Method and apparatus for trimming sensor output using a neural network engine
A sensor is provided, comprising: a first sensing element that is arranged to generate, at least in part, a first signal; a second sensing element that is arranged to generate, at least in part, a second signal; and a neural network circuit that is configured to output an adjusted signal based on the first signal and the second signal.
1. A sensor, comprising:
a first sensing element that is arranged to generate, at least in part, a first signal;
a second sensing element that is arranged to generate, at least in part, a second signal, the second signal being indicative of an environmental condition; and
a neural network circuit implementing a neural network that is configured to adjust the first signal based on the second signal, the neural network circuit including a control circuitry, a computation unit, a plurality of registers, and a first switching circuitry that is configured to couple the plurality of registers to the computation unit, the computation unit including an addition unit, a lookup table, and a second switching circuitry that is configured to couple to one another the plurality of registers, the addition unit, and the lookup table, the second switching circuitry being configured to provide, at least in part, a first data path that passes through the lookup table and extends from the addition unit to the plurality of registers, and a second data path that extends from the addition unit through the plurality of registers without passing through the lookup table, the control circuitry being configured to:
(i) cause the second switching circuitry to route, to the lookup table, each of a plurality of affine function values that are calculated by the addition unit,
(ii) cause the second switching circuitry to route, to a different one of the plurality of registers, each of a plurality of activation function values that are retrieved by the lookup table, each of the plurality of activation function values being retrieved based on a different one of the plurality of affine function values in a same clock cycle with the addition unit performing a last calculation of the affine function,
(iii) cause the second switching circuitry to route, to the addition unit, each of the plurality of activation function values from a different one of the plurality of registers, and
(iv) cause a calculation of a weighted sum of the plurality of activation function values so as to produce a value of the first signal that is corrected for changes in sensitivity of the first sensing element that are caused by the environmental condition,
wherein each of the plurality of activation function values corresponds to a different one of a plurality of neurons in a hidden layer of the neural network, and each of the plurality of affine function values also corresponds to a different one of the plurality of neurons in the hidden layer of the neural network.
2. The sensor of claim 1 , wherein the environmental condition includes at least one of temperature, humidity, and stress.
3. The sensor of claim 1 , wherein the first sensing element includes a Giant Magnetoresistance (GMR) element, and the second sensing element includes a temperature sensing element.
4. The sensor of claim 1 , wherein each of the plurality of activation function values is calculated in accordance with the equation of:
A =tan h ( Z )
where A is the activation function value, and Z is one of the plurality of affine function values based on which the activation function value is calculated.
5. A sensor, comprising:
a first sensing element that is arranged to generate, at least in part, a first signal, the first sensing element being one of a magnetic field sensing element, a pressure sensing element, or a light-sensing element;
a second sensing element that is arranged to generate, at least in part, a second signal, the second signal being indicative of an environmental condition; and
a neural network circuit implementing a neural network that is configured to adjust the first signal based on the second signal, the neural network circuit including a control circuitry, a computation unit, a plurality of registers, and a first switching circuitry that is configured to couple the plurality of registers to the computation unit, the computation unit including an addition unit, a lookup table, and a second switching circuitry that is configured to couple to one another the plurality of registers, the addition unit, and the lookup table, the second switching circuitry being configured to provide, at least in part, a first data path that passes through the lookup table and extends from the addition unit to the plurality of registers, and a second data path that extends from the addition unit through the plurality of registers without passing through the lookup table, the control circuitry being configured to:
(i) cause the second switching circuitry to route, to the lookup table, each of a plurality of affine function values that are calculated by the addition unit, each of the plurality of affine function values being a weighted sum of a sample of the first signal and a sample of the second signal,
(ii) cause the second switching circuitry to route, to a different one of the plurality of registers, each of a plurality of activation function values that are retrieved by the lookup table, each of the plurality of activation function values being retrieved based on a different one of the plurality of affine function values in a same clock cycle with the addition unit performing a last calculation of the affine function,
(iii) cause the second switching circuitry to route, to the addition unit, each of the plurality of activation function values from a different one of the plurality of registers, and
(iv) cause calculation of a weighted sum of the plurality of activation function values so as to produce a value of the first signal that is corrected for changes in sensitivity of the first sensing element that are caused by the environmental condition,
wherein each of the plurality of activation function values corresponds to a different one of a plurality of neurons in a hidden layer of the neural network, and each of the plurality of affine function values also corresponds to a different one of the plurality of neurons in the hidden layer of the neural network.
6. The sensor of claim 5 , wherein each of the plurality of activation function values is calculated in accordance with the equation of:
A =tan h ( Z )
where A is the activation function value, and Z is one of the plurality of affine function values based on which the activation function value is calculated.
7. The sensor of claim 5 , wherein the first sensing element includes a magnetic field sensing element.
8. The sensor of claim 5 , wherein the hidden layer is the only hidden layer in the neural network.
9. The sensor of claim 8 , wherein the second sensing element is integrated into the sensor or is separate from the sensor.
10. The sensor of claim 8 , wherein each of the affine function values is calculated in accordance with the equation of:
Z
=
[
W
1
W
2
]
[
V
1
V
2
]
+
B
where W 1 is a first weight value, W 2 is a second weight value, V 1 is a sample of the first signal, V 2 is a sample of the second signal, and B is a bias coefficient.
11. The sensor of claim 8 , wherein the environmental condition includes one of temperature, humidity, and stress.