Enhancing Deep Line Segment Detection and Performance Evaluation for Wood: A Deep Learning Approach with Experiment-Based, Domain-Specific Implementations

Abstract

In recent decades, wood structures have gained significant attention for their ecological benefits and architectural versatility. The performance of wood, a popular construction material, often depends on the integrity of its connections. This study focuses on bolted glulam timber connections, which are strong but prone to cracks that pose structural health challenges. Traditional crack evaluation methods are manual, time-consuming, and error-prone. To address these issues, this research proposes a two-stage performance evaluation method. In the first stage, an innovative approach called ‘Enhanced Deep Line Segment Detection’ (Deep LSD), a non-supervised machine learning technique, is used for crack detection without relying on large, annotated datasets, thus enhancing efficiency and adaptability. In the second stage, cyclic loading assays simulate varying damage stages to collect data and establish a correlation between crack states and connection damage. The Park and Ang damage model is employed within this framework to assess the extent of damage. The efficacy of enhanced deep LSD is confirmed by comparing detected crack areas with ground truth measurements, yielding a high R-squared value of 0.98 and a minimal error margin of 1.41. Additionally, a damage index based on the Chinese standard (GB/T 24335-2009) is used to classify damage across different connection groups, ensuring robustness and alignment with established practices.

1. Introduction

From the earliest stages of human civilization, wood has been a crucial resource. Due to its cost-efficiency and enduring nature, wood has become a key material in the construction industry. Modern construction practices are now embracing wood in a broader way. Its applications extend to diverse projects like residential and commercial buildings [1,2], bridges [3], and even sports stadiums [4]. Above all, wooden buildings can demonstrate similar strength and resilience to conventional steel and concrete structures [5].

According to an investigation of structural damage in recent earthquakes, wood-framed buildings also exhibit excellent seismic behavior [6]. In order to construct large timber structures, individual wood pieces are often joined together using different types of connections. Timber joints as connections among frame components represent vulnerable parts due to weakened cross-sections and have an important role in dealing with the structural performances of timber structures [7]. However, the deformation and shrinkage of timber members caused by various issues (i.e., excessive load, wind vibration, or earthquake) can lead to damage to timber connections over time [8], thus potentially leading to the collapse of the whole structure [9,10,11]. A prior investigation has demonstrated that the stability and safety of entire timber structures greatly depend on the integrity of connections [12], and the insufficient integrity of connections is the major cause of failures in timber structures [13]. In addition, some researchers noticed that the peak stress, rotational stiffness, and bearing capacity of joints deteriorate due to damages in timber connections, especially if the joints are cracked [14]. The strong orthotropic anisotropy and complex microstructure of timber lead to extremely complicated development laws of cracks. General cracks will not affect the safety of the structure, but under the effect of load and environment, after the general cracks further expand into dangerous cracks, they will cause local structural fractures, resulting in the continuous collapse of the overall structure [15,16]. Consequently, the implementation of careful maintenance and monitoring programs becomes essential to ensure the sustained performance and safety of wooden structures throughout their service life.

The conventional approach to wood defect detection primarily employs manual visual inspection [17]. This method has some drawbacks, including high operational costs, low efficiency, and a propensity for both missed detections and misclassifications. The emergence of machine vision technology presents a promising avenue for addressing these limitations and enhancing the efficiency of wood defect detection. Automated visual inspection systems offer a multitude of advantages, including improved detection accuracy and consistency, reduced labor costs, and diminished labor intensity [18]. The realm of defect detection has witnessed a surge in the utilization of deep learning techniques in recent years.

Among the diverse machine learning [19] and deep learning algorithms, object detection algorithms have emerged as a prominent choice for defect detection tasks. Within the realm of deep learning, object detection algorithms have garnered significant attention for their application in wood defect detection tasks [20,21]. These algorithms can be broadly categorized into two distinct classes based on their processing methodologies: one-stage and two-stage [22]. One-stage algorithms, such as SSD [23] and YOLO [24,25], have been successfully employed by prior researchers for wood defect detection. One-stage semantic segmentation-based models such as UNet were employed before by previous authors [26,27,28]. Conversely, two-stage algorithms, exemplified by Faster R-CNN [29] and Mask RCNN [30], were also used by a few other researchers.

One of the main challenges with one-stage and two-stage deep learning algorithms is their need for large, well-annotated datasets [31], which are expensive and time-consuming to develop. Detection accuracy is tied to the nature of the training data, making it difficult to detect new types of unseen defects. Therefore, there is a need for more robust approaches that do not depend on extensive training datasets.

Wood has several defects [32], but crack detection in wood is very challenging as it contains lot of similar patterns to cracks, which hinders detection accuracy [33]. Some researchers have explored various thresholding techniques for segmenting crack pixels from the background in images [34,35,36]. Line segment detectors (LSDs) are another widely used approach with a wide range of practical applications in various fields, such as in detecting lines associated with walls and doors to understand room layouts [37], identifying house structures from electrical plans [38], and contributing to satellite object detection by providing geometric features [39]. This study is among the first to incorporate this specific approach for wood crack detection.

Bolted glulam timber connections have been extensively recognized for their outstanding mechanical characteristics and their simple installation process. In this study, cyclic loading assays were administered to bolted glulam timber connections to collect varied crack data across different damage stages. To measure the extent of structural damage under seismic action, it is necessary to define damage evaluation indicators. At present, the damage evaluation of connections is mainly based on damage indicator models [40]. Numerous researchers have proposed damage indicator models for structural components, including models based on high-cycle fatigue hysteresis loops, displacement, energy accumulation, and the combined effects of displacement and energy accumulation. A few examples of these models include the Miner damage model [41,42], the Darwin damage model [43], the Gosain damage model [44], the Newmark damage model [45], the Ou Jinping damage model [46], the Chai damage model [47], the Kumar damage model [48], and the Park and Ang damage model [49,50]. Due to the specific nature of our experiment, the Park and Ang damage model was selected as the most suitable tool for assessing damage within the experiment’s framework.

In the experimental setup, grain and knots within wood imagery, which often mimic cracks, posed significant challenges for precise crack detection. The removal of such background distractions, while maintaining the fidelity of crack profiles, was found to be complex due to the difficulties in semantically distinguishing non-crack features. Subsequently, cracks were detected autonomously using the enhanced deep LSD method, and the detected crack areas were compared with ground truth measurement areas. Finally, a damage performance evaluation of the connections was conducted using a damage-index function based on the observed damage at each stage. This evaluation provided a quantitative measure of the structural integrity, allowing for a detailed assessment of how cracks impact the overall performance of the connections. The motivation for this evaluation was to ensure a robust and reliable method for predicting the lifespan and safety of wood structures, ultimately contributing to more durable and sustainable construction practices.

2. Experimental Tests

2.1. Material Properties

In the cyclic loading tests of the connections, the material properties of glued laminated timber, steel plates, and bolts play a crucial role in the mechanical performance of the connection. The timber species selected for the experiments is Douglas fir, the steel plates are made of Q345 steel, and the bolts are of grade 8.8. The material properties of glued laminated timber are significantly influenced by factors such as tree species, necessitating comprehensive material testing. This includes tests for compressive strength parallel to grain, compressive strength perpendicular to grain, shear strength parallel to grain, tensile strength perpendicular to grain, and embedment strength parallel and perpendicular to grain. The testing protocol followed the American test standard ASTM D143-14 [51]. Additionally, the density and moisture content of the timber were also measured. The material properties of the glued laminated timber are summarized in

Table 1. Property parameters of glulam materials.

Property of Material	Parallel to Grain	Perpendicular to Grain
Compressive strength (MPa)	44.03	7.58
Compressive modulus of elasticity (MPa)	11,534	313
Shearing strength (MPa)	2.16	-
Tensile strength (MPa)	-	2.03
Embedment yield strength (MPa)	39.66	19.09
Embedment ultimate strength (MPa)	45.08	27.57
Air-dry density (kg/m3)	527.83
Moisture content	10.46%

.

2.2. Specimen Configurations

In the cyclic loading tests on glued laminated timber beam–column connections, two types of connections were designed, as depicted in Figure 1. The cross-sectional dimensions of the beams were 350 mm by 200 mm, and the length was 1000 mm; the columns had cross-sectional dimensions of 350 mm by 300 mm. The spacing between bolts, as well as the edge and end distances of the beam–column connection specimens, complied with the relevant requirements outlined in the “Technical Code of Glued Laminated Timber Structures” GB/T 50708-2012 [52].

The classification of the cyclic loading tests on the connections is outlined in

Table 2. Connection testing set.

Specimen Number	Sapwood Thickness t1 (mm)	Connection Form	Bolt Diameter d (mm)	Steel Plate Thickness t (mm)
S1-1	95	Connection Form 1	16	10
S1-2	95	Connection Form 1	16	10
S2-1	95	Connection Form 2	20	10
S2-2	95	Connection Form 2	20	10

. The experiment is divided into two groups, designated as S1 and S2, each comprising two specimens. Specifically, Group S1 utilizes Connection Form 1, while Group S2 employs Connection Form 2.

2.3. Test Setup and Loading Protocol

The test setup for the experiment is illustrated in Figure 2. The test employs a cyclic loading method, utilizing six displacement transducers (LVDTs) to measure the rotational displacement of the connections. LVDT1 is attached to the loading end of the specimen to record the displacement at the specimen’s loading end. LVDT2 and LVDT5 are connected to the side ends of the steel plate, positioned at the horizontal cycle of two rows of bolts, to monitor the rotational behavior of the steel plate during loading.

LVDT3 and LVDT4 are symmetrically attached to the glued laminated timber components on either side of the steel plate, aligned with the midpoint between the two rows of bolts. LVDT6 is connected to the end of the glued laminated timber column to measure the horizontal displacement of the column. The cyclic loading tests are conducted in accordance with the FEMA 461 standard [53]. Displacement-controlled loading is employed for these tests.

2.4. Damage Modes

The experimental loading and resulting damage to the connections are depicted in Figure 3, Figure 4, Figure 5 and Figure 6. For Group S1 (Figure 3 and Figure 4), S1-1 and S1-2 emitted distinct sounds at the 21st and 23rd loading cycle, respectively, with noticeable lifting at the beam ends, followed by continuous noise in the subsequent cycle. Visible significant cracks appeared on the exterior of the connections at the 29th and 30th cycle, and by approximately the 31st level, the connections exhibited severe cracking and splitting through the structure.

For Group S2 (Figure 5 and Figure 6), the S2-1 and S2-2 made noise at the 22nd and 16th loading cycle, respectively. Visible significant cracks developed at the 25th and 27th cycle, and by the 29th and 31st cycle, there was severe crack propagation and penetration through the connections.

3. Tiny Crack Detection in Wood

This study enhanced the DeepLSD model [54], which is based on a simple UNet [55] encoder–decoder structure, and used the conventional LSD detector [56] as its base detector, leveraging image gradients to identify lines in an image. Before feeding the model, we employed a morphological operation to surpass the crack-like wood texture; later, the DeepLSD model detected the line and converted image gradient to line distance, which was filed to compute the length of the crack profile.

3.1. Repeated Texture Suppression

In our experimental wood, images often contain wood grain or knots that closely resemble cracks. This similarity challenges the model’s ability to accurately detect actual cracks. Therefore, we need to remove this background clutter from the pre-processed image while preserving the crack profile itself. However, identifying and removing such non-crack related features in a semantically meaningful way remains a significant challenge.

To address the challenge of distinguishing cracks from repetitive patterns like wood grain, we leverage morphological operations as a pre-processing step. These image processing techniques analyze the image and extract key features like borders and simplified outlines. This not only improves the representation of regional shapes but also aids in differentiating cracks from the background. Specifically, this study employs binary morphological operations, including opening and closing, to refine the image and make depth features more prominent.

Opening, by first eroding and then dilating the image, can help remove small wood grain patterns and smoothen crack edges, potentially making actual cracks more distinct. Conversely, closing, achieved through dilation followed by erosion, refines the shape of the crack profile in a smoother way, potentially by filling small gaps within the cracks.

Let us define an image as two-dimensional image array $I$, whose range is $[N_{\min },N_{\max }]$. As a function, $I:{ℝ}^2\to [N_{\min },N_{\max }]$. $E$ denotes the collection of neighborhood’s points centered at the origin. In our methodology, we exclusively examine structuring elements that remain unchanged through translation, thereby associating them directly with subsets of ${ℝ}^2$. These elements are designated as linear structuring elements when the subset corresponds to a line segment. Basic morphological operators are defined with respect to the structuring element $E$, scaling factor k, image $I$, and point $P_o\in {ℝ}^2$ as follows.

The erosion process is as follows,

$${\epsilon }_k(E,I)(P_0)=\underset{P\in P_0+k.E(P_0)}{\min }I(P)$$

(1)

Here, $\epsilon$ represents erosion, which reduces the image based on the minimum value in the neighborhood defined by E. Again, the dilation process is denoted below,

$${\delta }_k(E,I)(P_0)=\underset{P\in P_0+k.E(P_0)}{\max }I(P)$$

(2)

In Equation (2), δ represents dilation, which expands the image based on the maximum value in the neighborhood. The opening operations are as follows:

$${\gamma }_k(E,I)={\delta }_k(E,{\epsilon }_k(E,I))$$

(3)

Here, δ represents dilation, which expands the image based on the maximum value in the neighborhood. ${\gamma }_k$ represents opening, which is an erosion followed by a dilation.

$${\varphi }_k(E,I)={\epsilon }_k(E,{\delta }_k(E,I))$$

(4)

${\varphi }_k$ represents closing, which is a dilation followed by an erosion. This pre-processing step prepares the image for subsequent crack detection algorithms by enhancing the relevant features and suppressing background noise and repeated crack resemble texture.

3.2. Principle of Extracting Crack Segments

Xue et al. [57] proposed the concept of representing line segments using a spatial influence map. They introduced a method to generate a 2D directional map for each image pixel, indicating the proximity to the nearest point on a linear structure. This approach enables the representation of line segments as a continuous, 2-channel image, making it well suited for deep learning algorithms. The spatial proximity map (D) analyses each pixel within an image and calculates its distance to the nearest point on a line segment; the orientation field $O$ indicates the direction of the nearest line.

$$\begin{gathered} D=\sqrt{u^2+v^2} \\ O=\arctan {\displaystyle \frac{u}{v}}+{\displaystyle \frac{\pi }{2}}\mathrm{mod}\pi \end{gathered}$$

(5)

These two metrics can be easily derived from the 2D displacement field $(u,v)\in {ℝ}^{h\times \omega }\times {ℝ}^{h\times \omega }$, which points to the nearest location on a line, where $(h\times w)$ denote the size of the image in I range.

The initial Deep LSD architecture comprises multiple convolutional blocks, which perform dimensionality reduction on the initial image by a scale of 8 and then upscale it back to its original resolution. Dimensionality reduction is achieved through three consecutive 2 × 2 average pooling processes, while upscaling is performed using bilinear interpolation. A skip connection is employed before each dimensionality reduction layer and fused with the output of the equivalent upscaling layer. Convolutional layers and batch normalization [58] are sequentially followed by ReLU activation [31], except for the last layer of every layer. The two output layers employ ReLU activation for the distance field and Sigmoid activation for the angle field, omitting batch normalization.

The sigmoid activation function is applied to the angle field, mapping its values to a range between 0 and 1. This output is then scaled by π to produce an angle within the range of 0 to π. The distance field accurately captures small distances near lines, which is crucial for achieving high accuracy. The distance field branch concludes with a ReLU activation function. This activation ensures all output values are non-negative. The resulting output is a normalized distance field, denoted by $\widehat{d}$, with dimensions of height and width. To obtain the final distance field, a denormalization step is applied using the following formula:

$$\widehat{d}=r·e^{-{\widehat{d}}_n}$$

(6)

Here, r measured in pixels, establishes a zone surrounding every line in the image. This zone encompasses only those pixels that fall within a distance of r pixels from the line itself.

Now, let us define distance and direction fields into an alternative image gradient magnitude, m, and orientation angle, $\theta$.

$$\left\{\begin{array}{l}m=r-\widehat{d}\\ \theta =\widehat{O}-{\displaystyle \frac{\pi }{2}}\end{array}\right.$$

(7)

Here, $\widehat{O}$ is updated truth angle field, where $\widehat{O}\in {[0,\pi ]}^{h\times \omega }$. The predicted angle aligns with the orientation of the lines but remains perpendicular to the variations in image gradient. The predicted angle aligns with the orientation of the lines but remains perpendicular to the variations in image gradient. Maximum intensity value for any pixel situated directly on a line is determined by the parameter r.

Later, it is necessary to calculate the vanishing points (VPs) linked to the predicted line segments. This computation leverages the multi-model fitting algorithm known as Progressive-X [59]. A stringent inlier threshold is employed to ensure that only pertinent lines are associated with each vanishing point. Subsequently, the optimization process is conducted independently for each line. This optimization is formulated as a weighted, unconstrained least squares minimization to solve three distinct cost functions.

$$L={\lambda }_1{C}_1+{\lambda }_2{C}_2+{\lambda }_3{C}_3$$

(8)

Considering a sample of neighboring points, $P_i$, from n points, uniformly sampled along a line segment, E. The orientation of this line segment is represented by the angle ${\theta }_E$. Additionally, $v_E$ signifies the vanishing point associated with line segment E. To optimize the fitting process, three cost functions are detailed as shown here:

$$C_1={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{p_i\in p}\left(1-(\cos \left(\widehat{O}(p_i)-{\theta }_E)\right)\right)}$$

(9)

$$C_2={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{p_i\in p}\widehat{d}}(p_i)$$

(10)

$$C_3=dvp(E,v_E)$$

(11)

Here, $dvp$ quantifies the distance between a line and a vanishing point. As the experimental images in this study contain many small textures and lack horizontal cracks, the majority of cracks are oriented vertically. The sample of neighboring points, $P_i$, on line segment E is set to five to capture the exact crack profile. Additionally, a filter was employed to eliminate any horizontal crack patterns.

3.3. Detection Accuracy Evaluation

After detecting the crack area of all the experimental images, it is essential to conduct a quantitative assessment to evaluate the performance of the detection accuracy. However, visually determining the detection accuracy of the model is difficult. Therefore, evaluation metrics are necessary to evaluate the model’s effectiveness. For detection tasks, the performance of the model is usually indicated by the error between the detected crack area and the measured area by the experimental investigation. This study employs three standard evaluation metrics to measure the discrepancy between the detected area and the measured area by the experimental investigation, including Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), and determining coefficient (R²).

The MAE is the average of the absolute errors between actual area of a crack and detected area, which directly reflects the actual error situation of the model detection results. The mathematical equation is below.

$$MAE={\displaystyle \frac{1}{m}}{\displaystyle \sum _{j=1}^m|a_j-\widehat{a_j}|}$$

(12)

Here, m is the total number of experimental images, $a_j$ represents the actual detected area, and $\widehat{a_j}$ denotes the model detected area.

The RMSE is the square root of the mean squared difference between detected and actual areas. Lower RMSE values indicate lower prediction errors and better model performance. Larger errors are given higher weights. The relationship is denoted below.

$$RMSE=\sqrt{{\displaystyle \frac{1}{m}}}{\left(a_j-\widehat{a_j}\right)}^2$$

(13)

Determining coefficient (R²) is a metric used to assess how well the detected crack areas from the model fit the actual crack areas measured in the experiment. The equation is as follows:

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{j=1}^m\left(a_j-\widehat{a_j}\right)}}{{\displaystyle \sum _{j=1}^m\left(a_j-\overline{a}\right)}}}$$

(14)

A value of R² closer to 1 indicates a better fit between the model’s estimates and actual values, suggesting higher accuracy in the model’s crack detection.

3.4. Evaluation and Visualization

This study utilized 100 images as experimental data for crack detection by the model. Figure 7 presents a set of example detections.

The crack areas are highlighted in gold, while the green points mark the discontinuity or endpoint of a crack profile. Figure 7A demonstrates the model’s effectiveness in detecting crack areas, even with discontinuities. The model correctly places an endpoint and resumes detection from the new starting point. In Figure 7B–D, the thickness of the detected area line varies according to the crack width. Despite the presence of textures resembling cracks, the model successfully identifies the actual cracks.

To further validate the model’s crack detection accuracy, Figure 8 compares the detected areas with the actual measured areas for the 100 crack profiles.

The Figure 8 shows a high degree of overlap between the detected and measured areas, indicating the successful detection of nearly all of the cracks. This visual assessment suggests the model’s robustness in accurately detecting cracks.

Table 3. Model evaluation results.

Evaluation Metrics	Value
R2	0.98
MAE	0.68
RMSE	0.23

provides quantitative evaluation metrics for the model’s performance. An R-squared (R²) value of approximately 0.98 indicates a strong correlation between the predicted and actual crack areas. Additionally, the low Mean Absolute Error (MAE) of 1.41 and Root Mean Squared Error (RMSE) of 2.01 further support the model’s effectiveness in accurately detecting crack areas. This prediction accuracy achieved is primarily attributed to the employment of the enhanced deep LSD method. By effectively differentiating crack and non-crack features within complex wood imagery, this approach significantly enhances the precision of crack detection. This superior discriminative capability is crucial for accurately identifying and localizing cracks, ultimately leading to the model’s robust performance.

4. Performance Evaluation

4.1. Damage Indicator Model

The damage indicator for a structure or component, denoted by D, represents the ratio of a specific seismic response parameter to its allowable limit. A value of D equal to zero corresponds to no damage to the structure. Conversely, a value of D equal to one indicates complete structural collapse. The expression for calculating D is provided below [60].

$$D(x_1,x_2,x_3\dots )={\displaystyle \frac{D_r(x_1,x_2,x_3\dots )}{D_{ru}}}$$

(15)

The damage indicator function D(x₁, x₂, x₃…) combines various parameters that influence the damage severity of a structure or component, where x₁, x₂, x₃… are key parameters influencing the extent of damage. These parameters could include factors such as displacement, stress, energy dissipation, and material properties. Cumulative damage quantity D_r(x₁, x₂, x₃…) represents the total damage accumulated over time or under repeated loading cycles. It accounts for the cumulative effect of damage on the structural integrity. The critical cumulative damage quantity D_ru is the threshold value at which the structure or component can no longer withstand the applied loads and experiences complete failure.

Considering the actual damage state of the connections and the results of the reciprocating test data, the damage index D is defined within the range [0,1] for convenient damage evaluation. In this study, the Park and Ang model was selected as the damage model for the connections.

The Park and Ang damage model is a classic damage index model that considers the dual effects of energy and displacement. Its expression is as follows:

$$D_{PA}={\displaystyle \frac{S_{i,\max }}{S_u}}+\beta {\displaystyle \frac{{\displaystyle \int dE}}{F_y{S}_u}}$$

(16)

$${\mu }_{si}={\displaystyle \frac{S_i}{S_y}}$$

(17)

Within the equation, S_i,max represents the maximum displacement prior to the i th cycle; S_u denotes the ultimate displacement of the component under monotonic loading; and F_y represents the yield load of the component. β is the cyclic loading influence coefficient. For steel structures, β is typically set to 0.025. For concrete structures, β falls within the range of 0 to 0.85. The formula for β is as follows:

$$\beta =(-0.447+0.073\lambda +0.24n_0+0.314{\rho }_t)\times {0.7}^{{\rho }_{\omega }}$$

(18)

Here, λ represents the shear span ratio of the component, n denotes the axial compression ratio of the component, ρ_t represents the longitudinal reinforcement ratio of the component, and ρ_ω represents the volumetric reinforcement ratio of the component.

The Park and Ang damage model was primarily developed based on studies of seismic damage in reinforced concrete structures. The damage index, denoted by D_PA, quantifies the extent of damage. When D_PA is less than 0.4, it indicates repairable damage. When D_PA falls between 0.4 and 1.0, it signifies irreparable damage. When D_PA is greater than or equal to 1.0, it indicates structural collapse. The value of β can be determined by fitting experimental data to ensure that D_PA equals 1.0 when the structure collapses.

4.2. Experimental Damage Evaluation and Analysis

The damage index results for different loading cycles obtained through the Park and Ang model are shown in Figure 9. By combining the damage results of the connections under a different loading cycle in Section 2.4 of this paper with the damage indices under different loading cycles, the damage evaluation of the experimental connections can be obtained.

The Chinese standard, “Classification of Seismic Damage Cycle for Buildings and Structures” (GB/T 24335-2009) [61], defines the damage cycle for commonly used buildings and structures. The classification is based primarily on the extent of damage to load-bearing components, while also considering the extent of damage to non-load-bearing components and the ease of repair and the degree of loss of function. The classification consists of five cycles: no damage, minor damage, moderately damaged, severely damaged, and collapse.

For S1 group connections, when the damage index 0 < D < 0.4, the loading level is within 26 cycles. For S2 group connections, the loading level is within 24 cycles. As can be seen from Section 2.4, no obvious cracks were observed in the connections within this level range, but there was a continuous crackling sound at some loading cycles, which is predicted to be the connection of micro-damage inside the wood. Therefore, for this type of connection, the damage index range 0 < D < 0.4 is classified as essentially intact.

For S1 group connections, when the damage index 0.4 ≤ D < 0.65, the loading level is between 27 and 30 cycles. For S2 group connections, the loading level is between 25 and 28 cycles. Visible cracks and other damage phenomena begin to appear in the connections. Based on the form of crack occurrence, the connection damage state can be divided into two categories: minor damage and moderate damage. Referring to Section 2.4 and the damage state of the connections during the experimental loading process, when 0.4 ≤ D < 0.5, the connection damage state is defined as the minor damage state visualized in Figure 10; when 0.5 ≤ D < 0.65, the connection damage state is defined as the moderate damage state depicted in Figure 11.

When D ≥ 0.65, the loading level for S1 group connections is above 31 cycles, and the loading level for S2 group connections is above 29 cycles. At this time, the connections have all experienced different degrees of crack expansion, the crack widths are large, the cracks cannot be closed when the connection angle is restored, and there is even severe penetration when the connection is destroyed. These states can be defined as severe damage and collapse.

5. Conclusions

Wood construction relies on strong connections like bolted glulam timber connections, where even minor cracks can impact performance. Manual crack detection is inefficient and error-prone. This study conducted an experimental investigation on bolted glulam timber connections to gather diverse crack data at various damage stages. During the experiments, grain patterns and knots in the wood imagery often resembled cracks, creating significant challenges for accurate crack detection. Eliminating these background elements while preserving the integrity of the crack profiles was complex due to the difficulty in differentiating between crack and non-crack features. The Park and Ang damage model was chosen as the most appropriate tool for assessing damage within the experimental framework. Later, this study also proposes enhanced deep LSD, a non-supervised machine learning technique, for efficient automated crack detection, validated with cyclic loading assays and the Park and Ang damage model. The key outcomes of this research are outlined below.

• This study proposed an enhanced deep LSD, which eliminates the need for extensive training data, making it more efficient and adaptable for wood crack detection tasks. The enhanced deep LSD achieved a high R-squared value of 0.95 and a minimal error of 0.34, suggesting it has significant potential as a valuable tool for automated crack detection in bolted glulam timber connections.
• The observed damage exhibited a strong correlation with the Park and Ang damage index. Under cyclic loading, specimens demonstrated essentially intact behavior when the damage index (D) fell within the range of 0 ≤ D < 0.4. As the damage index increased to 0.4 ≤ D < 0.5, the damage state transitioned to minor. For specimens with a damage index in the range of 0.5 ≤ D < 0.65, the damage was classified as moderate. Finally, the states with a damage index exceeding 0.65 were defined as severe damage and collapse.
• Future work will involve exploring alternative surrogate models to benchmark the accuracy of the enhanced deep LSD method.