A Single-Machine Reliability Assessment Method for Control Systems Based on Convolutional Neural Networks
By combining physical and data models using convolutional neural networks, the problem of long life assessment time in traditional control systems has been solved, enabling fast and accurate single-machine reliability assessment, adapting to multiple failure modes and complex operating conditions, and improving assessment efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2025-12-30
- Publication Date
- 2026-06-02
AI Technical Summary
Traditional methods for assessing the lifespan of control systems rely on testing and simulation, which are time-consuming and difficult to adapt to batch single-machine assessments, random operating condition changes, and online monitoring. They also cannot effectively consider multiple failure modes and complex operating conditions.
By employing a convolutional neural network-based approach, combining physical and data models, point cloud data is generated through finite element simulation. The point cloud convolutional neural network is then trained as a surrogate model to achieve rapid and accurate reliability assessment of the control system.
It can complete the reliability assessment of a single-machine control system in milliseconds, improving the assessment efficiency and accuracy. It can adaptively learn the nonlinear coupling relationship between materials and structures, and achieve efficient reliability prediction under multiple failure modes and complex operating conditions.
Smart Images
Figure CN122131736A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for evaluating the reliability of a single control system based on a convolutional neural network, belonging to the field of single-machine life prediction for control systems. Background Technology
[0002] With the widespread application of modern electronic control systems, evaluating their reliability under different environments has become increasingly important. Control systems operating in complex environments are susceptible to performance degradation and even failure due to stress caused by temperature and vibration. Ensuring the long-term reliability and stability of control systems is particularly crucial in fields such as automation, intelligent manufacturing, and industrial equipment. Especially in military equipment, control systems face stringent requirements for long-term standby power-on and instantaneous operational use, demanding high reliability under complex environments of high temperature, high humidity, and vibration. Traditional methods for assessing the lifespan of control systems primarily rely on theoretical models and experimental verification. These methods suffer from limitations such as mismatch between models and actual operating conditions, inability to cover complex conditions, and significant deviations in assessment results due to environmental disturbances and load fluctuations.
[0003] Finite element method (FEM) simulation has numerous applications in control system life assessment. By setting environmental and material parameters, FEM simulation is used to conduct virtual experiments to evaluate the environmental parameters of the control system, thereby predicting its lifespan. However, FEM simulation suffers from high computational costs and long processing times. Furthermore, while existing physical models of failure can reflect the degradation process of a control system, they fail to effectively consider lifespan prediction under multiple failure modes and complex operating conditions. Therefore, there is an urgent need for a new technology that combines physical knowledge with data-driven approaches to achieve more accurate and efficient control system life assessment, especially for control systems under multiple failure modes and complex operating conditions. Summary of the Invention
[0004] Traditional control system life assessment relies on accelerated testing and time-consuming simulations, making it difficult to adapt to batch single-machine assessments, random operating condition changes, and online monitoring requirements. The purpose of this invention is to provide a single-machine reliability assessment method for control systems based on convolutional neural networks. This method integrates a physical model and a data model: the physical model (Physics of Fluids, PoF) provides mechanistic constraints and a damage evolution framework, while the point cloud convolutional neural network provides high-dimensional spatial feature learning and rapid prediction capabilities. This approach ensures that the model has a physical basis and also has the advantages of data-driven analysis, achieving accurate predictions while improving interpretability and reliability.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] This invention discloses a single-machine reliability assessment method for control systems based on convolutional neural networks. It constructs a finite element model of a single-machine control system component and samples the stress values or temperature information of discrete grid points, grid centers, and their respective points in the finite element numerical simulation to obtain the stress and temperature fields of the control system's point cloud model. By setting the finite element simulation environment configuration and performing Latin hypercube sampling on material parameters to obtain several sets of point cloud data, stress and temperature field data for training the point cloud convolutional neural network are obtained. Through training the constructed point cloud neural network, a rapid predictive proxy model of the control system's gravitational and temperature fields is obtained. The input is the material parameters of the control system, and the output is a point cloud model with gravity or temperature information. Fully supervised training is performed using point cloud data obtained from discrete sampling via finite element simulation to improve the accuracy and consistency of the surrogate model's predictions. The dropout layer inside the point cloud neural network also avoids overfitting of the surrogate model. The surrogate model is used to quickly predict the stress and temperature fields under various control system hardware materials. The stress and temperature information at each point in the point cloud model, along with the material's strength and temperature thresholds, are used to obtain the reliability index for a single experiment. Monte Carlo simulation is used to conduct a large number of experiments to achieve rapid reliability prediction, including the following steps:
[0007] Step 1: Construct a point cloud model of the control system components and sample the original stress and temperature data of the point cloud model;
[0008] By leveraging the mesh discretization characteristics of finite element simulation, large amounts of stress and temperature field data can be obtained quickly and inexpensively. This involves the following sub-steps:
[0009] Step 1.1: First, build a physical model of the control system unit in the finite element simulation software, and set the spatial coordinates of grid points and grid centers, as well as their initial stress and temperature properties.
[0010] As a preferred option, a sparser grid is selected in areas with fewer components, and a denser grid is selected in areas with more components.
[0011] Step 1.2: By setting different material parameters, perform batch finite element simulations of the stress field and temperature field, and extract point cloud data for training.
[0012] As a preferred option, only material parameters that have a significant impact on the actual lifespan of the control system and the finite element simulation results are considered.
[0013] Step 1.3: Perform Latin hypercube sampling on the finite element simulation experiment;
[0014] By extracting several sets of sample points, the finite element simulation software is further developed, and scripts are used to automatically read the material parameter values in the sample points and conduct batch experiments.
[0015] All material parameters are randomly sampled within a set range using a uniform distribution to enable the import of different material parameters;
[0016] Several sets of finite element analysis results for temperature fields and several sets of finite element analysis results for stress fields were obtained for different parameters.
[0017] Step 2: Construct a rapid prediction proxy model for stress and temperature fields based on point cloud neural networks;
[0018] The experimental data obtained in step one is used to train a point cloud convolutional neural network model. This process handles discrete and ordered point cloud data to ensure the consistency, accuracy, and timeliness of stress and temperature field predictions. Specifically, this includes the following sub-steps:
[0019] Step 2.1: Use a linear fully-connected (FC) layer to map low-dimensional material information to a higher-dimensional feature space, extract the implicit relationship between material parameters and stress / temperature fields, and establish the coupling relationship between parameters nonlinearly through the ReLU activation function;
[0020] The training data for the model comes from the point cloud data obtained by batch finite element simulation sampling in step one. The input is the material parameters of the simulated control system. The material parameters determine the response law of the final 3D stress / temperature field. Simple direct coupling of material features will result in insufficient feature extraction. FC dimension upscaling is used, as shown in equation (1):
[0021] (1)
[0022] in, For the first Material parameter characteristics of this experiment Features after FC dimensionality upgrade;
[0023] Step 2.2: Extend high-dimensional material features to the point cloud model;
[0024] Point cloud data, as an unordered collection of points in three-dimensional space, has randomness and rotation issues. The order in which the points are stored does not affect the spatial position or the geometric relationship of the model. However, the same point cloud model can be represented by completely different storage matrices, and the weights trained by the neural network are also completely different.
[0025] Multi-perceptrons (MLPs) based on full-joint (FC) do not possess rotation invariance and require fixed-dimensional input. Their sensitivity to the order of point cloud data storage severely affects the generalization ability and robustness of MLPs.
[0026] By introducing a 3D stress / temperature field, the positional relationships of point clouds are directly represented relative to each other in the 3D storage matrix, ensuring that the order of data points stored in the point cloud model does not affect the accuracy of the 3D set feature representation.
[0027] The multidimensional features of material parameters are extracted using FC and extended to a 3D stress / temperature field, as shown in Equation (2):
[0028] (2)
[0029] in, For a 3D stress / temperature field, its dimensions are... This provides information on the length (depth), width, and height of the temperature / stress field in space.
[0030] Step 2.3: Aggregate high-dimensional spatial features to obtain the physical parameters of each voxel in the stress / temperature field;
[0031] A 3D convolutional neural network is used to aggregate spatial features of the stress / temperature field. In this case, the input of the convolutional neural network is an FC-encoded vector that is expanded into a feature tensor of a three-dimensional field. ;
[0032] By sliding the convolution kernel in the depth, height, and width directions respectively, spatial features of the local neighborhood are extracted. The convolution process is shown in equation (3):
[0033] (3)
[0034] in, The 3D convolution kernel is set;
[0035] The training process of the model is realized by optimizing the parameters of the convolution kernel. The feature aggregation method can combine the coupling relationship between material parameters and structural spatial position, adaptively learn the transmission path and local change law of stress / temperature in three-dimensional structure, and realize accurate modeling of high temperature hot spot area and potential failure location inside the control system.
[0036] The features obtained after traversing the last convolutional kernel are the stress / temperature of the corresponding voxel;
[0037] The number of point clouds obtained by random mesh sampling in finite element simulation is not regular, i.e. The traversal region and convolution kernel of the 3D convolutional neural network must be a regular cubic structure. Point cloud mask constraints are used to retain only the effective sampling region, avoiding errors caused by regular voxel interpolation and improving the consistency between the prediction results and the actual physical structure.
[0038] Step 3: Reliability prediction of control systems based on surrogate models;
[0039] The surrogate model for rapid prediction of stress and temperature fields based on the point cloud model trained in step two is used for inference. A large number of random input samples are generated using the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress or temperature field response. Finally, a consistent and efficient reliability assessment of the single-machine control system is obtained. The specific steps include the following:
[0040] Step 3.1: Construct the statistical distribution of the uncertain input variables;
[0041] Uncertainty input variables include: material properties, environmental loading parameters, manufacturing errors, and operating conditions;
[0042] Statistical distributions include: Gaussian distribution, uniform distribution, or engineering experience distribution;
[0043] By using surrogate models to replace traditional finite element simulation or accelerated testing, stress / temperature response prediction under different material parameters and working conditions can be performed in multiple scenarios with extremely low computational cost, enabling rapid analysis of the reliability indicators of the control system.
[0044] Step 3.2: Determine the reliability of the control system;
[0045] Based on step 3.1, a large number of random input samples are generated using the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress field or temperature field response;
[0046] For a given stress failure criterion, the following reliability criterion is adopted: when the predicted stress at any point of the structure... Not exceeding the material strength threshold If the sample is functioning normally, it is considered to be in operation; otherwise, it is considered to be in failure.
[0047] The reliability index of the control system is shown in equation (4):
[0048] (4)
[0049] That is, as shown in equation (5):
[0050] (5)
[0051] in, This represents the number of samples that satisfy the failure criterion in the Monte Carlo simulation. This represents the total number of samples;
[0052] Through rapid surrogate model inference, stress response analysis of a single sample can be completed in milliseconds, enabling Monte Carlo simulation to be performed with extremely high efficiency. This solves the problem that traditional control system reliability analysis cannot perform large-sample statistics due to the long simulation analysis time.
[0053] The process also includes step four, which applies the reliability prediction results obtained in step three to the online operation and maintenance of the control system. The real-time environmental data and performance parameters of the actual operation of the control system are input into the surrogate model trained in step two, and the degradation trajectory with confidence intervals for future time periods is output. When the predicted degradation trajectory approaches the preset failure threshold, an early warning signal is automatically triggered and a maintenance instruction is generated. By utilizing the millisecond-level inference capability of the surrogate model, the problem of resource waste and high risk of post-maintenance in traditional periodic maintenance is solved, realizing the transformation from "planned maintenance" to "predictive maintenance" and ensuring the safe operation of the control system under random operating conditions.
[0054] Beneficial effects:
[0055] 1. The present invention provides a single-machine reliability assessment method for control systems based on convolutional neural networks. By using a "physical model + data model" fusion strategy, high-fidelity data is generated by combining finite element simulation with Latin hypercube sampling. Point cloud convolutional neural networks are trained as fast surrogate models. Under the premise of preserving physical mechanism constraints, the single prediction time is reduced to the millisecond level, which solves the technical problems of high time consumption and difficulty in balancing efficiency and physical accuracy in traditional methods, and effectively improves the assessment efficiency.
[0056] 2. The present invention provides a single-machine reliability assessment method for control systems based on convolutional neural networks. It uses a fully connected layer coupled 3D convolutional neural network architecture to process point cloud data. By extending the feature dimension to a three-dimensional field and using mask constraints to extract effective spatial features, it overcomes the disorder of point clouds and geometric error problems. It can adaptively learn the nonlinear coupling relationship between materials and structures, and improve the consistency and accuracy of predicting high-temperature hot spots and stress concentration areas inside the control system.
[0057] 3. The present invention provides a single-machine reliability assessment method for control systems based on convolutional neural networks. It employs a deep learning surrogate model to perform large-scale Monte Carlo simulations and performs rapid inference on massive random operating condition samples based on failure criteria. This method can quickly obtain convergent and statistically significant reliability indicators and failure probabilities, thereby achieving high-efficiency and full-probability reliability assessment of control systems under random operating conditions and parameter fluctuations. Attached Figure Description
[0058] Figure 1 This is a schematic diagram of the overall process of a single-machine reliability assessment method for a control system based on a convolutional neural network according to the present invention.
[0059] Figure 2 This is a schematic diagram of the network architecture of the fast prediction proxy model for stress and temperature fields based on convolutional neural networks of the present invention.
[0060] Figure 3 This is a graph showing the loss decrease during the training process of the proxy model in the embodiment;
[0061] Figure 4 This is a comparison chart of the reliability assessment results based on the proxy model in the embodiments. Detailed Implementation
[0062] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.
[0063] Example 1
[0064] This embodiment uses finite element simulation data of a single engine control unit of a certain type of aero-engine control system to obtain stress field and temperature field point cloud data of the single control unit in 150 experiments. Each point cloud contains 16,436 sampling points. Due to the large number of samples, only the stress data of the first 50 sample points of a single sampling are shown, as shown in Table 1.
[0065] Table 1
[0066]
[0067] In the above scenario, the single-machine reliability assessment method for control systems based on convolutional neural networks of this invention is applied to perform rapid reliability assessment of the control system, such as... Figure 1 As shown, it includes the following steps:
[0068] Step 1: Construct a point cloud model of the control system components and sample the original stress and temperature data of the point cloud model;
[0069] By leveraging the mesh discretization characteristics of finite element simulation, large amounts of stress and temperature field data can be obtained quickly and inexpensively. This involves the following sub-steps:
[0070] Step 1.1: First, build a physical model of the control system unit in the finite element simulation software Abaqus, and set the spatial coordinates of grid points and grid centers, as well as their initial stress and temperature properties.
[0071] Choose a sparser mesh for areas with fewer components and a denser mesh for areas with more components.
[0072] Step 1.2: In Abaqus, perform batch finite element simulations of the stress and temperature fields by setting different material parameters, and extract point cloud data for training. Specific material parameters include:
[0073] thermal conductivity Specific heat capacity Young's modulus Poisson's ratio ;
[0074] Based on the assembly characteristics of the circuit board assembly method in the airborne electronic control unit, the boundary condition for finite element simulation is determined to be four-end fixed support;
[0075] Based on the operating environment of the control unit, the stress load and temperature load settings are determined as follows:
[0076] Stress load settings: Random vibration direction is two directions parallel to the printed circuit board plane, overload is 4G, random action 20 times, total duration 1h, emission load direction is the normal direction of the printed circuit board plane, overload magnitude is 30G, action time is 70ms;
[0077] Temperature load settings: The time-varying temperature parameters are determined based on the temperature changes in the natural environment. From 0:00 to 6:00, the ambient temperature rises from 10℃ to 20℃ to simulate the morning warming process. From 6:00 to 12:00, the temperature continues to rise to 30℃ to simulate the midday high temperature. From 12:00 to 18:00, the temperature linearly decreases from 30℃ to 20℃ to simulate the afternoon cooling. From 18:00 to 24:00, the temperature further decreases to 10℃ to simulate the nighttime low temperature.
[0078] Step 1.3: Perform Latin hypercube sampling on the finite element simulation experiment;
[0079] Several sets of sample points were extracted to conduct secondary development of the simulation software. Python scripts were used to automatically read the material parameter values in the sample points and conduct batch experiments.
[0080] All material parameters are randomly sampled within a set range using a uniform distribution to enable the import of different material parameters;
[0081] For the finite element simulation of the temperature field, 1500 sets of sample points were selected and imported into the finite element simulation. For the 1500 simulation results, the temperature of each node was extracted as the sample output. For the finite element simulation of the stress field, the displacement of each node in each direction was extracted as the sample output.
[0082] Step 2: Construct a rapid prediction proxy model for stress and temperature fields based on point cloud neural networks;
[0083] The experimental data obtained in step one is used to train a point cloud convolutional neural network model. This process handles discrete and ordered point cloud data to ensure the consistency, accuracy, and timeliness of stress and temperature field predictions. Specifically, this includes the following sub-steps:
[0084] Step 2.1: Use a linear fully-connected (FC) layer to map low-dimensional material information to a higher-dimensional feature space, extract the implicit relationship between material parameters and stress / temperature fields, and establish the coupling relationship between parameters nonlinearly through the ReLU activation function;
[0085] In the embodiment, the network architecture of the surrogate model for fast prediction of stress and temperature fields based on convolutional neural networks is as follows: Figure 2 As shown, the material parameters from the input layer are mapped to a high-dimensional feature space through a fully connected layer (FC), and then expanded to form a 3D stress / temperature field tensor. Finally, 3D convolution is used for feature aggregation and prediction.
[0086] The training data of the model comes from the point cloud data obtained by batch finite element simulation sampling in step one. The input includes material parameters, consisting of only four material parameters: thermal conductivity, specific heat capacity, Young's modulus, and Poisson's ratio. The material parameters determine the response law of the final 3D stress / temperature field. Direct coupling using thinking features will result in insufficient feature extraction. FC dimension upscaling is used, as shown in equation (1):
[0087] (1)
[0088] in, For the first Material parameter characteristics of this experiment For the features after FC dimensionality upgrade, dimension C is set to 128 in this embodiment;
[0089] Step 2.2: Extend high-dimensional material features to the point cloud model;
[0090] Point cloud data, as an unordered collection of points in three-dimensional space, has randomness and rotation issues. The order in which the points are stored does not affect the spatial position and geometric relationship of the model. However, the same point cloud model can be represented by completely different storage matrices, and the weights trained by the neural network are also completely different.
[0091] Multi-perceptrons (MLPs) based on full-joint (FC) do not possess rotation invariance and require fixed-dimensional input. Their sensitivity to the order of point cloud data storage severely affects the generalization ability and robustness of MLPs.
[0092] A 3D stress / temperature field is introduced, and the positional relationships of the point cloud are directly represented relative to each other in the 3D storage matrix, ensuring that the order of data points stored in the point cloud model does not affect the accuracy of the 3D set feature representation.
[0093] The multidimensional features of material parameters are extracted using FC and extended to a 3D stress / temperature field, as shown in Equation (2):
[0094] (2)
[0095] in, For a 3D stress / temperature field, its dimensions are... This provides information on the length (depth), width, and height of the temperature / stress field in space.
[0096] In the example, the scale of the convolution space is set to... ;
[0097] Step 2.3: Aggregate high-dimensional spatial features to obtain the physical parameters of each voxel in the stress / temperature field;
[0098] A 3D convolutional neural network is used to aggregate spatial features of the stress / temperature field. In this case, the input of the convolutional neural network is an FC-encoded vector that is expanded into a feature tensor of a three-dimensional field. ;
[0099] By sliding the convolution kernel in the depth, height, and width directions respectively, spatial features of the local neighborhood are extracted. The convolution process is shown in equation (3):
[0100] (3)
[0101] in, The 3D convolution kernel is set;
[0102] The training process of the model is realized by optimizing the parameters of the convolution kernel. The feature aggregation method can combine the coupling relationship between material parameters and structural spatial position, adaptively learn the transmission path and local change law of stress / temperature in three-dimensional structure, and realize accurate modeling of high temperature hot spot area and potential failure location inside the control system.
[0103] In the embodiment, the dimension of the feature is set to ;
[0104] The features obtained after traversing the last convolutional kernel are the stress / temperature of the corresponding voxel;
[0105] The number of point clouds obtained by random mesh sampling in finite element simulation using Abaqus is not regular, i.e. In the example, the number of 3D points sampled is 16436. However, the traversal region and convolution kernel of the 3D convolutional neural network must be regular cubic structures. Point cloud mask constraints are used to retain only the effective sampling region, avoid the error caused by regular voxel interpolation, and improve the consistency between the prediction results and the actual physical structure.
[0106] In the embodiment, the loss descent curve during the surrogate model training process is as follows: Figure 3 As shown, the training loss and validation loss converge with the number of iterations as the number of training iterations (Epoch) increases. Both losses converge to zero within 50 training epochs, validating the effectiveness of model training and its fast convergence.
[0107] Step 3: Reliability prediction of control systems based on surrogate models;
[0108] The surrogate model for rapid prediction of stress and temperature fields based on the point cloud model trained in step two is used for inference. A large number of random input samples are generated using the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress or temperature field response. Finally, a consistent and efficient reliability assessment of the single-machine control system is obtained. The specific steps include the following:
[0109] Step 3.1: Construct the statistical distribution of the uncertain input variables;
[0110] Uncertainty input variables include: material properties, environmental loading parameters, manufacturing errors, and operating conditions;
[0111] Statistical distributions include: Gaussian distribution, uniform distribution, or engineering experience distribution;
[0112] By using surrogate models to replace traditional finite element simulation or accelerated testing, stress / temperature response prediction under different material parameters and working conditions can be performed in multiple scenarios with extremely low computational cost, enabling rapid analysis of the reliability indicators of the control system.
[0113] Step 3.2: Determine the reliability of the control system;
[0114] Based on step 3.1, a large number of random input samples are generated using the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress field or temperature field response;
[0115] For a given stress failure criterion, the following reliability criterion is adopted: when the predicted stress at any point of the structure... Not exceeding the material strength threshold If the sample is functioning normally, it is considered to be in operation; otherwise, it is considered to be in failure.
[0116] The reliability index of the control system is shown in equation (4):
[0117] (4)
[0118] That is, as shown in equation (5):
[0119] (5)
[0120] in, This represents the number of samples that satisfy the failure criterion in the Monte Carlo simulation. This represents the total number of samples;
[0121] Through rapid surrogate model inference, stress response analysis of a single sample can be completed in milliseconds, enabling Monte Carlo simulation to be performed with extremely high efficiency. This solves the problem that traditional control system reliability analysis cannot perform large-sample statistics due to the long simulation analysis time.
[0122] In the embodiments, the comparison of reliability assessment results based on the proxy model is as follows: Figure 4 As shown, the numerical differences between the finite element simulation results (Ground Truth) and the surrogate model prediction results (Prediction) under different point indices show that the root mean square error between the true value and the predicted value is 0.01. This indicates that the surrogate model prediction results of this invention have accuracy and consistency.
[0123] Step 4: The reliability prediction results of the single-engine control system of the aero-engine system are applied to the online operation and maintenance of the control system. The real-time environmental data and performance parameters of the actual operation of the control system are input into the surrogate model trained in Step 2, and the degradation trajectory with confidence intervals for future time periods is output. When the predicted degradation trajectory approaches the preset failure threshold, an early warning signal is automatically triggered and maintenance instructions are generated. By utilizing the millisecond-level inference capability of the surrogate model, the problems of resource waste and high risk of post-maintenance in traditional periodic maintenance are solved, realizing the transformation from "planned maintenance" to "predictive maintenance" and ensuring the safe operation of the control system under random operating conditions.
[0124] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for evaluating the reliability of a single control system based on a convolutional neural network, characterized in that, By constructing finite element models of single-machine control system components and sampling the stress values or temperature information of discrete grid points, grid centers, and their respective points in the finite element numerical simulation, the stress field and temperature field of the control system point cloud model are obtained. By setting the environment configuration of the finite element simulation and performing Latin hypercube sampling on the material parameters for several sets of point cloud data, stress field and temperature field data of the control system point cloud model used for training the point cloud convolutional neural network are obtained. By training the constructed point cloud neural network, a fast prediction surrogate model of the control system's gravitational field and temperature field is obtained. The input of the point cloud neural network is the material parameters of the control system, and the output is a point cloud model with gravitational or temperature information. Fully supervised training is performed using the point cloud data obtained by discrete sampling in finite element simulation to achieve the accuracy and consistency of the surrogate model prediction. The dropout layer inside the point cloud neural network also avoids overfitting of the surrogate model. The surrogate model is used to achieve fast prediction of the stress field and temperature field under various control system hardware materials. The stress information, temperature information, and material strength thresholds of each point in the point cloud model are used to obtain the reliability index of a single experiment. Monte Carlo simulation is used to conduct a large number of experiments to achieve rapid prediction of reliability.
2. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 1, characterized in that, Includes the following steps: Step 1: Construct a point cloud model of the control system components and sample the original stress and temperature data of the point cloud model; By utilizing the mesh discretization characteristics of finite element simulation, a large amount of stress and temperature field data can be obtained quickly and inexpensively. Step 2: Construct a rapid prediction proxy model for stress and temperature fields based on point cloud neural networks; The experimental data obtained in step one is used to train the point cloud convolutional neural network model to process discrete and ordered point cloud data, ensuring the consistency, accuracy and timeliness of stress field and temperature field predictions. Step 3: Reliability prediction of control systems based on surrogate models; The surrogate model for rapid prediction of stress and temperature fields based on the point cloud model trained in step two is used for inference. A large number of random input samples are generated by the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress or temperature field response. Finally, a consistent and efficient reliability assessment of the single-machine control system is obtained.
3. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 2, characterized in that, Step one involves constructing a point cloud model of the control system components and sampling the original stress and temperature data from the point cloud model, including the following sub-steps: Step 1.1: First, build a physical model of the control system unit in the finite element simulation software, and set the spatial coordinates of grid points and grid centers, as well as their initial stress and temperature properties. Step 1.2: By setting different material parameters, perform batch finite element simulations of the stress field and temperature field, and extract point cloud data for training. Step 1.3: Perform Latin hypercube sampling on the finite element simulation experiment; By extracting several sets of sample points, the finite element simulation software is further developed, and scripts are used to automatically read the material parameter values in the sample points and conduct batch experiments. All material parameters are randomly sampled within a set range using a uniform distribution to enable the import of different material parameters; Several sets of finite element analysis results for temperature fields and stress fields corresponding to different parameters were obtained.
4. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 2, characterized in that, Step two involves constructing a fast prediction proxy model for stress and temperature fields based on point cloud neural networks, including the following sub-steps: Step 2.1: Use a linear fully-connected (FC) layer to map low-dimensional material information to a higher-dimensional feature space, extract the implicit relationship between material parameters and stress / temperature fields, and establish the coupling relationship between parameters nonlinearly through the ReLU activation function; The training data for the model comes from the point cloud data obtained by batch finite element simulation sampling in step one. The input is the material parameters of the simulated control system. The material parameters determine the response law of the final 3D stress / temperature field. Simple direct coupling of material features will result in insufficient feature extraction. FC dimension upscaling is used, as shown in equation (1): （1） in, For the first Material parameter characteristics of this experiment Features after FC dimensionality upgrade; Step 2.2: Extend high-dimensional material features to the point cloud model; Point cloud data, as an unordered collection of points in three-dimensional space, has randomness and rotation issues. The order in which the points are stored does not affect the spatial position or the geometric relationship of the model. However, the same point cloud model can be represented by completely different storage matrices, and the weights trained by the neural network are also completely different. Multi-perceptron (MLP) based on full-joint (FC) does not possess rotation invariance and requires fixed-dimensional input. Its sensitivity to the order of point cloud data storage will severely affect the generalization ability and robustness of MLP. By introducing a 3D stress / temperature field, the positional relationships of point clouds are directly represented relative to each other in the 3D storage matrix, ensuring that the order of data points stored in the point cloud model does not affect the accuracy of the 3D set feature representation. The multidimensional features of material parameters are extracted using FC and extended to a 3D stress / temperature field, as shown in Equation (2): （2） in, For a 3D stress / temperature field, its dimensions are... This provides information on the length (depth), width, and height of the temperature / stress field in space. Step 2.3: Aggregate high-dimensional spatial features to obtain the physical parameters of each voxel in the stress / temperature field; A 3D convolutional neural network is used to aggregate spatial features of the stress / temperature field. In this case, the input of the convolutional neural network is an FC-encoded array that is expanded into a feature tensor of a three-dimensional field. ; By sliding the convolution kernel in the depth, height, and width directions respectively, spatial features of the local neighborhood are extracted. The convolution process is shown in equation (3): （3） in, The 3D convolution kernel is set; The training process of the model is realized by optimizing the parameters of the convolution kernel. The feature aggregation method can combine the coupling relationship between material parameters and structural spatial position, adaptively learn the transmission path and local change law of stress / temperature in three-dimensional structure, and realize accurate modeling of high temperature hot spot area and potential failure location inside the control system. The features obtained after traversing the last convolutional kernel are the stress / temperature of the corresponding voxel; The number of point clouds obtained by random mesh sampling in finite element simulation is not regular, i.e. The traversal region and convolution kernel of the 3D convolutional neural network must be a regular cubic structure. Point cloud mask constraints are used to retain only the effective sampling region, avoiding errors caused by regular voxel interpolation and improving the consistency between the prediction results and the actual physical structure.
5. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 2, characterized in that, Step three, the reliability prediction of the control system based on the surrogate model, specifically includes the following sub-steps: Step 3.1: Construct the statistical distribution of the uncertain input variables; Uncertainty input variables include: material properties, environmental loading parameters, manufacturing errors, and operating conditions; Statistical distributions include: Gaussian distribution, uniform distribution, or engineering experience distribution; By using surrogate models to replace traditional finite element simulation or accelerated testing, stress / temperature response prediction under different material parameters and working conditions can be performed in multiple scenarios with extremely low computational cost, enabling rapid analysis of the reliability indicators of the control system. Step 3.2: Determine the reliability of the control system; Based on step 3.1, a large number of random input samples are generated using the Monte Carlo sampling method and input into the deep learning surrogate model to obtain the corresponding stress field or temperature field response; For a given stress failure criterion, the following reliability criterion is adopted: when the predicted stress at any point of the structure... Not exceeding the material strength threshold If the sample is functioning normally, it is considered to be in operation; otherwise, it is considered to be in failure. The reliability index of the control system is shown in equation (4): （4） That is, as shown in equation (5): （5） in, This represents the number of samples that satisfy the failure criterion in the Monte Carlo simulation. This represents the total number of samples; By using fast surrogate model inference, stress response analysis of a single sample can be completed in milliseconds, enabling Monte Carlo simulation to be performed with extremely high efficiency. This solves the problem that traditional control system reliability analysis cannot perform large-sample statistics due to the long simulation analysis time.
6. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 1, characterized in that, Step four involves applying the reliability prediction results obtained in step three to the online operation and maintenance of the control system. Real-time environmental data and performance parameters from the actual operation of the control system are input into the surrogate model trained in step two, which outputs a degradation trajectory with a confidence interval for future time periods. When the predicted degradation trajectory approaches the preset failure threshold, an early warning signal is automatically triggered and a maintenance command is generated. By utilizing the millisecond-level inference capability of the surrogate model, the problem of resource waste and high risk of post-maintenance in traditional periodic maintenance is solved, realizing the transformation from "planned maintenance" to "predictive maintenance" and ensuring the safe operation of the control system under random operating conditions.
7. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 3, characterized in that, In step 1.1, a sparser mesh is selected in the parts with fewer components, and a denser mesh is selected in the parts with more components.
8. The single-machine reliability assessment method for a control system based on a convolutional neural network as described in claim 3, characterized in that: In step 1.2, only material parameters that have a significant impact on the actual lifespan of the control system and the finite element simulation results are considered.