When the behavior recognition algorithm is deployed on the edge computing device of the cloud edge collaborative intelligent monitoring system, due to the lack of a reasonable task offloading scheme, the computing resources of the system are distributed unevenly, which leads to unstable system operation power consumption and affects the speed and accuracy of recognition. To solve the above problems, a task offloading scheme based on the number of scene characters has been designed to optimize the operational stability and recognition effect of the cloud edge collaborative intelligent monitoring system. Firstly, the operating parameters of the intelligent monitoring system were collected, and its power consumption curve and recognition performance were determined. Next, a lightweight character number recognition module was designed, and the classification of monitoring tasks based on the number of scene characters was realized by programming. Then, the influence of different video sampling rates on the power consumption and recognition performance of the intelligent monitoring system was tested, and the optimal sampling rate allocation scheme was determined. Finally, the proposed task offloading scheme was tested on the intelligent monitoring system for the production line of Fuxing electric multiple units. The results showed that compared with the existing parallel task offloading scheme, the task offloading scheme based on the number of scene characters improved the average recognition accuracy of the intelligent monitoring system of the production line by 0.53%, reduced average delay by 1.56%, and reduced average power consumption by 14.47%, which effectively improved the operational stability of the system. The research results are of great significance for optimizing the operational stability and recognition effect of the cloud edge collaborative intelligent monitoring system, and can provide theoretical basis and engineering support for its performance improvement.

In order to achieve the goal of carbon peak and carbon neutrality, it is necessary to continuously promote the high-quality development of manufacturing enterprises and establish a new development pattern of energy saving and consumption reduction. In response to the new development needs, the energy structure of manufacturing enterprises needs to be continuously optimized, and the power energy management method directly affects the energy structure layout of manufacturing enterprises. Different functional workshops in manufacturing enterprises have different equipment operation modes and energy consumption characteristics. Many years of production and operation have accumulated a large amount of energy usage data, but these data have not been fully utilized, resulting in data isolation and seriously affecting the production efficiency of manufacturing enterprises. Therefore, there is an urgent need for a power energy management method for different working conditions of manufacturing enterprises. In order to solve the above problems, taking steam energy as an example, the steam prediction models for process and air-conditioning were constructed to realize the intelligent supply and management of steam to meet the production process requirements of manufacturing enterprises. Firstly, a steam prediction model for process based on work section division was proposed to forecast process steam consumption with strong periodic characteristics. After the model improvement, the average standard energy consumption for process was reduced by 5.12%. Then, a steam prediction model for modular air-conditioning based on hybrid deep learning and a steam prediction model for independent air-conditioning based on multiple scenarios were constructed. Through comparing with other prediction models, the effectiveness and accuracy of the proposed model were verified. The results showed that the proposed power energy prediction model had wide applicability and could be applied to the management of different power energy in other manufacturing enterprises after proper modification and adjustment. The research results are helpful for relevant manufacturing enterprises to make full use of historical energy usage data, achieve energy conservation and efficiency improvement in power energy, and provide strong support for the digital transformation and upgrading of manufacturing enterprises in China.

As a common application mode of multi-modal data registration, the registration of visible images and infrared images of power equipment has significant application value in substation inspection process. However, due to the differences in resolution, perspective and lighting conditions between data captured by inspection robots and humans, the registration of visible images and infrared images of power equipment faces considerable challenges. Aiming at the feature-level consistency between two types of images, a scale-invariant feature transform algorithm based on the power equipment edge point features was proposed. By constructing bilateral feature descriptors and incorporating the distance between bilateral feature points as a confidence constraint, the modal difference between visible images and infrared images was reduced, and the high-precision registration of multi-source heterogeneous image data of critical equipment for substation primary and secondary power distribution networks was achieved. The experimental results showed that the proposed method exhibited strong robustness against variations in resolution and perspective, delivering excellent performance in practical applications of multi-modal image registration for power equipment. The research results provide theoretical support and engineering guidance for enhancing the intelligence and reliability of power equipment inspection.

Aiming at the low intelligence level of main transformers in nuclear power plants, the fact that the operation and maintenance still mainly rely on manual inspection and regular maintenance, and the lack of real-time prediction and evaluation means of operating state, the digital twin of the main transformer based on fast calculation of coupling field has been built. Firstly, a refined digital model of the main transformer was established and the multi-physical field real-time simulation analysis was carried out to obtain the electromagnetic field and loss distribution cloud maps of its key components. Then, in order to solve the problems that the simulation of the flow field and temperature field of the main transformer took a long time and the real-time monitoring of operating states could not be realized, a fast calculation method of coupling field was proposed. Finally, the digital twin of the main transformer was constructed by combining the proposed method with the operating state data collected by various sensors, achieving multi-dimensional, full process and panoramic state monitoring and display for the physical main transformer. The results show that the construction of digital twins can realize the real-time monitoring and predictive evaluation of the operating state of main transformers under different working conditions, which provides convenience for the operation and maintenance inspection of the main transformer and contributes to the safe and reliable operation of nuclear power plants.

The propeller-motor matching design of electric propulsion system is a key link for high-altitude unmanned aerial vehicle (UAV) to achieve long-endurance flight. According to the full-profile power requirements of the high-altitude UAV, the overall parameter matching design for the motor and propeller of the electric propulsion system was carried out by using the power loss method and the single-point method, respectively, and the tensile force and power characteristics of the electric propulsion system were measured through ground static test. At the same time, the surrogate model based on BP (back propagation) neural network was established on the basis of the confirmed motor selection, in order to carry out the propeller-motor matching design and efficiency improvement for the electric propulsion system. The test results showed that the measured tensile force was consistent with the calculated tensile force, which indicated that the adopted overall parameter matching design method for the motor and propeller had high accuracy. Taking the climb profile with elevation angle of 5° as an example, the efficiency of the new propeller obtained by optimized matching through the surrogate model was improved by about 0.1, and the energy saving effect was remarkable. The relevant surrogate model can provide a powerful tool for the propeller-motor matching and optimization design of the UAV electric propulsion system in the scheme design and later engineering use stage, as well as the improvement of dynamics characteristics and economy.

Aiming at the problems of permanent magnet synchronous motors with high power density, large torque and small volume for vehicle driving, such as small effective heat dissipation area of traditional air-cooled structure and high temperature of internal components caused by electromagnetic loss during operation, an oil-air hybrid cooling method with natural air cooling of internal cavity oil-cooled chassis was proposed, to meet the temperature performance requirements of each component in the drive motor. The equivalent thermal network method was used to calculate the temperature of the stator winding, stator, permanent magnet and rotor in the drive motor under different working conditions, and the highest temperature of the drive motor appeared at the stator winding. Then, the temperature at the stator winding end of the drive motor was measured by experiment and compared with the simulation results. The relative error between the simulation results and the measured results was less than 5%. The results showed that the temperature of the stator winding and other components of the oil-air hybrid cooling drive motor under different working conditions dropped obviously and met the temperature performance requirements, which indicated that the oil-air hybrid cooling method had good heat dissipation performance and high cooling efficiency. The research results can provide reference for the development of heat dissipation systems for vehicle drive motors.

Aiming at the difficulties of on-line fault diagnosis of centrifugal pumps in nuclear power plants, a fault diagnosis method based on wavelet pack decomposition and random forest is proposed. Firstly, the wavelet pack decomposition was used to decompose the vibration signal in the radial vertical direction of the centrifugal pump motor drive end into three layers, and the sub-band energy features were extracted. Then, the time-domain statistical features were extracted based on the waveform data of centrifugal pump vibration signal, and combined with wavelet packet energy features as inputs for the random forest model. Finally, the random forest model was trained with centrifugal pump vibration dataset collected from vibration test, and the centrifugal pump fault diagnosis model was formed. This model was compared with machine learning models such as support vector machine, logistic regression, K-nearest neighbor and Gaussian Naive Bayes on the same centrifugal pump vibration dataset. The results showed that the constructed model could accurately identify different operating states of the centrifugal pump, such as normal operation, impeller damage, impeller blockage and motor bearing fault, and exhibited better classification performance. The fault diagnosis method based on wavelet packet decomposition and random forest can effectively extract features from vibration signals and realize fault classification, which has certain feasibility and effectiveness for on-line fault intelligent diagnosis of centrifugal pumps in nuclear power plants.

With the continuous development of new energy systems, the automation level of substation has a crucial impact on the stable operation of the power grid and the maintenance of metering equipment. The accurate acquisition of meter panel readings is one of the key links of achieving substation automation, which is of great significance to the status monitoring and fault diagnosis of substation metering equipment. However, due to the complexity of meter panel readings and the impact of various environmental factors such as light and angle, the automatic recognition of meter panel readings presents significant challenges. In order to solve this problem, an automatic recognition method for substation meter panel readings based on Fast R-CNN (regional convolutional neural network) and DeepLabV3+ was proposed. Firstly, the target detection technology based on Fast R-CNN was analyzed theoretically, and its training process was described in detail by using the data set of substation meter panel. Then, the semantic segmentation model of meter panel based on DeepLabV3+ and the reading calculation method were designed. Finally, the experiments of automatic identification of substation meter panel readings were conducted to verify the effectiveness and accuracy of the proposed method. The experimental results showed that the proposed method could recognize the readings of substation meter panel efficiently and accurately, and had good robustness. The automatic identification method for meter panel readings based on Fast R-CNN and DeepLabV3+ can improve the working efficiency, safety and reduce the operation and maintenance cost of substations, and further promote the intelligent process of power systems.

The development of new energy systems increasingly emphasizes the state monitoring and performance prediction of electronic equipment. The battery is an important part of the new energy system, and the accurate monitoring and prediction of its service life and performance is of profound significance to improve the performance of electronic equipment, reduce maintenance costs and enhance energy efficiency. However, due to the influence of various complex factors on the battery performance, predicting its remaining life remains a major challenge. To solve these problems, a new battery remaining life prediction model was proposed. Firstly, the in-depth theoretical research was conducted on residual neural network (ResNet), bidirectional long short-term memory (BiLSTM) network and multi-head self-attention (MHSA) mechanism. Then, based on the above theories, the battery remaining life prediction model based on MHSA-Res-BiLSTM was constructed, and its hyperparameters were optimized. Finally, the battery remaining life prediction experiment was carried out to verify the performance of the proposed MHSA-Res-BiLSTM network. The experimental results showed that the proposed model performed excellently in the prediction of battery remaining life. Compared with other prediction models, the proposed prediction model had lower mean absolute error and root mean square error. The battery remaining life prediction model based on MHSA-Res-BiLSTM has good predictive performance and convergence performance, which can provide theoretical and technical support for the health management of batteries in new energy systems.

The method of variable-speed driving variable pump is mostly adopted for the bidirectional flow changes of hydraulic system. However, the structure of variable mechanism is relatively complicated, and frequent bidirectional variable regulation will lead to low dynamic response speed. Therefore, a new mechanism for flow regulation of fixed displacement pump was proposed to realize bidirectional variable function. The duty cycle of fluid pulse width modulation was changed by the axial movement of the spool of bidirectional variable mechanism. The product of the rotating speed of the spool relative to the valve sleeve and the number of valve port on the spool determined the frequency of fluid pulse width modulation. The flow characteristics of the bidirectional variable mechanism were analyzed by simulation and test. The results showed that the bidirectional variable mechanism could achieve bidirectional flow control with duty cycle of 0 to 100%, ane there was only 1.8% deviation between the simulation and test results. The research results provide a new flow control method for hydraulic pump with fixed speed design to use under variable speed working condition.

In order to reduce the energy consumption of the load sensitive system of a fixed displacement pump, the valve group composed of load sensitive valves in the pump control system was replaced by a two-dimensional pulse width modulated rotary valve, and a load sensitive system based on fluid pulse width modulation was designed. According to the working principle of the load sensitive system, the AMESim simulation model of the system was established, the load sensitive characteristics and energy consumption of the system were simulated and analyzed, and an experimental platform was built for experimental verification. The results showed that the combination of two-dimensional pulse width modulated rotary valve and fixed displacement pump made the system load sensitive, so that the fixed displacement pump output power and load consumption power changed with the load, which reduced the overflow loss and improved the energy saving effect of the system.

In order to improve the stress condition of piling rig, a novel multi-stage luffing mechanism of piling rig was proposed. The mechanism could effectively reduce the bending moment of the front hydraulic cylinder and improve the stability of piling rig during operation. Firstly, the spatial coordinate system was established, and the forward and inverse solutions of the spatial position of the mechanism were analyzed, and then the velocity Jacobian matrix and acceleration Hessian matrix were derived, and the kinematics model of the mechanism was obtained. Secondly, after the mechanism parameters and the displacement function of the hydraulic cylinder were set according to the luffing conditions, the motion characteristics of the mechanism were analyzed using MATLAB software, the connection mode of the hydraulic cylinder contraction was determined, and the influence of the mechanism parameters on the luffing motion was further analyzed. Finally, the ADAMS simulation results of luffing motion of the mechanism were compared with the theoretical analysis results. The results showed that the kinematics model of the multi-stage luffing mechanism was correct. The mechanism could steadily luff the column from horizontal to vertical, the slider always moved upward during the luffing process, and the connecting rod after luffing was vertical to the front hydraulic cylinder. Accelerating contraction of the rear hydraulic cylinder while decelerating contraction of the front hydraulic cylinder was conducive to stable luffing. The research results can provide a reference for the dynamics analysis and optimal design of the novel multi-stage luffing mechanism of piling rig.

Predictive maintenance is one of the typical applications of digital twin, and data-driven is the main way to realize predictive maintenance. Aiming at the problems of difficult fault feature extraction and large deviation of prediction results in predictive maintenance, a predictive maintenance model building method based on the combination of VMD (variational mode decomposition) algorithm and HHT (Hilbert-Huang transform) algorithm was proposed. The time-domain features of the vibration signal were extracted by VMD+HHT algorithm, the data dimension was reduced by combining deep sparse auto-encoder (DSAE), support vector data description (SVDD) algorithm was used to form a health index curve, and a predictive maintenance model was established based on long short-term memory (LSTM) algorithm. The method was applied to the predictive maintenance of a hydraulic motor of a rotary drilling rig. The vibration signal of the motor housing was extracted, the predictive maintenance model of the hydraulic motor was constructed, and the validity and accuracy of the method were verified by test. The test results showed that adopting the predictive maintenance model based on DSAE+SVDD+LSTM algorithm could avoid the problems of mode aliasing and endpoint effect, the prediction accuracy could reach more than 90%, and the model had practical value. The research results can provide important reference for the construction of hydraulic component digital twin predictive maintenance application scenarios.

Large vertical mill is a grinding system with complex structure, multiple operating parameters, multiple monitoring and operating points, and multiple physical fields, and its performance is directly related to the structure. Based on the parametric design method of NX secondary development, CFD-DPM (computational fluid dynamics-discrete phase model) of LGM large vertical mill was established. The grating structure of the powder separator was analyzed and designed, and the influence of the rotor blade structure on the performance of the mill was discussed. A structure optimization design method for powder separator based on Kriging model was proposed and verified on Isight platform. The results showed that the rotor torque and the velocity of the airflow between rotor blades could be effectively improved by appropriately reducing the waist length and head length of the rotor blades and increasing their thickness and inclination angle within the parameter design range. The torque was increased by 2.91% and the flow rate by 9.76% after optimization, which proved the effectiveness of the proposed method. The research results provide a new theoretical basis and practical guidance for the structural design and optimization of vertical mill, and have high scientific research reference value and industrial application prospect.

Crane is one of the eight special equipment, the rationality of its structural design is crucial for the safe operation, quality improvement, and efficiency enhancement of the equipment. Therefore, an optimization design method of metal structure of bridge crane based on structural function derivative coefficients was proposed. Firstly, the interpretive structure model was used to accurately describe the interactions and influence relationships between different parts of the crane, the structure complexity of the crane was analyzed to identify and optimize the parameters that had a significant impact on the strength of metal structure of the crane. Secondly, based on the finite element simulation, the structural function derivative coefficients were combined with fitting functions to find the optimal combination of design parameters. Finally, a 300/100 t-30 m bridge crane was taken as an engineering example, and the mass of the whole crane was reduced by 2 035.113 kg through simulation, and the effectiveness of the proposed method was verified. The optimization design method of metal structure of crane based on structural function derivative coefficients solves the problem that the complex interactions between various parts of crane is ignored in the traditional optimization method, and can realize the lightweight design of metal structure on the basis of ensuring the safety of crane service.