Unsupervised Machine Monitoring System

Abstract

Disclosed herein are methods and systems for the unsupervised, non-intrusive monitoring of motorized machines on a shop floor. A statistical algorithm is used to train itself to distinguish between an inactive and an active state of each monitored machine. The systems and methods use inexpensive sensors, do not intervene in the operation of the machines upon installation, would not adversely affect the operation of the machines, if the monitoring system would fail, do give reliable reports, and do not require any more than negligible human training.

Claims

1. A method of monitoring activity and inactivity of a motor, the method comprising: non-intrusively attaching an analog sensor to at least one cable of an electric motor; applying unsupervised statistical analysis to an output signal of the sensor; computing statistical parameters of the signal when the motor is not active; and deriving, from the statistical computation, an indication of whether the motor is active.

2. The method of claim 1, wherein the statistical analysis is provided at least in part by an expectation-maximization algorithm.

3. The method of claim 1, wherein the analog sensor is a current sensor.

4. The method of claim 1, wherein the analog sensor is a voltage sensor.

5. The method of claim 1, wherein the analog sensor is a power sensor.

6. A motor activity monitoring system comprising: at least one analog sensor attachable to a phase cable of a motor; and a processor configured to apply an unsupervised statistical analysis of an output signal of the sensor and to produce an indication of whether the motor is active or inactive.

7. The motor of claim 6, wherein the statistical analysis is provided at least in part by an expectation-maximization algorithm.

8. The motor of claim 6, wherein the analog sensor is a current sensor.

9. The motor of claim 6, wherein the analog sensor is a voltage sensor.

10. The motor of claim 6, wherein the analog sensor is a power sensor.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The invention is described below in the appended claims, which are read in view of the accompanying description including the following drawings, wherein:

[0033] FIG. 1 compares characteristics of various types of sensors;

[0034] FIG. 2 provides a simplified block diagram of a system according to one embodiment of the invention.

[0035] FIG. 3 provides a simplified flowchart of a training process for use when operating the system of FIG. 2;

[0036] FIG. 4 provides a simplified flowchart of the monitoring process for use when operating the system of FIG. 2; and

[0037] FIG. 5 provides a plot of statistical distributions that embodiments of the invention, such as those of FIGS. 2-4, use in their operation.

DETAILED DESCRIPTION

[0038] FIG. 1, using information from the McLeay PhD thesis (see Background), compares characteristics to show advantages and disadvantages of several types of sensors.

[0039] Attention is now called to FIG. 2.

[0040] A machine 20 in this embodiment of the invention uses a three-phase motor 22 having leads 26 and a one-phase motor 24 having a lead 42. Two phases of motor 22 have clamping current sensors 28 and 30 installed thereon, and the one phase of motor 24 has clamping current sensor 32 installed thereon. The current sensors are standard split current transformer sensors, such as SCT016S, produced by YHDC from DUZHUANG INDUSTRIAL PARK, QINHUANGDAO CITY, HEBEI PROVINCE, CHINA, as a non-limiting example.

[0041] The three outputs of the current sensors 28, 30, 32 are each input to three ADC analog-to-digital converters (ADCs) 36, 34, and 35, respectively, and the digitized data is fed to a processor 38.

[0042] The processor 38 processes the data in accordance to the flowcharts described below and automatically learns to distinguish between a monitored motor's inactive state, when the motor is not working, and the active state, when the motor is working. The state of the machine, as determined by processor 38, is displayed on display 40.

[0043] Several machines may be monitored in parallel by processor 38, and under such configuration the display presents the active/inactive states of all the monitored machines.

[0044] Attention is now called to FIG. 3, which provides a flowchart representing the unsupervised learning algorithm implemented in one embodiment of the system.

[0045] A length of time for which sensor data are collected is chosen by one skilled in the art. In this embodiment, data from a clamping current sensor are collected 50 for a length of time equal to one-half hour. After the data collection, the unsupervised learning algorithm processes the data as follows to find a threshold value of processed current data that indicates whether the monitored machine is active or inactive.

[0046] After collecting the data of the unsupervised learning phase, the algorithm groups 52 the data according to the time increments in which they were sensed. A non-limiting example of such time increment is one lasting 100 milliseconds in duration. The samples of each time increment are processed 54 to produce a value that is used as the increment's variable of interest for the active/inactive state determination. This variable of interest is known as the “feature value,” the feature in this example based on the amount of current supplied to the monitored machine. The feature value can be the mean (average), the RMS value, the standard deviation, the median, and even the minimum or maximum values of the sampled data, as non-limiting examples.

[0047] During the unsupervised learning phase of the operation of the present method, the algorithm assumes that a histogram of the feature values based on data obtained while the machine is in an inactive state forms a Gaussian distribution. The algorithm receives 56 initial values for a mean and standard deviation of the feature values of the Gaussian distribution. These initial values should be set to differ sufficiently from the first calculated mean and standard deviation, as discussed below, to cause an iteration, as also discussed below. The initial values for the mean and standard deviation may both be set to be equal to the minimum of all the feature values computed during the unsupervised learning phase. Such would reduce the risk of samples coming from the active state being classified as inactive. Alternatively, the mean can be set equal to the highest feature value and standard deviation can be set equal to the minimum feature value. The threshold for determining active/inactive state status (as clear from the discussion below) would start high and decrease with subsequent iterations until convergence. During the processing described below, the sum of the mean and the standard deviation will decrease and eventually stabilize. If the result for a set of initial values is that there is not even one iteration in the computations described below, the values would need to be reset and the feature values processed again.

[0048] Although, for reasons of clarity, this section describes the step of setting initial values of mean and standard deviation after describing the step of calculating feature values, the initial values may alternatively be set even before the data sampling even begins. Setting initial values accordingly could use the judgement of one skilled in the art to determine initial values that allow iterations (see below) but preferably not require too many to deplete time and resources.

[0049] The algorithm determines 58 as follows which feature values were obtained from data sampled when the machine was inactive: A feature value found exceeding the value of the mean of the distribution plus a pre-determined multiple of the standard deviation of the distribution is classified as obtained when the machine was active. A feature value found below that value of mean plus standard deviation multiple is classified as obtained when the machine was inactive. An example standard deviation multiple to use for this embodiment is “two.”

[0050] The unsupervised learning algorithm then processes 60 the feature values considered in step 58 to be from the inactive state using an expectation maximization algorithm to converge to a new mean and the standard deviation.

[0051] Then, the algorithm determines 62 how much the newly-computed values of the mean and standard deviation of the normal distribution (signifying the inactive state) may have changed from the initial values, or alternatively since the last determination/iteration. If they changed more than a pre-determined value/tolerance set by one skilled in the art, the process returns to step 58 to perform another iteration of determining which feature values appear to have derived from data collected during an inactive state and updating the mean and standard deviation accordingly. The iteration continues, until the new mean and standard deviation and preceding mean and standard deviation differ by less than the pre-determined tolerance. At that time, the process ends.

[0052] Attention is now drawn to FIG. 4, which provides a flowchart representing the process of monitoring the active/inactive states of a machine. The embodiment is said to be operating at this stage in the “monitoring mode.”

[0053] The algorithm determines 70 active/inactive states using as a feature value threshold the measured normal distribution's mean plus a predetermined multiple of its standard deviations, these measured values being determined during the unsupervised learning mode, using for example the embodiment described above referencing the flowchart in FIG. 3. A non-limiting example of the predetermined multiple of standard deviations is “2.”

[0054] The algorithm obtains 72 sampled sensor data, for example, from a current sensor, and groups them 74 according to the time increments in which they were measured, the time increments having the same duration as those of the training phase. Then, the algorithm calculates 76 the value of the feature value, for example, the RMS value, and compares that value to the threshold. If 78 the calculated feature value exceeds the threshold, the machine is determined 82 to be active at the time of data sampling for that time increment. Otherwise, the machine is determined 80 to be inactive. This process of sampling, determining feature values of time increments, and comparison with a threshold may continue for as long as the user wants machine active/inactive status information.

[0055] Attention is now called to FIG. 5 showing the statistical distributions that the algorithms of the presently-described embodiments use.

[0056] The normal distribution 92 of the inactive state has a mean 94 and a standard distribution 96. Line 98 indicates the value of mean plus twice the standard deviation. This sum serves as the threshold for decision suitable for use in the above example embodiments.

[0057] Having thus described exemplary embodiments of the invention, it will be apparent that various alterations, modifications, and improvements will readily occur to those skilled in the art. Alternations, modifications, and improvements of the disclosed invention, although not expressly described above, are nonetheless intended and implied to be within spirit and scope of the invention. Accordingly, the foregoing discussion is intended to be illustrative only; the invention is limited and defined only by the following claims and equivalents thereto.