To connect the photovoltaic (PV) power generation to the traction power supply system for effective utilization and take into account the improvement of the power quality of the traction power supply system and the recycling of the regenerative braking energy of the train, a coordinated control method based on railway power conditioner (RPC) was proposed to connect the PV and battery system to the traction power supply system. Firstly, the PV and battery system based on RPC was connected to the traction power supply topology, and its power quality compensation mechanism was analyzed theoretically. On this basis, a combined coordinated control method for the PV and battery system connected to the traction power supply was proposed by taking full account of the operating conditions of the PV and battery system and the working conditions of traction, braking, no load, or idling. As a result, the coordinated control of “PV + battery + load”, the dynamic compensation of negative sequence, and the effective suppression of harmonics in the traction power supply system were realized. To verify the validity of the proposed system structure and the coordinated control method, the simulation model of the PV and battery system connected to the traction power supply system based on RPC was established in the RT-Lab real-time simulation system, in combination with the actual illumination and the data of traction conditions. The results show that the proposed system structure and coordinated control method can realize the effective access of the PV and battery system in the traction power supply system. The regenerative braking energy recovery rate is increased to 86.7%, and the power factor is increased by 6.4%. There is a big improvement in the current ratio of harmonics, especially high harmonics. Therefore, the proposed method can meet the demands of efficient PV power dissipation and comprehensive utilization, dynamic and comprehensive compensation of power quality, and regenerative braking energy recovery.

The deformation of the tunnel surrounding rock and the thickness of the loose zone are important considerations for tunnel structure design and an important reference basis for safe operation. However, the existing identification methods of the loose zone and measurement methods of tunnel surrounding rock deformation are mostly at the detection level, and long-term and real-time monitoring methods are inadequate. In this paper, a self-sensing fiber reinforced plastic (FRP) anchor embedded with optical fiber was developed, and an intelligent monitoring system for the tunnel surrounding rock based on the self-sensing FRP anchor embedded with optical fiber was proposed. The monitoring system can realize three-dimensional, round-the-clock, and real-time monitoring of surrounding rock deformation. Based on the monitoring data and theoretical analysis, a method to identify the loose zone of the tunnel surrounding rock was proposed. The intelligent monitoring method was applied to the Chentang tunnel in Guangzhou−shantou high-speed railway for the first time. The results show that the self-sensing FRP anchor can accurately detect the influence of field construction on surrounding rock deformation, and the steel frame has a prominent supporting effect on the tunnel surrounding rock. The monitoring data can reflect the force law of the self-sensing anchor in real time, so as to accurately determine the thickness of the loose zone of surrounding rock at different positions of the tunnel. The intelligent tunnel surrounding rock monitoring system based on the self-sensing FRP anchor will continuously and timely monitor the deformation of tunnel surrounding rock during the tunnel operation and provide a high-tech guarantee for the safety of the tunnel structure during the full life cycle.

The tunnel pressure wave excitation generated when the same train passes through the same tunnel repeatedly has the characteristics of similar forms, variable scales, and variable amplitudes. Since the current control strategies do not take into account this fixed form, an iterative learning control method based on higher-order feedback forgetting was proposed, so as to control the interior pressure fluctuation under the disturbance of the pressure wave of the tunnel with a fixed form. First, a mathematical model of air pressure transmission inside and outside high-speed trains was established, and the measured pressure data inside and outside the train were used for correction and verification. Secondly, pressure changes in the train were mitigated by controlling the valve of the train’s ventilation. The iterative learning control algorithm based on higher-order feedback forgetting was proposed, and the variable amplitude and variable scale processing methods were designed. Finally, a set of stochastic pressure waves with a fixed form was generated by using the measured pressure waves, and simulation analysis was carried out. The simulation results show that the iterative learning control algorithm based on higher-order feedback forgetting can make the pressure inside the train converge to below 200 Pa/s within 1 s after the 8th iteration period, and the RMSE value is reduced to below 15.0000% after the 4th iteration period under the excitation of the tunnel pressure wave with repetitive fixed form.

In order to reveal the effect of ventilation parameters and ground temperature on freezing damage in tunnels in cold regions, the three-dimensional unsteady numerical heat transfer control equations for the tunnel surrounding rock, lining, and airflow are developed based on heat transfer theory. The numerical heat transfer difference equations of different nodes are analyzed, and a three-dimensional temperature field numerical calculation model is established for tunnels in cold regions with high ground temperatures. The effect of mechanical ventilation velocity, mechanical ventilation time, and ground temperature on the anti-freezing length of tunnels in cold regions with high ground temperatures is studied based on numerical analysis. The results show that 1) without considering mechanical ventilation, the anti-freezing length of Zilashan tunnel will exceed 1 200 m and decrease by about 100 m when the ground temperature increases by every 5 ℃. 2) By considering the effect of mechanical ventilation, when the ground temperature is 10–30 ℃, ventilation at a speed of 2.5 m/s for 2.0 h per day can reduce the anti-freezing length by 215 m; the anti-freezing length decreases by about 20 m as mechanical ventilation velocity increases every 0.5 m/s under different ground temperatures, and the effect of ground temperature on the decay rate of anti-freezing length is slight; when the mechanical ventilation time is less than 2.0 h, increasing mechanical ventilation velocity under different ground temperature conditions has little effect on anti-freezing length.

To explore the aerodynamic load characteristics of trains under side wind environments, the model of the longitudinal and lateral components of wind velocity at a point moving with the train for any wind angle was set up. The computational method of aerodynamic loads of the train exposed to wind velocity with longitudinal and lateral components was studied. The aerodynamic load characteristics of the train at a speed of 200–400 km/h, mean wind speed of 10–35 m/s, and wind angle of 30°–150° were analyzed. The results show that under different wind angles, when the lateral component of wind velocity is considered, the aerodynamic load of high-speed trains fluctuates greatly, leading to an increase in the extreme value of instantaneous aerodynamic load acting on the train. The lateral component of wind velocity mainly affects the standard deviation of aerodynamic loads of the train, and the influence has much relationship with wind angle. The influence becomes smaller when the wind angle is closer to the critical wind angle and becomes larger when the wind angle is far away from the critical wind angle. Under different wind angles, the ratio of standard deviation to mean value of aerodynamic load of the train mainly depends on the yaw angle. The ratio is relatively large at wind angles of 30° and 150°, followed by wind angles of 60° and 120°, but it is relatively small at wind angles of 90°.

In the multi-objective aerodynamic optimization design of high-speed trains, the optimization efficiency of the surrogate model established using the traditional infill criterion is low when the initial sample points are few. To this end, a hybrid infill criterion (HIC) was proposed by combining the improved expectation infill criterion (EIC) and the Pareto solution infill criterion (PIC). Meanwhile, a Kriging surrogate model was established using the HIC method, and multi-objective aerodynamic optimization on the head shape of the high-speed train was conducted, with the minimum aerodynamic drag force of the leading car, the minimum aerodynamic drag and lift force of the rear car as the objectives. The single-objective Branin test function and the multi-objective Poloni test function were taken as examples, and the convergence speed of EIC, PIC, and HIC surrogate models was compared. The results show that the optimization efficiency of the HIC surrogate model is improved by 50.0% compared with the EIC and PIC surrogate models in the single-objective optimization. For the multi-objective test function, the efficiency of the HIC surrogate model is improved by 62.5% compared with the PIC surrogate model. Moreover, the HIC surrogate model is used to carry out the multi-objective aerodynamic optimization of the head shape of the high-speed train, and the optimal solution model obtained reduces the above three objectives respectively by 1.6%, 1.7%, and 3.0% compared with the original model. The heights of the nose, the coupler area, and the cab window of the optimal solution are all reduced. Meanwhile, the two lateral contour lines are retracted.

With the increase in the train speed, the contribution of the quadrupole noise source to the aerodynamic noise of the train increases. When the running speed of the high-speed maglev train reaches 600 km/h, it is necessary to consider the influence of the quadrupole noise source on the aerodynamic noise of the high-speed maglev train. The numerical simulation method of the aerodynamic noise of the high-speed maglev trains considering the quadrupole noise source was set up. The local extrapolation of the integral surfaces for the streamlined tail and head regions of the high-speed maglev train was carried out, and the influence of the quadrupole noise sources of the streamlined tail and head regions on the aerodynamic noise of the high-speed maglev train was explored. The study shows that the wake vortex of the high-speed maglev train will pass the downstream integral surface. The fully enclosed integral surface can not be used for the streamlined tail region, or otherwise, it will produce large spurious noise. The integral surface for the streamlined tail region needs to extend more towards the wake vortex region and remove the region through which the wake vortex passes. The contribution of the quadrupole noise source of the streamlined head region of the high-speed maglev train is small, and the integral surface for the streamlined head region can be taken as the streamlined head surface. When the high-speed maglev train runs at a speed of 600 km/h, the aerodynamic noise energy of the high-speed maglev train caused by the quadrupole noise source accounts for 42%.

In order to solve the local overheating and the energy waste of radiant floor heating systems in railway stations on the plateau with intensive solar radiation, a railway station on the plateau was studied, and a partition heating scheme of the radiant floor system was proposed according to the solar radiation distribution in the station. The building energy consumption simulation software, namely Energyplus was used to simulate the station, and both the indoor thermal environment and the heating capacity of the radiant floor system with or without the partition scheme were compared. Results show that when the radiant floor system adopts a unified design and operation control scheme for heating, the average operative temperature in the irradiated zone can reach up to 27 ℃ during the daytime, and the temperature difference with the unirradiated zone is more than 5 ℃. When the radiant floor system uses the partition heating scheme, the maximum operative temperature in the irradiated zone is 26 ℃. The indoor temperature distribution is more uniform, and the local overheating is alleviated. Under the premise of ensuring the thermal comfort of the station, the partition heating scheme of the radiant floor system can reduce the heating capacity in the irradiated zone during the heating season by 38.2%.

In order to accurately recognize the readings of digital instruments in the actual scene of substations, intelligently control substation security, and promote its intelligent development, the digital instruments in the substation are taken as the research object, and in view of real-time and accuracy, a lightweight YOLO-v4 model is proposed for the detection and recognition of digital instruments. Firstly, the digital instrument images captured from the Ordos substation are expanded by using the Albumentations framework, thus building an effective digital instrument data set for detection and recognition. After that, an efficient channel attention (ECA)-based deep separable convolution block (ECA-bneck-m) is constructed with attention mechanism, and further a lightweight YOLO-v4 model is proposed to conduct comparative experiments on model size and performance. Finally, experiments comparing model size and performance are performed. The results show that, the storage size of the model can be compressed by about 5 times nearly without loss of detection accuracy, and the processing speed of model can be increased from 24.0 frame/s to 36.9 frame/s, indicating that the proposed model can meet the requirements of real-time detection and recognition in the actual substation.

In order to investigate whether longitudinal ventilation can effectively control the diffusion of leaked hydrogen in tunnels and reduce the overpressure hazards of flammable hydrogen clouds, a square tunnel with a length of 100 m was studied. The hydrogen leakage and diffusion process in the tunnel were numerically simulated by using FLUENT software. Based on the classification of damage of combustible hydrogen cloud explosion overpressure to people and buildings, the effect of longitudinal ventilation on reducing the harm of hydrogen leakage and explosion in the tunnel was analyzed. The results show that under calculated conditions, the longitudinal ventilation velocity in the case of a hydrogen jet fire in the tunnel is 6.5–8.0 m/s. Longitudinal ventilation can effectively control the leakage and diffusion of hydrogen in the tunnel, but it fails to eliminate the explosion possibility and harm of hydrogen leakage in the tunnel.

In order to study the effect of the temperature on the emission of volatile organic compounds (VOCs) from typical non-metallic materials (heavy anti-corrosion coatings, floor covering, and glassfiber reinforced plastics) for railway passenger trains, tests were carried out to analyze the emission law of VOCs based on multi-gas-solid ratio method and data fitting method. First, the concentrations of VOCs under four different gas-solid ratios were measured, and key parameters (initial emittable concentration and diffusion coefficient) influencing the emission of VOCs were obtained. Then, with the help of vehicle manufacturing technology and application scenarios, the effect of different temperatures on the emission law of VOCs was studied. The results show that the physical and chemical properties of the materials and the temperature greatly affect the emission of VOCs. When the temperature rises from 16 to 55 ℃, the concentrations of benzene compounds and aldehydes decrease; the initial emittable concentration of heavy anti-corrosion coatings is decreased to 1.8%, and its diffusion coefficient is decreased. At the same time, the initial emittable concentrations of both floor covering and glassfiber reinforced plastics are decreased to less than 0.3%; the diffusion coefficient of floor covering is increased, while that of glassfiber reinforced plastics is decreased. Benzene compounds are the main components of VOCs, in which the styrene accounts for the largest proportion, and the ranking rule of toluene, ethylbenzene, and xylene is not obvious. Meanwhile, the benzene is not detected in the test. In the baking process of railway trains for environmental protection, it is suggested that the baking temperature of heavy anti-corrosion coatings should not be less than 55 ℃, and that of floor covering and glassfiber reinforced plastics should not be less than 45 ℃.

High-speed trains carry a large number of passengers and have a well-developed transportation network. However, the closed compartment environment is likely to cause the accumulation of pollutants. In order to improve the comfort and safety of the train, a full-scale compartment ventilation model of the train under full load conditions was established based on computational fluid dynamics (CFD). For the exhaust mode where the exhaust vent was located above the window, the coefficients of velocity non-uniformity, temperature non-uniformity, and energy utilization, as well as ventilation efficiency were used as the evaluation indexes of the train’s ventilation system. The effects of six air supply modes on the flow field characteristics and diffusion characteristics of respiratory pollutants in the compartment were comparatively studied, including perforated ceiling air supply, lower air supply, perforated ceiling air supply combined with lower air supply, local perforated ceiling air supply, side roof air supply, and local perforated ceiling air supply combined with side roof air supply. The results show that by adjusting the flow distribution ratio between the air vents, the air supply can flow evenly to both sides of the passenger compartment, thereby improving the temperature uniformity in the train. When the lower air supply is used, it helps to improve the energy utilization coefficient and ventilation efficiency of the ventilation system, which are as high as 1.38 and 1.21, respectively. However, it will deteriorate the riding comfort of the train. By studying the interaction of respiratory pollutants among passengers, it is found that the respiratory pollutants of passengers in column C tend to diffuse to the breathing area of passengers in column B, thereby aggravating cross-infection among passengers. Reducing the size of the air supply outlet of the perforated ceiling and using local perforated ceiling air supply mode can effectively alleviate this phenomenon and reduce the volumetric concentration of pollutants to 0.0019.

Double-chimney structure has aerodynamic interference effects under natural wind, which will induce large wind-induced vibrations, further threatening the safety of the structure. Reasonable calculation and prediction of wind-induced vibration responses are essential for wind resistance design of double-chimney structures. A double-chimney structure with a center distance equal to eight times its average diameter was experimentally tested. The wind tunnel tests of force measurement with a rigid model and vibration measurement with an elastic model were carried out. The test results were compared with the calculated values of the Chinese code, European code, and International Committee on Industrial Construction (CICIND) code to explore the wind-induced response characteristics of the double chimneys under different wind angles. The results demonstrate that in tandem arrangement, the chimney on the windward side shows the shielding and interference effects. Therefore, the bending moment at the bottom of the chimney on the leeward side is reduced, and the across-wind displacement is greater than that at other wind angles. The calculated values of wind-induced vibration coefficients are close to the test values due to the interference effects of the power station. When the height of the chimney exceeds the station, the calculated values are greater than the test values. In terms of the across-wind responses, the values calculated by the Chinese code are 37.1% larger than the test values. The values calculated by the European code are close to and only 6.9% smaller than the test values, but the values calculated by the CICIND code are 17.1% smaller than the test values.

To solve the design waste and idle operation of ventilation systems in urban tunnels caused by the continuous update of traffic vehicles, the influence of base emission factor update on tunnel fresh-air demand was studied, and two calculation methods of base emission factor in response to vehicle update were established. To begin with, the key parameters updating over time in the fresh-air demand formula were determined as base emission factors through theoretical analysis. Then, the influence of base emission factors and gradient-speed factor update on fresh-air demand was obtained based on quantitative analysis. Finally, according to foreign concepts and actual design experience, two calculation methods for base emission factors considering time were proposed. The results show that Chinese road tunnel ventilation design standards refer to the PIARC ventilation report when formulating the base emission factors, with the reasons explained. Compared with those in 2000, the base CO, NO_x, and PM emission limits in 2021 are reduced by 81.6%, 76.7%, and 97.9% respectively. The most unfavorable design method using vehicle pollutant emission limits as base emission factors is established, and the base emission factors of each pollutant in 2018 are calculated as follows: the base CO emission factor is 0.0011 m³/(veh·km), and the base PM emission factor is 0.4610 m²/(veh·km), which are reduced by 84.3% and 77.0% compared with China Standard in 2014. The research results establish and verify the calculation method of base emission factors, which provides a reference for the design of urban road tunnel ventilation systems.

The prediction of machining quality is a vital component of intelligent manufacturing and a prerequisite for achieving quality closed-loop control, playing an extremely important role in promoting the practical application of intelligent manufacturing systems. A brief review of the history of machining quality prediction reveals that scholars have mostly focused on the mechanism of the influence of a key component of the machine tool on machining quality, while research on the correlation between the coupling effects of machine components is rare. Based on the aforementioned challenges, firstly, seven types of factors that affect machining quality are analyzed, including tool geometry parameters, cutting parameters, cutting fluid type, thermal errors and deformations, degradation of CNC machine tool components, cutting chatter, and system characteristics. Subsequently, according to the different types of data and measurement methods for each factor, the monitoring and prediction methods of machining quality are divided into four categories, including machine vision measurement, power measurement, vibration measurement, and other measurement methods. The technical characteristics, limitations, and development trends of each method are then expounded. Finally, considering the deficiencies of various machining quality monitoring and prediction methods, this paper points out that research on material cutting mechanisms, data quality assessment methods, standards for constructing industry site databases, and intelligent representation and visualization of quality prediction information may be future development trends.

Since the weak fault feature of train axle-box bearings is difficult to be extracted in a wide frequency band, this paper proposes a weighted combined improved envelope spectrum (WCIES) for fault diagnosis based on the second-order cyclostationary of bearing fault signals. First, the fine demodulation of the vibration signal in the full frequency band is achieved by decomposing the vibration signal into the dual-frequency domain composed of spectral and cyclic frequencies through the spectral coherence algorithm. The candidate fault frequency of the bearing is identified based on the local feature of spectral coherence. Then, the 1/3-binary tree filter is applied to divide the spectral frequency into a series of narrowbands with different center frequencies and bandwidths, and the mode of spectral coherence is integrated along the spectral frequency in the narrow band to obtain the narrowband IES. Then, the CIES of each decomposition layer is constructed by taking the ratio of the energy of the candidate fault frequency in the narrowband IES as the diagnostic index. Finally, the weighted average of the WCIES of different decomposition layers is performed, and the WCIES of the bearing vibration signal is obtained. The research results show that the advantage of the proposed method is that it can fully integrate the bearing fault information distributed in different narrowbands and does not depend on the nominal fault period information. Compared with the existing methods, it can more effectively reveal the characteristic frequency and harmonic characteristics of bearing faults and has advantages in extracting and identifying weak faults of axle-box bearings.

A multi-fault feature extraction and matching method guided by variational mode decomposition (VMD) was proposed to address the difficulty in identifying and diagnosing composite faults in train wheelset bearing systems. Firstly, in order to avoid the pre-defined mode number relying on prior knowledge during operation and thus affecting the diagnosis results, the original axle-box vibration data are directly decomposed by VMD step by step, and the number of modes is 2–N. Secondly, the VMD intrinsic mode functions (VIMF) obtained by VMD are calculated to extract the VIMF with the largest correlation kurtosis; then, the determined VIMF is analyzed by square envelope analysis to extract the fault feature frequency. Finally, the proposed method is compared with the fast spectral Kurtogram method and the correlation Kurtogram method. The analysis of simulation signals and experimental data shows that the proposed method can completely avoids the problem of selecting the key parameter K in the VMD model, and can accurately and effectively extract the fault characteristics of wheelsets and bearings, respectively. Compared with the fast spectral Kurtogram method and the correlation Kurtogram method the proposed method can diagnose compound faults effectively, and the obtained fault feature harmonic components are more advantageous in quantity and signal-to-noise ratio.

In-situ tool-condition monitoring system based on machine vision realizes tool wear measurement and condition assessment without removing the tool. However, the system deployment parameters that are closely related to the quality of the tool image are rarely studied. To this end, a polynomial regression model based on improved information entropy is constructed to realize the optimal deployment of the tool-condition monitoring system. First, the adaptive threshold method is used to remove the interference of background elements in the captured tool image, and the imaging quality of the tool wear area is evaluated by the information entropy metric. Then, a polynomial regression model with respect to the camera working distance, exposure time, and the proposed evaluation metric is constructed to describe the mapping relationship between the deployment parameters and the proposed evaluation metric. Finally, the least squares method is used to solve the coefficients of the polynomial model and obtain the optimal deployment parameters. Orthogonal experiments are designed to ensure that the factor levels of independent variables cover the optimal deployment parameters. The experimental results show that there is a main effect relationship between the proposed evaluation metric and deployment parameters, such as working distance and exposure time, which is in line with the changing rule of optical imaging systems. Compared with nonlinear regression prediction models such as support vector machine, decision tree and K-nearest neighbor (KNN), the cubic polynomial regression model has the smallest prediction error, with its mean absolute error, mean square error, and root mean square error being 0.022631, 0.00068, and 0.026069, respectively. The measurement accuracy of the tool image captured under the optimal deployment parameters reaches 96.76%, increased by 0.74%, demonstrating that it meets the accuracy requirements of tool condition monitoring.

In order to reveal the reasons for the great difference of cracks in the lifting lugs at different positions of the under-chassis equipment of CRH380AL high-speed EMU, full-scale test on vibration acceleration and aerodynamic load are carried out. The dynamics model of the large-scale system with multiple coupling of under-chassis equipment, vehicle, wheel-rail, and railway line is established. The vehicle body and under-chassis equipment are established as elastomer models by finite element method. The wheel-rail subsystem and bogie subsystem are modeled by rigid multibody dynamics. The track irregularity spectrum is based on the measured data samples from Wuhan to Guangzhou. The aerodynamic load under the conditions of tunnel passing and tunnel intersection is numerically simulated by the eight-car aerodynamic model. The influence of elasticity, aerodynamic load, bolt stiffness and other factors of the vehicle body on the lifting lug reaction force of the under-chassis equipment is analyzed. The research shows that, there is a strong coupling behavior between the under-chassis equipment and the vehicle system. The mass distribution of the under-chassis equipment and the elastic coupling effect of the vehicle body lead to the maximum vertical dynamic load of No. 4 lifting lug, which corresponds to the highest proportion of on-site fault cracks. The aerodynamic load has a significant impact on the dynamic load of No. 8 lifting lug of the under-chassis equipment. The dynamic load of the lifting lug in the low-frequency domain increases with the increase of bolt stiffness, the vertical average dynamic load and the maximum dynamic load are 4 times and 6 times of the other two directions respectively. The dynamic analysis method based on the coupling of line and vehicle can provide theoretical support for the design of dynamic mechanical behavior of equipment under the vehicle and the optimization of fatigue performance.

In order to obtain the pure wear acoustic emission of diamond-coated mechanical seal, the denoising method based on empirical wavelet transform (EWT) and Kullback-Leibler divergence (KLD) was proposed. Firstly, filter bank was calculated with empirical wavelet transform on acquired acoustic emission signal. Then the filter bank was applied to both the acquired acoustic emission signal and background noise acoustic emission signal. The Kullback-Leibler divergences were calculated between the corresponding bands of two signals. The cumulative sum algorithm was employed to find a threshold for determining whether the corresponding band is used for signal reconstruction. The results show that the proposed method can effectively suppress the noise of acoustic emission signals under different working conditions and wear states, and effectively improve the signal-to-noise ratio of wear acoustic emission signals, especially weak wear signals. Compared with the traditional denoising methods, the proposed EWT-KLD method has stronger adaptability and stability for denoising of wear acoustic emission signal under different working conditions, which is of great significance for the monitoring early seal wear and the cumulative wear process of seal.

To solve the problem of limited generalization ability of fault diagnosis model caused by the lack of sufficient fault data characteristics of single railway high-speed train bogie, and to realize the diagnosis of bogie faults of multiple railway high-speed trains, a global bogie fault diagnosis method based on federated learning is proposed in this work. Firstly, according to the bogie vibration signals of each railway, the multi-scale convolution fusion algorithm is conducted locally to extract and fuse the fault features at different scales, and the bogie fault diagnosis model is established locally. On the premise of not divulging data privacy, the fault diagnosis models of all railways are aggregated by the third party, the weights of model parameter are adjusted, the fault diagnosis models are optimized, and finally the global fault diagnosis model of bogie is jointly trained by multiple railways. The experiments show that under the federated learning framework, the fault diagnosis accuracy of the global bogie fault diagnosis model is reach more than 93% for the railway participating in federated modeling, and more than 75% for the railway not participating in, which provides a practical scheme for the ‘data island’ problem in railway transportation.

Traditional machine learning methods (e.g., hand-coded feature extraction) are sensitive to light sources, equipment installation errors and other factors, which require repeated debugging and experiments and make it difficult to achieve automatic detection in large-scale production. Considering the above-mentioned problems, an in-situ roughness evaluation method is proposed to effectively enhance the efficiency and accuracy of the detection processes. Firstly, an enhanced candidate frame extraction operator for the histogram-of-gradient feature set with low sensitivity parameters is proposed to locate the milling workpiece, and the installation error is corrected using the point matching algorithm. Then, the focusing process of the industrial camera is optimized via the sharpness evaluation metrics. Finally, a lightweight convolutional neural network model for real-time computing at mobile terminals is constructed. The proposed method realizes the classification of surface textures of workpieces with different roughness values, and is experimentally verified on the end milling texture data set. Taking the times of multiplication and addition as the metrics, the performed experiments indicate that the number of floating-point operations (e.g., add and multiply) required for model inference is reduced by 55%, compared with the general convolutional neural network. In addition, the introduced cost-sensitive loss effectively improves the model’s stability to unbalanced data. Compared with the traditional machine learning methods, the accuracy of the proposed model is improved under the same experimental conditions (i.e., detection frame rate and image resolution), where the recall rate is increased by 21%, and the accuracy rate is enhanced by 8% simultaneously.

To enhance the robustness of axle-box bearing fault detection under complex interference noise, the envelope spectrum construction method for axle-box bearing fault diagnosis is investigated by cyclic spectral analysis and considering the distribution difference of bearing fault information and threshold denoising. Firstly, the frequency domain signal-to-noise ratio is proposed as a new measure for bearing fault information quantification to evaluate the fault-related information in different spectral frequency bands of the spectral coherence. Secondly, a fault characteristic information distribution function with spectral frequency as the variable is constructed and an information threshold is adaptively determined to identify the spectral frequency components that are rich in fault information and dominated by interference noise in the spectral coherence; further, a weight function based on the fault characteristic information distribution function and the information threshold is designed. Finally, a multiband weighted envelope spectrum is generated from the spectral coherence with the weight function and is used to detect different axle-box bearing faults by analyzing the bearing fault characteristic frequencies. The analysis results of the experimental data of railway axle-box bearings show that compared with typical spectral coherence-based envelope spectrum methods, the multiband weighted envelope spectrum can accurately detect the faults of the outer race, rolling element and inner race of the axle-box bearing under complex interference noise and can achieve higher performance quantification indicators (frequency domain signal-to-noise ratio and negentropy).

In order to realise the real-time detection of wheel out-of-roundness (OOR) for high-speed trains, the spectral characteristics and mapping relationship between the axlebox high-frequency vibration and the wheel OOR are studied. The amplitude and order of the wheel OOR are identified using the frequency-domain integration method. Firstly, manifestations of the wheel polygonisation, rail corrugation and track modes of China high-speed railways are investigated through static measurements and laboratory experiments. Secondly, the time-frequency characteristics and evolution principle of the bogie axlebox vibrations are obtained through the long-term operation performance tracking test of a high-speed train. Finally, taking the 20th order wheel polygonisation as the research object, an identification approach of the wheel OOR order and amplitude is proposed based on the frequency-domain integration method. The results show that the third-order bending modal frequency of the China railway track system Ⅱ (CRTS-Ⅱ) track slab is 592 Hz. When the train is running at a speed of 300 km/h, the response frequencies due to the 20th-order wheel polygonisation and the rail corrugation with a wavelength of 136 mm are 580 and 613 Hz, respectively. The dominant frequencies of rail mode, wheel polygonisation and rail corrugation are relatively concentrated. The amplitude of high-frequency vibration of axlebox increases with the increase of vehicle speed and mileage after re-profiling. Theoretically, the amplitude and order of the wheel OOR can be identified by the integration of accelerations in the frequency domain. The relative error between the identified 20th-order wheel OOR results based on the field-tested axlebox acceleration and the static measured values is less than 5%.

The variable operating speed of mechanical equipment bearings leads to blurred spectrum of vibration monitoring signals, which to some extent affects the accuracy of bearing fault diagnosis. The current tacholess order tracking technology works well when the rotational speed fluctuation is small and the rotational frequency harmonics do not overlap, but it is difficult to conduct analysis when the bearing rotational frequency harmonics overlap. Aiming at the above problems, this paper proposes a tacholess order tracking fault diagnosis method of bearings based on the generalized demodulation. First, the generalized Fourier transform and the improved cost function-based ridge extraction technology are used to accurately extract the rotational frequency harmonic components of the bearings. At the same time, the fast spectral kurtosis algorithm and a band-pass filter are used to de-noise the bearing vibration signal. Then, the denoised time-domain signal is converted into an angular-domain signal through angle resampling. Finally, the order spectrum information of the bearing is obtained through the envelope spectrum analysis, so as to identify the fault type of the bearing. In addition, the effectiveness of the proposed method is verified using the numerical simulation signal and the actual bearing monitoring signal. The results show that the proposed method has an error of less than 5% in reconstructing the phase and an accuracy of more than 94% in characterizing the order frequency of bearing faults, which can be used for the order tracking fault diagnosis of rolling bearings without rotational speed information.

The operating conditions of axle-box bearings of urban rail transit trains are complex and time-varying, and they often suffer from random external interferences. Correspondingly, the monitoring data of axle-box bearings contain a great amount of measurement noise and even abnormal data, thereby limiting the accuracy of prognostics models. To overcome the aforementioned problems, a monitoring data-driven dynamic multiple aggregation prediction method is proposed for forecasting the remaining useful life (RUL) of axle-box bearings of urban rail transit trains. In the proposed method, abnormal data are first automatically recognized and deleted by measuring the amplitude distribution similarity between signals in a short time. Then, various degradation curves can be fitted to predict the mean and variance of RUL by aggregating health indicators from different temporal scales. The proposed method is evaluated using vibration data from real monitoring systems of urban rail transit trains and accelerated degradation tests of rolling element bearings. The results show that the proposed method is able to effectively recognize the not a number (NaN) data and strong interference data, and as time goes on, the predictive RUL converges to the actual RUL gradually and the 95% confidence interval becomes narrower. Further, compared with the single exponential prognostics model and the hybrid prognostics model, the proposed method increases the mean of cumulative relative accuracy by 29.78% and 27.63% respectively, and improves the mean of convergence speed by 10.56% and 10.20% respectively.