequently, the mean values of wood attributes (ρ_b, w_i, M_i, and M_f) were analyzed and compared between six drying batches using analysis of variance (ANOVA) at a significance level of 0.95% (confidence level of 0.05%). In case of significant difference for each wood property in the ANOVA test, Duncan tests were implemented to elaborate on the differences in mean values between different drying batches.

Machine learning approaches

Among all influential parameters that affect M_f variation of a kiln's batch, M_i, M_t, and

, are most likely the key factors (Perre 2007). Several parameters, such as wood species, region, presence of juvenile and/or reaction wood, etc., affect the density (ρ_M) of the wood (Siau 1995). Hence, it potentially has a remarkable effect on the drying rate, and consequently, M_f. Additionally, w_i is correlated with M_i of timbers, i.e. the more w_i, the more M_i can be. Thus, four wood drying indices, M_i, M_t, ρ_b, and w_i, were considered to estimate M_f. Table 1 lists different network configurations used in this study.

Table 1

Features available in ML configurations.

Configuration	Number of included initial factors	List of excluded initial factors
A	Four	None
B	Three	ρ_b
C		w_i
D		M_i
E	Two	ρ_b and w_i
F		M_i and ρ_b
G		M_i and w_i

Approaches for moisture prediction

GMDH is a family of inductive algorithms for various data mining and pattern recognition problems. GMDH is a powerful approach for predicting the mechanical properties of wood and monitoring the wood manufacturing processes (Nasir et al. 2019a; Fathi et al. 2020b). This method can find the optimal architecture automatically, thanks to its inductive and self-organizing nature (Witczak et al. 2006). GMDH approach works according to a multilayer network of second-order polynomials. The quadratic neurons’ weights are tuned through the learning process, and the GMDH network creates all possible combinations of input pairs from m variables in the first hidden layer. Then, each quadratic neuron will get trained by the least-squares method. The neuron selection criterion is applied to every layer through the natural selection of ideas to the feasible network complexity. Correspondingly, the classification accuracy for each quadratic neuron is calculated by comparing the output of the polynomial models with the targets. Consequently, the neurons with the polynomial model fitness function below the preferred error level are kept, and the rest are left aside (Nasir, Nourian et al. 2019).

ANFIS is a neural network-based fuzzy system based on the Takagi–Sugeno–Kang model (Sugeno 1985). The ANFIS antecedent consists of fuzzy sets; however, linear equations make the consequence. ANFIS is similar to fuzzy inference system (FIS) and NN in several ways. On the one hand, ANFIS includes input fuzzification, rule inference, fire strength computation/implication, aggregation, and output defuzzification steps (like FIS). On the other hand, its input and output neurons represent the training and predicted values, respectively (similar to ANN). This system's membership functions and rules are positioned in the hidden layer. ANFIS consists of a neural structure in which its neurons are the FIS parameters that can be adjusted and optimized by ANN learning methods (Nasir et al. 2019b). A hybrid algorithm was used to train the generated ANFIS model, and the least-squares method in the feedforward algorithm was used to specify the output parameters. The mean square error between the model output and target data was calculated to evaluate the accuracy of the network. For training, validation, and test, a percentage breakdown of 60%, 20%, and 20% of the total dataset were considered, respectively. The fuzzy C-mean (FCM) clustering method generated the FIS structure applied in this study to make the Sugeno-type FIS model and determine the number of rules and membership functions (MFs). The number of clusters produced by the FCM is defined beforehand through trial-and-error. Yilmaz and Kaynar (2011) claimed that the ANFIS computing method is preferable over linear regressions because it benefits from fuzzy logic and neural-based structures. Masoudi et al. (2018) reported that ANFIS is a powerful ML tool in intelligent machine monitoring, especially when it comes to performance in predicting the cutting zone temperature in the turning process. It has been used to predict PCH and Douglas-fir's mechanical properties (Ayanleye et al. 2021) and monitor the wood machining process (Nasir and Cool 2020b). Abellan-Nebot and Subiron (2010) stated that ANFIS is a robust method when dealing with a medium sample size, which was the case in this study. Besides, they indicated that ANFIS could decrease the size and cost of the experiments during training a model.

SVR constitutes a new promising approach to data regression, and its algorithm is different than neural networks but quite like a support vector machine. SVR maps training data into the l-dimensional feature space that formulates an optimized hyperplane and demonstrates the nonlinear relation between input and output data, i.e. independent and dependent variables (Chen et al. 2017). The decision tree is popular for its intuitive nature, making it easier to interpret. It has been used to predict surface and internal checks formation in weathered thermally modified timber (Van Blokland et al. 2021a, 2021b). It was also employed to monitor the mechanical degradation of artificially weathered wood using the wave propagation data (Nasir, Fathi et al. 2021). The challenge of using a decision tree is finding the optimal tree size, which should be tuned during the pruning phase. CART (classification and regression trees) algorithm (Steinberg 2009) was used for decision tree modeling. The least squared error method was used for node splitting, and the optimal tree was set to be within one standard error of the tree with maximum R². Other hyperparameters were formed similar to those reported by Van Blokland et al. (2021a). Random forest is an ensemble learning algorithm based on bootstrap sampling, which aggregates the decision of many trees known as weak learners and uses a voting scheme to compute the model output. Schubert et al. (2020) used the random forest to predict the mechanical properties of wood fiber insulation boards. Also, it has been successfully employed for wood machining monitoring (Nasir et al. 2020; Nasir, Kooshkbaghi et al. 2021). The number of trees in the model was 500, and 81% of the training data were used to make the bootstrap sample size. The model was also validated by the bag method.

Results and discussion

Initial and final moisture distribution

The results of the ANOVA test for the M_i values of the kiln specimens revealed a significant difference (0.012) between groups at the confidence level of 95%. Duncan's results revealed that the second group's average M_i values differed from those of the first, fourth, and sixth groups. Moreover, the average w_i values of the third and fourth groups were statistically different (Figure 1). The average M_i for the six drying batches were 102.4%, 108.2%, 106.8%, 99.6%, 103.4%, and 101.3%, respectively.

Figure 1

The statistical difference for M_i between six drying batches. Letters above the columns show statistically significant differences.

ANOVA test for the M_f values of the kiln specimens also indicated a significant difference (0.012) between groups at the confidence level of 95%. Duncan's results showed an insignificant difference between the first and second groups and the fifth and sixth groups, which verified that M_t significantly impacted the statistical differences between M_f values of the groups. In contrast, the M_i variation insignificantly influenced differences between M_f values of the groups (Figure 2). The average M_f for the six drying batches were 21.1%, 20.3%, 15.8%, 16.1%, 11.7%, and 11.3%, respectively.

Figure 2

The statistical difference for M_f between six drying batches. Letters above the columns show statistically significant differences.

The reported

values in this study did not accord with the findings of previous studies possibly due to three reasons, i.e. determining different M_t (setpoints), applying different drying schedules, and pre-sorting. In addition, current research results did not accord with previous findings, which could be due to two reasons. The timbers used in this research were green and freshly procured, thus having very high M_f values. However, the timber population used in this research contained 2304 specimens, significantly larger than those used in previous studies. Table 2 summarizes the values of

reported in former studies.

Table 2

Statistical parameters for M_i in PCH reported in the literature.

Author (year)	Statistical parameters for M_i (%)
	Mean	SD	Maximum	Minimum
Current study	104	37	249	30
Shahverdi (2015)	88	42	204	30
Wada (2013)	81	–	175	30
Rohrbach (2008)	79	–	168	33
Bradic (2005)	61	–	167	17
Sackey et al. (2004)	62	–	186	34
Oliveira and Wallace (2001)	–	–	60	26
Hao and Avramidis (2004)		–	72	58

Initial weight distribution

Concerning current commercial-scale technology, sawmills cannot reliably measure the M and ρ_M of green timbers in their production lines. This underscores the importance of w_i as an input factor to predict M_f. Considerable differences in the w_i among kiln specimens probably originated from various M and ρ_M values (Siau 1995). ANOVA test for the M_f values of the kiln specimens indicated a significant difference (0.007) between groups at the confidence level of 95%. Duncan's results revealed that the average w_i values of the second and third runs were statistically greater than those of the other four runs (Figure 3). The average w_i for the six drying batches were 5.64, 5.84, 5.82, 5.65, 5.63, and 5.65 kg, respectively.

Figure 3

The statistical difference for w_i between six drying batches. Letters above the columns show statistically significant differences.

Basic density distribution

ρ_b is irrelevant to M because it defines as oven-dry weight over green volume. In addition, the three dimensions of the kiln specimens are almost equal and, therefore, have nearly the same V. However, the differences in the ρ_b among kiln specimens stemmed from the significant difference of ρ_b between western hemlock and fir, as well as, between juvenile and mature wood, as well as heartwood and sapwood. ANOVA test for the ρ_b values of the kiln specimens indicated a significant difference (0.007) between groups at the confidence level of 95%. Duncan's results demonstrated that the only significant difference was between the average ρ_b values of the fourth and fifth runs, which had the greater and smaller ρ_b values, respectively (Figure 4).

Figure 4

The statistical difference for ρ_b between six drying batches. Letters above the columns show statistically significant differences.

Table 3 summarizes the ρb values reported in previous studies, revealing that this study's results accord with previous findings. Fir dries faster because of its relatively lower ρ_b, and greater permeability than western hemlock (Avramidis and Oliveira 1993; Morris 1995).

Table 3

The statistical parameters for ρ_b in PCH reported in literature.

Author (year)	Statistical parameters for ρ_b (kg m^–3)
	Mean	SD	Maximum	Minimum
Current study	394	60	594	227
Nourian (2018)	455	22	424	517
Shahverdi (2015)	397	55	546	310
Wada (2013)	–	–	564	316
Rohrbach (2008)	380	–	570	198
Bradic (2005)	401	–	556	282
Li et al. (1997)	–	–	540	261

Wood moisture prediction

Table 4 lists the performance of the ML models for the different configurations (architecture) based on 4-fold cross-validation. As can be seen, ANFIS resulted in very high MSE values in architectures A (including M_t, M_i, w_i, and ρ_b), B (including M_t, M_i, w_i), and C (including M_t, M_i, and ρ_b), thus failed to predict M_f.

Table 4

Performance of the predictive models for different configurations.

Approach	MSE values for different configuration (%)
	A	B	C	D	E	F	G
GMDH	5.8	5.2	5.9	18.8	5.4	19.3	34.9
ANFIS	3089.6	204.3	198.3	14.2	10.0	5.8	7.3
SVR	6.1	6.2	6.1	14.3	6.1	20.5	36.5
Decision Tree	4.04	4.04	4.04	18.1	4.04	23.6	42.7
Random Forest	3.91	3.87	4.02	15.0	4.9	23.8	46.6

According to Table 4, models that do not use the M_i (model D, F, G) show a relatively lower performance indicating the importance of wood M_i in the accuracy of the prediction. The best performance was achieved using configuration B (including M_t, M_i, and w_i) and random forest modeling. The relative importance of the parameters used in the model is shown in Figure 5.

Figure 5

Relative importance of the parameters in configuration B used to train the random forest model.

Figure 5 shows that M_i and w_i account for the most and least important variables in the random forest model, respectively. Thus, rapid and reliable measuring of M_i is the key to predicting the M_f in kiln-drying. Excluding w_i, and using M_t and M_i (configuration E) also yields very similar results. However, in this case, the best performance is obtained using a decision tree followed by random forest, GMDH, and SVR. The R² (test data) achieved using the optima model (random forest – configuration B) is 0.934. Figure 6 shows that this R² depends on the number of trees in the random forest model. It can be seen that the model has poor performance with a low number of trees (N < 50). Figure 7 shows the actual M_f versus the fitted (predicted) M_f on the out-of-bag and test data.

Figure 6

The variation in R² with the number of trees in the random forest model.

Figure 7

The actual M_f versus the fitted (predicted) M_f on the out-of-bag and test data.

Overall, ANFIS failed to accurately predict the M_f. Despite the superior performance of GMDH over MLP and RBF ANN (Rahimi and Avramidis 2022), it still yielded a lower performance than random forest. The performance of SVR was comparable with that of GMDH, indicating a promising ability to predict M_f. While the best result was achieved using random forest, decision tree, GMDH, and SVR can also yield a comparable result when M_i is included in the input parameters in the ML model. Owing to the importance of M_i and w_i in the prediction of M_f, future research should investigate the rapid and reliable techniques for measuring the moisture of green wood. However, this may require different data acquisition methods and affect the choice of ML modeling, which is affected by the complexity of the data and the level of feature engineering before feeding the data into a predictive model (Nasir and Sassani 2021).

Conclusions

This study bridged some significant knowledge gaps on modeling M_f in kiln-dried PCH timbers. It was shown that the key parameter in developing a reliable predictive model is the accurate measurement of wood's M_i. Measuring the moisture content of green lumber can be challenging in sawmills. Future studies should focus on non-destructive moisture assessment in large-size green lumber. Despite the high performance of random forest, other methods such as decision tree, GMDH, or SVR yielded competitive results. Non-destructive measurement of initial moisture content by techniques such as infrared spectroscopy or ultrasound method would change the nature of the input data set. It may require re-evaluating the machine models to identify the one that best fits the input data. This study did not consider the effect of wood properties and drying conditions on the drying defect such as warping, cupping, and checking, to name a few. Future studies can focus on developing predictive tools for assessing the drying defects parallel to moisture prediction.

Footnotes

Acknowledgement

The use of the FPInnovations laboratory kiln in this project is highly appreciated.

Disclosure statement

No potential conflict of interest was reported by the author(s).

References

Abellan-Nebot

, Subiron

FR.

2010. A review of machining monitoring systems based on artificial intelligence process models. Int J Adv Manuf Technol. 47(1-4):237–257.

Avramidis

, Oliveira

1993. Influence of presteaming on kiln-drying of thick hem-fir lumber. For Prod J. 43(11/12):7–12.

Ayanleye

, Nasir

, Avramidis

, Cool

2021. Effect of wood surface roughness on prediction of structural timber properties by infrared spectroscopy using ANFIS, ANN, and PLS regression. Eur J Wood Wood Prod. 79(1):101–115.

Ayanleye

, Udele

, Nasir

, Zhang

, Militz

2022. Durability and protection of mass timber structures: a review. J Build Eng. 46:103731.

Bradic

2005. Impact of juvenile wood on the drying characteristics of Pacific Coast hemlock structural timber [M.Sc. Thesis] Vancouver (BC): University of British Columbia; p. 152.

Cabardo

SJ.

2007. Variability in physical and transport properties regarding drying behavior for regrowth and plantation blackbutt timber in New South Wales. Ph.D. dissertation University of Sidney, South Wales, Australia, p. 109.

Cai

, Oliveira

LC.

2004. Equalizing and conditioning spruce-pine-subalpine fir lumber. For Prod J. 54(12):126–131.

Ceylan

2008. Determination of drying characteristics of timber by using artificial neural networks and mathematical models. Drying Technol. 26(12):1469–1476.

Chen

, Xu

, Chu

, Li

, Wu

, Ni

, Bao

, Wang

2017. Short-term electrical load forecasting using the support vector regression (SVR) model to calculate the demand response baseline for office buildings. Appl Energy. 195(139):659–670.

10.

Cronin

, Abodayeh

, Caro-Corrales

2002. Probabilistic analysis and design of the industrial lumber drying process. Dry Technol. 20(2):307–324.

11.

Cronin

, Baucour

, Abodayeh

, Barbot Da Silva

2003. Probabilistic analysis of timber drying schedules. Dry Technol. 21(8):1435–1458.

12.

Elustondo

, Avramidis

2002. Stochastic numerical model for radio frequency vacuum drying of lumbers. Dry Technol. 20(9):1827–1842.

13.

Elustondo

, Avramidis

2005. Comparative analysis of three methods for stochastic lumber drying simulation. Dry Technol. 23(1-2):131–142.

14.

Elustondo

, Oliveira

2009. A method for optimizing lumber sorting before kiln-drying. For Prod J. 59(9):45–50.

15.

Fathi

, Kazemirad

, Nasir

2020a. A nondestructive guided wave propagation method for the characterization of moisture-dependent viscoelastic properties of wood materials. Mater Struct. 53(6):1–14.

16.

Fathi

, Nasir

, Kazemirad

2020b. Prediction of the mechanical properties of wood using guided wave propagation and machine learning. Constr Build Mater. 262:120848.

17.

, Avramidis

, Zhao

, Cai

2017. Artificial neural network modeling for predicting elastic strain of white birch disks during drying. Eur J Wood and Wood Prod. 75(6):949–955.

18.

Hao

, Avramidis

2004. Annual ring orientation effect and slope of grain in hemlock timber drying. For Prod J. 54(11): 41–49.

19.

Hornik

, Stinchcombe

, White

1989. Multilayer feedforward networks are universal approximators. Neural Netw. 2(5):359–366.

20.

, Avramidis

, Oliveira

, Hartley

1997. The effect of vertical air gaps, air velocities and fan revolutions on the drying characteristics of thick Pacific Coast hemlock lumber. Holzforschung. 51(4):381–387.

21.

Masoudi

, Sima

, Tolouei-Rad

2018. Comparative study of ANN and ANFIS models for predicting temperature in machining. J Eng Sci Technol. 13(1):211–225.

22.

Morris

PI.

1995. Pacific silver fir is the more-treatable component of hem-fir from coastal British Columbia. For Prod J. 45(9):37.

23.

Nasir

, Cool

2020a. Characterization, optimization, and acoustic emission monitoring of airborne dust emission during wood sawing. Int J Adv Manuf Technol. 109(9):2365–2375.

24.

Nasir

, Cool

2020b. Intelligent wood machining monitoring using vibration signals combined with self-organizing maps for automatic feature selection. Int J Adv Manuf Technol. 108(5):1811–1825.

25.

Nasir

, Cool

, Sassani

2019a. Intelligent machining monitoring using sound signal processed with the wavelet method and a self-organizing neural network. IEEE Rob Autom Lett. 4(4):3449–3456.

26.

Nasir

, Cool

, Sassani

2019b. Acoustic emission monitoring of sawing process: artificial intelligence approach for optimal sensory feature selection. Int J Adv Manuf Technol. 102(9-12):4179–4197.

27.

Nasir

, Fathi

, Kazemirad

2021. Combined machine learning–wave propagation approach for monitoring timber mechanical properties under UV aging. Struct Health Monit. 20(4):2035–2053.

28.

Nasir

, Kooshkbaghi

, Cool

2020. Sensor fusion and random forest modeling for identifying frozen and green wood during lumber manufacturing. Manuf Lett. 26:53–58.

29.

Nasir

, Kooshkbaghi

, Cool

, Sassani

2021. Cutting tool temperature monitoring in circular sawing: measurement and multi-sensor feature fusion-based prediction. Int J Adv Manuf Technol. 112(9):2413–2424.

30.

Nasir

, Nourian

, Zhou

, Rahimi

, Avramidis

, Cool

2019. Classification and characterization of thermally modified timber using visible and near-infrared spectroscopy and artificial neural networks: a comparative study on the performance of different NDE methods and ANNs. Wood Sci Technol. 53(5):1093–1109.

31.

Nasir

, Sassani

2021. A review on deep learning in machining and tool monitoring: methods, opportunities, and challenges. Int J Adv Manuf Technol. 115(9):2683–2709.

32.

Nourian

2018. Thermal modification of western hemlock (Tsuga heterophylla) [M.Sc. Thesis]. Vancouver (BC): University of British Columbia; p. 72.

33.

Oliveira

, Wallace

JW.

2001. Examining some drying/re-drying options for 105 mm × 105 mm green hem-fir lumber. Forintek Canada Corp. Report prepared for CFLA; p. 17.

34.

Park

, Lek

2016. Artificial neural networks: multilayer perceptron for ecological modeling. In: Sven Erik Jørgensen, editor. Developments in environmental modeling. Vol. 28. Elsevier; p. 123–140. https://doi.org/10.1016/B978-0-444-63623-2.00007-4.

35.

Perre

2007. Fundamental wood drying. In: European COST. France: Nancy; p. 366.

36.

Rahimi

, Avramidis

2022. Predicting moisture content in kiln dried timbers using machine learning. Eur J Wood Wood Prod. 80:681–692.

37.

Rahimi

, Avramidis

, Lazarescu

2021. Estimating moisture content variation in kiln dried pacific coast hemlock. Holzforschung. DOI:10.1515/hf-2021-0080

38.

Rahimi

, Faezipour

, Tarmian

2011. Drying of internal-check prone poplar lumber using three different conventional kiln drying schedules. J Indian Acad Wood Sci. 8(1):6–10.

39.

Rahimi

, Nasir

, Avramidis

, Sassani

2021. Wood moisture monitoring and classification in kiln-dried timber. Struct Control Health. 29(4): e2911.

40.

Sackey

, Avramidis

, Oliveira

LC.

2004. Exploratory evaluation of oscillation drying for thick hemlock timbers. Holzforschung. 58(4):428–433.

41.

Rohrbach

2008. Schedule and post-drying storage effects on Western hemlock squares quality [M.Sc. Thesis] Vancouver (BC): University of British Columbia; p. 113.

42.

Sablani

, Kacimov

, Perret

, Mujumdar

, Campo

2005. Noniterative estimation of heat transfer coefficients using artificial neural network models. Int J Heat Mass Transfer. 48(3-4):665–679.

43.

Salas

, Moya

2014. Kiln-, solar-, and air-drying behavior of lumber of Tectona grandis and Gmelina arborea from fast-grown plantations: moisture content, wood color, and drying defects. Dry Technol. 32(3):301–310.

44.

Schubert

, Luković

, Christen

2020. Prediction of mechanical properties of wood fiber insulation boards as a function of machine and process parameters by random forest. Wood Sci Technol. 54(3):703–713.

45.

Shahverdi

2015. Kiln drying optimization for quality hem-fir lumber [M.Sc. Thesis]. Vancouver (BC): University of British Columbia; p. 91.

46.

Shahverdi

, Oliveira

, Avramidis

2017. Kiln-drying optimization for quality pacific coast hemlock timber. Dry Technol. 35(15):1867–1873.

47.

Siau

JF.

1995. Wood: influence of moisture on physical properties. Blacksburg: Department of Wood Science and Forest Products, Virginia Polytechnic Institute and State University. p. 227.

48.

Steinberg

2009. ‘CART: classification and regression trees. In: Wu

, Kumar

, editors. The top ten algorithms in data mining. Boca Raton (FL): CRC Press; p. 179–202.

49.

Sugeno

1985. Industrial applications of fuzzy control. Amsterdam: Elsevier.

50.

Van Blokland

, Nasir

, Cool

, Avramidis

, Adamopoulos

2021a. Machine learning-based prediction of internal checks in weathered thermally modified timber. Constr Build Mater. 281:122193.

51.

Van Blokland

, Nasir

, Cool

, Avramidis

, Adamopoulos

2021b. Machine learning-based prediction of surface checks and bending properties in weathered thermally modified timber. Constr Build Mater. 307:124996.

52.

Wada

2013. Internal moisture movement in hem-fir timbers exposed to ambient conditions following kiln drying [M.Sc. Thesis] Vancouver (BC): University of British Columbia; p. 65.

53.

Watanabe

, Hayashi

, Kobayashi

2016. Probabilistic evaluation of the final moisture content of kiln-dried lumber using the bootstrap method. J Wood Sci. 62(6):479–486.

54.

Watanabe

, Kobayashi

, Matsushita

, Saito

, Kuroda

, Noshiro

2014. Application of near-infrared spectroscopy for evaluation of drying stress on lumber surface: a comparison of artificial neural networks and partial least squares regression. Dry Technol. 32(5):590–596.

55.

Watanabe

, Matsushita

, Kobayashi

, Kuroda

2013. Artificial neural network modeling for predicting final moisture content of individual Sugi (Cryptomeria japonica) samples during air-drying. J Wood Sci. 59(2):112–118.

56.

Witczak

, Korbicz

, Mrugalski

, Patton

RJ.

2006. A GMDH neural network-based approach to robust fault diagnosis: application to the DAMADICS benchmark problem. Control Eng Pract. 14(6):671–683.

57.

, Avramidis

2006. Prediction of timber kiln drying rates by neural networks. Dry Technol. 24(12):1541–1545.

58.

Yilmaz

, Kaynar

2011. Multiple regression, ANN (RBF, MLP), and ANFIS models for prediction of swell potential of clayey soils. Expert Syst Appl. 38(5):5958–5966.

59.

Zhou

, Rahimi

, Avramidis

, Fang

2020. Species- and moisture-based sorting of green timber mix with near infrared spectroscopy. BioResources. 15(1):317–330.