Multiphase motor and method for controlling a multiphase motor

Abstract

A multiphase motor (1) comprising: a stator (2), N coils (B1,…, BN) for generating a rotating magnetic field, a rotor (4) for generating torque, and N control modules (C1, …, CN). Each control module (Ci) comprises: a generator (RAi) for generating a phase-compensated signal to help reduce oscillation of the generated torque; a neural network (RBi) configured to generate a correction factor taking into account faults other than coil unavailability; a correction module (Ai) configured to apply the correction factor to the phase-compensated signal to produce a phase-corrected signal to further reduce oscillation; and an output interface (Oi) configured to drive a coil using the phase-corrected signal. Figure for the abstract: Fig. 7

Description

Title of the invention: Multiphase motor and method for controlling a multiphase motor. FIELD OF THE INVENTION

[0001] This disclosure relates to a multiphase motor and a method for controlling such a motor. STATE OF THE ART

[0002] A conventional multiphase motor comprises a stator, a rotor movable relative to the stator, and several coils. The coils are controlled by means of phase signals so as to jointly generate a magnetic field that rotates relative to the stator. The rotor is driven in rotation by this rotating magnetic field. The rotation of the rotor relative to the stator generates torque.

[0003] In a nominal operating condition of the multiphase motor, the phase signals used to control the coils to generate the rotating magnetic field generally have a sinusoidal waveform, and the torque generated by the multiphase motor remains substantially constant. By "substantially constant" is meant variations of less than 5%. This is true as long as no fault is present in the motor. When a fault occurs, the previously constant torque may begin to oscillate, the shape of the oscillation depending on the fault. To reduce torque oscillations, it has been proposed to modify the waveform of the phase signals. The phase signals that reduce these oscillations are called "optimal phase signals" or "optimal currents" when the phase signals are currents. To calculate the waveform of the optimal signals in real time, the MTPA (Maximum Torque Per Ampere) algorithm is commonly used..

[0004] However, these real-time calculations of optimal phase signals are particularly resource-intensive, especially when the number of motor coils is high, and require the use of a powerful centralized controller. Although vector methods exist and are much less resource-intensive, they generally do not account for torque oscillations that would result from defects other than an open phase. These methods are therefore not comprehensive. For example, the back electromotive force (back EMF) is induced by the passage of magnets in front of the coils, which generates a voltage. However, at high temperatures, the magnets generate a lower flux, which reduces the amplitude of the back EMF. The magnets can even irreparably lose their magnetization, which alters the back EMF. Current controllers do not take these effects into account, and it is not possible to compensate for them with known methods.

[0005] A less resource-intensive method was described in Nguyen's document, "Chapter 3#: Synchronous Machines in Degraded Mode#: Failure Mode Analysis and Optimal Power Supplies for Synchronous Actuators in the Presence of Faults," 2011. This method uses an AD ALINE-type neural controller that compensates for the oscillatory components of the torque by acting on the currents in the DQ basis of dummy machines. However, this method involves driving the motor in the DQ basis, which does not allow for independent control of the different coils. Description of the invention

[0006] One problem to be solved is to reduce the possible oscillations of a torque generated by a multiphase motor without requiring heavy calculations.

[0007] This problem is solved by a multiphase motor consisting of a stator; N coils of respective indices from 1 to N, with N > 3, the N coils being configured to jointly generate a rotating magnetic field from phase signals of respective indices from 1 to N; a suitable rotor to be driven in rotation relative to the stator under the effect of the rotating magnetic field so as to generate torque; and N control modules of respective indices from 1 to N to control the N coils in parallel. For any i from 1 to N, the control module of index i comprises: • an input interface for obtaining input data including: availability data indicating, for any i from 1 to N, whether the coil with index i is available or unavailable to contribute to generating the rotating magnetic field, a relative position between the rotor and the stator, and a torque error between a setpoint torque and a torque generated by the multiphase motor, • a generator configured to generate a phase-compensated signal of index i from availability and relative position data, the phase-compensated signal of index i being adapted to help reduce oscillation of the generated torque, • a neural network configured to generate, from availability data, relative position and torque error, a correction index i taking into account faults other than coil unavailability, • a correction module configured to apply the correction to the phase-compensated signal of index i so as to produce a phase-corrected signal of index i, the phase-corrected signal of index i being adapted to contribute to a further reduction of the oscillation of the generated torque, • an output interface configured to control the coil of index i using the phase-corrected signal of index i.

[0008] The multiphase motor, which is the first subject of this disclosure, may also include the optional features below.

[0009] Preferably, the neural network has been pre-trained with learning data relating to healthy multiphase motors, multiphase motors including unavailable coils and motors suffering from defects other than coil unavailability.

[0010] Preferably, the generator is another neural network.

[0011] Preferably the other neural network has been previously trained with training data relating selectively to healthy multiphase motors and multiphase motors including unavailable coils.

[0012] Preferably, the neural network or each neural network comprises only one hidden layer.

[0013] Preferably, the neural network or each neural network is configured to use a radial basis function as the activation function.

[0014] Preferably, the correction module is an adder configured to sum the phase-compensated signal of index i and the correction of index i.

[0015] Preferably, the multiphase motor further comprises: a linear hall effect sensor configured to acquire an analog signal varying continuously according to an orientation of the rotating magnetic field, and an additional neural network configured to determine the relative position between the rotor and the stator from the analog signal.

[0016] Preferably, the multiphase motor comprises N source modules of respective indices from 1 to N independent of each other, in which for any i from 1 to N, the source module of index i is configured to provide input data of index i to the control module of index i.

[0017] A second object of this disclosure is a method for controlling the multiphase motor described above. The method comprises implementing the following steps, carried out for any i from 1 to N: • obtain input data including: availability data indicating, for any i from 1 to N, whether the coil with index i is available or unavailable to contribute to generating the rotating magnetic field, a relative position between the rotor and the stator, and a torque error between a setpoint torque and a torque generated by the multiphase motor, • generate a phase-compensated signal with index i from the availability data and relative position, the phase-compensated signal with index i being adapted to contribute to a reduction of an oscillation of the generated torque, • generate, using a neural network and based on availability data, relative position and torque error, a corrective index i taking into account faults other than coil unavailability, • apply the correction to the phase-compensated signal of index i so as to produce a phase-corrected signal of index i, the phase-corrected signal of index i being adapted to contribute to a further reduction of the oscillation of the generated torque, • control the coil of index i using the phase-corrected signal of index i. DESCRIPTION OF THE FIGURES

[0018] Other features, objectives and advantages of the invention will become apparent from the following description, which is purely illustrative and not limiting, and which should be read in conjunction with the accompanying drawings on which:

[0019] Fig. 1 is a schematic cross-sectional view of a multiphase motor according to one embodiment.

[0020] Fig. 2 is a schematic representation of components of a multiphase motor according to one embodiment.

[0021] Fig. 3 is a schematic representation of a set of position detectors for a multiphase motor, according to a first embodiment.

[0022] Fig. 4 is a schematic representation of a set of position detectors for a multiphase motor, according to a second embodiment.

[0023] The [Fig.5] is a flowchart of steps of a method for detecting a relative position between a rotor and a stator of a multiphase motor, according to one embodiment.

[0024] The [Fig.6] includes a first curve representing the evolution of a detected position, a second curve representing the evolution of an actual position, and a third curve representing a difference between the first curve and the second curve.

[0025] Fig. 7 is a schematic representation of a control module for a multiphase motor coil, according to one embodiment.

[0026] The [Fig.8] is a flowchart of steps of a method for controlling a multiphase motor coil, according to one embodiment.

[0027] Throughout the figures, similar elements bear identical references. DETAILED DESCRIPTION OF THE INVENTION Multiphase motor#: general architecture

[0028] Figure [1] shows a schematic cross-section of a multiphase motor 1 in a plane (X,Y) defined by two axes X and Y, according to one embodiment.

[0029] The multiphase motor 1 can be installed in an aircraft. In this context, it can serve as a primary or secondary flight control actuator, as a pump actuator (for example, for fuel pumps), or even be used for aircraft propulsion.

[0030] The multiphase motor 1 comprises a stator 2 and a rotor 4 movable in rotation relative to the stator 2 around an axis Z perpendicular to the plane (X, Y).

[0031] The multiphase motor 1 comprises N coils B1 to BN, with N>3. The coils are fixed to the stator 2. In the embodiment illustrated in [Fig.1], we have N=9. In another embodiment, we have N=3 and the motor is then said to be three-phase.

[0032] The N coils are configured to jointly generate a rotating magnetic field relative to the stator 2, from phase signals with respective indices ranging from 1 to N. The rotating magnetic field results from a superposition of N magnetic fields respectively generated by the N coils. In the literature, the phase signals are sometimes more simply called "phases." The phase signals can be electric currents or electric voltages. As is known, the rotating nature of the magnetic field is obtained using periodic phase signals with phases shifted relative to each other.

[0033] The rotor 4 is designed to be driven in rotation relative to the stator 2 by the effect of the rotating magnetic field, so as to generate torque. To this end, the multiphase motor comprises a plurality of magnets 6 fixed to the rotor 4. The magnets 6 are arranged so as to cooperate magnetically with the coils B1 to BN. The magnets 6 have alternating polarities about the axis of rotation Z, in the sense that when traversing the magnets 6 around the axis of rotation Z of the rotor 4, one alternately encounters a magnet having a north pole oriented in a centripetal direction with respect to the axis Z (towards the stator), and a magnet having a south pole oriented in the centripetal direction.

[0034] With reference to [Fig. 2], the multiphase motor 1 also includes N control modules Cl to CN to control the N coils B1 to BN in parallel. Thus, each control module Ci selectively controls the coil Bi of the same index, and is not intended to control the other coils.

[0035] The multiphase motor 1 also includes N source modules indexed from 1 to N. For any i from 1 to N, the source module Si has the function of providing the control module Ci with data which it uses to control the coil Bi.

[0036] Each source module Si includes a coil availability detector Di and a torque sensor CPLi. In [Fig.2], these components are shown for the source module SI, but are also present in each of the other source modules S2 to SN.

[0037] The availability detector Di is configured to detect whether coils B1 to BN are available or unavailable. An "unavailable coil" is defined as a coil unable to generate a magnetic field that contributes to the formation of the rotating magnetic field. For example, a coil in "open phase" is an unavailable coil. A coil can also be unavailable when it is short-circuited. The availability detector is capable of performing this detection coil by coil, thus determining, among the N coils, which ones are available and which ones are unavailable. The coil availability detector is known from the prior art. By convention, the data produced by the availability detector is called "availability data," this availability data indicating for each coil whether it is available or not.

[0038] The CPLi torque sensor is configured to measure the torque generated by the multiphase motor 1 during the rotation of the rotor 4 relative to the stator 2. The CPLi torque sensor is known from the prior art. This sensor is, for example, a linear Hall effect sensor. Alternatively, the sensor can be an estimator operating from an analytical model.

[0039] Modules and method for detecting a relative position between the rotor and the stator

[0040] Each source module Si also includes a position detection module Pi.

[0041] With reference to [Fig. 3], the position detection module Pi comprises a linear Hall effect sensor Hi. As is known per se, a linear Hall effect sensor, sometimes also called an analog Hall effect sensor in the literature, differs from a binary or digital Hall effect sensor in its ability to acquire an analog signal, rather than a digital one. A digital signal is limited to two values, whereas an analog signal can take on more values, and therefore provides more detailed information.

[0042] The linear Hall effect sensor Hi is positioned in the multiphase motor to specifically acquire an analog signal varying continuously as a function of the orientation of the rotating magnetic field discussed previously, relative to the stator 2. The Hall effect sensor is fixed to the stator 2.

[0043] For example, the analog signal takes on a maximum value when the north pole of one of the magnets 6 is opposite the linear Hall effect sensor, and a minimum value when the south pole of one of the magnets 6 is opposite the Hall effect sensor. The reverse is also possible. During the rotation of the rotating magnetic field, the analog signal will oscillate continuously between the maximum and minimum values, due to the alternating arrangement of the magnets 6 fixed to the rotor 4. This oscillation constitutes a source of information. much richer than a simple "top" acquired by a binary Hall effect sensor, that is to say a simple alternation between two values ​​such as 0 and 1.

[0044] The Pi position detection module further includes an RPi neural network.

[0045] In this text, the term "neural network" refers to a neural network Artificial intelligence implemented by any processor, having been trained to perform a specific task. Thus, when we say that the engine includes a neural network, this implies that the engine includes the hardware necessary to implement this task.

[0046] The RPi neural network is configured to determine a relative position between the rotor 4 and the stator 2 from the analog signal provided by the analog Hall effect sensor Hi.

[0047] To perform this function, the RPi neural network underwent pre-training using training data. For example, this training is supervised, and the engine is instrumented with a resolver that is used as a truth source. The engine is started, the neural network provides a position, which is compared to a reference position provided by the resolver, which is assumed to be reliable. If the position provided by the RPi neural network deviates from a position provided by the resolver (assumed to be reliable), then the weights of the neural network are updated to reduce this deviation, for example, by the backpropagation method. Once the training is complete, the resolver is removed from the engine.

[0048] The RPi neural network preferably comprises a single hidden layer. In this case, the neuron comprises only three layers: an input layer, the hidden layer, and an output layer.

[0049] The hidden layer preferably comprises at least three neurons, for example 10 neurons.

[0050] The hidden layer is preferably fully connected to the input layer, which means that each neuron in the hidden layer will process the analog signal.

[0051] The hidden layer is preferably fully connected to the output layer, meaning that each neuron in the hidden layer is connected to the output layer.

[0052] Furthermore, the RPi neural network preferentially uses the ReLu function as its activation function. The advantage of this ReLu function is that it limits the computational load consumed by the neural network. However, other activation functions can be used, for example, sigmoid or hyperbolic tangent functions.

[0053] The position detector Pi further includes a smoothing unit Li arranged at the output of the RPi neural network. The smoothing unit Li is configured to apply smoothing to data produced by the neural network.

[0054] In practice, the RPi neural network and the Li smoothing unit can take the form of computer programs or parts of a single computer program executed by a processor of any structure. This processor may comprise one or more cores. For example, the processor is a microcontroller, a DSP, an FPGA, an ASIC, etc.

[0055] The linear Hall effect sensors H1 to HN are arranged around the axis of rotation of the rotor 4. Preferably, for any i from 1 to N, the linear Hall effect sensor of index i is located at least partially inside the coil of index i. In other words, at least one turn of the coil of index i extends around the linear Hall effect sensor of index i. In [Fig. 1], the linear Hall effect sensors H1 to Hn are represented by black rectangles.

[0056] In a decentralized embodiment of the position detectors, represented in [Fig.3], the neural network of index i is connected at the input only to the linear Hall effect sensor of index i.

[0057] However, in other embodiments, an RPi neural network can receive signals from several linear Hall effect sensors among the sensors H1 to HN. Figure 4 thus shows an embodiment in which any RPi neural network receives signals from each of the N sensors H1 to HN. In another embodiment, an RPi neural network can selectively receive signals from K adjacent Hall effect sensors, with 1 <K<N, par exemple K capteurs adjacents autour de l’axe de rotation Z.

[0058] A method for detecting a relative position between the rotor 4 and the stator 2 of the multiphase motor 1, implemented by the source module of index Pi, comprises the following steps.

[0059] It is assumed that the coils B1 to BN have been pre-controlled so as to jointly generate the rotating magnetic field described above, causing the rotor 4 to rotate relative to the stator 2. During this rotation, the orientation of the rotating magnetic field varies, and the magnets 6 move relative to the linear Hall effect sensor Hi.

[0060] In an acquisition step 100, the linear Hall effect sensor Hi acquires the analog signal described above, varying continuously as a function of the orientation of a magnetic field rotating with the rotor 4 relative to the stator 2.

[0061] In a determination step 102, the RPi neural network takes the analog signal as input, and generates on its basis a relative position between the rotor 4 and the stator 2.

[0062] The combination of the linear Hall effect sensor and the neural network makes it possible to accurately detect the relative position between the rotor 4 and the stator 2 of the motor multiphase 1, including when the rotor 4 is rotating at low speed relative to the stator 2.

[0063] Preferably, the relative position consists of Cartesian coordinates in a plane perpendicular to an axis of rotation of the rotor 4 with respect to the stator 2. For example, if working with normalized coordinates, the Cartesian coordinates consist of the cosine and sine of an angle indicating the relative orientation between the rotor 4 and the stator 2 in the plane considered. Such Cartesian coordinates have the advantage of evolving continuously over time, unlike the aforementioned angle which evolves from 0 to 2ir, then returns to 0 once the magnetic field has completed a full rotation around the axis of rotation of the rotor 4. The inventors have observed that this discontinuity in the value of the angle degrades the performance of the neural network, especially when the size of the neural network is limited.In comparison, it is much easier for the neural network to learn to estimate continuously changing coordinates, especially when the network size is limited.

[0064] The RPi neural network can generate the relative position between the rotor 4 and the stator 2 on the basis of a single value taken by the analog signal at a time t.

[0065] Alternatively, the neural network can generate this position based on several values ​​successively taken by the analog signal at different times, therefore not only a value at time t, but also previous values ​​of the signal, acquired at respective earlier times t-1, t-2,... tk and constituting a history of values. This history of values ​​allows the neural network to perceive the dynamics of the rotor 4, enabling it to improve its accuracy.

[0066] Alternatively or in addition, the RPi neural network can determine the relative position between the rotor and the stator on the basis of analog signals respectively provided by several linear Hall effect sensors among the N available, or even for example by the N sensors in the embodiment shown in [Fig.4].

[0067] The preceding steps are repeated over time, so that the RPi neural network is led to generate a succession of relative positions between the rotor 4 and the stator 2.

[0068] This sequence of positions may contain noise. As an example, Figure 6 shows the evolution of the sine of the angle determined by the neural network (top), the actual evolution of this sine of the angle (middle), and the difference between these two evolutions (bottom). The presence of peaks in this difference shows that the evolution of the sine determined by the neural network is subject to occasional noise.

[0069] To eliminate or at least reduce this noise, the smoothing unit Li implements, in step 104, a smoothing of the sequence of positions generated by the neural network. The inventors have found that such smoothing can, in certain cases, reduce An amplitude error at the peak level of 10% to 1%. For smoothing implementation, the smoothing unit Li can use a moving average or include a first-order recursive filter, which is simpler to implement in real time because it is less expensive. In principle, the signal is at the motor's rotational frequency, so it is preferable to filter only the high frequencies using a high-pass filter.

[0070] The position determined by the position detector Pi can advantageously be used by the control module Ci to control the coil Bi.

[0071] Modules and method for controlling the coils of the multiphase motor

[0072] With reference to [Fig.7], for any i from 1 to N, the control module Ci comprises an input interface INi, a phase generator RAi, a neural network RBi, a correction module Ai and an output interface Oi.

[0073] The input interface INi is suitable for receiving input data provided by the source module Si. This input interface can have any structure, for example wired or wireless radio.

[0074] The phase generator RAi is configured to generate a phase-compensated signal of index i from supplied availability data and a relative position between the rotor (4) and the stator (2). The phase-compensated signal could be used by the coil Bi to generate its magnetic field, contributing to the formation of the rotating magnetic field described above.

[0075] The compensated phase generator takes into account possible unavailabilities indicated by the availability data, so that the compensated phase signal helps to reduce an oscillation of the torque generated by the multiphase motor, this oscillation occurring in the event of unavailability of one or more of the N coils.

[0076] In one embodiment, the RAi phase generator implements the MTPA (Maximum Torque Per Ampere) algorithm, which is known per se, to achieve this objective. In the literature, phase signals produced using the MTPA algorithm are referred to as "optimal." Those skilled in the art will refer, for example, to the following documents for implementing this MTPA algorithm: • N, Nguyen, “Chapter 3#: Synchronous machines in degraded mode#: analysis of failure modes and optimal power supplies of synchronous actuators in the presence of faults.”, 2011. • X. Kestelyn and E. Semail, “A Vectorial Approach for Generation of Optimal Current References for Multiphase Permanent-Magnet Synchronous Machines in Real Time”, IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, vol. 58, no. 11, 2011.

[0077] In another embodiment, the RAi phase generator is a neural network that has been trained to fulfill the same objective as the MTPA algorithm. To train the RAi neural network, data from classical optimal phase signal determination methods described in the documents above are used as training data. The input and output parameters described so far are used. Specifically, the training data can take the form of an Excel spreadsheet, with each column containing one of the parameters that the network will receive as input and output in the future. Each row corresponds to a case. From this spreadsheet, the network receives the input columns from one row at a time and estimates the output, which is then compared to the actual output to train the network in a supervised manner.

[0078] The training data for the RAi neural network relates to healthy motors and motors including unavailable coils, in particular open-phase coils or coils connected to an open circuit. For example, if it is desired that the RAi neural network be able to correct situations in which two of the N coils are unavailable, the training data includes network input and output data illustrating this scenario. The classical MTPA algorithm is used as the truth source for updating the neural network weights during training. This update is classically performed by backpropagation of gradients.

[0079] The RAi neural network preferably comprises a single hidden layer. In this case, the neural network consists of three layers: an input layer, the hidden layer, and an output layer. The hidden layer comprises, for example, five neurons.

[0080] Preferably, the RAi neural network uses a radial basis function as its activation function, which has the advantage of being computationally inexpensive. This function offers performance comparable to the tanH function, for example, but allows for a drastic increase in the number of neurons and therefore the accuracy of the network.

[0081] The RBi neural network is configured to generate a patch for the phase-compensated signal index i. The objective of this patch is to further reduce the oscillation of the torque generated by the multiphase motor, taking into account faults other than coil unavailability. For example, a short-circuited coil still has the capacity to generate a magnetic field, but this magnetic field then becomes a disturbance that prevents the torque generated by the multiphase motor from remaining perfectly constant. Thus, a "short-circuited coil" type fault is a cause of torque oscillation.

[0082] To train the RBi neural network, training data are used that cover a wider variety of fault cases affecting torque than those covered by the training data for the RAi neural network. The RBi neural network training data includes patches that may These can be obtained by subtracting optimal phase signals determined using the MTPA algorithm from faulty phase signals. It should be noted that the MTPA algorithm can be used as a source of truth if the learning is supervised, in which case the algorithm is parameterized to correct faults other than unavailabilities, as taught in the document AG Yepes, I. Gonzalez-Prieto, O. Lopez, MJ Duran, and J. Doval-Gandoy, “A Comprehensive Survey on Fault Tolerance in Multiphase AC Drives, Part 2: Phase and Switch Open-Circuit Faults”, 2022.

[0083] For example, faults covered by the RBi neural network training data include: short-circuited coils, short-circuited switches, faults related to a back electromotive force under load that differs from a no-load back electromotive force, or significant reluctant torques. Other faults that can be trained include the beginnings of partial discharge marks, mechanical play, or changes in contact resistances within the motor.

[0084] The correction module Ai is configured to apply the correction to the phase-compensated signal of index i, so as to produce a phase-corrected signal of index i. The correction module is, for example, an adder.

[0085] The output interface Oi is in communication with the coil Bi, so as to be able to transmit the phase-corrected signal of index i to the coil Bi.

[0086] A method implemented to control a coil Bi comprises the following steps.

[0087] In step 200, the CPLi torque sensor acquires a torque signal representing a time evolution of the torque generated by the multiphase motor (1). When the motor is operating normally (in particular when all the coils Bi are available), this torque signal is assumed to be constant. However, in the presence of a fault (for example, if one or more coils are unavailable), the torque signal may begin to oscillate.

[0088] In step 204, the availability detector Di generates availability data indicating, for any i from 1 to N, whether the coil Bi is available or unavailable, that is, whether the coil Bi is capable of contributing to the generation of the rotating magnetic field driving the rotating rotor. When the coil Bi is available, it is possible to adjust the current flowing through it (amplitude and phase) in relation to a rotor position, which is not the case when the coil is unavailable. The availability data can thus take the form of a vector of N Booleans. The i-th Boolean component of the vector defines the state of the coil Bi. If this i-th component is 1, the coil Bi is considered available, and if this component is 0, the coil Bi is considered unavailable.

[0089] Furthermore, in a step 206, the position detector Pi determines a relative position between the rotor (4) and the stator (2), using for example the detection method which has been described previously.

[0090] Steps 200, 204, 206 can be carried out in any order or in parallel.

[0091] The source module Si provides the control module Ci with input data including: • the relative position between the rotor (4) and the stator (2), having been determined by the position detector Pi; • the torque measured by the CPLi torque sensor; • the availability data provided by the availability detector Di.

[0092] In a step 208, the phase generator RAi generates a phase-compensated index i signal from the availability data and the relative position between the rotor 2 and the stator 4. As previously stated, the phase-compensated index i signal would contribute, if used directly by the coil Bi, to reducing an oscillation of the torque generated by the motor in the event of unavailability of one or more of the coils B1 to BN.

[0093] In a step 210, the RBi neural network generates a correction for the phase-compensated signal of index i, based on availability data, the relative position between the rotor 2 and the stator 4, and an error between a setpoint torque and the torque generated by the multiphase motor received by the input interface.

[0094] It should be noted that the torque error can be calculated by the control module Ci, or alternatively calculated by the source module Si, in which case the source module Si directly provides this torque error instead of the torque measured by the torque detector CPLi.

[0095] In [Fig.8], the step of generating the correction 210 is presented as subsequent to the step 208 of generating the phase-compensated signal only to simplify the figure; in reality, these steps are carried out in parallel, in any order.

[0096] In a step 212, the correction module Ai applies the correction to the phase-compensated signal of index i, so as to produce a phase-corrected signal of index i. When this module is an adder, the phase-corrected signal of index i is obtained by summing the phase-compensated signal of index i and the correction.

[0097] In a step 214, the output interface Oi transmits the phase-corrected signal of index i to the coil Bi to control it.

[0098] It should be noted that the phase-compensated signal of index i, the correction and the phase-corrected signal of index i can be current signals or voltage signals.

[0099] Upon receiving the phase-corrected signal of index i, the coil Bi generates a magnetic field.

[0100] The preceding steps are carried out in parallel by the control modules Cl to CN, so that the coils B1 to BN generate magnetic fields which combine to form the rotating electric field causing the rotor (4) to rotate relative to the stator (2).

[0101] If the phase signals with respective indices from 1 to N, originating solely from the MTPA-simulating network, were directly transmitted to the coils B1 to BN, they would cause a reduction in the oscillation of the torque generated by the multiphase motor in the event of one or more coils being unavailable. However, the torque could continue to oscillate due to other faults in the motor. Therefore, phase signals with respective indices from 1 to N, which use the correction network in addition to the MTPA network, are used to control the coils B1 to BN. These signals take into account faults beyond simple coil unavailability, and thus allow for a further reduction of any potential oscillations in the torque generated by the multiphase motor. Other implementation methods

[0102] In the description above, it was assumed that each neural network used is trained in a supervised manner. This implies that a large amount of data is available. Alternatively, it is possible to train either of the neural networks used with unsupervised methods such as reinforcement learning, which allows training without training data, but based on a simple model.

Claims

Demands

1. Multiphase motor (1) comprising: a stator (2), N coils (Bl,..., BN) with respective indices ranging from 1 to N, with N>3, the N coils being configured to jointly generate a rotating magnetic field from phase signals of respective indices ranging from 1 to N, a rotor (4) suitable for being driven in rotation relative to the stator (2) under the effect of the rotating magnetic field so as to generate a torque, N control modules (Cl, ..., CN) with respective indices from 1 to N to control the N coils in parallel respectively, in which for any i from 1 to N, the control module (Ci) with index i comprises: an input interface (INi) to obtain Input data including: • availability data indicating, for any i from 1 to N, whether the coil with index i is available or unavailable to contribute to generating the rotating magnetic field, • a relative position between the rotor (4) and the stator (2), • a torque error between a setpoint torque and a torque generated by the multiphase motor (1), a generator (RAi) configured to generate a phase-compensated index i signal from availability and relative position data, the phase-compensated index i signal being adapted to help reduce an oscillation of the generated torque, a neural network (BN) configured to generate, from availability data, relative position and torque error, a • a correction index i taking into account faults other than coil unavailability, • a correction module (Ai) configured to apply the correction to the phase-compensated index i signal so as to produce a phase-corrected index i signal, the phase-corrected index i signal being adapted to contribute to a further reduction of the oscillation of the generated torque, • an output interface (Oi) configured to control the index i coil using the phase-corrected index i signal.

2. Multiphase motor (1) according to the preceding claim, wherein the neural network (RBi) has been pre-trained with learning data relating to healthy multiphase motors, multiphase motors including unavailable coils and motors suffering from defects other than coil unavailability.

3. Multiphase motor (1) according to any one of the preceding claims, wherein the generator (RAi) is another neural network.

4. Multiphase motor (1) according to the preceding claim, wherein the other neural network has been previously trained with learning data relating selectively to healthy multiphase motors and multiphase motors comprising unavailable coils.

5. Multiphase motor (1) according to any one of the preceding claims, wherein the neural network or each neural network comprises only one hidden layer.

6. Multiphase motor (1) according to any one of the preceding claims, wherein the or each neural network is configured to use a radial basis function as an activation function.

7. Multiphase motor (1) according to any one of the preceding claims, wherein the correction module (Ai) is an adder configured to sum the phase-compensated signal of index i and the correction of index i.

8. Multiphase motor (1) according to any one of the preceding claims, further comprising: • a linear hall effect sensor (Hi) configured to acquire an analog signal varying continuously according to an orientation of the rotating magnetic field, • an additional neural network (RPi) configured to determine the relative position between the rotor (4) and the stator (2) from the analog signal.

9. Multiphase motor (1) according to any one of the preceding claims, comprising: • N source modules of respective indices from 1 to N independent of each other, in which for any i from 1 to N, the source module of index i is configured to provide input data of index i to the control module of index i.

10. A method for controlling a multiphase motor, the multiphase motor comprising: • a stator, • N coils of respective indices from 1 to N, with N > 3, the N coils being configured to jointly generate a rotating magnetic field from phase signals of respective indices from 1 to N, • a rotor suitable for being driven in rotation relative to the stator under the effect of the rotating magnetic field so as to generate a torque, the method comprising carrying out the following steps, performed for any i from 1 to N: • ​​obtaining input data comprising: • availability data indicating, for any i from 1 to N, whether the coil of index i is available or unavailable to contribute to generating the rotating magnetic field, • a relative position between the rotor and the stator, • a torque error between a setpoint torque and a torque generated by the multiphase motor, generate (208) a phase-compensated index i signal from the availability and relative position data, the phase-compensated index i signal being adapted to contribute to a reduction of an oscillation of the generated torque, generate (210) by a neural network and from the availability, relative position and torque error data, a correction of index i taking into account faults other than coil unavailability, apply (212) the correction to the phase-compensated index i signal so as to produce a phase-corrected index i signal, the phase-corrected index i signal being adapted to contribute to a further reduction of the oscillation of the generated torque, control the coil of index i using the phase-corrected index i signal.