In indoor visible light communication （VLC） systems， the reliability of high-order modulation transmission is severely constrained by strict power emission limits and multipath fading. Additionally， the

complex channel environment leads to a significant mismatch between fixed shaping parameters and dynamic time-varying channel states. Moreover， traditional compressive sensing algorithms suffer from severe atom selection errors and low channel state information （CSI） reconstruction accuracy under low signal-to-noise ratio （SNR） and nonlinear distortion conditions. To address these issues and achieve the best balance between transmission reliability and information rate， this paper proposes a probabilistic shaping VLC system model that integrates deep neural networks （DNN） and compressive sampling matching pursuit （CoSaMP） channel estimation. The aim is to achieve high-precision CSI estimation in complex environments， dynamically optimize the statistical characteristics of high-order modulation signals， and adapt to the time-varying VLC channel.Firstly， orthogonal chirp division multiplexing （OCDM） is adopted as the modulation method， taking advantage of its energy diffusion characteristics in the time-frequency plane to effectively resist frequency-selective fading caused by multipath interference. On this basis， probabilistic shaping （PS） technology following the Maxwell-Boltzmann （MB） distribution is applied to the quadrature amplitude modulation （QAM） constellation. This technique reduces the average transmission power by lowering the probability of high-energy constellation points and increasing that of low-energy points. Secondly， to ensure the precise execution of adaptive probabilistic shaping， a channel estimation algorithm integrating DNN and CoSaMP is proposed. A fully connected DNN is constructed as a preprocessing module to correct features and reduce noise in the rough received pilot observation signals. The nonlinear mapping output by the DNN and the corrected signal are input into the CoSaMP algorithm to complete high-precision sparse channel reconstruction. Further， a closed-loop feedback mechanism is established based on the SNR extracted from the high-precision reconstructed CSI. To solve the n

on-convex optimization problem of maximizing the generalized mutual information （GMI） under average power constraints， the particle swarm optimization （PSO） algorithm is executed in the offline stage to generate a mapping lookup table of the best shaping factor and discrete SNR. In the online transmission stage， the system uses the real-time estimated SNR to query this table and dynamically adjust the MB distribution parameters， ensuring that the signal's statistical characteristics continuously match the instantaneous channel state. Finally， a hardware-in-the-loop experiment platform is built to comprehensively evaluate the system performance.Experimental results show significant improvements in multiple indicators. Firstly， increasing the probabilistic shaping factor from 0 to 0.6 effectively concentrates the signal energy， achieving a significant SNR gain at the bit error rate （BER）. Secondly， the integration of DNN and CoSaMP algorithms effectively corrects the atom selection bias at low SNR， reducing the normalized mean square error by one order of magnitude and achieving approximately 2 dB SNR gain from 16-QAM to 64-QAM modulation. Finally， the APS demonstrates strong environmental adaptability. When the channel SNR increases from 10 dB to 24 dB， the system autonomously adjusts the shaping factor， significantly increasing the modulation error ratio from 8.82 dB to 22.61 dB. At the BER of

<math id="M2"><msup><mrow><mn mathvariant="normal">10</mn></mrow><mrow><mo>-</mo><mn mathvariant="normal">3</mn></mrow></msup></math>

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=110277421&type=

https://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=110277448&type=

6.58090878

2.88636374

， compared with the joint DNN+Transformer channel estimation and probabilistic shaping， joint CoSaMP channel estimation and probabilistic shaping， joint shaping， fixed parameter probabilistic shaping， and no shaping schemes， the proposed system achieves approximately 0.5 dB， 1 dB， 1.5 dB， 3 dB， and 6 dB SNR gains， respectively， and achieves the highest generalized mutual information （GMI） and the best received eye diagram quality within the effective SNR range.The probabilistic shaping VLC system model integrating deep neural networks （DNN） and compressive sampling matching pursuit （CoSaMP） channel estimation proposed in this paper effectively overcomes the estimation bottleneck and dynamic mismatch problems in complex optical channels. This scheme， while reducing the noise sensitivity of high-order constellations， achieves an exact match between the source entropy and the instantaneous channel capacity， providing a highly robust new paradigm for enhancing the comprehensive transmission efficiency of the next-generation visible light communication networks.