Have you actually measured the voltage the LM35 is outputting with a multimeter? I would trust the LM35 far more than the ATmega's ADC to go completely down to 0 V.
From the graph you included, it looks like something is hitting a rail somewhere.
If it is indeed the LM35 outputting the errant voltages, try a pulldown (~50 kilohm?) on the output to ground (or better yet, -5 V).
You're measuring a very small voltage with the built-in ADC - 20 °C is only 200 mV. Furthermore, without a negative power rail, the sensor is only really good to 2 °C, and any bias current on the output (from the ADC) will probably affect the readings you get.
While I am not personally familiar with the specific product mentioned in the question, the description at that link clarifies at least some of what the microcontroller brings to the table:
- Conversion of a linear resistance change to a standard analogue current / voltage
signal from e.g. valves or linear movements with attached potentiometer.
- Suitable in applications with potentiometers that are not fully utilised as the 0 and 100% adjustments on the front can be adjusted individually without interacting.
In brief: A basic temperature sensor will demonstrate:
- Not necessarily linear resistance change against temperature (or voltage change against temperature, for thermocouples)
- For negative / positive temperature coefficient resistive sensors, a resistance relationship to temperature, which then needs some work to convert the output to a linear current relationship, or a linear voltage relationship, with the temperature
- Some interaction between set-point adjustments for minimum and maximum scale values if analog methods are used instead of programmable ones, i.e. changing the upper set-point will often affect the lower set point, etc.
What the microcontroller enables is an output that is linearized (could be a lookup table, some parametric algorithm, or use-your-imagination) and either a current value, used in industrial current loop analog communications, or a voltage value, used in many different systems. Again, from the link:
Analogue standard current / voltage output of 0/4...20 mA / 0/2...10 VDC. The output signal is proportional and linear to the value of the temperature or resistance value that influences the input.
The "Analogue standard" is with reference to industrial sensing standards.
Of course, both the linearization, and the conversion from a resistance representation of temperature to a current or voltage representation, could be done by an external device, either microcontroller based or analog. Sometimes such massaging of values is even done independently on a computer: Use of a spreadsheet for such purposes is not unknown!
However, for someone whose prime objective is to focus all available engineering time on an end product where the temperature sensing is merely an enabler, a component, devices like the one described are not overkill: How many work-hours of quality engineer time would balance out the premium paid for an industrial temperature sensing device with the flexibility described?
Best Answer
Limits include the sensor accuracy (calibration and interchangeability at the temperatures of interest), self-heating of the sensor, coupling of the sensor to the thing you're trying to measure, effect of the sensor on the thing you're trying to measure (heat input or heat loss, for example), instrument calibration and changes in that with time and environmental factors, and noise.
Noise (including quantization noise) is the only one that you can improve by averaging measurements, and not necessarily even that, depending on the type of noise.
Most errors sources are systematic, so you may get really low-noise measurement of a value that is not very accurate.