A Single-Machine Life Assessment Method for Control Systems Based on Physical Data Coupling

Abstract

This invention discloses a method for single-machine life assessment of control systems based on physical data coupling, belonging to the field of control system single-machine life prediction and health management technology. This invention applies PINN technology, combined with uncertainty quantification and multi-objective optimization strategies, to achieve deep synergy between physical constraints and data-driven approaches. Within a framework of deep learning and the deep integration of physical laws, a physical information neural network is introduced to simultaneously handle forward and inverse problems and organically combine data-driven approaches with physical constraints. By integrating key dynamic characteristics of the control system from reliability test data, historical operating data, and numerical simulation data, a predictive model for life assessment using a deep learning method is established, enabling control system life assessment. This invention achieves high-precision life prediction and reliability certification even under small sample conditions, quickly and cost-effectively assessing product life in complex and harsh environments, intuitively indicating risks, and guiding the formulation of operation and maintenance strategies.

Description

Technical Field

[0001] This invention relates to a method for assessing the lifespan of a single control system based on physical data coupling, belonging to the field of control system single-machine lifespan prediction and health management technology. Background Technology

[0002] As the "nerve center" of equipment operation, the control system directly determines combat effectiveness, production efficiency, and operational safety. Taking typical equipment such as spacecraft, industrial robots, and wind power generation systems as examples, the control system needs to operate for extended periods in extreme environments such as high temperature, high humidity, salt spray, and strong electromagnetic interference, while simultaneously enduring multiple stress coupling effects such as mechanical vibration, fatigue, and creep. This characteristic of "long-term storage and instantaneous operation" leads to a non-linear degradation trend in the performance of the control system over time. Traditional life assessment methods based on failure time, which rely on large sample statistical data and ignore physical degradation mechanisms, are no longer sufficient to meet the engineering requirements of high precision and high reliability. For example, after long-term storage, the drift of electronic component parameters or the accumulation of wear on mechanical parts in the control system of an aero-engine may cause sudden failures. Traditional methods, unable to capture the gradual degradation process, often lead to premature or delayed maintenance, increasing the total life cycle cost.

[0003] The limitations of traditional life assessment methods stem from the separation between data-driven and mechanistic modeling: while pure data-driven models can capture nonlinear characteristics under complex operating conditions, they lack the guidance of physical constraints and are prone to losing physical rationality under data sparsity or noise interference, resulting in prediction results that violate basic physical laws; while pure mechanistic models can reveal the physical essence of the degradation process, the complexity of the analysis increases exponentially when faced with high-dimensional nonlinear partial differential equation systems, and they rely on a large number of prior parameters, making it difficult to achieve accurate modeling under complex operating conditions.

[0004] Physical Information Neural Networks (PINNs) embed physical laws into the neural network training process as differential constraints, constructing a hybrid loss function that includes data fitting terms and physical residual terms. This achieves an organic unity between data-driven learning and physical law constraints. Embedding physical constraints reduces the model's dependence on labeled data, and the prior knowledge provided by the physical model can serve as a substitute for "unlabeled data," guiding the network to make reasonable extrapolations in sparse data regions. The optimization process of physical residuals ensures that the output conforms to known physical laws, giving the model a natural physical property of conservation, invariance, and symmetry. The prediction results are simultaneously constrained by both data and physical laws, enhancing the model's interpretability and facilitating the identification of failure causes and optimization of data acquisition schemes.

[0005] In engineering applications, PINN requires customized physical constraints tailored to the characteristics of control systems. For example, for aero-engine control systems, residual constraints are embedded from the thermal aging equations of electronic components and the fatigue crack propagation equations of mechanical parts to capture the electromechanical coupling degradation mechanism. For wind power generation systems, residual constraints from the wind speed-power coupling model are introduced to accurately predict the fatigue load distribution of the gearbox and generator under wind speed fluctuations. Through joint optimization of multi-physics residuals, PINN can predict the failure time 20%-30% earlier than a single data-driven model, providing a scientific basis for equipment maintenance decisions.

[0006] The data- and model-driven life assessment method integrates PINN theory and deep learning technology to achieve a two-way enhancement of "mechanism guiding data and data correcting mechanism". However, it still faces challenges such as insufficient depth of integration between physical mechanism and data and poor adaptability across operating conditions. It is urgent to break through technical barriers through interdisciplinary collaborative innovation to provide theoretical and methodological support for the reliable operation and industrial upgrading of high-end equipment in my country. Summary of the Invention

[0007] To address the challenges of low accuracy and insufficient fusion of physical mechanisms and data under complex operating conditions and small sample sizes, traditional methods aim to provide a single-machine life assessment method for control systems based on physical data coupling. This method introduces Physical Information Neural Network (PINN) technology and constructs a high-precision, highly adaptable life assessment model by fusing the physical mechanisms of control system degradation with multi-source data characteristics, thus providing a scientific basis for equipment maintenance decisions.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] This invention discloses a single-machine life assessment method for control systems based on physical data coupling. It applies PINN technology, combines uncertainty quantification and multi-objective optimization strategies to achieve deep synergy between physical constraints and data-driven approaches. Within a framework of deep integration of deep learning and physical laws, it introduces a physical information neural network to simultaneously handle forward and inverse problems and organically combine data-driven approaches with physical constraints. It integrates key dynamic characteristics of the control system from reliability test data, historical operating data, and numerical simulation data to establish a prediction method for life assessment using deep learning. The method includes the following steps:

[0010] Step 1: Construct a physical data-driven life assessment framework and establish a mathematical model of the control system degradation process;

[0011] This paper proposes a neural network model framework that can learn from data while being constrained by physical laws. This framework refers to a deep learning architecture that embeds physical laws into neural network training in the form of differential constraints. By constructing a hybrid loss function that includes data fitting terms and physical residual terms, it achieves a unity of physical laws and data-driven approaches. This ensures that the neural network maintains its powerful function approximation capability while guaranteeing that the prediction results conform to physical laws. It is based on a deep fusion of function approximation theory and numerical solutions of differential equations to facilitate subsequent coupled driving processes. Specifically, it includes the following sub-steps:

[0012] Step 1.1: Establish a physically guided degradation process model;

[0013] A partial differential equation model is used to describe the degradation process of key performance parameters of the control system.

[0014] In the time domain Under the given conditions, the degradation process model of key performance parameters is shown in equation (1):

[0015] (1)

[0016] Where V is voltage and t is time. The physical degradation mechanism is represented by F, which is a key performance parameter of the control system. For the physical parameters to be determined, The random fluctuation term represents the noise that the degradation process model did not capture;

[0017] As a preferred approach, physical degradation mechanisms include Ohm's law and material aging effects, and the physical parameters to be determined include activation energy and degradation rate.

[0018] Establish a mapping model using a neural network. The relationship shown in equation (2) is satisfied:

[0019] (2)

[0020] in, The voltage function with time as the independent variable. Here, η represents the neural network parameters, and η represents the observation noise.

[0021] Step 1.2: Construct a multilayer perceptron as a function approximator;

[0022] Build a basic neural network to learn the complex mapping relationship between performance parameters and time directly from data;

[0023] The basic neural network adopts a three-layer fully connected structure: the input layer is time t, and the Tanh activation function is used to ensure gradient smoothness; the output layer is the basic predicted value of the performance parameters.

[0024] The basic neural network, as the core function approximator, is responsible for capturing the nonlinear degradation trend presented in the data;

[0025] The architecture of the multilayer perceptron is shown in equation (3):

[0026] (3)

[0027] in, The basis functions and weight matrix are... Bias vector , For the hidden layer dimension;

[0028] Step 2: Embed physical constraints as residual terms and collaboratively optimize environmental parameters;

[0029] Environmental stress parameters are embedded as training variables into the basic neural network structure constructed in step 1.2. Through the deep integration of the physical model and the neural network, the synergistic driving force of physical mechanisms and data-driven approaches is achieved. Specifically, this includes the following sub-steps:

[0030] Step 2.1: Construct the physical information loss function;

[0031] By constructing a physical information loss function, the residuals of the physical model are incorporated as a constraint term into the overall loss function Loss, as shown in equation (4):

[0032] (4)

[0033] in, The mean square error between the network's predicted values ​​and the actual observed data. For physical equations Calculate the sum of the root squares of the residuals at each point on the graph. Weighting coefficients to balance the two losses;

[0034] By minimizing the overall loss function, the neural network is forced to satisfy physical laws while fitting the data;

[0035] Step 2.2: Embed and adaptively optimize environmental stress parameters;

[0036] Environmental stress is embedded into the model as a training variable.

[0037] To ensure its physical rationality ( This also allows the optimization process to be performed in logarithmic space, improving training stability, and preferentially applying a logarithmic transformation to environmental stress: ,in , These are the parameters for network learning.

[0038] The logarithmically transformed environmental stress and neural network weights are optimized together through backpropagation to achieve adaptive identification of environmental parameters without the need for manual calibration.

[0039] Step 2.3: Forward propagation of multiphysics coupling;

[0040] Based on the stress acceleration model in reliability physics, the prediction results are directly affected by the physical influence of environmental stress and time. In the final output of the neural network, the influence of environmental stress and time is explicitly coupled. The physical coupling of the forward propagation process adopts a product form, as shown in Equation (5):

[0041] (5)

[0042] in, The temperature acceleration factor is used, and a linear approximation of the Arrhenius model is employed. The humidity acceleration factor is used, employing the Peck model or its linear approximation. Let be the time degradation function, expressed in exponential form as exp(λt);

[0043] Step 3: Multi-stage collaborative training for small sample data;

[0044] The physical-data neural network model constructed in steps one and two is trained to ensure that the physical-data neural network model is stable and converges efficiently with a small amount of data, and to enhance the generalization ability of the physical-data neural network model.

[0045] In view of the characteristics of few-shot learning and the special nature of physical constraints, the entire training process based on few-shot data adopts a phased training strategy, which includes the following sub-steps:

[0046] Step 3.1: In the early stage of model training, a high initial learning rate is selected and the Adam optimizer is used to quickly bring the neural network parameters and physical parameters into the neighborhood of the global optimum. The focus is on the initial calibration of the physical parameters and quickly adjusting them to a reasonable order of magnitude so that the model can converge quickly in the early stage of training.

[0047] Step 3.2: After the model converges quickly, reduce the learning rate and perform a fine-grained search;

[0048] Differentiated learning rates are applied to physical parameters and network parameters, with a preferred learning rate for physical parameters to maintain the stability of physical meaning. The synergistic relationship between neural network weights and physical parameters is carefully adjusted to ensure that the model can accurately fit the data and strictly conform to physical laws.

[0049] Step 3.3: After the learning rate reduction and stabilization tuning process, gradient clipping and weight decay techniques are introduced to prevent gradient explosion and suppress overfitting.

[0050] By employing an early stopping mechanism and rolling back to the optimal parameters, the model parameters converge to a "flat" minimum region, enhancing the ability to generalize to unknown data.

[0051] Step 4: Uncertainty Prediction and Lifetime Assessment;

[0052] Based on the physical-data neural network trained in steps one, two, and three, the test data of various physical parameters of the control system's individual units are input into the physical-data neural network for uncertainty prediction and life assessment. This process includes the following sub-steps:

[0053] Step 4.1: Perform Monte Carlo Dropout sampling to quantify uncertainty;

[0054] During the prediction phase, the Dropout mechanism is kept on, and random forward propagation is performed at each test time point. Each propagation yields slightly different prediction values ​​due to Dropout, which characterizes the result distribution caused by the uncertainty of model parameters.

[0055] Step 4.2: Differentiate and visualize uncertainty;

[0056] Cognitive uncertainty is caused by the uncertainty of the model's own parameters. Through the above sampling variance assessment, this value will increase in sparse data regions. Random uncertainty is the noise inherent in the data, which cannot be reduced by the model. The final prediction result is the mean curve superimposed with a 95% confidence interval. When the data is sparse or long-term extrapolation prediction is performed, the confidence interval will automatically widen, intuitively indicating an increase in prediction risk.

[0057] Step 4.3: Remaining useful life assessment;

[0058] A failure threshold is set, and the predicted degradation trajectory is extrapolated on the time axis. The time point that intersects with the failure threshold is the predicted failure time. The difference between the current time and the predicted failure time is the remaining service life of the single device. The assessment result of the remaining service life has an uncertainty range.

[0059] Step 4.4: Model Validation;

[0060] To verify the accuracy of the lifespan prediction results, a comparative verification method was adopted to verify the accuracy. The first 60% of the measured data was selected as the training data, and the last 40% was selected as the comparison data to verify the accuracy of the model. The accuracy verification method was to calculate the difference between each corresponding pair of the original 40% of the data and the predicted last 40% of the data, and then take the average of all the differences to obtain the proportional error. The accuracy was then calculated based on the proportional error.

[0061] It also includes step five: For the spacecraft controller, through the uncertainty prediction and life assessment in step four, a prediction curve with a confidence interval is output, which can intuitively visualize the predicted "risk". The assessment results are directly integrated into the operation and maintenance management platform to help decision-makers understand the reliability of the prediction. In the stage with a wide confidence interval, it is recommended to increase the detection frequency or prepare a backup plan.

[0062] Beneficial effects:

[0063] 1. The present invention provides a single-machine life assessment method for control systems based on physical data coupling. By introducing a physical information neural network (PINN) to construct a hybrid driving framework, the physical degradation mechanism is embedded in the neural network loss function in the form of differential constraints. An assessment model with bidirectional enhancement of mechanism and data is established. Physical laws are used as "implicit data" to provide prior constraints, reducing the dependence on a large number of failure samples. High-precision life prediction and reliability certification can still be achieved under small sample conditions.

[0064] 2. The present invention provides a single-machine life assessment method for control systems based on physical data coupling. By explicitly coupling an environmental stress acceleration model in a neural network and using environmental stress parameters as training variables for adaptive joint optimization, key physical degradation parameters can be automatically and efficiently identified from the data without manual calibration. This improves the efficiency of accelerated life testing and the accuracy of parameter identification, enabling rapid and low-cost assessment of product life under complex and harsh environments.

[0065] 3. The present invention provides a single-machine life assessment method for control systems based on physical data coupling. Through a multi-stage collaborative training strategy for small samples, the model is trained using a hybrid loss function that includes data terms and physical residual terms. This ensures stable convergence of the model under limited data and outputs the mean value containing the predicted degradation trajectory. By combining Monte Carlo Dropout sampling, cognitive uncertainty and accidental uncertainty are quantified and distinguished. The prediction results are presented in the form of confidence intervals. Based on the quantitative assessment of risk, the confidence interval automatically widens when the data is sparse or long-term extrapolation is performed, so as to intuitively indicate the risk and guide the formulation of operation and maintenance strategies.

[0066] 4. The present invention provides a method for assessing the lifespan of a single control system based on physical data coupling. By solidifying and deploying a trained neural network model that combines data learning capabilities with physical constraint guarantees, a high-fidelity, interpretable digital twin of the single control system can be constructed. The digital twin can simulate and predict the lifespan evolution of the equipment under different operating conditions and usage strategies in virtual space, which can be used for scheme comparison and optimization, providing a reliable simulation verification platform for system-level task planning, health management, and optimization design.

[0067] 5. The present invention provides a single-machine life assessment method for control systems based on physical data coupling. By integrating a model framework with physical constraints and a training strategy, it achieves accurate modeling of the degradation process, providing early warning of potential faults 20%-30% earlier than a single data-driven model, allowing sufficient time for proactive intervention. By integrating the predicted remaining useful life (RUL) and its uncertainty range with the operation and maintenance system, it can automatically trigger early warnings and generate optimized condition-based maintenance plans, avoiding unplanned downtime. This elevates the operation and maintenance mode from reactive and periodic maintenance to precise predictive maintenance, improving the operational safety and life-cycle economic benefits of high-value control systems. Attached Figure Description

[0068] Figure 1 This forms the overall technical framework of the present invention;

[0069] Figure 2 This refers to the predicted single-machine life assessment result of the control system in step four of the present invention.

[0070] Figure 3 This refers to the accuracy of the single-machine life assessment of the control system predicted in step five of this invention. Detailed Implementation

[0071] 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.

[0072] Example 1

[0073] This embodiment includes voltage-time data of a single engine control unit of a certain type of aero-engine control system. The voltage data of the single engine was obtained by monitoring over a period of more than 3 months at 5-minute intervals, totaling 20,000 data points. Due to the large number of samples, only the initial and final data are shown, as shown in Table 1 and Table 2 respectively.

[0074] Table 1 Initial Voltage-Time Data for Single-Unit Generator-Control Unit Combination of Control System

[0075]

[0076] Table 2. End-of-term voltage-time data for the single-unit generator-control combination of the control system.

[0077] Under the above conditions, the life assessment method for a single-machine control system based on physical data coupling driven according to the present invention is applied to realize the life assessment of the control system, such as... Figure 1 As shown, it includes the following steps:

[0078] Step 1: Construct a physical data-driven life assessment framework and establish a mathematical model of the control system degradation process;

[0079] This paper proposes a neural network model framework that can learn from data while being constrained by physical laws. This framework refers to a deep learning architecture that embeds physical laws into neural network training in the form of differential constraints. By constructing a hybrid loss function that includes data fitting terms and physical residual terms, it achieves a unity of physical laws and data-driven approaches. This ensures that the neural network maintains its powerful function approximation capability while guaranteeing that the prediction results conform to physical laws. It is based on a deep fusion of function approximation theory and numerical solutions of differential equations to facilitate subsequent coupled driving processes. Specifically, it includes the following sub-steps:

[0080] Step 1.1: Establish a physically guided degradation process model;

[0081] Based on circuit theory and component aging theory, a partial differential equation model is used to describe the degradation process of key performance parameters of the control system.

[0082] In the time domain Under these conditions, the voltage degradation process model of the generator control unit is shown in equation (1):

[0083] (1)

[0084] Where V is voltage and t is time. The physical degradation mechanism is represented by V, which is the combined voltage of the single generator and controller in the control system. For the physical parameters to be determined, The random fluctuation term represents the noise that the degradation process model did not capture;

[0085] Establish a mapping model using a neural network. The relationship shown in equation (2) is satisfied:

[0086] (2)

[0087] in, Here, η represents the neural network parameters, and η represents the observation noise.

[0088] Step 1.2: Construct a multilayer perceptron as a function approximator;

[0089] Build a basic neural network to learn the complex mapping relationship between performance parameters and time directly from data;

[0090] The basic neural network adopts a three-layer fully connected structure: the input layer is time t, and the Tanh activation function is used to ensure gradient smoothness; the output layer is the basic predicted value of the performance parameters.

[0091] The basic neural network, as the core function approximator, is responsible for capturing the nonlinear degradation trend presented in the data;

[0092] The architecture of the multilayer perceptron (MLP) is shown in equation (3):

[0093] (3)

[0094] in, The basis functions and weight matrix are... Bias vector ;

[0095] The hidden layer has two layers, each with 32 neurons. The hidden layer dimension is... =32;

[0096] Step 2: Embed physical constraints as residual terms and collaboratively optimize environmental parameters;

[0097] Environmental stress parameters are embedded as training variables into the basic neural network structure constructed in step 1.2. Through the deep integration of the physical model and the neural network, the synergistic driving force of physical mechanisms and data-driven approaches is achieved. Specifically, this includes the following sub-steps:

[0098] Step 2.1: Construct the physical information loss function;

[0099] By constructing a physical information loss function, the residuals of the physical model are incorporated as a constraint term into the overall loss function Loss, as shown in equation (4):

[0100] (4)

[0101] in, Predicting voltage for the model Compared with the measured voltage in Table 1 The mean square error, Let be the mean square error of the residuals of the physical equation, where Obtained through automatic differentiation The physical loss weight is set to 0.5 in this example;

[0102] By minimizing the overall loss function, the neural network is forced to satisfy physical laws while fitting the data;

[0103] Step 2.2: Embed and adaptively optimize environmental stress parameters;

[0104] Environmental stress is embedded into the model as a training variable.

[0105] In this embodiment, assuming the storage environment temperature of the single unit is 55°C and the humidity is 40%RH, the temperature T and humidity H are parameterized as trainable variables, and a logarithmic transformation is performed on the environmental stress: ,in , These are the network learning parameters, initialized to... ;

[0106] The logarithmically transformed environmental stress and neural network weights are optimized together through backpropagation to achieve adaptive identification of environmental parameters without the need for manual calibration.

[0107] Step 2.3: Forward propagation of multiphysics coupling;

[0108] Based on the stress acceleration model in reliability physics, the prediction results are directly affected by the physical influence of environmental stress and time. In the final output of the neural network, the influence of environmental stress and time is explicitly coupled. The physical coupling of the forward propagation process adopts a product form, as shown in Equation (5):

[0109] (5)

[0110] in, =0.01, =0.5, These are trainable variables shared with the physical parameters in step 1.1;

[0111] Step 3: Multi-stage collaborative training for small sample data;

[0112] The physical-data neural network model constructed in steps one and two is trained to ensure that the physical-data neural network model is stable and converges efficiently with a small amount of data, and to enhance the generalization ability of the physical-data neural network model.

[0113] The 20,000 data points were split in a 6:4 ratio, with the first 12,000 data points used for training and the last 8,000 data points used for validation.

[0114] In view of the characteristics of few-shot learning and the special nature of physical constraints, the entire training process based on few-shot data adopts a phased training strategy, which includes the following sub-steps:

[0115] Step 3.1: Fast convergence period;

[0116] In this embodiment, the total number of rounds is set to 1000. In the first 300 rounds of the initial training of the model, a relatively high initial learning rate of 0.001 is selected, and the Adam optimizer is used to make the neural network parameters and physical parameters quickly enter the neighborhood of the global optimum. The focus is on the initial calibration of the physical parameters, which are quickly adjusted to a reasonable order of magnitude so that the model converges quickly in the early stage of training. After this stage, the loss function drops rapidly from the initial value of ~0.1 to ~0.01, and the physical parameter "λ" is initially calibrated to a positive value, indicating that the model has captured the downward trend of voltage.

[0117] Step 3.2: Fine-tuning period;

[0118] After the model converges quickly, from the 301st to the 800th round, the learning rate is reduced to 0.0001. Differential learning rates are used for the physical parameters and network parameters to finely adjust the physical parameters in order to maintain the stability of the physical meaning. The synergistic relationship between the neural network weights and the physical parameters is carefully adjusted to ensure that the model can accurately fit the data and strictly conform to the physical laws.

[0119] During this phase, the validation set loss continues to decrease, indicating that the model is not overfitting and has good physical consistency.

[0120] Step 3.3: Stabilization period;

[0121] After a period of stabilization and optimization by reducing the learning rate, gradient clipping with a threshold of 0.1 and weight decay techniques are introduced from round 801 to round 1000 to prevent gradient explosion and suppress overfitting.

[0122] An early stopping mechanism is adopted, with patience set to 100 rounds. If the validation set loss does not decrease for 100 consecutive rounds, training is stopped and the model is rolled back to the optimal parameters, so that the model parameters converge to a "flat" minimum region, thereby enhancing the ability to generalize to unknown data.

[0123] In this example, training is terminated in round 920 due to the early stopping mechanism and restored to the optimal parameters in round 820. At this point, the losses of both the training set and the validation set converge to a low level (~0.001).

[0124] Step 4: Uncertainty Prediction and Lifetime Assessment;

[0125] Based on the physical-data neural network trained in steps one, two, and three, the test data of various physical parameters of the control system's individual units are input into the physical-data neural network for uncertainty prediction and life assessment. This process includes the following sub-steps:

[0126] Step 4.1: Perform Monte Carlo Dropout sampling to quantify uncertainty;

[0127] During the prediction phase, the Dropout mechanism is kept on, and K=100 random forward propagations are performed for each test time point. Each propagation yields a slightly different prediction value due to Dropout, which characterizes the result distribution caused by the uncertainty of model parameters.

[0128] In the example, the voltage is predicted after t=10 hours, resulting in 100 slightly different predicted values, forming a distribution;

[0129] Step 4.2: Differentiate and visualize uncertainty;

[0130] Cognitive uncertainty is caused by the uncertainty of the model's own parameters. Through the above sampling variance assessment, this value will increase in sparse data regions. Random uncertainty is the noise inherent in the data, which cannot be reduced by the model. The final prediction result is the mean curve superimposed with a 95% confidence interval. When the data is sparse or long-term extrapolation prediction is performed, the confidence interval will automatically widen, intuitively indicating an increase in prediction risk.

[0131] Step 4.3: Remaining useful life assessment;

[0132] A failure threshold is set, and the predicted degradation trajectory is extrapolated on the time axis. The time point that intersects with the failure threshold is the predicted failure time. The difference between the current time and the predicted failure time is the remaining service life of the single device. The assessment result of the remaining service life has an uncertainty range.

[0133] In the example, the prediction results are as follows Figure 2 As shown, with time on the horizontal axis and health (or normalized value of key performance parameters) on the vertical axis, the discrete "observation" points represent historical monitoring data, the smooth "predicted mean" curve represents the model's best estimate of the performance degradation trend, the "95% confidence interval" shaded band surrounding the predicted mean curve, the width of which intuitively reflects the uncertainty of the model's prediction, and the horizontal "failure threshold (baseline)" is used to determine failure.

[0134] As time goes on, the predicted mean curve smoothly declines and eventually crosses the failure threshold, while the confidence interval is narrower in the data-dense region and gradually widens in the data-sparse future prediction region (right side).

[0135] The initial average voltage is 28.922V, and 90% of this is taken as the threshold, i.e., V_threshold=26.030V;

[0136] Extrapolation prediction: Extrapolate the trained model to the next 60 years (525,600 hours).

[0137] Result determination: The predicted average voltage curve will intersect with the failure threshold after t=21.20 years;

[0138] Remaining lifetime report: The predicted remaining lifetime of this power control combination at the current moment is 21.20 years, with a 95% confidence interval of [19.99 years, 23.41 years];

[0139] Step 4.4: Model Validation;

[0140] To verify the accuracy of the lifetime prediction results, a comparative verification method was adopted. The last 8000 verification data points were used for validation. The model's predicted values ​​for the last 27 time points were compared one by one with the corresponding actual measured values ​​in Table 1. The prediction accuracy is as follows: Figure 3 As shown, the mean absolute percentage error (MAPE) is 0.41%, and the prediction accuracy is 99.59%. The method of this invention achieves high-precision prediction under the condition of a small sample size of 3 months.

[0141] It also includes step five: For the spacecraft controller, through the uncertainty prediction and life assessment in step four, a prediction curve with a confidence interval is output, which can intuitively visualize the predicted "risk". The assessment results are directly integrated into the operation and maintenance management platform to help decision-makers understand the reliability of the prediction. In the stage with a wide confidence interval, it is recommended to increase the detection frequency or prepare a backup plan.

[0142] 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 lifespan of a single control system based on physical data coupling, characterized in that, By applying PINN technology and combining uncertainty quantification and multi-objective optimization strategies, a deep synergy between physical constraints and data-driven approaches is achieved. Within the framework of deep learning and the deep integration of physical laws, a physical information neural network is introduced to simultaneously process forward and inverse problems and organically combine data-driven approaches with physical constraints. By integrating control system reliability test data, historical operating data, and numerical simulation data, a prediction method based on deep learning is established for life assessment of the control system.

2. The method for single-machine life assessment of a control system based on physical data coupling as described in claim 1, characterized in that, Includes the following steps: Step 1: Construct a physical data-driven life assessment framework and establish a mathematical model of the control system degradation process; This paper proposes a neural network model framework that can learn from data and is constrained by physical laws. This framework refers to a deep learning architecture that embeds physical laws into neural network training in the form of differential constraints. By constructing a hybrid loss function that includes data fitting terms and physical residual terms, the framework achieves the unity of physical laws and data-driven approaches. This ensures that the neural network maintains its powerful function approximation ability while ensuring that the prediction results conform to physical laws. The framework is based on the deep integration of function approximation theory and numerical solutions of differential equations to facilitate subsequent coupling and driving processes. Step 2: Embed physical constraints as residual terms and collaboratively optimize environmental parameters; By embedding environmental stress parameters as training variables into the basic neural network structure constructed in step 1.2, the synergistic driving force of physical mechanism and data-driven approach is achieved through the deep integration of physical model and neural network. Step 3: Multi-stage collaborative training for small sample data; The physical-data neural network model constructed in steps one and two is trained to ensure that the physical-data neural network model is stable and converges efficiently with a small amount of data, and to enhance the generalization ability of the physical-data neural network model. In view of the characteristics of few-shot learning and the special physical constraints, the entire training process based on few-shot data adopts a phased training strategy. Step 4: Uncertainty Prediction and Lifetime Assessment; Based on the physical-data neural network trained in steps one, two, and three, the test data of various physical parameters of the control system unit are input into the physical-data neural network for uncertainty prediction and life assessment.

3. The method for single-machine life assessment of a control system based on physical data coupling as described in claim 1, characterized in that, Including step five: For spacecraft controllers, the uncertainty prediction and life assessment in step four outputs a prediction curve with a confidence interval, which can intuitively visualize the predicted "risk". The assessment results are directly integrated into the operation and maintenance management platform to help decision-makers understand the reliability of the prediction. In the stage with a wide confidence interval, it is recommended to increase the detection frequency or prepare a backup plan.

4. The method for assessing the single-machine lifespan of a control system based on physical data coupling as described in claim 2, characterized in that, Step one involves constructing a physical data-driven lifetime assessment framework and establishing a mathematical model of the control system degradation process, including the following sub-steps: Specifically, it includes the following sub-steps: Step 1.1: Establish a physically guided degradation process model; A partial differential equation model is used to describe the degradation process of key performance parameters of the control system. In the time domain Under the given conditions, the degradation process model of key performance parameters is shown in equation (1): （1） Where V is voltage and t is time. The physical degradation mechanism is represented by F, which is a key performance parameter of the control system. For the physical parameters to be determined, The random fluctuation term represents the noise that the degradation process model did not capture; Establish a mapping model using a neural network. The relationship shown in equation (2) is satisfied: （2） in, The voltage function with time as the independent variable. Here, η represents the neural network parameters, and η represents the observation noise. Step 1.2: Construct a multilayer perceptron as a function approximator; Build a basic neural network to learn the complex mapping relationship between performance parameters and time directly from data; The basic neural network adopts a three-layer fully connected structure: the input layer is time t, and the Tanh activation function is used to ensure gradient smoothness; the output layer is the basic predicted value of the performance parameters. The basic neural network, as the core function approximator, is responsible for capturing the nonlinear degradation trend presented in the data; The architecture of the multilayer perceptron is shown in equation (3): （3） in, The basis functions and weight matrix are... Bias vector , For the hidden layer dimension.

5. The method for assessing the single-machine lifespan of a control system based on physical data coupling as described in claim 2, characterized in that, In step two, physical constraints are embedded as residual terms and environmental parameters are collaboratively optimized, which specifically includes the following sub-steps; Step 2.1: Construct the physical information loss function; By constructing a physical information loss function, the residuals of the physical model are incorporated as a constraint term into the overall loss function Loss, as shown in equation (4): （4） in, The mean square error between the network's predicted values ​​and the actual observed data. For physical equations Calculate the sum of the root squares of the residuals at each point on the graph. Weighting coefficients to balance the two losses; By minimizing the overall loss function, the neural network is forced to satisfy physical laws while fitting the data; Step 2.2: Embed and adaptively optimize environmental stress parameters; Environmental stress is embedded into the model as a training variable. The logarithmically transformed environmental stress and neural network weights are optimized together through backpropagation to achieve adaptive identification of environmental parameters without the need for manual calibration. Step 2.3: Forward propagation of multiphysics coupling; Based on the stress acceleration model in reliability physics, the prediction results are directly affected by the physical influence of environmental stress and time. In the final output of the neural network, the influence of environmental stress and time is explicitly coupled. The physical coupling of the forward propagation process adopts a product form, as shown in Equation (5): （5） in, The temperature acceleration factor is used, and a linear approximation of the Arrhenius model is employed. The humidity acceleration factor is used, employing the Peck model or its linear approximation. The time degradation function is expressed in exponential form as exp(λt).

6. The method for single-machine life assessment of a control system based on physical data coupling as described in claim 2, characterized in that, Step three, the multi-stage collaborative training for small sample data, includes the following sub-steps: Step 3.1: In the early stage of model training, a high initial learning rate is selected and the Adam optimizer is used to quickly bring the neural network parameters and physical parameters into the neighborhood of the global optimum. The focus is on the initial calibration of the physical parameters and quickly adjusting them to a reasonable order of magnitude so that the model can converge quickly in the early stage of training. Step 3.2: After the model converges quickly, reduce the learning rate and perform a fine-grained search; Differentiated learning rates are applied to physical parameters and network parameters, with a preferred learning rate for physical parameters to maintain the stability of physical meaning. The synergistic relationship between neural network weights and physical parameters is carefully adjusted to ensure that the model can accurately fit the data and strictly conform to physical laws. Step 3.3: After the learning rate reduction and stabilization tuning process, gradient clipping and weight decay techniques are introduced to prevent gradient explosion and suppress overfitting. By employing an early stopping mechanism and rolling back to the optimal parameters, the model parameters converge to a "flat" minimum region, enhancing the ability to generalize to unknown data.

7. The method for assessing the single-machine lifespan of a control system based on physical data coupling as described in claim 2, characterized in that, Step four, uncertainty prediction and lifetime assessment, specifically includes the following sub-steps: Step 4.1: Perform Monte Carlo Dropout sampling to quantify uncertainty; During the prediction phase, the Dropout mechanism is kept on, and random forward propagation is performed at each test time point. Each propagation yields slightly different prediction values ​​due to Dropout, which characterizes the result distribution caused by the uncertainty of model parameters. Step 4.2: Differentiate and visualize uncertainty; Cognitive uncertainty is caused by the uncertainty of the model's own parameters. Through the above sampling variance assessment, this value will increase in sparse data regions. Random uncertainty is the noise inherent in the data, which cannot be reduced by the model. The final prediction result is the mean curve superimposed with a 95% confidence interval. When the data is sparse or long-term extrapolation prediction is performed, the confidence interval will automatically widen, intuitively indicating an increase in prediction risk. Step 4.3: Remaining useful life assessment; A failure threshold is set, and the predicted degradation trajectory is extrapolated on the time axis. The time point that intersects with the failure threshold is the predicted failure time. The difference between the current time and the predicted failure time is the remaining service life of the single device. The assessment result of the remaining service life has an uncertainty range. Step 4.4: Model Validation; To verify the accuracy of the lifespan prediction results, a comparative verification method was adopted to verify the accuracy. The first 60% of the measured data was selected as the training data, and the last 40% was selected as the comparison data to verify the accuracy of the model. The accuracy verification method was to take the difference between each corresponding pair of the original 40% of the data and the predicted last 40% of the data, and then take the average of all the differences to obtain the proportional error. The accuracy was then calculated based on the proportional error.

8. The method for single-machine life assessment of a control system based on physical data coupling as described in claim 4, characterized in that: In step 1.1, the physical degradation mechanisms include Ohm's law and material aging effects, and the physical parameters to be determined include activation energy and degradation rate.

9. The method for evaluating the single-machine lifespan of a control system based on physical data coupling as described in claim 5, characterized in that: In step 2.2, to ensure its physical rationality ( This also allows the optimization process to take place in logarithmic space, improving training stability and performing a logarithmic transformation on environmental stress: ,in , These are the parameters for network learning.

Images