Automated analysis of non-stationary machine performance
A method for monitoring at least one machine including causing at least a first sensor to acquire at least a first non-stationary signal from at least one machine operating in a non-stationary manner during at least one operational time frame, the first sensor(s) providing at least a first non-stationary output, causing at least a second sensor to acquire at least a second non-stationary signal from the machine(s) during the operational time frame(s), the second sensor(s) providing at least a second non-stationary output, fusing the first non-stationary output(s) with the second non-stationary output(s) to produce a fused output, extracting at least one feature from the first and/or second non-stationary signal(s) based on the fused output, analyzing the feature(s) to ascertain a state of health of the machine(s) and performing a repair operation, maintenance operation and/or modification of operating parameters of the machine(s) based on the analyzed state of health.
1 . A method for monitoring a robotic machine comprising:
causing at least a first sensor to acquire at least a first non-stationary signal from a servo-motor driving a component of a robot in a non-stationary manner during at least one operational time frame, said at least first sensor providing at least a first non-stationary output;
causing at least a second sensor to acquire at least a second non-stationary signal from said servo-motor during said operational time frame, said at least second sensor providing at least a second non-stationary output;
fusing, by a computer-implemented machine-learning model, said at least first non-stationary output with said at least second non-stationary output, wherein said fusing comprises modifying said at least first non-stationary output based on said at least second non-stationary output, to produce a fused output, said fusing, by said computer-implemented machine-learning model, further comprising ascertaining cross-correlation levels between said first non-stationary signal and said second non-stationary signal, by cross-correlating said first non-stationary signal with said second non-stationary signal, and grading said cross-correlation levels, based on weighting a correlation magnitude by a rotational speed of said servo-motor;
extracting, by said computer-implemented machine-learning model, at least one feature of at least one of said first and second non-stationary signals based on said fused output;
tracking, by feature tracking algorithms of said computer-implemented machine-learning model, said at least one feature over time; and
adjusting operational parameters of said servo-motor to optimize performance of said robot, based on said tracking,
wherein said fusing further comprises identifying, by said computer-implemented machine-learning model and based on said cross-correlation levels, a plurality of intervals within said non-stationary operating regime, and
wherein said adjusting operational parameters of said servo-motor to optimize said performance also comprises adjusting at least one of a sensor sampling frequency and a sensor sampling duration of said first and second sensors, based on said plurality of intervals.
2 . A method according to claim 1 , wherein said first non-stationary signal is of a same type as said second non-stationary signal.
3 . A method according to claim 1 , wherein said first non-stationary signal is of a different type to said second non-stationary signal.
4 . A method according to claim 2 , wherein said first non-stationary signal and said second non-stationary signal are vibration signals.
5 . A method according to claim 3 , wherein one of said first and second non-stationary signals is a vibration signal and another one of said first and second non-stationary signals is a magnetic signal.
6 . A method according to claim 1 , wherein said fusing, by said computer-implemented machine-learning model, further comprises:
applying a wavelet transform to both of said at least first and second different non-stationary signals;
generating a mask of said wavelet transform of one of said first and second non-stationary signals; and
multiplying said mask of said wavelet transform of said one of said first and second non-stationary signals by said wavelet transform of the other one of said first and second non-stationary signals.
7 . A method according to claim 6 , wherein said mask comprises an inverse mask.
8 . A method according to claim 1 , wherein:
said extracting further comprises extracting, by said computer-implemented machine-learning model, said at least one feature for at least some intervals of said plurality of intervals; and
said tracking further comprises tracking said at least one feature, for said at least some intervals of said plurality of intervals.
9 . A method according to claim 8 , wherein said tracking said at least one feature, by said computer-implemented machine-learning model, comprises:
training said computer-implemented machine-learning model by:
supplying, to said computer-implemented machine-learning model, a time series of said at least one feature extracted from said first and second signals;
predicting, by said computer-implemented machine-learning model, a predicted successive value of said at least one feature in said time series;
comparing, by said computer-implemented machine-learning model, said predicted successive value of said at least one feature to a measured successive value of said at least one feature;
updating, iteratively and by said computer-implemented machine-learning model, parameters of said machine-learning model, until a difference between said predicted successive value of said at least one feature and said measured successive value of said at least one feature is less than a given threshold; and
thereafter said training, operating said computer-implemented machine-learning model by:
predicting, by said computer-implemented machine-learning model, another predicted successive value of said at least one feature in said time series;
comparing, by said computer-implemented machine-learning model, said another predicted successive value of said at least one feature in said time series, to another measured successive value of said at least one feature; and
identifying, by said computer-implemented machine-learning model, a faulty operating state of said robot based on said comparing,
said adjusting operational parameters being based on said identified faulty operating state.
10 . A system for monitoring at least one machine comprising:
a first sensor operative to acquire at least a first non-stationary signal from a servo-motor driving a component of a robot in a non-stationary manner during at least one operational time frame, said at least first sensor being operative to provide at least a first non-stationary output;
a second sensor operative to acquire at least a second non-stationary signal from said servo-motor during said operational time frame, said at least second sensor being operative to provide at least a second non-stationary output;
a signal processor employing a machine-learning model operative to fuse said at least first non-stationary output with said at least second non-stationary output, based on modification of said at least first non-stationary output by said at least second non-stationary output, to produce a fused output, said machine-learning model being further operative to cross-correlate said first non-stationary signal with said second non-stationary signal, to ascertain cross-correlation levels between said first non-stationary signal and said second non-stationary signal, and to grade said cross-correlation levels, based on weighting on a correlation magnitude by a rotational speed of said servo-motor;
a feature extractor operative to extract, using said machine-learning model, at least one feature of at least one of said first and second non-stationary signals based on said fused output;
a feature tracker operative to track, using said machine-learning model, said at least one feature over time; and
a machine control module operative to adjust operational parameters of said servo-motor in order to optimize performance of said robot,
wherein said signal processor is further operative to identify, by said machine-learning model and based on said cross-correlation levels, a plurality of intervals within said non-stationary operating regime, and
wherein said control module is further operative to adjust operational parameters of said servo-motor in order to optimize performance of said robot by adjusting at least one of a sensor sampling frequency and a sensor sampling duration of said first and second sensors, based on said plurality of intervals.
11 . A system according to claim 10 , wherein said first sensor is of a same type as said second sensor.
12 . A system according to claim 10 , wherein said first sensor is of a different type to said second sensor.
13 . A system according to claim 11 , wherein said first and second sensors are vibration sensors.
14 . A system according to claim 12 , wherein one of said first and second sensors is a vibration sensor and another one of said first and second sensors is a magnetic sensor.
15 . A system according to claim 10 , wherein said signal processor being operative to fuse, by said machine-learning model, said first non-stationary output with said second non-stationary output comprises said signal processor being operative to:
apply a wavelet transform to both of said at least first and second different non-stationary signals;
generate a mask of said wavelet transform of one of said first and second non-stationary signals; and
multiply said mask of said wavelet transform of said one of said first and second non-stationary signals by said wavelet transform of the other one of said first and second non-stationary signals.
16 . A system according to claim 15 , wherein said mask comprises an inverse mask.
17 . A system according to claim 10 , wherein said signal processor is operative to grade said cross-correlation levels in accordance with:
1
S
tot
∫
1
W
(
f
rpm
,
t
)
S
1
*
(
t
)
·
S
2
(
t
+
τ
)
dt
where S i represents data from said first and second sensor, S tot is a sum of a product of an energy magnitude of said first and second sensors and W is an RPM weighted function.
18 . A system according to claim 10 , wherein:
said feature extractor is further operative to extract, by said machine-learning model, said at least one feature for at least some intervals of said plurality of intervals; and
said feature tracker is further operative to track, by said machine-learning model, said at least one feature, for said at least some intervals of said plurality of intervals.
19 . A system according to claim 18 , wherein said control module is further operative to adjust at least one of a sensor sampling frequency and a sensor sampling duration of said first and second sensors, based on said plurality of intervals.
20 . A system according to claim 18 , wherein said feature tracker being operative to track, by said machine-learning model, said at least one feature, comprises:
said feature tracker being operative to train said machine-learning model by being operative to:
supply, to said machine-learning model, a time series of said at least one feature extracted from said first and second signals;
predict, by said machine-learning model, a predicted successive value of said at least one feature in said time series;
compare, by said machine-learning model, said predicted successive value of said at least one feature to a measured successive value of said at least one feature; and
update, iteratively and by said machine-learning model, parameters of said machine-learning model, until a difference between said predicted successive value of said at least one feature and said measured successive value of said at least one feature is less than a given threshold; and
said feature tracker being operative to implement said machine-learning model by being operative to:
predict, by said machine-learning model, another predicted successive value of said at least one feature in said time series;
compare, by said machine-learning model, said another predicted successive value of said at least one feature in said time series, to another measured successive value of said at least one feature; and
identify, by said machine-learning model, a faulty operating state of said robot based on the comparison,
said control module being operative to adjust said operational parameters based on said identified faulty operating state.
21 . A method for monitoring at least one machine comprising:
causing at least a first sensor to acquire at least a first non-stationary signal from at least one machine operating in accordance with a non-stationary operating regime during at least one operational time frame, said at least first sensor providing at least a first non-stationary output;
causing at least a second sensor to acquire at least a second non-stationary signal from said at least one machine during said operational time frame, said at least second sensor providing at least a second non-stationary output;
fusing, by a computer-implemented machine-learning model, said at least first non-stationary output with said at least second non-stationary output, wherein said fusing comprises modifying said at least first non-stationary output based on said at least second non-stationary output, to produce a fused output, said fusing further comprising:
ascertaining, by said computer-implemented machine-learning model, cross-correlation levels between said first non-stationary signal and said second non-stationary signal, by cross-correlating said first non-stationary signal with said second non-stationary signal, and grading said cross-correlation levels, based on weighting a correlation magnitude by a rotational speed of said at least one machine; and
identifying, by said computer-implemented machine-learning model and based on said cross-correlation levels, a plurality of intervals within said non-stationary operating regime,
extracting, by said computer-implemented machine-learning model, at least one feature of at least one of said first and second non-stationary signals based on said fused output, for at least some intervals of said plurality of intervals;
tracking, by feature tracking algorithms of said computer-implemented machine-learning model, said at least one feature over time, for said at least some intervals of said plurality of intervals; and
adjusting operational parameters of said machine to optimize said machine performance, based on said tracking,
wherein said at least one machine comprises a servo-motor driving a component of a robot in a non-stationary manner,
wherein said adjusting operational parameters of said machine to optimize said machine performance also comprises adjusting at least one of a sensor sampling frequency and a sensor sampling duration of said first and second sensors, based on said plurality of intervals.
22 . A method according to claim 21 , wherein said tracking said at least one feature, by said computer-implemented machine-learning model, comprises:
training said computer-implemented machine-learning model by:
supplying, to said computer-implemented machine-learning model, a time series of said at least one feature extracted from said first and second signals;
predicting, by said computer-implemented machine-learning model, a predicted successive value of said at least one feature in said time series;
comparing, by said computer-implemented machine-learning model, said predicted successive value of said at least one feature to a measured successive value of said at least one feature;
updating, iteratively and by said computer-implemented machine-learning model, parameters of said machine-learning model, until a difference between said predicted successive value of said at least one feature and said measured successive value of said at least one feature is less than a given threshold; and
thereafter said training, operating said computer-implemented machine-learning model by:
predicting, by said computer-implemented machine-learning model, another predicted successive value of said at least one feature in said time series;
comparing, by said computer-implemented machine-learning model, said another predicted successive value of said at least one feature in said time series, to another measured successive value of said at least one feature; and
identifying, by said computer-implemented machine-learning model, an anomalous operating state of said at least one machine based on said comparing,
said adjusting operational parameters being based on said identified anomalous operating state.