Event-triggered model learning method based on physical information neural network

Abstract

The present invention discloses an event-triggered model learning method based on a physical information neural network, which relates to the field of artificial intelligence technology. Step one: construct a physical information neural network of the system to complete the model identification of the system; step two: perform online state estimation of the system in combination with an unscented Kalman filter based on the system model and real-time input and output; step three: use the state estimation result, combine the model to perform a one-step forward output prediction and calculate the prediction error, and determine whether the model needs to be updated through an event-triggered learning mechanism. The present invention adopts the above-mentioned event-triggered model learning method based on a physical information neural network, and proposes a physical information neural network by constructing a physical information neural network system identification method and an event-triggered learning method, combining the advantages of traditional mechanism analysis modeling methods and data-driven modeling methods, while taking into account the interpretability of the identification model and improving the accuracy of model identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an event triggering model learning method based on a physical information neural network. Background Art

[0002] With the development of industry and manufacturing, real-time monitoring of industrial systems is the premise to ensure their safe operation, and establishing an accurate and effective industrial system model is the first step in real-time monitoring. Existing methods for identifying complex industrial systems include mechanism analysis and data-driven methods. Although the mechanism analysis method has the advantages of interpretability and easy description of changes in the operation process, its model error is often large and the description of the system may involve more complex physical and chemical formulas, and there are certain difficulties in solving its parameters; while the data-driven method requires a large amount of historical operation data of the target system, and the model building process requires a lot of computing resources and time. On the other hand, for some industrial systems, the monitoring equipment and actuators of the system are far away, so the two need to use limited communication resources to achieve real-time monitoring of the system. A specific example is the drone system. When the drone is working, its actuator and monitoring system (remote control) are generally far away and the transmission distance and data are limited. Secondly, considering that the life of a general industrial system can reach several years or even decades, and the industrial system will age after a certain period of operation, it is necessary to decide whether to update the model according to the degree of aging of the system to reduce the error of real-time state estimation. A specific example is the battery management system of an electric vehicle. The health status of the battery will decrease as the number of charge and discharge cycles increases. To achieve real-time and accurate monitoring of the battery status, it is necessary to consider the aging of the battery and update the battery model when necessary to reduce the estimation error.

[0003] Therefore, firstly, the proposed physical information neural network is used to identify the industrial system so that the established model has both interpretability and accuracy. Secondly, an event-triggered learning method is designed to realize online system status monitoring and reduce the state estimation error caused by system aging, thereby ensuring the accuracy and effectiveness of system monitoring. Summary of the invention

[0004] The purpose of the present invention is to provide an event-triggered model learning method based on physical information neural network, to realize a hybrid modeling method combining physical information and neural network for industrial systems, to make the model interpretable while improving the modeling accuracy of the system; to complete the online monitoring of the system by estimating the system state through an unscented Kalman filter; finally, to perform a one-step forward output prediction based on the state estimation result, and to perform system aging assessment through weighted multi-step output prediction error and event triggering mechanism.

[0005] To achieve the above object, the present invention provides an event trigger model learning method based on physical information neural network, and the specific steps are as follows:

[0006] Step 1: Build the physical information neural network of the system to complete the model identification of the system;

[0007] Step 2: Based on the system model and real-time input and output, the unscented Kalman filter is used to estimate the system state online;

[0008] Step 3: Use the state estimation results and the model to make a one-step forward output prediction and calculate the prediction error, and use the event-triggered learning mechanism to determine whether the model needs to be updated.

[0009] Preferably, the physical model equation of the system in step 1 is as follows:

[0010] x k+1 =f(x k ,u k )+w k

[0011] y k =h(x k ,u k )+v k

[0012] Where: f(·,·) represents the time-invariant discrete dynamic system function, h(·,·) represents the measurement function of the system; x k =[x 1,k ,x 2,k ,...,x n,k ] T ∈R n is the state variable at time k, and its initial state is represented by x0; the control signal is defined as u k =[u 1,k ,u 2,k ,...,u m,k ] T ∈R m , the measured / controlled signal is defined as y k =[y 1,k ,y 2,k ,...,y p,k ] T ∈R p ;x 1,k ,x 2,k ,...,x n,k For the system, it is an observable state variable that cannot be directly measured. 1,k ,u 2,k ,...,u m,k For the system, it is a known deterministic input signal, y 1,k,y 2,k ,...,y p,k It is a measurable output signal for the system; w k and v k are used to represent process noise and measurement noise, respectively.

[0013] Preferably, the first step is to construct a neural network using a physical model of the system, and the specific steps are as follows:

[0014] One construction method is to design the neural network weight matrix used to identify the system through the physical equations of the system, so that the network training process can meet the physical constraints of the system, and finally the parameters of the physical model are obtained through the neural network;

[0015] Or transform part of the output of neurons in the hidden layer to make it satisfy the physical differential equations, or impose special network layers and connection relationships on these neurons, set physical constraints, and realize the physical information network construction of the system.

[0016] Preferably, the specific steps of step 2 are as follows:

[0017] (1) Initialize the model parameters according to the physical model equation of the system in step 1;

[0018] (2) Update the prior time of entering step k;

[0019] (3) Obtain the final state estimate of the kth step through the posterior measurement update; according to the input and output u of the system k ,y k , after the unscented Kalman filter algorithm, the online state of the system is completed Estimates.

[0020] Preferably, the specific steps of step three are as follows:

[0021] (1) Based on the state estimation result of step 2 Use the current model to make a forward output prediction and model prediction error;

[0022] (2) Predict output based on measurement and the actual output y k+1 The absolute value of the multi-step difference is used to set the event-triggered learning mechanism to determine whether the current model of the system is aging.

[0023] Therefore, the present invention adopts the event trigger model learning method based on physical information neural network of the above structure, which has the following beneficial effects:

[0024] (1) The present invention constructs a physical information neural network system identification method and an event-triggered learning method, and proposes a physical information neural network, which combines the advantages of traditional mechanism analysis modeling methods and data-driven modeling methods, while taking into account the interpretability of the identification model and improving the accuracy of model identification.

[0025] (2) The present invention combines the demand for real-time monitoring of system status with an unscented Kalman filter to complete online system status estimation based on the input and output of the system, thereby ensuring the timeliness of monitoring and the safety of system operation.

[0026] (3) The present invention takes into account factors such as sudden changes in the system's working environment and model aging. After completing the online state estimation, a one-step forward output prediction is performed based on the system identified model and system input to determine whether the average prediction error of multiple steps exceeds the trigger threshold. When the trigger threshold is exceeded, the event triggers the learning mechanism and the system model is re-identified to ensure the accuracy of the system and the effectiveness of monitoring. When the trigger threshold is not exceeded, the original system model is still used for online state estimation and output prediction.

[0027] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of an embodiment of an event trigger model learning method based on a physical information neural network of the present invention;

[0029] Figure 2 Voltage prediction and prediction error diagram of lithium battery physical information neural network model;

[0030] Figure 3 It is the online state estimation of lithium battery power and the estimation error diagram;

[0031] Figure 4 Update the previous step forward voltage prediction and prediction error graph for the aged lithium battery model;

[0032] Figure 5 One-step forward voltage prediction and prediction error plot updated for the aged lithium battery model. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0034] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0035] Example

[0036] See also Figure 1-5 The present invention provides an event trigger model learning method based on physical information neural network, and the specific steps are as follows:

[0037] (1) Construct the system's physical information neural network and complete the system model identification.

[0038] Considering similar systems with known physical models such as power grid energy management system, spacecraft power system and battery management system, the discrete equivalent physical model of such systems can be described as follows:

[0039] x k+1 =f(x k ,u k )+w k

[0040] y k =h(x k ,u k )+v k

[0041] Where f(·,·) represents the time-invariant discrete dynamic system function, and h(·,·) represents the measurement function of the system;

[0042] x k =[x 1,k ,x 2,k ,...,x n,k ] T ∈R n is the state variable at time k, and its initial state is represented by x0; the control signal is defined as u k =[u 1,k ,u 2,k ,...,u m,k ] T∈R m , the measured / controlled signal is defined as y k =[y 1,k ,y 2,k ,...,y p,k ] T ∈R p ;x 1,k ,x 2,k ,...,x n,k For the system, it is an observable state variable that cannot be directly measured. 1,k ,u 2,k ,...,u m,k For the system, it is a known deterministic input signal, y 1,k ,y 2,k ,...,y p,k It is a measurable output signal for the system; w k and v k are used to represent process noise and measurement noise, respectively.

[0043] The above-mentioned equivalent physical model equations are combined with data-driven neural networks. The specific method is to use the physical model to construct the neural network. One construction method is to design the neural network weight matrix used to identify the system through the physical equations of the system, so that the network training process can meet the physical constraints of the system, and finally the parameters of the physical model are obtained through the neural network; another method is to transform a part of the output of the neurons in the hidden layer to satisfy the physical differential equations or to impose special network layers and connection relationships on these neurons, set physical constraints, and thus realize the physical information network construction of the system. Both methods can alleviate the black box nature of the neural network and make some deep neural networks interpretable. Their ultimate goal is to obtain the physical model through the neural network, that is, to obtain the system's f(·,·) and h(·,·) functions.

[0044] (2) Based on the system model and real-time input and output, the unscented Kalman filter is used to perform online state estimation of the system.

[0045] According to step (1), after the physical model system identification is completed, the dynamic system function f(·,·) and the system measurement function h(·,·) are known. To monitor the system in real time, it is necessary to perform online state estimation using the unscented Kalman filter algorithm. Assume that the input and output of the monitored system at time k are u k ,y k , combined with the identified system model, the unscented Kalman filter algorithm for completing online state estimation is described as follows. To facilitate the algorithm description, the commonly used symbols and their meanings in the present invention are shown in Table 1 below.

[0046] Table 1 Common symbols and definitions

[0047]

[0048] First, according to the discrete equivalent physical model of the system in step (1), the model parameters are initialized. where E[·] represents the expected function.

[0049] Secondly, enter the k-th step of the prior time update, the specific steps are as follows:

[0050] The first step is to estimate the posterior state of the k-1th step Sampling to obtain Sigma point set and the weight matrix The meanings of n, α, κ, λ, and β are shown in Table 1.

[0051] The second step is to calculate the one-step state prediction of each point in the Sigma point set according to the system function f(·,·) And weighted to get the state prior estimate and the state prior covariance P k x- :

[0052]

[0053]

[0054] Where: T represents the transpose of the matrix.

[0055] Again, the final state estimate of step k is obtained through the posterior measurement update. The specific steps are as follows:

[0056] The first step is to calculate the one-step observation prediction value of each point in the Sigma point set according to the system's measurement function h(·,·) And weighted to obtain the measured predicted value and covariance P k y :

[0057]

[0058]

[0059] The second step is to calculate the prior state With the measured predicted value The cross covariance between And combined with the covariance P of the measured prediction k y Calculate the Kalman gain K k =P k xy (Pk y ) -1 .

[0060] The third step is to use the measured value y k Update to obtain the posterior state and the covariance matrix P k x :

[0061]

[0062]

[0063] Finally, according to the input and output u of the system k ,y k , after the unscented Kalman filter algorithm, the online state of the system is completed Estimates.

[0064] (3) Using the state estimation results, the model is combined to make a one-step forward output prediction and calculate the prediction error, and the aging model is updated through an event-triggered learning mechanism.

[0065] After the online state estimation of the system has been completed by step (2), considering that the system performance changes with the increase of working time or the change of working environment, the system model identified in step (1) may be aged or inaccurate, resulting in inaccurate online state estimation results. Therefore, an event trigger mechanism and a learning trigger are designed to update the aging model.

[0066] The present invention evaluates whether the current system is aged by using the state estimation result combined with the identified system model to perform a one-step output prediction. If the multi-step weighted output prediction error exceeds the set system safe operation output error range, it is considered that the system has aged or the working environment has changed significantly, and the original model is no longer applicable and needs to be updated according to step (1).

[0067] First, according to the state estimation result of step (2) Use the current model to make a forward output prediction to get the model prediction error. The specific steps are as follows:

[0068] According to the input u of step k and step k+1 k ,u k+1 And the state estimation result of step k The measured output of the system at step k+1 is The prediction is made as follows:

[0069]

[0070]

[0071] Second, predict the output based on the measurements and the actual output y k+1 The absolute value of the multi-step difference is used to set the event-triggered learning mechanism to determine whether the current model of the system is aged as follows:

[0072]

[0073] where γ learn is the binary trigger variable of the event-triggered learner, y i and are the actual measured value of the i-th step and the predicted value obtained by one-step forward prediction, and δ is the event trigger threshold set according to experience. At the same time, the prediction error window of h+1 step is selected to detect smaller continuous errors rather than larger errors of one-step window to avoid unnecessary system updates and waste of communication resources due to occasional large system prediction errors, thereby making the designed event trigger more robust. j (j=kh,...,k) is a weighting coefficient used to make the prediction error closer to the prediction window have a greater weight, so it satisfies:

[0074]

[0075] When γ learn =1, it means that the original system needs to be updated. At this time, the sensor of the actuator sends the measured system measurement information to the receiving end for model update, that is, the aged model is relearned in step (1). learn =0, the current system model is maintained for state estimation and then transmitted to the receiving end to implement online monitoring of the system.

[0076] See also Figure 2 , Figure 2 The effect of using physical information neural network to identify the model of lithium batteries is demonstrated. It can be seen that under the modeling method proposed in the present invention, the prediction error of the system is within a very small range, which proves the accuracy of the proposed model identification method.

[0077] See also Figure 3 , Figure 3 This is the result of online state estimation of the lithium battery power through the unscented Kalman filter combined with the identified lithium battery model. It can be seen that the estimation error is limited to within 1.5%, which once again proves the effectiveness of the system physical information neural network identification model and the accuracy of the online state estimation algorithm.

[0078] See also Figure 4-5 , Figure 4 as well as Figure 5It is the result of event-triggered learning and updating the model for an aged lithium battery system after multiple cycles of charge and discharge. Figure 4 It can be clearly seen from the one-step forward output predicted voltage error result that the predicted voltage error exceeds the set trigger threshold. At this time, the aging model will be updated when the event trigger condition is met. Figure 5 It is the one-step forward output voltage and error predicted for the lithium battery after relearning the physical information neural network. It can be clearly seen that the prediction error is controlled within the trigger threshold, which reflects the practicality of the event-triggered learning mechanism of event setting and ensures the accuracy of the battery's real-time state estimation and the effectiveness of the battery model.

[0079] Aiming at the problem of online state estimation of industrial systems that require real-time monitoring, the present invention proposes a system modeling method of a physical information neural network and performs system online state estimation based on an identification model, ensuring the accuracy of system modeling while making the model physically interpretable. By designing an event-triggered learning mechanism after the system performs online state estimation, it is ensured that the system will not affect the effectiveness of system monitoring due to aging, thereby ensuring the safety and continuity of subsequent system operation.

[0080] Therefore, the present invention adopts the event-triggered model learning method based on the physical information neural network of the above structure. By constructing a physical information neural network system identification method and an event-triggered learning method, the proposed physical information neural network combines the advantages of traditional mechanism analysis modeling methods and data-driven modeling methods, while taking into account the interpretability of the identification model and improving the accuracy of model identification; combined with the needs of real-time monitoring of the system status, combined with the unscented Kalman filter, the system status estimation is completed online according to the input and output of the system, ensuring the timeliness of monitoring and the safety of system operation.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. An event-triggered model learning method based on physical information neural network, characterized in that: The specific steps are as follows: Step 1: Build the physical information neural network of the system and complete the model identification of the system. Specifically, it includes designing the neural network weight matrix used to identify the system through the physical equation of the system, realizing that the network training process meets the physical constraints of the system, and finally obtaining the parameters of the physical model through the neural network. The physical model equation is as follows: x k+1 =f(x k ,u k )+w k y k =h(x k ,u k )+v k Where: f(·,·) represents the time-invariant discrete dynamic system function, h(·,·) represents the measurement function of the system; x k =[x 1,k ,x 2,k ,...,x n,k ] T ∈R n is the state variable at time k, and its initial state is represented by x0; the control signal is defined as u k =[u 1,k ,u 2,k ,...,u m,k ] T ∈R m , the measured / controlled signal is defined as y k =[y 1,k ,y 2,k ,...,y p,k ] T ∈R p ;x 1,k ,x 2,k ,...,x n,k For the system, it is an observable state variable that cannot be directly measured. 1,k ,u 2,k ,...,u m,k For the system, it is a known deterministic input signal, y 1,k ,y 2,k ,...,y p,k It is a measurable output signal for the system; w k and v k They are used to represent process noise and measurement noise respectively; Step 2: Based on the system model and real-time input and output, the unscented Kalman filter is used to estimate the system state online. The specific steps are as follows: (1) Initialize the model parameters according to the physical model equation of the system in step 1; (2) Update the prior time of entering step k; (3) Obtain the final state estimate of the kth step through the posterior measurement update; according to the input and output u of the system k ,y k , after the unscented Kalman filter algorithm, the online state of the system is completed estimates; Step 3: Use the state estimation results and the model to make a one-step forward output prediction and calculate the prediction error. Use the event-triggered learning mechanism to determine whether the model needs to be updated. The specific steps are as follows: (1) Based on the state estimation result of step 2 Use the current model to make a forward output prediction and calculate the model prediction error; (2) Predict output based on measurement and the actual output y k+1 The absolute value of the multi-step difference is used to set the event-triggered learning mechanism to determine whether the current model of the system is aging.

2. The event trigger model learning method based on physical information neural network according to claim 1 is characterized in that: The first step is to construct a neural network using the physical model of the system, and also includes the following method: Transform part of the output of neurons in the hidden layer to satisfy the physical differential equations or impose network layers and connection relationships on these neurons, set physical constraints, and realize the physical information network construction of the system.

Images