A method and system for fusing and estimating state variables of in-wheel motor driven electric vehicles
By employing a hierarchical collaborative estimation architecture and an improved robust filtering algorithm, combined with intelligent data fusion rules, the problem of multi-state integrated estimation for in-wheel motor electric vehicles was solved, achieving high-precision and robust state estimation and improving the stability and safety of vehicle control.
Patent Information
- Application Number
- CN202611290255.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-17
AI Technical Summary
Electric vehicles driven by hub motors are difficult to achieve high-precision estimation of multiple states. In particular, the accuracy of traditional methods tends to decrease in complex noise environments, and there is a lack of a collaborative estimation framework for multiple physical states such as vehicle, tires, and motors, resulting in poor system robustness.
A hierarchical collaborative estimation architecture is adopted, which combines an improved robust filtering algorithm and intelligent data fusion rules. Through multi-source data acquisition and preprocessing, multi-state integrated estimation is performed using an improved square root commensurate Kalman filter algorithm and the maximum correlation entropy criterion or adaptive forgetting factor mechanism. Furthermore, a three-stage fusion rule based on error threshold and a dynamic confidence interval based on slip rate-adhesion coefficient based on physical model are designed in the data fusion layer for correction.
It achieves integrated collaborative estimation of states across multiple physical domains, improving estimation accuracy and system stability, enhancing robustness and fault tolerance in complex noise environments, ensuring high reliability and real-time performance of critical state information, and meeting the performance requirements of real-time vehicle control.
Smart Images

Figure CN123116140A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electric vehicle control and state estimation technology, specifically relating to a fusion estimation method and system for state variables of an electric vehicle driven by a hub motor, applicable to electric vehicles using distributed hub motor drive. Background Technology
[0002] Hub motor electric vehicles eliminate the traditional mechanical transmission system, enabling independent control of the torque of each wheel and potentially improving vehicle dynamics. However, this also places higher demands on overall vehicle state perception. Key states such as center of gravity sideslip angle, yaw rate, vehicle speed, and road adhesion coefficient are difficult to measure directly or at low cost, while high-precision sensors are expensive. These states are crucial inputs for stability control and torque distribution, and their estimation accuracy directly affects control performance. Currently, model-based filtering methods are often used for state estimation, but hub motor vehicle systems are complex and highly nonlinear, and noise in the operating environment often exhibits non-Gaussian characteristics. Traditional methods are prone to accuracy degradation when noise changes abruptly or outliers exist. Furthermore, existing methods mostly focus on independent estimation of single states, lacking an integrated collaborative estimation framework for multiple physical states such as vehicle, tires, and motors. This leads to system redundancy, insufficient information utilization, and poor robustness in the event of sensor failure. Therefore, a method and system are needed that can adapt to complex noise, achieve high-precision integrated estimation of multiple states, and integrate multi-source information to improve reliability. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for the fusion estimation of state variables of in-wheel motor driven electric vehicles. This method achieves high-precision and high-reliability integrated estimation of key vehicle states by constructing a hierarchical collaborative estimation architecture, employing an improved robust filtering algorithm, and combining intelligent data fusion rules.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] In a first aspect, the present invention provides a method for fusing and estimating state variables of an electric vehicle driven by a hub motor, comprising the following steps:
[0006] Step S1: Multi-source data acquisition and preprocessing.
[0007] Measurement sensors include current sensors, voltage sensors, motor Hall effect sensors or resolvers, inertial measurement units, satellite navigation systems / real-time dynamic carrier phase differential systems (GNSS / RTK), and angle sensors. Raw data, including the three-phase current of the hub motor, is acquired through the sensor layer. , , With voltage , , Motor speed With rotor angle Vehicle yaw rate Longitudinal acceleration Lateral acceleration lateral angle Front wheel steering angle Heading angle provided by GNSS / RTK ,speed and latitude and longitude The coordinate transformation (Clark / Park transformation) of the motor electrical signal is performed to calculate electromagnetic torque, back electromotive force, etc.; the preliminary value of the vehicle's center of gravity sideslip angle and running trajectory are calculated using GNSS and inertial measurement unit data.
[0008] The measurement data preprocessing involves storing measurement data from the sensor layer, processing the relevant data, and finally sending the stored and processed data to the state estimation layer, model layer, and data fusion layer. The stored data includes: the three-phase current of the hub motor. , , Three-phase voltage of hub motor , , Hub motor load torque Hub motor speed and rotor angle Longitudinal velocity yaw rate latitude and longitude coordinates Heading angle lateral angle Front wheel steering angle The data processed includes: converting the three-phase current of the hub motor. , , and three-phase voltage , , Transform into a stationary coordinate system through coordinate transformation shaft current and voltage , , shaft current and voltage , In a rotating coordinate system shaft current and voltage , , shaft current and voltage , Calculate the three-phase back electromotive force based on the motor parameters. , , and electromagnetic torque By heading angle and yaw angle Calculate the vehicle's center of gravity sideslip angle Using the latitude and longitude coordinates during vehicle operation Calculate vehicle trajectory coordinates ; via hub motor speed and the longitudinal speed of the whole vehicle Calculate tire slip ratio Then, the road surface adhesion coefficient is determined using a physical model that corresponds to the slip ratio and the adhesion coefficient. Interval range.
[0009] Among them, the present invention innovatively develops an algorithm for converting vehicle latitude and longitude coordinates into plane coordinate system coordinates.
[0010] GNSS / RTK sensors used in in-wheel motor-driven vehicles can collect the vehicle's latitude and longitude coordinates during vehicle operation testing. To further study vehicle performance, it is necessary to analyze these latitude and longitude coordinates. Convert to test road direction The axis is perpendicular to the test road direction. shaft Cartesian coordinate system for calibrating vehicle trajectory coordinates The specific algorithm design for its conversion is as follows:
[0011] ① Define the vehicle's initial point as the origin of the coordinate system, corresponding to the origin of the latitude and longitude coordinates. The corresponding point of the Cartesian coordinate system. The vehicle's latitude and longitude coordinates were collected sequentially. .
[0012] ② Longitude value Convert to x-coordinate in a planar coordinate system Values, writing latitude and longitude coordinates as Calculate separately and distance .
[0013] ③ Latitude value Convert to ordinate in a planar coordinate system Values, writing latitude and longitude coordinates as Calculate separately and distance .
[0014] ④ Calculate separately in the latitude and longitude coordinate system To the origin distance They are respectively:
[0015] (1)
[0016] ⑤ Calculate separately in the latitude and longitude coordinate system and the origin Angle towards longitude They are respectively:
[0017] (2)
[0018] ⑥ Measure or calculate the angle between the test road and the longitude line. Calculate the corresponding Cartesian coordinate system coordinates They are respectively:
[0019] (3)
[0020] Step S2: Multi-state integrated collaborative estimation.
[0021] In the state estimation layer, an improved square root capacitive Kalman filter algorithm framework is used as the core estimator. To address the potential non-Gaussian or time-varying characteristics of measurement noise, a maximum correlation entropy criterion or an adaptive forgetting factor mechanism is introduced to dynamically update and correct the measurement noise covariance matrix, thereby enhancing the algorithm's robustness.
[0022] 1. The specific algorithm of the improved square root volume Kalman algorithm is as follows:
[0023] (1) Flowchart of the improved square root volume Kalman algorithm based on the maximum correlation entropy criterion:
[0024] ① Initialize the state vector And error covariance matrix .
[0025] (8)
[0026] ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. .
[0027] ③ Generate according to the volumetric Kalman rule State volume points :
[0028] (9)
[0029] in, The number of state variables. This is a preset set of volume points.
[0030] ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. .
[0031] ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0032] ⑥ Calculate the weighted center matrix of the state variables
[0033] (10)
[0034] ⑦ Weight the center matrix of the state variables With matrix A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. .
[0035] ⑧Use Generate updated state volume points ,
[0036] (11)
[0037] ⑨ Through measurement equations Propagate the updated state volume point , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. .
[0038] ⑩ The measurement noise covariance is updated using the maximum correlation entropy criterion to enhance robustness to non-Gaussian noise. The core of this approach is the introduction of a Gaussian kernel function. Process measurement residuals and dynamically calculate the weight matrix. , where m is the number of measurement variables. Thus, the influence of the measurement noise covariance matrix is equivalently corrected or reconstructed during the iteration process. Item ⑦ in (1) and item 10 Reorganization The matrix is used to update the covariance matrix.
[0039] (12)
[0040] For the measurement noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0041] The computational cost is to measure the weighted central moments and simultaneously update the weighted central matrix of the state variables.
[0042] (13)
[0043] Weighted center matrix of measurement With square root factor A new matrix is constructed, and the square root factor of the measurement error covariance is obtained through QR decomposition. .
[0044] Combine with the updated state quantity weighted center matrix With measurement weighted center matrix Calculate the square root factor of the cross-covariance matrix of state measurements. .
[0045] (14)
[0046] Update the filter gain matrix ,
[0047] (15)
[0048] Update the state matrix ,
[0049] (16)
[0050] Determine the convergence of the state matrix and whether the rate of change η of the state estimate is less than a set threshold ε:
[0051] If η≥ε, then the current Return to step 10 to perform a new measurement iteration;
[0052] If η < ε, then continue with the following steps to update the error covariance matrix.
[0053] matrix and Construct new moments and update the square root factor of the error covariance through QR decomposition. .
[0054] (2) Adaptive algorithm improved square root comprehension Kalman algorithm flow:
[0055] ① Initialize the state vector And error covariance matrix .
[0056] (17)
[0057] ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. .
[0058] ③ Generate according to the volumetric Kalman rule State volume points :
[0059] (18)
[0060] in, The number of state variables. This is a preset set of volume points.
[0061] ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. .
[0062] ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0063] ⑥ Calculate the weighted center matrix of the state variables
[0064] (19)
[0065] ⑦ Weight the center matrix of the state variables With matrix A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. .
[0066] ⑧Use Generate updated state volume points ,
[0067] (20)
[0068] ⑨ Through measurement equations Propagate the updated state volume point , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. .
[0069] ① Initialize the state vector And error covariance matrix .
[0070] (twenty one)
[0071] ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. .
[0072] ③ Generate according to the volumetric Kalman rule State volume points :
[0073] (twenty two)
[0074] in, The number of state variables. This is a preset set of volume points.
[0075] ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. .
[0076] ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0077] ⑥ Calculate the weighted center matrix of the state variables
[0078] (twenty three)
[0079] ⑦ Weight the center matrix of the state variables With matrix A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. .
[0080] ⑧Use Generate updated state volume points ,
[0081] (twenty four)
[0082] ⑨ Through measurement equations Propagate the updated state volume point , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. .
[0083] ⑩ Update the measurement noise covariance using an adaptive filtering formula.
[0084] (25)
[0085] For the measurement noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0086] The computational cost is to measure the weighted central moments and simultaneously update the weighted central matrix of the state variables.
[0087] (26)
[0088] Weighted center matrix of measurement With square root factor A new matrix is constructed, and the square root factor of the measurement error covariance is obtained through QR decomposition. .
[0089] Combine with the updated state quantity weighted center matrix With measurement weighted center matrix Calculate the square root factor of the cross-covariance matrix of state measurements. .
[0090] (27)
[0091] Update the filter gain matrix ,
[0092] (28)
[0093] Update the state matrix ,
[0094] (29)
[0095] matrix and Construct new moments and update the square root factor of the error covariance through QR decomposition. .
[0096] 2. Within the improved square root commensurate Kalman framework, three core estimation algorithm modules are run:
[0097] (1) Vehicle state estimation module: Based on the three-degree-of-freedom vehicle dynamics model and the magic formula tire model, the vehicle dynamics mathematical model and the magic tire formula provide the front and rear axle lateral stiffness for the vehicle state estimation algorithm. , longitudinal velocity Derivative, centroid side deflection angle Derivative, yaw rate Derivative. The measurement data acquisition and processing layer provides the steering wheel angle for the vehicle state estimation algorithm. Lateral acceleration Longitudinal acceleration Measured values. Design a vehicle estimation algorithm for longitudinal velocity. centroid side slip angle yaw rate Make an estimate:
[0098] ① Define the state vector for the vehicle state estimation algorithm ;
[0099] ② Define the input vector for the vehicle state estimation algorithm ;
[0100] ③ Define the measurement vector for the vehicle state estimation algorithm ;
[0101] ④ List the three-degree-of-freedom nonlinear mathematical model of the vehicle.
[0102] (30)
[0103] in For vehicles Moment of inertia of the axis.
[0104] Let the sampling time be Based on the first-order forward Euler method, the nonlinear discrete expression of the vehicle state equation is derived using a three-degree-of-freedom nonlinear mathematical model of the vehicle:
[0105] (31)
[0106] ⑤ Derive the vehicle's lateral kinematics equations
[0107] (32)
[0108] in, This refers to the lateral force exerted by the vehicle's tires.
[0109] The tire's lateral slip characteristics are within the linear range.
[0110] (33)
[0111] in, This refers to the front axle slip angle of the vehicle. This refers to the rear axle slip angle of the vehicle.
[0112] Vehicle front and rear axle slip angles and ,
[0113] (34)
[0114] in, The lateral speed of the vehicle, This is the distance from the vehicle's center of gravity to the front axle. This is the distance from the vehicle's center of gravity to the rear axle.
[0115] Vehicle center of gravity sideslip angle β,
[0116] (35)
[0117] The lateral acceleration is derived from formulas (25-28). equation,
[0118] (36)
[0119] Establish the measurement equations for the vehicle state estimation algorithm:
[0120] (37)
[0121] ⑥ Choose either the maximum correlation entropy square root capacitive Kalman or the adaptive square root capacitive Kalman flow, combined with the state vector. Measurement vector and input vector yaw rate of the state variable , centroid side slip angle Longitudinal speed Make an estimate.
[0122] (2) Road surface adhesion coefficient estimation module: The vehicle dynamics mathematical model and the magic tire formula provide the tire slip ratio for the Dugoff tire mathematical model. and sideslip angle Parameters such as these are provided. The Dugoff tire mathematical model provides longitudinal and lateral tire normalized forces for the road adhesion coefficient estimation algorithm. , The measurement data acquisition and processing layer provides the steering wheel angle for the road surface adhesion coefficient estimation algorithm. yaw rate Lateral acceleration Longitudinal acceleration Equal measured values. Design a road surface adhesion coefficient algorithm, for... Make an estimate:
[0123] ① Define the state vector for the road surface adhesion coefficient estimation algorithm ;
[0124] ② Define the input vector for the road surface adhesion coefficient estimation algorithm. ;
[0125] ③ Define the measurement vector for the road surface adhesion coefficient estimation algorithm. ;
[0126] ④ Assuming the road adhesion coefficient is considered constant over a short time, based on the first-order forward Euler method, the nonlinear discrete expression of the state equation for the road adhesion coefficient is derived:
[0127] (38)
[0128] ⑤ Establish the vehicle dynamics equations for normalized tire forces.
[0129] Write the Dugoff tire model in normalized form.
[0130] (39)
[0131] (40)
[0132] in, For boundary values, For longitudinal stiffness, This refers to the lateral stiffness.
[0133] Longitudinal movement:
[0134] (41)
[0135] Lateral movement:
[0136] (42)
[0137] Lateral motion:
[0138] (43)
[0139] in, , These are the leading front and rear axle distances of the vehicle's center of gravity, respectively. The wheelbase of the vehicle. Let δ be the moment of inertia of the vehicle about its vertical axis. ij This refers to the wheel steering angle.
[0140] Derivation of the measurement equation for the road surface adhesion coefficient estimation algorithm:
[0141] (44)
[0142] (45)
[0143] ⑥ Choose either the maximum correlation entropy square root capacitive Kalman or the adaptive square root capacitive Kalman flow, combined with the state vector. Measurement vector and input vector The coefficient of friction of the left front wheel on the road surface Coefficient of road surface adhesion for the right front wheel Coefficient of road surface adhesion for the left rear wheel Coefficient of road surface adhesion for the right rear wheel Make an estimate.
[0144] (3) Hub motor state estimation module: When the hub motor is driven by the space vector pulse width modulation (SVPWM) algorithm, the hub motor mathematical model provides the hub motor rotor position for the hub motor speed and rotor angle estimation algorithm. Derivative and rotational speed derivative, shaft current derivative, shaft current Derivative; the measurement data acquisition and processing layer calculates in real time through coordinate transformation. Current and voltage in the axial coordinate system , , , And based on this, the electromagnetic torque can be solved. The measurement data acquisition and processing layer provides load torque for the hub motor speed and rotor angle estimation algorithms. Measured values, as well as the aforementioned current and voltage. , , , Electromagnetic torque Calculated values, etc.
[0145] When the hub motor is driven by the six-step square wave algorithm, the hub motor mathematical model provides the hub motor rotor position for the hub motor speed and rotor angle estimation algorithm. Derivative and rotational speed Derivative, three-phase current of hub motor , , The derivative; real-time solution of electromagnetic torque in the measurement data acquisition and processing layer. and three back electromotive forces , , The measurement data acquisition and processing layer provides three-phase current and voltage data for the hub motor speed and rotor angle estimation algorithms. , , , , , and load torque Measured values, electromagnetic torque and three back electromotive forces , , Calculated values, etc.
[0146] 1) When the hub motor is driven by the SVPWM algorithm, the motor speed and rotor angle estimation algorithm is as follows:
[0147] ① Define the state vector based on the SVPWM drive mode hub motor speed and rotor angle estimation algorithm. ;
[0148] ② Define the input vector for the in-wheel motor speed and rotor angle estimation algorithm based on SVPWM drive mode. ;
[0149] ③ Set the measurement vector for the in-wheel motor speed and rotor angle estimation algorithm based on SVPWM drive mode. ;
[0150] ④ Derivation of the state equations for the speed and rotor angle of the hub motor in SVPWM drive mode
[0151] Differential equation of current in stationary coordinate system:
[0152] right By performing the inverse Park transformation on the voltage equation in the rotating coordinate system, we obtain the current differential equation of the hub motor in the stationary coordinate system.
[0153] (46)
[0154] in, Electric angular velocity, For electrical angle, Ψ f For stator flux linkage, , , These represent the mutual inductance between the three-phase phase inductance, phase resistance, and windings, respectively.
[0155] In-wheel motor dynamic equations:
[0156] (47)
[0157] Electromechanical angle differential equations of an electric motor:
[0158] (48)
[0159] Based on the first-order forward Euler method, the nonlinear discrete expression of the state equation of the hub motor speed and rotor angle based on SVPWM drive mode is derived:
[0160] (49)
[0161] ⑤ Establish measurement equations for the state estimation of wheel motor speed and rotor angle based on SVPWM drive mode:
[0162] (50)
[0163] ⑥ Choose either the maximum correlation entropy square root capacitive Kalman or the adaptive square root capacitive Kalman flow, combined with the state vector. Measurement vector and input vector For the speed of the hub motor based on SVPWM drive mode and rotor angle Make an estimate.
[0164] 2) When the hub motor is driven by the six-step square wave algorithm, the estimation algorithm for motor speed and rotor angle is as follows:
[0165] ① Define the state vector based on the six-step square wave drive mode hub motor speed and rotor angle estimation algorithm. ;
[0166] ② Define the input vector for the in-wheel motor speed and rotor angle estimation algorithm based on the six-step square wave drive mode. ;
[0167] ③ Set the measurement vector based on the six-step square wave drive mode hub motor speed and rotor angle estimation algorithm. ;
[0168] ④ Derivation of the state equations for the speed and rotor angle of the hub motor in the six-step square wave drive mode
[0169] Three-phase voltage winding equations:
[0170] (51)
[0171] in,
[0172] (52)
[0173] In the formula, This is the motor damping coefficient. The moment of inertia of the motor. is the number of magnetic pole pairs.
[0174] (53)
[0175] Based on formulas (44-46) and the first-order forward Euler method, the nonlinear discrete expressions for the state equations of the hub motor speed and rotor angle based on the six-step square wave drive mode are derived:
[0176] (54)
[0177] ⑤ Establish measurement equations for the state estimation of hub motor speed and rotor angle based on the six-step square wave drive mode:
[0178] (55)
[0179] ⑥ Choose either the maximum correlation entropy square root capacitive Kalman or the adaptive square root capacitive Kalman flow, combined with the state vector. Measurement vector and input vector For the speed of the hub motor based on the six-step square wave drive mode and rotor angle Make an estimate.
[0180] Step S3: Intelligent Fusion of Multi-Source Information. In the data fusion layer, a rule-based fusion algorithm is designed to fuse the estimated value from step S2 with the direct measurement / calculation value from step S1, further improving the accuracy and reliability of the output state.
[0181] 1. For vehicle status and motor status: A three-stage fusion rule based on error threshold is adopted. The error between the estimated value and the measured value is calculated. If the error is less than the set threshold, the average value is taken. If the error is between 1 and 3 times the threshold, dynamic weighted fusion is performed (the estimated value moves closer to the measured or calculated value). If the error exceeds 3 times the threshold, the measured or calculated value is determined to be abnormal, and the estimated value is fully trusted.
[0182] (1) The multi-source data fusion algorithm for vehicle status is as follows:
[0183] ① Calculate the errors between the estimated and measured (or calculated) vehicle condition values, including the yaw rate error. Centroid side slip angle error and longitudinal velocity error ,
[0184] (56)
[0185] ② Set an error threshold. To improve fusion accuracy, the error threshold for vehicle state variables can be set to 1% of the estimated value of each state (the error threshold for each state variable can be adjusted according to actual requirements), including the yaw rate error threshold. ( ), centroid side slip angle error threshold ( and longitudinal velocity error threshold ( And compare it with the corresponding state error value to determine the data fusion method.
[0186] ③ If the error is less than the error threshold, i.e. , , This indicates that the estimated value is close to the measured value. The final fused data is the average of the estimated and measured (or calculated) values. The fused yaw rate is... , centroid side slip angle and longitudinal velocity They are respectively,
[0187] (57)
[0188] ④ If the error is greater than the error threshold but less than three times the error threshold (the error threshold multiple can be adjusted according to actual requirements), that is... 3 , This indicates that at least one of the estimated and measured (or calculated) values is inaccurate, and the estimated value needs to be moved closer to the measured (or calculated) value. The measured (or calculated) value and the estimated value are dynamically weighted based on the ratio of a threshold to the error, resulting in the fused yaw rate. , centroid side slip angle and longitudinal velocity They are respectively,
[0189] (58)
[0190] ⑤ If the error is greater than three times the error threshold, i.e. , 3 , 3 Fully trusted estimates, fused yaw rate , centroid side slip angle and longitudinal velocity They are respectively,
[0191] (59)
[0192] (2) The multi-source data fusion algorithm for motor status is as follows:
[0193] ① Calculate the errors in the estimated and measured values of the hub motor's speed and rotor angle, including speed error. Rotor angle error ,
[0194] (60)
[0195] ② Set error thresholds. The error threshold for the hub motor state variables can be 1% of the estimated value of each state (the error threshold for each state variable can be adjusted according to actual requirements), including the speed error threshold. Rotor angle error threshold The data fusion method is determined by comparing the data with the corresponding state error value.
[0196] ③ If the error is less than the error threshold, i.e. , This indicates that the estimated value is close to the measured value. The final fused data is the average of the estimated and measured values. The fused motor speed is... Rotor angle They are respectively,
[0197] (61)
[0198] ④ If the error is greater than the error threshold but less than three times the error threshold, i.e. 3 This indicates that at least one of the estimated and measured values is inaccurate, and the estimated value needs to be moved closer to the measured value. The resulting motor speed... Rotor angle They are respectively,
[0199] (62)
[0200] ⑤ If the error is greater than three times the error threshold, i.e. , 3 Fully trust the estimated values, the merged motor speed Rotor angle They are respectively,
[0201] (63)
[0202] 2. For the road surface adhesion coefficient: Establish a dynamic confidence interval for slip ratio-adhesion coefficient based on a physical model. Compare the estimated adhesion coefficient with this interval: if it falls within the interval, the estimated value is directly adopted; if it exceeds the upper or lower limits of the interval, it is corrected using the interval boundary values. Multi-source data fusion algorithm for road surface adhesion coefficient:
[0203] ① Establish slip ratio through experimental data or authoritative sources With road surface adhesion coefficient The physical model of the corresponding relationship forms the dynamic confidence interval of the road surface adhesion coefficient.
[0204] (64)
[0205] Lower limit of dynamic confidence interval Upper limit of dynamic confidence interval The proportionality coefficients 0.9 and 1.1 can be adjusted and modified according to actual requirements and effects.
[0206] ②If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient Within the confidence interval (20), it indicates that the estimated value is consistent with the working condition reflected by the slip ratio, and the estimated value is determined to be reliable. The fused pavement adhesion coefficient is directly taken as the estimated value.
[0207] (65)
[0208] ③ If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient If the value exceeds the upper limit of the confidence interval (20), it indicates that the estimated value is too high. The fused road adhesion coefficient should directly take the upper limit value.
[0209] (66)
[0210] ④ If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient If the value is below the lower limit of the confidence interval (20), it indicates that the estimated value is too low. The lower limit value is directly taken after fusion of the road adhesion coefficient.
[0211] (67)
[0212] Step S4: Model Layer Support. The method relies on the accurate mathematical models provided by the model layer, including the hub motor mathematical model, the Dugoff tire model, the Magic Formula tire model, and the three-degree-of-freedom vehicle dynamics model, which provide the basis for the state equations and measurement equations of the above estimation modules.
[0213] Secondly, the present invention provides a fusion estimation system for implementing the above method, characterized in that the system includes a physical hardware layer, which is composed of multiple dedicated hardware modules:
[0214] Data acquisition hardware module: responsible for communicating with the sensor interface, acquiring and caching raw data in real time, and has a first processing module and a first storage module.
[0215] State estimation hardware module: This is the core computing unit. It acquires preprocessed data through a high-speed data channel and has a built-in second processing module, second storage module, and a model coprocessor that has embedded the hub motor, vehicle, and tire models. It is used to execute the improved square root commensurate Kalman algorithm.
[0216] Data fusion hardware module: responsible for executing fusion rules, including a third processing module and a third storage module.
[0217] System arbitration and interconnection module: It is used to coordinate the data flow and control commands between the above three modules. It adopts a network on-chip (NoC) interconnection architecture, and the processing and storage units of each module are connected as independent nodes to achieve high-speed, low-latency data interaction.
[0218] The beneficial effects of this invention are as follows:
[0219] This invention achieves integrated collaborative estimation of states across multiple physical domains, significantly improving estimation accuracy and system consistency. For the first time, this invention constructs a unified state estimation framework integrating three core systems: vehicle, tire, and motor. Through an improved square-root capacitive Kalman filter algorithm, multiple key state variables, such as longitudinal vehicle speed, center-of-gravity sideslip angle, yaw rate, four-wheel road adhesion coefficient, hub motor speed, and rotor angle, are collaboratively calculated within the same filtering iteration process. This method fully utilizes the dynamic coupling relationships between the state variables of each subsystem (e.g., motor torque affects wheel slip ratio, slip ratio is related to adhesion coefficient, and adhesion coefficient determines tire force and affects vehicle motion), avoiding the error isolation and accumulation problems caused by traditional independent modular estimation. This makes the estimated state values more physically self-consistent, improving overall estimation accuracy and system consistency.
[0220] An advanced robust filtering mechanism is introduced, greatly enhancing the estimation stability and adaptability in complex noise environments. Addressing the multi-source, time-varying noise such as electromagnetic interference and road impacts encountered by hub-motor electric vehicles in actual operation, this invention innovatively introduces the maximum correlation entropy criterion or adaptive forgetting factor within the square root capillary Kalman framework. The maximum correlation entropy criterion, by weighting the measurement residuals with a Gaussian kernel function, effectively suppresses the negative impact of outliers and impulse noise on the estimation process; the adaptive mechanism dynamically tracks and corrects the statistical characteristics of the measurement noise. This design enables the algorithm not only to be optimal for Gaussian noise but also to handle complex non-Gaussian noise environments with ease. Even when sensor signals are subjected to brief, strong interference or drastic changes in road conditions, the estimation process remains stable, preventing filter divergence and ensuring reliable estimation under all weather and operating conditions.
[0221] A multi-layered intelligent data fusion strategy was constructed, significantly improving the reliability of state output and the system's fault tolerance. This invention transcends the limitations of simply relying on models or sensors, designing a three-source information fusion layer of "estimation-measurement-computation". For vehicle and motor states, a three-stage fusion rule based on physical constraint error thresholds is adopted, which can intelligently determine the reliability of sensor data and smoothly transition or eliminate outliers. For the road adhesion coefficient, a unique approach utilizes a slip ratio-adhesion coefficient physical model to generate dynamic confidence intervals for correction, ensuring that the estimation results always conform to basic physical laws. This fusion mechanism enables the system to automatically rely on highly reliable model estimates and other sensor information for compensation when some sensors (such as GNSS signal loss or resolver interference) temporarily fail or their performance degrades, achieving excellent fault tolerance and ensuring the continuity and high reliability of critical state information output.
[0222] A highly efficient dedicated hardware system architecture is proposed, perfectly meeting the stringent performance requirements of real-time vehicle control. To facilitate the engineering implementation of the complex algorithms described above, this invention features a modular and parallelized dedicated hardware system. By dividing data acquisition, state estimation (including model computation), and data fusion into independent hardware modules and using on-chip networks (NoC) for high-speed interconnection, pipelined operation and parallel computing of the processing flow are achieved. In particular, the state estimation module integrates a model coprocessor specifically for solving nonlinear models of vehicles, tires, and motors, significantly accelerating core computations. This architecture significantly reduces communication latency between functional units, ensuring that the entire multi-state fusion estimation loop can be completed in a very short time (e.g., <10ms), fully meeting the real-time requirements of vehicle chassis electronic control systems (such as ESP and DYC) for state feedback signals, thus laying the hardware foundation for high-performance control.
[0223] This invention provides comprehensive, accurate, and reliable status information input for advanced vehicle control systems, directly enhancing overall vehicle performance. The high-precision, high-refresh-rate fused status parameters output by this invention form the cornerstone of advanced functions such as vehicle dynamics control, torque vector distribution, and energy management optimization. Precise perception of the center of gravity sideslip angle and road adhesion coefficient allows the vehicle stability control system to intervene earlier and more accurately before the loss of control threshold; accurate estimation of the motor rotor position and speed improves the efficiency and smoothness of the motor control itself. Therefore, the implementation of this invention can significantly enhance the handling stability, driving safety, and energy efficiency of in-wheel motor electric vehicles, unleashing the full potential of distributed drive systems. Attached Figure Description
[0224] Figure 1 This is a flowchart illustrating the overall architecture of the method of the present invention.
[0225] Figure 2This is a flowchart of the improved square root capacitive Kalman filter algorithm (maximum correlation entropy) in this invention.
[0226] Figure 3 The flowchart shows the improved square root commensurate Kalman filter algorithm (adaptive algorithm) in this invention.
[0227] Figure 4 This is a flowchart of the vehicle state fusion estimation process in this invention.
[0228] Figure 5 This is a flowchart of the road surface adhesion coefficient fusion estimation process in this invention.
[0229] Figure 6 This is a flowchart illustrating the fusion estimation process of hub motor speed and rotor angle in this invention.
[0230] Figure 7 This is a flowchart of the algorithm for converting vehicle latitude and longitude coordinates to planar coordinates in this invention.
[0231] Figure 8 This is a schematic diagram of the hardware system of the present invention. Detailed Implementation
[0232] To make the technical solution, innovations, and practicality of this invention clearer, the specific implementation methods of the method and system described in this invention will be described in detail below with reference to the accompanying drawings. This section aims to provide a feasible engineering implementation scheme, and those skilled in the art can make appropriate adjustments without departing from the core ideas of this invention.
[0233] I. Overall System Implementation Process
[0234] Figure 1 The overall architecture and data flow of the method of this invention are demonstrated. It is divided into a sensor layer, a measurement data acquisition and processing layer, a state estimation layer, a model layer, and a data fusion layer. In implementation, the system needs to be built according to the closed-loop process of "sensing-processing-estimation-fusion-output".
[0235] Sensor layer:
[0236] The system is equipped with a complete sensor suite, including current and voltage sensors for each wheel hub motor, Hall effect or rotary transformers, a vehicle inertial measurement unit, a steering angle sensor, a GNSS / RTK module, and a cornering sensor. All sensor data is transmitted in real time to the system's data acquisition hardware module via a high-speed bus (such as a CAN bus or automotive Ethernet).
[0237] Measurement data acquisition and processing layer:
[0238] Data acquisition: Current and voltage sensors are used to acquire the phase current of the hub motor. , , and phase voltage , , ;Measure hub motor speed using Hall effect sensors or rotary transformers and rotor angle The hub motor torque sensor collects the load torque. The inertial measurement unit is used to measure the yaw rate of a vehicle. Lateral acceleration Longitudinal acceleration Vehicle yaw angle Steering angle sensor measures front wheel steering angle GNSS / RTK measurement of vehicle heading angle Longitudinal speed and latitude and longitude coordinates .
[0239] Measurement data processing: This involves verifying, filtering, performing coordinate transformation (Clark / Park transformation), and calculating intermediate quantities on the raw data. Specifically, this includes processing the three-phase current of the hub motor. , , and three-phase voltage , , Transform into a stationary coordinate system through coordinate transformation shaft current and voltage , , shaft current and voltage , In a rotating coordinate system shaft current and voltage , , shaft current and voltage , Calculate the three-phase back electromotive force based on the motor parameters. , , and electromagnetic torque By heading angle and yaw angle Calculate the vehicle's center of gravity sideslip angle Using the latitude and longitude coordinates during vehicle operation Calculate vehicle trajectory coordinates ; via hub motor speed and the longitudinal speed of the whole vehicle Calculate tire slip ratio Then, the road surface adhesion coefficient is determined using a physical model that corresponds to the slip ratio and the adhesion coefficient. Interval range. This step provides standardized input for subsequent estimation and fusion.
[0240] State estimation layer:
[0241] The maximum correlation entropy criterion or an improved square root capillary Kalman filter framework is used to run the following three core estimation modules in parallel.
[0242] Vehicle State Estimation Module: The vehicle dynamics mathematical model and the magic tire formula provide the front and rear axle lateral stiffness for the vehicle state estimation algorithm. , longitudinal velocity Derivative, centroid side deflection angle Derivative, yaw rate Derivative. The measurement data acquisition and processing layer provides the steering wheel angle for the vehicle state estimation algorithm. Lateral acceleration Longitudinal acceleration Equal measured values. Based on the above vehicle dynamics model, estimate the longitudinal vehicle speed. , centroid side slip angle yaw rate .
[0243] Road surface adhesion coefficient estimation module: Vehicle dynamics mathematical model and Magic Tire formula provide tire slip ratio for Dugoff tire mathematical model. and sideslip angle Parameters such as these are provided. The Dugoff tire mathematical model provides longitudinal and lateral tire normalized forces for the road adhesion coefficient estimation algorithm. , The measurement data acquisition and processing layer provides the steering wheel angle for the road surface adhesion coefficient estimation algorithm. yaw rate Lateral acceleration Longitudinal acceleration Equivalent measured values. Based on the Dugoff tire model and vehicle dynamics, the four-wheel road adhesion coefficient is estimated. .
[0244] In-wheel motor state estimation module: When the in-wheel motor is driven by the Space Vector Pulse Width Modulation (SVPWM) algorithm, the mathematical model of the in-wheel motor provides the rotor position of the in-wheel motor for the in-wheel motor speed and rotor angle estimation algorithm. Derivative and rotational speed derivative, shaft current derivative, shaft current Derivative. The measurement data acquisition and processing layer calculates it in real time through coordinate transformation. - Current and voltage in the axial coordinate system , , , And based on this, the electromagnetic torque can be solved. This measurement data acquisition and processing layer provides load torque for the hub motor speed and rotor angle estimation algorithm. Measured values, as well as the aforementioned current and voltage. , , , Electromagnetic torque Calculated values; when the hub motor is driven by the six-step square wave algorithm, the hub motor mathematical model provides the hub motor rotor position for the hub motor speed and rotor angle estimation algorithm. Derivative and rotational speed Derivative, three-phase current of hub motor , , The derivative of the electromagnetic torque is calculated in real time at the measurement data acquisition and processing layer. and three back electromotive forces , , The measurement data acquisition and processing layer provides three-phase current and voltage data for the hub motor speed and rotor angle estimation algorithms. , , , , , and load torque Measured values, electromagnetic torque and three back electromotive forces , , Calculated values. Based on the mathematical model of the motor, the speed of each motor is estimated. and rotor position .
[0245] Model layer:
[0246] Provide real-time and accurate mathematical models (hub motor model, Dugoff / magic formula tire model, and three-degree-of-freedom vehicle model) to support the above three estimation modules.
[0247] Data fusion layer:
[0248] Intelligent fusion of state estimates with sensor measurements (or pre-processed calculations) is performed.
[0249] For vehicle and motor status: a three-stage fusion rule based on an error threshold is adopted. The error between the estimated and measured values is calculated. If the error is less than a set threshold, the average value is taken; if the error is between 1 and 3 times the threshold, dynamic weighted fusion is performed (the estimated value converges towards the measured or calculated value); if the error exceeds 3 times the threshold, the measured or calculated value is deemed abnormal, and the estimated value is fully trusted. During vehicle status data fusion, the measurement data acquisition and processing layer uses heading angle... and yaw angle Calculate the difference to determine the vehicle's sideslip angle. Then, the calculated centroid sideslip angle , collected yaw rate and longitudinal speed The data is passed to the data fusion layer and compared with the longitudinal vehicle speed of the state estimate. , centroid side slip angle yaw rate During data fusion, the measurement data acquisition and processing layer provides the data fusion layer with the rotor position of the hub motor. With rotational speed Equal measured values and state estimates of rotational speed and rotor position To integrate.
[0250] For the road surface adhesion coefficient, a dynamic confidence interval fusion based on the slip ratio-adhesion coefficient physical model is used. The measurement data acquisition and processing layer utilizes the rotational speed of each wheel hub motor. and longitudinal speed Calculate the slip ratio of each tire Transmitted to the data fusion layer, and compared with the estimated road surface adhesion coefficient. To integrate.
[0251] Finally, the output is a fused set of high-precision, high-reliability state variables. It is used by upper-level vehicle control systems (such as ESP and torque vectoring).
[0252] II. Implementation of Core Algorithm Steps
[0253] The implementation of the improved square root capacitive Kalman filter is as follows: Figure 2 and Figure 3 :
[0254] After initialization, the algorithm enters a periodic iteration phase. Within each cycle, it strictly follows... Figure 2 (Taking the maximum correlation entropy criterion as an example) or Figure 3 The process shown (using the adaptive forgetting factor as an example):
[0255] (1) Flowchart of the improved square root volume Kalman algorithm based on the maximum correlation entropy criterion:
[0256] ① Initialize the state vector And error covariance matrix .
[0257]
[0258] ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. .
[0259] ③ Generate according to the volumetric Kalman rule State volume points :
[0260]
[0261] in, The number of state variables. This is a preset set of volume points.
[0262] ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. .
[0263] ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0264] ⑥ Calculate the weighted center matrix of the state variables
[0265]
[0266] ⑦ Weight the center matrix of the state variables With matrix A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. .
[0267] ⑧Use Generate updated state volume points ,
[0268]
[0269] ⑨ Through measurement equations Propagate the updated state volume point , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. .
[0270] ⑩ The measurement noise covariance is updated using the maximum correlation entropy criterion to enhance robustness to non-Gaussian noise. The core of this approach is the introduction of a Gaussian kernel function. Process measurement residuals and dynamically calculate the weight matrix. , where m is the number of measurement variables. Thus, the influence of the measurement noise covariance matrix is equivalently corrected or reconstructed during the iteration process. Item ⑦ in (1) and item 10 Reorganization The matrix is used to update the covariance matrix.
[0271]
[0272] For the measurement noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0273] The computational cost is to measure the weighted central moments and simultaneously update the weighted central matrix of the state variables.
[0274]
[0275] Weighted center matrix of measurement With square root factor A new matrix is constructed, and the square root factor of the measurement error covariance is obtained through QR decomposition. .
[0276] Combine with the updated state quantity weighted center matrix With measurement weighted center matrix Calculate the square root factor of the cross-covariance matrix of state measurements. .
[0277]
[0278] Update the filter gain matrix ,
[0279]
[0280] Update the state matrix ,
[0281]
[0282] Determine the convergence of the state matrix and whether the rate of change η of the state estimate is less than a set threshold ε:
[0283] If η≥ε, then the current Return to step 10 to perform a new measurement iteration;
[0284] If η < ε, then continue with the following steps to update the error covariance matrix.
[0285] matrix and Construct new moments and update the square root factor of the error covariance through QR decomposition. .
[0286] (2) Adaptive algorithm improved square root comprehension Kalman algorithm flow:
[0287] ① Initialize the state vector And error covariance matrix .
[0288]
[0289] ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. .
[0290] ③ Generate according to the volumetric Kalman rule State volume points :
[0291]
[0292] in, The number of state variables. This is a preset set of volume points.
[0293] ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. .
[0294] ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0295] ⑥ Calculate the weighted center matrix of the state variables
[0296]
[0297] ⑦ Weight the center matrix of the state variables With matrix A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. .
[0298] ⑧Use Generate updated state volume points ,
[0299]
[0300] ⑨ Through measurement equations Propagate the updated state volume point , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. .
[0301] ⑩ Update the measurement noise covariance using an adaptive filtering formula.
[0302]
[0303] For the measurement noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. .
[0304] The computational cost is to measure the weighted central moments and simultaneously update the weighted central matrix of the state variables.
[0305]
[0306] Weighted center matrix of measurement With square root factor A new matrix is constructed, and the square root factor of the measurement error covariance is obtained through QR decomposition. .
[0307] Combine with the updated state quantity weighted center matrix With measurement weighted center matrix Calculate the square root factor of the cross-covariance matrix of state measurements. .
[0308]
[0309] Update the filter gain matrix ,
[0310]
[0311] Update the state matrix ,
[0312]
[0313] matrix and Construct new moments and update the square root factor of the error covariance through QR decomposition. .
[0314] Key steps and implementation points:
[0315] Noise covariance update is the core of the improved square root capacitive Kalman algorithm in this invention: if using Figure 2 The process requires calculating the measurement residuals, obtaining the weight matrix using a Gaussian kernel function, and updating the measurement noise covariance accordingly. If adopted Figure 3 The process then uses a formula incorporating a forgetting factor to directly iterate and update. .
[0316] QR decomposition and square root update: The algorithm uses QR decomposition multiple times to update the square root factor of the error covariance, which is numerically more stable than directly updating the covariance matrix and can be achieved by calling hardware math libraries or dedicated instructions.
[0317] Implementation of vehicle state fusion estimation:
[0318] During implementation, the maximum correlation entropy square root capillary Kalman or adaptive square root capillary Kalman algorithm is first selected to determine the vehicle's yaw rate. , centroid side slip angle Longitudinal speed Make an estimate, and then calibrate the error threshold through experiments. In each period, the error between the estimated value and the measured or calculated value is calculated, and then based on... Figure 4 The decision-making logic selects one of three fusion strategies: "averaging", "dynamic weighting", or "taking the estimated value", to obtain the final output. .according to Figure 4 The specific process for vehicle state estimation fusion is as follows:
[0319] ① Define the state vector for the vehicle state estimation algorithm ;
[0320] ② Define the input vector for the vehicle state estimation algorithm ;
[0321] ③ Define the measurement vector for the vehicle state estimation algorithm ;
[0322] ④ Let the sampling time be... Based on the first-order forward Euler method, the nonlinear discrete expression of the vehicle state equation is derived:
[0323]
[0324] in For vehicles Moment of inertia of the axis.
[0325] ⑤ Establish the measurement equations for the vehicle state estimation algorithm:
[0326]
[0327] ⑥ Select the improved square root capacitive Kalman filter process, combined with state vector Measurement vector and input vector yaw rate of the state variable , centroid side slip angle Longitudinal speed Make an estimate.
[0328] ⑦ Calculate the errors in the estimated and measured (or calculated) vehicle condition values, including the yaw rate error. Centroid side slip angle error and longitudinal velocity error ,
[0329]
[0330] ⑧ Set error thresholds, including yaw rate error thresholds. ( ), centroid side slip angle error threshold ( and longitudinal velocity error threshold ( And compare it with the corresponding state error value to determine the data fusion method.
[0331] ⑨ If the error is less than the error threshold, i.e. , , This indicates that the estimated value is close to the measured value. The final fused data is the average of the estimated and measured (or calculated) values. The fused yaw rate is... , centroid side slip angle and longitudinal velocity They are respectively,
[0332]
[0333] ⑩ If the error is greater than the error threshold but less than three times the error threshold, i.e. 3 , This indicates that at least one of the estimated and measured (or calculated) values is inaccurate, and the estimated value needs to be moved closer to the measured (or calculated) value. The measured (or calculated) value and the estimated value are dynamically weighted based on the ratio of a threshold to the error, resulting in the fused yaw rate. , centroid side slip angle and longitudinal velocity They are respectively,
[0334]
[0335] If the error is greater than three times the error threshold, that is , 3 , 3 Fully trusted estimates, fused yaw rate , centroid side slip angle and longitudinal velocity They are respectively,
[0336]
[0337] Implementation of road surface adhesion coefficient fusion estimation:
[0338] During implementation, the maximum correlation entropy square root capacitive Kalman or adaptive square root capacitive Kalman algorithm is first selected to evaluate the road surface adhesion coefficient. Estimate the coefficient of friction. The key to implementing road surface adhesion coefficient fusion is establishing an accurate slip ratio-adhesion coefficient ratio. Physical model boundaries. The upper and lower limits of the confidence interval can be fitted using bench test data. And store it as a lookup table or function. During fusion, the estimated values are... With current slip ratio Comparing the corresponding confidence intervals, if If the estimated value falls within the interval, it is directly adopted; if it exceeds the upper or lower limit of the interval, it is corrected according to the interval boundary values. Figure 5 The specific process for road surface adhesion coefficient estimation and fusion is as follows:
[0339] ① Define the state vector for the road surface adhesion coefficient estimation algorithm ;
[0340] ② Define the input vector for the road surface adhesion coefficient estimation algorithm. ;
[0341] ③ Define the measurement vector for the road surface adhesion coefficient estimation algorithm. ;
[0342] ④ Assuming the road adhesion coefficient is considered constant over a short time, based on the first-order forward Euler method, the nonlinear discrete expression of the state equation for the road adhesion coefficient is derived:
[0343]
[0344] ⑤ Establish the measurement equation for the road surface adhesion coefficient estimation algorithm:
[0345]
[0346]
[0347] in, , These are the leading front and rear axle distances of the vehicle's center of gravity, respectively. The wheelbase of the vehicle. Let δ be the moment of inertia of the vehicle about its vertical axis. ij This refers to the wheel steering angle.
[0348] ⑥ Select the improved square root capacitive Kalman filter process, combined with state vector Measurement vector and input vector The coefficient of friction of the left front wheel on the road surface Coefficient of road surface adhesion for the right front wheel Coefficient of road surface adhesion for the left rear wheel Coefficient of road surface adhesion for the right rear wheel Make an estimate.
[0349] ⑦ Establish slip ratio through experimental data or authoritative sources. With road surface adhesion coefficient The physical model of the corresponding relationship forms the dynamic confidence interval of the road surface adhesion coefficient.
[0350]
[0351] ⑧ If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient Within the confidence interval (20), it indicates that the estimated value is consistent with the working condition reflected by the slip ratio, and the estimated value is determined to be reliable. The fused pavement adhesion coefficient is directly taken as the estimated value.
[0352]
[0353] ⑨ If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient If the value exceeds the upper limit of the confidence interval (20), it indicates that the estimated value is too high. The fused road adhesion coefficient should directly take the upper limit value.
[0354]
[0355] ⑩ If in slip ratio Under the given conditions, the estimated road surface adhesion coefficient If the value is below the lower limit of the confidence interval (20), it indicates that the estimated value is too low. The lower limit value is directly taken after fusion of the road adhesion coefficient.
[0356]
[0357] In-wheel motor state fusion ( Figure 6 ):
[0358] The implementation method is similar to vehicle state fusion. First, the maximum correlation entropy square root capacitive Kalman or adaptive square root capacitive Kalman algorithm is selected to measure the wheel hub motor speed. and rotor position Estimate the error thresholds for the rotational speed and rotor angle. .according to Figure 6 The process intelligently fuses the measured values and model estimates of the hub motor to improve the reliability of motor status information:
[0359] (1) When it is determined that the hub motor adopts SVPWM drive mode
[0360] ① Define the state vector based on the SVPWM drive mode hub motor speed and rotor angle estimation algorithm. ;
[0361] ② Define the input vector for the in-wheel motor speed and rotor angle estimation algorithm based on SVPWM drive mode. ;
[0362] ③ Set the measurement vector for the in-wheel motor speed and rotor angle estimation algorithm based on SVPWM drive mode. ;
[0363] ④ Based on the first-order forward Euler method, derive the nonlinear discrete expression of the state equations for the speed and rotor angle of the hub motor based on the SVPWM drive mode:
[0364]
[0365] ⑤ Establish measurement equations for the state estimation of wheel motor speed and rotor angle based on SVPWM drive mode:
[0366]
[0367] ⑥ Select the improved square root capacitive Kalman filter process, combined with state vector Measurement vector and input vector For the speed of the hub motor based on SVPWM drive mode and rotor angle Make an estimate.
[0368] (2) When it is determined that the hub motor is driven by a six-step square wave algorithm
[0369] ① Define the state vector based on the six-step square wave drive mode hub motor speed and rotor angle estimation algorithm. ;
[0370] ② Define the input vector for the in-wheel motor speed and rotor angle estimation algorithm based on the six-step square wave drive mode. ;
[0371] ③ Set the measurement vector based on the six-step square wave drive mode hub motor speed and rotor angle estimation algorithm. ;
[0372] ④ Based on the first-order forward Euler method, the nonlinear discrete expression of the state equation of the hub motor speed and rotor angle based on the six-step square wave drive mode is derived:
[0373]
[0374] ⑤ Establish measurement equations for the state estimation of hub motor speed and rotor angle based on the six-step square wave drive mode:
[0375]
[0376] ⑥ Select the improved square root capacitive Kalman filter process, combined with state vector Measurement vector and input vector For the speed of the hub motor based on the six-step square wave drive mode and rotor angle Make an estimate.
[0377] (3) The estimated values and measurement data of the in-wheel motor SVPWM drive mode and the six-step square wave drive mode are fused respectively:
[0378] ① Calculate the errors in the estimated and measured values of the hub motor's speed and rotor angle, including speed error. Rotor angle error ,
[0379]
[0380] ② Set error thresholds, including speed error thresholds. Rotor angle error threshold The data fusion method is determined by comparing the data with the corresponding state error value.
[0381] ③ If the error is less than the error threshold, i.e. , This indicates that the estimated value is close to the measured value. The final fused data is the average of the estimated and measured values. The fused motor speed is... Rotor angle They are respectively,
[0382]
[0383] ④ If the error is greater than the error threshold but less than three times the error threshold, i.e. 3 This indicates that at least one of the estimated and measured values is inaccurate, and the estimated value needs to be moved closer to the measured value. The resulting motor speed... Rotor angle They are respectively,
[0384]
[0385] ⑤ If the error is greater than three times the error threshold, i.e. , 3 Fully trust the estimated values, the merged motor speed Rotor angle They are respectively,
[0386]
[0387] Implementation of coordinate transformation algorithm:
[0388] To analyze vehicle trajectories, the latitude and longitude coordinates output by GNSS need to be transformed to the plane coordinate system in which the vehicle is traveling. During implementation, strict adherence to… Figure 7 Programming the algorithm flow shown:
[0389] ① Define the vehicle's initial point as the origin of the coordinate system, corresponding to the origin of the latitude and longitude coordinates. The corresponding point of the Cartesian coordinate system. The vehicle's latitude and longitude coordinates were collected sequentially. .
[0390] ② Longitude value Convert to x-coordinate in a planar coordinate system Values, writing latitude and longitude coordinates as Calculate separately and distance .
[0391] ③ Latitude value Convert to ordinate in a planar coordinate system Values, writing latitude and longitude coordinates as Calculate separately and distance .
[0392] ④ Calculate separately in the latitude and longitude coordinate system To the origin distance They are respectively:
[0393]
[0394] ⑤ Calculate separately in the latitude and longitude coordinate system and the origin Angle towards longitude They are respectively:
[0395]
[0396] ⑥ Measure or calculate the angle between the test road and the longitude line. Calculate the corresponding Cartesian coordinate system coordinates They are respectively:
[0397]
[0398] Summary of key steps:
[0399] First, determine the origin of the trajectory plane coordinates. and its corresponding latitude and longitude origin Then, for each new latitude and longitude coordinate... Calculate its longitude distance, latitude distance, total distance, and angle with the longitude line relative to the origin in sequence. Finally, based on the angle between the test road and the longitude line... (This can be determined by initially calibrating a path segment). Using the coordinate rotation formula in step ⑥, calculate the coordinates of this point in the vehicle trajectory plane coordinate system. .
[0400] III. Specific Implementation of Hardware System
[0401] Figure 8 The dedicated hardware architecture of the system of the present invention is shown, and its implementation is as follows:
[0402] Data acquisition hardware module:
[0403] Implementing devices:
[0404] A high-performance microcontroller (MCU, such as TI TMS320F28379D) or FPGA can be used as the core processor. This module corresponds to... Figure 1It comprises the "sensor layer" and the "measurement data acquisition and processing layer" functionalities. It connects to all physical sensors via multiple interfaces (ADC, SPI, CAN, UART) to achieve synchronous data acquisition. Its internal first processing module is responsible for running preliminary preprocessing algorithms (such as coordinate transformation), and its first dedicated storage module (such as SRAM, Flash) is used to cache raw data and preprocessing results.
[0405] State estimation hardware module:
[0406] Implementing devices:
[0407] This is the computing core; a high-performance multi-core SoC (such as NVIDIA Jetson AGX Orin) or a heterogeneous "MCU+FPGA" platform is recommended. This module corresponds to... Figure 1 The “state estimation layer” and the “model layer”.
[0408] The second processing module (such as the CPU core of the SoC) is responsible for scheduling and execution. Figure 2 / Figure 3 The main flow of the improved square root capacitive Kalman filter is shown.
[0409] The model coprocessor is an innovative design and can be a hardware acceleration IP implemented in a SoC's GPU, DSP core, or FPGA. It specifically embeds a solver for the nonlinear equations of hub motors, tires, and vehicle dynamics models, which can greatly accelerate the calculation of state prediction and measurement prediction steps to meet real-time requirements.
[0410] The second dedicated storage module is used to store the filter's state vector, covariance matrix, and model parameters.
[0411] Data fusion hardware module:
[0412] Implementing devices:
[0413] This can be achieved using a standalone MPU (such as an ARM Cortex-A53) or a dedicated core within the state estimation hardware module. This module corresponds to... Figure 1 The "data fusion layer". Its third processing module is responsible for executing... Figure 4 , Figure 5 , Figure 6 The fusion rule algorithm shown includes a third dedicated storage module for storing error thresholds, confidence interval parameters, etc. The data fusion hardware module provides fused state variables to the vehicle controller, human-machine interface, and data recorder.
[0414] System arbitration and interconnection module:
[0415] Implementation architecture:
[0416] Employ a network-on-chip (NoC) interconnect architecture, such as using Arteris FlexNoC IP cores.
[0417] Function: such as Figure 8 As shown, the NoC acts as a high-speed communication backplane, connecting the processors, memory, and peripheral interfaces of the three hardware modules into a unified network. It is responsible for packet routing, transmission arbitration, and real-time scheduling between modules, ensuring efficient and low-latency collaboration among the three pipeline stages of data acquisition, state estimation, and data fusion, avoiding bus congestion, and is crucial for achieving real-time system performance.
[0418] The above specific embodiments demonstrate that, through software and hardware co-design, the present invention effectively combines the multi-level algorithm architecture shown in the accompanying drawings with a dedicated hardware system, achieving high-precision, high-reliability, and real-time fusion estimation of multiple state variables of hub motor electric vehicles, thus achieving the intended purpose of the invention.
Claims
1. A fusion estimation method for state variables of an electric vehicle driven by a hub motor, characterized in that, The fusion estimation of state variables is achieved based on the sensor layer, measurement data acquisition and processing layer, state estimation layer, data fusion layer, and model layer, including the following steps: (1) The sensor layer acquires multi-source measurement data of vehicle operation through measurement sensors and transmits it to the measurement data acquisition and processing layer; (2) The measurement data acquisition and processing layer stores, preprocesses and transforms the received measurement data, and sends the processed data to the state estimation layer, data fusion layer and model layer respectively. (3) The state estimation layer adopts the square root capacitive Kalman filter algorithm framework, introduces the maximum correlation entropy criterion or adaptive forgetting factor to dynamically update and correct the measurement noise covariance matrix, and constructs an anti-interference and adaptive improved square root capacitive Kalman filter algorithm. Based on vehicle stability control logic, the coupling relationship of three types of variables is defined, with each level performing its function independently and variable data being linked and coupled. Then, an improved algorithm is used to achieve accurate state estimation: among which, the vehicle yaw rate... , centroid side slip angle Longitudinal speed The vehicle state estimation parameters are core control variables for determining the vehicle's driving stability and handling characteristics; the hub motor rotor position With rotational speed The estimated parameters for the in-wheel motors are the underlying execution variables for vehicle drive, braking, and power output control; the four-wheel road adhesion coefficient. The parameters for estimating the road surface adhesion coefficient are key road condition constraint variables for vehicle stability control. Under the vehicle stability control system, three types of variables form a closed-loop collaborative constraint relationship: the in-wheel motor's operating state directly determines the vehicle's power output characteristics and is the core factor affecting the overall vehicle stability; the vehicle's real-time stability state inversely constrains the in-wheel motor's operating range and corrects the power output control logic; the road surface adhesion coefficient limits the vehicle's extreme stability conditions and the in-wheel motor's safe output boundary, providing operating condition constraints for overall vehicle stability control and motor execution control. Based on the strong coupling mechanism of the variables in the vehicle stability control, the state estimation layer uses an improved square root capacitive Kalman filter algorithm to perform integrated calculation and estimation of vehicle state parameters, hub motor parameters, and road adhesion coefficient parameters. Finally, the high-precision estimated values of each state variable are output to the data fusion layer, providing accurate data support for the upper-level vehicle stability collaborative control. (4) The data fusion layer fuses the received state estimate with the measured and calculated values of the corresponding state variables according to the preset error threshold or dynamic confidence interval rules, and outputs the final state variable estimation result. (5) The model layer provides mathematical model support for the integrated collaborative estimation of the state estimation layer through the built-in hub motor model, tire model and vehicle dynamics model; (6) The model layer receives the state variable estimates output by the state estimation layer, performs online correction on the hub motor model, tire model and vehicle dynamics model, and feeds back the corrected model parameters to the state estimation layer to form a two-way closed-loop interaction.
2. The method for fusing and estimating state variables of an electric vehicle driven by a hub motor according to claim 1, characterized in that, The improved square root capacitive Kalman filter algorithm includes a basic filtering process and a noise covariance adaptive correction process, as detailed below: (1) Basic square root commensurate Kalman filter algorithm flow: ① Initialize the state vector And error covariance matrix ; ; ② Regarding the error covariance matrix Perform Cholesky decomposition to obtain its square root factor. ; ③ Generate according to the volumetric Kalman rule State volume points ,in The number of state variables; ; in, This is a preset set of volume points; ④ Through state equations Propagate volume points to obtain the predicted point vector. And calculate the predicted state mean vector. ; ⑤ The process noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. ; ⑥ Calculate the weighted center matrix of the state variables; ; ⑦ Weight the center matrix of the state variables Square root matrix of process noise covariance A new matrix is constructed, and the square root factor of the prediction error covariance is updated through QR decomposition. ; ⑧ Utilize the updated prediction error covariance square root factor Generate new state volume points ; ; ⑨ Through measurement equations Propagation of new state volume points , to obtain the measurement prediction value And calculate the mean vector of the measured predicted values. ; ⑩ Measurement noise covariance matrix Perform Cholesky decomposition to obtain its square root matrix. ; The computational cost is to measure the weighted central moments and simultaneously update the weighted central matrix of the state variables. ; Weighted center matrix of measurement Square root matrix of measurement noise covariance A new matrix is constructed, and the square root factor of the measurement error covariance is obtained through QR decomposition. ; Combined with the updated state quantity weighted center matrix Weighted center matrix of measurement Calculate the square root factor of the cross-covariance matrix of state measurements. ; ; Update the filter gain matrix ; ; Update the state matrix ; ; matrix and A new matrix is constructed, and the square root factor of the error covariance is updated through QR decomposition. ; (2) Measurement noise covariance correction process based on the maximum correlation entropy criterion: Introducing the Gaussian kernel function Process measurement residuals and dynamically calculate the weight matrix. Where m is the number of measurement variables; thus, the influence of the measurement noise covariance matrix is equivalently corrected or reconstructed during the iteration process; the ⑦ terms in (1) are... and item 10 Reorganization The matrix is used to update the covariance matrix: ; For the corrected measurement noise covariance matrix Perform Cholesky decomposition to obtain the updated square root factor. Replace the square root matrix of the original measurement noise covariance in the basic process. After completing the state matrix update, determine the rate of change of the state estimator. Is it less than the set threshold? ,like If so, return to the correction process and iterate again. If so, then the update step of the square root factor of error covariance is performed; (3) Adaptive measurement noise covariance correction process: The measurement noise covariance matrix is updated using an adaptive filtering formula. : ; For the updated measurement noise covariance matrix Perform Cholesky decomposition to obtain the updated square root factor. Replace the square root matrix of the original measurement noise covariance in the basic process. .
3. The method for fusing and estimating state variables of an electric vehicle driven by a hub motor according to claim 1 or 2, characterized in that, The integrated and coordinated estimation and fusion of vehicle yaw rate, center of gravity sideslip angle, and longitudinal speed specifically includes: (1) Vehicle state estimation algorithm design: ① Set the state vector as ,in For longitudinal vehicle speed, For the centroid side slip angle, This refers to the yaw rate; ② Set the input vector as ,in For the front wheel steering angle, For longitudinal acceleration, Lateral acceleration, For front axle lateral stiffness, Rear axle lateral stiffness, where As both a measurement input variable and an output variable, data fusion using an improved Kalman filter algorithm can enhance [the measurement efficiency]. Accuracy; ③ Set the measurement vector as ; ④ Let the sampling time be... Based on the first-order forward Euler method, the nonlinear discrete expression of the vehicle state equation is derived: ; ; in For vehicles moment of inertia of the shaft For vehicle curb weight, This is the distance from the vehicle's center of gravity to the front axle. This is the distance from the vehicle's center of gravity to the rear axle. ⑤ Establish the measurement equations for the vehicle state estimation algorithm: ; ⑥ By improving the square root capacitive Kalman filter algorithm, the vehicle yaw rate is... , centroid side slip angle Longitudinal speed Make an estimate; (2) Vehicle status data fusion rules: ① Calculate the error between the estimated state value and the corresponding measured or calculated value, including the yaw rate error. Centroid side slip angle error Longitudinal speed error ; ; ② Preset error thresholds for each state variable based on physical constraints. To improve fusion accuracy, the error threshold for vehicle state variables is set to 1% of the estimated value of each state, including the yaw rate error threshold. Centroid side slip angle error threshold Longitudinal vehicle speed error threshold ; ③ If the error is less than the corresponding error threshold, i.e. , , The estimated value is fused with the measured or calculated value using an arithmetic mean. ; ④ If the error is greater than the corresponding error threshold but less than three times the error threshold, i.e. , , The measured or calculated value and the estimated value are dynamically weighted and fused based on the ratio of the threshold to the error. ; ⑤ If the error is greater than or equal to three times the error threshold, i.e. , , If the measured or calculated value is deemed abnormal, the estimated value is directly adopted as the fusion result. 。 4. The method for fusing and estimating state variables of an electric vehicle driven by a hub motor according to claim 1 or 2, characterized in that, The integrated collaborative estimation and fusion of the four-wheel road adhesion coefficient specifically includes: (1) Design of road surface adhesion coefficient estimation algorithm: ① Set the state vector as ,in , , , These are the road adhesion coefficients for the front left, front right, rear left, and rear right wheels, respectively. ② Set the input vector as ,in For the front wheel steering angle, For longitudinal tire normalized force, This refers to the normalized force of the lateral tires. ③ Set the measurement vector as ,in For longitudinal acceleration, Lateral acceleration, This refers to the yaw rate; ④ Assuming the road surface adhesion coefficient is constant over a short period of time, based on the first-order forward Euler method, derive the nonlinear discrete expression of the road surface adhesion coefficient state equation: ; ⑤ Establish the measurement equation for the road surface adhesion coefficient estimation algorithm: Longitudinal motion equation: ; Lateral motion equations: ; Equation of yaw motion: ; Normalized form of the Dugoff tire model, ; ; ; in, For longitudinal tire normalized force, For the lateral tire normalized force, For lateral stiffness, For longitudinal stiffness, For the slip angle of each wheel, For each wheel, the longitudinal slip ratio, The speed influence factor corrects for the effect of tire slip speed on tire longitudinal force; These are boundary values used to distinguish the linear region of the tire; Based on the above formula, derive the measurement equation for the road surface adhesion coefficient estimation algorithm. ; ; in, , These are the distances from the vehicle's center of gravity to the front and rear axles, respectively. The wheelbase of the vehicle. Let be the moment of inertia of the vehicle about its vertical axis. This refers to the wheel steering angle; ⑥ By improving the square root capacitive Kalman filter algorithm, the road adhesion coefficient of the left front, right front, left rear, and right rear wheels is analyzed. , , , Make an estimate; (2) Rules for fusing road surface adhesion coefficient data: ① The relationship between the road surface adhesion coefficient and the tire slip ratio is utilized. The μ-λ function is used to calculate the theoretical value of the road surface adhesion coefficient through the slip ratio, and a dynamic confidence interval for the road surface adhesion coefficient is established. ; ; Where D1, D2, and D3 are fitting parameters, calibrated based on experiments; the lower limit of the dynamic confidence interval... Upper limit of dynamic confidence interval ; ②If the estimated road surface adhesion coefficient Falling on the corresponding slip ratio Within the dynamic confidence interval, the estimated value is directly adopted as the fusion result; ; ③ If the estimated road surface adhesion coefficient If the result exceeds the upper limit of the dynamic confidence interval, the upper limit of the interval will be used as the fusion result. ; ④ If the estimated road surface adhesion coefficient If the value is below the lower limit of the dynamic confidence interval, the lower limit of the interval is used as the fusion result; 。 5. The method for fusing and estimating state variables of an electric vehicle driven by a hub motor according to claim 1 or 2, characterized in that, The integrated and coordinated estimation and fusion of rotor position and speed of the hub motor specifically includes: (1) For hub motors driven by space vector pulse width modulation (SVPWM), the estimation algorithm is designed as follows: ① Set the state vector as ,in , In a stationary coordinate system , shaft current, For hub motor speed, This refers to the rotor position of the hub motor; ② Set the input vector as ,in , In a stationary coordinate system , shaft voltage, For electromagnetic torque, This is the load torque; ③ Set the measurement vector as ; ④ Based on the first-order forward Euler method, derive the nonlinear discrete expression of the corresponding state equation: ; in, For motor phase resistance, For motor phase inductance, This represents the number of pole pairs of the motor. For the stator flux linkage of the motor, For the electric angle of the motor, This refers to the moment of inertia of the motor. ⑤ Establish the corresponding measurement equation: ; ⑥ By improving the square root commensurate Kalman filter algorithm, the rotational speed of the hub motor is... With rotor position Make an estimate; (2) For the hub motor with six-step square wave drive mode, the estimation algorithm is designed as follows: ① Set the state vector as ,in , , This refers to the three-phase current of the hub motor. For hub motor speed, This refers to the rotor position of the hub motor; ② Set the input vector as ,in , , For the three-phase voltage of the hub motor, , , It consists of three opposite electromotive forces. For electromagnetic torque, This is the load torque; ③ Set the measurement vector as ; ④ Based on the first-order forward Euler method, derive the nonlinear discrete expression of the corresponding state equation: ; ⑤ Establish the corresponding measurement equation: ; ⑥ By improving the square root commensurate Kalman filter algorithm, the rotational speed of the hub motor is... With rotor position Make an estimate; (3) Hub motor status data fusion rules: ① Calculate the speed error between the estimated and measured speed values. Rotor angle error between estimated and measured values ; ; ②Preset a speed error threshold based on physical constraints Rotor angle error threshold To improve fusion accuracy, the error threshold for motor state variables is set at 1% of the estimated value of each state. ③ If the error is less than the corresponding error threshold, i.e. , The estimated and measured values are fused using an arithmetic mean. ; ④ If the error is greater than the corresponding error threshold but less than three times the error threshold, i.e. , The measured and estimated values are dynamically weighted and fused based on the ratio of the threshold to the error. ; ⑤ If the error is greater than or equal to three times the error threshold, i.e. , If the measured value is determined to be abnormal, the fused value is taken as the state estimate; 。 6. The method for fusing and estimating state variables of an electric vehicle driven by a hub motor according to claim 1, characterized in that, The measurement data acquisition and processing layer also includes an algorithm for converting vehicle latitude and longitude coordinates to plane coordinates, with the following specific steps: ① Define the vehicle's initial point as the origin of the coordinate system, corresponding to the origin of the latitude and longitude coordinates. and the origin of the Cartesian coordinate system The latitude and longitude coordinates of the vehicle were collected sequentially during its operation. , ; ② Longitude value Convert to x-coordinate values in a plane coordinate system, and write latitude and longitude coordinates as... , , Calculate using Vincenty's formula respectively , and distance ; ③ Latitude value Convert to the ordinate value in a plane coordinate system, and write the latitude and longitude coordinates as... , , Calculate using Vincenty's formula respectively , and distance ; ④ Calculate separately in the latitude and longitude coordinate system , To the origin distance : ; ⑤ Calculate separately in the latitude and longitude coordinate system , and the origin Angle towards longitude , : ; ⑥ Measure or calculate the angle between the test road and the longitude line. By performing coordinate rotation transformation, the coordinates of each data collection point in the direction of the test road are calculated. The axis and perpendicular test road direction are The trajectory coordinates of the axis in the Cartesian coordinate system , : 。 7. A fusion estimation system for state variables of an electric vehicle driven by a hub motor, used to implement the fusion estimation method according to any one of claims 1-6, characterized in that, The system includes a physical hardware layer, which consists of the following dedicated hardware modules: (1) The data acquisition hardware module is connected to the current sensor, voltage sensor, Hall / resolver sensor, inertial measurement unit, GNSS / RTK system and angle sensor of the sensor layer through analog signal interface, SPI / I2C, CAN / CAN FD and NMEA / RTCM3 protocol. It includes a first processing module and a first dedicated storage module for real-time collection, preprocessing, caching of raw sensor data and marking timestamps. The preprocessing includes filtering, outlier removal, unit conversion and calling sensor calibration parameters stored in the first dedicated storage module to complete data calibration. (2) The state estimation hardware module is connected to the data acquisition hardware module via an on-chip high-speed data channel based on the AXI4 protocol. It includes a second processing module for executing the improved square root commensurate Kalman filter algorithm, a second dedicated storage module for caching model parameters and intermediate filtering variables, and a model coprocessor that has a fixed hub motor model, vehicle dynamics model and tire model. The model coprocessor is implemented using ASIC / NPU and is used for hardware-accelerated matrix operations in the filtering prediction step. (3) A data fusion hardware module, which is connected to the data acquisition hardware module and the state estimation hardware module through an on-chip NoC node interface, includes a third processing module for executing multi-source data fusion rules and a third dedicated storage module for storing fusion rule parameters and fusion results. The multi-source data fusion rules include weighted fusion and federated filtering. The data fusion hardware module receives the estimation results with confidence information output by the state estimation hardware module and the redundant observation data output by the data acquisition hardware module to generate the final fused state variables. (4) System arbitration and interconnection module, used to coordinate the data flow and instructions between the data acquisition hardware module, the state estimation hardware module and the data fusion hardware module. It includes an on-chip network NoC interconnection architecture, and the processing module and storage module of each hardware module are connected to the on-chip network as independent nodes. The NoC interconnection architecture adopts a 2×2 Mesh topology, the routing algorithm adopts XY dimensional order routing, and the arbitration strategy is fixed priority arbitration, in which the state estimation hardware module has the highest priority, the data fusion hardware module has the second highest priority, and the data acquisition hardware module has the lowest priority, so as to ensure the real-time performance of the key control flow. The system arbitration and interconnection module includes an on-chip network switching matrix and a system arbitrator, and realizes the transmission and scheduling of data and control signals between nodes through the AXI4-Stream / AXI4-Lite protocol. The output of the data fusion hardware module is connected to the vehicle controller, human-machine interface and data recorder via CAN / CAN FD, Ethernet or LVDS interface, respectively, and outputs the encapsulated fused state variables.
