Algorithms and implementations to overcome practical issues in active noise control systems

Abstract

Noise control has garnered increasing attention, in the face of growing urbanization. To deal with the noise issue, the active noise control (ANC) technique is revisited in the thesis. The ANC technique utilizes a sensor to acquire the noise information and synthesizes the anti-noise by driving the actuator to cancel the noise. Compared to the traditional passive noise control approach, it shows the apparent advantages of the cost, size, and easy deployment. Hence, ANC is widely used in active control headphone and in automotive interiors. However, the ANC technique still possesses many practical challenges, such as the nonlinearity of the secondary path, the placement of the error sensor, and the soaring computational requirements with the increasing scale of multichannel active noise control (MCANC) systems. To effectively solve these practical issues, the thesis firstly reviews existing optimization techniques and subsequently proposes unique algorithms and methods.

The distortions caused by the saturation of the audio amplifier play the central role in the nonlinearity of ANC. These distortions seriously influence noise reduction performance and often result in the divergence of the adaptive algorithm. One of the most effective ways is to constrain the power of the output signal to ensure that the audio amplifier operates linearly and avoids saturation. To guarantee optimal noise reduction performance of the ANC system under output constraint, a two-gradient direction FxLMS algorithm is proposed and validates. The algorithm automatically switches the gradient direction to reduce the power gain of the control filter when the ANC system exceeds the output-power constraint. An alternative method is to utilize the leaky FxLMS algorithm adaptively restraining the magnitude of the control filter. With the proposed optimal leak factor suggested in the thesis, the leaky FxLMS algorithm gains optimal control under output constraint. The numerical simulations and the experiments verify the performance of both algorithms.

In some ANC applications, it is impractical to place the error sensor in the desired location. To overcome this issue, the virtual sensing ANC technique that is capable of projecting the quiet zone to the desired position from the physical microphone is reexamined. In the thesis, the virtual sensing technique is extended from single-channel ANC to the multichannel ANC. The multichannel virtual sensing ANC (MVANC) is shown to achieve the same noise attenuation performance as MCANC, which places the real error sensors at the desired positions. Furthermore, the physical limitation of the MVANC approach is determined by a sensor-actuator analysis and experimentally validated in a small chamber.

To alleviate the computational complexity of FxLMS-based approaches, selective fixed-filter active noise control (SFANC) approach is proposed in the thesis that obtains a series of filters and stores into a database in the preliminary stage. When dealing with a specific type of primary noise, the SFANC algorithm selects a suitable pre-trained filter from the database based on the frequency-band-match method. The simulations show that SFANC obtains satisfactory noise reduction performance in different applications. Moreover, since SFANC does not contain the adaptive update process, it saves a considerable amount of computations and achieves higher robustness.

Finally, the thesis explores a series of approaches, which can optimize the structure of the digital processing algorithm implemented on the FPGA platform. Firstly, the systolic FxLMS, which significantly increase the sampling rate and throughput rate of the ANC system, is proposed in the thesis. For realizing a large channel count MCANC on FPGA, a multiple-parallel-branch with folding structure is proposed. This architecture is used to implement a 24-channel MCANC and successfully balances the usage of the computation resource and the maximum operating speed. This 24-channel MCANC is deployed in an open window whereby the noise reduction performance sufficiently validates the effectiveness of the proposed architecture.