System and method including neural net for tool break detection

• US 5,579,232 A
• Filed: 03/09/1995
• Issued: 11/26/1996
• Est. Priority Date: 03/29/1993
• Status: Expired due to Fees

• Current Assignee: General Electric Company
• Original Assignee: General Electric Company
• Inventors: Martin, Kenneth M., Carmichael, Jerry H., Tong, David W., Diei, Edward N.
• Primary Examiner(s): Gordon, Paul P.
• Assistant Examiner(s): OAKES, BRIAN
• Application Number:
  US08/401,333
• Time in Patent Office: 628 Days
• Field of Search: 364/474.15-474.21, 364/148, 364/551.01, 364/551.02, 364/554, 395/902-906, 395/22, 395/23, 395/50, 395/54, 395/60, 395/77, 340/680, 340/683
• US Class Current: 700/175
• CPC Class Codes: G05B 19/4065 Monitoring tool breakage, life or condition
• CPC Class Codes: G05B 2219/33268 D-A, A-D
• CPC Class Codes: G05B 2219/33296 ANN for diagnostic, monitoring
• CPC Class Codes: G05B 2219/34047 Dsp digital signal processor
• CPC Class Codes: G05B 2219/37245 Breakage tool, failure

Abstract

A system and method for monitoring vibrations of a cutting tool uses a neural network for classifying signal features as break or non-break or, in another embodiment, as non-break or abnormal. A vibration signal is produced by an accelerometer, positioned to sense vibrations at the tool-workpiece interface. The signal is pre-processed to extract low frequency machining noise and detect the energy in a higher frequency band. The signal is then sampled and segments of the digitized signals are processed by digital logic into feature vectors for input to a trained neural net having two output nodes for classification. The use of a neural net provides performance improvement and economies over previously known heuristic methods of signal analysis.

Claims

1. A system for detecting tool break events while machining a workpiece comprising:

means for generating electrical signals representing vibrations at an interface between a tool and said workpiece;

analog signal processing means for producing from said signals, unipolar, signal energy-versus-time analog waveforms;

digital processor means for a) sampling and digitizing said unipolar output signals, b) screening said sampled and digitized signals for alerting spikes, and c) calculating a plurality of feature vectors for digitized signals at, preceding, and following said alerting spikes, said plurality of feature vectors including mean values, slopes of mean values, standard deviations, and slopes of standard deviations; and

a trained neural net for receiving said plurality of feature vectors at an input layer of said trained neural net, said neural net trained iteratively by applying inside data representative of a normal event and outside data representative of an abnormal event, the neural net using an adjustable bias parameter for biasing correct classification of inside and outside data during training and for forming an acceptable class boundary about said inside data and outside data, said trained neural net using said acceptable class boundary to classify said plurality of features as at least one of a break event, a non-break event, and an abnormal event.

2. The system of claim 1 further comprising break indicator means for indicating a tool break event.

3. The system of claim 1 wherein said electrical signals are ultrasonic and low frequency accelerometer signals.

4. The system of claim 1, wherein said trained neural net is classifier net.

5. A method for detecting tool break events while machining a workpiece comprising the steps of;

generating electrical signals representing vibrations at an interface between a tool and said workpiece;

producing from said signals, unipolar, signal energy-versus-time analog waveforms;

sampling and digitizing said unipolar output signals;

screening said sampled and digitized signals for alerting spikes;

calculating a plurality of feature vectors for digitized signals at, preceding, and following said alerting spikes, said plurality of feature vectors including mean values, slopes of mean values, standard deviations, and slopes of standard deviations;

inputting said plurality of feature vectors to an input layer of a neural net, said neural net trained iteratively by applying inside data representative of a normal event and outside data representative of an abnormal event, the neural net using an adjustable bias parameter for biasing correct classification of inside and outside data during training and for forming an acceptable class boundary about said inside data and outside data; and

using said acceptable class boundary for classifying said plurality of feature vectors as at least one of a break event, a non-break event, and an abnormal event.

6. The method of claim 5 further comprising the step of indicating a tool break event.

7. The method of claim 5, wherein said trained neural net is a classifier net.