Underwater Wireless Optical Communication Method Based on I-SC-FDM Technology and Sparse Weight Loading DNN

Abstract

The present invention discloses an underwater wireless optical communication method based on I-SC-FDM technology and sparse weight loading DNN, which includes a transmitting end and a receiving end. The transmitting end uses interleaved single-carrier frequency division multiplexing technology to greatly reduce the complexity and maximum peak-to-average power ratio required by traditional orthogonal frequency division multiplexing. The receiving end uses a specially designed sparse weight pre-loading neural network equalizer to overcome the problem of slow convergence of traditional neural networks with random weight initialization, and at the same time further improves the system bit error performance. The present invention can effectively improve the communication capacity of the underwater wireless optical communication system. The low-complexity transmitting end and the fast-converging receiving end are conducive to the actual system hardware deployment, and further improve the practicability of the communication system.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater wireless optical communication, and particularly to an underwater wireless optical communication method based on I-SC-FDM technology and sparse weight loading DNN. Background Technique

[0002] In recent years, underwater wireless optical communication technology has received extensive attention due to its characteristics such as high speed, low latency, low cost, high flexibility and security. Due to the existence of the underwater low-attenuation optical window of 400-550nm, the use of blue-green lasers to achieve long-distance high-bandwidth underwater optical communication has gradually become a research hotspot. At the same time, in order to make full use of the limited bandwidth of the system to achieve high-speed transmission, the orthogonal frequency division multiplexing (OFDM) technology with high spectral efficiency and inter-symbol interference resistance has been gradually introduced into the underwater wireless optical communication system. However, the traditional OFDM modulation and demodulation technology is relatively complex, which is not conducive to the hardware deployment of the actual system, and its high peak-to-average power ratio (PAPR) restricts the amplification ability of the signal during transmission, which is not conducive to long-distance communication, and even causes more serious nonlinear effects, further deteriorating the communication performance. Facing the above problems, there is an urgent need for an easy-to-implement technology with lower complexity, smaller PAPR and improved performance of the underwater wireless optical communication system. The joint signal processing technology based on interleaved single-carrier frequency division multiplexing (I-SC-FDM) and sparse weight loading deep neural network (DNN) proposed by the present invention can better solve the above difficulties. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: by introducing I-SC-FDM technology, the complexity and PAPR of the transmitter of the traditional OFDM technology are reduced to meet the simplification requirements of the actual deployment of the underwater wireless optical communication system. At the same time, sparse weight loading DNN is used to achieve signal equalization and demodulation at the receiver. Compared with the DNN with random weight initialization, the introduction of sparse weights improves the convergence speed of the equalizer and can obtain better bit error performance than traditional equalization methods. The joint application of I-SC-FDM technology and sparse weight loading DNN reduces the system complexity while improving the performance of the communication system, facilitating the hardware deployment of the actual system.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is:

[0005] An underwater wireless optical communication method based on I-SC-FDM technology and sparse weight loading DNN specifically includes the following steps:

[0006] Step 1, build a transmitter of an underwater wireless optical communication system based on I-SC-FDM technology.

[0007] 1.1. First, the binary sequence is first subjected to quadrature amplitude modulation (QAM), and then a complex signal is generated through discrete Fourier transform spread (DFT-spread), and the expression is:

[0008]

[0009] In the above formula, x(k) is the QAM signal, X(n) is the signal after DFT-spread, and N is the number of DFT points.

[0010] 1.2. Then, X(n) is interleaved and assigned to the odd positions of the 2N-point IDFT, which can be expressed as:

[0011] Y = [0, X(0), 0, X(1),..., 0, X(N - 1)] T

[0012] 1.3. Then, it undergoes a 2N-point IDFT transformation, which can be expressed as:

[0013]

[0014] Noting the conjugate symmetry of the discrete Fourier transform pair, it can be further obtained that:

[0015]

[0016] It can be seen from the above formula that the first half of the complex interleaved FDM signal y(m) obtained after N-point DFT, interleaved assignment, and 2N-point IDFT can be simply regarded as a phase rotation transformation of the QAM signal x(m), and the latter part can be simply obtained by taking the opposite of the first half. Therefore, the Nlog2N complex multiplications required for the original N-point DFT and 2N-point IDFT can be simplified to N complex multiplications, saving computing resources and being more conducive to hardware deployment. Based on this, the final real-valued I-SC-FDM can be realized by continuously transmitting the real part and the imaginary part of y(m), and can be expressed as:

[0017]

[0018] In the above formula,

[0019]

[0020] It can be found that the real-valued I-SC-FDM signal has a cosine / sine envelope characteristic, so it has a lower PAPR than the traditional discrete Fourier transform spread OFDM (DFT-s OFDM, also known as L-SC-FDM) with a Gaussian envelope characteristic, is more conducive to the amplification of the signal at the transmitter, and has stronger anti-nonlinear characteristics.

[0021] Step 2, build the receiver of an underwater wireless optical communication system based on sparse weight loading DNN.

[0022] 2.1, Based on the simplicity of I-SC-FDM modulation, if the signal remains unchanged after passing through the channel, then the ideal receiver signal processing and transmitter signal processing should be inverse processes. Under such an assumption, the received signal first needs to undergo a real-to-complex transformation, which can be expressed as:

[0023] P rx =[I N×N |jI N×N N×2N ·S rx

[0024] In the above formula, S rx is the received I-SC-FDM signal. In the above assumption, it is consistent with the transmitted signal S tx . I N×N is the N-order identity matrix. Then P rx needs to undergo an anti-symmetric transformation, which can be expressed as:

[0025] Q rx =[I N×N |-I N×N T 2N×N ·P rx

[0026] 2.2, Next, the 2N-point DFT and N-point IDFT are respectively applied to Q rx to generate the QAM signal. Using the simplification process at the transmitter, this transformation can be expressed as:

[0027] R rx =[E N×N |-E N×N N×2N ·Q rx

[0028] In the above formula,

[0029]

[0030] The obtained R rx is then subjected to QAM demapping to obtain the restored binary sequence. Combining the above formulas, the transmission matrix of the receiver can be obtained, expressed as:

[0031]

[0032] ​​​As can be seen from the above formula, ideally, the received signal can obtain the corresponding QAM symbol by passing through only one transmission matrix. However, in an actual system, the received signal is severely interfered by the channel, so that channel estimation and signal equalization are indispensable steps.

[0033] 2.3. Based on this, considering the good fitting ability of the neural network, this transmission matrix can be used for the weight initialization of the DNN equalizer, thus forming a joint equalization-demodulation scheme. First, construct an extended transmission equation based on the above formula:

[0034] R tx =W N ·W N-1 ·...·W1·S rx

[0035] In the above formula, W N is equal to T rx . W i (i < N) is a sparse matrix centered on the identity matrix and can be expressed as:

[0036]

[0037] In the above formula, 0 is the zero matrix. The size of the W i matrix can be freely adjusted according to requirements, which introduces more coefficient weight spaces. Obviously, by adding an activation function and a bias, the extended transmission equation can be further transformed into a DNN model, which can be expressed as:

[0038]

[0039] In the above formula, f i is the activation function, and b i is the bias vector. It can be noted that different from the DNN with general random weight initialization, the above DNN model inserts a large number of preset sparse weights, thus constituting the sparse weight loading DNN equalizer required at the receiving end.

[0040] To verify the effectiveness of this method, a wireless optical communication experiment was carried out under the condition of a 7m water tank. The performance advantages of I-SC-FDM compared with traditional OFDM and DFT-spread OFDM under different received optical powers and different transmission rates were verified respectively. At the same time, the advantage of the sparse weight loading DNN equalizer in performance improvement was also verified. Generally speaking, the effective combination of I-SC-FDM and the sparse weight loading DNN effectively reduces the system complexity and the signal PAPR, and at the same time further improves the performance of the communication system, making it more suitable for the hardware deployment of the actual underwater wireless optical communication system.

[0041] The beneficial effects of the present invention are:

[0042] 1. The present invention is compatible with all current underwater wireless optical communication systems. At the receiving end, only the traditional digital signal processing process needs to be replaced with a sparse weight loading DNN equalizer.

[0043] 2. The transmitting end of the present invention uses the I-SC-FDM technology to achieve lower complexity and PAPR, making the entire communication system simpler, with less computational difficulty and more convenient actual hardware deployment.

[0044] 3. The receiving end of the present invention uses a sparse weight loading DNN equalizer to replace multiple original digital signal processing modules in the traditional system. Essentially, it considers the optimal solution of the system and also has the excellent non-linear fitting ability of DNN, which can achieve better bit error performance.

[0045] 4. The sparse weight loading scheme introduced by the DNN equalizer at the receiving end of the present invention further improves its convergence speed, which is beneficial to the rapid training and deployment of the actual system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a structural schematic diagram of the present invention;

[0047] Figure 2 is a comparison diagram of PAPR of different modulation schemes at the transmitting end of the present invention;

[0048] Figure 3 is a structural diagram of the sparse weight loading DNN equalizer at the receiving end of the present invention;

[0049] Figure 4 is an experimental result diagram of the present invention;

[0050] Figure 5 is a comparison diagram of DNN convergence speed of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0051] The following describes in detail the specific implementation manners of the present invention with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0052] Embodiment 1, an underwater wireless optical communication method based on I-SC-FDM technology and sparse weight loading DNN, specifically including the following steps:

[0053] Step 1, as shown on the left, build the transmitting end of an underwater wireless optical communication system based on I-SC-FDM technology. The binary sequence first undergoes quadrature amplitude modulation (QAM), and then a complex signal is generated through discrete Fourier transform spread (DFT-spread), which can be expressed as: Figure 1 The binary sequence first undergoes quadrature amplitude modulation (QAM), and then a complex signal is generated through discrete Fourier transform spread (DFT-spread), which can be expressed as:

[0054]

[0055] In the above formula, x(k) is the QAM signal, X(n) is the signal after DFT-spread, and N is the number of DFT points. Then X(n) is interleaved and assigned to the odd positions of the 2N-point IDFT, which can be expressed as:

[0056] Y = [0, X(0), 0, X(1),..., 0, X(N - 1)] T

[0057] Then, through the 2N-point IDFT transformation, it can be expressed as:

[0058]

[0059] Noting the conjugate symmetry of the discrete Fourier transform pair, it can be further obtained that:

[0060]

[0061] It can be seen from the above formula that the first half of the complex interleaved FDM signal y(m) obtained after N-point DFT, interleaved assignment, and 2N-point IDFT can be simply regarded as a phase rotation transformation of the QAM signal x(m), and the latter part can be simply obtained by taking the opposite of the first half. Therefore, the Nlog2N complex multiplications required for the original N-point DFT and 2N-point IDFT can be simplified to N complex multiplications, that is Figure 1 The "phase rotation" module shown at the transmitter saves computing resources and is more conducive to hardware deployment. Based on this, the final real-valued I-SC-FDM can be realized by continuously transmitting the real and imaginary parts of y(m), which can be expressed as:

[0062] S tx = [y re (0),..., y re (m), |y im (0),..., y im (m)] T 2N×1

[0063] In the above formula,

[0064]

[0065] It can be found that the real-valued I-SC-FDM signal has a cosine / sine envelope characteristic. Therefore, it has a lower PAPR than the traditional discrete Fourier transform spread OFDM (DFT-s OFDM, also known as L-SC-FDM) with a Gaussian envelope characteristic, is more conducive to the amplification of the signal at the transmitter, and has stronger anti-nonlinear characteristics. Specifically, Figure 2Shows the PAPR distribution of traditional OFDM, DFT-spread OFDM, and I-SC-FDM. It can be seen that the complementary cumulative distribution function of I-SC-FDM is smaller than that of the other two modulation formats.

[0066] Next, the real-valued I-SC-FDM signal successively passes through a signal generator, an amplifier, a bias, and is then applied to a laser to achieve electro-optic conversion, and is received by a detector at the receiving end to achieve opto-electric conversion, and then captured by an oscilloscope.

[0067] Step 2, as Figure 1 shown on the right, build the receiving end of an underwater wireless optical communication system based on a sparse weight loading DNN equalizer. Due to the simplicity of I-SC-FDM modulation, if the signal remains unchanged after passing through the channel, then the ideal receiving-end signal processing and transmitting-end signal processing should be inverse processes. Under such an assumption, the received signal first needs to undergo a real-to-complex transformation, which can be expressed as:

[0068] P rx =[I N×N |jI N×N N×2N ·S rx

[0069] In the above formula, S rx is the received I-SC-FDM signal. In the above assumption, it is consistent with the transmitted signal S tx . I N×N is the N-order identity matrix. Then P rx needs to undergo an anti-symmetric transformation, which can be expressed as:

[0070] Q rx =[I N×N |-I N×N T 2N×N ·P rx

[0071] Next, the 2N-point DFT and the N-point IDFT act on Q rx respectively to generate a QAM signal. Using the simplification process at the transmitting end, this transformation can be expressed as:

[0072] R rx =[E N×N |-E N×N N×2N ·Q rx

[0073] In the above formula,

[0074]

[0075] The obtained R above​​​rx After QAM inverse mapping, the restored binary sequence can be obtained. Combining the above equations, the transmission matrix at the receiving end can be obtained and expressed as:

[0076]

[0077] It can be seen from the above equation that in the ideal case, the received signal only needs to pass through a transmission matrix to obtain the corresponding QAM symbol. However, in an actual system, the received signal will be severely interfered by the channel, so that channel estimation and signal equalization are indispensable steps. Based on this, considering the good fitting ability of the neural network, this transmission matrix can be used for the weight initialization of the DNN equalizer, thus forming a joint equalization-demodulation scheme. First, an extended transmission equation is constructed based on the above equation:

[0078] R tx =W N ·W N-1 ·...·W1·S rx

[0079] In the above equation, W N is equal to T rx . W i (i < N) is a sparse matrix centered on the identity matrix and can be expressed as:

[0080]

[0081] In the above equation, 0 is the zero matrix. The size of the W i matrix can be freely adjusted according to requirements, and it introduces more coefficient weight spaces. Obviously, by adding activation functions and biases respectively, the extended transmission equation can be further transformed into a DNN model, which can be expressed as:

[0082]

[0083] In the above equation, f i is the activation function, and b i is the bias vector. It can be noted that different from the DNN with general random weight initialization, the above DNN model inserts a large number of preset sparse weights, thus constituting the sparse weight loading DNN equalizer required at the receiving end. As Figure 3 shown, in actual implementation, a sparse weight loading DNN with two hidden layers is used to achieve equalization and demodulation at the receiving end. The non-linear connection of the first hidden layer uses the "tanh" activation function, and the rest use linear connections.

[0084] As Figure 4As shown, in the actual sink experiment verification, the received optical power at the receiving end was set to -25 dBm, -30 dBm, and -35 dBm respectively. It can be seen that I-SC-FDM has the best bit error rate performance. Specifically, when the received optical power is the relatively high -25 dBm, the bit error is mainly caused by the remaining non-linear damage. When the baud rate is relatively low, the performance of I-SC-FDM is the same as that of DFT-spread OFDM. However, as the received optical power further decreases and the baud rate further increases, the total signal-to-noise ratio decreases. At this time, I-SC-FDM with a higher PAPR has better bit error performance. At the bit error rate threshold of 3.8×10 -3 When the received optical power is -25 dBm, the maximum baud rate that I-SC-FDM can achieve is increased by 57.1% compared with traditional OFDM and by 22.2% compared with DFT-spread OFDM. When the received optical power is -30 dBm, the maximum baud rate that I-SC-FDM can achieve is increased by 42.9% compared with DFT-spread OFDM.

[0085] As Figure 5 shown, it shows the convergence of the DNN equalizer with and without sparse weight loading at baud rates of 150 MBd and 250 MBd. Specifically, when the received optical power is the relatively high -25 dBm and the baud rate is 150 MBd, considering the bit error rate threshold of 3.8×10 -3 the number of iterations required for the DNN with sparse weights added is only 11.1% of that without DNN added. When the baud rate is 250 MBd, the number of iterations required is still only 16.2%. When the received optical power is -30 dBm, in the case of 150 MBd, the number of iterations required for the DNN with sparse weights added is only 10.3% of that without DNN added. For 250 MBd, considering the number of iterations required to converge to a stable state, the number of iterations required for the DNN with sparse weights added is 80.3% of that without DNN added. The convergence situation is similar when the received optical power is -35 dBm. Generally speaking, a low signal-to-noise ratio makes it difficult for a DNN with traditional random weight initialization to quickly learn the characteristics of the signal, while the DNN with sparse weight loading pre-sets relatively correct weights at the beginning, clarifies the convergence direction, makes the convergence speed faster, and effectively saves the computing time and resources required for actual system deployment.

Claims

1. An underwater wireless optical communication method based on I-SC-FDM technology and sparse weight loading DNN, characterized in that it specifically includes the following steps: Step 1, build the transmitter of an underwater wireless optical communication system based on I-SC-FDM technology; 1.1, First, the binary sequence undergoes quadrature amplitude modulation, and then a complex signal is generated through discrete Fourier expansion. The expression is: In the above formula, x(k) is the QAM signal, X(n) is the signal after DFT-spread, and N is the number of DFT points; 1.2, Then X(n) is interleaved and assigned to the odd positions of the 2N-point IDFT. The expression is: Y = [0, X(0), 0, X(1),..., 0, X(N - 1)] T (2) 1.3, and then through the 2N-point IDFT transform, the expression is: Through the conjugate symmetry of the discrete Fourier transform, it is further obtained that: The first half of the complex interleaved FDM signal y(m) obtained after N-point DFT, interleaved allocation, and 2N-point IDFT is regarded as the phase rotation transformation of the QAM signal x(m), and the second half is obtained by taking the opposite of the first half; Therefore, the Nlog2N complex multiplications required for the original N-point DFT and 2N-point IDFT are simplified to N complex multiplications. The final real-valued I-SC-FDM is realized by continuously transmitting the real and imaginary parts of y(m), and the expression is: S tx = [y re (0),..., y re (m), |y im (0),..., y im (m)] T 2N×1 (5) In the above formula, where x re (m) is the real part of x(m), and x im (m) is the imaginary part of x(m), y re (m) is the real part of y(m), and y im (m) is the imaginary part of y(m); Step 2, build the receiving end of the underwater wireless optical communication system based on the sparse weight loading DNN; 2.1, the received signal first undergoes a real-to-complex transformation, and the expression is: P rx = [I N×N |jI N×N N×2N ·S rx (7)​ In the above formula, S rx is the received interleaved SCFDM signal, which, in the above assumption, is identical to the transmitted signal S tx ; I N×N is the N-order identity matrix, and then P rx is subjected to an anti-symmetric transformation and is expressed as: Q rx = [I N×N |-I N×N T 2N×N ·P rx (8)​ 2.2, the 2N-point DFT and the N-point IDFT respectively act on Q rx to generate a QAM signal. Using the simplified process at the transmitter, this transformation is expressed as: R rx = [E N×N |-E N×N N×2N ·Q rx (9)​ In the above formula, The obtained R above rx Then, through QAM inverse mapping, the restored binary sequence is obtained, and finally the transmission matrix at the receiving end is obtained. The expression is: 2.3, construct an extended transmission equation: R tx = W N ·W N-1 ·...·W1·S rx (12) In the above formula, W N is equal to T rx , and W i (i < N) is a sparse matrix with the identity matrix as the center, expressed as: In the above formula, 0 is the zero matrix, and W i The size of the matrix can be freely adjusted; by adding an activation function and a bias, the extended transfer equation is converted into a DNN model, expressed as: In the above formula, f i is the activation function, and b i is the bias vector.

Images