FI83577C - Adaptive detection method for quantized signals - Google Patents

Description

1 83577

Adaptive detection method for quantized signals

The invention relates generally to an adaptive detection method for quantized signals whose nominal values at different points in time form a finite set of discrete states, wherein the discrete state of a signal at a given point in time is interpreted by comparing the signal sample value to a set of reference states.

10 Since the transmission path of each communication system is in practice always non-ideal, there is some signal distortion in all communication, which may make it difficult to detect the signal at the receiving end. The transmission path may cause very large random variations and variations and distortions in the representations (signal values) of the states of the discrete-15 path to be transmitted, which cause errors in expression or make expression impossible.

For example, in the transmission of quantized signals by radio or by carrier cable technology, digital codes are often transmitted using modulation depths corresponding to binary values, which are thus quantized analog quantities. In this case, the transmission path causes large and relatively fast level fluctuations of the modulated waves.

Therefore, an adaptive reception method (equalizer) must generally be used in data transmission, in which the sensitivity of the detection method automatically follows the level variations of the transmitted signal.

Various automatic level control methods are well known, which usually tend to normalize the carrier intensity in reception. However, the problem remains that even if the carrier variations could be compensated, the modulation depth in the reception may vary due to inaccurate or malfunctioning of the transmit and receive circuits, varying media characteristics, or when the signals reach different positions in the receiver. down the paths, with signals arriving with different delays interfering with each other.

A large number of different adaptive equalizers are also known, the transmission function of which tends to approximate the inverse function of the transmission channel transmission function, whereby they partially correct the distortions caused by the transmission channel signal before the signal is detected and restored to digital form.

These solutions aim to correct the signal before the signal states are detected, but do not aim to correct the detection event itself.

It is an object of the invention to provide a new adaptive detection method for quantized signals which reduces the effect of changes and distortions caused by the transmission path on the interpretation of the signal states.

This is achieved by an adaptive detection method of the type described in the introduction, which according to the invention is characterized in that the actual signal states of the signal to be detected are adaptively corrected towards the values of the reference numbers used in the detection. This is made possible by the fact that, although the transmission path causes potentially large random variations and distortions in the signal states, these changes are relatively continuous in time.

According to the invention, at least the reference number selected as the identification result is corrected on the basis of the distance between the sample value and the reference number after each sample value, or each reference number is corrected on the basis of group averages of sample values grouped by identification results.

According to one aspect of the invention, the aim is to determine, by the smallest possible number of samples of the received signals, whether these samples contain statistical groupings, the group means or other averages of 3,83577 could then be monitored and used as new reference numbers for expressing different quantized states. In this way, the reference numbers follow the signal levels quite closely, thus improving the accuracy of the expression.

5 In some situations, e.g. when reception begins or the connection is lost, the reference values of the different groups determined as mentioned above may, due to large control corrections, be misaligned and locked in, causing serious long-term interference in reception due to continuous misinterpretation of the signal.

According to another aspect of the invention, the principle of self-organizing feature maps is used to correct for comparison figures. The detection method using the self-organizing feature maps 15 tends to restore the magnitude and topology ratios of the reference values according to the corresponding ratios prevailing in the signal values, even if the order of magnitude of the reference numbers is temporarily incorrect.

According to a third aspect of the invention, both the mean-based detection method and the self-organizing feature maps are used alternately. Such a solution strikes the best compromise between adaptive accuracy and the ability to restore proper order.

With the detection method according to the invention, the accuracy of the detection can be increased and the accuracy requirements of previous correctors can be reduced or even such correctors can be replaced. The benefit obtained by the method increases as the number of different states of the signal increases, because even a very small disturbance can then cause errors in expression. The method according to the invention is computationally quite easy to implement and enables efficient correction even at high transmission rates, so that a suitable detector can be implemented as a signal processor or microcircuit application 35 more advantageously than conventional equalizers. The speed of current signal processors is not good enough to process high speed signals.

The implementation of the method can be performed by digital or analog calculation methods.

5 It has applications in, for example, digitally modulated data transmission by radio and carrier cable technology, measurement and control technology, medical technology, computer technology, and in general when transmitting quantized signals.

The invention will now be described in more detail by means of exemplary embodiments with reference to the accompanying drawings, in which Figure 1 shows a one-dimensional feature map, Figure 2 illustrates one embodiment of the feature map of Figure 1, Figure 3 shows the operation of a feature map method by way of example, Figure 4 shows a two-dimensional feature map, an apparatus in which the method according to the invention is applied.

In its simplest form, the method according to the invention is applicable to a one-dimensional signal (e.g. a multilevel signal) whose values are quantized for a range [a, a + nq], where a is the lowest value of the interval 25, n is an integer, and q is the quantization interval. Thus, an attempt has been made to give the signals one of the values a + hq, where h is the order state index of the signal state, 0,1, ..., n, but the transmission path converts it to the random value a + hq + fcj ^, where Sh is a random error (possibly oral). 30 greater than the quantization interval q). However, it is assumed that when different factors causing variations occur in the same circuits or in the same transmission path, the successive values of 6h are so correlated that the absolute value of their difference \ Sh-t ^ .x | is greater than q / 2 with only a very small probability of to-35.

In this case, the received signal value passing through the transmission path is represented by the numerical value x = a + hq + £ h, where h and £ h are the order state of the signal state and the random variation of the state, respectively.

5 Identification of signal conditions (expression):

The signal states of the signal are identified by comparing them with certain reference values, incl. Let us first assume that the comparison figures mt have been calculated. Using some suitable distance measure d = d (x, mt), for example 10 d = | x-m1 |, the nearest reference value mc is determined, which is just the identification result: d (x, mc) = min {d (x, mt)}. (1) i 15

Update of benchmarks

Although the initial reference values mt achieve a satisfactory detection result of the link, the accuracy of the detection can be improved when the reference values are continuously corrected according to the changes caused by the transmission path to the received signal and its signal states.

The signal states of the received signal may vary within wide limits due to various interfering factors, but in the short term the relative or topological relationships of the signals are maintained.

The principle of the present method is to automatically monitor the presence of statistical clustering in these samples with the smallest possible number of received signals and to produce adaptive reference numbers or vectors that best match these ... 30 groups, which can then be used to identify quantized signal states.

This can be accomplished using conventional grouping methods or vector algorithms. Such is the grouping of k averages used in 35 e.g. digital signal technology and pattern recognition, which is shown in the reference: [1] J. Makhoul, S. Roucos, H. Gish, "Vector Quantization in Speech Coding", Proc, IEEE, Vol. 73, p. 1551-5 1588, 1985.

The following describes an update of the reference figures of the method according to the invention, in which the grouping of the mean of k is used.

Updating comparison figures based on group averages Fri-10:

In the k-averaging method described in reference [1], a series of successive signal samples are collected such that each possible signal state occurs in them several times. The samples are divided into categories according to the identification result described in the previous section-15 sa and the average of each category is formed. The new reference numbers mA can now be these averages calculated from the samples, or, for example, some weighted linear combination of the average reference values calculated from the old reference numbers and the samples. When the method is repeated continuously when new signals are received, the reference numbers follow the variations of the signal values. The figures are therefore updated here at regular intervals.

To update the reference numbers based on the distance between the signal sample and the reference number:

The update of the reference figures can also be performed after each signal sample, in which case the experimental identification of the sample according to formula (1) must first be performed in order to find the nearest reference value mc. This and .30 only this is updated: mc (t + 1) = mc (t) + a (x (t) -mc (t)) (2) m ^ t + 1) = m ^ t) when i φ c (3) 35 Thus, here mA = mt (t) is the value of the reference number just li 7 83577 before obtaining the signal sample at time (t) (t = 0,1,2 ... is the so-called discrete time variable). The parameter a (0 <α <1) is a factor that determines the relative magnitude of the corrections. The corrected reference number is obtained by adding to the original 5 reference numbers the difference between the signal sample x and the original reference number multiplied by a: 11a.

Using these mA update methods of the reference numbers, the reference numbers of the different signal states are normally very closely monitored to follow the group averages of the representations of the signal states.

However, these update figures are still partly deficient, as they run the risk that the algorithm-defined reference values mt may be in the wrong order of magnitude-15 due to large control corrections and locked due to the control method used, causing severe, long-term reception disturbances due to discrete conditions. constantly misinterpreted. The adaptive method must therefore be self-correcting in the sense that it seeks to maintain 20 and restore the correct magnitude and topological ratios of the reference values according to the corresponding ratios in the signal values, even if the ratios of the reference numbers are temporarily erroneous.

In the preferred embodiment of the method according to the invention, the so-called self-organizing feature maps with such a feature.

To update the benchmarks using the self-organizing map description to restore the correct order:

The principle of self-organizing map mapping in the one-30 dimensional case is illustrated in Figure 1, which shows a set of processing units, hereinafter referred to as nodes. Each node has a reference number m £, i - 0.1, ..., n stored, which is used to identify and indicate the signal state of the received signal x. In Figure 1, the lines between the nodes 8,83577 depict topological neighborhood relationships, i.e., which nodes are topologically closest to each other. By the topological order of the nodes in this case it is meant that the reference numbers mt closest to the intended values are at the adjacent nodes of the map description of Fig. 1, i.e. belonging to each other's topological neighborhood. However, in an erroneous state, this order may be temporarily different.

The topological map description of Fig. 1 can also be reset according to Fig. 2, in which the nodes, i.e. the processing units, are connected in parallel to the signal line and the same signal x is connected to each input. Each processing unit has an output to which it produces an output signal when the signal state of the signal x is closest to the reference number mt of that node. For example, if the signal x is a multilevel signal, the processing units could be comparators, for example.

However, the present disclosure is intended only to facilitate an understanding of the invention. Self-organizing topological feature maps and their use are described in more detail in the following publications, which are incorporated herein by reference: [2] T. Kohonen, "Clustering, Taxonomy, and Topology-25 cal Maps of Patterns," Sixth International Conference on

Pattern Recognition, Munich, Germany, October 19-22, 1982, ss. 114-128.

[3] T. Kohonen, "Self-Organization and Associative Memory," Springer-Verlag, Series in Information Sciences, 30 Vol. 8, Berlin-Heidelberg-New York-Tokyo, 1984, p.

The disadvantage of the previously presented method of updating the reference figures is that if for some reason, for example due to a malfunction or a calculation error, the reference figures mt obtain values with magnitudes different from the original 35 (mt> mj although i <j), this wrong order li 9 83577 is retained. Instead, the method described below can automatically restore the correct order. Proof of this has been done in reference [3] and its cross-references, and is also illustrated here by means of Example 5.

In the method, the signal states of the received signal are identified according to Equation (1).

The update of the reference numbers is performed according to the following correction rule (adjustment rule, grouping rule): both the reference number mc selected as the identification result and the reference numbers of the neighboring node points of the corresponding node c are corrected closer to the value of x. That is, 15 mt (t + 1) = m1 (t) + Aix-n ^ it)), ie Nc (4) mi (t + 1) = mt (t), i £ Nc (5) where mA = mi (t) is the value of the reference number just before receiving the signal sample at time t (t = 0,1,2, ... is the so-called dis-20 cretaceous time variable) and m ^ t + 1) is the new reference number to be used at the next sampling time. The parameter a is a factor that determines the relative magnitude of the corrections. N. is an index set consisting of node c and its topological neighbors. Figure 1 shows the nodes of line-25a that are topological neighbors of each other.

Above, the parameter a = a (t) always satisfies the condition 0 <α <1, but it can be varied as needed, and Nc = Ne (t) is an index set comprising all nodes in the neighborhood of sol-30 m in the distance of a given radius. When the radius is equal to or larger than one node spacing, the correction rule described above can effectively return to the reference number mt the same order of magnitude as the received signal states have, but leaves a certain bias in the values of mt, which will be considered later.

10 83577 When the radius is zero (Nc = {c}), the correction rule again tries to give the reference numbers the correct group averages as possible, but cannot restore the correct order. This situation corresponds to the 5 correction rules presented in Equations 2 and 3.

A compromise must be defined between these two characteristics. In practice, different choices of a = a (t) and Ne Nc (t) have to be applied periodically as needed, at least when the data transfer is interrupted and continues for a further 10.

Often, the signal states of a transmission can be considered to vary randomly with the same probability of occurrence. This is the case, for example, if the analog signals are quantified in the so-called by delta modulation, and the accumulated bits are collected into size-15; the numerical values of these codes, with which the transmission is then modulated, are usually the so-called a pseudo-random-numbers. Then there is an unambiguous numerical correspondence between the comparison figures produced by the correction rule and the group averages (see reference [3]). Then, if the 20 received signal values x are pretreated in the correction rule with the expression b1 + d1x specific to each reference number, where and dA are standard parameters, the bias left by the above correction rule can be largely compensated and more accurate 25 values can be obtained for the reference numbers. This same method can be used at other times when the probability density p (x) of the signal x is an arbitrary but known function, in which case the bias-compensating parameters bd and dr can still be calculated for the preprocessing expression. The final weighted correction rule to be used. typically of the form mA (t + 1) = mt (t) + aibj + dA x-m1 (t)), ie Nc (6) 35 m1 (t + 1) = mt (t), i £ Nc (7) 11 83577

Example 1 This example illustrates the ability of the improved preprocessed correction rule according to Equations (6) and (7) to define topologically ordered reference-5 numbers mx that quickly follow the transmission set {x = x (t), t = 0,1,2, ... } general variations. We assume that there are 4 signal states, and they occur randomly with the same probability. In this example, the parameters b0 = bj = b2 = b3 = 0, d0 = 0.7, dx = 1, d2 = 1, d3 = 1.14 have been used. Fig. 3 to Fig. 3 The continuous curves drawn show the possible signal states, the magnitudes of which here vary according to an arbitrarily chosen law. The actual values of the signal x are extracted at different times from these four curves sa-randomly. At separate points The plotted ku-15 plots show how the corresponding reference numbers follow these variations. The example has been a = 0.1 and N = ({max (0, c-1), c, min (3, c + 1)}, i.e. only the nearest neighbors have been included in the neighborhood of node c according to Figure 1. The example also shows , how the order of the mouths of the reference numbers m3-20, which was initially incorrect (1-0-3-2), becomes correct after about 30 samples (0-1-2-3) .The random deviations of the reference numbers are due to the presence of different signal states. in signal x completely randomly, in which case the reference figures are also corrected in random order.

The one-dimensional map description presented above and illustrated in Figure 1 can be easily generalized to multi-dimensional. Such a case occurs when, for example, the transmission has two or more separate signal variables due to some combined modulation mode. In the following, we assume that there are simultaneous signal states (hj and h2) in the transmission and the corresponding transmissions are xx * ax + h1q1 and x2 * a2 + h2 q2, where hx and h2 are integers and the corresponding quantization intervals of q: and q2. In this case 35, the transmission is represented by the vector X = (x:, x2) = X (t) i2 83577 and the corresponding reference numbers A, mi 2 as vectors = (m ^, n> i 2) = Mt (t). The identification rule can be selected to compare distances according to one of the vector norms. Identification: 5 d (X, Mc) = min {d (X, M1)} i where d (·) can be, for example, some Minkowski distance-10 measure with a special case of the Euclidean distance;

Updating reference vectors (without preprocessing expression): M * (t + 1) = Mt (t) + p (X (t) - M, (t)), ie N 'e 15 M, (t + 1) = M , (t), i * N 'e where β = β (ΐ) is another scalar (0 <β <1) and N'c = N'c (t) is a neighborhood set in a two-dimensional node network, e.g. according to Figure 4 . In Figure 4, the neighborhood set N’c of node c 20 includes the surrounding nodes closest to node c, as illustrated by a solid line. In this case, the topological order of the nodes can be e.g. such that the reference numbers m ± χ of the reference vectors in the nodes increase from top to bottom and the reference numbers mt2 increase from left to right.

The rules described above illustrate the general principle that the method must be kept the same even if the topology of the node set is chosen differently, and the definition of the neighbors N. and N'c used in the configuration would differ from the previous examples as long as they aim at the topological arrangement described above. Also, the parameters a - a (t) and β = β (ΐ) do not have to be the same for all nodes, but can be generalized to some 35 functions of the distances calculated from point c.

Il i3 83577

In general, when the number of signal variables x is n, each reference vector has n reference numbers m1, respectively.

It should be emphasized that the adaptive reference numbers of the detection-5 method described herein are determined solely by the metric topological relationships of the different signal states of the detected transmission, without the need to indicate the identity or significance of any signal state in the transmitted signal. This is therefore nonsupervised learning, which is characteristic of all clustering methods and most adaptive signal processing systems.

Various Embodiments This method thus includes two essential sub-15 methods: 1. A method by which the comparison numbers of different signal states are made to follow the group averages of the representations of the states. 2. A method that returns the correct magnitudes to the reference figures if, for some reason (malfunction, incorrect correction, etc.), they mix.

20 The operations required by methods 1 and 2 can be carried out numerically, for example with so-called signal processor architectures, or analog circuits. The choice depends on the accuracy required, the computational speed, and 25 cost factors; the particular trade-off depends on the prevailing industrial development of the various circuit technologies. For example, T. Korhonen, K. Torkkola, M. Shozahai, J. Kangas, O. Ventä, "Microprocessor Implementation of a Large Vocabulary Speech Recognizer and 30 Phonetic Typewriter for Finnish and Japanese", Proc. Of the European Conference on Speech Technology, The signal process-sorcer solution described in CEP Consultants Ltd, Edinburgh, 1987, pp.377-380, which has been used for speech recognition, is technically possible for this purpose, but unnecessarily complicated and expensive. The correction rule 83 577 can also be implemented by the classical analog-computer technology described, for example, in G.A. Korn, T.M. Korn, "Electronic Analog and Hybrid Computers", McGraw-Hill, New York, 1964, or equivalent electronic, optical 5, or optoelectronic constructs.

If the method is applied to data transmission by radio or cable using a modulated carrier, the detection at the receiver can be very simple; it is not subject to high accuracy requirements of 10 because adaptive adjustment automatically seeks to correct nonlinearities and instabilities. In fact, it is sufficient for the detector that the magnitude correlations of the detected signal states are maintained in the short term, as the grouping analysis used automatically finds and adapts to the accumulations of transmissions corresponding to these states.

Figure 5 shows a typical block diagram of an apparatus using the method. Part 1 is a means for modulating and transmitting signals, Part 2 is a means for receiving and detecting (demodulating) signals 20, Part 3 is a means whose reference numbers follow the signal states received with the highest possible accuracy (e.g. by the k-average method), and Part 4 is a means whose reference figures are determined, for example, according to correction rules based on self-organizing map descriptions, whereby the reference figures are able to automatically return to the correct order of magnitude and topology. Part 5 is a means for coupling a signal to either Part 3 or Part 4, and Part 6 (controller) is a means for controlling the operation of Part 5. 30 Part 6 may receive information from the hand control, the transmission receiving device (part 3), which indicates the start and interruptions of the transmission, or the interpretation means (parts 3 and 4), which may indicate when the interpretation is constantly incorrect. In addition, the part 5 can be connected to the parts 3 and 35 4 at regular or otherwise predetermined intervals, i.e. by a timer.

The mt or initial values of the reference figures in parts 3 and 4 can be set to pre-calculated values at the start of the operation. Also, the controller, part 6 can set the ver-5 values mA or MA to the appropriate default values.

The advantage of the described apparatus is that at the beginning of reception, after interruptions or in bad error situations, the reference numbers can be restored to the correct order by connecting the signal to part 4. When the reception is established 10, the signal can be connected to part 3, which calculates the reference values more accurately. This optimizes both accuracy and recovery from interference.

The adaptive and self-correcting signal detection method according to the invention is suitable for use in the transmission of any data, in which the nominal values of the signals to be transmitted at different times form a finite set of discrete states.

The signals can directly represent digital information, e.g. in a transmission within the device. 20 The method can be applied when digital signals or states are converted to analog quantities and transmitted as such, and then again analyzed, expressed and encoded by this method into digital form. One area of application is cellular radio (cellular), in which speech or other analog signals are first digitized by delta modulation, converted to an analog format (e.g., QAM), transferred in this format, converted back to digital format, and then reconstructed into speech.

30 However, the method to be described is not limited to any particular modulation method or standard in data transmission, but is generally intended to distinguish quantized states.

The method can also be applied to optical signals or devices that read digital signals from magnetic or optical memories.

The accompanying figures and the related description are intended to illustrate the present invention only. The details of the method according to the invention may vary within the scope of the appended claims.

Claims (10)

An adaptive detection method for quantized signals whose nominal values at different times 5 form a finite set of discrete states, wherein the discrete state of the signal at a given time is interpreted by comparing the sample value of the signal to a set of reference numbers based on signal state identification. correcting the values of the reference numbers used in the expression towards the actual signal states of the signal to be detected. A method according to claim 1, characterized in that at least the reference number selected as the identification result is corrected either on the basis of the distance between the sample value and the reference number 15 after each sample value or each reference number is corrected on the basis of group averages of sample values grouped by identification results. A method according to claim 1 or 2, wherein the signal has two or more separate co-occurring signal variables, each with a finite number of discrete states, characterized in that the reference numbers are reference vectors. Method according to Claim 2 or 3, characterized in that successive signal values of the signal are collected in such a number that each possible detection result occurs several times, the samples are divided into groups according to the detection results and a group average is formed for each group, the new reference numbers being taken at regular intervals. calculated group means or some linear combination of these calculated group means and previous reference numbers. Method according to Claim 2 or 3, characterized in that after each sample value 35, 18 83577 sample values are corrected for the reference value selected as the identification result. Method according to claim 2 or 3, characterized in that the reference numbers are updated by monitoring the magnitude and topological ratios of the different signal states of the signal and determining reference numbers which follow the same metric-topological relationships. Method according to Claim 6, characterized in that the reference figures are determined by means of self-organizing feature maps. A method according to claim 1, 2, 3, 5, 6 or 7, characterized in that after each sample value, both the reference number selected as the identification result and a predetermined number of other reference numbers forming a predetermined topological neighborhood set of the selected reference number are corrected. Method according to Claim 8, characterized in that the reference figures are corrected in the direction of the sample value by an amount which is proportional to the distance between the sample value and the respective reference number multiplied by a parameter between 0 and 1. Method according to claim 9, characterized in that before correcting each reference number, the sample value is treated with the expression bi + d1x specific to each reference number, where x is the sample value, 25 and d £ are constant parameters and i is the order index of the reference number. 19 83577