Using fuzzy neural network as a tool for multi-sensor data fusion reduces the cost of the detection system and improves the accuracy of detection. It is applied to the detection part of the self-developed forecasting system for the discharge state of WEDM, and it is proved to be effective and feasible on SCX-II CNC wire-cutting machine. EDM is a non-traditional processing method in which a certain amount of medium is processed by electro-corrosion of pulse discharge between the tool electrode and the workpiece electrode. In order to ensure the normal processing and improve the processing efficiency as well as the fault diagnosis and fault prediction of the machine tool, it must be based on advanced testing links. Among them, the detection of the discharge state of the machining gap is one of the important research contents for fault diagnosis and monitoring. At present, the detection methods of the gap discharge state mainly include radio frequency signal detection, gap average voltage detection, breakdown delay detection, and discharge state time percentage detection. The process is very complicated, and many factors affect the processing precision and the smooth progress of the processing. Experiments show that it is difficult for a single sensor to correctly reflect the processing state, and the development of multi-sensor information fusion is an inevitable path. Sensor information fusion refers to the synthesis of multi-sensor information that has been integrated to form an expression of a certain feature of the environment. the way. The integrated and integrated multi-sensor information has the characteristics of information redundancy, information complementarity, real-time information, and low-cost information. For the specific problem of this paper, it has strong fault fusion and System refactoring capabilities. Information fusion requires strong capabilities and effective algorithms, and computers and artificial neural networks provide strong support for this.
On the basis of considering the characteristics of WEDM, aiming at the information fusion problem in CNC wire-cut machining fault prediction, a kind of information based on peak voltage and peak current in electric discharge machining is proposed, and multi-sensor fusion based on neural network is applied. Technology, which is based on decision-level integration and feature-level integration. At the same time, considering the characteristics of the neural network and the ambiguity of the processing state, the fuzzy system is introduced into the neural network, and the fuzzy neural network method is adopted in the decision-level information fusion system. According to the requirements of processing state monitoring, a reasonable fuzzy neural network model is designed. Then, based on the fusion information, the discharge state is classified and identified according to certain rules. Due to the use of multi-sensor information fusion for target recognition, the limitations of a single sensor can be avoided and the impact of each sensor uncertainty can be reduced. Therefore, the introduction of multi-sensor data fusion technology makes the monitoring module reduce the requirements of the sensing component and the acquisition circuit in hardware, which greatly reduces the system cost. The classifier realized by software can be used directly after being trained, and the classification effect is ideal.
The hardware design of multi-sensor information fusion, the experimental system consists of using the computer to analyze and utilize the discharge gap voltage and gap current. This paper uses a WV332Q1 pulsating DC voltage sensor and a WBI332Q1 pulsating DC current sensor. They measure the gap voltage and gap current in real time and convert them to a standard DC voltage output. It has high precision, high isolation, low drift and wide temperature range. Its response time is 350ms, which can fully meet the pulse power requirements of machine tools. And the input and output circuits are completely isolated, and the output signal is sent to the HY-6040 type A/D capture card to convert the analog quantity into a digital quantity. A fuzzy neural network classifier is constructed by software to realize the identification of three states of WEDM.
The experimental method collects several hundred sets of data for three states of no-load, normal spark discharge and short circuit under the condition of constant power frequency and discharge voltage, and then sequentially changes the power frequency and discharge voltage to obtain the data of each state. , using statistical methods to determine the mean and variance.
Fuzzy neural network design, design method Fast wire cutting process has three basic discharge states: no-load, spark discharge and short circuit. From the voltage and current characteristics of the spark discharge state, it is known that the peak voltage and peak value of the pulse discharge are required to determine the machining state. The two kinds of information of current are combined with each other. If the two sensors are independently identified, there may be cases where a certain discharge state is misidentified or unrecognizable.
In this system, in order to further process the collected information by computer, this paper uses the digital values ​​of peak voltage and peak current obtained by A/D conversion. Firstly, the two feature quantities are pre-processed, that is, the membership values ​​of the three kinds of discharge states (open circuit, spark discharge, short circuit) respectively at a certain time are obtained as the input of the neural network; the output is fusion The voltage and current after the two sensors belong to the membership value of various discharge states. This paper chooses a widely used feedforward artificial neural network as a model, with 6 neurons as the input of the network; the output of the network is 3 neurons, which represent the detected processing state belongs to 3 types of membership. . That is, the fuzzy neural network model is used to realize the nonlinear mapping relationship between peak voltage and peak current and three processing states. At the same time, the fusion of the two sensor information is realized, which overcomes the limitation of the single sensor. The model is shown in Figure 2. In the specific application, through the comparison of several different model structures, the number of hidden layers of the model is determined to be 1, and the number of hidden layer neurons is 8.
The membership function of the membership function and the decision-making principle using the linguistic variable value are generally triangular or trapezoidal. This is because the fuzzy process is insensitive to the shape of the membership function of the linguistic variable, but only has a certain sensitivity to the scope of the membership function. The membership function of the trapezoid is used in this system, as shown in Figure 3. The mathematical formula is as follows: μ is the membership value of the sensor I measured by the j classifier; x is the actual characteristic value measured by the sensor I; x is determined by statistical calculation by experimental data.
For the interpretation decision of the obtained membership value, the following rules are determined: (1) the target category should have the greatest degree of membership; (2) the membership value of the target category must be greater than a certain threshold, which is specified as 0.7. (3) The difference between the target category and the membership values ​​of other categories must be greater than a certain threshold, which is specified as 0.2.
Experimental and experimental results analysis, experimental processing (1) with the selected sensor to monitor the peak voltage and peak current between the wire and the workpiece.
(2) The acquired sensor signal is input to the computer via the A/D conversion card.
(3) The obtained peak voltage and peak current are fuzzified by the determined membership function, providing a standardized form for the input and pattern discrimination of the neural network.
(4) Training phase: According to the obtained membership values ​​belonging to the three states and the corresponding actual state, the training samples are composed; for the determined neural network structure and learning method, the training is repeated until the desired performance requirements are met.
(5) Working stage: Apply the trained neural network to the state online classification, and use the value of the peak voltage and the peak current to be processed by the membership function as the input of the trained network. After the association of the neural network, Obtain the membership value of the three states at this moment; reuse the decision rule to determine which processing state belongs.
Experimental Results Analysis In order to verify the effectiveness of the fusion system, 10 sets of data were selected for neural network training and verification. The first six groups of data were used as teacher samples to train the above-designed fuzzy neural network, and several different learning methods for BP network were compared and compared. The learning method using the momentum and adaptive learning rate was obtained. More satisfactory results. The training steps and the mean squared relationship diagram are shown in Figure 4. Then, using the last four sets of data to verify, the network can not only get the correct conclusion for the discharge state that can be judged by a single sensor, but also make a correct judgment on the discharge state that cannot be correctly judged by a single sensor, and it is satisfactory. result.
(a) BP network error convergence graph (step 22200) (b) Add momentum and BP network error convergence graph using adaptive learning rate (step 6006) From the experimental results, it can be seen that only a single sensor is used to classify the discharge state. Sometimes it is inevitable that the correct conclusion cannot be drawn. At the same time, due to the high frequency of the collected signals, in order to obtain a smaller distortion reduction signal, the hardware facilities of the acquisition system are required to have higher performance, so that the cost of the development system is greatly improved; and the multi-sensor fusion technology can be used to reduce the The requirements of the acquisition system, give full play to the advantages of multiple information complementation, and achieve the system requirements with lower hardware costs. The statistical analysis of the experimental data can ensure the correctness of the design membership function and the process parameters can be selected within a wide range. On this basis, by pre-processing the collected peak voltage and peak current, the training speed of the neural network can be improved, and the required precision can be achieved with only a limited number of training steps. In the experiment, the authors found that the main state of the normal spark discharge and the short circuit state are difficult to distinguish with a single sensor. Through the verification of the above data sets, we can see the ability to apply highly nonlinear mapping of neural networks. As long as the teacher sample is selected properly, the correct classification result can be obtained.
(Finish)
Wire Cut EDM is fast, effective, and can be used to machining virtually any electrically conductive metal material, including Bronze, Copper, Tungsten, Carbon graphite, Carbon steel, Stainless steel, High alloy steel, Hard alloy, Inconel, Kovar, Titanium and many others. Computer-automated CNC equipment and advanced CAD/CAM programming guarantee our Wire Cut EDM to meet even the tightest tolerances.
We have been developed to prove an ideal mix of Speed, Accuracy, Surface Finsh and Low electrode Consumption.
Single Cutting Wire Cut EDM,High Speed Wire Cut EDM,CNC Wire Cutting EDM Machine,EDM Wire Cut
Jiangsu Sailing Intelligent Equipment Co., Ltd. , https://www.sailingcnc.com