专利汇可以提供Method and apparatus for implementing a low cost, intelligent controller for a switched reluctance machine专利检索,专利查询,专利分析的服务。并且A controller for a switched reluctance machine utilizing a feedforward neural network in combination with either a fuzzy logic controller or a proportional-integral controller to provide output control signals (e.g., turn-ON angle, turn-OFF angle and peak current) for controlling the energization of a switched reluctance machine. In an alternate embodiment, a fuzzy logic controller is utilized by itself to control a switched reluctance machine.,下面是Method and apparatus for implementing a low cost, intelligent controller for a switched reluctance machine专利的具体信息内容。
I claim as my invention:1. A controller for a switched reluctance machine comprising:a fuzzy-logic control circuit that: (i) receives as an input a signal representing the difference between a desired operating characteristic of the switched reluctance machine and an actual operating characteristic of the switched reluctance machine; and (ii) provides at its output an intermediate control signal;a multi-layer neural network circuit that receives at its input the intermediate control signal from the fuzzy-logic control circuit and provides at its output control signals for controlling energization of the switched reluctance machine.2. The controller of claim 1 wherein the fuzzy-logic control circuit and the neural network are implemented through appropriate programming of a digital processor.3. The controller of claim 1 wherein the input signal representing the difference between a desired operating characteristic of the switched reluctance machine and an actual operating characteristic of the switched reluctance machine is a speed error signal representing the difference between the desired rotational speed of the machine and the actual rotational speed of the machine.4. The controller of claim 3 wherein the fuzzy logic control circuit receives as a second input signal representing the trend in the speed error signal.5. The controller of claim 4 wherein the output of the fuzzy logic controller is a signal corresponding to the desired output torque of the switched reluctance machine.6. The controller of claim 1 wherein the neural network comprises:a front end tapped delay line to adaptively filter the input of the neural network and provide a filtered output signal.7. The controller of claim 6 wherein the neural network further comprises a two-layer feed forward neural network that has been trained using a training data set derived from the operational characteristics of at least one switched reluctance machine.8. The controller of claim 6 wherein the neural network has been trained using a training set derived from the operational characteristics of a plurality of switched reluctance machines having different sizes and operating characteristics.9. The controller of claim 6 wherein the neural network comprises an eight neuron hidden layer and a three neuron output layer.10. The controller of claim 1 wherein the neural network receives as an input, in addition to the intermediate control signal from the fuzzy-logic control circuit, a signal representing the desired operating characteristic of the switched reluctance machine.11. The controller of claim 1 wherein the neural network further receives as an input a signal representing the actual operating characteristic of the switched reluctance machine.12. The controller of claim 1 wherein the output control signals from the neural network define at least one control angle for controlling the switched reluctance machine.13. A controller for a switched reluctance machine comprising:a proportional-integral control circuit that receives as an input a speed error signal representing the difference between the desired operating speed of the switched reluctance machine and the operating speed of the switched reluctance machine, the proportional-integral controller providing output signals that are proportional to the sum of a component proportional to the intergral of the input and a component proportional to the input; anda multi-layer neural network circuit that receives at one input the outputs from the proportional-integral control circuit, the neural network providing at its output control signals for controlling energization of the switched reluctance machine.14. The controller of claim 13 wherein the neural network receives as an input, in addition to the outputs from the proportional-integral controller, a signal representing the desired operating speed of the switched reluctance machine.15. The controller of claim 13 wherein the neural network receives as an input, in addition to the outputs from the proportional-integral controller, a signal representing the actual operating speed of the switched reluctance machine.16. The controller of claim 13 wherein the neural network comprises a front end tapped delay line to adaptively filter the input of the neural network.17. The controller of claim 13 wherein the neural network comprises a two-layer feed forward neural network having an eight neuron hidden layer and a three neuron output layer.
FIELD OF THE INVENTION
This invention relates to the field of electronic controllers for controlling a switched reluctance machine.
BACKGROUND OF THE INVENTION
Electric motors turn electrical energy into mechanical energy to produce work. Electric motors work by applying a voltage across one or more inductors, typically a machine winding, thereby energizing the inductor(s) to produce a resultant magnetic field. Mechanical forces of attraction or repulsion caused by the magnetic field cause a rotor in an electric motor to move. The efficiency of the electrical motor depends in part on the timing and magnitude of each application of voltage to the motor. Timing of the voltage being applied is particularly important in the case of switched reluctance motors.
Historically, the switched reluctance motor was thought to be incapable of competing effectively with other types of motors. More recently, however, a better understanding of motor design and the application of electronically controlled switching has resulted in a robust switched reluctance drive capable of high levels of performance over a wide range of sizes, powers and speeds. Note that the term ‘motor’ is used here, but it will be appreciated by those skilled in the art that the term covers the same machine in a generating mode unless a particular distinction is made.
The switched reluctance motor is generally constructed without conductive windings or permanent magnets on the rotating part (called the rotor) and includes electronically-switched windings carrying unidirectional currents on the stationary part (called the stator). Commonly, pairs of diametrically opposed stator poles may be connected in series or parallel to form one phase of a potentially multi-phase switched reluctance motor. Motoring torque is developed by applying voltage to each of the phase windings in a predetermined sequence that is synchronized with the angular position of the rotor so that a magnetic force of attraction results between poles of the rotor and stator as they approach each other. Similarly, generating action is produced by positioning the voltage pulse in the part of the cycle where the poles are moving away from each other.
The general theory of design and operation of switched reluctance motors is well known and discussed, for example, in
The Characteristics Design and Applications of Switched Reluctance Motors and Drives
, by Stephenson and Blake and presented at the PCIM '93 Conference and Exhibition at Nuremberg, Germany, Jun. 21-24, 1993.
The control of the torque of a switched reluctance machine (and hence speed and rotor position) is made possible by the positioning and wave-shaping of the excitation currents in the motor windings with respect to the rotor positions. In certain known switched reluctance machine systems, the positioning and shaping of the excitation currents is controlled through the use of peak current values and turn-ON and turn-OFF angles where: (i) a peak current value represents a desired peak current level over a given interval of rotor rotation; (ii) a turn-ON angle defines a point during the rotor's rotation when voltage is applied to the phase winding (referred to as the “turn-ON angle”); and (iii) a turn-OFF angle defines a point during the rotor's rotation when the application of voltage to the winding is halted. Various control schemes (e.g., pulse width modulation, bang-bang current chopping and others) may be used to control the voltage that is applied to a phase winding during the intervals defined by the turn-ON and turn-OFF angles for that phase winding.
The relationships of the control inputs for a switched reluctance machine system (e.g., turn-on angles, turn-off angles, current levels, current wave-shapes, etc.) with the outputs (speed, torque, rotor position, etc.) are complicated, and a full utilization of all the control inputs to maximize performance represents a non-linear many-to-many mapping problem that is difficult for conventional control design. As such, in known switched reluctance motor systems, this complex relationship is typically implemented through the use of a relatively complex circuit that stores signals representative of the relationship between the speed, torque demand, and turn-ON and turn-OFF angles of the motor. Circuits of this type are commonly referred to as “control law tables.”
In many known controllers for switched reluctance motors, the control law table comprises a circuit that includes turn-ON and turn-OFF angle information for various rotor speed and torque demand combinations. In most systems, the information that is stored in the control law table is derived empirically through a process commonly known as “characterization” in which the appropriate turn-ON and turn-OFF angles required to produce the torque demand are determined for a number of different rotor speeds. The empirically derived information is then stored in the control-law table, sometimes together with information from non-tested speeds and torque demands which has been interpolated from the empirically derived information.
One disadvantage of conventional controllers that use control law tables is the necessity of providing for the control law circuitry. For a switched reluctance machine system that works in a wide speed range and a wide torque range and for stability purpose, the control law table must be densely populated, which means that a large data memory is needed. Moreover, because each control law table is the result of the characterization of a particular switched reluctance machine, a control law based controller is a motor-specific controller and cannot be properly used with other switched reluctance machines having designs and sizes different from the machine used to generate the data in the control law table. As such, control law based controllers adds to the cost and time required to develop a new control system, in that, the process of characterizing the machine used in the system may be required for each new motor and controller.
The present invention provides a circuit and a method for controlling a switched reluctance motor without the use of large, costly control law tables. The present invention also provides a compact and generalized controller and control technique for intelligent control of a switched reluctance machine that satisfies desired control requirements and that has lower development and construction costs and better performance than conventional switched reluctance controllers and control techniques.
SUMMARY OF THE INVENTION
Aspects of the present invention provide a compact and generalized technique for intelligent control of a switched reluctance machine that can be easily implemented in a low cost digital signal processor or microcontroller. Such aspects of the present invention allow for the construction of a low-cost high-performance controller.
In accordance with one embodiment of the present invention, a controller for a switched reluctance machine is provided that includes a fuzzy-logic control circuit that: (i) receives as an input a signal representing the difference between a desired operating characteristic of the switched reluctance machine and the actual operating characteristic of the switched reluctance machine; and (ii) provides at its output an intermediate control signal corresponding to the desired output torque of the reluctance machine; and (iii) a multi-layer neural network circuit that receives at its input the intermediate control signal from the fuzzy-logic control circuit and a signal representing the rotational speed of the switched reluctance machine and provides at its output control signals for controlling energization of the switched reluctance machine.
In accordance with another embodiment of the present invention, a controller for a switched reluctance machine is provided that includes a proportional-integral control circuit that receives as input a speed error signal representing the difference between the desired operating speed of the switched reluctance machine and the operating speed of the switched reluctance machine, the proportional-integral controller providing output signals that is the sum of a component proportional to the input and a component proportional to the integral of the input; and a multi-layer neural network circuit that receives at one input the outputs from the proportional-integral control circuit and at another input a signal representing the rotational speed of the switched reluctance machine, the neural network providing at its output control signals for controlling energization of the switched reluctance machine.
In accordance with yet another embodiment of the present invention, a fuzzy logic control circuit for controlling the energization of a switched reluctance machine is provided where the control circuit includes fuzzification means for receiving a signal corresponding to the difference between a desired output characteristic of the switched reluctance machine and the actual output characteristic of the machine and for generating a fuzzy set representing the input signal; a knowledge-based fuzzy inference engine that receives at its inputs the fuzzy set generated by the fuzzification means, the inference-engine generating a fuzzy output control set utilizing a rule-base that corresponds to the general operating characteristics of switched reluctance machines; and de-fuzzification means for receiving the output of the fuzzy inference engine and providing output control signals for controlling the energization of the switched reluctance machine.
Other exemplary embodiments of the present invention and other features of the present invention will be apparent to one of ordinary skill in the art having the benefit of this disclosure.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1
illustrates a reluctance machine and a machine controller constructed according to aspects of the present invention.
FIG. 2
illustrates one embodiment of a switched reluctance machine that may be used in the system of FIG.
1
.
FIGS. 3
a
-
3
b
illustrate in greater detail one exemplary construction of a controller that may be used to construct the system of FIG.
1
.
FIG. 4
illustrates, in functional block form, the operation of one novel control approach that may be implemented in a data processor forming part of the controller of
FIGS. 3
a
and
3
b.
FIGS. 5
a
-
5
c
provide a detailed illustration of one of the constructions of an embodiment of a fuzzy logic block that may be used to implement the control approach of FIG.
4
.
FIGS. 6
a
-
6
c
illustrate components that may be used to form a neural network that may be used to implement the control approach of FIG.
4
.
FIGS. 7
a
-
7
c
illustrate one exemplary training data set that may be used to train the neural network of
FIGS. 6
a
-
6
c.
FIG. 8
illustrates a controller similar to that illustrated in
FIG. 4
in which the fuzzy logic block is replaced with a conventional proportional-integral (“PI”) block.
FIG. 9
illustrates a still further embodiment of the present invention comprising a controller for a switched reluctance machine that is similar to that illustrated in
FIG. 4
but that does not utilize a neural network.
DETAILED DESCRIPTION OF THE INVENTION
Turning to the drawings and, in particular, to
FIG. 1
, a switched reluctance machine system
10
is disclosed. The system
10
includes a switched reluctance machine
12
and a machine controller
14
. The switched reluctance machine comprises a stationary stator defining a number of salient stator poles, and one or more phase windings wound about the poles. The machine
12
also includes a rotor defining a number of rotor poles that is adapted to rotate relative to the stator when electric power is supplied to the phase windings of the machine
14
. The precise construction of the switched reluctance machine
12
(e.g., in terms of the number of stator and rotor poles, the number of phase windings, or the size of the machine) is not essential to the present invention. In general, any switched reluctance machine may be used to practice and implement the subject matter described herein.
FIG. 2
illustrates one embodiment of switched reluctance machine
12
that includes twelve stator poles
20
, eight rotor poles
21
and three phase windings. The winding configuration for only one phase is shown to have four of its coils
22
surrounding four stator poles, and the other two phases are just displaced by one or two stator pole pitches. When the stator pole group of a phase is not aligned to rotor poles, coils of the phase can be excited by applying electric power to the coils to create a magnetic flux path through the excited stator poles, air-gaps and the nearest rotor poles. The flux path established by the energization of the phase coils
22
with the rotor in the illustrated position are generally illustrated by lines
26
. Due to the tendency for the reluctance of the flux path to minimize, the rotor poles are attracted to these stator poles, hence torque is produced. Then when the rotor poles become aligned or nearly aligned with these stator poles, excitation to the active coils is removed so that torque is not produced in the reversed direction; instead, the coils of the adjacent stator poles that belong to the second phase are excited so that the rotor poles are attracted to the second group of stator poles, since they are not aligned.
By selectively exciting the stator pole group to attract rotor poles, continuous motion and continuous torque are produced.
Referring back to
FIG. 1
, machine controller
14
receives electric power from a power source such as the power utility (not shown) through input power lines
16
, and applies such electric power to the phase windings of the switched reluctance machine
12
in a controlled manner so as to control the operation of the machine
12
. The controller
14
may be used to control the output torque of machine
12
or, alternatively, the rotational speed of the rotor of machine
12
or the position of the rotor of machine
12
.
FIGS. 3
a
and
3
b
illustrate in greater detail one exemplary construction of controller
14
. Referring to
FIG. 3
a,
the illustrated controller includes a control board containing a data processor having memory
30
. The data processor
30
receives as inputs various command control signals (e.g., desired speed, desired torque, etc.) and feedback signals (actual speed, actual torque, rotor position, etc.) on input lines
31
. The control and feedback signals may be preprocessed (e.g., amplified, converted from analog to digital or from serial to parallel) before being provided to the data processor
30
by circuit elements not illustrated in
FIG. 3
a.
Data processor
30
may be constructed using any of a number of microcontrollers, microprocessors, application-specific integrated circuits and other devices that may be constructed and/or programmed to digitally implement the control techniques described herein. The selection and programming such a data processor
30
will be apparent to those of ordinary skill in the art having the benefit of this disclosure.
As will be described below, the data processor
30
processes the input signal to generate control signals in the form of gate drive signals that are used to control the application of electric power to the phase windings of the electric machine. In the illustrated example, the data processor
30
generates gate drive signals Q
1
, Q
2
, Q
3
, Q
4
, Q
5
and Q
6
that are applied to gate drivers. Gate drivers
32
convert the gate drive signals from the data processor
30
into drive signals suitable for controlling the conductivity of power semiconductor switching devices such as those illustrated in
FIG. 3
b.
Referring to
FIG. 3
b,
a power control circuit
33
for controlling the application of electric power to the phase windings of the switched reluctance machine
12
is illustrated. As illustrated, the power control circuit includes a rectifier
35
that receives input power through power lines
16
. The rectifier circuit
35
converts the input power (which in the illustrated embodiment is in the form of AC power) into DC power that is regulated to some extent by a DC bus capacitor
34
. Each phase winding of the switched reluctance machine
12
is coupled to the DC bus through semiconductor switching devices Q
1
-Q
6
(e.g., semiconductor switching devices such as MOSFETS, IGBTs, etc.) and two diodes
36
(both in serial with the phase winding). The control inputs to the switches Q
1
-Q
6
are coupled to the outputs from the data processor
30
to control the state of the switches. The switches Q
1
-Q
6
may be placed in a conductive (or “ON”) state or a non-conductive (or “OFF”) state.
For each phase winding, the state of its associated switches determines the energization of the phase winding: turning on the switch pair puts current in the connected winding, while turning off the switch pair removes current through the pair of free-wheeling diodes. The turning on of one switch and the turning off of the other allows current to freewheel through the phase winding and the conductive switch. In the illustrated embodiment, the turning on and off instants of the phase winding are synchronized by data processor
30
with the rotor position. The rotor position may be either directly measured by an optical or hall-effect magnetic sensor or calculated by an estimation algorithm using voltage and current measurements or utilizing other approaches for measuring, calculating or estimating rotor position as is known in the art. Alternate embodiments are envisioned, where the turning on and turning off instances are not synchronized with actual or calculated rotor positions and the instances are determined without reference to actual rotor position.
As described above, the data processor
30
is programmed to implement a novel control technique corresponding to certain aspects of the present invention.
FIG. 4
illustrates, in functional block form, the operation of one novel control approach that may be implemented in data processor
30
. The functions represented by the various blocks of
FIG. 4
may be implemented through appropriate programming of data processor
30
or through appropriate hardware. The implementation of the function blocks in FIG.
4
through software (programming of a data processor) or hardware (discrete circuit components) will be within the abilities of those of ordinary skill having the benefit of this disclosure.
Referring to
FIG. 4
, the illustrated control system
40
receives as inputs a desired speed signal
41
and a signal
42
representing the actual rotational speed of the switched reluctance machine
12
. The signal
42
representing the actual rotational speed of the switched reluctance machine
12
is derived from the output of a sensor set
43
. In the illustrated embodiment, the sensor set
43
includes a rotor position detector or rotor position estimator for providing an output
44
representing the angular position of the rotor relative to the stator. The precise form of the rotor position detector (or rotor position estimator) is not critical and any appropriate sensor or method may be used to provide the rotor position signal
44
.
In the illustrated embodiment, the rate of change of the rotor position signal
44
is determined by function block
45
to provide signal
42
that, as described above, corresponds to the actual rotational speed of the rotor.
In the illustrated embodiment, sensor set
43
also includes current sensors coupled to the phase windings of reluctance machine
12
. These current sensors provide output signals Ia, Ib, Ic that correspond to the currents in the various phase windings. Any appropriate current sensors may be used.
As illustrated in
FIG. 4
, the signal
42
representing the actual rotational speed of the rotor is compared to a signal
41
representing the desired rotational speed of the motor at comparator
46
to provide at the output of the comparator a speed error signal. The speed error signal is applied as one input to fuzzy logic block
50
. The speed error signal is also applied to a trending block
49
that provides at its output a signal corresponding the trend of the speed error signal over time. The trend output from trending block
49
is applied as a second input to fuzzy logic block
50
.
In the illustrated embodiment, fuzzy logic block
50
follows a two-input single-output design where the two inputs correspond to the speed error and its trend as described above and the output corresponds to a torque demand signal. In general, the construction of the fuzzy logic block
50
may follow general fuzzy logic controller structure as described in, for example, L. Tsoukalas, R. Urig,
Fuzzy and Neural Approaches in Engineering,
John Wiley & Sons, Inc. (1997). While a general fuzzy logic structure may be used to implement fuzzy logic block
50
, the fuzzy logic rules and fuzzification/de-fuzzification methods are preferably based on the dynamic characteristics of a switched reluctance machine.
FIG. 5
a
provides a more detailed illustration of one embodiment of the fuzzy logic block
50
. Referring to
FIG. 5
a,
the inputs to the fuzzy logic block
50
are applied to a normalizing block
51
. Normalizing block
51
is utilized in the illustrated embodiment because the knowledge-base for the fuzzy logic block
50
was developed using scaled data. The normalization block
51
should be selected to accommodate the differences between the reluctance machine
12
to which controller
14
is coupled and the scaled data used to develop the knowledge base of the fuzzy logic block.
Once the input signals are normalized by normalizing block
51
they are applied as inputs to fuzzification block
52
. Fuzzification block
52
“fuzzifies” the input signals using any appropriate fuzzification method. One method that may be used to fuzzify the input signals is the singleton fuzzification method which interprets an input as a fuzzy set with the membership function equal to zero except at the point where the membership function equals one. This fuzzification method may be used with standard triangular membership functions such as those illustrated in
FIG. 5
b.
From left to right, the exemplified fuzzy sets that have triangular membership functions have the linguistic values of “NEGATIVE LARGE,” NEGATIVE MEDIUM,” “NEGATIVE SMALL,” “ZERO,” “POSITIVE SMALL,” “POSITIVE MEDIUM,” “POSITIVE LARGE,” respectively. For the inputs, the two fuzzy sets at both ends have an open membership function range extending to infinity while for the output the two fuzzy sets at both ends have a closed membership function range since the output (i.e., desired output torque) has a limited range.
The output of fuzzification block
52
is a fuzzy set that is applied as an input to fuzzy inference engine
53
. The fuzzy inference engine utilizes conventional fuzzy logic techniques and fuzzy rules provided by a knowledge base
54
to generate a fuzzy output set that is applied to a defuzzification block
55
. The fuzzy rules from knowledge base
54
are based on the characteristics of a switched reluctance machine. One exemplary set of such fuzzy rules is provided in
FIG. 5
c,
which are fuzzy rules that may be utilized with various variable speed rotating machines. The linguistic values of the fuzzy sets are used in the fuzzy rule table in
FIG. 5C
, including NL (Negative Large), NM (Negative Medium), NS (Negative Small), ZE (Zero), PS (Positive Small), PM (Positive Medium), PL (Positive Large). Therefore, as an example, the first entry in the table reads: If input #
1
is NL and input #
2
is NL, then output is PL. The use of fuzzy rules, such as are illustrated in
FIG. 5
c
to implement a fuzzy logic function, is described in greater detail in Tang and Xu,
Fuzzy Logic Application for Intelligent Control of a Variable Speed Drive,
IEEE/PES (1994, Winter/New York).
The fuzzy output set from fuzzy inference engine
53
is applied as an input to defuzzification block
55
which converts the fuzzy output set to a single desired torque output signal. In the illustrated embodiment, the defuzzification block
55
utilizes a standard Center of Gravity defuzzification method to provide the appropriate torque demand output signal. The torque output signal from defuzzification block
55
may be de-normalized in de-normalization block
56
to provide a de-normalized torque demand signal that more closely relates to the torque production range of the controlled motor and represents a torque that is achievable.
Referring back to
FIG. 4
, the torque demand output from fuzzy logic block is one input to a feedforward neural network
47
. A second input to the feedforward neural network
47
is either the speed demand signal or the actual speed signal. The choice between the two signals has less effect on the machine performance in steady-state, while in the transient-state, generally the use of the speed demand signal results in faster dynamics while the use of the actual speed signal results in better stability. In general, neural network
47
receives the speed signal (either speed demand signal or actual speed signal) and torque demand signals and provides at its output signals corresponding to appropriate turn-ON and turn-OFF angles and a desired peak current value. The turn-ON, turn-OFF and peak current signals are then applied to a gating signal generator
48
that also receives as an input the rotor position and phase current information from sensor set
43
. The construction and operation of gating signal generator
48
may utilize conventional switched reluctance control circuits for receiving rotor position, turn-ON, turn-OFF, peak current, and actual phase current information and generating gating signals Q
1
-Q
6
. The gating signals from the gating signal generator
48
are applied to gate drivers
32
to control the application of electric power by power circuit
33
to the phase windings as described above. The construction of gating signal generator
48
is well within the ability of those of ordinary skill in the art having the benefit of this disclosure and will not be further discussed herein.
FIG. 6
a
generally illustrates the construction of an exemplary neural network
47
that may be used to implement one controller constructed according to teachings of the present invention. Referring to
FIG. 6
a
the illustrated neural network comprises three parts: normalizing of inputs and de-normalizing of outputs, tapped delay line that allows for adaptive filtering of the input signals to the neural network, the basic neural network. The normalization adjusts for the speed range and torque range of the controlled motor, and the de-normalization adjusts for the ranges of the control angles and the range of the peak current of the controlled motor.
The tapped delay line forms a type of finite impulse response filter that may be used to improve the performance of the neural network
47
.
FIG. 6
b
generally illustrates an exemplary tapped delay line
60
that may be used to form the first part of neural network
47
. Although only one tapped delay line
60
is illustrated, it will be appreciated that a tapped delay line may exist for each of the inputs to the neural network
47
. Referring to
FIG. 6
b,
the tapped delay line includes a first delay layer
61
that provides delayed input signals where the number of delayed input signals corresponds to the number of delay elements in delay layer
61
. The outputs from each delay element are then multiplied by varying weighting factors to produce weighted output delayed signals. While the weighting factors may be selected according to any appropriate process, in general the weighting factors for the present and more recent input signals should be greater than those for more distance signals. Alternately, the weighting factors may be selected through the use of training process akin to those used to train neural networks. The weighted output delayed signals are all applied to a summing block
62
that sums the various signals. A bias may be added as an input to the summing block to allow for easy tuning or adjustment of the tapped delay line
60
.
The output of the summing block
62
may be applied to a transfer function block
63
to implement an appropriate transfer function to produce an adaptively filtered input for application to the basic portion of the neural network.
While the embodiment of
FIGS. 6
a
-
6
b
utilizes a tapped delay line as described above, it should be understood that alternate embodiments may eliminate the tapped delay line and apply the torque demand and speed signals directly to a structure such as that described herein as the basic portion of neural network
47
.
While the precise construction of the basic portion of neural network
47
(referred to hereafter as the “neural network”) may vary, preferably, it is designed and constructed by following the following process: (1) select a topology for the neural network; (2) design the neural network and data structures; (3) obtain training sets for training the neural network; (4) preprocess the training data; (5) train the network using the training data; (6) validate the training of the neural network; and (7) evaluate the performance of the neural network.
The structure of the neural network should be based on the evaluation of various candidate topologies and various trials. For real-time control purpose, one preferred design utilizes a feedforward back-propagation network with two layers, i.e., one hidden layer and one output layer. One such neural network is illustrated in
FIG. 6
c.
In the illustrated embodiment of
FIG. 6
c,
there are only eight neurons in the hidden layer and three neurons in the output layer. Elements of a neuron include weights, a bias, a summer and a transfer function. The transfer functions utilized in neural network are the hyperbolic tangent sigmoid function and the linear function, for the neurons in the hidden layer and the ones in the output layer, respectively. Higher number of neurons (such as twelve neurons in the hidden layer) may give better modeling precision and further minimize the Mean Square Error on training data sets when compared to the network of
FIG. 6
c;
however, for the control of a typical switched reluctance motor, the network of
FIG. 6
b
can learn correctly off-line (for the training data used) and perform adequately on-line without the necessary overhead that would be required to implement a network having greater than eight neurons. The construction of a two layer neural network as described above will be within the ability of one of ordinary skill having the benefit of this disclosure.
While many training approaches may be used to train the neural network of
FIG. 6
c,
one preferred method is a training method utilizing the Levenberg-Marquardt algorithm and multiple training data sets. Other training algorithms may be used such as the gradient descent with momentum approach, the conjugate gradient approach and the quasi-Newton approach.
The training data for training the neural network may be obtained from laboratory testing of a plurality of switched reluctance machines of potentially differing sizes and constructions. For example, one exemplary training data set may be obtained from two motors that are 10-15 HP (about 50 N.m thermally continuous torque at 1500 rpm), both in NEMA 215T industrial frame size. A more desirable approach, and one that would likely lead to a neural network more adapted to handle a variety of motor constructions and sizes would be to train the neural network utilizing data sets obtained from a number of differing switched reluctance machines having a number of different sizes and constructions. When a large number of machines are utilized to obtain test data which is then highly correlated with redundancy, rapid convergence during the training process may be enhanced by the following Principle Component Analysis: (i) orthogonalizing the obtained data sets; (ii) arranging the orthogonal components of the sets; and (iii) truncating those data points that contribute little to variations in the collected data.
In general, the training data should be collected with the aim of modeling the relationship between the control inputs (i.e., turn-on angle, turn-off angle, peak current level) and the outputs (speed, torque). Because there are many possible combinations of the control inputs to obtain the same outputs, the sets of relationship that give maximum system efficiency should be collected throughout the speed and torque ranges to obtain a training data set. One exemplary training data set collected over such a range is illustrated in
FIGS. 7
a
-
7
c
which illustrate training data in terms of the speed/torque/turn-ON angle (
FIG. 7
a
); the speed/torque/turn-OFF angle (
FIG. 7
b
) and the speed/torque/peak current signal (
FIG. 7
c
). Before the training process is performed, the training data may be preprocessed by sequencing and normalizing the raw test data to standard normalized values handled by the neural network.
With the training data, the trained neural network will exhibit a control characteristic that tends to optimize the system's efficiency. Likewise, the sets of relationship that give lowest torque ripples or lowest acoustic noises can be collected and used alternatively through the speed and torque ranges to obtain the training data set. System efficiency, torque ripples and acoustic noises are all measurable quantities representing motor performance.
In general, the neural network may be trained by first initializing the weights and biases of the network and then applying test data in the form of network inputs to generate network outputs. The weights and biases should then be adjusted to minimize the mean square error between the network outputs and the target outputs that are part of the test data. The gradient of the mean square error may be calculated utilizing backpropogation and the weights and biases may be adjusted to move in the direction of a negative gradient. The process of adjustment continues until the above mean square error falls within a small convergence limit.
Once a neural network is trained as described herein, it will not need to be retrained for different switched reluctance motor designs; however, the normalizing factors may need to be adjusted according to the ranges of the inputs and outputs as described earlier.
Once the network is trained, the operation of the system may be verified by utilizing the controller
14
containing the trained network to control a switched reluctance machine under the speed and torque conditions utilized to generate the test data, and under certain untested conditions due to the fact that the trained network has some limited ability of extrapolation. Dynamic performance can then be evaluated, including transient response, stability and robustness.
While the fuzzy-neural control system described above has many advantages over conventional controllers, alternate controllers are envisioned in which the fuzzy logic block
50
is replaced with a conventional proportional-integral (“PI”) block that receives at its input the speed error signal and provides as output signal control signal that is the sum of a component proportional to the input and a component proportional to the integral of the input. An exemplary embodiment of such a controller, including PI block
70
is illustrated in FIG.
8
. In the embodiment of
FIG. 8
, the neural network
47
should be trained using inputs like that provided by PI block
70
.
FIG. 9
illustrates a still further embodiment of the present invention comprising a controller for a switched reluctance machine that is similar to that discussed above in connection with
FIG. 4
but that does not utilize a neural network. Instead, the system of
FIG. 9
utilizes a fuzzy logic controller
80
alone to receive the speed error and speed error trend data and directly provide the turn-ON, turn-OFF and peak current control signals to the gating signal generator
48
. Because the fuzzy logic controller
80
of
FIG. 9
provides the turn-ON, turn-OFF and peak current control signals directly, it is important that the knowledge-base of the system
81
implement appropriate relationships between the inputs of the fuzzy logic controller and the outputs. In general, some relationships that may be implemented include the relationship that as the speed error and speed error trend signals are positive and large, the value of the peak current will increase, the value of the turn-ON angle will decrease and the value of the turn-OFF angle will increase (so as to produce larger torque). Another relationship that may be implemented is that as the speed error and speed error trend signals are negative and their absolute values are large, the value of the peak current will decrease, the value of the turn-ON angle will increase and the value of the turn-OFF angle will decrease so as to produce less torque. Still another relationship that may be implemented is that as the speed error and speed error trend signals remain approximately zero, the values of the peak current, turn-ON angle and turn-OFF angle will remain approximately unchanged. There are many other relationships in the similar format that indicate the relative values of the inputs and the degrees of increase or decrease of the outputs, such as “POSITIVE LARGE,” “POSITIVE MEDIUM,” “POSITIVE SMALL,” “ABOUT ZERO,” “NEGATIVE SMALL,” “NEGATIVE MEDIUM,” “NEGATIVE LARGE.” The full fuzzy rule table contains all the relationships using the fuzzy variables representing such degrees of membership. The implementation of the rule table, membership functions and fuzzy data processing will be within the abilities of those of ordinary skill having the benefit of this disclosure.
The above description of several embodiments is made by way of example and not for purposes of limitation. The present invention is intended to be limited only by the spirit and scope of the following claims.
标题 | 发布/更新时间 | 阅读量 |
---|---|---|
一种基于机器学习的棒束子通道热工水力特性预测方法 | 2020-05-08 | 170 |
高阶邻域混合的网络表示学习方法及装置 | 2020-05-12 | 831 |
一种信息中心网络中基于深度学习的自适应拥塞控制方法 | 2020-05-15 | 731 |
一种舰载机导引路径规划方法 | 2020-05-12 | 403 |
一种非线性动态质量系统预测方法 | 2020-05-11 | 274 |
一种基于土壤重金属污染途径分析的风险评估方法 | 2020-05-12 | 14 |
一种用于分布式光伏发电系统短期出力的预测方法 | 2020-05-11 | 419 |
一种基于动态注意力机制的多事件视频描述方法 | 2020-05-14 | 931 |
基于无线传感器网络的双馈风电机组风能集成控制方案 | 2020-05-08 | 469 |
文本处理模型的运行方法、装置、电子设备、及存储介质 | 2020-05-14 | 95 |
高效检索全球专利专利汇是专利免费检索,专利查询,专利分析-国家发明专利查询检索分析平台,是提供专利分析,专利查询,专利检索等数据服务功能的知识产权数据服务商。
我们的产品包含105个国家的1.26亿组数据,免费查、免费专利分析。
专利汇分析报告产品可以对行业情报数据进行梳理分析,涉及维度包括行业专利基本状况分析、地域分析、技术分析、发明人分析、申请人分析、专利权人分析、失效分析、核心专利分析、法律分析、研发重点分析、企业专利处境分析、技术处境分析、专利寿命分析、企业定位分析、引证分析等超过60个分析角度,系统通过AI智能系统对图表进行解读,只需1分钟,一键生成行业专利分析报告。