WO2015185074A1

WO2015185074A1 - Neural-network based selective upstream filtering by frequency modulation

Info

Publication number: WO2015185074A1
Application number: PCT/EP2014/003294
Authority: WO
Inventors: Chao Lin
Original assignee: Continental Automotive France; Continental Automotive Gmbh
Priority date: 2014-06-05
Filing date: 2014-12-09
Publication date: 2015-12-10
Also published as: FR3022082A1

Abstract

Method for selective dynamic filtering of channels in the FM band, the method comprising the steps: /a/ selecting a first frequency channel (F0) carrying signals from a station of interest, /b1/ applying an input filter (11) centred on F0 to the received signals, performing a transposition into a baseband and digitizing (12) the filtered signal, /b2/ applying a Fourier transform (14) to obtain a spectrum of the filtered signal in the baseband, /c/ introducing the points of the spectrum obtained into a neural network to identify a particular selective digital filter from a plurality of predefined digital filters (8), and applying the selective digital filter chosen, /d/ continuously repeating steps /b1/ /b2/ and /c/ until a change in the preferred channel occurs.

Description

Upstream selective frequency modulation filtering based on neural network

The present invention relates to vehicle audio systems and particularly to filtering methods for dynamic selective filtering of frequency modulation (FM) broadcast channels.

More specifically, we are interested in audio systems capable of receiving and demodulating electromagnetic waves broadcasting radio programs on the frequency modulation.

In the frequency modulation band, useful channels separated by 200 kHz or 300 kHz can commonly be found in terms of carrier frequency, but it also happens that two adjacent channels have carrier frequencies separated only by 100 kHz, which poses a problem of discrimination and selective filtering to focus the reception on one of the channels and reject the signals of the other channel.

In known systems, a more or less narrow input filter is applied depending on the reception conditions which are estimated after demodulation, that is to say on the audio quality; the width of the filter is controlled by a feedback loop.

But it happens that this feedback loop can introduce unnecessary instabilities, even harmful, in the system, especially in case of changes in FM content and / or the presence of adjacent channels or various electromagnetic disturbances and / or changes in conditions of reception.

It has therefore appeared a need to provide a more efficient solution for selectively filtering the input signals, in order to ensure optimized reception of the channel of interest even with a nearby adjacent channel.

According to the invention, there is provided a method for dynamic selective filtering of electromagnetic waves broadcasting radio programs on the frequency modulation, the method comprising the steps of:

lai select a first frequency channel (F0) carrying signals of a station of interest intended to be listened to,

/ b1 / applying to the received signals an input filter centered on the frequency channel, performing a baseband transposition, and digitizing the filtered signal, / b2 / applying a Fourier transform, to obtain a spectrum of the band-filtered signal basic,

Here, introduce the points of the spectrum obtained in a neural network, to identify a particular selective digital filter among a plurality of predefined digital filters, and apply the selective digital filter chosen, 161 to repeat continuously the steps / b1 / / b2 / and Here up to 'to a change of frequency channel. Through the use of the highly parallel process of the neural network, this allows the most relevant particular digital filter to be selected from among a plurality of predefined, selective digital filters, without applying complex calculation rules.

Demodulated audio signals are thus obtained which reproduce the signals diffused by the first channel of interest by strongly canceling or limiting the disturbances induced by the presence of a second disturbing adjacent channel, or the presence of interference outside the audio content.

Note also that, thanks to the transposition to the baseband, all digital processing is advantageously in the baseband, quite easily.

In various embodiments of the method according to the invention, one or more of the following provisions may also be used:

The neural network is advantageously a multi-layered unidirectional neuron network (also called multilayer "forward") with preferably four layers; so that one uses a conventional and well-controlled neural network structure.

• The plurality of predefined digital selective filters can comprise between 20 and 50 predefined digital selective filters (typical value 16 filters); thus, no need to recalculate / parameterize a filter in real time, we choose a predefined filter already set.

• The spectrum curve resulting from the Fourier transform can comprise between 128 and 1024 points (typical value 256 points); the characteristics of the spectrum of the received signal can thus be treated with great finesse.

· The weighting coefficients of the neural network can be determined by field test campaigns; therefore, the parameterization of the coefficients in the neural network is established once this initial learning period has been completed.

• The weighting coefficients of the neural network are determined by retro propagation of the error gradient; a well-known and well-known learning mechanism is used in the field of neural networks.

• Digitization, Fourier transform and neural network processing are preferably repeated at least every 500 ps. This provides rapid decision means so as to adjust the decision very dynamically. • The input filter has a width of about 300 kHz. This advantageously eliminates upstream all the channels whose carrier is separated by 200 kHz or more from the carrier of interest.

The invention also relates to an audio system capable of receiving and demodulating electromagnetic waves broadcasting radio programs on the frequency modulation, remarkable in that it is configured to implement the method according to one of the preceding characteristics; such a system has the advantages mentioned above.

Other aspects, objects and advantages of the invention will appear on reading the following description of one of its embodiments, given by way of non-limiting example. The invention will also be better understood with reference to the accompanying drawings in which:

FIG. 1 schematically represents a partial electromagnetic spectrum in the FM band and its evolution over time, with a channel of interest and a potentially disruptive adjacent channel,

FIG. 2 represents a block diagram of an audio system and according to the invention,

FIG. 3 diagrammatically represents an electromagnetic spectrum, transposed from that of FIG. 1 to the baseband,

FIG. 4 illustrates different cases of the spectral curve resulting from the Fourier transform, in particular its variations over time,

FIG. 5 diagrammatically represents the respective audio contents of the channel of interest and of a potentially disruptive adjacent channel, over time,

FIG. 6 represents a neural network type structure used in the proposed method.

In the different figures, the same references designate identical or similar elements.

Figures 1 and 2 illustrate an audio system 1 for a vehicle, also called car radio, capable of receiving and demodulating electromagnetic waves broadcasting radio programs on the FM ("frequency modulation"). With regard to radio programs, as known per se, the principle of frequency modulation is based on the use of a frequency deviation with respect to a carrier frequency (F0) for encoding the audio signals to be broadcast. The frequency deviation is proportional to the encoder audio signal.

A plurality of channels may be broadcast over the entire FM band (88 MHz - 108 MHz in many countries, 76 MHz - 90 MHz in Japan). Each channel typically has a useful bandwidth of about 240 kHz, i.e. +/- 120 kHz around the carrier; this corresponds to a starting composite audio spectrum (Mono + Stereo + RDS baseband) bounded by 60 kHz, modulating the carrier according to its amplitude and its spectral content to give a channel of at most 240 kHz wide.

It is therefore understood that when two channels are separated, in terms of carrier frequency F0, by 100 kHz, there may be a slight overlap of signals on the spectrum, a problem to which the present invention seeks to provide a solution. .

In the remainder of this document, the frequency F0 designates the carrier frequency of a first channel, also called channel of interest, that the user of the audio system has selected, this user wishing to listen to the radio program broadcast by this first channel. carrier frequency F0. We note that the spectral content of this first channel evolves over time, mainly because of the modulating audio content that also evolves over time.

The spectrum is not always centered on F0, and in particular the location of the maximum 71 moves around F0 continuously according to the audio content.

FIG. 1 illustrates the evolution of the spectral content of the carrier interest channel F0 as well as the spectral content of a second channel F2, which will be referred to here as the interfering channel. The second channel has a carrier frequency F2 not far from the carrier frequency F0, for example the difference can be 100 kHz. The disruptive channel may be located either above (as illustrated) or below the frequency of the first channel F0. The spectral content of the second channel also evolves with time independently of the spectral content of the first channel.

In FIG. 2, the audio system 1 comprises an analog front part 2, a digital central part 4 with a logic control unit 3 (digital core), and an analog downstream part with an amplifier 19 and loudspeakers 9. By elsewhere, the audio system includes a user interface 6 with a display screen and a touchpad or physical buttons (not shown); This allows the user to choose the station / channel of interest.

The electromagnetic signals are received by an antenna device 10, the signals are then filtered by an input filter 1. Advantageously, the input filter can be completed by a frequency change operation (offset F0) to transpose baseband signals. The signals from the filtering are then digitized by a digital analog converter 12.

The further processing of the signal is performed by digital operations, including storage in a buffer memory of Data 13. From the buffered data, a Fourier transform analysis 14 is performed which outputs a power spectral density curve which will be discussed in detail later. In addition, the buffered data is digitally filtered by a digital filter of programmable characteristics. A filter width control block 16 controls the selection of the filter characteristics based on the content analysis of the power spectral density curve.

All or part of the operations of buffering, Fourier transform, digital filtering can be performed by a DSP-type dedicated circuit 5. The term "DSP" comes from the English "Digital Signal Processor" and designates a signal processor digital.

Note that the input filter 1 1 has a width of the order of 200 to 300 kHz centered on F0, which makes it possible to eliminate the signals coming from the channels remote from F0, but does not make it possible to eliminate correctly the signals of a nearby adjacent channel whose spectral density may overflow within the input filter.

As illustrated in FIG. 3, in the case where the second potentially interfering channel F2 has a carrier shifted at 100 kHz from F0, there is a spectral overlap area where the curves intersect.

The signals downstream of the input filtering, after digitization and spectral transposition, then Fourier transform processing are illustrated in FIG. 4.

It should be understood that with reference to FIGS. 3 and 4, they illustrate the baseband processing i.e. after frequency transposition of -F0.

Analysis of the spectral density curve at the output of block 14 makes it possible to identify local minima, which are generally symptomatic of the presence of an adjacent channel that interferes with the channel of interest F0.

If no minimum is identified, as for example in the case of curve 93 in FIG. 4, the width of the digital filter can then be relatively large and preferably deduced from the spectral content of the channel of interest.

In Figure 4, in dotted lines, in another case, there is a local minimum 91 b to the left of the center frequency, with a boss 91a symptomatic of the presence of an adjacent channel on the left.

Similarly, in another case, in the case of curve 92, there is a local maximum (bump) 92a to the right of a local minimum 92b symptomatic of the presence of an adjacent channel on the right.

Typically, the digitization and the Fourier transform are repeated at least every 500 με. Advantageously according to the present invention, it will avoid performing a conventional analysis of the extremums of the spectrum curve, whose profile can change rapidly over time; such an analysis of the extremums may indeed require significant computing power. Above all, it turns out that under certain extreme conditions (which were not foreseen by the rules), a classical analysis of the spectrum curve is not enough to find the most suitable filter.

Instead of such an analytical study, it is advantageous to use a neural network structure, which has the characteristic of being able to perform a large number of parallel treatments.

The role of the neural network is to identify a predefined and parameterized filter in advance, said filter being determined as the most suitable among a plurality of predefined filters available in a memory space 8.

For this purpose, several digital centered filters of different widths are defined in advance, for example each having a width of +/- 10 kHz relative to its neighbor, (or +/- 20 kHz); the plurality of predefined filters may also include non-centered filters, i.e. with a left terminal positioned at a distance of zero different from the right terminal.

In practice, the plurality 8 of predefined digital filters can thus comprise an amount of between 20 and 50 filters, typically for example 16 filters. In Figure 6 is shown an example of neural network structure. It is a multilayer structure (or "levels"), of unidirectional type (also called "forward" or "feedforward") that is to say that each neuron of a given layer produces a result (an output) according to a weighted function of signals output from one or more neurons of the previous layer (left in the figure), there is no feedback or feedback / iteration in such a structure. It is therefore a network of neurons simple to implement, reliable and commonly used.

In the illustrated example, the first layer on the left contains as many input neurons 61 as J points in the spectrum curve from the fast Fourier transform. This number J of points can be understood in practice between 256 and 1024 points, typically 256 points.

Each neuron of the first layer (squares) 61 may be connected at the output (reference 91) to several neurons of the second layer (among K neurons) 62, for example between 10 and 50 or more.

Similarly, each neuron of the second (round) layer 62 may be connected at the output (reference 92) to several neurons of the third layer (among M neurons) 63, for example between 10 and 50 or more. Finally, each neuron of the third layer (triangles) 63 may be connected at the output (reference 93) to several neurons of the fourth and last layer (stars) 64, for example between 10 and 50 or more.

We can choose as activation function of each neuron either a sigmoid function or a hyperbolic tangent function.

The number of outputs of each layer decreases with the layer number that converges the network. Preferably we have J> K> M> N.

In the last layer, there are as many neurons N as predefined filters in the storage space 8. In the example illustrated, it is therefore the neuron that has the highest output that will be designated to identify the filter 15 the most adequate.

For each neuron, the weighting coefficients affecting each of the links from the neurons of the previous layer are obtained by a learning mechanism, starting from default values.

These weights are established by a learning method, supervised or not, during field test campaigns; a test vehicle equipped with the audio system and according to the invention thus traverses a wide variety of geographical areas with different conditions for broadcasting the FM channels, the different results produced being evaluated, that is to say, rated or classified, by a qualified professional, which allows to refine the value of the weights.

As an alternative, the weighting coefficients can be learned by passing predefined test sets each consisting of a pair (input spectrum, output selected filters), these test sets possibly being achieved by previous field trial campaigns.

In practice, the learning process will lead to weighting coefficients that will tend to choose a rather narrow filter when the content of the channel of interest has a narrow spectrum or when the spectrum has one or more lateral bumps (eg 91a , 92a), while the weighting coefficients will tend to lead to a rather wide filter choice when the content of the channel of interest has a broad spectrum without lateral bump.

With reference to FIG. 5, during the first illustrated time period 81, the spectral density curve resulting from the Fourier transform does not reveal a lateral hump and consequently the neural network selects a digital filter 151 which corresponds to a width W by default for example 100 kHz wide.

At the beginning of the second illustrated time period 82, the spectral curve comprises a lateral hump induced by the presence of the adjacent channel F2 whose spectral content has approached the central frequency of interest F0. The curve treatment of the spectral density resulting from the Fourier transform by the neuron network leads to choosing a digital filter 152 which is tighter; and the output of the neural network selects a filter increasingly tight.

At the beginning of the third illustrated time period 83, as the interference between the two channels decreases, the output of the neural network selects a filter less and less tight.

At the beginning of the fourth illustrated time period 84, the spectrum becomes classic again without lateral hump, and the output of the neural network selects a standard width filter.

At the following time period 85, we note that the neural network selects a filter 155 off-center to the left to follow the spectral content of the channel of interest.

At the illustrated time period 86, the spectrum again shows a right side bump and therefore the output of the neural network selects a narrower filter.

During periods of time 87, 88, the spectrum becomes classic again without lateral hump, and the output of the neural network selects a filter of standard width.

Claims

A method of dynamic selective filtering of electromagnetic waves broadcasting radio programs over frequency modulation (FM), the method comprising the steps of:

/ b1 / applying to the received signals an input filter (11) centered on F0, performing a baseband transposition, and digitizing (12) the filtered signal,

/ b2 / applying a Fourier transform (14), to obtain a spectrum of the baseband filtered signal,

Here introduce the points of the spectrum obtained in a neural network, to identify a particular selective digital filter among a plurality of predefined digital selective filters, and apply the selected selective digital filter, to continually repeat the steps / b1 / / b2 / and Here until a frequency channel change,

so that the most suitable predefined filter can be selected very quickly thanks to the parallel operations performed by the neural network.

2. The method of claim 1, wherein the neural network is a multilayer unidirectional neural network, preferably with four layers.

3. Method according to one of claims 1 or 2, wherein the plurality of predefined digital selective filters comprises between 20 and 50 predefined digital selective filters.

4. Method according to any one of claims 1 to 3, wherein the spectrum curve from the Fourier transform comprises between 128 and 1024 points.

The method of any one of claims 1 to 4, wherein the weighting coefficients of the neural network are determined by field trial campaigns.

6. Method according to any one of claims 1 to 4, wherein the weighting coefficients of the neural network are determined by retro propagation of the error gradient.

7. Method according to any one of claims 1 to 6, wherein the digitization, the Fourier transform and the treatment by the neural network are repeated at least every 500 με.

The method of any one of claims 1 to 7, wherein the input filter (11) has a width of about 300 kHz.

An audio system capable of receiving and demodulating electromagnetic waves broadcasting radio programs on the frequency modulation, characterized in that it is configured to implement the method of any one of the preceding claims.