Since the term linear does not appear in the question and the current answers, let me offer a complementary perspective.
A kernel in this acceptation (especially for images, which don't always follow linear rules, think about occlusion or saturation) is an array that is applied, somehow, on any input data. One often distinguishes linear and non-linear kernel (as one has linear and nonlinear filters, even is the terminology might seem improper).
Let us start from the linear point-of-view in the most specific sense: the filter array is applied as a convolution. Then @MarcusMuller's answer is perfect: a set, an array of linear filters, applied to input data as convolutions to yield several separate output data.
Most additional scalar linear operation (like the sum, the average, a weighted combination) on the output would be "useless": as they commute, summing the output is equivalent to summing the three filters in one single filter, and performing only one single convolution on the data.
Which leads us back to the objective in your comment; traditionally, a linear (analysis, I will come back on that later) filter-bank (FB) is used to split or separate data into components, often with separate spectra or a narrower ocontent (low-, mid- or high-frequencies for a three-band filter-bank). Or to merge different data streams into others, with a wider spectrum. So a generic multi-input-multi-output (MIMO) FB takes one or several inputs, filters them into one or several outputs. One then distinguishes analysis or synthesis filter banks.
Generally, recombining the outputs from an analysis FB drifts away from the separation objective. But a single filter is also a filter-bank (not very interesting per se though).
But sometimes, this can be more efficient (computationally for instance).
Now, having narrower/wider outputs invites to rate variations, like downsampling and upsampling before or after the filters. To me, the most accepted sense of a filter bank is a bank of linear filters optionally combined with (linear, but not shift-invariant) upsampling or downsampling operations. And it is somewhat related to linear transforms, allowing expansion or shrinkage on the number of coefficients (they can be critical, oversampled or undersampled).
Then, people extend the notion to nonlinearity: filters can be nonlinear (like the median) and the kernels are interpreted as weights applied to a chunk of data. Or the data can be combined in non-linear ways, with $\min$, $\max$, AND or OR...
But in your case, as Marcus said, I'd bet on three standard filtered outputs. But in this case, there is not relationship between filters (except their kernel size), and what is powerful in the filter-bank theory is the linkage between the filters, and how one can optimize them. Now a couple of pointers: