A soft measurement method and system for complex electromechanical systems based on the LAC-T model

Abstract

The present application relates to the technical field of state monitoring and health management of complex electromechanical systems, and particularly relates to a complex electromechanical system soft measurement method and system based on an LAC-T model; the method takes a physical mechanism mapping F as a constraint, combines a Lyapunov actor-critic-teacher (LAC-T) framework, realizes convergence and feasible region control of parameter estimation through a double-channel mechanism tracking error and a teacher reference item, and guarantees smoothness and anti-bias of an estimated trajectory through stability / maximum entropy target training, so as to output reliable parameter sequences and observation estimates, support subsequent state monitoring, control optimization and RUL and other PHM applications. The present application is suitable for complex electromechanical equipment in a multi-sensor telemetry scene of an aero-engine, a compressor, a rail vehicle, intelligent manufacturing line equipment and the like.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring and health management technology for complex electromechanical systems, and in particular to a soft measurement method and system for complex electromechanical systems based on the LAC-T model; applicable to complex electromechanical equipment in multi-sensor telemetry scenarios such as aero-engines, compressors, rail vehicles, and intelligent manufacturing production line equipment. Background Technology

[0002] Soft sensing aims to infer key state variables and intrinsic parameters that are difficult to measure directly from readily available process / telemetry data. In complex electromechanical systems (such as aero-engines, compressors, rail vehicles, and smart manufacturing production line equipment), it is often used to estimate "hidden parameters" that are invisible or cannot be directly measured by sensors, in order to support condition monitoring, process control, and health management. However, such objects generally have characteristics such as strong nonlinearity, strong coupling, variable operating conditions, asynchronous sampling, and significant noise interference, which makes the traditional soft sensing approach face obvious bottlenecks in engineering applications.

[0003] The existing technologies can be broadly classified into three categories: (1) Mechanism model / observer type: Based on prior mechanisms, state space models or mechanism mapping functions F are established, and Luenberger / EKF / UKF, sliding mode or unknown input observers are used for parameter and state estimation. This type of method has interpretability and controllability when the mechanism is sufficient and the disturbance can be modeled, but it is highly sensitive to model structure and parameter identification; when model mismatch, sensor noise and sudden changes in operating conditions coexist, estimation divergence, hysteresis and drift are likely to occur, making it difficult to work stably online for a long time. (2) Data-driven / statistical learning type: PLS / SVR / random forest, deep networks (such as LSTM, TCN, Transformer) are used to directly learn the mapping of "telemetry / target quantity" from historical data, which has the advantages of flexible deployment and friendliness to complex nonlinearity; however, it often relies on the assumption of approximately independent and identically distributed (iid) and consistent operating conditions, and is more sensitive to domain migration, distribution drift and missing measurement anomalies, and lacks physical consistency constraints, which easily gives estimates that exceed the feasible domain. (3) Hybrid modeling and adaptive identification: The simplified mechanism is coupled with the data model, and subspace identification, recursive least squares or Bayesian filtering are introduced for online updates, taking into account a certain degree of interpretability and adaptability; however, in the case of strong noise, strong time variation and multiple solutions (multiple parameter combinations produce approximately equivalent outputs), the problem of non-unique parameters and time discontinuity may still occur, which leads to the instability of soft measurement results in health assessment and control closed loop.

[0004] In recent years, some studies have attempted to introduce reinforcement learning (RL) to improve parameter inversion and adaptive capabilities. However, mainstream RL frameworks mostly focus on "optimal control policies." When directly applied to soft measurement (treating "actions" as parameter estimates), they face three prominent challenges: First, rewards constructed solely from "output tracking errors" are prone to propagation over time when observation estimation biases exist, causing parameter sequence jitter and discontinuities. Second, the exploration and high-variance gradient updates of RL may amplify the impact of noise, lacking stability and explicit constraints on the feasible region, making it difficult to ensure that online identification does not diverge. Third, while engineering often involves some physical priors or empirical intervals, existing solutions lack sufficient support for effectively injecting these "mentors / priors" into the learning process and guiding parameters to converge in a physically interpretable direction. Furthermore, in health management processes, although unsupervised state division and stage identification can be performed using time series models such as HMM / HSMM / HDP-HMM, if the intrinsic parameter trajectories obtained by soft measurement are unstable or driven by noise, the state boundaries will become unreliable, thus affecting subsequent threshold judgments, early warnings, and life assessments. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a soft measurement method for complex electromechanical systems based on the Lyapunov Actor-Critic-Mentor (LAC-T) model. This method can identify implicit degradation parameters that cannot be directly measured online, stably, and continuously under multi-condition noise environments. Using the physical mechanism mapping F as a constraint, and combining it with the Lyapunov Actor-Critic-Mentor (LAC-T) framework, the method achieves convergence and feasible region control of parameter estimation through dual-channel mechanism tracking error and mentor reference terms. Furthermore, stability / maximum entropy objective training ensures smooth estimation trajectory and robustness against bias, thereby outputting reliable parameter sequences and observation estimates. This supports subsequent state monitoring, control optimization, and PHM applications such as RUL (Restricted Utility Model). It is applicable to complex electromechanical equipment in multi-sensor telemetry scenarios such as aero-engines, compressors, rail vehicles, and intelligent manufacturing production line equipment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A soft measurement method for complex electromechanical systems using the LAC-T model, comprising the following steps:

[0008] S1) Construct the physical system mapping relationship F, and given the input vector and the previous observation estimate, output the observation vector estimate. ;

[0009] S2) Latent parameter identification is modeled as a decision-making process based on the LAC-T model framework, and the actor-environment-critic-mentor linkage mechanism is used to output the identified value of the latent degradation parameter at time t. ;

[0010] S3) tracking error The LAC-T model framework is trained to supervise the signal, so that the parameter trajectory remains continuous and stable when observation bias exists;

[0011] S4) Enter (y) during the online phase t u t Output the implicit degenerate parameter sequence Real-time identification is achieved, y t u represents the true value of the telemetry data vector at time t. t The input to the physical system of the complex electromechanical system at time t.

[0012] As a preferred option, given the input vector Under the condition of output observation vector estimation Where s and p represent the dimensions of the physical system input and observation vectors, respectively, and the identification parameter value θ t Reduce output observation vector estimation Compared with the true value of the observed vector y t+1 The error between them is described as shown in equations (1) and (2):

[0013] (1),

[0014] (2),

[0015] in, represent the input vector and observation vector estimates of the physical system at time t+1 and time t, respectively; F is the physical mechanism mapping between the system's internal state, input, and observation.

[0016] Alternatively, the tracking error is as shown in equation (3):

[0017] (3),

[0018] (4),

[0019] (5),

[0020] (6),

[0021] Where: e1 is the same as the observation vector estimate in equation (2), representing the tracking error based on the previous observation vector estimate; e2 represents the tracking error based on the previous observation vector true value; e3 represents the cumulative observation vector estimate error; ||.| is the vector Euclidean norm; T is the current end time of the sequence under consideration.

[0022] As a preferred option, the LAC-T model framework satisfies:

[0023] ,

[0024] ,

[0025] ,

[0026] ,

[0027] Where A, E, C, and T represent the mapping relationships within the actor, environment, critic, and mentor modules, respectively; t+1 Let y represent the input vector of the complex electromechanical system's physical system at time t+1. t+1 as well as This represents the true and estimated values ​​of the telemetry data vector of a complex electromechanical system at time t+1. It is the telemetry data vector estimate of the physical system at time t. The implicit degradation parameter identification value of the actor module at time t+1 is represented by r, the evaluation of the physical system telemetry data by the environment module is represented by v, and the estimated value of the implicit degradation parameter by the critic module is represented by v. Evaluation, These are implicit degradation parameter reference values ​​provided by the mentor module.

[0028] As a preferred embodiment, the state space of the LAC-T model is shown in equation (11):

[0029] (11),

[0030] in, and y represents the true and estimated values ​​of the telemetry data vector at time t. t+1 and u represents the true and estimated values ​​of the telemetry data vector at time t+1. t+1 The physical system input for the complex electromechanical system at time t+1;

[0031] The action space in the LAC-T model is shown in equation (12):

[0032] (12)

[0033] in, This represents the estimated value of the implicit degradation parameter of a complex electromechanical system at time t+1.

[0034] As a preferred approach, LAC-T model training includes two parts: loss function design and model training strategy.

[0035] First, to meet the requirement of identifying implicit degradation parameters in complex electromechanical systems, the loss function is designed as shown in equation (13):

[0036] (13)

[0037] Among them, u t+1 y represents the telemetry data vector of the physical system of a complex electromechanical system at time t+1. t+1 as well as Let represent the true value vector and the estimated value vector of the telemetry data of the physical system of the complex electromechanical system at time t+1, respectively. This represents the estimated value vector of the telemetry data vector of the physical system of a complex electromechanical system at time t. This represents the identified value of the implicit degradation parameter of the actor module at time t+1;

[0038] Secondly, guiding the latent degradation parameters towards a direction with physical meaning requires the tutor module and actor module to simultaneously identify the latent degradation parameters; guiding the actor module to identify the direction of the latent degradation parameters is part of the fitting process, and the loss function for this part is shown in Equation (14):

[0039] (14)

[0040] in, θ represents the identified value of the implicit degradation parameter of the actor module at time t. t * It is a reference value for the implicit degradation parameter, and MSE represents the mean square error between the identification results of the two implicit degradation parameters;

[0041] Finally, the loss function obtained under the current state space vector s and action space vector a is shown in equation (15):

[0042] L r =μL1+(1-μ)L2(15)

[0043] Where μ represents the weights of the two loss functions; within the framework of reinforcement learning, L r This represents the loss function.

[0044] As a preferred approach, the actor module is trained based on the critic module's evaluation of the state space and action space to improve the mapping relationship between the state space and action space; the LAC-T model is trained based on the Lyapunov function and the maximum entropy principle, and the training objective function is shown in equation (16):

[0045] (16)

[0046] Where D is the empirical data pool; β is the policy entropy coefficient; λ is the Lagrange multiplier corresponding to the learning rate; Lr target The reference loss generated by the target network; s ' The state after the transition; A(s) ' ) for in s ' The policy output at the location; α3c is the Lyapunov / stability regularization term.

[0047] As a preferred approach, soft updates are used for the LAC-T model:

[0048] ,

[0049] Where: ϕ L These are the current network parameters; τ represents the target network parameters; τ represents the soft update coefficient.

[0050] Furthermore, the present invention also provides a complex electromechanical system remaining lifetime prediction system for implementing the method, comprising: a processor, a memory, and program modules running thereon, the modules including: a data acquisition and preprocessing module; a physical environment model module E; an actor module A; a critic module C; a mentor module T; and a training and inference module; enabling the system to output an implicit degradation parameter estimation sequence and remaining lifetime results.

[0051] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the computer to implement the method described thereon.

[0052] Furthermore, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0053] This invention, by employing the aforementioned technical solution, reconstructs parameter estimation in soft sensing as an "intrinsic parameter imitation" problem, and introduces a dual-channel mechanism tracking error and a mentor consistency term within the LAC-T framework to jointly constitute a composite loss L. r This mechanism suppresses the time contagion of single-step observation bias and the drift caused by nonlinear multiple solutions, thereby realizing the implicit degradation parameter. The system achieves continuous, smooth, and physically feasible convergence within the domain; Lyapunov stability regularization and maximum entropy training objective provide stable boundaries and robustness for policy updates, reducing divergence and oscillations in online estimation; state-action design fully utilizes information from previous and subsequent time steps and input stimuli to enhance robustness against noise, missing data, and operational condition changes; combined with experience pooling and soft updates, it can continuously optimize in real-time scenarios with low computational cost, shortening convergence time and reducing estimation variance. The overall effect is that, under conditions of multiple operational conditions, strong noise, and model mismatch, the implicit degradation parameter sequences obtained by soft measurement are more robust, interpretable, and exhibit better monotonicity / gradual variation, significantly improving the reliability and consistency of subsequent PHM tasks such as state monitoring, control decision-making, and RUL. Attached Figure Description

[0054] Figure 1 A schematic diagram illustrating the problems of observation data tracking and internal parameter simulation.

[0055] Figure 2 A framework for identifying implicit degradation parameters of complex electromechanical systems based on the LAC-T model.

[0056] Figure 3 shows the results of identifying implicit degradation parameters in the simulation data experiment. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0058] I. Identification of Implicit Degradation Parameters in Complex Electromechanical Systems Based on LAC-T Model

[0059] 1.1 Definition of the Latent Degradation Parameter Identification Problem

[0060] Typical physical systems can be represented by quintuples.<x,u,y,F,θ> Let θ represent the internal state vector of the physical system, u represent the input vector of the physical system, y represent the observation vector output by the physical system, F represent the mapping relationship between the physical system's state vector, input vector, and observation vector, and θ be the internal parameter of the mapping relationship F. The core objective of identifying implicit degradation parameters in complex electromechanical systems is to reasonably estimate the internal parameter θ of the physical model using x, u, and y, based on constraints such as physical information and statistical characteristics.

[0061] To identify the implicit degradation parameters of complex electromechanical systems, this chapter treats the complex electromechanical system as a physical system, telemetry data as the observation vector y of the physical system, and the implicit degradation parameters as parameters θ of the mapping relationship F between the complex electromechanical system and the physical system, and tracks them in real time. The specific steps can be described as follows: given the input vector u, the physical system F... t ∈R s Under the given conditions, output the observation vector. Where s and p represent the dimensions of the physical system input and observation vectors, respectively, as described in equations (1) and (2). To achieve the goal of tracking observation data, it is necessary to identify the parameter values. Reduce output observation vector estimation Compared with the true value of the observed vector y t+1 The error between them.

[0062] (1),

[0063] (2),

[0064] in, These represent the input vector of the physical system at time t+1 and the estimated observation vector at time t, respectively.

[0065] However, the core objective of existing physical system observation vector tracking problems is to accurately and in real-time track the physical model's observation vector y, thus weakening the accuracy and stability of the implicit parameter θ itself. A schematic diagram of this type of approach is shown below. Figure 1 As shown in (a), during the observation vector tracking process, regardless of the estimated value at time t... Compared with the true value y t Whether there is a deviation or not, as long as the true value y of the observed vector at time t+1 can be accurately tracked. t+1 The parameter values ​​identified at time t All are considered correct. Because the observed vector estimation cannot guarantee high accuracy with every moment in the true observed vector sequence, the identified parameter values... The sequence is unstable. To overcome the inherent limitations of the physical system output tracking problem, this chapter extends the observation vector tracking problem into an intrinsic parameter imitation problem by introducing a new tracking error, ensuring that the parameter θ... t Stability and accuracy of identification.

[0066] An intuitive description of the internal parameter imitation problem is as follows: Figure 1 (b) represents the physical system mapping relationship. F By observing the true value y of the vector at previous and subsequent time points t and y t+1 To determine the accuracy of the internal parameter identification from time t to time t+1, a new tracking error is introduced as shown in equation (3):

[0067] (3),

[0068] (4),

[0069] (5),

[0070] (6),

[0071] Compared with the tracking errors described by equations (2) and (3), the improved tracking error considers the case of multiple observation vector estimates, increasing the stability of the implicit parameter estimation. Among them, e1 is the same as the observation vector estimate in equation (2), representing the tracking error based on the previous observation vector estimate; e2 represents the tracking error based on the previous observation vector true value, which eliminates the influence of the large deviation of the previous observation vector estimate on the implicit parameter estimation; e3 represents the cumulative observation vector estimation error, which measures the overall estimation error of the observation vector sequence and ensures the stability of the implicit degenerate parameter estimation.

[0072] 1.2 LAC-T Model

[0073] Based on the definition of the latent degradation parameter estimation problem type, this chapter proposes a Lyapunov Actor-Critic-Tutor (LAC-T) model within the reinforcement learning framework to identify latent degradation parameters in complex electromechanical systems. The model structure is as follows: Figure 2 As shown. First, the LAC-T model is derived from the Lyapunov Actor-Critic (LAC) model, which includes three modules: Environment, Actor, and Critic. In the context of identifying implicit degradation parameters of complex electromechanical systems, the Environment module represents the physical model of the complex electromechanical system. Based on the identified degradation parameter values, the physical model input vector, and the telemetry data from the previous moment, it estimates the telemetry data at the current moment, which is the observation vector of the physical model of the complex electromechanical system. The Actor module is responsible for identifying the implicit degradation parameters of the complex electromechanical system at the current moment based on the telemetry data. The Critic module is responsible for evaluating the correctness of the identification of implicit degradation parameters at the current moment. According to the internal parameter imitation problem defined above, the Critic module uses the tracking error defined in Equation (3) as the basis for judging the correctness of the identification of implicit degradation parameters and assists the Actor module in adjusting the strategy for identifying implicit degradation parameters.

[0074] The aforementioned LAC model identifies implicit degradation parameters by tracking telemetry data from complex electromechanical systems, essentially solving a system of equations. However, due to the nonlinear components in the physical mechanism model of complex electromechanical systems, the mapping relationship between implicit degradation parameters and telemetry data exhibits multiple solutions. This multiple-solution phenomenon may lead to multiple numerical solutions for the identified implicit degradation parameters of complex electromechanical systems simultaneously satisfying the requirements of tracking telemetry data. However, from a physical perspective, implicit degradation parameters with physical characteristics cannot have multiple values ​​at the same time, and because they possess slowly varying characteristics, they represent a continuously changing time series. To avoid the interference of multiple solutions in implicit degradation parameter identification, this chapter improves the LAC model by introducing a mentor module, adding new physical mechanism constraints to provide a reference for the actor module, guiding the direction of implicit degradation parameter identification, and improving the accuracy of implicit degradation parameter identification.

[0075] Based on the above design, the basic principle of the LAC-T model can be expressed by equations (7) to (10):

[0076] (7),

[0077] (8),

[0078] (9),

[0079] (10)

[0080] Where A, E, C, and T represent the mapping relationships within the actor, environment, critic, and mentor modules, respectively. t+1 Let y represent the input vector of the complex electromechanical system's physical system at time t+1. t+1 as well as This represents the true and estimated values ​​of the telemetry data vector of a complex electromechanical system at time t+1. It is the telemetry data vector estimate of the physical system at time t. The implicit degradation parameter identification value of the actor module at time t+1 is represented by r, the evaluation of the physical system telemetry data by the environment module is represented by v, and the estimated value of the implicit degradation parameter by the critic module is represented by v. Evaluation, θ t *These are reference values ​​for the implicit degradation parameters provided by the mentor module. It must be noted that although the mentor module introduces a new mapping relationship between telemetry data and implicit degradation parameters, it does not need to provide accurate identification values ​​for these parameters. Therefore, the mentor module can be either a pre-trained model or a simple physical mapping relationship between telemetry data and implicit degradation parameters. In the LAC-T model, the main functions of the mentor module's implicit degradation parameter identification reference values ​​are: 1) to limit the range of values ​​for the implicit degradation parameter identification of the actor module, improving the convergence of parameter identification; 2) to provide direction for the implicit degradation parameter identification of the actor module, overcoming the phenomenon of multiple solutions at the mathematical level.

[0081] 1.3 State Space and Action Space Design

[0082] To apply the LAC-T model to the identification of implicit degradation parameters in complex electromechanical systems, the state space s and action space a of the model need to be designed. Based on the previous analysis, the problem of identifying implicit degradation parameters in complex electromechanical systems can be viewed as a problem of imitating parameters within the physical system. Therefore, to identify implicit degradation parameters in complex electromechanical systems, the actor module needs to consider both the actual and estimated values ​​of telemetry data at different times. The state space of the LAC-T model is shown in equation (11):

[0083] (11),

[0084] Among them, y t and y represents the true and estimated values ​​of the telemetry data vector at time t. t+1 and u represents the true and estimated values ​​of the telemetry data vector at time t+1. t+1 The physical system input for the complex electromechanical system at time t+1.

[0085] The action space of the LAC-T model represents the corresponding actions performed by the actor module based on the state space quantities at the current moment, which serve as the input to the environment module and drive the output of the state space at the next moment. In the identification of implicit degradation parameters in complex electromechanical systems, the actor module estimates the values ​​of implicit degradation parameters based on telemetry data from previous and subsequent moments and feeds them back into the physical model of the complex electromechanical system as the basis for tracking telemetry data at the next moment. Therefore, the action space in the LAC-T model proposed in this chapter is shown in Equation (12):

[0086] (12)

[0087] in, This represents the estimated value of the implicit degradation parameter of a complex electromechanical system at time t+1.

[0088] 1.4 Estimation Process of Implicit Degradation Parameters for Complex Electromechanical Systems

[0089] The estimation process of implicit degradation parameters for complex electromechanical systems revolves around the LAC-T model and mainly includes two stages: model training and online parameter identification. The LAC-T model training mainly includes two parts: loss function design and model training strategy. First, the LAC-T model needs to consider not only the accuracy of telemetry data vector estimation but also whether the identification of implicit degradation parameters can maintain consistency and continuity when the telemetry data estimation deviates at the current moment. Therefore, to meet the requirements of implicit degradation parameter identification for complex electromechanical systems, the loss function is designed as shown in equation (13):

[0090] (13)

[0091] Among them, u t+1 y represents the telemetry data vector of the physical system of a complex electromechanical system at time t+1. t+1 as well as Let represent the true value vector and the estimated value vector of the telemetry data of the physical system of the complex electromechanical system at time t+1, respectively. This represents the estimated value vector of the telemetry data vector of the physical system of a complex electromechanical system at time t. This represents the identified value of the implicit degradation parameter of the actor module at time t+1.

[0092] Secondly, to guide the training of the latent degradation parameters in a direction with physical meaning, the mentor module and the actor module need to identify the latent degradation parameters synchronously. Essentially, guiding the actor module to identify the direction of the latent degradation parameters is part of the fitting process; therefore, this part of the loss function is shown in equation (14):

[0093] (14)

[0094] in, θ represents the identified value of the implicit degradation parameter of the actor module at time t+1. t * It is a reference value for the implicit degradation parameter, and MSE represents the mean square error between the identification results of the two implicit degradation parameters.

[0095] Taking both types of errors into account, the loss function for the current state space vector s and action space vector a can be obtained as shown in equation (15):

[0096] L r =μL1+(1-μ)L2(15)

[0097] Where μ represents the weights of the two loss functions. Within the framework of reinforcement learning, L... rThe loss function is typically viewed as the reward given by the environment module under the current state and action. It serves as the basis for the critic module to judge the value of the current action, guides the actor module to estimate the direction of the latent degradation parameters, and limits the range of latent degradation parameter identification to improve the convergence of parameter identification, thereby further improving the accuracy and stability of latent degradation parameter identification.

[0098] Based on the evaluation function design, the actor module is trained according to the critic module's evaluation of the state space and action space, thereby improving the mapping relationship between the state space and action space. The LAC-T model is trained based on the Lyapunov function and the maximum entropy principle, and the training objective function is shown in equation (16):

[0099] (16)

[0100] Where λ and β are the progressive learning rate and the positive Lagrange operators controlling the entropy stability of the policy, respectively, in reinforcement learning. This objective training function guarantees the smoothness of parameter estimation. Therefore, based on the designed reward function and learning policy, the training process of the LAC-T model is as follows.

[0101] Input: Model learning rate Environmental Model E

[0102] Output: Optimal characteristic parameter estimation model

[0103] Step 1: Randomly initialize the Lyapunov network L(s,a) and the policy function A(a|s);

[0104] Step 2: Initialize the model parameters of Actor Module A and Critic Module C. ;

[0105] Iteration:

[0106] Step 3: Sample the initial state vector s o ;

[0107] At each time t:

[0108] Step 4: Estimate the characteristic parameter θ from actor module A based on strategy π. t As the action vector a t ;

[0109] Step 5: Combine the observation vectors s before and after observations. t+1 The input is fed into the environment model E to obtain the observation vector s. t+1 The estimated quantity and reward value r t , forming a path (s t ,a t ,r t ,s t+1 Store it in data pool D;

[0110] stop

[0111] Each update iteration:

[0112] Step 6: Collect path pairs from data pool D and update L, π and the corresponding multiplication operators using gradient descent;

[0113] Step 7: Update the strategy: ;

[0114] Stop updating and iterating

[0115] Stop iteration.

[0116] II. Case Analysis

[0117] 2.1 Case Study: Simulation Data Verification

[0118] 2.2 Validation of the effectiveness of latent degradation parameter identification based on the LAC-T model

[0119] To verify the effectiveness of the LAC-T model in identifying implicit degradation parameters of complex electromechanical systems, this case study selects digital simulation data of a complex electromechanical system, using motor torque coefficient, motor current data, and motor speed under mechanical degradation failure of a high-speed rotor system as data sources. Specifically, the motor torque coefficient of the high-speed rotor system is used as an implicit degradation parameter, while the motor current and motor speed are used as observable telemetry data. To fully verify the stability and generalization ability of the LAC-T model proposed in this chapter under different degradation modes, this case study selects six sets of simulation data with different parameter value ranges and degradation rates. The experimental results for torque coefficient identification are as follows: Figure 3 As shown in the figure, the comparison between the estimated value and the actual value of the torque coefficient parameter in the figure shows that for different parameter value ranges and degradation rates, the LAC-T model can estimate the torque coefficient parameter value well based on telemetry data and physical mechanism model. The experimental results prove that the LAC-T model proposed in this chapter has the ability to identify implicit degradation parameters online.

[0120] Based on the demonstration that the LAC-T model has the ability to identify hidden degradation parameters, this case study selects the physical mechanism method based on expert knowledge and the LAC model as comparisons. At the same time, mean square error, mean absolute error and normalized root mean square error are selected as quantitative evaluation indicators to measure the effect of hidden degradation parameter identification, thus verifying the superiority of the LAC-T model over existing parameter identification methods.

[0121] Table 1 Comparison of Identification Errors for Implicit Degradation Parameters (Motor Torque Coefficient of High-Speed ​​Rotor System)

[0122]

[0123] Table 1 shows a comparison of parameter identification errors for different methods. As can be seen from the table, the LAC model, compared to the LAC-T model proposed in this chapter, achieves a lower identification error value based on the physical model. This comparison indicates that, due to the limitations of human subjective experience, the physical model cannot fully describe the mapping relationship between implicit degenerate parameters and telemetry data, making it difficult to meet the needs of identifying implicit degenerate parameters in complex physical systems. The reinforcement learning-based method, by combining physical mechanisms and data, improves the accuracy of implicit degenerate parameter identification. Furthermore, according to the experimental results in Table 1, the LAC-T model achieves a lower implicit degenerate parameter identification error compared to the LAC model. This result shows that the LAC-T model, by designing a mentor module to introduce more physical information constraints, effectively guides the direction of implicit degenerate parameter identification, further improving the accuracy of parameter identification.

[0124] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

[0125] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0130] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0131] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A soft-sensing method for a complex electromechanical system based on LAC-T model, characterized in that, The method comprises the following steps: S1) constructing a physical system mapping relationship F, outputting an observation vector estimation given an input vector and a previous time observation estimation ; S2) modeling the implicit parameter identification as a decision-making process based on the LAC-T model framework, using an actor-environment-critic-tutor linkage mechanism to output the implicit parameter identification value at time t ; S3) with tracking error The LAC-T model framework is trained for supervisory signals such that the parameter trajectories remain continuous and stable in the presence of observation bias; S4) input (y t , u t ) in online phase, output hidden parameter sequence Real-time recognition is completed, y t represents the true value of the telemetry data vector at time t, and u t is the input of the physical system of the complex electromechanical system at time t; Given an input vector , the output observation vector estimate is obtained under the condition that t , where s and p denote the dimensions of the physical system input and observation vector, respectively, and the identified parameter value t is such that the error between the output observation vector estimate and the true value of the observation vector t+1 is minimized; this is described by equations (1) and (2): （1）， （2）， wherein, respectively represent the input vector of the physical system at time t+1 and the observation vector estimate at time t; F is the physical mechanism mapping between the internal state of the system, the input and the observation; Alternatively, the tracking error is shown as formula (3): （3）， （4）， （5）， （6）， Wherein: e1 is the same as the observation vector estimation in formula (2), representing the tracking error based on the observation vector estimation value at the previous moment; e2 represents the tracking error based on the true value of the observation vector at the previous moment; e3 represents the cumulative observation vector estimation error; ‖.‖ is the vector Euclidean norm; T is the end time of the sequence currently considered; The state space of the LAC-T model is shown as formula (11): （11）， where y t with denote the true and estimated values of the telemetry data vector at time t, y t+1 with denote the true and estimated values of the telemetry data vector at time t+1, u t+1 is the physical system input of the complex mechatronic system at time t+1; The action space of the LAC-T model is shown as formula (12): （12）， wherein, represents the estimate of the hidden parameter of the complex mechatronic system at time t+1.

2. The method of claim 1, wherein, The LAC-T model framework satisfies: （7）， （8）， （9）， （10）， wherein A, E, C, T represent the mapping relationship inside the actor, environment, critic and tutor modules respectively; u t+1 represents the input vector of the complex mechatronic system physical system at t+1 time, y t+1 and represents the true value and the estimated value of the telemetry data vector of the complex mechatronic system physical system at t+1 time, is the telemetry data vector estimation of the physical system at t time, represents the implicit parameter identification value of the actor module at t+1 time, r represents the evaluation of the telemetry data of the physical system by the environment module, v represents the evaluation of the implicit parameter estimation value by the critic module, is the implicit parameter reference value provided by the tutor module.

3. The method of claim 1, wherein, The LAC-T model training includes two parts of loss function design and model training strategy. Firstly, in order to meet the needs of the complex mechanical and electrical system implicit parameter identification, the loss function design is shown as formula (13): (13), wherein u t+1 represents the telemetry data vector of the complex electromechanical system physical system at time t+1, y t+1 and respectively represent the true value vector and the estimated value vector of the telemetry data vector of the complex electromechanical system physical system at time t+1, represents the estimated value vector of the telemetry data vector of the complex electromechanical system physical system at time t, represents the identified value of the hidden parameter of the actor module at time t; Secondly, the guide implicit parameters to the direction with physical meaning training, need to guide the teacher module and the actor module to identify the implicit parameters simultaneously; Guide the actor module to identify the implicit parameter direction belongs to the fitting process, this loss function part is shown as formula (14): (14), wherein, represents the recognition value of the implicit parameter of the actor module at time t, is the reference value of the implicit parameter, and MSE represents the mean square error between the two recognition results of the implicit parameter. Finally, the loss function under the current state space vector s and the action space vector a is shown as formula (15): （15）， wherein μ is a weight value of the two loss functions; under the framework of reinforcement learning, L r represents a loss function.

4. The method of claim 1, wherein, The actor module is trained according to the evaluation of the critic module on the state space and the action space, and the mapping relationship between the state space and the action space is improved; The LAC-T model is trained based on Lyapunov function and maximum entropy principle, and the training target function is shown as formula (16): （16）， where D is an empirical data pool; β is a policy entropy coefficient; and λ is a Lagrange multiplier corresponding to a learning rate; a reference loss generated by the target) network; is the state after transition; is the policy output at is the policy output at is a Lyapunov / stability regularization term.

5. The method of claim 1, wherein, The experience pool and soft update mechanism are used for training the LAC-T model: (i) Sampling and storing the four-tuple (st, at, rt, st+1) ∈ D; (ii) Network parameter soft update: ， wherein: is a current network parameter; is a target network parameter; is a soft update coefficient, D is an empirical data pool.

6. A system for residual life prediction of a complex electromechanical system implementing the method of any one of claims 1 to 5, comprising: The processor, the memory and the program modules running thereon, the modules comprising: Data acquisition and preprocessing module; Physical environment model module E; Actor module A; Critic module C; Teacher module T; Training and reasoning module.

7. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to realize the method of any one of claims 1-5.

8. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed by the processor to realize the method of any one of claims 1-5.

Images