Sort
Technical CommunicationTool monitoring system is generally composed of signal detection, feature extraction and state recognition. Its key technologies include intelligent sensing, information fusion, signal processing and intelligent learning decision.
I. Intelligent sensing
1.1 Intelligent sensing technology
Sensors are used more and more widely in modern engineering. In tool monitoring, sensors convert physical quantities of the cutting process (such as cutting force) into electrical signals for subsequent processing. The electrical signals collected must reflect the cutting process.
At the same time, in the online detection, the sensor is used with the machine tool, which requires its easy installation and strong anti-interference. Different monitoring signals have different advantages and disadvantages.
The same type of monitoring signal has different sensitivity to tool wear under different working conditions. In tool monitoring, the diversity of cutting, the variability of cutting parameters, the randomness and fuzziness of tool dullness are contradictory with the monotonous monitoring signals. To solve this problem, multi-sensor information fusion is needed to realize tool monitoring.
Multi-sensor information fusion is fundamentally different from single sensor signal processing. Multi-sensor information is more complex and can be integrated at different levels. After fusion, the information is redundant, complementary, real-time and low cost, and the multi-sensor fusion information is more comprehensive and accurate.
1.2 Development trend of intelligent sensor technology
Information acquisition develops from single sensor to multi-sensor, and feature extraction develops from single characteristic value to multi-characteristic value. Multi-sensor information collection has become a trend.
Multi-sensor acquisition of multiple signals, multi-parameter intelligent decision; Develop high sensitivity, compact structure, easy installation, anti-interference sensor; It develops in the direction of multi - sensor information fusion, especially multi - sensor information fusion based on neural network.
Second, multi-sensor information fusion based on neural network
Information fusion requires effective algorithm and strong data processing. Neural network and computer meet these requirements respectively, so information fusion is widely used in tool monitoring. The combination of multi-sensor information fusion and neural network forms a multi-parameter and multi-model system, which has a broad prospect in tool monitoring.
Multi-sensor information fusion based on neural network has the following advantages:
2.1 Information is stored on the connection weight and connection structure of the network, and the form is unified, which is easy to build knowledge base and management.
2.2 Neural network increases fault tolerance. When the sensor fails or the detection fails, the fault tolerance function of the neural network allows the detection system to work normally and output reliable information.
2.3 The neural network has the function of self-learning and self-organization, and can self-adapt to detect the change of environment and the uncertainty of detection information.
2.4 The neural network has a parallel mechanism and can process information quickly to meet the needs of real-time processing.
Three, signal processing
3.1 Signal processing technology
Signal processing is to analyze and process the collected signal, obtain the characteristic value, make decision analysis of the characteristic value, and achieve the purpose of monitoring. Tool monitoring signal processing method is very rich, there are time domain analysis, frequency domain analysis, time frequency analysis, statistical analysis, intelligent analysis, neural network and so on. Traditional signal processing multi - centralized time domain or frequency domain analysis. In recent years, signal processing methods are gradually developing towards time - frequency analysis and intelligence.
3.1.1 Fourier Transform
It is more beneficial to analyze the characteristics and properties of transient and changeable time domain signals by converting them to frequency domain, and Fourier transform is an important method for frequency domain analysis. The discrete Fourier transform (DFT) is discretized in both time and frequency domains and can be analyzed by computer. The fast Fourier transform (FFT) improves the efficiency of the DFT by one to two orders of magnitude. Fourier transform uses spectral characteristics to analyze time domain signals, but it also has its limitations.
3.1.2 Wavelet analysis
Wavelet analysis is a time-frequency analysis method of multi-resolution analysis. It can characterize the local characteristics of signal in both time domain and frequency domain. Its window size is fixed but its shape is variable. Wavelet and wavelet packet can analyze weak faults, and can be used to detect transient anomalies in normal signals.
Continuous wavelet transform is a feature extraction method, which can reflect the features of the original information by extracting the features of the signal on the time-frequency scale, but can not accurately reflect the energy of the signal. The wavelet analysis based on multi-resolution has the variable time or frequency resolution and can accurately reflect the energy of the signal. The wavelet packet analysis based on the signal decomposes the signal into the same bandwidth, end to end band, and has high time-frequency resolution for both high and low frequency.
3.1.3 Generalized adaptive wavelet
Generalized adaptive wavelet analysis refers to the adaptive selection of one or several parameters in wavelet analysis according to the characteristics of the signal and according to a certain algorithm in order to obtain the best analysis results. These parameters include wavelet basis, wavelet decomposition scale, translation coefficient and weighting coefficient.
3.2 Development trend of signal processing technology
From a single feature extraction to a variety of feature extraction direction. Feature extraction of information is developing towards multi-model feature extraction.
Four, intelligent learning decision
Intelligent learning and decision making provide an effective method to solve difficult problems in process monitoring.
4.1 Intelligent learning decision technology
At present, the intelligent technology used in tool wear damage monitoring system includes pattern recognition, expert system and neural network.
4.1.1 Pattern recognition
Nonlinear decision function is the earliest application of pattern recognition in tool monitoring. Fuzzy pattern recognition is a branch of fuzzy mathematics in pattern recognition. Tool wear and damage state is a fuzzy concept, so fuzzy pattern recognition technology can be applied.
4.1.2 Expert System
In tool monitoring, the expert system is mainly used to manage and reason cutting data. However, because it is difficult to acquire knowledge and requires knowledge base, and the cost of establishing knowledge base is high, it is not easy to be accepted by tool monitoring system.
4.1.3 Artificial neural network
Artificial Neural Network (ANN or NN) simulates the human brain nervous system from the perspective of bionics, so that the machine has the perception, learning and reasoning abilities similar to the human brain. The neural network has the following advantages: parallel structure, fusion of multiple signals; Strong knowledge acquisition ability; Associative reasoning and self-adjustment; Nonlinear mapping. In addition, neural network also has many shortcomings: slow learning speed; Local convergence is easy to occur in the learning process. Some fuzzy signals cannot be processed; The structure is difficult to understand.
4.1.4 Wavelet neural network
Wavelet analysis is an effective tool for signal analysis and processing. The neural network can realize the nonlinear mapping between input and output effectively and has the ability of self-learning and pattern recognition.
Therefore, the combination of wavelet analysis and neural network can make use of the good time-frequency localization property of wavelet transform and the self-learning function of neural network, so that it has good approximation and fault tolerance ability.
There are two kinds of combination of wavelet and neural network: loose type and compact type. However, the existing wavelet neural network learning algorithm is mainly gradient descent, and the gradient algorithm convergence is slow.
4.1.5 Fuzzy neural network
Fuzzy Neural Network (FNN) is a neural network based on fuzzy reasoning, which can identify the non-deterministic problems of multi-signal input effectively and accurately characterize the fuzziness of tool wear. FNN is an effective method for complex signal feature recognition. It has fast learning speed, refined algorithm, and can meet the real-time requirements of tool monitoring.
4.2 Development trend of intelligent learning decision-making
4.2.1 Neural network is suitable for application in the integrated environment with large amount of information, high integration and good real-time performance, and pays more attention to unsupervised learning application. Therefore, the development of intelligent learning decision technology will increasingly depend on neural networks.
4.2.2 Fusion of neural network and fuzzy system, genetic algorithm infiltration into neural network.
5. Deficiencies and improvements of tool monitoring system
Tool wear and damage often happen in a moment, the monitoring system is very high requirements. At present, the domestic tool monitoring system has the following deficiencies:
5.1 The software system of tool monitoring has poor operation stability, can not timely analyze and process dynamic characteristic signals, and has poor universality and scalability.
5.2 Parameter extraction method of feature signal needs to be improved.
In the future, the improvement direction of tool monitoring system is: using the same development platform and language, sharing source code, modular development monitoring software, and making it bigger and stronger together. With real-time monitoring and advance prediction as the goal, various signal processing methods are selected reasonably according to the specific situation, and characteristic parameters are extracted.