EP0337868B1 - Method and apparatus for signal discrimination - Google Patents

Description

The present invention relates to discrimination between signals, in particular for recognizing the presence of a program audio-frequency signal having a determined bandwidth in an input signal capable of being affected by substantially stationary noise.

The invention finds a particularly important application in broadcasting or transmission networks having centers to which arrive and from which signals carrying broadcasting or television programs in more or less significant number depending on geographic location and importance of the center in the network. The larger centers are generally operated by an agent on site. But others are tele-operated from a central station. In all cases, incidents occurring with signals distributed by a center must be detected as quickly as possible, so that the measures necessary to ensure continuity and quality of service can be taken.

A particularly serious incident is the accidental interruption of the program, which must be detected very quickly. It is also desirable to monitor the parameters influencing the quality of broadcasting and transmission throughout the routing of programs to the listener. In automated centers, verification can only be ensured by a local automated system, which, depending on its degree of development, will command the required maneuvers or will limit itself to transmitting an alarm to a central station, hence the required instructions (go to an emergency route, diversion, ...) will be provided to the local machine.

Various methods are currently used to detect the presence of a signal. None allows perfectly achieving two results to a certain extent contradictory, namely the detection of all incidents interrupting the program and the absence of false alarms, and this automatically.

The usual detection criteria are based on the comparison between the amplitude of the transmitted or broadcast signal and a threshold: if the signal present has an amplitude below a certain predetermined threshold, it is considered that there is a defect and it is concluded that the absence of an audio signal if the fault persists beyond a determined period.

On this basis, a method has already been used which consists in monitoring the signal level, measured by a peak voltmeter or a VU meter. But the inertia of these systems masks the presence of low energy program signals so that the operator must verify the presence or absence of a signal by listening or direct observation. And all the methods based on the comparison between a signal amplitude and a threshold are faulted in many cases: if for example a sinusoidal measurement or alignment signal is transmitted instead of an audio program signal, no fault is detected; the presence of program signals of low amplitude (corresponding for example to orchestra pianissimi) is interpreted as an absence of signal; programs such as radio games or movies with long ranges of very low level audio-frequency signals can also give rise to nuisance alarms. In the case of very polluted signals, the noise level can be such that the threshold must be chosen at a high level and it is then no longer possible to detect with certainty the presence of a signal. The useful signal can even be completely drowned in noise.

The present invention aims to provide a method and a device for recognizing the presence of a program audio-frequency signal (the term audio does not imply that these are signals representing sounds) which meet the requirements of practice better than those previously known, in particular by rendering them to a large extent free from the disadvantages of previous systems; it aims in particular to determine, with a high degree of safety, the interruptions of the program signal. For this, the invention is based on the observation that the program signals have stochastic properties which differentiate them from the noise created by the most frequent parasitic sources. For stations placed at a fixed station, for example, the noise characteristics are statistically unchanging or at least change slowly over time, unlike the characteristics of an audio-program frequency signal. In other words, noise generally has properties of ergodicity and stationarity which are not those of a useful signal.

The invention therefore proposes a discrimination method according to claim 1. The sampling frequency is advantageously at least equal to the Shannon frequency when it is also desired to perform an analysis of the signal, for example to determine its quality.

For a different application, namely the discrimination between silence, speech and data transmitted in the voice band in telephony, the article by Y. Yatsuzuka "Highly sensitive speech detector and high speed voiceband data discriminator in DSI-AD CPM system", IEEE transactions on communications, Vol. COM-30, n ° 4, April 1982, pages 739-750 describes a process according to which periods of silence are detected by comparing the ratio of energies over two consecutive blocks to a threshold. The presence of unvoiced phonemes is detected by a high frequency of zero crossings.

The above-defined method involves a decision in the sense of statistics and uses variations in the distribution function of the amplitude values or, in a more elaborate version, variations in the distribution of these values as a decision criterion to detect the absence of a program signal, caused for example by a connection break upstream of the location where the detection is carried out.

If one simply seeks to detect the absence of the program audio signal, the sampling frequency of the input signal can be relatively low. But it is also possible to implement the method for determining in addition a certain number of parameters representative of the quality of the signal and of its characteristics, which makes it possible in particular to determine either the alterations that the signal may undergo during its subsequent delivery to the recipient while remaining of acceptable quality, or corrections to be made. In this case, sampling should be done at least at the Shannon frequency.

The invention also provides a device making it possible to implement the above-defined method, according to claim 9.

• la figure 1 est un synoptique de principe d'un dispositif suivant un mode de réalisation de l'invention ;
• la figure 2 montre une répartition possible d'amplitudes ;
• la figure 3 montre un mode possible de détermination de seuil.

The invention will be better understood on reading the following description of a particular embodiment, given by way of non-limiting example. The description refers to the accompanying drawings, in which:

• Figure 1 is a block diagram of a device according to an embodiment of the invention;
• Figure 2 shows a possible distribution of amplitudes;
• Figure 3 shows a possible mode of threshold determination.

The device, the basic structure of which is shown in FIG. 1, is intended to continuously monitor the presence of three analog program signals and a digital signal, these numbers not being limiting.

Each analog signal is assigned a channel comprising, from the corresponding source, a coupler stage 10, an input stage 12 intended to ensure the impedance matching, the symmetry and the gain control of the channel, and a blocking sampler 14. The input stage can moreover ensure band filtering and pre-emphasis of the audio-frequency signal when necessary. It will first be assumed that the blocking sampler operates at least at the Shannon frequency (for example at 32 kHz or 48 kHz in the case of a signal having a bandwidth between 40 Hz and 15 kHz). A lower sampling frequency can be adopted when simply trying to provide an alarm in the absence of an audio signal.

The output signals of the blocking samplers 14 of all the analog channels are applied to an analog multiplexer 16 which can be provided either for interleaving them, or for processing successively the different ways. They attack an analog / digital converter 18 which converts the amplitude of each sample into a byte representing this amplitude in digital form. The sample-and-hold units 14, the analog multiplexer 16 and the analog / digital converter are clocked by a sequencer 20 provided with a clock and supplying, on an output 21, control signals to select one of the channels at any time.

In the case where the device also receives a digital data signal, the corresponding channel 22 is applied to a digital data decoder 24 providing the digital value of the amplitude of the samples in a form compatible with that provided by the analog / digital converter (CAN) 18. A digital multiplexer 26 receives the digital output signals from the CAN 18 and from the decoder 24 and supplies them on a common output 28. This digital multiplexer receives control signals on the one hand from the sequencer 20, on the other hand part of the decoder 24 which performs clock recovery for this purpose. If the signals to be monitored are all analog, it is possible to dispense with components 24 and 26.

The processing of the signals downstream from the digital multiplexer 26 is the same regardless of the channel from which they originate and regardless of the chosen mode of detection of the absence of signal (determination of the variation in distribution of the amplitude or variation of the amplitude distribution function). A sorting and storage operation is carried out, comparable to a multi-channel analysis. All the samples taken on the same channel during a constant time interval Δt are divided into a certain number n of channels spaced by a determined pitch (1 dB for example) and the samples accumulated in each channel during the time Δt are counted. For this the multichannel analyzer 30 can include a set of comparators 32 and a random access memory 34. The comparators can be reduced to a programmable transcoding network.

Each of the n channels is assigned a corresponding address location in memory 34. The zero order channel corresponds to the minimum amplitude Amin that we want to take into account. Channel n-1 corresponds to the maximum amplitude Amax. In the case of signals intended for the transmission of sound programs, it is often possible to adopt a dynamic of 52 dB, divided into n = 52 channels spaced by 1 dB; channel 0 may correspond to samples of amplitude equal to or less than -30 dBu. Channel 51 will then correspond to samples of amplitude equal to or greater than +22 dBu. Each time a sample whose amplitude corresponds to the order channel i is detected, the content of the corresponding position of the address memory i is incremented by one. Since the input signals are digital, the multichannel analysis can be carried out very simply using a programmable transcoding read-only memory 32 which corresponds, to each digital sample value, the address of one of the n channels. The increment order is sent by a circuit 38 in a conventional manner.

At the end of the time Δt, each memory location contains a number indicating the number of times the signal sample has had a determined amplitude. There is thus a distribution of the amplitudes of the signal during the recording time Δt, which can for example be of the type shown in FIG. 2 and this for each signal to be supervised.

The processing for detecting the absence of a signal (and possibly for evaluating the quality of the signal) is provided by a computer 36. The memory 34 must be addressable in read and write by the read only memory 32, in read and write by the computer 36. It will first be assumed that the processing aims to determine the presence or absence of a signal by taking into account the distribution of the amplitudes.

In the case where the received signal comprises a useful audio program signal, the amplitude distribution and the average amplitude frequently vary over time. For example, one can have distribution curves of the genus designated by I and II over two successive time intervals. To obtain significant data, values between 0.8 s and 5 s can be frequently adopted for the time intervals Δt, advantageously approximately 1 s for spoken programs and approximately 3 s for music programs.

The detection by comparison of the distribution of the amplitudes could for example be done by determination, by the computer 36, of the correlation factor between several successive records, for example between two distributions stored in memory 34 for two successive time intervals Δt.

It is possible to implement the invention in a simpler way which by calculating a correlation factor, for example by determining the percentage of samples whose amplitude exceeds a threshold during each time interval Δt (this is ie by determining the fraction of time Δt during which the signal exceeds the threshold).

• 1/ Déterminer la valeur maximale A_j prise par le signal dans l'intervalle de temps Δt_j et la fonction de répartition F_j(X) de l'amplitude, c'est-à-dire la probabilité P pour qu'un échantillon du signal ait une amplitude dont la valeur X est inférieure à une valeur de référence donnée x, qui peut être fixée une fois pour toutes ou calculée sur chaque intervalle de temps.
  On utilisera souvent une valeur fixe lorsqu'on souhaite faire un évaluation de qualité. Par exemple on sait qu'une bande magnétique, pour ne pas donner de distorsion excessive, doit avoir un niveau de sortie absolu ne dépassant pas un seuil. Avec les bandes ayant des caractéristiques courantes on peut alors adopter x = 18 dBu. De façon similaire, pour vérifier l'absence de risque de distorsion à la sortie d'un émetteur, on peut risque de distorsion à la sortie d'un émetteur, on peut adopter un seuil de 12 dBu.
  Dans le second cas, qui sera celui de la discrimination entre signal de bruit et signal utile, la valeur de x est calculée pour chaque intervalle de temps à partir de la valeur maximale A_j de l'amplitude dans l'intervalle de temps correspondant Δ_j. On peut par exemple adopter, en dehors des périodes de défaut,
  
  ${\text{x}}_{\text{j}}{\text{= A}}_{\text{j}}{\text{- 6 dB pour A}}_{\text{j}}\text{> -20 dBu}$
  
  
  ${\text{x}}_{\text{j}}{\text{= A}}_{\text{j}}{\text{- 3 dB pour A}}_{\text{j}}\text{≦ -20 dBu}$
  
  
  
     On maintiendra par contre x_j à une valeur constante aussi longtemps que le calculateur signale un défaut, quelle que soit l'évolution de la valeur maximale A de l'amplitude.
  Dans ce cas la première étape revient à déterminer la fonction F_j(X) :
  $\text{Fj(x) = P[X < x]}$
  
  avec Amin ≦ x ≦ Amax P[X < x] représente, pour chaque intervalle de temps Δx, la fraction du temps pendant laquelle X est inférieur au seuil x.
  Le CAN 18 peut notamment être prévu tel que l'on ait les valeurs Amin = -30 dBu et Amax = +22 dBu mentionnées plus haut.
• 2/ Effectuer, à la fin de l'enregistrement d'ordre j+1, les mêmes calculs que précédemment pour obtenir l'amplitude maximale A_j+1 et la fonction de répartition F_j(X).
• 3/ Comparer les deux résultats obtenus, mémorisés en mémoire centrale du calculateur.

A first embodiment using this approach consists in programming the computer for:

• 1 / Determine the maximum value A _j taken by the signal in the time interval Δt _j and the distribution function F _j (X) of the amplitude, i.e. the probability P for a sample of the signal has an amplitude whose value X is less than a given reference value x, which can be fixed once and for all or calculated over each time interval.
  We will often use a fixed value when we want to make a quality assessment. For example, we know that a magnetic tape, in order not to give excessive distortion, must have an absolute output level not exceeding a threshold. With bands having common characteristics we can then adopt x = 18 dBu. Similarly, to verify the absence of risk of distortion at the output of a transmitter, we can risk of distortion at the output of a transmitter, we can adopt a threshold of 12 dBu.
  In the second case, which will be that of the discrimination between noise signal and useful signal, the value of x is calculated for each time interval from the maximum value A _j of the amplitude in the corresponding time interval Δ _j . We can for example adopt, outside the default periods,
  
  ${\text{x}}_{\text{j}}{\text{= A}}_{\text{j}}{\text{- 6 dB for A}}_{\text{j}}\text{> -20 dBu}$
  
  
  ${\text{x}}_{\text{j}}{\text{= A}}_{\text{j}}{\text{- 3 dB for A}}_{\text{j}}\text{≦ -20 dBu}$
  
  
  
  On the other hand, x _{j will} be kept at a constant value as long as the calculator signals a fault, whatever the evolution of the maximum value A of the amplitude.
  In this case, the first step is to determine the function F _j (X):
  $\text{Fj (x) = P [X <x]}$
  
  with Amin ≦ x ≦ Amax P [X <x] represents, for each time interval Δx, the fraction of the time during which X is below the threshold x.
  The CAN 18 can in particular be provided such that the values Amin = -30 dBu and Amax = +22 dBu mentioned above are obtained.
• 2 / Carry out, at the end of the order recording j + 1, the same calculations as above to obtain the maximum amplitude A _{j + 1} and the distribution function F _j (X).
• 3 / Compare the two results obtained, stored in the computer's central memory.

In a simple implementation mode, only the distribution functions F _{j + 1} and F _j are compared. If the absolute value of difference between them (or between their representations in the form of percentages of time) is less than a determined threshold, the computer signals a fault (indication of pre-alarm). Expressed as a percentage of time, the threshold can often be 0.5%. However, it may be necessary to increase it in an environment of impulse noise up to approximately 2% to avoid absence of detection.

If the difference between F _{j + 1} and F _j exceeds the threshold, the computer proceeds to the processing of the following records to calculate F _{j + 1} and F _{j + 2} .

If the fault persists for a given time (for example ten consecutive seconds) the computer 36 emits an absence of modulation alarm.

   Cette seconde solution, plus sûre, ne complique pas notablement le programme et sera en générale utilisée. Dans les deux cas la valeur du seuil x est conservée aussi longtemps qu'un défaut est signalé.In a more elaborate embodiment, the decision criterion is twofold and it is represented by the flowchart:

This second, more secure solution does not significantly complicate the program and will generally be used. In both cases the value of the threshold x is kept as long as a fault is reported.

.Figure 3 shows an example of how x is determined for a particular signal. The first measurement is supposed to be made between instants t0 and t1. The maximum amplitude is then A1. The selected reference level is $\text{x1 = A1 - 6dB}$

.

: on maintient alors x à la valeur x1.The second measurement, between t1 and t2, shows that $\text{| A1 - A2 | <-3 dBu}$

: we then maintain x at the value x1.

.The measurement continues thus showing a fault each time. At time t5 an alarm is triggered. From this instant, a new calculation of x occurs, even in the event of a slight modification of the maximum value of the amplitude. The measurement carried out over the time interval Δt between t5 and t6 shows an amplitude A6. We then adopt $\text{x6 = A6 - 6 dB}$

.

   Les valeurs successives ainsi obtenues sont traitées par le même type de comparaison que F_j(X) et F_j+1(X).A second embodiment making it possible to reduce the calculation time, consists in calculating the percentage of time during which X has been greater than the value x over the time interval Δt. The computer is programmed to calculate the function:

$\text{1-F (X) = P [X> x]}$

The successive values thus obtained are treated by the same type of comparison as F _{j (X)} and F _{j + 1 (X)} .

The value of this function can be used to develop other types of alarm, for example an overmodulation alarm in the case where the program signal has too large an amplitude, and leads to overmodulation.

The computer can also be provided to differentiate a noise from a sinusoidal signal, for example from a measurement or chain alignment signal, by analyzing the distribution of the amplitudes: it suffices in this case to compare the distribution of amplitude with stored distributions corresponding to the sinusoidal signals capable of being transported.

Still other solutions are possible: in particular, the maximum amplitude of the signal can be used during the successive Δt intervals as a criterion.

In all cases, the alarm signals can be transmitted on a link of any kind, in particular wired or wireless.

Claims (9)

1. A method of recognising the presence of a fluctuating audio frequency program signal in an input signal affected by substantially steady noise, which comprises: sampling the input signal; subjecting the samples to an amplitude analysis over successive time intervals of a predetermined duration sufficient to obtain a significant distribution of amplitudes in the course of each time interval; and comparing the distributions of the amplitudes obtained for successive time intervals, or simply the functions of distribution of the amplitude values representing said distributions, in order to identify the presence of the program by using the stochastic properties of the program signal, whose amplitude distribution and mean amplitude frequently vary in time.
2. A method according to Claim 1, wherein the analysis consists in determining the maximum amplitude during the intervals of time and in identifying a program signal by the fact that the variations in maximum amplitude in the course of consecutive intervals of time are higher than a given value.
3. A method according to Claim 1, wherein the comparison consists in determining the distribution function of the amplitudes by multi-channel analysis and in identifying a program signal by the fact that variations in the distribution function of the amplitudes in the course of successive intervals of time are higher than a given value.
4. A method according to Claim 3, wherein the distribution function of the amplitudes is derived from the percentage of time in the course of each interval during which the amplitude is higher (or lower) than a predetermined threshold.
5. A method according to Claim 4, wherein the threshold is fixed.
6. A method according to Claim 4, wherein the threshold is constituted for a predetermined time by the maximum value of the amplitude of the signal during the preceding time interval, reduced by a fixed value lying between 3 and 6dB.
7. A method according to Claim 6, in which, if a fault is detected, the threshold value is maintained until an alarm is triggered following the detection of a predetermined number of faults over successive intervals of time.
8. A method according to any of the preceding Claims, comprising sampling the signal at a frequency at least equal to the Shannon frequency with the object of analyzing the quality of the signal.
9. An apparatus for recognising the presence of an audio frequency program signal having a particular pass-band in an input signal affected by steady noise, comprising:

   - means for sampling the input signal,

   - multi-channel analyzing means (30) for storing the distribution of the amplitudes of the signal provided by the sampling means over a predetermined interval of time, and

   - computational means (36) enabling the comparison of the distribution stored in the multi-channel analyzer during two successive time intervals and generating an alarm signal if the comparison indicates that the input signal is stationary.

Images