Artificial sensor sample generation
A maintenance and monitoring infrastructure can collect real machine samples and train failure analysis models to identify operational anomalies, indicative of a machine failure. Failure analysis models can be trained and more robustly tested and developed by using artificial samples that can simulate what a real monitor would have measured had the real monitor been deployed in a set of selected circumstances (a scenario). A synthetic sample generator (SSG) can include a seed development cycle stage (SDCS) and a synthesis execution stage (SES). The SDCS can generate a signal, from which the SES can generate a multi-phenomena synthetic signal, simulating a scenario, where a monitor would have likely yielded the generated synthetic signal, or a similar signal. The synthetic signal data can be used to train failure analysis models.
1. A method comprising:
receiving a scenario of machine operations, the machine operations impacted by a plurality of phenomena;
determining sources of the phenomena;
gathering training datasets of samples of the machine operations, corresponding to the phenomena;
generating a seed from each training datasets, the seed comprising a statistical representation of an independent and isolated source of a phenomena;
selecting seeds, based on the scenario;
generating single-sourced synthetic samples from a selected seed, and applying one or more physical constraints when generating the single-source synthetic samples to preserve a total energy of the generated single-source synthetic samples relative to a corresponding real signal;
combining the single-sourced synthetic samples, generating multi-phenomena synthetic samples simulating real samples from the machine operations, as impacted by the plurality of the phenomena;
training a failure prediction model with the multi-phenomena synthetic samples;
generating vibration data by an accelerometer of the sensor attached to a machine;
applying the generated vibration data to the trained failure prediction model; and
determining the occurrence of an operational anomaly of the machine by the trained failure prediction model based on the applied vibration data.
2. The method of claim 1 , further comprising:
attaching the sensor to the machine, wherein the sensor comprises:
a housing;
a housing lid;
wireless communication circuitry;
the accelerometer, wherein the accelerometer is configured to measure and transmit both magnitude and spectral data of vibrations of the machine;
a battery; and
a microcontroller configured to reduce battery consumption by controlling the operation of the sensor from a hibernation mode that reduces power to components of the sensor to a normal operation mode in which the sensor obtains the vibration data by the accelerometer;
wherein the housing lid encloses the housing and seals the wireless communication circuitry, the accelerometer, the microcontroller and the battery; and
obtaining the vibration data by the sensor on a scheduled sampling interval.
3. The method of claim 2 , further comprising:
receiving by the sensor, operational parameters of timing and frequency of when and how the sensor should collect data from the machine; and
obtaining the vibration data according to the received operational parameters.
4. The method of claim 1 , further comprising:
applying a physical constraint of phase continuity when generating the single-source synthetic samples to reduce or minimize unnatural discontinuities in the generated single-source synthetic samples.
5. The method of claim 1 , further comprising:
applying frequency constraints to reduce or minimize over-amplification of harmonics in the generated single-source synthetic samples.
6. The method of claim 1 , wherein a total spectral energy of the generated single-source synthetic sample is matched with a total spectral energy of a corresponding real signal.
7. The method of claim 1 , further comprising:
obtaining a discrete Fourier transform (DFT) of samples in the training dataset, the DFT comprising Fourier coefficients;
modeling each Fourier coefficient as a random variable, wherein the random variable is a frequency-dependent distribution of samples in a training dataset;
sampling the modeled Fourier coefficients; and
performing inverse discrete Fourier transform with the sampled modeled Fourier coefficients.
8. A non-transitory computer-storage medium that stores executable program instructions that, when executed by one or more computing devices, configure the one or more computing devices to perform operations comprising:
receiving a scenario of machine operations, the machine operations impacted by a plurality of phenomena;
determining sources of the phenomena;
gathering training datasets of samples of the machine operations, corresponding to the phenomena;
generating a seed from each training datasets, the seed comprising a statistical representation of an independent and isolated source of a phenomena;
selecting seeds, based on the scenario;
generating single-sourced synthetic samples from a selected seed, and applying one or more physical constraints when generating the single-source synthetic samples to preserve a total energy of the generated single-source synthetic samples relative to a corresponding real signal; and
combining the single-sourced synthetic samples, generating multi-phenomena synthetic samples simulating real samples from the machine operations, as impacted by the plurality of the phenomena;
training a failure prediction model with the multi-phenomena synthetic samples;
generating by a sensor attached to a machine, vibration data by an accelerometer of the sensor;
applying the generated vibration data to the trained failure prediction model; and
determining the occurrence of an operational anomaly of the machine by the trained failure prediction model based on the applied vibration data.
9. The non-transitory computer storage of claim 8 , wherein the operations further comprise:
applying a physical constraint of phase continuity when generating the single-source synthetic samples to reduce or minimize unnatural discontinuities in the generated single-source synthetic samples.
10. The non-transitory computer storage of claim 8 , wherein the operations further comprise:
applying frequency constraints to reduce or minimize over-amplification of harmonics in the generated single-source synthetic samples.
11. The non-transitory computer storage of claim 8 , wherein a total spectral energy of the generated single-source synthetic sample is matched with a total spectral energy of a corresponding real signal.
12. The non-transitory computer storage of claim 8 , wherein the operations further comprise:
obtaining a distribution for each training dataset; and
based on the distribution and the samples in the training dataset, generating synthetic samples.
13. The non-transitory computer storage of claim 8 , wherein the operations further comprise:
obtaining a discrete Fourier transform (DFT) of samples in the training dataset, the DFT comprising Fourier coefficients;
modeling each Fourier coefficient as a random variable, wherein the random variable is a frequency-dependent distribution of samples in a training dataset;
sampling the modeled Fourier coefficients; and
performing inverse discrete Fourier transform with the sampled modeled Fourier coefficients.
14. A system comprising one or more processors, wherein the one or more processors are configured to perform operations comprising:
receiving a scenario of machine operations, the machine operations impacted by a plurality of phenomena;
determining sources of the phenomena;
gathering training datasets of samples of the machine operations, corresponding to the phenomena;
generating a seed from each training datasets, the seed comprising a statistical representation of an independent and isolated source of a phenomena;
selecting seeds, based on the scenario;
generating single-sourced synthetic samples from a selected seed, and applying one or more physical constraints when generating the single-source synthetic samples to preserve a total energy of the generated single-source synthetic samples relative to a corresponding real signal; and
combining the single-sourced synthetic samples, generating multi-phenomena synthetic samples simulating real samples from the machine operations, as impacted by the plurality of the phenomena;
training a failure prediction model with the multi-phenomena synthetic samples;
generating vibration data by an accelerometer;
applying the generated vibration data to the trained failure prediction model; and
determining the occurrence of an operational anomaly of the machine by the trained failure prediction model based on the applied vibration data.
15. The system of claim 14 , wherein the operations further comprise:
applying a physical constraint of phase continuity when generating the single-source synthetic samples to reduce or minimize unnatural discontinuities in the generated single-source synthetic samples.
16. The system of claim 14 , wherein the operations further comprise:
applying frequency constraints to reduce or minimize over-amplification of harmonics in the generated single-source synthetic samples.
17. The system of claim 14 , further comprising:
a sensor comprising:
a housing;
a housing lid;
wireless communication circuitry;
the accelerometer, wherein the accelerometer is configured to measure and transmit both magnitude and spectral data of vibrations of the machine;
a battery; and
a microcontroller configured to reduce battery consumption by controlling the operation of the sensor from a hibernation mode that reduces power to components of the sensor to a normal operation mode in which the sensor obtains the vibration data by the accelerometer;
wherein the housing lid encloses the housing and seals the wireless communication circuitry, the accelerometer, the microcontroller and the battery.
18. The system of claim 14 , wherein the operations further comprise:
obtaining a distribution for each training dataset; and
based on the distribution and the samples in the training dataset, generating synthetic samples.
19. A method comprising:
receiving a scenario of machine operations, the machine operations impacted by a plurality of phenomena;
determining sources of the phenomena;
gathering training datasets of samples of the machine operations, corresponding to the phenomena;
generating a seed from each training datasets, the seed comprising a statistical representation of an independent and isolated source of a phenomena;
selecting seeds, based on the scenario;
generating single-sourced synthetic samples from a selected seed;
combining the single-sourced synthetic samples, generating multi-phenomena synthetic samples simulating real samples from the machine operations, as impacted by the plurality of the phenomena;
training a failure prediction model with the multi-phenomena synthetic samples;
attaching a sensor to a machine, wherein the sensor comprises:
a housing;
a housing lid;
wireless communication circuitry;
an accelerometer, wherein the accelerometer is configured to measure and transmit both magnitude and spectral data of vibrations of the machine;
a battery; and
a microcontroller configured to reduce battery consumption by controlling the operation of the sensor from a hibernation mode that reduces power to components of the sensor to a normal operation mode in which the sensor obtains vibration data by the accelerometer;
wherein the housing lid encloses the housing and seals the wireless communication circuitry, the accelerometer, the microcontroller and the battery;
receiving by the sensor, operational parameters of timing and frequency of when and how the sensor should collect data from the machine;
obtaining the vibration data, via the accelerometer, according to the received operational parameters;
applying the vibration data to the trained failure prediction model; and
determining the occurrence of an operational anomaly of the machine by the trained failure prediction model based on the applied vibration data.
20. The method of claim 19 , further comprising:
applying one or more physical constraints when generating the single-source synthetic samples to:
preserve a total energy of the generated single-source synthetic samples relative to a corresponding real signal;
to reduce or minimize unnatural discontinuities in the generated single-source synthetic samples; or
to reduce or minimize over-amplification of harmonics in the generated single-source synthetic samples.