Rock-filled concrete (RFC) is a dam construction technique that heavily relies on engineering machinery. Because of the conservative design concept, currently, many transverse joints exist in RFC gravity dams, resulting in a narrow working space and a small radius for mechanical rotation. The Dagutai RFC gravity dam innovatively adopts a design with fewer joints, and it has the longest section (134 m) among all RFC gravity dams. The reservoir has been in operation for more than 3 years after its impoundment. To determine the safety state of the Dagutai gravity dam during the impoundment operation period, this paper researched the dam's working behavior under different loads, using dam temperature, seepage pressure, and displacement monitoring results, along with the temperature stress simulation of the finite element method (FEM). The results revealed that the dam body's temperature and seepage pressure are within normal limits, and the dam's longest section has no obvious cracking or seepage, indicating that the dam performed well during storage and operation. RFC's adiabatic temperature rise is low, as evidenced by the temperature rise of the most unfavorable position in the long section being less than 10℃. The section length of the RFC gravity dam can be appropriately extended, but the large lifting surfaces near the foundation should not be constructed during hot seasons, to control the initial temperature and reduce temperature stress during construction. Compared to standard dam sections, the long section's stress state meets the requirements, and the maximum stress can reach 2.5 MPa under empty or full reservoir conditions. Reinforcing mesh and short joints should be installed in the upstream impermeable layer. Overwintering in the upstream surface above water and the downstream surface can cause high tensile stress. If a gravity RFC dam with no joints is built, the large tensile stress at both ends of the dam should be considered. Even when the overloading factor is 10.0, there was no yield failure through the upstream and downstream of the dam, indicating that the overloading safety of the dam is high and the dam stability is good. This paper's research can provide scientific guidance for the structure design of RFC gravity dams in the future.

Rock-filled concrete (RFC) is a heterogeneous material composed of large rocks and self-compacting concrete (SCC), and its temperature distribution before and after pouring differs significantly from that of conventional concretes. The temperature variation data of RFC in different seasons, at different locations, and before and after pouring were obtained by conducting on-site experiments for temperature monitoring, and the summary of laws and theoretical analysis was performed on this basis. Under various spatial and temporal conditions, the nonuniformity distribution of the rockfill temperature before RFC pouring was discussed. Quantitative calculations revealed that the influence depth of the air temperature on the equivalent rockfill can be as high as 0.9 m. In summer, the maximum temperature difference between the top and bottom of the rockfill can reach 13℃, and the phase lag of the equivalent rockfill temperature is approximately 3 h compared to the air temperature. The temperature variation regularity of the RFC after pouring in different seasons was summarized. According to the research results, there was a rapid temperature exchange between the rockfill and the SCC, especially in the first 8 h after pouring, until both temperatures were uniform. During the temperature rise process, the rockfill will absorb some of the SCC's hydration heat, thereby reducing the overall temperature rise of the RFC. The temperature difference caused by solar radiation at the upper and lower reaches of the lift surface is about 3℃. The results of this study have important implications for numerical simulations of RFC temperature and possible engineering temperature control measures.

Cemented riprap underwater technology uses good antidispersion performance and high-fluidity characteristics of underwater self-protecting concrete or mortar to fill the gaps between underwater blocks to form cement with a definite structural strength. This paper covers the first instance of cemented riprap underwater technology being applied to the field of scour protection for offshore wind power monopile foundations. Based on the cemented riprap underwater technology, this paper proposes anti-scouring measures for offshore wind power monopile rock-filled concrete, and uses a 1:13 large-scale wave current physical model test to study the anti-scouring performance of underwater structure around piles. Tests have shown that the plum-shaped pouring method can form a complete cemented riprap underwater protective structure. This protective structure can maintain good integrity and stability. Under extreme wave and current conditions with a return period of 50 and 100 years, the maximum scour depths of the sand around the monopile are 4.00 and 6.24 m, and the scour pit widths are 6.50 and 13.75 m, respectively. After the cemented riprap underwater protective structure, the maximum scour depths of the soil around the pile are 2.96 and 4.42 m, and the scour pit widths are 2.65 and 4.55 m, respectively. The above results provide new ideas for the protection measures and scour restoration of offshore wind power monopile foundations.

Key thermal parameters of self-compacting, rock-filled, and normal concrete of the rock-filled concrete gravity dam of Shibahe Reservoir in Renhuai City, Guizhou Province, are inverted based on its design data and temperature-monitoring data. On this basis, the finite element simulation method is used to simulate the working behavior during the construction of the dam, including the dam pouring, concrete hardening, and meteorological change processes. Next, the distribution and evolution law of the temperature field and stress field of the dam are analyzed to evaluate the overall safety of the dam. The results show that the temperature process of the dam body obtained by simulation inversion can truly reflect the temperature variation of the dam body. The maximum temperature in the dam body is typically reached within 3—5 d, and the temperature rise of hydration heat is between 4—7 ℃. The maximum temperature of pouring concrete in summer is about 35.0—39.5 ℃, while that in winter is about 25—28 ℃. Although the partial tensile stress on the spillway section surface is large in the low-temperature season, the stress in other parts does not exceed the concrete strength, and the dam safety can be guaranteed. The result indicates that pouring concrete on the large warehouse surface of the rock-filled concrete gravity dam is feasible.

Most traditional distributed deep learning training systems have been based on parameter servers which have centralized communication architectures that face serious communication bottlenecks due to the large amounts of communications and AllReduce communication frameworks which have decentralized communication architectures that cannot store the entire model due to the large number of parameters. This paper presents PS-Hybrid, a hybrid communication framework, for large deep learning recommendation model training which decouples the communication logic from the embedded parameters and other parameters. Tests show that this prototype system achieves better performance than previous parameter servers for recommendation model training. The system is 48% faster than TensorFlow-PS with 16 computing nodes.

Hilbert encoding and decoding are fundamental steps in many Hilbert curve based applications. However, existing algorithms are not very efficient when the data distribution is skewed. This paper shows that for a coordinate with the specific first m orders, the code of the first m orders is a multiple of its corresponding first order code. For a code with the specific first m orders, the coordinate of the first m orders is a multiple of its corresponding first order coordinate. These findings were used to develop an algorithm that skips the first m orders of the Hilbert encoding (SFO-HE) and another algorithm that skips the first m orders of the Hilbert decoding (SFO-HD). These algorithms exploit efficient bit operations and fast bit set detections to improve the encoding and decoding efficiencies for data skewed to the 4 corners of the Hilbert space. Extensive tests show that these two algorithms have good skewness adaptability and outperform existing algorithms on specific skewed data.

Big data processing is being widely used in academia and industry to handle DNN-based inference workloads for fields such as video analyses. In such cases, multiple parallel inference tasks in the big data processing system repeatedly load the same, read-only DNN model so the system does not fully utilize the GPU resources which creates a bottleneck that limits the inference performance. This paper presents a model sharing technique for single GPU cards that enables sharing of the same model among various DNN inference tasks. An allocator is used to make the model sharing technique work for each GPU in the distributed environment. This method was implemented in Spark on a GPU platform in a distributed data processing system that supports large-scale inference workloads. Tests show that for video analyses on the YOLO-v3 model, the model sharing reduces the GPU memory overhead and improves system throughput by up to 136% compared to a system without the model sharing technique.

The key issue for cross-modal retrieval using cross-modal Hashing is how to maximize the consistency of the semantic relationship for heterogeneous media data. This paper presents a self-supervised deep semantics-preserving hashing network (UDSPH) that generates compact Hash codes using an end-to-end architecture. Two modality-specific hashing networks are first trained for generating the Hash codes and high-level features. The semantic relationship between different modalities is then measured using cross-modal attention mechanisms that maximize preservation of the local semantic correlation. Multi-label semantic information in the training data is used to simultaneously guide the training of two modality-specific Hashing networks by self-supervised adversarial learning. This constructs a deep semantic hashing network that preserves the semantic association in the global view and improves the discriminative capability of the generated Hash codes. Tests on three widely-used benchmark datasets verify the effectiveness of this method.

Ethylene is an essential petrochemical industry product produced in a complex steam cracking process. Fast, accurate predictions of ethylene cracking depths depend on accurate naphtha cracking models. This paper compares three machine learning models based on a support vector regression (SVR), a k-nearest neighbor regression, and an extreme gradient boosting (XGBoost) to predict the ethylene cracking depth. Several industrial datasets are screened to identify the critical variables controlling the process using the density-based spatial clustering of applications with noise (DBSCAN) and a local abnormal factor detection algorithm. These three models are then trained and combined into an ensemble model to provide better predictions. The ensemble model combines the advantages of the three models and reduces the overfitting, the sensitivity to noise and other shortcomings. The ensemble model then has better prediction stability and generalization ability. The ensemble model predictions have R²=0.955 and an average absolute percentage error of about 0.23%, which is sufficient for process research and industrial applications.

A smart system predicting the remaining useful life (RUL) of fan belts can ensure the safety of chemical engineering processes and reduce human labor. In this study, the features extracted from the vibration signals of the belt are fed into a stacked autoencoder (SAE) to obtain the health indicator (HI). Then, a polynomial is fitted using the HI to calculate the RUL of the belt. The destructive experiment is conducted to shorten the data collection period, and a loss function is constructed to fit the destructive experiment. This study also improved the structure of the SAE, i.e., the inner product and the activation operation of the middle layer are separately done in two neurons. This bounds the HI in [0,1] and preserves more information from the encoding layers. Compared with traditional SAEs, the proposed method reduces the prediction errors in real cases. The errors of the predicted RUL are bounded by 1 d in the last 50% of the sections of all three datasets. The trained model also performs well when employed on an online test.

As one of the most promising carbon reduction methods, solvent-based post-combustion carbon capture is expected to provide clean use of fossil fuels. Low-carbon coal-fired power plants with carbon capture scheduling can significantly advance the goal of "carbon neutrality". However, few studies have considered the impact of electricity price fluctuations and electricity consumption changes on power plant coupled with carbon capture device. This study integrated the characteristics of power plants and their carbon capture into an information gap decision theory (IGDT) model to analyze the uncertainties in the combined system. A deterministic scheduling model for the integrated power plant and carbon capture device was used with IGDT to describe the load demand uncertainty in the real-time market. Two certainty scheduling models were developed based on the users risk adverse attitude with the optimized scheduling decision based on either robustness or opportunity. The scheduling schemes of integrated power plant and carbon capture device with these two risk attitudes were used for a case study to demonstrate the reliability and effectiveness of the model.

In ball-end milling of thin-plate parts, the chatter stability is analyzed using a three-dimensional multiple degrees of freedom milling dynamic model. The influence of the circular arc area left by the previous cutting on the tool– workpiece contact area is considered. The cutting force coefficients are identified by the average cutting force method. The modal parameters of the tool and workpiece are obtained via the hammer experiment. The stability lobe diagram (SLD) at each measurement point of the workpiece is obtained using the full-discretization method. Both theoretical and experimental results show that the variations in the dynamic characteristics at each point of the workpiece will change the SLD. Moreover, several closed unstable "island" regions are detected in the SLD. The unstable "island" region is attributed to the modal stiffness, modal frequency, and modal damping ratio of the workpiece.

The performance of an internal combustion engine is related to the lubrication state of the crankshaft. In this paper, a thickness measurement system is developed using an ultrasonic reflection signal to monitor the change of oil film thickness in real time. Based on acoustic theory, the relationship between oil film thickness and acoustic reflection coefficient in multilayer media is derived, and the application methods of the spring, resonance, and phase shift models are given. For the crankshaft, the oil film thickness ranges from 0 to 50 μm, an ultrasonic double crystal straight probe with a center frequency of 1.056 MHz is used to generate and receive the ultrasonic signal, and a long axis with four annular grooves with different depths is processed to obtain the functional relationship between the oil film thickness and the spectral peak of the acoustic reflection signal. To calculate the oil film thickness, the fitted linear model is embedded into LabVIEW software. The test shows that the relative error of the measurement system is less than 10%, and the oil film thickness can be measured well.

A monocular method is proposed to estimate the parameters of uncooperative targets whose inertia is equal crosswise, to solve the problem that uncooperative targets without communication, such as failed satellites, cannot provide their rotational motion parameters or dynamic parameters. First, a dynamic parameter model is given, and then an analytic expression for target parameters is created. On this basis, taking the monocular image sequence as an input, the noisy result of simultaneous localization and mapping (SLAM) techniques visual odometry is filtered multiple times, and rotational speed vectors are fitted to a double-cone model. Eventually, targets' rotational kinetic and dynamic global optimal parameters can be determined. The simulation results demonstrate that the method can estimate parameters efficiently and accurately, and the relative errors are less than 0.9%.

The LuGre model is an advanced friction model that describes dynamic friction characteristics. However, the imported unobservable bristle deformation complicates accurate, efficient identification of the model parameters. The traditional LuGre parameter identification method requires the motion system to use torque control and requires a significant computational load. The traditional LuGre parameter identification method is not applicable to some systems, such as multiple degree-of-freedom manipulators, that use the position control mode. Therefore, this paper presents a modified LuGre parameter identification method based on an area-specific analysis of the friction torque-velocity curve. Shape factors are defined to quantify the shape features of the Stribeck peak and the hysteresis loop which are then used for the LuGre parameter identification. The identification result is better than the local optimal solution of the PSO method. Simulations and hardware identification tests on a manipulator verify the effectiveness of this method, which requires fewer experiments, has better parameter identification accuracy and more accurately predicts the manipulator joint torque than the pure PSO method.

An auto-calibration method was developed for planar four-cable parallel robots based on visual measurements. The end-effector pose is obtained using a camera fixed on the end-effector that measures the pose of AprilTags on storage racks. Geometric parameter identification and error compensation are then used for rapid calibration of the planar four-cable parallel robot. The system model is based on a kinematics model of the planar four-cable parallel robot for storage and logistics tasks. Then, the kinematics calibration method is used with the pose measuring method to calibrate the robot. The calibration accuracy was verified using both computer simulations and tests. The results show that this rapid calibration method provides positioning accuracy of less than 1 mm, which meets the accuracy requirement for the storage and stacking tasks.

Visual sensing is a key technology for online detection of welding groove size parameters, and welding torch positions and postures in intelligent welding systems. Accurate visual sensing requires accurate calibration of the internal visual sensor parameters. This paper describes an integrated calibration method for the internal parameters of a visual sensor based on combined laser structured lights. The method is based on an ordinary checkerboard calibration board with the calibration system extracting the centerline of the laser line in the image using the skeleton thinning method and Hough line detection. The system then determines the three-dimensional coordinates in the camera coordinate system of the points on the laser center line of the calibration board to fit the parameters in the laser structured light plane equation. This integrated calibration method improves the calibration accuracy, efficiency and convenience. Tests measuring the welding groove size on a flat workpiece gave mean and repeated measurement errors of the welding groove size of not more than 0.04 mm, which verifies that the sensor calibration accuracy meets the needs for welding groove size measurements.

Sensor accuracy in special environments can be very limited due to closed systems and magnetic interference. For example, sensors on wall climbing robots can experience accumulation of autonomous positioning errors with time. The paper presents an autonomous positioning method for wall climbing robots based on an external RGB-D camera and a robot-mounted inertial measurement unit (IMU). This method uses the target tracking method with a deep learning and kernelized correlation filter (KCF) for preliminary positioning. A normal direction projection method is then used to locate the center on the top of the robot shell for the robot position positioning. The system determines the normal, the roll angle and the heading of the robot with a series EKF filter calculating the roll angle, pitch angle and heading to estimate the robot attitude. Tests show that the wall climbing robot positioning error is within 0.02 m, the heading error and the roll angle error for the attitude estimate are both within 2.5°, and the pitch angle error is within 1.5°. This system effectively improves the wall climbing robot positioning accuracy.

A sealing leakage test system for gas-liquid mixtures was developed to measure oil leakage from shaft seals on large marine diesel engine crankcases. The air and oil leakage rates were then measured for various rotational speeds and the leakage reducing function of the oil slinger, labyrinth grooves and oil return channels were studied quantitatively. A brush seal design was then developed that was appropriate for the large shaft diameters and large clearances in crankcase shaft seals. A porous media model of the brush seal was then used to analyze the influence of the brush seal on the leakage and the flow field. The effect of the brush seal on the leakage was also studied on the sealing leakage test system. The results show that the brush seal reduces the air leakage rate by 76% and the oil leakage rate by 79%.

Some key coated joint bolts on helicopter rotor systems had poor fatigue lives. This study investigated the microhardness distribution in these bolts using the scanning electron microscope (SEM) and electron back-scattered diffraction (EBSD) tests. The tests showed that the deteriorated layers induced by turning, machining and shot blasting overlapped at the bolt substrate surface, which negatively impacted the bonding strength between the WC-10Co4Cr coating and the substrate. The poor fatigue life was found to be mainly due to premature cracking of the substrate-coating interface. The grinding depth was then increased to eliminate the overlapping deteriorated layers to more uniformly distribute the microhardness between the coating and the bolt substrate which lengthened the fatigue life 3 folds relative to the original bolt. This result is mainly due to the removal of the overlapping deteriorated layers which improves the hardness as well as the bolt surface homogeneity and relieves the hardness gradient between the substrate and the coating, which significantly improves the adhesion between the coating and the substrate.

The stability of a completely restrained 3-DOF cable-driven parallel robot with four long-span cables, due to the flexibility and unidirectional restraint characteristics of the cables, as well as the influence of the large-span cable sags, faces severe challenges. This paper establishes the stability evaluation model and stability sensitivity analysis model for the robot, explores and analyzes the influence of the end-effector positions and the cable tensions on the stability of the robot. Firstly, based on kinematics and dynamics model of the robot, the stability position influence factor and the cable tension influence factor are proposed, and furthermore, the stability evaluation model is established. Secondly, the gray correlation analysis method is used to establish the stability sensitivity analysis model for the robot, and it is proposed to use the correlation degree to study and measure the influence degree of the end-effector positions and cable tensions on the stability of the robot. Finally, the established stability evaluation model and sensitivity analysis model are simulated for a cable-driven camera robot with four long-span cables. The research results show that the stability of the camera robot is more sensitive to cable tension influence factors. Among them, the stability has the smallest correlation to the y-direction displacement of the camera platform. This research provides guidance for the robust optimization design of the motion trajectory and motion control for the robot.