Abstract: Abstract: With the progress of agricultural technology, agricultural machinery is developing towards the direction of large power, high speed and high efficiency. Due to structural complexity and harsh working conditions, the gear in the agricultural machinery is prone to failure. If the fault cannot be detected accurately, it will result in great economic losses and even much more serious casualties. Yet, the fault feature information is not only weak, but also usually drowned in strong background noise. As a result, it is a challenging project to extract the fault feature information precisely. Different from the traditional de-noising method, the traditional frequency-shifted and re-scaling stochastic resonance method can divert the noise energy to the weak characteristic signal. So it can make the weak characteristic signal increased when the noise was reduced. Then it can precisely realize the aim of detecting the weak characteristic signal which is drowned in strong background noise. Meanwhile, it breaks the limitation of the traditional stochastic resonance due to the frequency-shifted and re-scaling method involved. So, it can be applied under the conditions of large parameters. However, the traditional frequency-shifted and re-scaling stochastic resonance method just does the optimization on the single parameter and neglects the interaction between individual parameter. It cannot sufficiently make full use of the advantages of stochastic resonance in the weak feature signal extraction. So an improved frequency-shifted and re-scaling stochastic resonance method is proposed in this paper. In the proposed method, the signal-noise ratio (SNR) is chosen as the optimization objective, and the improved fish swarm algorithm is used for synchronous optimization of the frequency-shifted and re-scaling stochastic resonance system parameters. With the progress of the optimization, the precision and speed of the traditional fish swarm algorithm will become lower and lower due to the fixed view and step. So an improved method is taken in this paper and the details are as follows. In the improved fish swarm algorithm, each fish calculates the average of the distances between all other fishes and itself. Then the average value is taken as the visual field of itself, and meanwhile the 0.1 times visual field is taken as the step of itself. In this paper, in order to verify the validity of the method proposed, a numerical simulation is performed firstly. The numerical simulation results show that the method can divert the noise energy to the weak characteristic signal and effectively detect the weak feature information under strong background noise. Secondly, the experimental investigation of a gear with normal teeth and broken teeth is made. We find that compared with the processed vibration signals of the normal teeth and broken teeth, the fault is prone to distinguish. And we can judge that the fault occurs on little gear, which is in agreement with the experimental condition. The experimental results show that the method proposed can not only validly detect the weak information under strong background noise, but also make a diagnosis for the gear fault. Finally, by comparing this proposed method with the traditional frequency-shifted and re-scaling stochastic resonance method and the wavelet threshold de-noising method, we find that this method is much more valid in weak feature detection under strong background noise.