Azure流分析-计算线性回归

Question

I am trying to calculate some linear regression with Azure Stream Analytics.

The Setup: There is a Sensor sending temperature and time to an IoT-Hub, the Stream-Analytics is listening to the IoT-Hub with a SlidingWindow and should calculate the trend of the temperature in this window (with linear regression).

An event looks like this:

{"deviceId":"sensor_1","temp":322.3376736446427,"time":1517500183940}

The SQL I got so far is this:

WITH Step1 AS
(
select time AS x, avg(time) over () AS x_bar,
       temp AS y, avg(temp) over () AS y_bar
FROM inputStream
GROUP BY SlidingWindow(second,10) 
),
Step2 AS (
    SELECT (sum((x - x_bar) * (y - y_bar)) / sum((x - x_bar) * (x - x_bar))) AS slope,
           max(x_bar) AS x_bar_max,
           max(y_bar) AS y_bar_max    
    FROM Step1
),
FinalStep AS (
    SELECT  slope, 
            y_bar_max - x_bar_max * slope AS intercept
    FROM Step2
)

SELECT * INTO outputEventHub FROM FinalStep

The original Linear Regression SQL-Template from here looked like this:

select slope, 
       y_bar_max - x_bar_max * slope as intercept 
from (
    select sum((x - x_bar) * (y - y_bar)) / sum((x - x_bar) * (x - x_bar)) as slope,
           max(x_bar) as x_bar_max,
           max(y_bar) as y_bar_max    
    from (
        select x, avg(x) over () as x_bar,
               y, avg(y) over () as y_bar
        from ols) s;
) 

If you know some better way to do this without Stream Analytics I am also fine. Please share your thoughts and thanks in advance!

Answer 1

ASA doc has an example of two variable regression. https://docs.microsoft.com/en-us/azure/stream-analytics/stream-analytics-high-frequency-trading