Autonomous bulldozers represent cutting-edge technology in field earthwork construction, and semantic segmentation is vital for their environmental understanding. However, challenges remain due to complex interference, specific engineering information, and unstructured objects within construction scenes. To address this issue, we develop a semantic segmentation network for LiDAR-camera data, termed the Multi-Modal Assisting Fusion Network (MMANet). To adapt to the complex interference in field earthwork construction, we implement a pixels-to-points joint augmentation technique that back-projects redundant image data into 3D formats, enhancing the network’s resistance to noise data. Furthermore, to mitigate the impact of vibration interference, we propose a graph-feature aggregating module that extracts additional features from point clouds. Finally, we present a multi-modal knowledge distillation fusion module, which improves the network’s understanding of engineering information and unstructured objects by leveraging cross-modal knowledge. To validate MMANet, we have established a field LiDAR-camera semantic segmentation dataset for earthwork construction consisting of 16 scenes, which, to our knowledge, is the first of its kind. MMANet demonstrates superior performance, achieving mean Intersection over Union (mIoU) scores of 80.6% and 79.5% on the nuScenes validation and test datasets, respectively, and an mIoU of 76.1% on the field dataset—surpassing U2MKD, MSeg3D, 2DPASS, EPMF, PMF, RandLA-Net, and Cylinder3D.

Current static and dynamic analysis research on earth-rock dams mainly relies on staged construction information and dam model from the design phase, lacking consideration of actual terrain and layered filling processes. To address these issues, this paper proposes a static and dynamic analysis method for earth-rock dams based on layered filling 4D parametric modeling under intelligent monitoring of layered filling construction. Based on layered filling construction information obtained from an intelligent monitoring system, this study develops two key aspects: To begin with, a 4D refined layered modeling method for earth-rock dams using UAV tilt photography and B-Rep (Boundary Representation) parametric modeling. This approach employs Delaunay triangulation to achieve UAV-based modeling of actual dam terrain, and establishes a 4D dam model integrated with construction information through B-Rep boundary reconstruction of triangular meshes and CATIA parametric constraint rules that match layered construction scales. In addition, a finite element static and dynamic analysis method based on construction layered loading is proposed. This method progressively applies loads and material parameters according to actual filling progress to simulate pre-earthquake static conditions, followed by seismic analysis using these initial conditions, thereby enabling static-dynamic analysis considering complex layered construction processes. Application in an earth-rock dam project in Southwest China demonstrates that the static analysis results under layered loading are consistent with the actually monitored settlement trends. The maximum calculated settlement of the dam body is 2230 mm, representing a 3% reduction in deviation compared to the traditional staged loading method. In the dynamic analysis, the acceleration response at the dam crest exhibits asymmetry between the left and right banks, with an amplification factor of 1.92, achieving a 9.5% improvement over the traditional method. The proposed method fully considers complex layered filling processes in earth-rock dam analysis, effectively improves result accuracy, and provides a novel approach for dam structural safety assessment.

Accurate assessment of the compaction quality of sandy gravel materials remains challenging because representative features are difficult to extract from roller acceleration signals. Existing intelligent compaction (IC) indicators are largely derived from handcrafted spectral features and therefore cannot fully capture the multi-periodicity, local cyclic patterns, and long-range temporal dependencies embedded in the signals. To address this limitation, this study proposes a multi-period 2D temporal convolution and hierarchical dependency learning framework for automatic compaction quality assessment. The framework first identifies dominant periods and folds one-dimensional acceleration sequences into a two-dimensional temporal representation, enabling joint characterization of intra-period and inter-period variations. It then employs 2D temporal convolution to learn local cyclic features and a binary tree-structured hierarchical dependency module to extract multiscale short- and long-range temporal dependencies. Finally, the learned features are mapped to dry density through a multilayer perceptron, and a new compaction quality index, MPNT, is established. Field validation on a large concrete-faced rockfill dam in northwest China with 105 in-situ test points demonstrates that MPNT achieves the highest correlation with rolling passes (R2 =0.8544) and dry density (R2 =0.8882), and improves prediction accuracy by 14.34% compared with ECP, while outperforming CMV, CCV, THD, and W. MPNT also exhibits superior robustness under complex working conditions.

Non-road construction equipment (NRCE) emissions are one of the major sources of urban carbon footprints and atmospheric pollution, particularly in large-scale civil engineering projects. Achieving high accuracy, costeffectiveness, and real-time monitoring of NRCE emissions remains a challenge, particularly given the limitations of current indirect technologies, which face difficulties in transient capture, data fusion, model interpretability, and data redundancy. To address these limitations, this study proposes an interpretable multimodal fusion virtual sensor framework, providing an efficient and economical alternative to expensive portable emissions measurement systems (PEMS). Key innovations include: (1) A signal-physics-guided heterogeneous multimodal feature extraction architecture is developed, utilising temporal convolutional networks (TCN) for local temporal steady-state features extraction from engine and environmental parameters, Transformers for kinematic parameters modelling, and VGGish for capturing high-frequency transient audio features. (2) An embedded multi-level interpretable fusion mechanism is established, employing adaptive gating units to dynamically regulate modality weights and quantifying input feature attribution via output gradients. (3) An interpretability-driven stepwise sensitivity pruning strategy is proposed to eliminate redundant input features with low contributions. Experimental results demonstrate that the framework achieves mean R 2 values for predictive performance exceeding 0.98 across 11 metrics derived from 5 primary pollutants, with a single prediction latency of only 0.7201 s on an industrial-grade embedded controller. Quantitative analysis indicates that the proposed framework reduces hardware deployment costs by approximately 99% compared to PEMS. This study provides a robust and transparent methodology for the large-scale, cost-effective, and real-time monitoring of NRCE emissions, effectively facilitating green construction and sustainable building management.

Accurate fracture modeling helps clarify the mechanical behavior of rock masses. However, traditional statistical and random simulation methods ignore correlations among fracture parameters and surface roughness and assume uniformly random fracture networks within the domain, while current advanced generative models suffer from training instability and limited ability to capture complex spatial distributions. We propose a denoising diffusion probabilistic model (DDPM)-based framework that learns the multivariate joint distribution of fracture parameters—dip direction, dip angle, trace length, aperture, roughness, and center coordinates (x,y,z)–to generate DFNs with realistic attributes while maintaining stable and efficient training. The method derives and integrates fracture center positions within a sampling window into joint generation for location optimization, aiming to improve the representation of spatial variability in fractures. Fractal-dimension analysis is coupled with non-uniform rational B-spline (NURBS) tensor products to model rough fracture surfaces. Validations on real engineering data under the presented implementation, using descriptive statistics, Kullback–Leibler (KL) divergence and Wasserstein distance, show that the proposed method achieves lowest deviation among baselines for dip direction, dip angle, trace length, aperture, and roughness. The mean aggregated KL and Wasserstein distances across five parameters and three random seeds are reduced by at least 59.41

The deformation evolution and its underlying mechanisms are crucial for understanding the clay compaction process, which is essential for assessing the compaction quality of core walls in rockfill dams. However, numerical analysis using the discrete element method (DEM) struggles to accurately capture the large-strain deformation of clay, particularly in representing viscous and hydraulic effects from the viscoelastic-plastic properties during rapid compaction. To address this, this study proposes a hydro-mechanical viscoelastic-plastic (HMVP) contact model. The model begins with a hardening-dependent force-strain law governing the steadystate response. It incorporates the viscous resistance governed by cohesion and compression rate to characterize viscosity, and the saturation-driven excess pore pressure to capture delayed fluid dissipation. The model is calibrated and validated against a series of quick-shear triaxial tests, particularly under high confining pressure up to 2 MPa, reproducing both the stress-strain response and morphology evolution of triaxial specimens. Key findings emerge: (1) the central region of triaxial specimen exhibits significant plastic deformation with increasing water content; (2) the steady-state support, governed by the hardening level, provides the effective compression resistance of soil skeleton; (3) viscous resistance in low-moisture specimen suppresses effective support enhancement and exhibits notable rate effects; and (4) the fluid dissipation volume and the resulting excess pore pressure in high-moisture specimen exhibit significant spatial heterogeneity, controlled by the localized drainage paths, which govern morphology evolution.

The discrete element method (DEM) provides critical advantages in analyzing clay hardening by capturing particle alignment. However, persistent computational distortion induced by particle overlaps has impeded accurate volumetric evolution during hardening. This study proposes a reconstructed methodology featuring three sequential innovations: (1) a solid-phase conservation development framework resolving the geometric penetration via rigid particle and compressive contact; (2) the hardening-evolutive-elastic-plastic (HEEP) contact model characterizing hardening behaviours through the hardening, friction, adhesion and plasticity modules; (3) a porosity mapping approach enabling explicit modeling via quantitative correlation between simulated and measured porosity. Validation across various experimental conditions including moisture, density and loading-unloading paths confirms the HEEP model's reliability, with calibration results across a broad moisture range (13%-25%) achieving an average RMSE below 0.004. The hardening exponent n governs the morphology of consolidation curve, while existing compressed gap delta e quantifies hardening level. The plasticity parameters precisely replicate elastic recovery and plastic rebound under varied hardening level, exhibiting significant moisture dependency.

The discrete element method (DEM) is widely adopted for investigating cohesive soil mechanisms at the microscale due to its capacity to directly capture particle-scale kinematics. However, its extension to the simulation of macroscopic behavior remains challenging. Applying insights from the microscale by merely upscaling particle sizes leads to incomplete physical mechanisms and poor response accuracy. To overcome these limitations, this study proposes a mesoscopic Hardening Plastic Cohesive (HPC) contact model, formulated through several advancements: (1) A solid-phase conservation framework accurately quantifies compressed void volume, which resolves persistent errors in porosity calculation; (2) Compression-plasticity-related parameters drive the soil's compressive strength and rebound collectively by governing the hardening level; (3) The friction coefficient between mesoscopic elements evolves with the saturating growth in the coordination number of their underlying microparticles during densification; (4) A tension-fracture-healing mechanism for cohesive soils is incorporated into the cohesion interaction against plastic yielding. High accuracy (average RMSE=0.0035) confirms the model's performance at 13%-25% moisture content by experimental validation. The key insights through Shapley Additive Explanations (SHAP) analysis reveal a decoupled control mechanism: hardening exponent nh governs the overall compressive strength, while elastic ratio lambda e determines rebound magnitude. Critically, the saturation-increasing friction mechanism, controlled by the friction exponent nf, is essential for correcting the flattening feature in the late stage of consolidation curves. The cohesive modulus Cc and cohesive exponent nc further enhance simulation accuracy, particularly in the early stage of consolidation. Ultimately, the optimized calibration workflow with clear physics-informed characterization of HPC model achieves a balance between calibration efficiency and high-fidelity results.

In complex construction environments, autonomous bulldozers require high-quality path planning. Traditional non-learning algorithms such as artificial potential fields and A* algorithm often struggle to handle dynamic obstacles and efficiency demands. As a representative Deep Reinforcement Learning method, the Soft Actor-Critic (SAC) algorithm demonstrates superior policy diversity and convergence stability compared to deterministic policy methods, providing an inherent advantage in planning challenging policies for difficult environments. Nevertheless, SAC also suffers from intrinsic limitations: sparse rewards, limited generalization capabilities, unstable initial policies, and insufficient exploration of extreme actions. These issues lead to convergence at local optima, failing to satisfy practical demands. To address these challenges, this study proposes a highly exploratory and adaptive SAC algorithm. First, enhancing the exploration capability of the policy through the integration of Random Network Distillation (RND), which utilizes intrinsic reward generated by state novelty to circumvent conservatism and local optima. Second, a training strategy based on Curriculum Learning is designed—starting from simple scenarios and gradually increasing complexity—to stabilize policies and accelerate convergence. Third, the reward function is designed by incorporating exploration components and combined with curriculum learning to balance exploration and exploitation. The proposed method significantly enhances exploration efficiency, policy learning efficiency, and problem-solving capabilities in real-world complex architectural environments, while providing high-quality paths for practical deployment. Ablation studies demonstrate that incorporating RND achieves 100% path planning success rates across diverse environments, whereas task success rates decline significantly with increasing environmental difficulty without RND. The curriculum learning mechanism provides robust guarantees for policy learning efficiency, boosting efficiency by 34.09% and 26.73% in specific environments, respectively. Compared to baseline algorithms, the proposed method outperforms both PPO and TD3 in policy adaptability, learning efficiency, and environmental complexity.

The rapid and accurate detection of concrete sand moisture content (MC) is crucial for ensuring concrete quality. However, existing unimodal detection methods are constrained by limited representative features and lack robustness. Multimodal operations often involve simple concatenation of features from different modalities, lacking potential interactivity among features. To address this issue, a novel robust cross-modal integration fusion model, which uses five branches to extract the features of images, near-infrared spectrum, and dielectric constant and a multilevel cross-modal integration fusion network to fuse these features, is proposed for the rapid detection of MC in concrete sand. Specifically, the multilevel cross-modal integration fusion network comprises a feature attention module, a cross-modal self-attention fusion module, and an integrated output module. The feature attention module enhances the feature representation from each modality, reducing the interference from redundant features and noise. The cross-modal self-attention fusion module employs a residual self-attention mechanism to deeply mine and fuse interactions between modalities while retaining low-level features, improving model accuracy and stability. The integrated output module is utilized to obtain more robust prediction results. The results show that the proposed model outperforms unimodal, traditional multimodal, and cross-modal methods on our concrete sand dataset, achieving excellent and robust prediction results for both machine-made sand (root mean square error ERMS=0.458, coefficient of determination R2=0.983, and residual predictive deviation DRP=7.900) and natural sand (ERMS=0.705, R2=0.984, and DRP=7.931). The detection time was within 71 s, significantly enhancing the detection frequency and efficiency, which provides a reliable solution for the rapid detection of MC in concrete sand.

Autonomous bulldozers are pivotal in earthwork construction, yet path planning in transition operations remains challenging due to complex scenarios and minimized steering control requirements. To enable effective path planning, this paper presents a deep reinforcement learning (DRL) framework coupled with a topological scene graph, termed earthwork construction scene graph (ECSG)-DRL. ECSG simplifies environment representation by focusing on task-related elements, reducing irrelevant complexity. Value points are designed in ECSG for action space representation, restricting unnecessary steering by targeting essential obstacle-avoidance positions. The DRL network is customized with dynamic action spaces to leverage value points' guidance. Validated in simulation and real-world scenarios, ECSG-DRL outperforms other methods (PPO, D3QN, TD3, SG-D3Q, ALN-DSAC, D3QN-PER, DDQN-CNN, A*, BINN, LPA*, and BIT*), achieving an 4.00 % improvement in simulation path distance, and 3.91 % and 16.83 % in path distance, 3.61x and 4.89x in steering angle, 2.01x and 89.14 % in steering frequency across static and dynamic real-world scenarios.

Large-scale earthwork transportation encounters queuing congestion and dynamic uncertainties, while existing methods ignore complex traffic behaviors and exhibit limited responsiveness and generalization. This paper proposes a multi-task Deep Reinforcement Learning (DRL) framework for the dynamic scheduling of large fleets across supply sites and traffic networks. In the framework, multiple agents interact in complex environments modeled by discrete-event simulation, utilizing long short-term memory networks that consider queuing behaviors and dynamic trends of transportation systems to allocate rational materials, supply sites, and routes collaboratively, with an invariant update strategy to balance generalization and task-specific optimization during training. Experiments demonstrate that the model generates dynamic schedules within 7 min, reducing transportation time by 24 %. The trained agent can adapt to the changing transportation demand in complex construction environments and enhance transportation efficiency. This paper demonstrates the potential of DRL in scheduling more complex construction projects and promoting real-time lean control of modern logistics.

Seepage in earth-rock dams is a critical factor affecting their safety and stability. Timely analysis of monitoring data and accurate seepage pressure prediction are essential for ensuring dam safety. However, current research primarily focuses on image-form Euclidean data for multi-sequence prediction, which limits the exploration of pairwise relationships among monitoring sequences and fails to adequately extract local semantic information in the temporal dimension. Moreover, previous studies often overlook the influence of environmental factors on seepage pressure. To address these issues, this study proposes a multi-scale spatiotemporal prediction model for seepage pressure using an adaptive graph neural network under dual-modal data-driven conditions. Considering the complex spatial relationships between multi-point seepage-pressure and environmental monitoring, adopting a dual-modal data-driven adaptive graph neural-network model as the basic architecture effectively represents the high-dimensional complex spatial topological relationship of monitoring sequences in non-Euclidean space, significantly improving the accuracy of seepage-pressure prediction. Furthermore, for the dynamic features of each sequence at different time scales, a combination of multi-scale convolutional kernels and temporal attention mechanisms is used to extract multi-scale features, addressing the lack of local semantic information in previous point-level inputs. Finally, the proposed model is applied to a dam-engineering project in Southwest China, demonstrating superior performance in spatiotemporal seepage pressure prediction compared to classical methods. Furthermore, the consideration of environmental quantity factors and multi-scale temporal features improved the proposed model by approximately 30% and 50% compared with models that do not account for these factors, respectively.

ABSTRACT Rapid urban flood mapping is crucial for timely risk alerts and emergency relief. Machine learning (ML)‐based mapping models emerge as a promising approach for fast, accurate inundation forecasts. However, current ML models often use precipitation features as inputs and predict maximum flood depth for all grid cells of a specific region simultaneously. This special design improves their prediction efficiency but limits their application in new regions. This study aims to create a highly adaptable, rapid urban maximum flood water depth mapping model based on the random forest regression algorithm and the extreme gradient boosting algorithm. Our mapping model additionally incorporates terrain and land‐use features, besides the precipitation feature, as input variables and generates the maximum water depth only for a grid cell in each mapping. Thus, it can be unchangeably applied to the grid cells in a new area when the model is fully trained. In the case study of Shenzhen, China, our ML‐based mapping model demonstrated excellent mapping ability in both training and validation sets. The coefficient of determination ( R 2 ) is consistently greater than or close to 95%. Furthermore, it revealed good generalization ability when directly applied to a new rainfall event ( R 2 = 0.875) and a new area ( R 2 = 0.810). Meanwhile, the time cost of the mapping model is less than 3 s, meeting the requirement for real‐time mapping. These results indicate that this highly adaptable model, once appropriately trained, can be applied to rapid urban flood severity mapping, which significantly reduces its use cost in urban flood management.

Enhancing positioning accuracy in rolling machinery is vital for quality and construction efficiency. To mitigate random noise interference in deep and narrow valleys, a multi-source positioning information fusion method utilizing an improved robust Kalman filter is proposed. This method adaptively selects optimal observations from GNSS, Robotic Total Station (RTS) and Ultra Wide Band (UWB) data, compensates for location deviation and data loss from noise interference, thus improving data robustness. The Kalman filter is improved by incorporating a thick tail Laplace distribution to dynamically adjust noise covariance, overcoming challenges with large random errors in data fusion and improving the robustness. Engineering tests show this method can adapt to complex and harsh environments in deep and narrow river valleys, with a compensation rate of over 97.33 % for data offset and loss issues, reducing localization offset rates by 7.72 % and loss rates by 1.64 % compared to single-method approaches, effectively improving the robustness, accuracy, and completeness of real-time monitoring results.

Accurate prediction of PM2.5 in underground powerhouse caverns group (UPCG) is of great significance for safeguarding the health of construction personnel and optimizing ventilation energy consumption. However, existing studies on PM2.5 prediction have not considered the impact of wind fields on the diffusion of PM2.5 or the spatio-temporal multi-scale information, which limits the accuracy of prediction models. To address these issues, this study proposed an improved Spatio-Temporal Graph Convolutional Network (STGCN) for predicting PM2.5 in UPCG under construction ventilation. Specifically, skip connections were incorporated between two feature pyramids to extract multi-scale spatio-temporal information from environmental features. Then, the PM2.5 diffusion distance map based on the Gaussian diffusion model was proposed as the adjacency matrix. Additionally, a Transformer-based spatio-temporal block fusion model was proposed to build a more efficient STGCN. The results demonstrate that the proposed model achieves smaller MAE and RMSE, as well as superior R2, compared to Transformer and CNN-LSTM in both single-step and multi-step prediction tasks. Ablation experiments confirmed the effectiveness of each proposed module. The model accurately predicts PM2.5 concentrations in UPCG, providing reliable support for ventilation requirements regarding decision-making.

Accurate forecasting of carbon dioxide (CO2) emissions from heavy construction machinery in large-scale infrastructure projects presents a viable avenue for mitigating climate change. However, most current studies neglect the intricate operating conditions of such machinery. Moreover, in complex construction environments with high-frequency vibrations and heavy dust, sensor failures leading to random data loss significantly increase the difficulty of emission prediction. Especially when high-emission periods account for a small proportion, existing methods struggle to effectively capture CO2 emission peaks. To address these challenges, this study proposes an improved Inverted-Transformer (iTransformer) multimodal interval prediction model for accurate CO2 emission forecasting in heavy construction machinery under conditions of random sensor failures. The model incorporates a Mixture of Experts (MoE) mechanism within the iTransformer framework, which adaptively adjusts expert weights for missing modalities, thereby reducing prediction errors caused by random data loss. Additionally, to capture localized CO2 emission peaks, this study introduces a Peak Capture Loss (PCL) function, which adjusts incremental emissions between adjacent time steps by supervising the differences between generated sequences, enabling the model to track abrupt emission variations. The Bootstrap method is also utilised to quantify and estimate uncertainty in the CO2 emission Interval Prediction. Case studies reveal that the proposed model achieves high prediction accuracy (coefficient of determination (R2) = 0.99), especially across various data missing rates (5 %, 10 %, 15 %), with the average R2 value increasing by approximately 6 %. This provides a novel approach for predicting emissions of heavy construction machinery in large-scale infrastructure projects.

Discontinuities in rock masses significantly influence their mechanical properties and are critical for engineering applications, making it essential to thoroughly understand their geometric parameters. 3D point clouds serve as fundamental data for efficiently and accurately analyzing discontinuity orientations. In this context, a novel semi-automated method that employs a Nutcracker Optimization Algorithm-improved Probabilistic Neural Network (NOA-PNN) is proposed. The NOA enables the PNN to quickly identify the optimal smoothing factor, balancing both accuracy and efficiency. This method utilizes not only normal vectors, but also point coordinates, curvature, and density, incorporating a broader set of features to accurately identify points on discontinuities. The NOA-PNN model, trained on manually selected samples, swiftly identifies discontinuity sets while efficiently filtering out noise. Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is then used to extract single discontinuities within each set. Each discontinuity is fitted to a plane using a Principal Component Analysis (PCA)-based least squares method, facilitating the measurement of their spatial geometric parameters. Validation through two cases demonstrated that the proposed method achieved an average deviation of less than 5° in both dip direction and dip angle, exhibiting potential advantages in terms of accuracy and efficiency when compared to other important studies or software. This method significantly improves computational efficiency and achieves satisfactory results with only a small number of randomly selected samples. Its low requirements for sample quality and operator expertise make it highly operable and easily adaptable for practical engineering applications.

Field transport productivity analysis is crucial for scheduling large-scale earth-rock works. Although camera surveillance facilitates the monitoring of transportation activities, disjoint views from sparse cameras result in discontinuous monitoring. To address this issue, a single-camera tracking with cascade R-CNN is used for target detection, and an improved TransReID for appearance feature extraction. Subsequently, these features were utilized by the WDA-Tracker algorithm to associate targets across discontinuous camera views. Enhancements in the improved TransReID include multi-scale information extraction, and the logic information building module mitigates the substantial scene variation between nonadjacent cameras, which affects the consistency of the target appearance. The dataset incorporating self-made truck images had a Rank-1 accuracy of 90.1 %, outperforming the original TransReID accuracy of 86.2 %, and experiments at a hydropower site with three cameras showed a multi-camera tracking accuracy of 98.98 %. This method accurately calculates transport productivity and provides information on construction progress and dispatch plans.

The operational safety of robotic rollers is of paramount importance, particularly in the challenging construction environment of dam construction sites. However, factors like low-illumination and intense vehicle vibrations can critically impair obstacle tracking and decision-making processes. To address this issue, this study proposes an improved BOTSORT multi-object tracking algorithm using feature-level fusion of millimeter-wave radar and camera sensors. Initially, by utilizing convolutional and PS-ROI align networks, radar and camera data are merged into feature maps, which are then processed by the improved BOTSORT algorithm using YOLOv8 instead of YOLOX for precise obstacle detection in low-illumination conditions. Additionally, an unscented Kalman filter module is employed to predict nonlinear motion of objects within the image during vibrations, while radar data refines the target association process, improving tracking accuracy under severe vibration conditions. A case study of a large-scale hydropower project demonstrates that the proposed method achieves 61.7 % mAP and 76.5 % MOTA, outperforming other obstacle detection and multi-object tracking algorithms. The proposed method improves the safety and reliability of robotic rollers under low-illumination and severe vibration working conditions.