Underwater optical wireless communication system: Deep learning CNN with NOMA-based performance analysis

This research is looking forward improving the performance for underwater optical wireless communication (UOWC) by applying a Non-orthogonal multiple access (NOMA) technique. We also get the benefit of the advantage the transmission based on convolutional neural network hybrid with a long short-term memory cell. The relays selection and power optimization are two main parameters to enhance the UOWC system performance. In this work, we suppose a pairing method for NOMA nodes. By replacing the inner dense connections with convolution layers, this model is proposed to overcome high complexity and over fitting to improve the model performance. The obtained performance for sum rates show that NOMA outperforms the orthogonal multiple access system by ~ 6%. Applying a step-by-step sub-optimization algorithm (SSOPA) yields better results than using fixed power allocation (FPA), while using a global optimal power allocation algorithm (GOPA) increases the sum rates over both FPA and SSOPA. It is found that the improvement when using GOPA combined with CNN approach enhances the performance of sum rates by ~ 2.5% than using the independent-relay-aided NOMA (ICNOMA) for UOWC. The GOPA improvement is 1.2%, 2.5%, 8.7% over FPA and is 0.12%, 0.34%, 2.09% over SSOPA, for clear, pure, and coastal water, respectively. The ICNOMA outperforms both ordinary NOMA (ONOMA) and cooperative NOMA (CNOMA) without independent relay nodes. The ICNOMA achieves an improvement over ONOMA and CNOMA by 20.4% and 3.2%, respectively.

1. A CNNLSTM model is utilized to reallocate the power for a relay aided NOMA-UOWC network. Moreover, this increases the efficiency of spectrum use for underwater optical wireless communication (UOWC). 2. The introduction of several situations, including ICNOMA sum rates versus transmitted power and FPA, SSOPA, and GOPA sum rates depending on various values of the scintillation coefficient ( 2 Sci ).
3. The sum rates versus depth of transmitter for ICNOMA the FPA, SSOPA, and GOPA algorithms are presented in clear ocean water, pure ocean water, and coastal water. 4. The sum rates for four users (1 and 2, 1 and 3, and 1 and 4) are determined for ONOMA, CNOMA, and ICNOMA. 5. Our proposed CNNLSTM model is applied to solve the complexity, over fitting problem and reduce the computational time, hence enhancing the performance of our framework.
The remainder of the paper is structured as follows. The system model used is explained in Sect. 2. Section 3 demonstrates the power allocation algorithms applied in this paper. The methodologies used are listed in Sect. 4. The obtained results and analysis based on simulation and assessment parameters are displayed and discussed in Sect. 5. The findings are concluded in Sect. 6, which also lists suggestions for a future work.

System model
The communication range constrains the deployment of nodes in UOWC while for long range between nodes gets unstable links of optical communication. We assume the transmission of UOWC depending on the AF relay, where main advantage for our system is easy implementation and low complexity. Figure 1 shows the architecture of the NOMA in UOWC system with nodes of receiving N simultaneously with relaying node M (Liang et al. 2022).
Currently, the studies depend on NOMA either for two users or more than one user can be achieved via adding a new user to the group, where the performance of multi-user paired NOMA depends on the channel quality of the supplementary users, according to the NOMA application. Also, miscommunication probability rise when increasing number of users.

Sum rate-based power optimization
In this selection, we discuss the power allocation used to increase the energy efficiency. The main reason for using the power allocation algorithms is the high cost of the secondary replacement of node batteries. Three different power allocation algorithms; FPA, GOPA, and SSOPA are combined with CNN approach to enhance the performance of sum rates.
The mechanism of using FPA algorithm determines the percentage ratio for all pairing users NOMA (Randrianantenaina et al. 2020;Liang et al. 2022). The FPA algorithm  (Liang et al. 2022) 436 Page 4 of 14 allocates a specific transmitted power for each user. This method is very simple and easily implemented. For getting implementation easily, we use simple algorithms.
The main parameter for the optimization solution is the step ratio configuration, where it is preferred to get this parameter in average value not high or low. The mean distribution method is used for power allocation within the source node.
For GOPA algorithm, the transmitted power is constant and data rate is maximized using optimization objective (Ghosh et al. 2022;Bi et at. 2022;Song et al. 2021). The total transmission power consists of the transmitted power of relay nodes and transmitter. There is a similar transmitted power for NOMA group.
Although, the complexity increases when increasing the number of nodes, GOPA is considered the optimal method via traversing the overall system. The SSOPA has a lower complexity computation and can optimize the allocation of the power between receiving nodes and relays by determining the used parameters professionally.
In this paper, we use the coefficient of the power allocation in Liang et al. (2022) to guarantee achieving a quality of service (QoS) of the user in NOMA scheme and to compare between our results in this search and that found in Liang et al. (2022). In particular, the channel state information (CSI) between the transmitter and the relays, and between the relays and receiving nodes is comprehensively considered.

Dataset
Many algorithms are used to generate the datasets; the sum rates are the common factors for these relations. To use CNN, one should prepare several datasets to utilize this model in an accurate way. These datasets used in the DL are existing from (Liang et al. 2022), where it is considered a main reference paper for algorithms and to compare the results.
The data sets are based on relation between the sum rates and different parameters as follows: 1. Dataset with sum rates and both OMA and NOMA. 2. Dataset with sum rates and transmitted power for different algorithms as FPA, SSOPA, GOPA. 3. Dataset with sum rates and depth of transmitter for different algorithms as FPA, SSOPA, GOPA. 4. Dataset with sum rates and transmitted power for three techniques: ONMA, CNOMA, ICNOMA.

Proposed CNN hybrid with LSTM
There is an urgent need for more adaptable and flexible solutions to increase the efficiency of spectrum use for underwater optical wireless communication (UOWC). In our proposed framework, relay selection and power reallocation based on CNNLSTM are the golden keys in increasing our framework performance and decreasing complexity.
To solve various problems, pre-trained models are trained on a large benchmark dataset. For the sum rates and transmitted power, in this study, the CNNLSTM model is utilized to overcome the high complexity and over fitting, leading to improve the sum rates, spectrum efficiency. Different scenarios are introduced, OMA and NOMA sum rates versus transmitted power, FPA, SSOPA and GOPA sum rates based on the scintillation coefficient. Moreover, sum rates versus distance are proposed in clear ocean water, pure ocean water and coastal water for FPA, SSOPA and GOPA algorithms. Furthermore, ONOMA, CNOMA and ICNOMA sum rates are calculated between users (1,2), users (1, 3) and, users (1, 4) as shown in Fig. 2, which shows the location of the users as receiving nodes and the depth of the transmitters.

CNN model
The CNN may have similar parameters. Each layer that makes up a CNN has the power to turn one volume into another through a differentiable function (Alzubaidi et al. 2021;Abdeltawab et al. 2019;Abdellah and Koucheryavy 2020). Many different layers types are used in CNN. While the input layer maintains the raw input data for the image, the convolutional layer performs the dot product operation across all filters and picture patches to compute the output volume. The activation function layer subjects the output of the first layer or the convolutional layer element by element. Additionally, the pool layer is in charge of reducing the volume and improving processing performance, both of which are input to the CNN with the main objective of avoiding any kind of over fitting.

LSTM model
The data can be kept in the LSTM for a very long time. It is utilized for planning, categorization, and prediction and is based on time series data. The LSTM chain structure is made up of various memory units called cells and four neural networks. These cells are in charge of knowledge retention, whereas gates are responsible for memory change. The Forget gate allows information to be removed when it is no longer useful. The Input gate is in responsible of filling the cell with essential data. The output gate is used to mine cells for useful data. The mathematical description of the model is (Awan et al. 2020;Salama et al. 2022aSalama et al. , 2022b (1) where Z t , u t , m t and K t are the Input gate, Forget gate, Output gate, storage unit and hidden vectors, respectively, W ix , W fx , W fh , W fc , W ox , W oh , W oc are the weight matrices, b i , b f , b c , b 0 are the deviation of the LSTM obtained during the training process, and ⊙ are the sigmoid function and multiplication, respectively.

CNN hybrid with LSTM mode
In order to capture and process spatial, temporal, and external information, a unique CNN-LSTM (Kataria et al. 2019;Lu et al. 2020;Zhang et al. 2020) is employed in this research. A CNNLSTM architectural combination has an initial convolutional layer that is in charge of taking the input. This method generates an output, which is then pooled to a smaller dimension and sent into an LSTM layer. The proposed framework combines an LSTM model and a CNN hybrid. This LSTM model makes use of convolutional layers, which are useful for capturing the information. Due to simultaneous convolution processing in the suggested structure, using our approach also provides cost reduction. Dropout is also used to avoid over fitting, which enhances the performance of our model. The CNNLSTM parameters are described in Table 1, and Fig. 3.

Results and discussion
Here, we present simulation results that improve the effectiveness of different algorithms underwater. Applying the CNN approach with using the datasets of Sect. 4.1, we obtained the following results which verify that our model outperforms the results of (Liang et al. 2022), we suppose line of sight (LoS) communication links. The parameters settings are summarized in Table 2.
The demonstration diagram Fig. 4 shows, (a) the ordinary NOMA (ONOMA), where no relay used, shown in Fig. 4a, and the cooperative NOMA (CNOMA), Fig. 4b, assuming that the near-end node has both signal forwarding and signal reception functions for NOMA system. Figure 4c shows the CNOMA but added and independent rely aided which is called ICNOMA.
The simulation results and a discussion of the predetermined model are presented the following figures. As shown in Fig. 5, there is a relation between the sum rates and the transmitted power for the two schemas OMA and NOMA. NOMA outperforms of sum rates performance over OMA by about 6%. Also, the NOMA with using CNN approach outperforms that used in (Liang et al. 2022) by ~ 1.4%. Figure 6 shows the results for using different values of 2 Sci , indicting the scintillation coefficient with three different algorithms applied with CNN approach, FPA, SSOPA, GOPA. It is noted that using SSOPA achieves better results than using FPA, while using GOPA increases the sum rates over both FPA and SSOPA. For (a) 2 Sci = 0.1, the GOPA outperforms FPA and SSOPA by ~ 0.38% and 4.7%, respectively. For (b) 2 Sci = 0.4, GOPA outperforms FPA and SSOPA by ~ 0.5% and 4.5%, respectively, and for (c) 2 Sci = 1, GOPA outperforms FPA and SSOPA by ~ 4% and 5.9%, respectively.  We deduce that increasing 2 Sci reduces the sum rates. Comparing our results with that in (Liang et al. 2022), we find the improvement of using GOPA combined with CNN approach increases the sum rate performance by ~ 1.6%, 1.05% and 2.5%, respectively, when using the ICNOMA_UOWC (a) 2 Sci = 0.1, (b) 2 Sci = 0.4, (c) 2 Sci = 1.0. The GOPA, SSOPA and FPA algorithms are used with the CNN approach to find the relation between depth of transmitter and sum rates for three different types of water (a) clear ocean water, (b) pure ocean water, and (c) coastal water. Figure 7 shows that the GOPA algorithm achieves the best performance among other algorithms for three different types of water. The GOPA achieves an improvement of 1.2%, 2.5%, 8.7% than FPA, and 0.12%, 0.34%, 2.09% than SSOPA, for the types of water; clear, pure and coastal water, respectively. Also, comparing our results with that in (Liang et al. 2022), we find that using GOPA with CNN approach outperforms the old algorithm without using DL by ~ 0.67%,1.53%,3.63% for clear, pure, coastal water, respectively.

Conclusion
Applying the CNNLSTM with different power allocation algorithms improves the performance of UOWC systems. The obtained results reveal that NOMA outperforms OMA by ~ 6%. Also, applying the SSOPA algorithm achieves better results than FPA, while using GOPA increases the sum rates over both FPA and SSOPA. I t is also found that GOPA combined with CNN enhances the performance of sum rates by ~ 2.5% than the ICNOMA_UOWC. The GOPA achieves 1.2%, 2.5%, 8.7% improvement than FPA, and 0.12%, 0.34%, 2.09% than SSOPA, for the clear, pure, and coastal water, respectively. The ICNOMA outperforms the sum rates of both ONOMA and CNOMA by 20.4% and 3.2% respectively.
Author contributions WMS, MHA, ESM have directly participated in the planning, execution, and analysis of this study. WMS drafted the manuscript. All authors have read and approved the final version of the manuscript.
Funding Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB). The authors did not receive any funds to support this research.

Data availability
The data used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations
Competing interest The authors declare that they have no competing interest.