An embedded AI-assisted nonlinear system state and fault joint estimation method

Abstract

The application discloses an embedded AI-assisted nonlinear system state and fault joint estimation method. The method establishes a discrete-time nonlinear augmented model containing the object running state, the sensor fault state and the unknown input state, performs extended Kalman filter recursion to obtain innovation, uses the innovation of continuous multiple cycles to construct a timing window, and uses a neural network pre-deployed on an embedded low-computing-power processing platform to generate dynamic process noise covariance parameters and dynamic measurement noise covariance parameters according to the timing window to adaptively adjust the filter gain, and finally outputs the joint estimation result of the system state, the sensor fault and the unknown input. The application can consider the real-time performance, stability and deployability of the low-computing-power embedded platform, realizes the joint estimation of the nonlinear system state and the fault under unknown noise covariance, and provides fault information support for system fault-tolerant control.

Description

Technical Field

[0001] This invention belongs to the field of nonlinear system fault estimation technology, and in particular relates to an embedded AI-assisted method for joint estimation of nonlinear system state and fault. Background Technology

[0002] Nonlinear systems with closed-loop control are widely present in various modern industrial systems. Sensor failures during operation can severely impair system performance and even lead to system instability. Compared to traditional fault detection or isolation, sensor fault estimation aims to dynamically and continuously reconstruct fault signals, thereby providing quantitative underlying information support for fault-tolerant control and performance recovery.

[0003] In real-world operating environments, nonlinear systems are often affected by the coupling effects of unknown inputs (such as time-varying loads and external disturbances), parameter drift, and random noise. Traditional model-driven methods, such as the Extended Kalman Filter (EKF), can achieve state observation of nonlinear systems to a certain extent. However, under complex operating conditions, due to changes in operating conditions and environmental disturbances, the covariance of the system's process noise and measurement noise is usually unknown and varies over time. The traditional EKF uses a fixed covariance matrix, which is difficult to adapt to the dynamic changes in these noise statistical characteristics, and is prone to filter gain misalignment, estimation bias, and even divergence.

[0004] With the development of artificial intelligence technology, combining deep learning technology with filtering algorithms based on physical models has become an effective way to improve the estimation accuracy and robustness of nonlinear systems. By extracting features from real-time operating data through neural networks and dynamically adjusting the noise covariance of the filter, nonlinear disturbances can be adaptively captured while preserving the interpretability of the physical model.

[0005] However, existing algorithms of this kind are mostly adapted to high-performance computing platforms with ample computing power, without fully considering the hardware deployment constraints in industrial settings. In actual measurement and control devices and industrial actuators, the main control unit (such as a DSP (Digital Signal Processor)) is severely limited by computing power, power consumption, and high real-time sampling rate, making it difficult to support online inference of complex neural networks. Therefore, there is a need for an intelligent joint estimation scheme for sensor faults in nonlinear systems that can effectively handle unknown inputs and time-varying noise covariance and can be executed in real time on low-computing-power embedded platforms. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the prior art and propose an embedded AI-assisted joint estimation method for the state and faults of nonlinear systems. This method can achieve joint estimation of system state and faults, and the neural network used satisfies both interpretability and low-computing-power deployment.

[0007] To achieve the above objectives, this invention proposes an embedded AI-assisted joint estimation method for the state and faults of nonlinear systems, comprising the following steps: S1. Establish a discrete-time nonlinear augmented model that includes the object's running state, sensor fault state, and unknown input state; S2. Based on the nonlinear augmented model, perform recursive estimation to complete state prediction, covariance prediction, measurement prediction, and innovation calculation. S3. Construct and update the new information timing window using new information from multiple consecutive cycles; S4. Based on the innovation timing window and the pre-trained neural network model deployed on the embedded low-computing-power processing platform, generate dynamic process noise covariance parameters and dynamic measurement noise covariance parameters, and adjust the prior covariance matrix and innovation covariance matrix in the state prediction stage and the posterior update stage respectively according to the dynamic process noise covariance parameters and dynamic measurement noise covariance parameters to adjust the extended Kalman filter gain. S5. Based on the extended Kalman filter gain and innovation information, perform posterior state and covariance updates, and output system state estimation, sensor fault estimation and unknown input estimation results.

[0008] Preferably, in step S1, the augmented state vector includes the object's operating state, sensor fault state, and unknown input state; the sensor fault state and unknown input state are described using a slowly changing model, a random walk model, or a preset evolution model; the augmented model includes an augmented state transition function and an augmented measurement function, the augmented state transition function being jointly constructed by the original system state transition function, the fault state evolution function, and the unknown input state evolution function. This preferred scheme, by unifying the augmented modeling of the object's operating state, sensor fault state, and unknown input state, and using a slowly changing model or a random walk model to describe the evolution of faults and loads, ensures the completeness of the joint estimation of state and fault, providing an accurate model foundation for subsequent recursive estimation.

[0009] Preferably, step S2 includes: performing state prediction using the posterior augmented state estimate from the previous sampling period; calculating the augmented state transition Jacobian matrix and performing covariance prediction based on the Jacobian matrix and the posterior covariance matrix from the previous sampling period; performing measurement prediction on the predicted state based on the augmented measurement function; and calculating the difference between the actual measured value and the measurement prediction value as an innovation vector. This preferred scheme, through the prediction-update recursive structure of the standard extended Kalman filter, can achieve recursive estimation of the augmented state using model and measurement information without directly measuring faults and loads, and generate an innovation vector containing information about system malfunctions.

[0010] Preferably, in step S3, the innovation timing window is a fixed-length, multi-channel timing window used to store the innovation vectors of the most recent consecutive sampling periods; when the window is not full, it is filled with default values; a circular buffer mechanism is used for window updates; when the number of effective innovation frames in the innovation timing window does not reach a preset threshold, the default static covariance parameter is used to perform filtering recursion, and after the window is full, it switches to the neural network dynamic covariance adjustment mode. This preferred scheme, by constructing a fixed-length, multi-channel innovation timing window, enables the neural network to perceive the temporal evolution of innovation rather than relying solely on the instantaneous value of a single frame; the circular buffer mechanism reduces the computational complexity of single-frame updates; and the default covariance transition strategy before window establishment ensures the numerical stability of the system at startup.

[0011] Preferably, step S4 specifically includes: inputting the innovation timing window into a pre-trained neural network model deployed on an embedded low-computing-power processing platform; outputting process noise adjustment parameters and measurement noise adjustment parameters from the neural network model; subjecting the process noise adjustment parameters and measurement noise adjustment parameters to preset range constraints to obtain dynamic process noise covariance parameters and dynamic measurement noise covariance parameters; correcting the prior covariance matrix based on the dynamic process noise covariance parameters, and constructing the innovation covariance matrix based on the dynamic measurement noise covariance parameters; and applying the corrected prior covariance matrix... The difference matrix and the innovation covariance matrix are used to calculate the extended Kalman filter gain. The dynamic process noise covariance parameter is only used to adjust the diagonal elements of the prior covariance matrix corresponding to the object's operating state in the augmented state vector. The diagonal elements of the prior covariance matrix corresponding to the sensor fault state and the unknown input state do not participate in the dynamic adjustment. The object's operating state includes d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angle. The sensor fault state includes A-phase current sensor bias fault and B-phase current sensor bias fault. The unknown input state includes load disturbance. In this preferred scheme, features are extracted from the innovation timing window through a neural network and the covariance adjustment parameter is dynamically output. Then, bounded mapping is used to ensure that the covariance parameter is always within a reasonable range, so that the filter gain can be adaptively adjusted according to the changes in operating conditions. At the same time, only the operating state is dynamically adjusted while keeping the basic covariance of the fault and load states unchanged. This ensures that the evolution of the fault and load states depends only on the continuous correction of model assumptions and measurement information, avoiding excessive intervention in fault estimation.

[0012] Preferably, the neural network model is a lightweight neural network model pre-trained offline based on sample data. After parameter fixing, the neural network model is deployed in an embedded low-computing-power processing platform. During online estimation, model training is not performed; only output parameters for adjusting the extended Kalman filter noise covariance are generated based on the innovation time-series window. This preferred solution significantly reduces the computational burden during online execution through pre-offline training and parameter fixing, enabling real-time execution of neural network inference on a low-computing-power DSP platform. Simultaneously, the strategy of not training online ensures the determinism and stability of the estimation process.

[0013] Preferably, the method runs on an embedded low-computing-power processing platform and employs an asynchronous execution structure that decouples the front-end real-time filtering chain from the back-end neural network inference chain. The extended Kalman filter recursion for the current sampling period uses the dynamic process noise covariance parameter and dynamic measurement noise covariance parameter from the previous published time. The information from the current sampling period is used to generate the dynamic process noise covariance parameter and dynamic measurement noise covariance parameter for subsequent sampling periods. This preferred scheme, by decoupling the front-end filtering recursion from the back-end neural network inference in time, ensures that the extended Kalman filter recursion is not blocked by the neural network inference, thereby meeting the high real-time requirements of the low-computing-power embedded platform while maintaining estimation accuracy.

[0014] Preferably, step S5 specifically includes: weighting and correcting the innovation information using the Kalman gain to obtain a posterior state estimate; updating the prior covariance matrix based on the Kalman gain, the measurement Jacobian matrix, and the measurement noise covariance to obtain a posterior covariance matrix; performing symmetry processing and positive definite lower bound protection on the posterior covariance matrix; extracting the system state estimate, sensor fault estimate, and unknown input estimate from the posterior state estimate, with the sensor fault estimate output as a continuous fault evolution quantity. This preferred scheme achieves optimal estimation of the augmented state through weighted correction of the innovation using the Kalman gain, effectively suppresses numerical divergence through symmetry processing and positive definite lower bound protection of the covariance matrix, and outputs the sensor fault estimate as a continuous evolution quantity, which can intuitively reflect the development trend and relative amplitude of the fault, providing quantitative fault information support for fault-tolerant control.

[0015] Preferably, the posterior covariance update adopts the Joseph-stable form to maintain the positive semi-definiteness of the covariance matrix in a finite-word-length floating-point arithmetic environment; the symmetry processing averages the posterior covariance matrix with its transpose; and the positive definite lower bound protection restricts the diagonal elements of the posterior covariance matrix to above a preset positive threshold. This preferred scheme, by structurally guaranteeing the positive semi-definiteness of the covariance matrix through the Joseph-stable form, combined with symmetry processing and positive definite lower bound protection, effectively suppresses the non-positive definite divergence of the covariance matrix caused by the accumulation of rounding errors in a finite-word-length floating-point arithmetic environment, significantly improving the numerical robustness of the filtering algorithm.

[0016] This invention also proposes an embedded AI-assisted joint estimation system for the state and faults of a nonlinear system, comprising: The data input module is used to receive the measured quantities and control inputs of the object; The augmented state real-time estimation module is used to perform extended Kalman filter recursion at fixed periods and output state estimates, fault estimates, load estimates and information about the current sampling period; The news timing window module is used to store news information for multiple consecutive sampling periods and update it in a circular buffer manner. The neural network covariance adjustment module is used to output dynamic process noise covariance parameters and dynamic measurement noise covariance parameters according to the innovation timing window; The covariance exchange module is used to publish the dynamic process noise covariance parameters and the dynamic measurement noise covariance parameters to the augmented state real-time estimation module for use in subsequent sampling periods. The background task scheduling module is used to asynchronously trigger the inference calculation of the neural network covariance adjustment module while the augmented state real-time estimation module performs filtering recursion.

[0017] The beneficial effects of this invention are: This invention ensures the completeness of joint state and fault estimation by augmenting the nonlinear operating state, sensor fault state, and disturbance state. Furthermore, it combines an information window-driven neural network to dynamically adjust the process noise covariance and measurement noise covariance, thereby adapting to changes in operating conditions and effectively improving the estimation accuracy and robustness in complex environments. At the same time, it adopts an asynchronous architecture that decouples real-time estimation in the foreground and neural network inference in the background to dynamically adjust the covariance parameter, which significantly enhances the system's deployability and real-time performance on low-computing-power embedded platforms.

[0018] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0019] Figure 1This is a flowchart of the embedded AI-assisted joint estimation method for nonlinear system state and fault described in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the state estimation, fault estimation, and load estimation effects of a permanent magnet synchronous motor under the condition of bias fault of phase A current sensor in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the state estimation, fault estimation, and load estimation effects of a permanent magnet synchronous motor under the condition of a B-phase current sensor bias fault in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the state estimation, fault estimation, and load estimation effects of a permanent magnet synchronous motor under the condition of simultaneous bias fault of current sensors in phases A and B in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0021] See Figure 1 This invention addresses the issues of time-varying noise statistical characteristics and coupling between load disturbances and sensor bias faults under complex operating conditions. It proposes an embedded AI-assisted joint estimation method for the state and faults of nonlinear systems, with the following specific steps: S1. Establish a discrete-time nonlinear augmented model that includes the object's operating state, sensor fault state, and unknown input state: Consider the following discrete-time nonlinear system: (1) in, , and These represent the system status, system control input, and measurement output, respectively. The problem is a sensor malfunction in the system. and The noise is uncorrelated, zero-mean Gaussian white noise. It's process noise. The measurement noise, process noise covariance matrix, and measurement noise covariance matrix are respectively... and . It is a discrete-time nonlinear state transition function. For measurement functions. , and These are the allocation matrices for unknown inputs or external disturbances, process noise, and sensor faults, respectively. , Assignment matrix for control inputs and unknown inputs.

[0022] In this embodiment, the discrete-time model of a permanent magnet synchronous motor with a current sensor fault under time-varying conditions, where the load disturbance is unknown, the noise covariance matrix is ​​unknown, is written in the form shown in (1), where , , , These represent the system state, control input, and unknown input of the permanent magnet synchronous motor, respectively. , , , These represent the stator current, rotor mechanical angular velocity, and mechanical angle in the dq coordinate system, respectively. It is the d-axis stator current. It is the q-axis stator current. and These are the stator voltages in the dq coordinate system, respectively. It is the d-axis stator voltage. It is the q-axis stator voltage. This indicates load disturbance.

[0023] It should be noted that the sensors of the permanent magnet synchronous motor in this embodiment mainly include sensors for measuring the current of phase A and phase B. Therefore and These represent unknown bias faults in phase A and phase B current sensors, respectively. The measurement equations are composed of the corresponding current states, the equivalent bias of the fault in the dq domain, and the measurement noise. The measurement vector is: (2) Based on the state equation of the augmented dynamic model of the permanent magnet synchronous motor, it is discretized using a fixed sampling period Ts with first-order Euler discretization. The discretized state equation is as follows: (3) In the formula, , , , , , and These represent the number of pole pairs, permanent magnet flux linkage, moment of inertia, coefficient of friction, d-axis stator inductance, q-axis stator inductance, and stator resistance of the permanent magnet synchronous motor, respectively. The motor operating state in the k-th sampling period is: (4) In the formula, It is zero-mean Gaussian white noise. This is noise in the d-axis stator current measurement. This is q-axis stator current measurement noise. It is noise from mechanical angular velocity measurement. and Fault biasing of phase A and B current sensors , The equivalent fault offset mapped to the dq coordinate system through coordinate transformation is given in the following equation (7). The fault state and load state adopt a random walk form, which can be expressed as: (5) In the formula , and These represent the process noise corresponding to fault-related augmented states and load-related augmented states, respectively, thus describing the uncertainty and random fluctuations in the aforementioned slowly changing process.

[0024] To accurately reflect the impact of current sensor faults in the filter's measurement prediction model, a coordinate transformation is used to map the fault bias in the abc domain to the dq rotating coordinate system, thereby establishing a nonlinear measurement equation incorporating the fault effect. The three-phase fault bias is mapped to... In a stationary coordinate system, the equivalent components of the fault offset in a two-phase stationary coordinate system are: (6) After further Park transformation, the equivalent fault bias in the dq rotating coordinate system is: (7) Based on the above dq-domain mapping results of the fault bias, a nonlinear measurement equation incorporating the fault effect is constructed. The nonlinear measurement equation is defined as follows: (8) In the formula, the predicted values ​​of the stator currents on the d-axis and q-axis are equal to the corresponding current states plus the equivalent fault bias in the dq domain, thus directly embedding the influence of the current sensor fault into the measurement model; the predicted values ​​of the mechanical angular velocity and mechanical angular position are directly equal to the corresponding state values. Taking the first-order partial derivative of the nonlinear measurement function shown in equation (8) with respect to the state vector, the measurement Jacobian matrix is ​​obtained, and its key partial derivative term is: (9) Therefore, the complete measurement Jacobian matrix can be written as: (10) The partial derivative terms in equation (10) are designed to assist in the extended Kalman filter in determining the source of bias.

[0025] In the update process, by correctly identifying whether the current measurement deviation is caused by changes in the actual state or by sensor failure, the decoupled estimation of the operating state and the fault state is achieved.

[0026] S2. Based on the aforementioned nonlinear augmented model, perform recursive estimation to complete state prediction, covariance prediction, measurement prediction, and innovation calculation: Based on the augmented dynamic model, an extended Kalman filter is used to recursively estimate the motor state, fault state, and load state to obtain innovation information. At each sampling time, the extended Kalman filter operates according to its standard "prediction-update" recursive structure: state prediction, covariance prediction, and measurement prediction, thereby calculating the innovation vector.

[0027] Using the posterior state estimate from the previous time step and the discretized state equation, a one-step prediction of the current state is performed. The state prediction equation is: (11) In equation (11), This is the posterior state estimate from the previous time step. The control input vector from the previous time step. It is formed by adding random walk to the discrete equation (first 4 dimensions) of equation (3).

[0028] The posterior estimation error covariance matrix from the previous time step is propagated to obtain the prior covariance matrix for the current time step, which is used to quantify the uncertainty of the prior state prediction. The covariance prediction equation is as follows: (12) in The basic process noise covariance matrix represents the dynamic increment of the previous cycle. It can be taken as a diagonal matrix: (13) The basic process noise covariance matrix shown in equation (13) The adjustment amount of the superimposed neural network dynamic output is used to inject uncertainty caused by modeling uncertainty and changes in operating conditions during the prediction stage; the diagonal elements corresponding to the A-phase current sensor bias fault, the B-phase current sensor bias fault and the load disturbance are kept as base values ​​and do not participate in dynamic adjustment, so as to ensure that the evolution of fault and load state depends only on the model random walk assumption and the continuous correction of measurement information.

[0029] Based on the prior state prediction value, it is mapped to the measurement space through a nonlinear measurement function to obtain the measurement prediction value at the current moment. The measurement prediction equation is: (14) Substituting the prior state predictions into the nonlinear function, the resulting measurement predictions reflect the expected outputs of each measurement channel under the current prior state estimation. The innovation vector can be defined as: (15) In the formula, each element corresponds to the difference between the actual measured value and the model prediction value on the four measurement channels: d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angular position.

[0030] S3. Construct and update the news timing window using news information from multiple consecutive periods: In step S2 above, a new information vector containing four-dimensional channel information is generated at each sampling time. In order for the neural network in step S4 to perceive the temporal evolution of the new information rather than relying solely on the instantaneous value of a single frame, the new information vectors from multiple consecutive sampling times are organized into a fixed-length multi-channel temporal window, and an efficient update is achieved using a circular buffer.

[0031] Write the current information vector into a fixed-length window, denoted as W. Then the current window is: (16) When the window is not full, it is filled with the default value of zero. This applies to the number of valid frames. Initially, the system will operate using the default static covariance parameters. Once the window is fully established, it will switch to the neural network dynamic covariance adjustment mode. This strategy ensures both numerical stability at startup and the ability for subsequent dynamic adaptation.

[0032] By using a circular buffer mechanism to update the information window, the write operation of each frame of information only requires one assignment and one modulo operation, without moving historical data, thereby reducing the computational complexity of updating the information window per frame and effectively adapting to the real-time requirements of low-computing-power DSP platforms.

[0033] S4. Based on the aforementioned information timing window and the pre-trained neural network model deployed on the DSP platform, generate dynamic process noise covariance parameters and dynamic measurement noise covariance parameters, and adjust the extended Kalman filter gain accordingly.

[0034] The information window is fed as input into a lightweight neural network model that has been pre-trained offline and deployed on a DSP platform. The network outputs dynamic process noise covariance parameters and dynamic measurement noise covariance parameters, which are used to adaptively adjust the gain matrix of the extended Kalman filter in the subsequent step S5.

[0035] New information window Input Neural Network The dynamic covariance parameters at the current moment are obtained: (17) The neural network output first yields the original output vector: (18) The first four elements are used to generate the diagonal elements of the dynamic process noise covariance, corresponding to the d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angle, respectively. The last four elements are used to generate the diagonal elements of the dynamic measurement noise covariance, corresponding to the four measurement channels: d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angular position, respectively. After bounded mapping, we obtain: (19) (20) in and It is a bounded mapping function.

[0036] The covariance diagonal elements are obtained by mapping the quantized output of the neural network, and the output is constrained within a preset physical range to ensure that the covariance parameter is positive and does not exceed a reasonable range, thereby avoiding filter divergence caused by extreme values. Then, the dynamic process noise covariance matrix is ​​constructed: (twenty one) In equation (21), the first four diagonal elements are dynamically output by the neural network, corresponding to the process noise adjustment amounts for the four operating states: d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angle, respectively. The last three diagonal elements are set to zero, indicating that the A-phase current sensor bias fault, B-phase current sensor bias fault, and load disturbance do not participate in the dynamic adjustment, ensuring that the evolution of the A-phase current sensor bias fault state, B-phase current sensor bias fault state, and load disturbance state is only driven by the model random walk assumption and measurement updates. The total measurement noise covariance is: (twenty two) The basic measurement noise covariance matrix and the dynamic measurement noise covariance matrix are as follows: (twenty three) Calculate the new information covariance matrix: (twenty four) The Kalman gain is: (25) When the neural network detects an anomalous pattern within the innovation time window, it automatically amplifies the values ​​of the corresponding diagonal elements in the prior covariance matrix by increasing the process noise covariance adjustment for that state. This allows the filter to assign higher trust weights to the measurement information in subsequent gain calculations. Conversely, when the system is running smoothly, the neural network outputs a smaller adjustment, and the filter tends to trust the model's predictions more. This achieves adaptive filtering capability while preserving the complete physical structure of the extended Kalman filter.

[0037] S5: Based on the extended Kalman filter gain and innovation information, perform posterior state and covariance updates, and output system state estimates, sensor fault estimates, and unknown input estimates. The posterior state is updated to: (26) The posterior covariance update uses the Joseph stable form: (27) in Compared to the standard form Joseph form by Symmetrically applied to both sides of the prior covariance matrix, and the positive definite contribution term of the measurement noise after gain transformation is explicitly accumulated. This structure ensures the positive semidefiniteness of the posterior covariance matrix, thereby effectively suppressing the problem of nondefinite divergence of the covariance matrix due to the accumulation of rounding errors in a finite word length floating-point arithmetic environment.

[0038] After the update is complete, symmetry transformation and positive definite lower bound protection are performed on the posterior covariance matrix: (28) (29) in This is a positive definite lower bound protection threshold to prevent diagonal elements from approaching zero due to numerical degradation. After completing the above posterior update, from the posterior state vector... The output quantity is extracted from it. The posterior state vector is: (30) The first four components of equation (30) represent the operating state estimation of the permanent magnet synchronous motor, while the last three components represent the fault and load state estimation. Fault and load estimates are extracted from the state vector: , , (31) Under normal operating conditions with no faults and It fluctuates around zero; when the current sensor experiences a bias fault... and It will deviate from zero and converge to the true value of the fault bias.

[0039] To meet the real-time operation requirements of low-computing-power DSP platforms, this embodiment... Figure 1 Based on the illustrated method flow, an engineering execution approach is adopted that decouples the front-end real-time filtering chain from the back-end neural network inference chain. The front-end real-time filtering chain reads the measured quantities and control inputs in the current sampling period, performs extended Kalman filtering recursion using the dynamic process noise covariance parameters and dynamic measurement noise covariance parameters from the previous published time, and outputs state estimation, fault estimation, and load estimation results. The back-end neural network inference chain generates new dynamic process noise covariance parameters and dynamic measurement noise covariance parameters based on the updated information timing window, and publishes them to the front-end real-time filtering chain in subsequent sampling periods through a secure exchange mechanism.

[0040] During system initialization, the innovation timing window is set to the default state, and the system operates using the default covariance parameters until the window is fully established. To improve the numerical stability and operational reliability of filtering recursion on low-computing-power DSP platforms, this embodiment performs symmetry transformation and positive definite lower bound protection on the covariance matrix, normalizes the angle states, and imposes preset range constraints on the neural network output and input innovations.

[0041] To illustrate the effectiveness and feasibility of this invention, the following description, based on the operating results of the permanent magnet synchronous motor and the DSP board, further explains the invention.

[0042] Table 1: Parameters of Permanent Magnet Synchronous Motor

[0043] Table 1 shows the system parameters of the permanent magnet synchronous motor used in the experiment. The control strategy adopted for the permanent magnet synchronous motor is field-oriented control, which uses dual-loop PI (Proportional-Integral) control (speed control loop and current control loop) to achieve mechanical angular velocity tracking control. By setting different mechanical angular velocity reference values, load torque, and fault types, multiple types of neural network samples are obtained. The sample length is selected as T=10000s, and the fault types mainly include ramp signals, constant signals, and sinusoidal signals.

[0044] The software-based method and DSP execution results are compared. The comparison methods include the theoretical software implementation, real-time board-side updates, and board-side covariance updates at equal intervals. To more intuitively compare the two-end estimation results, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are introduced for calculation and analysis. (32) The comparison results are shown in Table 2. The software incremental update has the best computing power results. The errors of the board-end interval update and the incremental update are almost the same, and both can jointly estimate the status and fault.

[0045] Table 2. RMSE and MAE of the estimation results obtained through software-side incremental updates, DSP-side incremental updates, and DSP-side interval updates.

[0046] Figures 2 to 4 The estimation results are shown under the following conditions: bias fault of phase A current sensor, bias fault of phase B current sensor, and simultaneous bias fault of phase A and phase B current sensors. Figures 2 to 4 It can be seen that under different fault conditions, the present invention can simultaneously output the motor operating status, sensor fault evolution, and load disturbance estimate; the software estimation results and the DSP board operation results have a consistent overall trend, indicating that the method can maintain stable online estimation capability on a low-computing-power DSP platform; for the DSP board operation results, when using interval updates of dynamic covariance parameters, the estimation results maintain good consistency with the real-time update method, effectively improving the onboard operation effectiveness of the method while ensuring estimation accuracy, further demonstrating that the method effectively ensures the efficiency and accuracy of fault feature estimation.

[0047] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A method for joint estimation of state and fault of a nonlinear system with embedded AI assistance, characterized in that... Includes the following steps: S1. Establish a discrete-time nonlinear augmented model that includes the object's operating state, sensor fault state, and unknown input state. The nonlinear system is a system whose system state evolves through a discrete-time nonlinear state transition function, whose measured output is associated with the system state through a measurement function, and which is affected by process noise and measurement noise. The nonlinear augmented model includes an augmented state vector, which includes the object's operating state, sensor fault state, and unknown input state. The object's operating state includes the d-axis stator current, q-axis stator current, rotor mechanical angular velocity, and mechanical angle. The sensor fault state includes the A-phase current sensor bias fault and the B-phase current sensor bias fault. The unknown input state includes load disturbance. S2. Based on the nonlinear augmented model, perform recursive estimation to complete state prediction, covariance prediction, measurement prediction, and innovation calculation. S3. Construct and update the new information timing window using new information from multiple consecutive cycles; The innovation timing window is a fixed-length, multi-channel timing window used to store the innovation vectors of the most recent consecutive sampling periods; when the window is not full, it is filled with the default value; a circular buffer mechanism is used to update the window; when the number of effective innovation frames in the innovation timing window does not reach the preset threshold, the default static covariance parameter is used to perform filtering recursion, and after the window is full, it switches to the neural network dynamic covariance adjustment mode. S4. Based on the innovation timing window and the pre-trained neural network model deployed on the embedded low-computing-power processing platform, generate dynamic process noise covariance parameters and dynamic measurement noise covariance parameters. Then, based on these parameters, correct the prior covariance matrix and innovation covariance matrix in the state prediction stage and the posterior update stage, respectively, to adjust the extended Kalman filter gain. Specifically, step S4 includes: The new information timing window Input Neural Network The dynamic covariance parameters at the current moment are obtained: The neural network output first yields the original output vector: ,in Represents the original output vector The first four elements are used to generate the diagonal elements of the dynamic process noise covariance, corresponding to the d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angle, respectively. Represents the original output vector The last four elements are used to generate the diagonal elements of the dynamic measurement noise covariance, corresponding to the four measurement channels: d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angular position, respectively; after bounded mapping, we obtain: ; ;in, and It is a bounded mapping function; The covariance diagonal elements are obtained by mapping the quantized output of the neural network, and the output is restricted within a preset physical range to ensure that the covariance parameter is positive and does not exceed a reasonable range, thereby avoiding filter divergence caused by extreme values; then, the dynamic process noise covariance matrix is ​​constructed: ,in, The process noise adjustment quantities, dynamically output by the neural network, correspond to four operating states: d-axis stator current, q-axis stator current, mechanical angular velocity, and mechanical angle. The last three diagonal elements are set to zero, indicating that A-phase current sensor bias fault, B-phase current sensor bias fault, and load disturbance do not participate in the dynamic adjustment, ensuring that the evolution of the A-phase current sensor bias fault state, B-phase current sensor bias fault state, and load disturbance state is driven only by the model random walk assumption and measurement updates. The total measurement noise covariance is: The basic measurement noise covariance matrix and the dynamic measurement noise covariance matrix are as follows: ; Calculate the new information covariance matrix: ,in, To measure the Jacobian matrix, The prior covariance matrix, Calculate the total measurement noise covariance matrix; calculate the extended Kalman filter gain: ,in Indicates the extended Kalman filter gain; S5. Based on the extended Kalman filter gain and innovation information, perform posterior state and covariance updates, and output system state estimation, sensor fault estimation and unknown input estimation results.

2. The embedded AI-assisted joint estimation method for the state and fault of a nonlinear system according to claim 1, characterized in that, In step S1, the augmented state vector includes the object's running state, sensor fault state, and unknown input state; the sensor fault state and unknown input state are described by a slowly changing model, a random walk model, or a preset evolution model; the augmented model includes an augmented state transition function and an augmented measurement function, and the augmented state transition function is composed of the original system state transition function, the fault state evolution function, and the unknown input state evolution function.

3. The embedded AI-assisted joint estimation method for nonlinear system state and fault according to claim 2, characterized in that, Step S2 includes: performing state prediction using the posterior augmented state estimate from the previous sampling period; calculating the augmented state transition Jacobian matrix and performing covariance prediction based on the augmented state transition Jacobian matrix and the posterior covariance matrix from the previous sampling period; performing measurement prediction on the predicted state based on the augmented measurement function; and calculating the difference between the actual measured value and the measured predicted value as the innovation vector.

4. The embedded AI-assisted joint estimation method for nonlinear system state and fault according to claim 1, characterized in that, The neural network model is a lightweight neural network model pre-trained offline based on sample data. After the neural network model is fixed with parameters, it is deployed on an embedded low-computing-power processing platform. During the online estimation process, model training is not performed. Instead, output parameters for adjusting the noise covariance of the extended Kalman filter are generated based on the innovation time series window.

5. The embedded AI-assisted joint estimation method for the state and fault of a nonlinear system according to claim 1, characterized in that, The method runs on an embedded low-computing-power processing platform and adopts an asynchronous execution structure that decouples the front-end real-time filtering chain from the back-end neural network inference chain. The extended Kalman filter recursion of the current sampling period uses the dynamic process noise covariance parameter and the dynamic measurement noise covariance parameter of the previous published time. The information of the current sampling period is used to generate the dynamic process noise covariance parameter and the dynamic measurement noise covariance parameter used in subsequent sampling periods.

6. The embedded AI-assisted joint estimation method for the state and fault of a nonlinear system according to claim 1, characterized in that, Step S5 specifically includes: The extended Kalman filter gain is used to weight and correct the innovation information to obtain the posterior state estimate. The prior covariance matrix is ​​updated based on the extended Kalman filter gain, the measurement Jacobian matrix, and the measurement noise covariance to obtain the posterior covariance matrix; The posterior covariance matrix is ​​symmetricized and protected with a positive definite lower bound. The system state estimation result, sensor fault estimation result, and unknown input estimation result are extracted from the posterior state estimation value, and the sensor fault estimation result is output as a continuous fault evolution quantity.

7. The embedded AI-assisted joint estimation method for the state and fault of a nonlinear system according to claim 6, characterized in that, The posterior covariance matrix update adopts the Joseph stable form to maintain the positive semidefiniteness of the covariance matrix in a finite word length floating-point operation environment; the symmetry processing averages the posterior covariance matrix with its transpose; the positive definite lower bound protection restricts the diagonal elements of the posterior covariance matrix to above a preset positive threshold.

8. An embedded AI-assisted joint estimation system for the state and faults of a nonlinear system, characterized in that, The system is used to implement the method of claim 1, including: The data input module is used to receive the measured quantities and control inputs of the object; The augmented state real-time estimation module is used to perform extended Kalman filter recursion at fixed periods and output state estimates, fault estimates, load estimates and information about the current sampling period; The news timing window module is used to store news information for multiple consecutive sampling periods and update it in a circular buffer manner. The neural network covariance adjustment module is used to output dynamic process noise covariance parameters and dynamic measurement noise covariance parameters according to the innovation timing window; The covariance exchange module is used to publish the dynamic process noise covariance parameters and the dynamic measurement noise covariance parameters to the augmented state real-time estimation module for use in subsequent sampling periods. The background task scheduling module is used to asynchronously trigger the inference calculation of the neural network covariance adjustment module while the augmented state real-time estimation module performs filtering recursion.

Images