Optimizing Data Rates for Visible Light Communication with Carrierless Amplitude and Phase Modulation

As the radio spectrum becomes increasingly crowded, visible light communication (VLC), a promising alternative for high-speed wireless communications, is gaining traction. VLC uses a light-producing device (such as an LED) to transmit signals and a photodiode to receive them. It offers a number of advantages over radio frequency data transmission, including a wide, license-free spectrum, better security, and immunity to electromagnetic interference.

At the INSA Rennes Institute of Electronics and Telecommunications (IETR), we have been evaluating various modulation techniques for VLC systems. Our goal is to build these systems using low-cost components and optimize their data rates under standard illuminations levels—the lighting in a typical office environment, for example. On a recent project, we modeled, built, and tested a VLC system that employed spectrally efficient, carrierless amplitude and phase (CAP) modulation to achieve reliable, high-throughput transmission: 184 Mbps with a bit error rate (BER) below 10^-3. Modeling and simulating this system in MATLAB^® with Communications Toolbox™ enabled us to verify our experimental setup, compare simulated and measured results, and visualize key performance metrics.

Modeling the VLC System

One advantage that CAP modulation has over discrete multitone (DMT) and other modulation schemes is the simplicity of its transceiver structure. The transmitter requires a pair of orthogonal pulse shaping filters, which are paired with a set of matched filters on the receiver side (Figure 1). An interesting property of CAP is that the signal spectrum is shifted around a low frequency $f_c$ such that the resulting spectrum remains in the baseband domain; i.e., the lower side band extends all the way to $f=0$. The frequency upconversion is implicit as part of the pulse shaping filtering operation, thus obviating the mixing with a carrier produced by a local oscillator.

Ease of implementation was also a key factor in our decision to use MATLAB to model and simulate our VLC setup. Our research group works extensively with MATLAB, and the group has built up a large library of reusable MATLAB code that we accessed as needed for this project. Further, Communications Toolbox, which is widely used for the development of radio frequency communications systems, also works well for VLC systems, with algorithms and functions that saved us time as we modeled each of the system’s three main components: the transmitter, the receiver, and the channel. 

In our transmitter model, the bits to be transmitted are first converted to QAM symbols, which are then mapped onto their in-phase and quadrature components. Both components are then upsampled before they are passed through orthogonal filters and then summed together. All these functions—including the modulation, upsampling, and summing, as well as the in-phase and quadrature filtering—were implemented using just a few lines of MATLAB code. For carrierless operation, the impulse responses of the in-phase filters and quadrature filters are realized by multiplying a standard square-root raised cosine filter response by a cosine and a sine, respectively.

The receiver mirrors the transmitter, with matching filters, downsampling, and summing of the signals. The combined signal is then passed through an equalizer to mitigate the interference between symbols stemming from the LED frequency selective response before being sent to the QAM demodulator. Like the transmitter, the receiver model was also straightforward to implement in MATLAB. 

For the channel model, we used a pseudo-noise (PN) sequence to estimate the impulse response of our actual VLC channel. The two dominant sources of noise in the system are the thermal noise inherent in the receiver circuitry and the shot noise induced in the photodiode by both the optical signal intensity and the ambient light. We modeled both noise sources as white Gaussian noise in MATLAB, with parameters based on signal-to-noise ratio measurements taken on the actual receiver. 

Once we had modeled the transmitter, channel, and receiver, we ran simulations to evaluate the performance of the entire system. As part of these evaluations, we generated constellation diagrams to check the quality of our simulated received signal (Figure 2).

Hardware Implementation and Comparison with Model

Of course, our research was not limited to VLC modeling and simulation; we were primarily focused on the real-world implementation of a VLC transceiver using CAP modulation. Our experimental setup included an arbitrary waveform generator (AWG), a low-cost white LED, a lens to concentrate the light signal, a single Silicon PIN (S-PIN) photodiode, and a real-time oscilloscope to capture and store the signal for further analysis in MATLAB (Figure 3).

To verify the results from our VLC experimental setup, we compared them against results from our MATLAB simulations. Specifically, we looked at BER as a function of throughput. As expected, BER increased with the data rate. Our experimental and simulation results followed similar trajectories, with the real-world results showing somewhat higher error rates. We attribute these differences to nonlinearities in the LED, which we have not yet accounted for in our MATLAB model (Figure 4).

Modeling the Impact of LED Nonlinearity

When comparing the constellation diagrams in simulation with that of the experiment, we noticed distortions on the edges of the received CAP constellations due to the nonlinear behavior of the LED, which impact the overall system performance. In fact, the electro-optical transfer characteristic relating the input current and the radiated optical power is a nonlinear function. In addition, the effects of the LED nonlinearity are frequency-dependent and therefore exhibit a memory effect that grows with the signal bandwidth.

To model the impact of the LED nonlinearity in simulation, we studied two models with different complexities. The first one is the Hammerstein model, which consists in a memoryless polynomial followed by a first-order low-pass filter (Figure 5). The advantage of this model is its simplicity, since the coefficients of the polynomial function and the first-order low-pass filter can easily be measured.

The second model we investigated is the Volterra model based on the Volterra series expansion. The coefficients of the Volterra series, also referred to as kernels, are not determined in a straightforward manner. We employed an adaptive algorithm that minimizes the error between the received signal in the experiment and the transmitted signal to extract the coefficients of the Volterra series up to the second order (Figure 6). 

We then generated the constellation diagrams in simulation with the two nonlinearity models and compared them with that of the experiment (Figure 7). We observed similar slight distortions on the top right and bottom left of the constellations. Moreover, the received constellation in the experiment is slightly noisier. While it has the virtue of being simple, the Hammerstein model lacks accuracy for the case of large signal bandwidth. On the other hand, the Volterra model exhibits a more accurate representation of the nonlinearity impact of the LED.

Plans for Further Research

We are continuing our evaluation of modulation schemes—including CAP, DMT, and pulse amplitude modulation (PAM)—for optimized VLC systems. In particular, we are currently implementing a post-distortion algorithm that will help mitigate the adverse effects that stem from nonlinearities in the LED. Next, we will begin applying multiple-input multiple-output (MIMO) techniques to VLC communications.

As we continue to improve the range, throughput, and robustness of our VLC design, we will examine the practical implementation of these designs in real-world applications. One such application is the wireless data links between overhead light fixtures in an office environment and devices embedded in or connected to workstations.