A dynamic parameter monitoring method for low-level system digital controller

Abstract

The present invention relates to the field of electronic data processing technology and discloses a dynamic parameter monitoring method for a low-level system digital controller. The method comprises constructing an augmented state vector including the physical state quantity of a controlled object and the controller parameter to be monitored, establishing a state space model; calculating a priori state estimate of the current cycle; calculating a process noise covariance matrix Q and a priori error covariance matrix of the current cycle; calculating an innovation vector; calculating its kurtosis statistic to adjust a forgetting factor and updating a weighted measurement noise covariance matrix R; calculating the Mahalanobis distance of the innovation vector, and if it exceeds a preset threshold, activating a parallel branch processing mechanism to evaluate and prune low-probability branches based on new measurement values ​​within a subsequent predetermined cycle; and extracting the current value of the parameter to be monitored from the posterior state estimate of the current cycle. The present invention enables the state prediction link to better match the actual uncertainty of the system under different working conditions, thereby improving the accuracy of model prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic data processing, in particular to a dynamic parameter monitoring method of a low-level system digital controller. Background Art

[0002] In low-level digital control systems such as motor drives and power electronic converters, accurate online monitoring of key system physical parameters (such as motor winding resistance and inductance, or capacitors and inductors in power filters) is crucial for improving system control performance, enabling fault diagnosis, and enabling predictive maintenance. These parameters often drift due to changes in operating temperature, material aging, or varying operating conditions. If the controller still uses initially calibrated parameters, control accuracy will decrease, system efficiency will decline, and even oscillation or instability may occur. Therefore, developing efficient and accurate online parameter monitoring methods has become a research hotspot in this field. State observer technology, particularly the Kalman filter (KF) and its variants based on state-space models, has been applied to such parameter identification problems because it can fuse model predictions with actual measurements to provide optimal state estimates in noisy environments.

[0003] However, traditional Kalman filters face numerous challenges in practical applications. First, the standard KF and its extended forms (such as the EKF and UKF) typically assume that the system's process and measurement noise are Gaussian white noise with known and fixed statistical properties. However, in practical low-level control systems, the system model itself is inherently nonlinear and uncertain. Furthermore, the noise characteristics introduced by actuators and sensors are often time-varying and non-Gaussian, and may even be interspersed with sudden strong pulses or outliers caused by electromagnetic interference, communication errors, and other factors. Using fixed noise covariance matrices Q and R not only makes it difficult to achieve optimal filtering results, but also leads to a significant decrease in estimation accuracy and even filter divergence when the actual noise significantly mismatches the preset values. While research on adaptive Kalman filters (AKFs) has addressed this issue, existing methods either fail to respond quickly enough to noise variations or struggle to effectively distinguish between model mismatch and measurement anomalies. Furthermore, when dealing with non-Gaussian outliers, simple removal or suppression strategies can easily result in the loss of useful information, compromising the continuity and stability of parameter estimates. Therefore, there is an urgent need for a new parameter monitoring method that can dynamically adapt to time-varying, non-Gaussian noise environments and robustly handle measurement anomalies. Summary of the Invention

[0004] The present invention provides a dynamic parameter monitoring method for a low-level system digital controller to solve the problem in the prior art that when actual noise seriously mismatches a preset value, the parameter monitoring accuracy is significantly reduced.

[0005] The dynamic parameter monitoring method of the low-level system digital controller of the present invention comprises the following steps:

[0006] Construct an augmented state vector containing the physical state of the controlled object and the parameters to be monitored by the controller, and establish a nonlinear discrete-time state space model; in each control cycle, calculate the prior state estimate of the current cycle based on the state space model and the posterior state estimate of the previous cycle; use the nonlinear mapping function driven by the specific physical state component in the prior state estimate to calculate the process noise covariance matrix Q, and combine the posterior error covariance matrix of the previous cycle to calculate the prior error covariance matrix of the current cycle; obtain the real-time measurement vector and calculate the innovation vector; based on the sequence of the innovation vector in the preset time window, calculate its kurtosis statistic to adjust a forgetting factor, and then update the weighted measurement noise covariance matrix online. difference matrix R; calculate the Mahalanobis distance of the innovation vector and compare it with the preset threshold: if it exceeds the preset threshold, the parallel branch processing mechanism is started, in which one branch uses the expanded measurement noise covariance matrix for measurement update, and the other branch skips the measurement update of the current cycle, and evaluates and prunes the low-probability branch based on the new measurement value in the subsequent predetermined cycle; if the Mahalanobis distance does not exceed the preset threshold or after the parallel branch processing is completed, the Kalman gain is calculated using the updated process noise covariance matrix Q and the measurement noise covariance matrix R, and the measurement update is performed to obtain the posterior state estimate and posterior error covariance matrix of the current cycle; the current value of the parameter to be monitored is extracted from the posterior state estimate of the current cycle.

[0007] Preferably, the construction of the augmented state vector including the physical state of the controlled object and the parameters to be monitored by the controller includes: p and the controller parameter vector x to be monitored θ Cascade to form an augmented state vector x=[x p T , x θ T ] T , where T represents the transpose operation and the controller parameters are modeled as a random walk process.

[0008] Preferably, the process noise covariance matrix Q is calculated by using a nonlinear mapping function driven by a specific physical state component in the prior state estimate, including: extracting the prior state estimate x k|k-1 The motor speed component ω in k|k-1 , calculate the process noise covariance matrix Q of the current cycle through the preset exponential function k , Q k =Q0×exp(α×|ω k|k-1 |), where Q0 is the basic process noise covariance matrix, α is a positive adjustment coefficient, and exp() is an exponential function.

[0009] Preferably, the sequence of innovation vectors in a preset time window is calculated to adjust a forgetting factor, and then the weighted measurement noise covariance matrix R is updated online, including: setting a time window of length N to store the innovation vector sequence of the latest N cycles; calculating the kurtosis statistic κ of the innovation vector sequence k ; Calculate the forgetting factor λ of the current cycle through the S-type function k ,λ k =λ min +(λ max -λ min ) / (1+exp(-β(κ k -κ0))), where λ min ,λ max , β, κ0 are all preset constants, exp() is an exponential function; the formula R k =λ k ×R k-1 +(1-λ k )×(e k ×e k T ) Update the measurement noise covariance matrix R k , where e k is the current innovation vector.

[0010] Preferably, the starting parallel branch processing mechanism includes the following steps: calculating the innovation covariance matrix S k =H k ×P k|k-1 ×H k T +R k , where H k is the Jacobian matrix of the measurement function h for the state x, P {k|k-1} is the prior error covariance matrix, R k is the measurement noise covariance matrix; and calculates the Mahalanobis distance squared of the innovation vector d k 2 =e k T ×S k -1 ×e k , and compare it with the threshold value γ determined according to the chi-square distribution; where e k is the current innovation vector, S k is the innovation covariance matrix; if d k 2 If it exceeds γ, the first branch and the second branch are generated.

[0011] Preferably, one of the branches uses the expanded measurement noise covariance matrix to perform measurement update, and the other branch skips the measurement update of the current cycle, including: the first branch uses the measurement noise covariance matrix R amplified by the expansion factor c>1 k' =c×R k Perform measurement update, the second branch skips the measurement update of the current cycle, and makes its posterior state estimate equal to the prior state estimate.

[0012] Preferably, the evaluating and pruning of low-probability branches based on new measurement values ​​in a subsequent predetermined period includes: in the next control period k+1, based on the new measurement values, calculating Gaussian likelihood function values ​​for the two branches respectively, and selecting the only branch with a higher likelihood value to continue subsequent iterations, and pruning the other branch.

[0013] Preferably, the nonlinear mapping function is a preset S-type or exponential mapping function.

[0014] Preferably, the low-level system is a permanent magnet synchronous motor control system, and the physical state quantity of the controlled object is the stator dq axis current i d with i q The controller parameter to be monitored is the stator resistance R s and dq axis inductance L d 、L q .

[0015] Preferably, the establishment of a nonlinear discrete-time state space model includes: based on the voltage equation of the permanent magnet synchronous motor, and regarding the controller's monitored parameters as a random walk process that changes slowly over time, using the Euler method discretization method to establish a nonlinear state equation and a measurement equation.

[0016] The beneficial effects of the present invention are as follows: the present invention can significantly improve the accuracy and robustness of parameter estimation. By establishing the intrinsic correlation between process noise and the physical state of the system, the state prediction link can better match the actual uncertainty of the system under different working conditions, thereby improving the accuracy of model prediction. At the same time, by continuously analyzing the statistical characteristics of the measurement residual sequence, the time-varying, non-Gaussian measurement noise can be effectively identified and quantified, and the trust weight of the measurement data can be adjusted accordingly, effectively suppressing the impact of random interference on the estimation results. Especially for sudden strong interference or measurement outliers, the parallel processing and evaluation pruning mechanism not only avoids the information loss and stability problems caused by the rough elimination of data, but also prevents the serious pollution of state estimation by abnormal data, ensuring the continuity, reliability and accuracy of parameter monitoring results under complex working conditions and harsh electromagnetic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1A schematic flow chart of a dynamic parameter monitoring method for a low-level system digital controller provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0019] like Figure 1 As shown, the dynamic parameter monitoring method of the low-level system digital controller provided by the embodiment of the present invention specifically includes the following steps:

[0020] S1. Construct an augmented state vector containing the physical state of the controlled object and the parameters to be monitored by the controller, and establish a nonlinear discrete-time state space model.

[0021] Specifically, taking the permanent magnet synchronous motor control system as an example, the physical state quantity of the controlled object is the stator dq axis current i d with i q , the controller parameter to be monitored is the stator resistance R s and dq axis inductance L d 、L q The augmented state vector x is constructed as a column vector containing these five quantities. Based on the voltage equation of the motor, and considering the parameters as a random walk process that changes slowly over time, a discretization method such as the Euler method is used to establish a vector of the form x k =f(x k-1 ,u k-1 )+w k-1 The nonlinear state equation and z k =h(x k )+v k The measurement equation of k-1 is the control input, i.e., the dq axis voltage, z k The output is the dq axis current, f and h are the corresponding nonlinear functions, w k-1 and v k are noise vectors.

[0022] S2. In each control cycle, first, based on the state space model and the posterior state estimate of the previous cycle, the prior state estimate of the current cycle is calculated; then, using a nonlinear mapping function driven by a specific physical state component in the prior state estimate, the process noise covariance matrix Q is calculated online, and combined with the posterior error covariance matrix of the previous cycle, the prior error covariance matrix of the current cycle is calculated.

[0023] Specifically, the prior state estimate x k|k-1 Through the state equation f(x k-1|k-1 ,uk-1 ) is calculated. Then, the current amplitude in the prior state estimation is selected as the driving physical quantity, and the current amplitude is mapped to the process noise covariance matrix Q through a preset S-type or exponential mapping function. k The variance value of the parameter random walk model term in is used to adjust the uncertainty of the model and the Q value when the load increases. Finally, the state transfer matrix F is calculated. k That is, the Jacobian matrix of f for state x, and through the formula P k|k-1 =F k ×P k-1|k-1 ×F k T +Q k The prior error covariance matrix is ​​calculated.

[0024] S3. Obtain a real-time measurement vector and calculate an innovation vector. At the same time, based on the sequence of the innovation vector within a preset time window, adjust a forgetting factor by calculating its kurtosis statistic, and then update the weighted measurement noise covariance matrix R online.

[0025] Specifically, the real-time measurement value z of the dq axis current is obtained from the sensor. k , and calculate the innovation vector e k =z k -h(x k|k-1 ). Maintain a first-in-first-out queue of length N to store the most recent N innovation vectors. In each cycle, calculate the sample kurtosis value of the data sequence in the queue. According to the deviation degree of the kurtosis value from the theoretical kurtosis value of Gaussian distribution, adjust the forgetting factor λ through an exponential decay function. k When the kurtosis deviates significantly from zero, indicating the presence of thick tails or impulse noise, reduce λ k Otherwise, increase λ k Finally, we use the forgetting factor λ k The recursive formula R k =λ k ×R k-1 +(1-λ k )×(e k ×e k T ) to update the measurement noise covariance matrix R.

[0026] S4. Calculate the Mahalanobis distance of the innovation vector and compare it with a preset threshold. If the Mahalanobis distance exceeds the preset threshold, initiate a parallel branch processing mechanism, wherein one branch uses the expanded measurement noise covariance matrix for measurement update, and the other branch skips the measurement update of the current cycle. In subsequent predetermined cycles, the low-probability branch is evaluated and pruned based on the new measurement value.

[0027] Specifically, calculate the innovation covariance matrix S k =H k ×P k|k-1 ×H k T +R k , where H k is the Jacobian matrix of the measurement function h for the state x. Then calculate the square value of the Mahalanobis distance d k 2 =e k T ×S k -1 ×e k , and compared with the threshold γ set based on the chi-square distribution. k 2 If it is greater than γ, the current measurement is determined to be an outlier and two filtering branches are generated: Branch 1, R k Multiplying by a large expansion factor, such as 100, weakens the weight of the abnormal measurement and then performs a normal measurement update. In branch two, the current measurement is completely discarded, and its posterior estimate is directly equal to the prior estimate. Over the following M control cycles, the two branches are independently filtered, and the likelihood probability is calculated based on their respective predicted and actual measurements. After M cycles, the branch with the high cumulative likelihood probability is retained, and the branch with the low probability probability is pruned.

[0028] S5. If the Mahalanobis distance does not exceed the preset threshold or after the parallel branch processing is completed, the Kalman gain is calculated using the updated process noise covariance matrix Q and the measurement noise covariance matrix R, and the measurement update is performed to obtain the posterior state estimate and the posterior error covariance matrix of the current cycle.

[0029] Specifically, the steps are standard extended Kalman filter measurement update process. If the Mahalanobis distance is within the limit, or the only surviving branch has been selected in the parallel processing, the state and covariance information of the branch are used to update the K value through the formula K k =P k|k-1 ×H k T ×S k -1 Calculate the Kalman gain. Then, by x k|k =x k|k-1 +K k ×e k Update the state estimate and pass P k|k =(IK k ×H k )×P k|k-1 Update the error covariance matrix to obtain the posterior estimate of the current cycle.

[0030] S6. Extracting the current value of the parameter to be monitored from the posterior state estimate of the current cycle.

[0031] Specifically, after completing the measurement update, the posterior state estimate vector x k|k is a column vector containing five elements. According to the definition of the augmented state vector in the first step, directly reading the third, fourth and fifth elements of the vector can respectively obtain the stator resistance R in the current cycle. s , d-axis inductance L d and q-axis inductance L q The best estimate of .

[0032] In an optional embodiment, the step of constructing an augmented state vector including the physical state of the controlled object and the controller parameter to be monitored includes: cascading the physical state vector x_p of the controlled object and the controller parameter vector x_θ to be monitored to form an augmented state vector x=[x p T , x θ T ] T , where T represents the transpose operation and the controller parameters are modeled as a random walk process. Specifically, in the permanent magnet synchronous motor control application, the physical state vector x p It can include four variables such as dq axis current and motor speed. At the same time, if it is necessary to identify the motor stator resistance and inductance, two parameters that vary with temperature or working conditions, online, then the controller parameter vector x θ These two parameters are included. Through cascading operations, a six-dimensional augmented state vector x is formed. This enables the Kalman filter to not only estimate the motor's operating state but also track and correct controller parameter drift caused by factors such as heating in real time, thereby improving the robustness of the control system.

[0033] In an optional embodiment, a nonlinear mapping function driven by a specific physical state component in the prior state estimate is used to calculate the process noise covariance matrix Q online, including: extracting the prior state estimate x k|k-1 The motor speed component ω in k|k-1 , calculate the process noise covariance matrix Q of the current cycle through the preset exponential function k , Q k =Q0×exp(α×|ω k|k-1|), where Q0 is the basic process noise covariance matrix, α is a positive adjustment coefficient, and exp() is an exponential function. Specifically, when the motor is running at high speed, the uncertainty of its system model usually increases significantly, for example, mechanical vibration and electrical harmonics will be stronger. This method realizes dynamic compensation of model uncertainty by associating the process noise covariance matrix with the motor speed. For example, the basic process noise covariance matrix Q0 is set to a diagonal matrix, the adjustment coefficient α is 0.01, and when the motor speed increases from static to 3000 revolutions per minute, the exponential function will make Q k The value of increases dynamically, which makes the filter trust the new measurement data more in the high-speed area, thereby effectively suppressing the divergence of state estimation and ensuring the estimation accuracy and stability in a wide speed range.

[0034] In an optional embodiment, based on the sequence of the innovation vector in a preset time window, a forgetting factor is adjusted by calculating its kurtosis statistic, and then a weighted measurement noise covariance matrix R is updated online, including: setting a time window of length N to store the innovation vector sequence of the latest N cycles; calculating the kurtosis statistic κ of the innovation vector sequence k ; Calculate the forgetting factor λ of the current cycle through the S-type function k ,λ k =λ min +(λ max -λ min ) / (1+exp(-β(κ k -κ0))), where λ min ,λ max , β, κ0 are all preset constants, exp() is an exponential function; the formula R k =λ k ×R k-1 +(1-λ k )×(e k ×e k T ) Update the measurement noise covariance matrix R k , where e k is the current innovation vector. Specifically, this method aims to diagnose and suppress measurement anomalies in real time, such as transient sensor failure or strong electromagnetic interference. Kurtosis is an indicator that measures the degree of peak distribution of data. When the measurement system is normal, the innovation vector sequence is approximately Gaussian distributed, and its kurtosis value is around 3. Once an outlier or outlier appears, the kurtosis value will increase sharply. For example, when the time window N is set to 50, when the calculated kurtosis statistic κ k When the value jumps from 3 to 8, the sigmoid function will forget the factor λ. k Rapidly increase from a low value such as 0.97 to close to λ max 0.995. This makes the update R k Time, History Rk-1 The weight of is much greater than the weight of the current abnormal innovation term, which effectively blunts the impact of the measurement outlier on the state estimation and ensures the robustness of the filter.

[0035] In an optional embodiment, starting the parallel branch processing mechanism includes the following steps:

[0036] a) Calculate the innovation covariance matrix S k =H k ×P k|k-1 ×H k T +R k , where H k is the Jacobian matrix of the measurement function h for the state x, P k|k-1 is the prior error covariance matrix, R k is the measurement noise covariance matrix; and calculates the Mahalanobis distance squared of the innovation vector d k 2 =e k T ×S k -1 ×e k , and compare it with the threshold value γ determined according to the chi-square distribution; where e k is the current innovation vector, S k is the innovation covariance matrix;

[0037] b) If d k 2 If it exceeds γ, the first branch and the second branch are generated. The first branch uses the measurement noise covariance matrix R amplified by the expansion factor c>1 k' =c×R k Perform measurement update, the second branch skips the measurement update of the current cycle, and makes its posterior state estimate equal to the prior state estimate;

[0038] c) In the next control cycle k+1, based on the new measurement value, the Gaussian likelihood function values ​​are calculated for the two branches respectively, and the only branch with the higher likelihood value is selected to continue the subsequent iterations, and the other branch is pruned.

[0039] Specifically, the parallel branch processing mechanism is used to process measurement data with high uncertainty to avoid filter divergence caused by a single misjudgment. Assuming the measurement dimension is 2, the chi-square distribution table can be used to set the threshold γ to 9.21, corresponding to a 99% confidence level. When the reading of a current sensor suddenly jumps, resulting in the Mahalanobis distance squared d k 2 The calculated value is 15, which exceeds the threshold and starts the branch. The first branch assumes that this is a valid data with large noise, and R kThe prediction is multiplied by the expansion factor c, for example, 5, and then corrected. The second branch assumes the data is completely invalid and discards it directly. At the next sampling moment, new measurements are used to evaluate which of the two branches best explains the current situation. For example, if the prediction of the first branch is more consistent with the new measurement and its Gaussian likelihood value is higher, the result of the first branch is adopted and the calculation of the second branch is terminated, thus achieving prudent handling of abnormal events and rapid recovery.

[0040] The implementation principle of the dynamic parameter monitoring method of the low-level system digital controller of the embodiment of the present invention is: by establishing the intrinsic correlation between process noise and the physical state of the system, the state prediction link can better match the actual uncertainty of the system under different working conditions, thereby improving the accuracy of model prediction. In addition, by continuously analyzing the statistical characteristics of the measurement residual sequence, the time-varying, non-Gaussian measurement noise can be effectively identified and quantified, and the trust weight of the measurement data can be adjusted accordingly, effectively suppressing the impact of random interference on the estimation results. Moreover, for sudden strong interference or measurement outliers, the parallel processing and evaluation pruning mechanism avoids the information loss and stability problems caused by the rough elimination of data, and also prevents the serious pollution of abnormal data on the state estimation. Even in complex working conditions and harsh electromagnetic environments, the parameter monitoring results are still reliable.

[0041] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for monitoring dynamic parameters of a low-level system digital controller, characterized in that: The following steps are involved: Construct an augmented state vector containing the physical state of the controlled object and the parameters to be monitored by the controller, and establish a nonlinear discrete-time state space model; In each control cycle, the prior state estimate of the current cycle is calculated based on the state space model and the posterior state estimate of the previous cycle; The process noise covariance matrix Q is calculated using a nonlinear mapping function driven by a specific physical state component in the prior state estimate. The prior error covariance matrix of the current cycle is calculated by combining the posterior error covariance matrix of the previous cycle. Acquire a real-time measurement vector and calculate an innovation vector. Based on the sequence of innovation vectors within a preset time window, calculate their kurtosis statistics to adjust a forgetting factor, and then update the weighted measurement noise covariance matrix R online. Calculate the Mahalanobis distance of the innovation vector and compare it with a preset threshold. If the threshold is exceeded, a parallel branch processing mechanism is initiated. One branch uses the expanded measurement noise covariance matrix for measurement update, while the other branch skips the measurement update in the current cycle. In subsequent predetermined cycles, low-probability branches are evaluated and pruned based on new measurement values. If the Mahalanobis distance does not exceed the preset threshold or after the parallel branch processing is completed, the Kalman gain is calculated using the updated process noise covariance matrix Q and the measurement noise covariance matrix R, and the measurement update is performed to obtain the posterior state estimate and the posterior error covariance matrix of the current cycle; the current value of the parameter to be monitored is extracted from the posterior state estimate of the current cycle.

2. The dynamic parameter monitoring method of the low-level system digital controller according to claim 1 is characterized in that: The construction of the augmented state vector including the physical state quantity of the controlled object and the controller's monitored parameters includes: p and the controller parameter vector x to be monitored θ Cascade to form an augmented state vector x=[x p T , x θ T ] T , where T represents the transpose operation and the controller parameters are modeled as a random walk process.

3. The dynamic parameter monitoring method of the low-level system digital controller according to claim 1, characterized in that: The method of calculating the process noise covariance matrix Q by using a nonlinear mapping function driven by a specific physical state component in the prior state estimate includes: extracting the prior state estimate x k|k-1 The motor speed component ω in k|k-1 , calculate the process noise covariance matrix Q of the current cycle through the preset exponential function k , Q k =Q0×exp(α×|ω k|k-1 |), where Q0 is the basic process noise covariance matrix, α is a positive adjustment coefficient, and exp() is an exponential function.

4. The dynamic parameter monitoring method of the low-level system digital controller according to claim 1 is characterized in that: The method is based on the sequence of innovation vectors in a preset time window, and its kurtosis statistics are calculated to adjust a forgetting factor, and then the weighted measurement noise covariance matrix R is updated online, including: setting a time window of length N, storing the innovation vector sequence of the latest N cycles; calculating the kurtosis statistic κ of the innovation vector sequence k ; Calculate the forgetting factor λ of the current cycle through the S-type function k ,λ k =λ min +(λ max -λ min ) / (1+exp(-β(κ k -κ0))), where λ min ,λ max , β, κ0 are all preset constants, exp() is an exponential function; the formula R k =λ k ×R k-1 +(1-λ k )×(e k ×e k T ) Update the measurement noise covariance matrix R k , where e k is the current innovation vector.

5. The dynamic parameter monitoring method of the low-level system digital controller according to claim 1 is characterized in that: The starting of the parallel branch processing mechanism includes the following steps: Calculate the innovation covariance matrix S k =H k ×P k|k-1 ×H k T +R k , where H k is the Jacobian matrix of the measurement function h for the state x, P k|k-1 is the prior error covariance matrix, R k is the measurement noise covariance matrix; and calculates the Mahalanobis distance squared of the innovation vector d k 2 =e k T ×S k -1 ×e k , and compare it with the threshold value γ determined according to the chi-square distribution; where e k is the current innovation vector, S k is the innovation covariance matrix; if d k 2 If it exceeds γ, the first branch and the second branch are generated.

6. The dynamic parameter monitoring method of the low-level system digital controller according to claim 5, characterized in that: One of the branches uses the expanded measurement noise covariance matrix to perform measurement update, and the other branch skips the measurement update of the current cycle, including: the first branch uses the measurement noise covariance matrix R amplified by the expansion factor c>1 k' =c×R k Perform measurement update, the second branch skips the measurement update of the current cycle, and makes its posterior state estimate equal to the prior state estimate.

7. The method for monitoring dynamic parameters of a low-level system digital controller according to claim 6, wherein: The evaluating and pruning of low-probability branches based on new measurement values ​​in a subsequent predetermined cycle includes: in the next control cycle k+1, based on the new measurement values, respectively calculating Gaussian likelihood function values ​​for the two branches, and selecting the only branch with a higher likelihood value to continue subsequent iterations, and pruning the other branch.

8. The method for monitoring dynamic parameters of a low-level system digital controller according to claim 1, wherein: The nonlinear mapping function is a preset S-type or exponential mapping function.

9. The method for monitoring dynamic parameters of a low-level system digital controller according to claim 1, wherein: The low-level system is a permanent magnet synchronous motor control system, and the physical state quantity of the controlled object is the stator dq axis current i d with i q The controller parameter to be monitored is the stator resistance R s and dq axis inductance L d 、L q .

10. The dynamic parameter monitoring method of the low-level system digital controller according to claim 9, characterized in that: The establishment of a nonlinear discrete-time state space model includes: based on the voltage equation of the permanent magnet synchronous motor, and regarding the controller's monitored parameters as a random walk process that changes slowly over time, using the Euler method discretization method to establish a nonlinear state equation and a measurement equation.

Images