How do I resolve a pose-dependent noisy response during IMU sensor fusion with imufilter?

Hi, I'm using imufilter to fuse accelerometer and gyroscope data. I previously was having issues with drift so I decided to use the envelope of the gyroscope signal to reset the filter (fuse) when the gyroscope measurement was low. This works fairly well, but now I sometimes get sections of the output orientation signal that show lots of noise. This noise affects all the sensors for the same section of signal (I have recordings from 9 sensors), EDIT: It seems to happen when the accelerometer values are around (9.8, 0, 0), so when one particular axis is aligned with gravity

Here is the code I have been using (I have played round with it but this is basically what I did):

figure,
for n = 1:9
    reset(fuse);
    [gyro_env_upper,gyro_env_lower] = envelope(IMU(n).gyro,100,'rms');
    gyro_env_upper = gyro_env_upper - mean(gyro_env_upper(1:50,:));
    gyro_env_lower = gyro_env_lower - mean(gyro_env_lower(1:50,:));
    for m = 1:length(IMU(n).gyro)
        if ((abs(gyro_env_upper(m,:))<0.002) & (abs(gyro_env_lower(m,:))<0.002))
            reset(fuse)
        end
        q(m,:) = fuse(IMU(n).acc(m,:),IMU(n).gyro(m,:));
    end
    axang = quat2axang(q);
    plot(axang(:,4))
    hold on 
end

fuse is a declaration of an imufilter, IMU(n).acc and IMU(n).gyro are sets of raw data I have saved in a struct

Here is an example where the noise is present for all 9 sensors at the beginning of the recording only (this is the axis-angle convention, but the noise also affects the quaternion convention)