Neural network IS-95 rate determination

[0002] 1. Field of the Invention

[0003] This invention relates to decoding in a spread spectrum, code division, multiple access (CDMA) communication system and, more particularly, to an apparatus and method of determining a data rate of an encoded user communication for purposes of controlling decoding of the user communication using a viterbi decoder.

[0004] 2. Related Art

[0005] The CDMA system is interference limited, i.e. signals from one user create interference to the other users. Performance of the system depends on the total interference level. The system is designed so that every user transmits as little power as possible and only when it is necessary, thereby allowing the whole system to support more users.

[0006] In the IS-95 standard, the data rate depends on the speech activity and can change from frame to frame. However, the receiver does not know the data rates of the received frames. Referring to FIG. 1, a prior art CDMA receiver is shown. The user communication signal is received by a RAKE receiver and passed to three averaging circuits 14a, 14b, and 14cconnected in parallel where an average of pairs, 4 tuples, and 8 tuples, respectively, is made. The output of the RAKE receiver is also supplied to the input of a first viterbi decoder 16a. The output of the averaging circuit 14ais supplied to the input of a viterbi decoder 16b. The output of the averaging circuit 14bis supplied to the input of a viterbi decoder 16c. The output of the averaging circuit 14cis supplied to the input of a viterbi decoder 16d. One output from each of the viterbi decoders 16a, 16b, 16c, and 16dis supplied to a selector circuit 24. Another output from each of the viterbi decoders 16a, 16b, 16c, and 16dis supplied through a separate encoder 18a, 18b, 18cand 18dand a separate bit error counter 20a, 20b, 20cand 20d, respectively, to a comparator 22. The comparator 22 outputs a signal to the selector 24 indicative of the which bit error counter has the least bit errors and the selector 24 then selects the output of the corresponding viterbi decoder 16a, 16b, 16c, and 16das the output.

[0007] Therefore, conventionally, in order to decode the received frame correctly, the received frame has to be decoded four times using viterbi decoding for the four data rates 1.2 kbps, 2.4 kbps, 4.8 kbps and 9.6 kbps in rate set 1 or 1.8 kbps, 3.6 kbps, 7.2 kbps and 14.4 kbps in rate set 2. Then the four decoded frames are re-encoded and the re-encoded frames are compared with the original received frame. The embedded CRC bits or the number of bit corrections will then tell which sub-rate was transmitted. However, this method requires high complexity and power consumption because the receiver has to decode the received frame four times.

[0008] Various attempts to avoid this problem have been proposed. Channel decoding power consumption at the mobile station can be reduced by determining the sub-rate (“rate determination”) prior to applying the viterbi decoder. For such methods, the viterbi decoder is applied once per frame. Papers which describe these methods are: E. Cohen and H. Lou, “Multi-rate Detection for the IS-95 CDMA forward traffic channels”, IEEE Global Telecommunications Conference,November 1995 (hereinafter referred to as the “Cohen-Lou method”); E. Cohen and H. Lou, “Multi-rate Detection for the IS-95A CDMA Forward Traffic Channels Using The 13 kbps Speech Coder”, IEEE International Conference on Communications,1996; H. Kwon, S. Ratanamahatana, and J. Shim, “One Viterbi Decoder with Data Rate Estimation for IS-95 CDMA Wireless Communications”, IEEE,1997. Patents which describe such processes are U.S. Pat. Nos. 5,509,020 and 5,796,757. However, such prior art methods are often complex, requiring many arithmetic operations which require more power and are difficult to implement in actual practice with acceptable accuracy.

[0009] An apparatus for determining a data rate of a received, encoded signal, according to the invention includes a deinterleaver for deinterleaving the recorded, encoded signal and outputting a frame of deinterleaved symbols, a feature calculation circuit for applying a plurality of different algorithms to the frame of deinterleaved symbols to produce a corresponding plurality of output feature values which are indicative of a degree of repetition of the deinterleaved symbols, and a neural network for processing the plurality of output feature values according to a predetermined set of weights to produce a plurality of output rate determination values y₁, y₂, . . . y_n, each of which corresponds to a different data rate, where m≦y≦M, n is an integer and m and M are predetermined minimum and maximum values, respectively. A rate detection circuit compares the plurality of output rate determination values y₁, y₂, . . . y_nand selects a rate corresponding to the largest output rate determination value y₁, y₂, . . . y_n. Where the encoded signal is a CDMA signal, then there are 384 deinterleaved symbols per frame and m=0 and M=1.

[0010] Preferably, the data rates include a full data rate, a half data rate, a quarter data rate, and an eighth data rate. The feature calculation circuit calculates four output feature values, x₁, x₂, x₃, and x₄according to the following algorithms: $x_1=\frac{\sum _{k=0}^{191}\left|b_{2k}+b_{2k+1}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_2=\frac{\sum _{k=0}^{95}\left|b_{4k}+b_{4k+1}+b_{4k+2}+b_{4k+3}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_3=\frac{\sum _{k=0}^{47}\left|b_{8k}+b_{8k+1}+b_{8k+2}+b_{8k+3}+b_{8k+4}+b_{8k+5}+b_{8k+6}+b_{8k+7}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_4=\frac{\sum _{k=0}^{383}\left|\left|b_k\right|-\frac{1}{384}\sum _{j=0}^{383}\left|b_j\right|\right|}{\sum _{k=0}^{383}\left|b_k\right|}$

[0011] where x₁indicates the degree of consistent repetition within every 2 deinterleaved symbols, x₂indicates the degree of consistent repetition within every 4 deinterleaved symbols, x₃indicates the degree of consistent repetition within every 8 deinterleaved symbols, and x₄indicates a noise to signal+noise ratio.

[0012] In the preferred embodiment, each weight in the neural network between x_j; and a sigmoid function representative of y_iis w_ij(i references the row, j the column):

[0013] where c_iis a bias value.

[0014] Another embodiment of the invention is an apparatus for decoding a received, encoded data signal having a plurality of data rates and which includes the above described rate determination apparatus.

[0015] Still another embodiment of the invention is an improved user terminal for receiving a radio frequency (RF) signal that conveys a spread spectrum, code division, multiple access (CDMA) user communication, the user communication being transmitted in frames of data bits, individual ones of the frames being transmitted with different data rates which are selected from a plurality of data rates, the user terminal including a deinterleaver for deinterleaving the recorded, encoded signal and outputting frames of deinterleaved symbols, and a decoder for decoding each frame of the user communication at an appropriate data rate, wherein the improvement is an apparatus for determining a data rate as described above and using the determined data rate to control the decoder.

[0016] A method, according to the present invention, for determining a data rate of a received, encoded signal uses the steps of deinterleaving the recorded, encoded signal and outputting a frame of deinterleaved symbols b₀, b₁, b₂, . . . b_N, where N is an integer, applying a plurality of different algorithms to the frame of deinterleaved symbols to produce a corresponding plurality of output feature values which are indicative of a degree of repetition of the deinterleaved symbols, and using a neural network to process the plurality of output feature values according to a predetermined set of weights to produce a plurality of output rate determination values y₁, y₂, . . . y_n, (n is an integer), each of which corresponds to a different data rate. Each y is between m and M, where m and M are predetermined minimum and maximum values, respectively. Further steps are comparing the plurality of output rate determination values y₁, y₂, . . . y_nand selecting a rate corresponding to the largest output rate determination value y₁, y₂, . . . y_n. In the method of the preferred embodiment n=4, N=384, m=0 and M=1.

[0017] Where the data rates may include a full data rate, a half data rate, a quarter data rate, and an eighth data rate, the output feature value calculation step calculates four output feature values, x₁, x₂, x₃, and x₄according to the algorithms referred to above. Each weight in the neural network between x_jand y_iis a value w_ij(i references the row, j the column) given in the table above.

[0018] A method of decoding a received, encoded data signal having a plurality of data rates, includes the above described steps and further includes the step of decoding the encoded data signal at the determined data rate, preferably using a viterbi decoder. The encoded signal can be a spread-spectrum code division multiple access (CDMA) signal.

[0019] The foregoing and other objectives, features and advantages of the invention will be more readily understood upon consideration of the following detailed description of certain preferred embodiments of the invention, taken in conjunction with the accompanying drawings.

[0020] FIG. 1 is a block diagram of a conventional CDMA receiver.

[0021] FIG. 2 is a block diagram of a CDMA receiver according to the present invention.

[0022] FIG. 3 is a diagram of the neural network topology utilized in the neural network depicted in FIG. 2.

[0023] FIG. 4 is a simplified diagram of the neural network topology utilized in a preferred embodiment of the present invention.

[0024] FIG. 5 is a diagram depicting an RS1 rate determination performance comparison for all sub-rates.

[0025] FIG. 6ais a diagram depicting an RS1 rate determination performance comparison for full rate.

[0026] FIG. 6bis a diagram depicting an RS1 rate determination performance comparison for half rate.

[0027] FIG. 6cis a diagram depicting an RS1 rate determination performance comparison for quarter rate.

[0028] FIG. 6dis a diagram depicting an RS1 rate determination performance comparison for eighth rate.

[0029] FIG. 7 is a diagram depicting an RS1 rate determination performance comparison for all sub-rates without using the fourth input x₄.

[0030] Referring now to FIG. 2, the receiver according to the present invention includes a deinterleaver 30 for deinterleaving the recorded, encoded signal and outputting a frame of deinterleaved symbols b₀, b₁, b₂, . . . b_N, where N is an integer. The output of the deinterleaver 30 is supplied to a feature calculation circuit 34 for applying a plurality of different algorithms to the frame of deinterleaved symbols to produce a corresponding plurality of output feature values which are indicative of a degree of repetition of the deinterleaved symbols. The output of the feature calculation circuit 34 is supplied to a neural network 36 for processing the plurality of output feature values according to a predetermined set of weights to produce a plurality of output rate determination values y₁, y₂, . . . y_n, each of which corresponds to a different data rate, where m≦y≦M, n is an integer and m and M are predetermined minimum and maximum values, respectively. The neural network 36 includes a rate detection circuit for comparing the plurality of output rate determination values y₁, y₂, . . . y_nand selecting a rate corresponding to the largest output rate determination value y₁, y₂, . . . y_n. The selected rate from the neural network 36 is supplied to a viterbi decoder 38 which decodes the frame of deinterleaved symbols from the deinterleaver 30 at the selected rate.

[0031] In the new method of neural network rate determination (herein referred as NNRD) according to the present invention and for rate set 1,4, output feature values (x₁, x₂, x₃, x₄) are computed for input into a neural network. These output feature values are computed from a frame of de-interleaved symbols b₀. . . b₃₈₃as follows: $x_1=\frac{\sum _{k=0}^{191}\left|b_{2k}+b_{2k+1}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_2=\frac{\sum _{k=0}^{95}\left|b_{4k}+b_{4k+1}+b_{4k+2}+b_{4k+3}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_3=\frac{\sum _{k=0}^{47}\left|b_{8k}+b_{8k+1}+b_{8k+2}+b_{8k+3}+b_{8k+4}+b_{8k+5}+b_{8k+6}+b_{8k+7}\right|}{\sum _{k=0}^{383}\left|b_k\right|}$ $x_4=\frac{\sum _{k=0}^{383}\left|\left|b_k\right|-\frac{1}{384}\sum _{j=0}^{383}\left|b_j\right|\right|}{\sum _{k=0}^{383}\left|b_k\right|}$

[0032] The output feature values x₁, x₂, x₃and x₄range between 0 and 1. High values of x₁, x₂, and x₃indicate consistent repetition within every 2 symbols, 4 symbols and 8 symbols respectively. The feature x₄serves as an indication of the noise to signal+noise ratio.

[0033] The weights of the neural network 36 are computed (or “trained”) to minimize the classification error. The network topology used is shown in FIG. 3. During training, the activation functions used at the output nodes are sigmoid functions (f(x)=1/(1+e^−x)). The four outputs (y₁, y₂, y₃, y₄) correspond to the rate determination decision for the 4 sub-rates (full, half, quarter and eighth, respectively). These outputs range in value between 0 and 1, where a higher value indicates that the frame from which x₁, x₂, x₃and x₄were calculated is more likely to be of the corresponding sub-rate. Hence, the rate determination decision would be rate k when y_k=MAX{y₁, y₂, y₃, y₄} (and k=1 for full, 2 for half, 3 for quarter and 4 for eighth rate).

[0034] A single output neural network and a multiple hidden layer neural were considered and evaluated. In a single output, the value is used to discriminate the data rate. However the performance of the single output neural network (which produced single output values of 0.1, 0.3, 0.6, 0.9 to indicate full, half, quarter, and eighth data rates, respectively) was much inferior to that of the 4 output neural network shown in FIG. 3. Multiple hidden layers did noticeably improve performance for low signal-to-noise ratio data.

[0035] The neural network weights were calculated by online backpropagation training using 42,000 patterns. Each pattern consists of the network input vector [x₁, x₂, x₃, x₄] and target output vector [y₁, y₂, y₃, y₄]. These input and output vectors were obtained from generated simulated frames of de-interleaved symbols. The frames' rates were uniformly selected from all sub-rates. The original symbols were randomly generated such that the probability of a +1 value is 0.5 and the probability of a −1 value is 0.5. Gaussian white noise was added so that the ratio of energy per symbol to noise (E_s/N_o) varied uniformly between −3 and 7 dB, with E_s/N_oconstant within each frame. Power control bits are ignored.

[0036] Referring now to FIG. 4, the specific implementation of the neural network 36 is straightforward. The output feature values x₁, x₂, x₃, and x₄are supplied to nodes 42, 44, 46, and 48, respectively, of the neural network topology 40. Each output feature value is multiplied by different weights and supplied to each of the output nodes 50, 52, 54, and 56 which are summing nodes. Separate bias values c₁, c₂, c₃, and c₄are supplied to the nodes 50, 52, 54, and 56, respectively. The outputs z₁, z₂, z₃, and z₄from the nodes 50, 52, 54, and 56, respectively, are supplied to a rate detection portion 58 of the neural network 36 which determines which of the outputs z₁, z₂, z₃, and z₄is the maximum value and outputs a rate determination signal corresponding to which of the outputs is a maximum value.

[0037] One huge simplification can be obtained by the fact that only the relative magnitude of the signal into the output nodes 50, 52, 54, and 56 is needed to make the rate determination. For the inputs into the output node activation function z₁, z₂, z₃, and z₄, the output is normally computed as y_k=sigmoid(z_k)=1/(1+e^−zk). Since the sigmoid function is monotonic increasing, when z_k=MAX{z₁, z₂, z₃, z₄}, then y_k=MAX{y₁, y₂, y₃, y₄}. Thus the sigmoid function of the neural network does not need to be evaluated at all!

[0038] The implementation of the rate determination neural network is given below in Matlab code.

[0039] The simplified implementation is shown in FIG. 4. Note that the weights can be scaled arbitrarily to simplify the multiplication. Additionally, the multiplies can be approximated using shifting and adding/subtracting.

[0040] Each weight in the neural network between x_jand z_iis w_ij(where i references the row, and j references the column):

[0041] For example, with reference to FIG. 4 and Table 1 above, an input x₁supplied to node 42 is multiplied by weight −14.8069 and supplied to node 50 where it is combined with a bias value 20.1445. Similarly, input x₂is multiplied by weight 0.446381 and also supplied to the node 50, and so forth. The input x₁is also multiplied by the weight 34.818 and supplied to the node 52, and so forth.

[0042] The numbers of operations required to perform the rate determination by the neural network 36 are:

[0043] For comparison, the prior art Cohen-Lou method uses 3453 additions, 1344 absolute values and 3 divisions.

[0044] The performance of the rate determination method of the present invention was evaluated by generating simulated frames of rate set 1 deinterleaved symbols and adding white gaussian noise. The symbol values and sub-rate were randomly selected as in the neural network training. The Cohen-Lou rate determination method and the NNRD of the present invention were used for each randomly generated frame of de-interleaved symbols. FIG. 5 and shows the rate determination error rate for all sub-rates and FIGS. 6A-6B show the rate determination error rates for individual sub-rates respectively. The improved performance is due to different statistics and use of neural network instead of limited decision rules of Cohen-Lou.

[0045] Even without using the fourth input x₄into the neural network for reduced complexity and training accordingly, the performance is still better than that of the Cohen-Lou method as shown in FIG. 7.

[0046] Alternative realizations of NNRD include:

[0047] different neural network topologies;

[0048] omission of power control bits in x₁, x₂, x₃or x₄computations;

[0049] biased neural network training for actual speech data, no quarter rate, actual distribution of E_s/N_o;

[0050] using shifting, adding and subtracting to efficiently realize the multiplications of the neural network;

[0051] tuning for Rate Set 2;

[0052] NNRD is best applicable for IS-95 or IS-95B mobile station channel decoder and can be used for a CDMA 2000 mobile station.

[0053] Although the present invention has been shown and described with respect to preferred embodiments, various changes and modifications are deemed to lie within the spirit and scope of the invention as claimed. The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims which follow are intended to include any structure, material, or acts for performing the functions in combination with other claimed elements as specifically claimed.