1. Field of the Invention
The invention relates to a method for interference cancellation in a cellular TDMA (Time Division Multiple Access) communication system according to pre-characterizing part of claim 1. Further, the invention relates to a communication device for interference cancellation in a cellular TDMA communication system according to pre-characterizing part of claim 14.
2. Description of the Related Art
Currently, single antenna co-channel interference cancellation (SAIC) is being a hot topic, especially for the Global System for Mobile communications/Enhanced Data (GSM/EDGE) downlink. In field trails, tremendous capacity gains have recently been demonstrated [1] 3G Americas, “SAIC and synchronized networks for increased GSM capacity,” www.3gamericas.org, September 2003, particularly for synchronized networks in urban areas. As a consequence, the upcoming GSM/EDGE release will be tightened with respect to interference cancellation.
One of the most challenging tasks is the design of the interference canceller, especially if due to cost, volume, power consumption, and design aspects only one receive antenna is available. Most interference cancellers fail if the number of receive antennas do not exceed the number of co-channels [2]J. H. Winters, “Optimum combining in digital mobile radio with co-channel interference,” IEEE J. Sel. Areas Commun. vol. 2, no. 4, pp. 528-539, 1984, [3] A. M. Kuzminskiy and D. Hatzinakos, “Semiblind estimation of spatio-temporal filter coefficients based on a training-like approach,” IEEE signal Processing Letters, vol. 5, no. 9, pp. 231-233, September 1998, [4]A. M. Kuzminskiy, C. Luschi, and P. Strauch, “Comparison of linear and MLSE-spatio-temporal interference rejection combining with an antenna array in a GSM System,” in Proc. IEEE Veh. Techn. Conf. '00, Tokyo, Japan, pp. 172-176, May 2000.
A receiver structure with linear filtering and decoupled non-linear detection has been studied in [5] D. Giancola, U. Girola, S. Parolari, A. Picciriello, U. Spagnolini, and D. Vincenzoni, “Space-time processing for time varying co-channel interference rejection and channel estimation in GSM/DCS systems,” in Proc. IEEE Veh. Techn. Conf. '99, May 1999, [6] S. Arnyavistakul, J. H. Winters, and N. R. Sollenberger, “Joint equalization and interference suppression for high data rate wireless systems,” IEEE J. Sel. Areas Commun. vol. 18, no. 7, pp. 1214-1220, July 2000, [7] R. Friedman and Y. Bar-Ness, “Combined channel-modified adaptive array MMSE canceller and Viterbi equalizer,” in Proc. IEEE Veh. Techn. Conf. '01, pp. 209-213, 2001, [8] A. L. F. de Almeida, C. M. Panazio, F. R. P. Cavalcanti, and C. E. R. Fernandes, “Space-time processing with a decoupled delayed decision-feedback sequence estimator,” in Proc. IEEE Veh. Techn. Conf. '02, Birmingham, Ala., USA, May 2002, [9] K. Kim and G. L. Stüber, “Interference cancelling receiver for range extended reception in TDMA cellular systems,” in Proc. IEEE Veh. Techn. Conf. '03, October 2003 using multiple receive antennas. A special case of linear filtering in conjunction with a memory-less detector has been investigated in [10] B. R. Petersen and D. D. Falconer, “Suppression of adjacent-channel, co-channel, and intersymbol interference by equalizers and linear combiners,” IEEE Trans. Commun. vol. 42, no. 12, pp. 3109-3118, December 1994, [11] H. Trigui and D. T. M. Slock, “Cochannel interference cancellation within the current GSM standard,” in Proc. IEEE ICUPC '98, pp. 511-515, 1998, [3], and [4] for the case of multiple receive antennas. Further, there is known Kammeyer's pre-filter [12] K. D. Kammeyer, “Time truncation of channel impulse responses by linear filtering: A method to reduce the complexity of Viterbi equalization,” AEÜ, vol. 48, no. 5, pp. 237-243, 1994, which is designed in order to truncated a single InterSymbol Interference (ISI) channel, but which does not take Co-Channel Interference (CCI) into account. A similar pre-filter has also been proposed by Al-Dhahir and Cioffi [13] N. Al-Dhahir and J. M. Cioffi, “MMSE decision feedback equalizers: Finite-length results,” IEEE Trans. Inform. Theory, vol. 41, no. 4, pp. 961-975, July 1995, for the case of Gaussian noise plus interference. In the present case, focus shall be on a few numbers of dominant interferers, which is the typical case in cellular networks, i.e., on non-Gaussian interference. [14] R. Meyer, R. Schober, and W. Gerstacker, EP 1221780 A1, published July 2002 discloses a pre-filter with real-valued processing, being designed according to the principles established in [11].
An object of the invention is to improve a method and a communication device for interference cancellation in a cellular TDMA communication system.
This object is solved by a method for interference cancellation in a cellular TDMA (Time Division Multiple Access) communication system according to the features of claim 1. Further, the object is solved by a communication device for interference cancellation in a cellular TDMA communication system according to the features of claim 14.
According to the preferred embodiment there is provided a method for interference cancellation of received data being received via a communication channel in a cellular communication system and having corresponding channel coefficients, especially channel impulse response coefficients, said method comprising the steps of linear filtering of said received data, and thereafter of executing a detection, especially non-linear detection to get detected data by especially non-linear signal processing. Further, there are determined filter coefficients for linear pre-filtering of said received data to suppress non-Gaussian interference when filtering said received data.
Further, according to the preferred embodiment there is provided a communication device for communicating with another device in a cellular communication network, said communication device comprising a receiver unit for receiving of data sent from said other device via a communication channel and receiving data sent from at least one further device via a disturbing, especially via an interfering channel, said channels having corresponding channel coefficients, and at least one processing unit for interference cancellation of said received data and for linear filtering of said received data and, thereafter, for detection, especially non-linear detection to get detected data by especially non-linear signal processing. Further, at least one of said at least one processing units being designed for determining of filter coefficients for linear pre-filtering of said received data to suppress non-Gaussian interference when filtering said received data, wherein said processing unit is designed especially for performing a method according to any of preceding claims.
Advantageous embodiments are subject matter of dependent claims.
Advantageously, there is provided a method, comprising the step of processing of pre-filtered data by non-linear detection to get detected data, especially to get detected symbols.
Advantageously, there is provided a method comprising the steps of determining of optimized channel coefficients and using these optimized channel coefficients as coefficients in a decision-feedback equalization to get detected data. Such a method is preferred when performing said decision-feedback equalization after the step of pre-filtering of data. Further, such a method is preferred when said optimized channel coefficients being determined from said filter coefficients and said channel coefficients, especially being determined as a convolution between said filter coefficients and said channel coefficients.
Advantageously, there is provided a method claim, comprising the steps of performing of an auto-correlation of said received data and a cross-correlation between a sequence of said received data and a sequence of known symbols for said determining of said filter coefficients. Especially, such method is preferred when said sequence of known symbols being part of sent data corresponding to a part of said received data or being determined from said received data.
According to a first sub case there is preferred a method comprising the step of calculation of an auto-correlation matrix using knowledge of at least one desired channel transmitting desired data and using knowledge of at least one disturbing, especially interfering channel, transmitting disturbing data, said knowledge being determined by channel estimation.
According to a second sub case there is preferred a method comprising the step of calculation of an auto-correlation matrix using only knowledge of at least one desired channel transmitting desired data, said knowledge being determined by channel estimation.
According to a third sub-case there is preferred a method comprising the step of calculation of an auto-correlation matrix using the received samples, rather than the knowledge of a desired channel transmitting desired data or the knowledge of a disturbing, especially interfering channel, transmitting disturbing data.
Advantageously, there is provided a method comprising the step of determining the optimized channel coefficients for a non-linear equalizer by convolution of the filter coefficients with the originally estimated channel coefficients of the desired data.
Advantageously, there is provided a method comprising the step of estimating optimized channel coefficients for non-linear equalization using the pre-filtered signal and using channel estimation.
Advantageously, there is provided a method comprising the step of processing of the pre-filtered data using a memory-less detector, using a backwardly directed filter of a MMSE-DEE (MMSE-DFE: Minimum Mean Square Error-Decision-Feedback Equalizer) in connection with a memory-less detector, or using a non-linear equalizer.
Preferred is a communication device and corresponding method, wherein said at least one processing unit being designed as a linear pre-filter to suppress non-Gaussian interference followed by a non-linear detector to cancel intersymbol interference.
Especially, a receiver structure is preferred being suitable for single-antenna co-channel interference cancellation in cellular TDMA networks. The receiver consists of a novel linear pre-filter followed by a non-linear detector. The task of the pre-filter is to suppress non-Gaussian interference, especially to suppress co-channel interference, whereas the task of the non-linear detector is to cancel intersymbol interference. The pre-filter is designed according to the principle of minimum mean square error decision-feedback equalization, without solving an eigenvalue problem. The non-linear detector may be an arbitrary non-linear equalizer, in the simplest case just a memory-less detector. The receiver structure is compatible with conventional TDMA receivers ignoring co-channel interference. Receiver and corresponding method are applicable in cellular communication systems, e.g. like GSM.
Especially, focus is on a receiver structure, where Co-Channel Interference (CCI) reduction is separated from InterSymbol Interference (ISI) suppression. CCI reduction is done by means of a time-domain linear filter, followed by a non-linear detector in order to perform ISI suppression of the desired users channel. For example, non-linear detector is constructed as a trellis-based, tree-based, graph-based equalizer or as a feedback filter of a decision-feedback equalizer. Besides interference suppression, the linear filter is able to truncate the overall channel impulse response seen by the subsequent non-linear detector. Advantageously, this receiver structure is compatible with conventional TDMA (Time Division Multiple Access) receivers ignoring CCI.
The method and the receiver structure are advantageously in some respects. A linear co-channel interference canceller is typically less sensitive with respect to model mismatch compared to a multi-user-based co-channel interference canceller. Model mismatch may be due to asynchronous interference, non-perfect channel estimation, EDGE interferer in a GSM environment or vice versa, etc.
Besides interference suppression, the linear filter is able to truncate the overall channel impulse response seen by the decoupled non-linear detector. The memory length of the overall channel impulse response is a design parameter.
Since the non-linear detector is matched to the possibly truncated desired users channel only, it is much simpler than a multi-user-based co-channel interference canceller.
In contrast to a similar receiver structure in [5], [6], [7], [8], [9] using multiple receive antennas, focus of preferred embodiment is on single-antenna interference cancellation, which is more demanding than space-time processing. Still, multiple antennas are optional according to design of preferred embodiment. The preferred pre-filter may be interpreted as a generalization of Kammeyer's pre-filter [12], which is designed in order to truncated a single ISI channel, but which does not take CCI into account. According to preferred embodiment focus is on a few number of dominant interferer, which is the typical case in cellular networks, i.e., on non-Gaussian interference, especially on co-channel interference.
An embodiment will be described hereinafter in further detail with reference to drawings.
The preferred embodiment relates to a method and communication device for Single Antenna Interference Cancellation (SAIC) for cellular TDMA networks by means of decoupled linear filtering and especially non-linear detection.
As can be seen from
The communication network provides further base stations BS2, BS3 each having a communication cell c2, c3, said communication cells c1-c3 being arranged in an overlapping manner. A mobile station MS1 moving from a first cell c1 to a second cell c2 changes from a first base station BS1 to a second base station BS2 after reaching handover regions ho of the overlapping cells. However, before reaching a handover region ho the mobile station MS receives interfering signals i11, i12, i21 from the other base stations BS2, BS3. Especially, the interfering signals are signals sent from the other base stations BS2, BS3 on the same channel to further mobile stations MS*, MS′ in their communication cells c2, c3. Therefore, the mobile station receives co-channel interference disturbing the data received from their own first base station BS1. In addition, the mobile station MS1 receives noise n, especially white Gaussian noise. Usually, the base stations BS1-BS3 of a cellular communication network are connected via a base station controller BSC to other components of the communication network. To avoid co-channel interference base station controller BSC assigns different channels to neighboring cells c1-c3. According to a preferred embodiment same channels shall be used in neighboring cells. In the following it is proposed to use sequence estimation, in the case of GSM a delayed-decision feedback sequence estimation to improve reconstruction of data received by the mobile station MS1 to reduce co-channel interference.
Data y(k) received by the mobile station MS can be written as:
The first terms describe the signals s1, s2, . . . received via communication paths i.e. communication channels s1, s2 from the first base station BS1. These signal components s1, s2 shall be used by sequence estimation to reconstruct data an originally sent by the first base station BS1. For performing this sequence estimation a channel impulse response h1 is used. The channel impulse response h1 can be determined from received data or from received special symbols, e.g. a training sequence.
Second sums and terms correspond to the signal components received via interfering communication paths i11, i12, i21 from second and third base stations BS2, BS3. These interfering signal components i11, i12, i21 and in addition further noise components n(k) has to be distinguished by the sequence estimation procedure. Especially, the signal components s1, s2 and the interfering signals components i11, i12, i21 has been sent as interfering data bj by corresponding base stations BS1, BS2, BS3 using a co-channel especially using same frequency f(sl)=f(ij,l)−f(n). For performing this sequence estimation a channel impulse response gj,l is used. The channel impulse response gj,l can be determined from received data or from received special symbols, e.g. a training sequence.
As can be seen from
As can be seen from
According to the preferred embodiment there are determined, especially computed filter coefficients ŵ. As can be seen from
Further, the estimated data a[k−k0] are used in decision-feedback equalizer DFE for further processing together with optimized channel impulse response {tilde over (h)}=({tilde over (h)}1,w. . . , {tilde over (h)}L,b). The estimated data are delayed up to L times by delaying units z−1. The estimated data a being delayed one time are processed with first component {tilde over (h)}1,w, estimated data a being delayed two times are processed with second component of optimized channel impulse response {tilde over (h)}, and so on. All in such a way processed data are added. Added data are subtracted from next pre-filtered data value z[k] before next estimation step.
Summarized, as can be seen from
As can be seen from
In the following there is offered an introduction of the equivalent discrete-time channel model under consideration, a short description of the conventional design rule for implementing a linear filter according to [5], [7], [8]. Then, a design rule according to preferred embodiment for implementing the linear filter in the presence of non-Gaussian interference is presented, which is much simpler and which leads to further insights. In particular, three different options for training the filter are presented. Finally, numerical results are presented and conclusions are drawn.
The equivalent discrete-time channel model considered is given as
where y(i)[k]εC is a k-th baud-rate output sample of an i-th polyphase channel, L is an effective memory length of the discrete-time ISI channel model, hl(i)[k]εC are the channel coefficients of the i-th polyphase channel of a desired user MS (E{∥h(i)[k]2∥}=1), gj,l(i)[k]εC are the channel coefficients of the i-th polyphase channel of the j-th interferer, J is the number of interferer, a[k] and bj[k] is a k-th independent identically distributed (i.i.d.) data symbol of the desired user and the j-th interferer, respectively, both randomly drawn over an M-ary alphabet (E{a[k]}=E{bj[k]}=0, E{a[k]2}=E{bj[k]2}=1), n(i)[k]εC is the k-th sample of a Gaussian noise process (E{n(i)[k]}=0, E{n(i)[k]2}=N0/Es), k is a time index, and K is a number of M-ary data symbols per burst. All random processes are assumed to be mutually independent. If the signal bandwidth does not exceed twice the Nyquist bandwidth, two samples per symbol are sufficient. This scenario is assumed in the following. The corresponding polyphase channels are labeled i=1 and i=2, respectively. The channel coefficients h(i)[k]:=[h0(i)[k], . . . , hL(i)[k]]T and gj(i)[k]:=[gj,0(i)[k], . . . , gj,1(i)[k]]T comprise pulse shaping, the respective physical channel, analog receive filtering, the sampling phase, and the sampling frequency. Without loss of generality, the effective memory length, L, is assumed to be the same for all co-channels. Some coefficients may be zero. In case of square-root Nyquist receive filtering and baud-rate sampling, the Gaussian noise processes of all polyphase channels are white. This case is assumed in the following. The equivalent discrete-time channel model is suitable both for synchronous as well as asynchronous TDMA networks, because time-varying channel coefficients are considered.
Conventional design of the linear filter is presented to highlight differences and to introduce design according to preferred embodiment. For simplicity, a synchronous TDMA network and low speeds of stations MS are assumed in the following for the purpose of filter design. In conjunction with frequency hopping, the second assumption corresponds to a so-called block fading channel model. Correspondingly, the time index k can be dropped for convenience. In case of asynchronous bursts, the channel coefficients are only piece-wise constant, i.e., the time index cannot be dropped.
As a design criterion for the linear filter, in exemplarity a maximization of the Signal to Interference plus Noise Ratio (SINR) at the output of the linear filter is chosen:
where the maximization is jointly with respect to the filter coefficients, {tilde over (w)}, and the overall channel coefficients including filtering of the desired user, {tilde over (h)}w.
The filter coefficients can be written in vector form as {tilde over (w)}=[{tilde over (w)}(1)T,{tilde over (w)}(2)T]T, where the filter coefficients of each polyphase channel are given by {tilde over (w)}(i)=[{tilde over (w)}0(i), {tilde over (w)}1(i), . . . {tilde over (w)}L
Note that further oversampling does not lead to any performance improvement in contrast to increasing the number of receive antennas. Specifically, the degree of diversity is not further enhanced.
The output samples of the analog receive filter can be written in vector form as y=[y(1)T,y(2)T]T, where y(i)=[k(i)[k], y(i)[k−1], . . . , y(i)[k−Lw]]T and 0≦k≦K−1.
Hypotheses of the overall channel coefficients including filtering of the desired user are denoted as {tilde over (h)}w=[{tilde over (h)}0,w, {tilde over (h)}1,w, . . . {tilde over (h)}L′,w]T, where L′ is the effective memory length. Note that {tilde over (h)}w is symbol spaced. If the filter length Lw is sufficiently large, the filter truncates or is capable to truncate the overall impulse response, i.e., L′≦L. As discussed when describing preferred embodiment, for a well-designed filter in conjunction with a suitable non-linear detector such as a trellis-based, tree-based, or graph-based equalizer, L′ is a design parameter. It can be interpreted as the desired memory length of the overall impulse response.
The corresponding data vector is given by a=[a[k−k0], a[k−k0−1], . . . , a[k−k0−L′]]T, where k0 is the decision delay of the linear filter, which is often chosen as k0=(Lw+1)/2
Recall that the coefficients of the linear filter, {tilde over (w)}, and the channel coefficients, {tilde over (h)}w, are jointly optimized. The optimum filter and channel coefficients with respect to the maximum SINR are denoted as ŵ and ĥw, respectively. According to the preferred embodiment these optimum filter and channel coefficients ŵ and ĥw, respectively, has to be determined.
A maximization of (2) corresponds to a minimization of the cost function
C:=E{{tilde over (w)}Hy−{tilde over (h)}wHa|2},
that is
The minimum of C is obtained when the following conditions apply:
By using the Wirtinger derivative of the cost function C we obtain
Ryy:=E{yyH} denotes the 2(Lw+1)×2(Lw+1) auto-correlation matrix of the received samples and Rya:=E{ayH} denotes the (L′+1)×2(Lw+1) cross-correlation matrix between the data sequence and the received samples. Combining (5) and (6), there is obtained the optimum filter coefficients minimizing the cost function C as
ŵH=ĥwHRyaRyy−1. (7)
By substituting (7) into (4), there is obtained
Cmin=ĥwH[Raa−RyaRyy−1RyaH]ĥw, (8)
where Raa:=E{aaH}=I is the (L′+1)×(L′+1) auto-correlation matrix of the data symbols a[k], which is an identity matrix due to the assumption of i.i.d. data, 0≦k≦K−1.
According to conventional concept, equation (8) leads to an eigenvalue problem:
[−RyaRyy−1RyaH]ĥw=λminĥw, (9)
where λmin is the smallest eigenvalue of the matrix [I−RyaRyy−1RyaH]. The optimum channel coefficient vector ĥw is equivalent to the eigenvector corresponding to the smallest eigenvalue λmin of the matrix [I−RyaRyy−1RyaH]. To eliminate the trivial result ĥw=0, a certain constraint has to be defined. Often, the overall impulse response is (i) chosen to be monic, i.e., ĥ0,w=1, or (ii) normalized according to ∥ĥw∥2=1. In coded systems it may be more suitable, however, to use a linear filter with amplification one.
An insertion of the channel coefficients ĥw in (7) gives the optimum filter coefficients ŵ. Hence, traditionally, in the first step the channel coefficients and in the second step the filter coefficients are computed. The symbol-spaced channel coefficients ĥw are provided to the decoupled non-linear detector.
The main problem of the conventional filter design discussed so far is the computational complexity. It appears to be very difficult to solve the eigenvalue problem (9), even by means of a Cholesky factorization as proposed in [5].
According to the preferred embodiment, in the following there is presented as an alternative another filter design, which is much simpler to solve and which leads to further insights.
A preferred method for calculating the coefficients of the linear filter is based on an MMSE-DFE equalizer. It is shown in the following that the feedforward filter of a fractionally-spaced MMSE-DFE equalizer is equivalent to the desired linear filter. The solution can be written in closed form. The computational complexity is much less than solving the corresponding eigenvalue problem. The filter design gives insight with respect to impulse response truncation.
The proposed solution generalizes the known receiver published in [12], which does not consider interference cancellation and which applies a symbol-spaced feedforward filter.
In order to derive the filter coefficients, there is considered the fractionally-spaced MMSE-DFE shown in
If correct decisions are assumed, the feedforward filter and the feedback filter of the MMSE-DFE are adapted so that the cost function
is minimized with respect to {tilde over (w)} and {tilde over (h)}w. The vectors {tilde over (w)} and y are the same as defined above. In (10), the data sequence is defined as a=[a[k−k0−1], . . . , a[k−k0−L′]]T, and the overall channel coefficients of the desired user are defined as {tilde over (h)}w=[{tilde over (h)}1,w, {tilde over (h)}2,w, . . . , {tilde over (h)}L′,w]T. The parameter k0 is the decision delay and L′ is the effective memory length of the overall channel impulse response of the desired user. k0 and L′ are design parameters. Note that the MMSE design criterion is the same as maximizing the SINR as considered above, if the overall impulse response is monic, cf. (3) and (10).
The Wirtinger partial derivatives, which minimize the cost function C, are given as follows:
Assuming i.i.d. data, after some calculation the following relations are obtained:
ŵHRyy=ĥwHRya+ryaH and (12)
ĥwH=ŵHRyaH. (13)
As defined before, Ryy:=E{yyH} denotes the 2(Lw+1)×2(Lw+1) auto-correlation matrix of the received samples y(i)(t) and Rya:=E{ayH} denotes the L′×2(Lw+1) cross-correlation matrix between the data a(i) and the received samples y(i)(t), and the 2(Lw+1) vector is defined as rya=[rya[k0], rya[k0−1], . . . , rya[k0−2Lw]]T:=E{a*[k−k0]y}. Upon insertion of (13) in (12) the optimum filter coefficients are obtained as
ŵH=ryaH[Ryy−RyaHRya]−1. (14)
The corresponding channel coefficient vector h is obtained by inserting (14) into (13). Hence, in the proposed solution the filter coefficients are computed prior to the channel coefficients. The symbol-spaced channel coefficients ĥw are provided to the decoupled non-linear detector. According to (13), (17) and (18), the overall channel coefficients ĥw can be written as a convolution between the filter coefficients, ŵ, and the channel coefficients, h.
The solution discussed so far is suitable for complex symbol alphabets such as 8-PSK (PSK: Phase Shift Keying). For one dimensional symbol alphabets such as M-PAM (M-ary-Pulse Amplitude Modulation) or derogated GMSK (Gaussian Minimum Shift Keying), the performance can be improved by applying real-valued processing as proposed in [11].
According to a first case (A) there is performed a calculation of correlation matrices in case of available channel estimates.
In the conventional solution as well as in the proposed preferred solution, in particular the auto-correlation and cross-correlation matrices Ryy:=E{yyH} and Rya:=E{ayH} have to be computed for each burst. This computation can be done in a straightforward manner by taking the expected values E{yyH} and E{ayH} without applying any channel knowledge.
However, in case of i.i.d. data there can be observed a significant performance improvement compared to using a short training sequence if channel estimates are used. According to the first case, we assume that channel estimates for the desired user as well as for the interferer are available. This assumption is suitable for synchronous TDMA networks e.g. using defined training sequences. According to a second case, we assume that channel estimates are only available for the desired user, a scenario that is more suitable for asynchronous TDMA networks. According to a third case, we assume that no channel estimates are available for the filter design.
Consider the auto-correlation matrix Ryy defined as
In case of i.i.d data, the elements of the 2(Lw+1)×2(Lw+1) auto-correlation matrix can be calculated as follows:
where σn2 is the noise variance. Correspondingly, the elements rya(i,j) of the L′×2(Lw+1) cross-correlation matrix
can be calculated as
Finally, the elements of the cross-correlation vector rya can be computed as
Note that for each burst estimates of h and gj, 1≦j≦J and an estimate of the noise variance σn2 have to be known at the receiver. The degradation is small, however, if a constant noise variance, e.g. σn2=0.001, is assumed. Though an explicit knowledge of the data is not necessary, training sequences are helpful in order to compute the channel estimates.
Second case (B) regards to a calculation of correlation matrices in case that channel estimates are not available for the interferer.
Particularly in asynchronous TDMA networks, an estimation of the channel coefficients of the interferer is difficult. In case that estimates of the channel coefficients of the desired user are available, but estimates of the channel coefficients of the interferer are not available, the auto-correlation matrix Ryy can be approximated as follows:
The main idea is to partition Ryy in the form
Ryy=E{ydydH}+E{y1+ny1+nH}=Ryyd+Ryyi+n, (20)
where yd(i)[k],
corresponds to the desired user and yi+n(i)[k]
corresponds to the interferer plus noise in case of two interferers iε={1,2}. The elements of the auto-correlation matrix Ryyd can determined for several possible cases and can be written as
The elements of the auto-correlation matrix Ryyi+n can be approximated correspondingly as
where y1+n(i)[k]:=y(i)[k]−yd(i)[k]. The parameter K′ is the length of the training sequence if training symbols are used in order to calculate Ryyi+n. If tentative decisions are used in order to calculate Ryai+n, the parameter K′ is the burst length. Note that no additional noise variance estimation is needed.
Third case (C) regards to a calculation of correlation matrices in case that no channel estimates are available.
In the case that co-channel interference is strong, channel estimates for the desired user are poor in synchronous and asynchronous networks. Also, channel estimates for the interferer are difficult to obtain in asynchronous networks. For these reasons, the following novel solution is proposed, where no channel estimation and no estimation of the noise variance is necessary for the design of the linear filter. In a first step, there are computed the filter coefficients
ŵH=ryaH[Ryy−RyaHRya]−1, (25)
where Ryy, Rya and ryaH are computed as follows.
Again, the elements of the auto-correlation matrix can be approximated as
A parameter K′ is the length of the training sequence if training symbols are used in order to calculate Rya. If tentative decisions are used in order to calculate Rya, the parameter K′ is the burst length.
Correspondingly, the elements of the cross-correlation matrix Rya can be approximated as
Finally, the elements of the cross-correlation vector rya can be approximated as
In a second step, the convolution
yw:=wHy (29)
is performed.
In a third step, a conventional LS channel estimation is done given yw and a. Due to the interference suppression, the channel estimation error is reasonably small. Finally, MLSE of the desired user is done given yw and the estimated channel coefficients.
For the numerical results presented in
In
Not filled spots and boxes show values according to performing signal processing according to preferred embodiment. As a reference, performance results for the conventional receiver are included as well as black filled spots and boxes. The performance/complexity trade-off is quite remarkable. Particularly, note that in
Numerical results shown for the popular GSM system illuminate significant performance improvements. The performance may be further improved (i) by iterative processing and (ii) by introducing a weighting factor ρ, which is a function of the SINR, in conjunction with
Ryy=Ryyd+ρRyyi+n (30)
in order to avoid performance degradation compared to the conventional receiver at high SINR. The weighting factor may be obtained by a table-look-up. Although oversampling is assumed, degradation in case of a symbol-spaced implementation is not serious.
When determining, i.e. calculating the auto-correlation matrix it is possible to use knowledge of at least one desired channel transmitting desired data from first base station BS1 and using knowledge of at least one disturbing, especially interfering channel transmitting disturbing data from e.g. second and third base stations BS2, BS3 said knowledge being determined by channel estimation in the decision-feedback equalizer DFE. According to second case B, calculation of the auto-correlation matrix is done using only knowledge of at least one desired channel transmitting desired data. According to third case C, calculation of the auto-correlation matrix is done using neither knowledge of a desired channel transmitting desired data nor using knowledge of a disturbing, especially interfering channel transmitting disturbing data.
When designing the filter coefficients for pre-filtering there can be considered only one or a plurality of receiving antennas. Further, the filter coefficients can consider data received according to the timing of the symbols or can consider oversampling. The filter coefficients can be designed as real or complex coefficients. Further, the filter coefficients for pre-filtering can be improved by an iterative method.
In addition, it is possible to estimate the SINR for controlling the design of the filter coefficients to avoid a degradation compared with a conventional receiver in case of high SINR.
Further, it is possible to process the output signal of the pre-filter process i.e. the pre-filtered data z using a memory-less detector, using a backwardly directed filter of a MMSE-DFE equalizer together with a memory-less detector, or using a non-linear equalizer, e.g. a Trellis tree or graph based equalizer.
Especially, it is preferred to use channel coefficients, especially coefficients of channel impulse response, of a non-linear equalizer corresponding to the coefficients of an MMSE-DFE backwardly directed filter.
Advantageously, the channel coefficients of the non-linear equalizer can be determined or calculated by convolution of the coefficients of the pre-filter with the originally estimated channel coefficients of the desired signal. Further, it is possible to estimate the channel coefficient of the non-linear equalizer from the pre-filtered received signal with aid of a channel estimator.
Summarized, a decoupled linear filter/non-linear detector was developed. The purpose of the fractionally-spaced linear filter is to suppress CCI and to shorten the overall impulse response. The purpose of the non-linear detector is ISI rejection for the desired user. The receiver structure is compatible with conventional TDMA receivers.
The proposed solution generalizes the known receiver published in [12]. Particular emphasis is on adaptation techniques for the linear filter taking channel knowledge of the desired user and, optionally, of the interferer into account. Also, the case of no channel estimation for the purpose of designing the linear filter is considered.
Number | Date | Country | Kind |
---|---|---|---|
EP04007211 | Mar 2004 | EP | regional |