Communication apparatus, learning apparatus, communication system, control circuit, storage medium, and step update method

Abstract

A communication device is provided with: a linear equalization unit (401) that linearly equalizes a received signal; a tap coefficient adjustment unit (402) that adjusts the tap coefficient (403) on the basis of the step size (405); and a step length learning unit (404) that learns a step length (405), the step length learning unit (404) having: a plurality of neural network layers that calculate an updated tap coefficient on the basis of an initial tap coefficient or an updated tap coefficient output from a preceding stage, a reception signal, and a reference signal (307), and that hold an internal parameter; a learning processing unit that learns a learned error function as a mean square error between a tap coefficient based on a least square solution calculated from the received signal and the reference signal (307) and an updated tap coefficient output from the last stage of the plurality of neural network layers, and updates the internal parameter; and an internal parameter collection unit that updates the step size (405) on the basis of the internal parameter.

Description

Technical Field

[0001] The present disclosure relates to a communication device, a learning device, a communication system, a control circuit, a storage medium, and a step size update method for performing equalization processing. Background Art

[0002] Conventionally, in a wireless communication system, waveform distortion caused by delay spread of a propagation path significantly degrades transmission performance. Therefore, equalization processing for reducing the influence of waveform distortion in a receiver is required. Among representative processes of equalization processing, there is linear equalization in the time domain. Linear equalization can be implemented by a transversal filter, and an equalized output is obtained by multiplying a sampled received signal by filter tap coefficients. In linear equalization, an equalization process of successively updating filter tap coefficients is called adaptive equalization. In adaptive equalization, there is a parameter called a step size that determines the degree of update of filter tap coefficients at each repetition. For example, the following technique is disclosed in Non-Patent Document 1: In adaptive equalization, a fixed-value step size is used to update filter tap coefficients.

[0003] Prior Art Documents

[0004] Non-Patent Documents

[0005] Non-Patent Document 1: "Adaptive Equalization" by S.U.H. QURESHI in Proceedings of the IEEE, vol. 73, no. 9, pp. 1349 - 1387, Sept. 1985 Summary of the Invention

[0006] Problems to be Solved by the Invention

[0007] However, according to the above prior art, in the method of updating filter tap coefficients using a fixed-value step size, although the step size can be determined empirically, there is a problem that it is difficult to obtain the optimal value empirically. Although it is also possible to perform a comprehensive search to determine the optimal step size, the number of searches becomes huge depending on the number of conditions to be considered, the granularity during the search, etc.

[0008] The present disclosure has been made in view of the above circumstances, and an object thereof is to obtain a communication device that can reduce errors in adaptive equalization by adjusting the step size of adaptive equalization.

[0009] Means for Solving the Problems

[0010] To solve the above problems and achieve the object, the communication device of the present disclosure includes: a linear equalization unit that performs linear equalization on a received signal; a tap coefficient adjustment unit that adjusts tap coefficients used in the linear equalization according to a step size; and a step size learning unit that performs learning of the step size. The step size learning unit is characterized by including: a plurality of neural network layers that respectively calculate updated tap coefficients based on a prescribed initial tap coefficient or updated tap coefficients output from a previous stage, the received signal, and a reference signal that is a prescribed signal sequence, and hold internal parameters used in the calculation; a learning processing unit that performs learning by setting an error function during learning as the mean square error between the tap coefficients based on the least squares solution calculated from the received signal and the reference signal and the updated tap coefficients output from the last stage of the plurality of neural network layers, and updates the internal parameters; and an internal parameter collection unit that updates the step size based on the internal parameters collected from the plurality of neural network layers.

[0011] Advantages of the Invention

[0012] The communication device of the present disclosure exhibits the following effects: By adjusting the step size of the adaptive equalization, it is possible to reduce the error in the adaptive equalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 FIG. is a diagram showing a structural example of the communication device according to Embodiment 1.

[0014] Figure 2 FIG. is a diagram showing a structural example of the transceiver processing unit included in the communication device according to Embodiment 1.

[0015] Figure 3 FIG. is a diagram showing a structural example of the reception processing unit included in the transceiver processing unit according to Embodiment 1.

[0016] Figure 4 FIG. is a diagram showing a structural example of the equalization processing unit included in the reception processing unit according to Embodiment 1.

[0017] Figure 5 FIG. is a diagram showing a structural example of the linear equalization unit included in the equalization processing unit according to Embodiment 1.

[0018] Figure 6 FIG. is a diagram showing a structural example of the tap coefficient adjustment unit included in the equalization processing unit according to Embodiment 1.

[0019] Figure 7 FIG. is a diagram showing a structural example of the step size learning unit included in the equalization processing unit according to Embodiment 1.

[0020] Figure 8 FIG. is a diagram showing a structural example of the NN (Neural Network) layer included in the step size learning unit according to Embodiment 1.

[0021] Figure 9 This is a diagram showing a structural example of the step size learning unit included in the equalization processing unit of Embodiment 3.

[0022] Figure 10 This is a diagram showing a structural example of the reception processing unit included in the transmission / reception processing unit of Embodiment 5.

[0023] Figure 11 This is a diagram showing a structural example of the equalization processing unit included in the reception processing unit of Embodiment 5.

[0024] Figure 12 This is a diagram showing a structural example of the tap coefficient adjustment unit included in the equalization processing unit of Embodiment 6.

[0025] Figure 13 This is a diagram showing a structural example of the NN layer included in the step size learning unit of Embodiment 6.

[0026] Figure 14 This is a diagram showing a structural example of the linear equalization unit included in the equalization processing unit of Embodiment 7.

[0027] Figure 15 This is a diagram showing a structural example of the communication device of Embodiment 8.

[0028] Figure 16 This is a diagram showing a structural example of the transmission / reception processing unit included in the communication device of Embodiment 8.

[0029] Figure 17 This is a diagram showing a structural example of the reception processing unit included in the transmission / reception processing unit of Embodiment 8.

[0030] Figure 18 This is a diagram showing a structural example of the equalization processing unit included in the reception processing unit of Embodiment 8.

[0031] Figure 19 This is a diagram showing a structural example of the communication device of Embodiment 9.

[0032] Figure 20 This is a diagram showing a structural example of the equalization processing unit included in the reception processing unit of Embodiment 10.

[0033] Figure 21 This is a diagram showing a structural example of the NN layer included in the step size learning unit of the learning device of Embodiment 10.

[0034] Figure 22 This is a diagram showing a structural example of the communication system in the case of collaborative learning by M communication devices of Embodiment 11.

[0035] Figure 23It is a flowchart showing the operation of the communication device according to Embodiment 12.

[0036] Figure 24 It is a diagram showing a structural example of a processing circuit when the processing circuit of the communication device according to Embodiment 12 is implemented by a processor and a memory.

[0037] Figure 25 It is a diagram showing an example of a processing circuit when the processing circuit of the communication device according to Embodiment 12 is composed of dedicated hardware. Detailed Embodiments

[0038] Hereinafter, a communication device, a learning device, a communication system, a control circuit, a storage medium, and a step size update method according to embodiments of the present disclosure will be described in detail with reference to the drawings.

[0039] Embodiment 1

[0040] Figure 1 It is a diagram showing a structural example of the communication device 100-1 according to Embodiment 1. The communication device 100-1 has a transmission / reception processing unit 101 and a control unit 102 that controls the transmission / reception processing unit 101. In Figure 1 It also shows the communication device 100-2 that is the communication partner of the communication device 100-1. Although not shown in Figure 1 The structure of the communication device 100-2 may be the same as that of the communication device 100-1 or different from that of the communication device 100-1.

[0041] Figure 2 It is a diagram showing a structural example of the transmission / reception processing unit 101 included in the communication device 100-1 according to Embodiment 1. The transmission / reception processing unit 101 has a transmission processing unit 201 and a reception processing unit 203. The transmission processing unit 201 transmits a transmission signal 202 to the communication device 100-2, and the reception processing unit 203 receives a reception signal 204 from the communication device 100-2.

[0042] Figure 3 It is a diagram showing a structural example of the reception processing unit 203 included in the transmission / reception processing unit 101 according to Embodiment 1. The reception processing unit 203 has a pre-equalization processing unit 301, an equalization processing unit 303, a post-equalization processing unit 305, and a reference signal generation unit 306.

[0043] Figure 4 It is a diagram showing a structural example of the equalization processing unit 303 included in the reception processing unit 203 according to Embodiment 1. The equalization processing unit 303 has a linear equalization unit 401, a tap coefficient adjustment unit 402, a step size learning unit 404, and an input destination control unit 406.

[0044] Figure 5FIG. is a diagram showing a structural example of the linear equalization unit 401 included in the equalization processing unit 303 according to Embodiment 1. The linear equalization unit 401 has a transversal filter structure. The linear equalization unit 401 includes L-1 delay elements 501-1 to 501-L-1, L multipliers 502-1 to 502-L, a coefficient distributor 503, and an adder 506. Regarding L that determines the number of the delay elements 501-1 to 501-L-1 and the number of the multipliers 502-1 to 502-L, it can be freely set according to the desired communication performance and the like in the communication device 100-1.

[0045] Figure 6 FIG. is a diagram showing a structural example of the tap coefficient adjustment unit 402 included in the equalization processing unit 303 according to Embodiment 1. The tap coefficient adjustment unit 402 adjusts the tap coefficient 403 of the adaptive equalization performed by the equalization processing unit 303. The tap coefficient adjustment unit 402 includes L-1 delay elements 601-1 to 601-L-1, L multipliers 602-1 to 602-L, a coefficient distributor 603, an adder 606, a subtractor 608, and a tap coefficient update unit 610.

[0046] Figure 7 FIG. is a diagram showing a structural example of the step size learning unit 404 included in the equalization processing unit 303 according to Embodiment 1. The step size learning unit 404 includes a plurality of NN layers, that is, a plurality of neural network layers, and calculates the step size 405 of the adaptive equalization performed by the equalization processing unit 303. The step size learning unit 404 includes an internal parameter collection unit 700, an NN control unit 701, a deep NN unit 705 having K NN layers 704-1 to 704-K, and a learning processing unit 708. Regarding the number K of the NN layers 704-1 to 704-K, it can be freely set according to the desired communication performance and the like in the communication device 100-1.

[0047] Figure 8 FIG. is a diagram showing a structural example of the NN layer 704-k included in the step size learning unit 404 according to Embodiment 1. Here, k is an integer from 1 to K. The NN layer 704-k is used when learning the step size 405 of the adaptive equalization performed by the equalization processing unit 303. The NN layer 704-k includes an inner product calculator 800, a subtractor 802, multipliers 804 and 806, an adder 808, and an internal parameter holding unit 809-k.

[0048] Next, the operation of the communication device 100-1 will be described. As Figure 1 shown, in the communication device 100-1, the transceiver processing unit 101 performs transmission and reception processing in communication with the communication device 100-2 under the control of the control unit 102. As Figure 2As shown, in the transceiver processing unit 101, the transmission processing unit 201 transmits the transmission signal 202 to the communication device 100-2, and the reception processing unit 203 receives the reception signal 204 from the communication device 100-2.

[0049] As Figure 3 shown, in the reception processing unit 203, the pre-equalization processing unit 301 performs signal processing required before equalization, i.e., before the equalization processing in the equalization processing unit 303, and outputs the pre-equalization signal 302 to the equalization processing unit 303. The signal processing required before equalization can be any method as long as it outputs the pre-equalization signal 302 suitable for the processing in the equalization processing unit 303. For example, it can be correction of circuit mismatch in the transceiver, correction of Doppler frequency shift, frequency conversion, analog-to-digital conversion, sampling rate conversion, level adjustment, etc. Here, the transceiver refers to the communication device 100-1. Circuit mismatch in the transceiver includes, for example, carrier frequency offset, frequency deviation, phase noise, nonlinearity of the amplifier, etc. Frequency conversion is, for example, down-conversion. Sampling rate conversion is, for example, up-sampling, down-sampling, etc. Level adjustment is, for example, amplification, attenuation, etc. Additionally, when the reception signal 204 is a signal that can be used in the equalization processing in the equalization processing unit 303, i.e., when the equalization processing unit 303 can perform equalization processing on the reception signal 204, the reception processing unit 203 may not have the pre-equalization processing unit 301. In this case, the pre-equalization signal 302 input to the equalization processing unit 303 becomes the reception signal 204. The same applies to the reception processing unit described later.

[0050] The reference signal generation unit 306 outputs a predetermined signal sequence (e.g., pilot signal) as the reference signal 307 to the equalization processing unit 303. A predetermined signal sequence is, for example, a PN (Pseudorandom Noise) sequence, Gold sequence, M sequence, ZC (Zadoff-Chu) sequence, etc., and can be any as long as it is suitable for the processing in the equalization processing unit 303.

[0051] The equalization processing unit 303 obtains the pre-equalization signal 302 from the pre-equalization processing unit 301, obtains the reference signal 307 from the reference signal generation unit 306, performs a predetermined processing using the pre-equalization signal 302 and the reference signal 307, and then outputs the equalized signal 304 to the post-equalization processing unit 305. The predetermined processing in the equalization processing unit 303 will be described later.

[0052] The equalization post-processing unit 305 obtains the post-equalization signal 304 from the equalization processing unit 303, and performs the processing required after equalization on the post-equalization signal 304, which is the equalization processing in the equalization processing unit 303. The processing required after equalization is, for example, sampling rate conversion, level adjustment, symbol determination, error correction, etc. The sampling rate conversion is, for example, upsampling, downsampling, etc. The level adjustment is, for example, amplification, attenuation, etc.

[0053] As Figure 4 shown, in the equalization processing unit 303, the input destination control unit 406 has three internal states: a linear equalization stage, a tap coefficient adjustment stage, and a learning stage. According to each stage, the pre-equalization signal 302 obtained from the pre-equalization processing unit 301 is distributed to the equalization unit input signal 407-1, the tap adjustment signal 407-2, and the learning signal 407-3 and output. For example, within the time range of receiving a data signal, the input destination control unit 406 outputs the pre-equalization signal 302 as the equalization unit input signal 407-1 to the linear equalization unit 401. In addition, within the time range of receiving a known sequence such as a preamble, the input destination control unit 406 outputs the pre-equalization signal 302 as the tap adjustment signal 407-2 to the tap coefficient adjustment unit 402. In addition, within the time range of receiving learning data, the input destination control unit 406 outputs the pre-equalization signal 302 as the learning signal 407-3 to the step size learning unit 404. In addition, as described above, when the pre-equalization signal 302 is the received signal 204, the equalization unit input signal 407-1, the tap adjustment signal 407-2, and the learning signal 407-3 also become the received signal 204. The same applies to the equalization processing unit described later.

[0054] In addition, these three internal states are not mutually exclusive and can also hold simultaneously. For example, within the time range of receiving a known sequence, the input destination control unit 406 outputs the pre-equalization signal 302 as the tap adjustment signal 407-2 to the tap coefficient adjustment unit 402 and also outputs it as the learning signal 407-3 to the step size learning unit 404, and outputs it to two places simultaneously. In this way, the equalization processing unit 303 can adjust the tap coefficient 403 while performing the learning of the step size 405. In addition, when the receiving processing unit 203 does not have the pre-equalization processing unit 301, the equalization unit input signal 407-1, the tap adjustment signal 407-2, and the learning signal 407-3 are the same as the received signal 204.

[0055] The linear equalization unit 401 uses the tap coefficient 403 obtained from the tap coefficient adjustment unit 402 to perform linear equalization on the equalization unit input signal 407-1, that is, the pre-equalization signal 302. As Figure 5As shown, in the linear equalizer unit 401, the delay element 501-1 delays the equalizer input signal 407-1 obtained from the input destination control unit 406 by one specified time and outputs it as the signal 500-1. Here, one specified time is, for example, a value that can be set according to the clock period, sampling period, etc. of the circuit of the equalization processing unit 303. Similarly, the delay element 501-2 delays the signal 500-1 by one specified time and outputs it as the signal 500-2, and the delay element 501-L-1 delays the signal 500-L-2 by one specified time and outputs it as the signal 500-L-1. The operations of the L-1 delay elements 501-1 to 501-L-1 are all the same.

[0056] The coefficient distributor 503 distributes the tap coefficients 403 obtained from the tap coefficient adjustment unit 402 to the multipliers 502-1 to 502-L and outputs them as the tap coefficients 504-1 to 504-L. The multipliers 502-1 to 502-L respectively perform complex conjugate multiplication of the equalizer input signal 407-1 and the signals 500-1 to 500-L-1 with the tap coefficients 504-1 to 504-L and output them as the signals 505-1 to 505-L. The complex conjugate multiplication is specifically described taking the multiplier 502-1 as an example. When the equalizer input signal 407-1 is set as X, the tap coefficient 504-1 is set as W, and the signal 505-1 is set as Y, the complex conjugate multiplication is based on Equation (1). Among them, (W*) is the complex conjugate of W. The adder 506 calculates the sum of the signals 505-1 to 505-L and outputs it as the equalized signal 304.

[0057] Y = (W*) × X...(1)

[0058] The tap coefficient adjustment unit 402 adjusts the tap coefficients 403 used in the linear equalization of the linear equalizer unit 401 according to the step size 405 obtained from the step size learning unit 404. As Figure 6 shown, in the tap coefficient adjustment unit 402, the delay element 601-1 delays the tap adjustment signal 407-2 obtained from the input destination control unit 406 by one specified time and outputs it as the signal 600-1. Here, one specified time is, for example, a value that can be set according to the clock period, sampling period, etc. of the circuit of the equalization processing unit 303. Similarly, the delay element 601-2 delays the signal 600-1 by one specified time and outputs it as the signal 600-2, and the delay element 601-L-1 delays the signal 600-L-2 by one specified time and outputs it as the signal 600-L-1. The operations of the L-1 delay elements 601-1 to 601-L-1 are all the same.

[0059] The coefficient distributor 603 distributes the updated tap coefficients 611 obtained from the tap coefficient update unit 610 to the multipliers 602-1 to 602-L, and outputs them as tap coefficients 604-1 to 604-L. The multipliers 602-1 to 602-L respectively perform complex conjugate multiplication of the tap adjustment signal 407-2 and the signals 600-1 to 600-L-1 with the complex conjugates of the tap coefficients 604-1 to 604-L, and output them as signals 605-1 to 605-L. The complex conjugate multiplication process in the tap coefficient adjustment unit 402 is the same as the complex conjugate multiplication process in the linear equalization unit 401. The adder 606 calculates the sum of the signals 605-1 to 605-L and outputs it as a signal 607. The subtractor 608 subtracts the signal 607 from the reference signal 307 and outputs it as a signal 609.

[0060] The tap coefficient update unit 610 holds the "predetermined initial tap coefficients", and at the start of processing, holds the "predetermined initial tap coefficients" as the current tap coefficients internally and outputs them as the updated tap coefficients 611. When the tap coefficient update unit 610 receives the tap adjustment signal 407-2 from the input destination control unit 406, it starts the process of holding the tap adjustment signal 407-2 and holds it internally for L specified times. The process of holding the tap adjustment signal 407-2 in the tap coefficient update unit 610 becomes a FIFO (First-Input First-Output) process, and at the next timing after obtaining it for L specified times, the first tap adjustment signal 407-2 is deleted. The tap coefficient update unit 610 uses the tap adjustment signal 407-2 held internally for L specified times, the signal 609 obtained from the subtractor 608, the step size 405 obtained from the step size learning unit 404, and the current tap coefficients held internally to calculate the next-period tap coefficients. When the l-th signal of the tap adjustment signal 407-2 held internally by the tap coefficient update unit 610 for L specified times is Y(l), the signal 609 is E, the step size 405 is μ, the l-th component of the current tap coefficients is W(l), and the next-period tap coefficients are V(l), their relationship is based on Equation (2). Here, (E*) represents the complex conjugate of E. The tap coefficient update unit 610 outputs the next-period tap coefficients as the updated tap coefficients 611, and replaces the current tap coefficients with the next-period tap coefficients. The tap coefficient adjustment unit 402 repeats this series of processes K times. Here, K is a value predetermined according to the length of the reference signal 307, etc., and can be freely set.

[0061] V(l) = W(l) + (μ × Y(l) × (E*))...(2)

[0062] The step learning unit 404 performs learning of the step 405 and outputs it to the tap coefficient adjustment unit 402. As Figure 7 shown, in the step learning unit 404, when the index of the learning signal 407-3 at the start of learning is set to i by the NN control unit 701, the i-th learning signal 407-3 is set to C(i), and the k-th reference signal 307 is set to D(k), the combination of the learning signals C(i + k), C(i + k - 1), C(i + k - 2), …, C(i + k - L + 1) and the reference signal D(k) is output as the k-th layer data set 702-k to the NN layer 704-k, and the "predetermined initial tap coefficient" is output as the initial tap coefficient 703 to the NN layer 704-1. Here, it is assumed that the "predetermined initial tap coefficient" is the same value as the "predetermined initial tap coefficient" held in the tap coefficient adjustment unit 402. The NN control unit 701 outputs the layer data sets 702-1 to 702-K to the NN layers 704-1 to 704-K.

[0063] The learning processing unit 708 outputs the initial values of the predetermined internal parameters 810 as the update parameters 709-1 to 709-K to the NN layers 704-1 to 704-K.

[0064] The deep NN unit 705 is a multi-layer neural network composed of the NN layers 704-1 to 704-K. In the deep NN unit 705, the k-th NN layer 704-k Figure 8 performs an inner product calculation on the signal 706-k-1 from the previous-level NN layer 704-k-1 and the learning signals C(i + k), C(i + k - 1), …, C(i + k - L + 1) in the layer data set 702-k through the inner product calculator 800 shown in

[0065] S = Σ_(j = 0)^(j = L - 1)(P(j)*) × C(i + k - j)……(3)

[0066] to obtain the signal 801. When the j-th signal of the signal 706-k-1 is set to P(j) and the signal 801 is set to S, the processing of the inner product calculator 800 is based on Equation (3). Here, (P(j)*) represents the complex conjugate of P(j).

[0067] The subtractor 802 subtracts the signal 801 from the reference signal D(k) in the layer data set 702-k and outputs it as the signal 803.

[0068] B(j) = (A*) × C(i + k - j) …… (4)

[0069] The multiplier 806 outputs the multiplication result of the internal parameter 810 obtained from the internal parameter holding unit 809-k and the signal 805 as the signal 807. When the internal parameter 810 is set to μ, the j-th component of the signal 805 is set to B(j), and the j-th component of the signal 807 is set to U(j), the processing of the multiplier 806 is based on Equation (5).

[0070] U(j) = μ × B(j) …… (5)

[0071] The adder 808 outputs the addition result of the signal 706-k-1 and the signal 807 as the signal 706-k. When the signal 706-k-1 is set to P(j), the j-th component of the signal 807 is set to U(j), and the signal 706-k is set to Q(j), the processing of the adder 808 is based on Equation (6).

[0072] Q(j) = P(j) + U(j) …… (6)

[0073] In the NN layers 704-1 to 704-K, the NN layers 704-1 to 704-K-1 output the signals 706-1 to 706-K-1 through the above processing, and the last NN layer 704-K outputs the NN output 706-K through the above processing.

[0074] The internal parameter holding unit 809-k holds the internal parameter 810 to be learned, and when the update parameter 709-k is obtained from the learning processing unit 708, the held internal parameter 810 is updated to the update parameter 709-k.

[0075] As Figure 7 shown, when the NN control unit 701 sets the i-th learning signal 407-3 to C(i) and the k-th reference signal 307 to D(k), it outputs the combination of the learning signals C(i + K), C(i + K - 1), C(i + K - 2), …, C(i - L + 1) and the reference signals D(0), D(1), …, D(K - 1) as the target data set 714.

[0076] The learning processing unit 708 uses the target data set 714 to calculate the data vector c according to Equation (7). Here, "^T" represents transpose.

[0077] c(k) = [C(i + k), C(i + k - 1), C(i + k - L + 1)]^T …… (7)

[0078] The learning and processing unit 708 calculates the correlation matrix R and the correlation vector r according to Equations (8) and (9) using the data vector c and the reference signals D(0), D(1), …, D(K-1).

[0079] R = Σ_(k = 0)^(k = K-1) c(k)c^H(k) ……(8)

[0080] r = Σ_(k = 0)^(k = K-1) (D(k)*) × c(k) ……(9)

[0081] The learning and processing unit 708 calculates the target tap coefficient vector u according to Equation (10) based on the correlation matrix R and the correlation vector r. Here, "^(-1)" represents the inverse matrix operation.

[0082] u = R^(-1)r ……(10)

[0083] The target tap coefficient vector u is the least squares (LS) solution that can be calculated using the target data set 714. The learning and processing unit 708 uses the mean square error (MSE) between the target tap coefficient vector u and the NN output 706-K as an error function to perform learning of the internal parameters 810 held by the internal parameter holding units 809-1 to 809-K in each NN layer 704-1 to 704-K of the deep NN unit 705. When the learning and processing unit 708 updates the internal parameters 810 in each NN layer 704-1 to 704-K during learning, it outputs the updated parameters 709-1 to 709-K to update the internal parameters 810 held by the internal parameter holding units 809-1 to 809-K. The learning and processing unit 708 performs learning of the internal parameters 810, for example, by the stochastic gradient descent method and the error backpropagation method.

[0084] After satisfying the predetermined learning end condition, the internal parameter collection unit 700 collects the internal parameters 810 in the NN layers 704-1 to 704-K as the step sizes 707-1 to 707-K of each layer and outputs them as the step size 405.

[0085] Thus, in the step-size learning unit 404, the NN layers 704-1 to 704-K respectively calculate the updated tap coefficients based on the specified initial tap coefficients 703 or the updated tap coefficients output from the previous NN layer 704-k-1, i.e., the signal 706-k-1, the learning signal 407-3 included in the layer data set 702-k, i.e., the pre-equalization signal 302, and the reference signal 307 which is a specified signal sequence included in the layer data set 702-k, and maintain the internal parameters 810 used in the calculation. The learning processing unit 708 sets the error function during learning as the mean square error between the tap coefficients based on the least squares solution calculated from the learning signal 407-3, i.e., the pre-equalization signal 302 and the reference signal 307, and the updated tap coefficients output from the last stage of the NN layers 704-1 to 704-K, i.e., the NN output 706-K, and updates the internal parameters 810. The internal parameter collection unit 700 updates the step size 405 based on the internal parameters 810 collected from the NN layers 704-1 to 704-K, i.e., the step sizes 707-1 to 707-K of each layer.

[0086] As described above, in the transceiver processing unit 101 of the communication device 100-1, in the equalization processing unit 303 of the reception processing unit 203, the step-size learning unit 404 performs learning of the step size 405 for updating the tap coefficients 403. Thus, even when the number of updates of the tap coefficients 403 is limited to K times, the optimal tap coefficients 403 can be adjusted, and the error in the equalization processing can be minimized. In addition, in the equalization processing unit 303, the step-size learning unit 404 sets the error function during learning as the mean square error between the target tap coefficient vector u of the LS estimation solution, i.e., the NN output 706-K, and the NN output 706-K. Thus, the deterioration of the convergence characteristics during learning can be prevented. The equalization processing unit 303 adjusts the step size 405 of the adaptive equalization, and the error in the adaptive equalization can be reduced.

[0087] Embodiment 2

[0088] In the above Embodiment 1, the learning of the deep NN unit 705 in the step-size learning unit 404 is performed by the stochastic gradient descent method and the error backpropagation method. However, any method that can perform the learning of the multi-layer neural network can be used, and the implementation means is arbitrary. For example, as the learning of the deep NN unit 705, the step-size learning unit 404 can also use AdaGrad, momentum, etc. instead of the stochastic gradient descent method, and can also perform learning using mini-batches.

[0089] As described above, the learning method of the step-size learning unit 404 has freedom. Thus, in order to improve the learning performance or reduce the computational load during learning, the learning method can be adjusted.

[0090] Embodiment 3

[0091] In the above Embodiment 1 to Embodiment 2, the learning of the deep NN unit 705 in the step learning unit 404 is performed for all layers in one go. In Embodiment 3, then, a case of incremental learning in which the NN layers 704-1 to 704-K are added one by one while learning will be described.

[0092] Figure 9 FIG. is a structural example of the step learning unit 404 included in the equalization processing unit 303 of Embodiment 3. The step learning unit 404 includes an internal parameter collection unit 900, an NN control unit 901, a deep NN unit 905, and a learning processing unit 910. The deep NN unit 905 includes K NN layers 904-1 to 904-K and K-1 switches 906-1 to 906-K-1. The number K of the NN layers 904-1 to 904-K can be freely set according to the desired communication performance and the like in the communication device 100-1.

[0093] Next, the operation of the step learning unit 404 will be described. In the step learning unit 404, similar to the processing of the NN control unit 701 in Embodiment 1, the NN control unit 901 outputs layer data sets 902-1 to 902-K and outputs initial tap coefficients 903.

[0094] The learning processing unit 910 outputs the initial values of the predetermined internal parameters 810 as update parameters 911-1 to 911-K. In addition, the learning processing unit 910 outputs all of the SW control signals 912-1 to 912-K-1 as meaningless to the switches 906-1 to 906-K-1.

[0095] The switches 906-1 to 906-K-1 are switches that change the output destinations of the signals 907-1 to 907-K-1, that is, the input signals, obtained from the NN layers 904-1 to 904-K-1. The switches 906-1 to 906-K-1 output the input signals as signals 908-1 to 908-K-1 when the SW control signals 912-1 to 912-K-1 obtained from the learning processing unit 910 are meaningful, and output the input signals as signals 909-1 to 909-K-1 when the SW control signals 912-1 to 912-K-1 are meaningless.

[0096] The deep NN section 905 is a multi-layer neural network composed of NN layers 904-1 to 904-K. The operations of each of the NN layers 904-1 to 904-K are the same as those of the NN layer 704-k in Embodiment 1. The NN layers 904-1 to 904-K output signals 907-1 to 907-K-1 to the switches 906-1 to 906-K-1. In addition, the last-stage NN layer 904-K of the NN layers 904-1 to 904-K outputs an NN output 907-K to the learning processing section 910.

[0097] The output processing of the target data set 914 in the NN control section 901 is the same as the output processing of the target data set 714 in the NN control section 701 of Embodiment 1. In addition, the learning in the learning processing section 910 is the same as the learning in the learning processing section 708 of Embodiment 1. However, since the effective NN layer is only the NN layer 904-1, the learning processing section 910 only performs learning of the internal parameters 810 held by the internal parameter holding section 809-1 of the NN layer 904-1. When the learning processing section 910 updates the internal parameters 810 of the NN layer 904-1 during learning, it outputs an update parameter 911-1 to the NN layer 904-1 to update the internal parameters 810 held by the internal parameter holding section 809-1.

[0098] After the learning processing section 910 satisfies the end condition of the individual learning of the predetermined incremental learning, it makes the SW control signal 912-1 meaningful, makes the SW control signals 912-2 to 912-K-1 meaningless, and repeatedly performs the output processing of the target data set 914 in the NN control section 901 to the learning in the learning processing section 910. After the learning processing section 910 satisfies the end condition of the individual learning of the predetermined incremental learning, it makes the SW control signals 912-1 and 912-2 meaningful, makes the SW control signals 912-3 to 912-K-1 meaningless, and repeatedly performs the output processing of the target data set 914 in the NN control section 901 to the learning in the learning processing section 910. Similarly, the learning processing section 910 makes the SW control signals 912-3 and later meaningful one by one, and repeatedly performs the output processing of the target data set 914 in the NN control section 901 to the learning in the learning processing section 910. After the learning processing section 910 makes the SW control signal 912-K-1 meaningful and satisfies the end condition of the individual learning of the predetermined incremental learning, the internal parameter collection section 900 collects the internal parameters 810 in the NN layers 904-1 to 904-K as the layer step sizes 913-1 to 913-K and outputs them as the step size 405.

[0099] As described above, the step learning unit 404 learns the step 405 through incremental learning. By adding the NN layer 904-k layer by layer during learning, the step learning unit 404 can prevent overfitting that occurs when learning deeper NN layers all at once.

[0100] Embodiment 4

[0101] In the above Embodiment 3, the step learning unit 404 uses the switches 906-1 to 906-K-1 and the SW control signals 912-1 to 912-K-1 to implement the process of adding the NN layer 904-k layer by layer during incremental learning. However, as long as the same process can be performed, the implementation means is arbitrary. For example, the learning processing unit 910 can also set the update parameter 911-k output for the invalid NN layer 904-k to 0.

[0102] As described above, the step learning unit 404 controls the validity of the NN layer 904-k only by the value of the update parameter 911-k. Thus, the switches 906-1 to 906-K-1 can be reduced, the circuit scale can be suppressed, and the processing load can be alleviated.

[0103] Embodiment 5

[0104] In the above Embodiments 1 to 4, the reference signal 307 is used in the adjustment of the tap coefficient 403, the learning of the step 405, etc. Next, in Embodiment 5, the following case will be described: when a signal equivalent to the reference signal can be generated from the equalized signal, the equalized signal is used in the adjustment of the tap coefficient, the learning of the step, etc.

[0105] Figure 10 FIG. is a structural example diagram of the reception processing unit 203 included in the transmission / reception processing unit 101 according to Embodiment 5. The reception processing unit 203 includes a pre-equalization processing unit 1000, an equalization processing unit 1002, a post-equalization processing unit 1004, and a reference signal generation unit 1006.

[0106] Figure 11 FIG. is a structural example diagram of the equalization processing unit 1002 included in the reception processing unit 203 according to Embodiment 5. The equalization processing unit 1002 includes a linear equalization unit 1101, a tap coefficient adjustment unit 1102, a step learning unit 1104, and an input signal storage memory 1106.

[0107] Next, the operation of the reception processing unit 203 will be described. The pre-equalization processing unit 1000 performs pre-equalization processing on the received signal 204 and outputs the pre-equalized signal 1001. The processing of the pre-equalization processing unit 1000 is the same as that of the pre-equalization processing unit 301 in Embodiment 1. The post-equalization processing unit 1004 obtains the equalized signal 1003 from the equalization processing unit 1002, and as processing required after equalization, for example, performs sample rate conversion, level adjustment, symbol determination, error correction, etc., and outputs hard decision bit information, soft decision bit information, error-corrected bit information, etc. as a posteriori information 1005 to the reference signal generation unit 1006. The sample rate conversion is, for example, upsampling, downsampling, etc. The level adjustment is, for example, amplification, attenuation, etc. The reference signal generation unit 1006 generates a reference signal 1007 based on the a posteriori information 1005 obtained from the post-equalization processing unit 1004. The reference signal generation unit 1006, for example, restores data symbols based on the hard decision bit information, soft decision bit information, and error-corrected bit information, etc., and outputs the restored data symbols as the reference signal 1007. In this way, the post-equalization processing unit 1004 performs post-equalization processing on the equalized signal 1003 obtained by linearly equalizing the pre-equalized signal 1001 by the equalization processing unit 1002. The reference signal generation unit 1006 generates a reference signal 1007 based on the a posteriori information 1005 obtained by the post-equalization processing of the post-equalization processing unit 1004.

[0108] In the equalization processing unit 1002, the pre-equalized signal 1001 obtained from the pre-equalization processing unit 1000 is input to the linear equalization unit 1101 and the input signal storage memory 1106. The input signal storage memory 1106 accumulates the pre-equalized signal 1001, and outputs the pre-equalized signal 1001 accumulated at the time corresponding to the reference signal 1007 as the tap adjustment signal 1107-2 and the learning signal 1107-3. The tap coefficient adjustment unit 1102 and the step size learning unit 1104 perform the same processing as the tap coefficient adjustment unit 402 and the step size learning unit 404 in Embodiments 1 to 4. The step size learning unit 1104 learns the step size 1105 based on the learning signal 1107-3 and the reference signal 1007. The tap coefficient adjustment unit 1102 adjusts the tap coefficient 1103 based on the tap adjustment signal 1107-2, the reference signal 1007, and the step size 1105. The linear equalization unit 1101 linearly equalizes the pre-equalized signal 1001 using the tap coefficient 1103 and outputs the equalized signal 1003.

[0109] As described above, the reception processing unit 203 utilizes the pre-equalized signal 1001 in the generation of the reference signal 1007. Therefore, it is not necessary to include a certain signal sequence specified in advance in the transmission signal 202, and thus the transmission rate can be increased.

[0110] Embodiment 6

[0111] In the above-described Embodiments 1 to 5, the LMS is used for updating the tap coefficients. In Embodiment 6, cases where various adaptive algorithms such as Normalized LMS (NLMS: Normalized Least Mean Square), Affine Projection Algorithm (APA), and Recursive Least Squares (RLS) are used for updating the tap coefficients will be described.

[0112] Figure 12 FIG. is a diagram showing a structural example of a tap coefficient adjustment unit 402 included in the equalization processing unit 303 according to Embodiment 6. The tap coefficient adjustment unit 402 includes an a priori estimation error calculation unit 1200 and a tap update processing unit 1201.

[0113] Figure 13 FIG. is a diagram showing a structural example of an NN layer 704-k included in the step size learning unit 404 according to Embodiment 6. The NN layer 704-k includes an internal parameter holding unit 1300-k and a linear operation processing unit 1301-k. That is, although not shown, the step size learning unit 404 includes internal parameter holding units 1300-1 to 1300-K and linear operation processing units 1301-1 to 1301-K in the NN layers 704-1 to 704-K.

[0114] Next, the operation of the equalization processing unit 303 will be described. As Figure 12 shown, in the tap coefficient adjustment unit 402, the a priori estimation error calculation unit 1200 calculates an estimation error 1202 using a tap adjustment signal 407-2, a reference signal 307, updated tap coefficients 1203, and a step size 405, and outputs it to the tap update processing unit 1201. The tap update processing unit 1201 calculates updated tap coefficients 1203 using the tap adjustment signal 407-2, the reference signal 307, the estimation error 1202, and the step size 405, and outputs it to the a priori estimation error calculation unit 1200. The tap coefficient adjustment unit 402 repeats this series of processes K times. Here, K is a value predetermined according to the length of the reference signal 307 or the like and can be freely set.

[0115] In the NN layer 704-k of the step size learning unit 404, the linear operation processing in the linear operation processing unit 1301-k is the same as the linear operation processing in the a priori estimation error calculation unit 1200 and the tap update processing unit 1201 in the tap coefficient adjustment unit 402. The internal parameter holding unit 1300-k has the same structure as the internal parameter holding unit 809-k Figure 8 shown, and outputs internal parameters 1310 to the linear operation processing unit 1301-k.

[0116] As described above, the equalization processing unit 303 generalizes the processing in the tap coefficient adjustment unit 402 into the repetition of the calculation of the estimation error 1202 and the tap update processing, and replaces these processes with a linear operation process having internal parameters 1310 in the NN layer 704-k. Thus, the equalization processing unit 303 can adjust the tap coefficients using various adaptive algorithms such as the affine projection method (APA) and the recursive least squares method (RLS), and can perform the learning of the internal parameters 1310 used in these adaptive algorithms in the step size learning unit 404.

[0117] Embodiment 7

[0118] In the above Embodiments 1 to 6, the linear equalization unit performs linear processing. In Embodiment 7, a case where the linear equalization unit performs widely linear processing will be described.

[0119] Figure 14 FIG. is a diagram showing a structural example of the linear equalization unit 401 included in the equalization processing unit 303 of Embodiment 7. The linear equalization unit 401 includes L-1 delay elements 1400-1 to 1400-L-1, L signal distributors 1402-1 to 1402-L, 2L multipliers 1404-1 to 1404-2L, a coefficient distributor 1405, and an adder 1408.

[0120] Next, the operation of the linear equalization unit 401 will be described. In the linear equalization unit 401, the signal distributor 1402-1 directly outputs the equalization unit input signal 407-1 obtained from the input destination control unit 406 as the signal 1403-1, and outputs the complex conjugate signal of the equalization unit input signal 407-1 as the signal 1403-2. The delay elements 1400-1 to 1400-L-1 output the signals 1401-1 to 1401-L-1 to the signal distributors 1402-2 to 1402-L. The signal distributors 1402-2 to 1402-L also perform the same operation as the signal distributor 1402-1. That is, the signal distributors 1402-1 to 1402-L output the signals 1403-1 to 1403-2L to the multipliers 1404-1 to 1404-2L.

[0121] The coefficient distributor 1405 distributes the tap coefficients 403 obtained from the tap coefficient adjustment unit 402 to the multipliers 1404-1 to 1404-2L, and outputs them as tap coefficients 1406-1 to 1406-2L. The multipliers 1404-1 to 1404-2L respectively perform complex conjugate multiplication of the signals 1403-1 to 1403-2L and the tap coefficients 1406-1 to 1406-2L, and output them as signals 1407-1 to 1407-2L. The adder 1408 calculates the sum of the signals 1407-1 to 1407-2L, and outputs it as the equalized signal 304.

[0122] As described above, the linear equalization unit 401 performs generalized linear processing as linear equalization. The linear equalization unit 401 also performs filter processing on the complex conjugate of the equalization unit input signal 407-1. Thus, for example, it is possible to equalize a Circular signal having a correlation between the real part and the imaginary part of the signal, such as IQ imbalance.

[0123] Embodiment 8

[0124] In the above Embodiments 1 to 7, the learning of the step size is processed inside the communication device 100-1. In Embodiment 8, a case where the step size is learned outside the communication device is described.

[0125] Figure 15 FIG. is a diagram showing a structural example of the communication device 1500-1 according to Embodiment 8. The communication device 1500-1 includes a control unit 1501 and a transceiver processing unit 1502. The transceiver processing unit 1502 is connected to a learning device 1503 located outside the communication device 1500-1. In Figure 15 Also shown is a communication device 1500-2 that is the communication partner of the communication device 1500-1. Although not described in Figure 15 it is omitted, the structure of the communication device 1500-2 may be the same as the structure of the communication device 1500-1 or may be different from the structure of the communication device 1500-1. In addition, the learning device 1503 includes a step size learning unit 1505. Further, the communication device 1500-1 and the learning device 1503 constitute a communication system 1510.

[0126] Figure 16 FIG. is a diagram showing a structural example of the transceiver processing unit 1502 included in the communication device 1500-1 according to Embodiment 8. The transceiver processing unit 1502 includes a transmission processing unit 1601 and a reception processing unit 1602. The transmission processing unit 1601 transmits a transmission signal 1603 to the communication device 1500-2, and the reception processing unit 1602 receives a reception signal 1604 from the communication device 1500-2.

[0127] Figure 17 This is a diagram showing a structural example of the reception processing unit 1602 included in the transmission / reception processing unit 1502 according to Embodiment 8. The reception processing unit 1602 includes a pre-equalization processing unit 1701, an equalization processing unit 1703, a post-equalization processing unit 1705, and a reference signal generation unit 1706. The pre-equalization processing unit 1701 performs pre-equalization processing on the received signal 1604 and outputs a pre-equalization signal 1702. The reference signal generation unit 1706 generates a reference signal 1504-2. The equalization processing unit 1703 outputs a learning signal 1504-1 to the learning device 1503, and performs equalization processing on the pre-equalization signal 1702 using the reference signal 1504-2 obtained from the reference signal generation unit 1706 and the step size 1504-3 obtained from the learning device 1503, and outputs a post-equalization signal 1704. The pre-equalization processing unit 1701, the post-equalization processing unit 1705, and the reference signal generation unit 1706 perform the same operations as Figure 3 the pre-equalization processing unit 301, the post-equalization processing unit 305, and the reference signal generation unit 306 shown.

[0128] Figure 18 This is a diagram showing a structural example of the equalization processing unit 1703 included in the reception processing unit 1602 according to Embodiment 8. The equalization processing unit 1703 includes a linear equalization unit 1801, a tap coefficient adjustment unit 1802, and an input destination control unit 1804. During the time range when a data signal is received, the input destination control unit 1804 outputs the pre-equalization signal 1702 as an equalization unit input signal 1805-1 to the linear equalization unit 1801. Further, during the time range when a known sequence such as a preamble is received, the input destination control unit 1804 outputs the pre-equalization signal 1702 as a tap adjustment signal 1805-2 to the tap coefficient adjustment unit 1802. Further, during the time range when learning data is received, the input destination control unit 1804 outputs the pre-equalization signal 1702 as a learning signal 1504-1 to the learning device 1503. The tap coefficient adjustment unit 1802 adjusts the tap coefficient 1803 using the reference signal 1504-2, the step size 1504-3, and the tap adjustment signal 1805-2. The linear equalization unit 1801 performs linear equalization on the equalization unit input signal 1805-1 using the tap coefficient 1803 and outputs a post-equalization signal 1704.

[0129] Next, the operation of the communication device 1500-1 will be described. In the communication device 1500-1, the transceiver processing unit 1502 transmits the learning signal 1504-1 and the reference signal 1504-2 to the learning device 1503. In the learning device 1503, the step-size learning unit 1505 performs learning of the step size 1504-3 used in the communication device 1500-1. The communication device 1500-1 adjusts the tap coefficients 1803 used in linear equalization according to the step size 1504-3, and performs linear equalization on the equalizer input signal 1805-1, that is, the pre-equalization signal 1702. When the step-size learning unit 1505 receives the learning signal 1504-1 and the reference signal 1504-2 from the transceiver processing unit 1502, it uses the learning signal 1504-1 and the reference signal 1504-2 to perform learning processing, and outputs the learned step size 1504-3 to the transceiver processing unit 1502.

[0130] The step-size learning unit 1505 of the learning device 1503 has the same structure as the step-size learning unit 404 or the step-size learning unit 1104 of the communication device 100-1 in Embodiments 1 to 7, and performs the same operations. That is, the step-size learning unit 1505 performs learning of the step size 1504-3 through the same processing as the processing in which the step-size learning unit 404 performs learning of the step size 405 or the step-size learning unit 1104 performs learning of the step size 1105 in Embodiments 1 to 7. In this way, since the step-size learning unit 1505 performs the same operations as the step-size learning unit 404 or the step-size learning unit 1104, the detailed structure and operations of the step-size learning unit 1505 are omitted. In the communication system 1510, the learning device 1503 outside the communication device 100-1 has the step-size learning unit that the communication device 100-1 has in Embodiment 1 and the like.

[0131] As described above, the communication device 1500-1 performs equalization processing using the step size 1504-3 obtained from the learning device 1503. Since the communication device 1500-1 uses an external learning device 1503 to perform learning of the step size 1504-3, compared with the communication device 100-1 in Embodiment 1 and the like, the load based on the learning processing can be reduced, and the learning accuracy can be improved by using a higher-performance external learning device 1503.

[0132] Embodiment 9

[0133] In the above Embodiment 8, learning of the step size 1504-3 is performed outside the communication device 1500-1. In Embodiment 9, a case where a wireless communication network is used in the signal exchange between the communication device and the learning device will be described.

[0134] Figure 19FIG. is a diagram showing a structural example of the communication device 1900-1 according to Embodiment 9. The communication device 1900-1 includes a control unit 1901 and a transceiver processing unit 1902. In Figure 19 it, the communication device 1900-2 which is the communication partner of the communication device 1900-1 is also illustrated. The transceiver processing unit 1902 is connected to the access point 1903 via the wireless link 1904. The access point 1903 is connected to the learning device 1905. The access point 1903 is a device used in a wireless communication network such as an access point of a wireless LAN (Local Area Network), a base station in a public mobile line, etc. In addition, the communication system 1910 is constituted by the communication device 1900-1, the access point 1903, and the learning device 1905.

[0135] Next, the operation of the communication device 1900-1 will be described. Similar to the transceiver processing unit 1502 of the communication device 1500-1 in Embodiment 8, the transceiver processing unit 1902 of the communication device 1900-1 transmits the learning signal 1905-1 and the reference signal 1905-2 as the signal 1906 to the learning device 1905 via the wireless link 1904 and via the access point 1903. The learning device 1905 includes a step size learning unit 1908, and the step size learning unit 1908 has the same function as the Figure 15 step size learning unit 1505 shown. Similar to the step size learning unit 1505 of Embodiment 8, when the step size learning unit 1908 receives the learning signal 1905-1 and the reference signal 1905-2 from the transceiver processing unit 1902, it performs learning processing using the learning signal 1905-1 and the reference signal 1905-2, and outputs the learned step size 1905-3 as the signal 1907 to the transceiver processing unit 1902.

[0136] As described above, the learning device 1905 communicates with the communication device 1900-1 through wireless communication via the wireless communication network. When the communication device 1900-1 uses the external learning device 1905 to learn the step size 1905-3, it transmits and receives the signals required for learning via the wireless link 1904. Thus, the communication device 1900-1 and the learning device 1905 do not need to be located in the same place, and the communication device 1900-1 can move.

[0137] Embodiment 10

[0138] In the above Embodiment 9, the communication device 1900-1 transmits the transmission signal 1603 as the learning signal 1905-1 to the learning device 1905. In Embodiment 10, the case where the communication device 1900-1 conceals the learning signal will be described.

[0139] Figure 20This is a diagram showing a structural example of the equalization processing unit 1703 included in the reception processing unit 1602 according to Embodiment 10. The equalization processing unit 1703 includes a linear equalization unit 2001, a tap coefficient adjustment unit 2002, an input destination control unit 2004, and a learning signal generation unit 2006. During the time range when the data signal is received, the input destination control unit 2004 outputs the pre-equalization signal 1702 as the equalization unit input signal 2005-1 to the linear equalization unit 2001. In addition, during the time range when a known sequence such as a preamble is received, the input destination control unit 2004 outputs the pre-equalization signal 1702 as the tap adjustment signal 2005-2 to the tap coefficient adjustment unit 2002. Further, during the time range when the data for learning is received, the input destination control unit 2004 outputs the pre-equalization signal 1702 as the learning signal 2005-3 to the learning signal generation unit 2006. The learning signal generation unit 2006 generates the learning signal 1905-1 using the reference signal 1905-2 and the learning signal 2005-3, and outputs it to the learning device 1905. The tap coefficient adjustment unit 2002 adjusts the tap coefficient 2003 using the reference signal 1905-2, the step size 1905-3, and the tap adjustment signal 2005-2. The linear equalization unit 2001 linearly equalizes the equalization unit input signal 2005-1 using the tap coefficient 2003, and outputs the post-equalization signal 1704.

[0140] Figure 21 This is a diagram showing a structural example of the NN layer 704-k included in the step size learning unit 1908 of the learning device 1905 according to Embodiment 10. The NN layer 704-k includes an internal parameter holding unit 2109-k, an inner product calculator 2100, a subtractor 2102, a multiplier 2106, and an adder 2108.

[0141] Next, the operation of the equalization processing unit 1703 will be described. When the i-th learning signal 2005-3 is set as C(i) and the k-th reference signal 1905-2 is set as D(k), the learning signal generation unit 2006 calculates the data vector c(k) according to the combination of C(i+K), C(i+K-1), C(i+K-2),..., C(i-L+1) and D(0), D(1),..., D(K-1) in accordance with Equation (7). In addition, the learning signal generation unit 2006 calculates the correlation matrix R(k) and the correlation vector r(k) using the data vector c(k) and D(0), D(1),..., D(K-1) in accordance with Equations (11) and (12), and outputs the correlation matrix R(k) and the correlation vector r(k) as the learning signal 1905-1.

[0142] R(k) = c(k)c^H(k)......(11)

[0143] r(k) = (D(k)*) × c(k)……(12)

[0144] The learning device 1905 receives the learning signal 1905-1 from the communication device 1900-1. In the learning device 1905, the NN control unit 701 of the step-size learning unit 1908 calculates the target tap coefficient vector u according to Equation (13) based on the correlation matrix R(k) and the correlation vector r(k) included in the learning signal 1905-1. In addition, the NN control unit 701 combines the index k of the NN layers 704-1 to 704-K and outputs the correlation matrix R(k) and the correlation vector r(k) included in the learning signal 1905-1 as the layer dataset 702-k.

[0145] u = (Σ_(k = 0)^(k = K - 1)R(k))^(-1)(Σ_(k = 0)^(k = K - 1)r(k))……(13)

[0146] In the NN layer 704-k, the internal parameter holding unit 2109-k holds the internal parameter 2110 to be learned, and when obtaining the update parameter 709-k from the learning processing unit 708, updates the held internal parameter 2110 to the update parameter 709-k. Since the step-size learning unit 1908 has the NN layers 704-1 to 704-K, it has the internal parameter holding units 2109-1 to 2109-K. The inner product calculator 2100 sets the signal 706-k-1 obtained from the previous layer and calculates the inner product Z with the correlation matrix R(k) included in the layer dataset 702-k according to Equation (14), and outputs it as the signal 2101.

[0147] Z = R(k)W(k - 1)……(14)

[0148] The subtractor 2102 takes the difference r(k) - Z between the signal 2101 obtained from the inner product calculator 2100 and the correlation vector r(k) included in the layer dataset 702-k, and outputs it as the signal 2105.

[0149] The multiplier 2106 multiplies the signal 2105 obtained from the subtractor 2102 and the internal parameter 2110 output from the internal parameter holding unit 2109-k, and outputs it as the signal 2107.

[0150] The adder 2108 calculates the sum of the signal 706-k-1 and the signal 2107 obtained from the multiplier 2106, and outputs it as the signal 706-k.

[0151] As described above, the communication device 1900-1 calculates the correlation matrix R(k) and the correlation vector r(k) based on the pre-equalization signal 1702 and the reference signal 1905-2 that are learning signals 2005-3, and transmits the correlation matrix R(k) and the correlation vector r(k) as learning signals 1905-1 to the learning device 1905. The step-size learning unit 1908 of the learning device 1905 performs learning of the step size 1905-3 using the correlation matrix R(k) and the correlation vector r(k). The communication device 1900-1 sets the learning signal 1905-1 output to the outside as the correlation matrix R(k) and the correlation vector r(k), rather than the received signal itself, and it is not easy to estimate the received signal itself based on the learning signal 1905-1. Thus, learning in an external device can be utilized while maintaining the concealment of the received signal.

[0152] Embodiment 11

[0153] In the above Embodiment 10, learning is performed up to one communication device 1900-1. In Embodiment 11, federated learning using multiple communication devices will be described.

[0154] Figure 22 FIG. is a structural example diagram of a communication system 2210 in the case of performing federated learning by M communication devices 2200-1 to 2200-M according to Embodiment 11. The communication system 2210 includes communication devices 2200-1 to 2200-M, an access point 2202, and a learning device 2203. Among the communication devices 2200-1 to 2200-M, the communication device 2200-m is the m-th communication device. m is an integer from 1 to M. The communication devices 2200-1 to 2200-M are connected to the access point 2202 via wireless links 2201-1 to 2201-M. The wireless links 2201-1 to 2201-M are the same as the Figure 19 wireless link 1904 shown. The access point 2202 is connected to the learning device 2203.

[0155] Next, the operations will be described. The communication devices 2200-1 to 2200-M transmit the reference signal 1905-2 and the learning signal 1905-1 required for learning to the learning device 2203 via the respective wireless links 2201-1 to 2201-M and through the access point 2202. The step learning unit 2206 of the learning device 2203 learns using the signals 2204 received from one or more of the communication devices 2200-1 to 2200-M, namely, the reference signal 1905-2 and the learning signal 1905-1, and calculates the common step size 1905-3 for the communication devices 2200-1 to 2200-M. The learning device 2203 outputs the calculated step size 1905-3 as a signal 2205 and transmits it to the communication devices 2200-1 to 2200-M via the access point 2202 and through the wireless links 2201-1 to 2201-M.

[0156] As described above, the step learning unit 2206 obtains the learning signal 1905-1, i.e., the pre-equalization signal 1702, and the reference signal 1905-2 from the plurality of communication devices 2200-1 to 2200-M and learns the step size 1905-3. The step learning unit 2206 learns using the data of the plurality of communication devices 2200-1 to 2200-M. Thus, it is possible to efficiently collect and utilize a plurality of data required for learning, and the learning accuracy can be improved.

[0157] Embodiment 12

[0158] In Embodiment 12, taking the communication device 100-1 of Embodiment 1 as an example, the operations of the communication device 100-1 will be described using a flowchart. In addition, the hardware structure of the communication device 100-1 will be described.

[0159] Figure 23 It is a flowchart showing the operations of the communication device 100-1 of Embodiment 12. In the communication device 100-1, the pre-equalization processing unit 301 of the reception processing unit 203 performs pre-equalization processing on the received signal 204 (step S1). In the equalization processing unit 303, the step learning unit 404 learns the step size 405 using the reference signal 307 and the learning signal 407-3 (step S2). The tap coefficient adjustment unit 402 adjusts the tap coefficient 403 using the reference signal 307, the step size 405, and the tap adjustment signal 407-2 (step S3). The linear equalization unit 401 performs linear equalization on the equalization unit input signal 407-1 using the tap coefficient 403 (step S4) and outputs the equalized signal 304. The post-equalization processing unit 305 of the reception processing unit 203 performs post-equalization processing on the equalized signal 304 (step S5).

[0160] Next, the hardware structure of the communication device 100-1 will be described. In the communication device 100-1, the transceiver processing unit 101 and the control unit 102 are implemented by a processing circuit. The processing circuit may be a processor and a memory that execute a program stored in the memory, or may be dedicated hardware. The processing circuit is also referred to as a control circuit.

[0161] Figure 24 FIG. is a diagram showing a structural example of a processing circuit 90 in a case where the processing circuit of the communication device 100-1 that implements Embodiment 12 is implemented by a processor 91 and a memory 92. Figure 24 The processing circuit 90 shown is a control circuit and includes a processor 91 and a memory 92. When the processing circuit 90 is constituted by the processor 91 and the memory 92, each function of the processing circuit 90 is implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in the memory 92. In the processing circuit 90, the processor 91 reads and executes the program stored in the memory 92, thereby implementing each function. That is, the processing circuit 90 has a memory 92 that stores a program whose result is to execute the processing of the communication device 100-1. This program can also be said to be a program for causing the communication device 100-1 to execute each function implemented by the processing circuit 90. This program can be provided by a storage medium storing the program or by other means such as a communication medium.

[0162] The above program can also be said to be the following program, which causes the communication device 100-1 to perform the following steps: a linear equalization step, in which the linear equalization unit 401 performs linear equalization on the received signal, i.e., the equalization unit input signal 407-1; a tap coefficient adjustment step, in which the tap coefficient adjustment unit 402 adjusts the tap coefficient 403 used in the linear equalization according to the step size 405; and a step size learning step, in which the step size learning unit 404 performs learning of the step size 405. In the step size learning step, the program causes the communication device 100-1 to perform the following steps: an operation step, in which the NN layers 704-1 to 704-K respectively calculate the updated tap coefficient based on the specified initial tap coefficient 703 or the updated tap coefficient output from the previous-stage NN layer 704-k, i.e., the signal 706-k, the learning signal 407-3 included in the layer data set 702-k, i.e., the pre-equalization signal 302, and the reference signal 307 which is a specified signal sequence included in the layer data set 702-k, and maintain the internal parameters 810 used in the calculation; a learning step, in which the learning processing unit 708 sets the error function during learning to the mean square error between the tap coefficient based on the least squares solution calculated from the learning signal 407-3, i.e., the pre-equalization signal 302, and the reference signal 307, and the updated tap coefficient output from the last stage of the NN layers 704-1 to 704-K, i.e., the NN output 706-K, and updates the internal parameters 810; and an update step, in which the internal parameter collection unit 700 updates the step size 405 based on the internal parameters 810 collected from the NN layers 704-1 to 704-K, i.e., the layer step sizes 707-1 to 707-K.

[0163] Here, the processor 91 is, for example, a CPU (Central Processing Unit), a processing device, an arithmetic device, a microprocessor, a microcomputer, or a DSP (Digital Signal Processor). In addition, the memory 92 is, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable ROM), an EEPROM (registered trademark) (Electrically EPROM), a magnetic disk, a floppy disk, an optical disk, a high-density disk, a mini disk, or a DVD (Digital Versatile Disc).

[0164] Figure 25This is a diagram showing an example of the processing circuit 93 in the case where the processing circuit of the communication device 100-1 implementing Embodiment 12 is constituted by dedicated hardware. Figure 25 The illustrated processing circuit 93 is, for example, a single circuit, a composite circuit, a programmed processor, a parallel-programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a component formed by combining them. Regarding the processing circuit, part of it can be implemented using dedicated hardware, and part of it can be implemented using software or firmware. In this way, the processing circuit can implement the above-described various functions through dedicated hardware, software, firmware, or a combination thereof.

[0165] The structures shown in the above embodiments are examples, and can be combined with other known technologies, can combine the embodiments with each other, and can also omit or change part of the structure without departing from the gist.

[0166] Reference Signs Explanation

[0167] 100-1, 100-2, 1500-1, 1500-2, 1900-1, 1900-2, 2200-1 to 2200-M: Communication devices; 101, 1502, 1902: Transceiving processing units; 102, 1501, 1901: Control units; 201, 1601: Transmission processing units; 202, 1603: Transmission signals; 203, 1602: Reception processing units; 204, 1604: Reception signals; 301, 1000, 1701: Pre-equalization processing units; 302, 1001, 1702: Pre-equalization signals; 303, 1002, 1703: Equalization processing units; 304, 1003, 1704: Post-equalization signals; 305, 1004, 1705: Post-equalization processing units; 306, 1006, 1706: Reference signal generation units; 307, 1007, 1504-2, 1905-2: Reference signals; 401, 1101, 1801, 2001: Linear equalization units; 402, 1102, 1802, 2002: Tap coefficient adjustment units; 403, 504-1 to 504-L, 604-1 to 604-L, 1103, 1406-1 to 1406-2L, 1803, 2003: Tap coefficients; 404, 1104, 1505, 1908, 2206: Step size learning units; 405, 1105, 1504-3, 1905-3: Step sizes; 406, 1804, 2004: Input destination control units; 407-1, 1805-1, 2005-1: Equalization unit input signals; 407-2, 1107-2, 1805-2, 2005-2: Tap adjustment signals; 407-3, 1107-3, 1504-1, 1905-1, 2005-3: Learning signals; 500-1 to 500-L-1, 505-1 to 505-L, 600-1 to 600-L-1, 605-1 to 605-L, 607, 609, 706-1 to 706-K-1, 801, 803, 805, 807, 907-1 to 907-K-1, 908-1 to 908-K-1, 909-1 to 909-K-1, 1401-1 to 1401-L-1, 1403-1 to 1403-2L, 1407-1 to 1407-2L, 1906, 1907, 2101, 2105, 2107, 2204, 2205: Signals; 501-1 to 501-L-1, 601-1 to 601-L-1, 1400-1 to 1400-L-1: Delay elements; 502-1 to 502-L, 602-1 to 602-L, 804, 806, 1404-1 to 1404-2L, 2106: Multipliers; 503, 603, 1405: Coefficient distributors; 506, 606, 808, 1408, 2108: Adders; 608, 802, 2102: Subtractors; 610: Tap coefficient update unit;611, 1203: Tap coefficients after update; 700, 900: Internal parameter collection unit; 701, 901: NN control unit; 702-1 to 702-K, 902-1 to 902-K: Layer data sets; 703, 903: Initial tap coefficients; 704-1 to 704-K, 904-1 to 904-K: NN layers; 705, 905: Deep NN unit; 706-K, 907-K: NN outputs; 707-1 to 707-K, 913-1 to 913-K: Step sizes for each layer; 708, 910: Learning processing unit; 709-1 to 709-K, 911-1 to 911-K: Updated parameters; 714, 914: Target data set; 800, 2100: Inner product calculator; 809-1 to 809-K, 1300-1 to 1300-K, 2109-1 to 2109-K: Internal parameter holding units; 810, 1310, 2110: Internal parameters; 906-1 to 906-K-1: Switches; 912-1 to 912-K-1: SW control signals; 1005: Posterior information; 1106: Input signal save memory; 1200: A priori estimation error calculation unit; 1201: Tap update processing unit; 1202: Estimation error; 1301-1 to 1301-K: Linear operation processing units; 1402-1 to 1402-L: Signal distributors; 1503, 1905, 2203: Learning devices; 1510, 1910, 2210: Communication systems; 1903, 2202: Access points; 1904, 2201-1 to 2201-M: Wireless links; 2006: Learning signal generation unit.;

Claims

1. A communication device, characterized in that, the communication device has: a linear equalization unit that performs linear equalization on a received signal; a tap coefficient adjustment unit that adjusts tap coefficients used in the linear equalization according to a step size; and a step size learning unit that performs learning of the step size, wherein the step size learning unit has: a plurality of neural network layers that respectively calculate updated tap coefficients based on prescribed initial tap coefficients or updated tap coefficients output from a previous stage, the received signal, and a reference signal that is a prescribed signal sequence, and hold internal parameters used in the calculation; a learning processing unit that sets an error function in the learning as a mean square error between tap coefficients based on a least squares solution calculated from the received signal and the reference signal and updated tap coefficients output from a last stage of the plurality of neural network layers, performs learning, and updates the internal parameters; and an internal parameter collection unit that updates the step size based on the internal parameters collected from the plurality of neural network layers.

2. The communication device according to claim 1, characterized in that, the step size learning unit performs learning of the step size by incremental learning.

3. The communication device according to claim 1 or 2, characterized in that, the communication device has: an equalization post-processing unit that performs equalization post-processing on an equalized signal obtained by the linear equalization unit performing linear equalization on the received signal; and a reference signal generation unit that generates the reference signal based on posterior information obtained through the equalization post-processing.

4. The communication device according to any one of claims 1 to 3, characterized in that, the linear equalization unit performs generalized linear processing as the linear equalization.

5. A learning device, characterized in that, the learning device has a step size learning unit that performs learning of a step size used in a communication device, the communication device adjusts tap coefficients used in linear equalization according to the step size, and performs the linear equalization on a received signal, wherein the step size learning unit has: a plurality of neural network layers that respectively calculate updated tap coefficients based on prescribed initial tap coefficients or updated tap coefficients output from a previous stage, the received signal, and a reference signal that is a prescribed signal sequence, and hold internal parameters used in the calculation; a learning processing unit that sets an error function in the learning as a mean square error between tap coefficients based on a least squares solution calculated from the received signal and the reference signal and updated tap coefficients output from a last stage of the plurality of neural network layers, performs learning, and updates the internal parameters; and an internal parameter collection unit that updates the step size based on the internal parameters collected from the plurality of neural network layers.

6. The learning device according to claim 5, characterized in that, communication with the communication device is performed through wireless communication via a wireless communication network.

7. The learning device according to claim 6, characterized in that, The step learning unit obtains the received signal and the reference signal from a plurality of the communication devices to perform learning of the step size.

8. The learning device according to claim 6 or 7, wherein, the communication device calculates a correlation matrix and a correlation vector based on the received signal and the reference signal, and sends the correlation matrix and the correlation vector to the learning device, and the step size learning unit uses the correlation matrix and the correlation vector to perform learning of the step size.

9. A communication system, wherein, the communication system includes: the learning device according to claim 5; and a communication device that performs equalization processing using the step size obtained from the learning device.

10. A control circuit for controlling a communication device, wherein, the control circuit causes the communication device to perform the following processing: performing linear equalization on a received signal; adjusting tap coefficients used in the linear equalization according to a step size; and performing learning of the step size, in the learning of the step size, the control circuit causes the communication device to perform the following processing: in a plurality of neural network layers, respectively calculating updated tap coefficients based on a prescribed initial tap coefficient or an updated tap coefficient output from a previous stage, the received signal, and a reference signal that is a prescribed signal sequence, and holding internal parameters used in the calculation; setting an error function in the learning to be a mean square error between a tap coefficient based on a least squares solution calculated from the received signal and the reference signal and an updated tap coefficient output from the last stage of the plurality of neural network layers, and updating the internal parameters; and updating the step size according to the internal parameters collected from the plurality of neural network layers.

11. A storage medium storing a program for controlling a communication device, wherein, the program causes the communication device to perform the following processing: performing linear equalization on a received signal; adjusting tap coefficients used in the linear equalization according to a step size; and performing learning of the step size, in the learning of the step size, the program causes the communication device to perform the following processing: in a plurality of neural network layers, respectively calculating updated tap coefficients based on a prescribed initial tap coefficient or an updated tap coefficient output from a previous stage, the received signal, and a reference signal that is a prescribed signal sequence, and holding internal parameters used in the calculation; setting an error function in the learning to be a mean square error between a tap coefficient based on a least squares solution calculated from the received signal and the reference signal and an updated tap coefficient output from the last stage of the plurality of neural network layers, and updating the internal parameters; and updating the step size according to the internal parameters collected from the plurality of neural network layers.

12. A step size update method, wherein, the step size update method includes the following steps: a linear equalization step in which a linear equalization unit performs linear equalization on a received signal; Tap coefficient adjustment step, in which a tap coefficient adjustment unit adjusts the tap coefficients used in the linear equalization according to a step size; And Step size learning step, in which a step size learning unit performs learning of the step size, The step size learning step includes the following steps: Operation step, in which multiple neural network layers respectively calculate updated tap coefficients based on the specified initial tap coefficients or the updated tap coefficients output from the previous stage, the received signal, and a reference signal that is a specified signal sequence, and retain the internal parameters used in the calculation; Learning step, in which a learning processing unit performs learning with the error function in the learning set as the mean square error between the tap coefficients based on the least squares solution calculated from the received signal and the reference signal and the updated tap coefficients output from the last stage of the multiple neural network layers, and updates the internal parameters; And Update step, in which an internal parameter collection unit updates the step size according to the internal parameters collected from the multiple neural network layers.

Images