CN106338264B - The method for diagnosing faults of hybrid vehicle switching magnetic-resistance BSG position sensors - Google Patents

Abstract

The invention discloses a fault diagnosis method for a switched reluctance BSG position sensor used in a hybrid vehicle. The electric energy output by a storage battery provides a current value i for the switched reluctance BSG after passing through a power converter, and the real-time flux linkage value ψ and current value i Input the wavelet neural network position estimator, the wavelet neural network position estimator outputs the estimated rotor position angle The estimated rotor position angle and the actual rotor position angle θ are input into the fault diagnosis module as input signals, and the fault diagnosis module is and θ are processed to obtain the residual R _i , and the residual R _i is compared with the set threshold T _i to judge the fault type; the wavelet neural network algorithm is used to train the input samples, and the good wavelet transform is fully utilized The time-frequency localization characteristics and the fast self-learning, high robustness and fault-tolerant capabilities of the neural network, compared with the traditional neural network, the convergence speed is faster and the accuracy rate is higher.

Description

Fault Diagnosis Method of Switched Reluctance BSG Position Sensor Used in Hybrid Electric Vehicle

technical field

The invention belongs to the technical field of hybrid power, in particular to a fault diagnosis technology for a position sensor of a belt-driven starter generator (hereinafter referred to as BSG) for hybrid power vehicles.

Background technique

Hybrid electric vehicles have outstanding advantages such as high efficiency and low pollution. Generators are one of the key components of hybrid electric vehicles, and generators are required to work efficiently, stably and reliably. The traditional car starter and generator are two separate parts, but BSG integrates the starter and generator. When the car starts, the BSG quickly drags the engine to the idle speed as the starter. When the car is driving normally or decelerating BSG is used as a generator to charge the car power supply and electrical equipment. BSG replaces the traditional car generator, which not only simplifies the engine design, reduces the weight of the vehicle, but also reduces fuel consumption and pollution emissions. At present, BSGs are mostly hybrid excitation claw pole motors, permanent magnet motors and induction motors. However, for hybrid excitation claw pole motors, it is difficult to obtain high torque at low speeds and the rotor structure is complex, which is not conducive to high-speed operation; for permanent magnet motors, Due to the existence of permanent magnet materials, the cost is high and the stability in high temperature and high magnetic field environments is difficult to guarantee; for induction motors, the speed regulation performance is poor, and it is difficult to carry out precise control, and the requirements for the control system of the motor are relatively high. The switched reluctance motor is suitable for high-speed operation and harsh environments due to its simple and firm structure, low cost and high reliability. It is adopted by BSG and is called switched reluctance BSG.

For switched reluctance BSG, in order to detect its rotor position, a position sensor is installed on it, and the position sensor outputs commutation information. The position sensor is the key to the correct commutation of the switched reluctance BSG. If the position signal fails, the commutation logic of the motor will be disordered, resulting in a decrease in the torque output capability of the motor, and the motor speed will drop or be zero. Therefore, in order to improve the reliability of the switched reluctance BSG system and enable the motor to output commutation information at the correct rotor position, it is necessary to quickly and accurately detect and diagnose faults on the position sensor.

The fault types of the position sensor include: total failure fault, accuracy degradation, fixed deviation fault and drift deviation fault. The first two types of faults are relatively easy to find and can be dealt with in time, but fixed deviation faults and drift deviation faults are not easy to find faults, which will cause a series of unpredictable problems during the fault occurrence process, making the control system unable to function normally for a long time effect.

Most of the existing position sensor fault diagnosis technologies are fault diagnosis methods based on analytical models, which use neural network fault observers to generate residuals, and then analyze the residuals to diagnose faults. However, the fault diagnosis method based on the analytical model needs to establish the mathematical expression of the system model, which is almost difficult to achieve for the severely nonlinear switched reluctance motor control system, and the neural network fault observer requires the system under various fault states. However, in the actual control process, there are very few fault samples.

Wavelet neural network is a neural network model based on wavelet transform. It can not only absorb the good localization characteristics of wavelet transform in time domain and frequency domain, but also take advantage of the advantages of self-adaptive, self-learning and robustness of neural network. , so that the wavelet neural network can be widely used in pattern recognition, nonlinear science, fault diagnosis and so on.

The existing flux linkage detection technology mostly adopts an indirect measurement scheme, that is, the flux linkage characteristics are indirectly calculated by measuring the voltage and current of a certain phase winding. The flux linkage calculation formula is: In the formula, ψ _i (t) is the flux linkage of a certain phase, and u _i , i _i , R _i represent the voltage, current and resistance of a certain phase, respectively. However, this method is difficult to meet the practical engineering application of real-time online detection of flux linkage.

Contents of the invention

The purpose of the present invention is to address the problems existing in the prior art, to provide a fault diagnosis method for a switched reluctance BSG position sensor for a hybrid vehicle, to diagnose the fault of the position sensor in real time, quickly and accurately, and to provide an efficient solution for the switched reluctance BSG system. Reliable job security.

The technical scheme adopted by the fault diagnosis method of the switched reluctance BSG position sensor for hybrid electric vehicles in the present invention is: the position sensor detects and outputs the actual rotor position angle θ of the switched reluctance BSG, and the electric energy output by the battery is converted into a switch after passing through the power converter The reluctance BSG provides the current value i, and also includes the following steps:

A. Connect the wavelet neural network position estimator to the output end of the flux linkage acquisition module after off-line training, and use the flux linkage acquisition module to obtain the real-time flux linkage value ψ of the switched reluctance BSG;

B. The real-time flux linkage value ψ and the current value i are input into the wavelet neural network position estimator as input signals, and the wavelet neural network position estimator outputs the estimated rotor position angle after processing ψ and i

C. The estimated rotor position angle and the actual rotor position angle θ are input into the fault diagnosis module as input signals, and the fault diagnosis module is and θ are processed to get the residual R _i , and the residual R _i is compared with the set threshold T _i to judge the fault type.

Further, in step C, if R _i ≤ T _i , it is judged that the position sensor is working normally; if R _i >T _i , it is judged that the position sensor is faulty.

If the residual R _i is greater than the threshold T _i and the value of R _i is constant after a certain time t _i , it is judged that the position sensor has a fixed deviation fault at the fault occurrence time t _i ; if the residual R _i is at a certain time t After _i appears greater than the threshold T _i and the difference between R _i and T _i becomes larger and larger, it is judged that the position sensor has a drift deviation fault at the fault occurrence time t _i .

Further, in step A, the method for off-line training of the wavelet neural network position estimator is: first obtain the relationship curve of flux linkage-current-rotor position angle under the normal operation state of the switched reluctance BSG, and form the initial sample set {i _a , i _b , i _c , i _d , ψ _a , ψ _b , ψ _c , ψ _d , θ _a , θ _b , θ _c , θ _d }, offline training on the initial sample set; i _a , i _b , i _c , _id represent the phase currents of A, B, C and D phases of the switched reluctance BSG respectively, ψ _a , ψ _b , ψ _c , ψ _d represent the magnetic fluxes of the A, B, C and D phases of the switched reluctance BSG respectively Chain, θ _a , θ _b , θ _c , θ _d represent the position angles of A, B, C, and D phases of the switched reluctance BSG, respectively.

The beneficial effects of the present invention are:

1. The present invention adopts the wavelet neural network algorithm to train the input collection samples, fully utilizes the good time-frequency localization characteristics of the wavelet transform and the fast self-learning, high robustness and fault-tolerant capabilities of the neural network, compared with the traditional neural network , the convergence speed increases and the accuracy rate increases. At the same time, when training the wavelet neural network, only the collected samples under the normal state of the system are needed, which overcomes the problem of lack of faulty samples of the position sensor.

2. Use the magnetic sensor installed on the switched reluctance BSG to directly measure the flux linkage and use the flux linkage acquisition module to obtain real-time flux linkage values, which can quickly and accurately input real-time data online. It is highly practical and has a wide range of engineering application values.

3. The present invention uses the position angle output by the wavelet neural network position estimator and the actual rotor position angle output by the position sensor to perform residual processing, and compares the obtained residual accuracy with the set threshold to determine the fault type. The control principle of the method is simple, and it only needs to be realized through software programming without other hardware devices, and the cost is low, and it is easy for engineering realization.

Description of drawings

Figure 1 is a block diagram of a position sensor fault diagnosis system for a switched reluctance BSG for a hybrid vehicle;

Fig. 2 is the flow chart of the position sensor fault diagnosis method of hybrid electric vehicle with switched reluctance BSG of the present invention;

Fig. 3 is a flux linkage characteristic curve of a certain phase of the switched reluctance motor BSG in Fig. 1;

Fig. 4 is a fixed deviation fault diagram of the switched reluctance BSG in Fig. 1;

FIG. 5 is a fault diagram of a drift deviation of the switched reluctance BSG in FIG. 1 .

In Figure 1: 1. Measured object; 2. Rotor position signal estimation and fault diagnosis module; 3. Battery; 4. Power converter; 5. Engine; 6. Clutch; 7. Transmission; 11. Switched reluctance BSG; 12. Position sensor; 21. Wavelet neural network position estimator; 22. Flux link acquisition module; 23. Fault diagnosis module; 24. Magnetic sensor.

Detailed ways

As shown in Figure 1, the switched reluctance BSG11 and the position sensor 12 constitute the measured object 1 for fault diagnosis. The position sensor 12 is installed on the switched reluctance BSG11 to detect the rotor position of the switched reluctance BSG11 and output the actual rotor position angle θ signal Give the rotor position signal estimation and fault diagnosis module 2. The rotor position signal estimation and fault diagnosis module 2 is connected in series behind the position sensor 12 of the measured object 1, and receives the actual rotor position angle θ signal output by the position sensor 12 to diagnose the fault type of the position sensor 12.

The switched reluctance BSG11 is respectively connected to the engine 5 and the power converter 4 , the engine 5 is connected to the transmission 7 through the clutch 6 , and the power converter 4 is a DC/DC converter connected to the battery 3 of the vehicle power supply. When the car is started, the switched reluctance BSG11 works as a starter. At this time, the electric energy output by the battery 3 provides an appropriate current value i for the switched reluctance BSG11 after being acted on by the power converter 4. At the same time, the switched reluctance BSG11 provides rotation for the engine 5. The power ω and the rotational power ω drive the car to start after passing through the clutch 6 and the transmission 7; when the car is running normally or decelerating, the switched reluctance BSG11 works as a generator, and at this time the switched reluctance BSG11 receives the rotational power ω provided by the engine 5 and Generate electricity, and the obtained current value i provides electric energy for the storage battery 3 after being acted on by the power converter 4 .

The rotor position signal estimation and fault diagnosis module 2 is composed of a wavelet neural network position estimator 21 , a flux linkage acquisition module 22 , a fault diagnosis module 23 and a magnetic sensor 24 .

The wavelet neural network position estimator 21 is trained using an off-line training method. The switched reluctance BSG11 establishes a finite element model through the electromagnetic field software ANSOFT, and obtains the relationship curve of flux linkage-current-rotor position angle under the normal operation state of the switched reluctance BSG11, as shown in Figure 3, and thus forms the initial sample Set {i _a , i _b , ic , i _d , ψ _a , ψ _b , ψ _c , ψ _d , θ _a , θ _b , θ _{c ,} θ _d }, where i _a , i _b , _ic , i _d represent the phase currents of A, B, C and D phases of switched reluctance BSG11 respectively, ψ _a , ψ _b , ψ _c , ψ _d represent the flux linkages of A, B, C and D phases of switched reluctance BSG11 respectively, θ _a , θ _b , θ _c , θ _d represent the position angles of A, B, C, and D phases of the switched reluctance BSG11 respectively, and the initial sample set is trained by the offline training method. The wavelet neural network of the wavelet neural network position estimator 21 adopts a four-layer structure, including an input layer, an output layer and a hidden layer respectively, and the hidden layer is selected as a two-layer structure. The number of nodes in each hidden layer is 8, and then the number of nodes in each hidden layer is gradually added and deleted through experiments by using the step-by-step growth method and step-by-step pruning method, and finally the number of nodes in the two hidden layers is 10 and 8 respectively. In addition, the Mexican hat (Mexican hat) wavelet function is used as the neuron activation function of the hidden layer node, and the wavelet function can be expressed as: where h(x) is the Mexican hat wavelet function, and x is the time constant. After determining the structure of the wavelet neural network, the Powell (Powell) algorithm is used to train the wavelet neural network. The Powell algorithm can be expressed as: M ^k+1 = M ^k -[K ^T K+μI] ^-1 K ^T d, where , M ^k+1 is a vector composed of all the weight coefficients of the wavelet neural network when the number of iterations is k+1; M ^k is a vector composed of all the weight coefficients of the wavelet neural network when the number of iterations is k; M is the vector composed of all the weight coefficients of the wavelet neural network; μ= ¹⁰⁴ is the precision coefficient; k is the current iteration number; K is the Jacobian matrix of the first-order derivative of the error of the weight coefficient of the wavelet neural network; K ^T is the wavelet The Jacobian transpose matrix of the first-order derivative of the error of the weight coefficient of the neural network; T is the matrix transpose; I represents the identity matrix; d is the error of the wavelet neural network for the weight coefficient. The calculation formula of each element in the Jacobian matrix K is: Where i=1,2,3,…,m, m is the number of input variables; j=1,2,3,…,n, n is the number of input variables; K _ij (M ^k+1 ) is k+ The function of the Jacobian matrix K at time 1, f _i (M ^k ) is the difference between the expected output and the actual output of the wavelet neural network at time k, and f _i (M ^k+1 ) is the expected output and the actual output of the wavelet neural network at time k+1 Difference, is a vector composed of all the weight coefficients of the wavelet neural network at time k, It is a vector composed of all the weight coefficients of the wavelet neural network at time k+1.

The output end of the magnetic sensor 24 is connected to the input end of the flux linkage acquisition module 22 , and the trained wavelet neural network position estimator 21 is connected in series to the output end of the flux linkage acquisition module 22 . The magnetic sensor 24 is installed on the switched reluctance BSG11, directly measures the real-time flux linkage of the switched reluctance BSG11, and uses the flux linkage acquisition module 22 to acquire the real-time flux linkage value ψ of the switched reluctance BSG11. The flux linkage value ψ output by the flux linkage acquisition module 22 and the current value i output by the power converter 4 are used as the input signal of the trained wavelet neural network position estimator 21, and the flux linkage value ψ and the current value i pass through the wavelet neural network position Estimator 21 outputs the estimated rotor position angle after processing

The fault diagnosis module 23 will estimate the rotor position angle and the actual rotor position angle θ output by the position sensor 12 as an input signal, and the estimated rotor position angle Perform residual processing with the actual rotor position angle θ, and compare the obtained residual R _i with the set threshold T _i to determine the fault type, see Figure 2.

The fault diagnosis module 23 defines the residual where f represents R _i and The functional relationship of R _i is compared with the threshold T _i that determines the normal state and fault state of the position sensor 12 . If R _i ≤ T _i , it can be judged that the position sensor 12 is working normally; if R _i >T _i , it can be judged that the position sensor 12 is faulty. Referring to Fig. 4, if the residual R _i is greater than the threshold T _i and the value of R _i is constant after a certain time t _i , it can be judged that the position sensor 12 has a fixed deviation fault at the fault occurrence time t _i . Referring to Fig. 5, if the residual R _i is greater than the threshold T _i after a certain time t _i and the difference between R _i and T _i becomes larger and larger, it is judged that the position sensor 12 has drifted at the time t _i of the failure Deviation failure. Set up fault alarms for different types of faults to quickly diagnose fault types in real time.

According to the above, the present invention can be realized. Other changes and modifications made by those skilled in the art without departing from the spirit and protection scope of the present invention are still included in the protection scope of the present invention.

Claims (4)

1. a kind of method for diagnosing faults of hybrid vehicle switching magnetic-resistance BSG position sensors, position sensor detection is simultaneously defeated The actual rotor position angle θ for going out switching magnetic-resistance BSG is switching magnetic-resistance after power inverter by the electric energy of accumulator output BSG provides current value i, it is characterized in that further comprising the steps of：

A, it to being serially connected in the output end of magnetic linkage acquisition module after wavelet neural network position estimator off-line training, is obtained using magnetic linkage Modulus block obtains the real-time magnetic linkage value ψ of switching magnetic-resistance BSG；

Method to wavelet neural network position estimator off-line training is：First obtain under switching magnetic-resistance BSG normal operating conditions Magnetic linkage-electric current-rotor position angle relation curve, composition original training set { i_a, i_b, i_c, i_d, ψ_a, ψ_b, ψ_c, ψ_d, θ_a, θ_b, θ_c, θ_d, to original training set off-line training；i_a, i_b, i_c, i_dThe phase current of A, B, C, D phase of switching magnetic-resistance BSG, ψ are indicated respectively_a, ψ_b, ψ_c, ψ_dThe magnetic linkage of A, B, C, D phase of switching magnetic-resistance BSG, θ are indicated respectively_a, θ_b, θ_c, θ_dIndicate respectively switching magnetic-resistance BSG A, B, the position angle of C, D phase；The wavelet neural network of wavelet neural network position estimator includes input layer, output layer and implies Layer, hidden layer is double-layer structure, in addition uses wavelet function h (x)=(1-x²)e^-x/2Neuron as hidden layer node encourages Function, x are time constant,

B, wavelet neural network position estimator is inputted using the real-time magnetic linkage value ψ and current value i as input signal, it is small Wave neural network position estimator exports the rotor position angle of estimation after handling ψ and i

C, by the rotor position angle of estimationWith actual rotor position angle θ as input signal input fault diagnostic module, therefore Hinder diagnostic module pairWith θ residual error R is obtained as residual noise reduction_i, by residual error R_iWith the threshold value T of setting_iIt compares and carrys out failure judgement class Type.

2. method for diagnosing faults according to claim 1, it is characterized in that：In step C, if Ri≤Ti, judge that position passes Sensor works normally；If Ri ＞ Ti, judge that position sensor is faulty；If residual error R_iAt a time t_iIt is more than later Threshold value T_iAnd R_iIt is worth color constancy, then judges that in failure moment t occurs for position sensor_iDroop failure has occurred；If residual error R_iAt a time t_iOccur being more than threshold value T later_iAnd R_iWith T_iThe increasing phenomenon of difference, then judge position sensor therefore Moment t occurs for barrier_iDrift bias failure has occurred.

3. method for diagnosing faults according to claim 1, it is characterized in that：Using algorithm M^k+1=M^k-[K^TK+μI]^-1K^TD pairs Wavelet neural network is trained, M^k+1For when iterations are by k+1 the entirety of the weight coefficient of wavelet neural network form Vector；M^kFor the vector that the entirety of the weight coefficient of wavelet neural network forms when iterations are by k；K is Wavelet Neural Network The Jacobian matrix of the first derivative of the error of network weight coefficient；T is matrix transposition；μ=10⁴For quality coefficient；I is unit square Battle array；D is error of the wavelet neural network for weight coefficient.

4. method for diagnosing faults according to claim 3, it is characterized in that：The meter of each element in the Jacobian matrix K Calculating formula isM, n is input variable number； K_ij(M^k+1) be k+1 moment Jacobian matrixes K function, f_i(M^k) it is k moment wavelet neural network desired outputs and reality output Difference, f_i(M^k+1) be k+1 moment wavelet neural network desired output and reality output difference,For k moment wavelet neural networks Weight coefficient the vector that is formed of entirety, M^k+1By k+1 moment wavelet neural networks weight coefficient entirety form to Amount.