In recent years, the technology of machine tools diagnostics has made great strides forward. In particular, with the rapid development of artificial intelligence, many powerful intelligent diagnostic techniques have been made available to meet the needs of diagnosing modern complex machine tools.
With industry changes and increased market demands, in addition to the development of high-efficiency and high-precision machine tool hardware in recent years, the developmental focus of machine tool providers has moved towards intelligent machine tools. The development of intelligent software is the most critical technology for machine tool diagnostic among all which is currently provided by software vendors. The most significant development in intelligent machine tools is the monitoring of machine performance. For example, the prediction of the failure of components or performance aging leads to early precautionary maintenance processes which reduces additional costs or loss of credit caused by such failures. While machine is operating, it vibrates and such vibration under normal operation tends to be minimal. With aging, the degree of vibration may increase upto the point of failure. Determining the amount of physical signals for an excellent or extraordinary feature is carried out using the machine’s collection of physical signals followed by an analysis conducted by professionals so as to judge the reasons for the abnormality behind the amount of physical signals. As a result of the various levels of knowledge, experience, and methods of analysis adopted by each expert.
Neural networks and expert systems are two popular methods and have been used independently in practice with varying degrees of success.However, neural networks and expert systems are two quite different approaches to diagnostics. They have different properties as well as advantages and disadvantages with regard to the diagnosis of faults. Neural networks are based on numeric computations and algorithms, while expert systems are based on symbolic and heuristic reasoning. Neural networks have capabilities of association, memorization, error-tolerance, self-adaptation and multiple complex pattern processing. On the other hand, they cannot explain their own reasoning behavior and cannot diagnose new faults (those not already made availablepreviously in training the networks). Expert systems have obvious knowledge representation forms that make knowledge easy to manage. Compared with neural networks, expert systems have the ability to explain their reasoning behavior and can diagnose new faults using knowledge bases. However, self-learning is still a problem and computation time can be lengthy for difficult diagnostics tasks. It is therefore sensible to combine neural networks with expert systems for machine tool diagnostics as the advantages of one approach can outweigh the disadvantages of the other. This scheme can be effective in modern machine tool fault diagnosis.In the majority of cases, neural networks and expert systems were developed separately with one way data passages from neural networks to expert systems, and with interaction between the user and the two models. Some of them permit both components to receive data concurrently.There is a new scheme for integrating neural networks and expert systems is introduced for machine tool fault diagnosis.In this case the two models are completely integrated as a single unit. The integrated system has been implemented and simulated on an existing machining centre.
Machine Tool Diagnosis Based on Expert Systems
Expert systems have been adopted extensively for machine tool diagnosis. They employ expert reasoning methods and computer models to solve problems, and draw similar conclusions to those derived by experts.Using expert systems, diagnostic goals can usually be achieved by specific knowledge representation forms and problem solving techniques, and machine fault sources can be deduced in conjunction with machine conditions.An expert system is composed of a database, a knowledge base and a reasoning engine. When diagnosing a machine tool fault, the reasoning engine carries out reasoning under the guidance of reasoning mechanisms, according to the fault data in the database, and diagnostic knowledge in the knowledge base.
Although expert systems have been applied to machine tool diagnosis with some success, problems have been experienced, predominantly due to the following limitations:
1) Knowledge acquisition is difficult; this acquisition is one of the major bottlenecks in developing expert systems, and requires long periods of time and is costly.
2) It is difficult to completely mimic the human thought process.
3) The knowledge base is never complete; when a new fault is encountered, to which no formal knowledge is available, expert systems become ineffective.
One way of overcoming these limitations is to improve the system’s ability for automatic knowledge acquisition, self-learning, association and memorization. This is where the strengths of neural networks can be incorporated.
Machine Tool Diagnosis Based on Neural Networks
Traditional neural network techniques may be used to diagnose faults in relatively straightforward simple machines with single processes, single faults or gradually occurring faults. However, modern machine tools usually operate multiple processes, and manifest multiple faults, some of which can be catastrophic. Moreover, these machines are complex and highly automated.
, xn]^T is constructed, where xi is the observed value of the i-th feature. These feature values are obtained through signal analysis and processing during the operation of the machine. The detailed procedure includes signal measurement, signal processing and feature extraction. A class set C= {c1, c2…ck} is defined for each given set of machine conditions, and an output space Y= [y1, y2…ym]^T, where ci is the i-th class and yj is the j-th output respectively. A specific class is represented through the output space Y.¼A feature space (input space) X=[x1, x2,
According to the input and output spaces defined above, and the hidden layers and their neurons are selected. The number of input neurons is equivalent to the number of the features selected. The number of output neurons is equal to the dimension of the output space. The parameters in the networks such as weight w^011 are estimated by training, after a set of samples are obtained. In other words, if S={s1, s2…sk} is a sample set of a specific class ci, the neural networks are trained through every pair of input-output (si,ci), where i=1, 2, …, m;
The decision constraint is established automatically after the training process has converged. The neural networks are then able to classify machine conditions for any input xn e X.
The selection of the activation function is also important. The most frequently used function is the sigmoid function, and is given by:
Sigmoid function (Logistic curve)
Wheretis the sum of inputs of a certain neuron and S(t) is the output of the neuron.
To use neural networks to diagnose machine tool faults, there must be a corresponding output neuron for each fault class. Therefore, when a new fault not available previously during the training process is presented, the diagnosis system is incapable of diagnosing the fault. Inaddition, neural networks cannot explain their own reasoning behavior, but this can be supplemented using expert system.
The Integration of Neural Networks and Expert Systems
In developing expert systems for machine fault diagnosis, the acquisition of diagnostic knowledge usually becomes a major bottleneck due to:
The indirect method of interviewing domain experts who would use “common sense” or anecdotal evidence;
Insufficient training, making it difficult for engineers to interview domain experts efficiently and effectively, and for programmers to codify their knowledge;
Little or no expert knowledge available in many scenarios.
Many algorithms or models have been employed to automatically extract diagnostic knowledge from training samples. The automatic generation of knowledge bases saves time, expense and human resources and can be applied to situations with only data but no human expert is available. To improve the accuracy of diagnosis, an approach has been created which automatically generates the knowledge base using neural networks and integrates neural networks with expert systems.When the neural networks are unable to make a diagnostic decision on a new fault because of a lack of data, the meta-system will activate the expert system to diagnose the fault by using deep knowledge. Deep knowledge is defined as knowledge about the machine structure, behavior, and function. In this case, the meta-system acts as a data classifier. At runtime, it classifies incoming data or symptoms and sends the results to the neural networks or expert system for further diagnosis. The final diagnostic result is taken as a fault sample which is used to train the neural network. The learning mechanism is then initiated by the meta-system to extract new diagnostic knowledge from the training sample.
To diagnose fault characteristics accurately in complex modern machines, multiple parameters or signals must be collected and analyzed. Generally these parameters or signals can be classified into three categories:
- Machine and process status, which is collected by a multi-sensor and multi-parameter based condition monitoring system;
- Signals in machine controllers, which can be acquired through an I/O interface between machine controllers and the diagnosis system computer;
- Symptoms observed by machine operators, which are entered to the diagnosis system database through the users’ interface.
Training samples are obtained by representing experiential knowledge in a numerical format. The experiential knowledge is previously acquired from experiences, fault history and theoretical fault analysis provided by maintenance personnel for each machine. When preparing samples those typical fault cases and valid fault signs, which are critical to the success of the diagnostic networks, must be selected.
Error backpropagation (BP) is the most widely used learning algorithm since it is relatively straightforward to implement and, most importantly, often outperforms other methods. Here learning procedure consists of both forward-propagation and back-propagation. During forward-propagation, all information is entered at the input layer and processed at the hidden layers, and finally transferred to the output layer. The status of neurons at each layer only affects the status of those at the next layer. If the expected output cannot be obtained at the output layer, it will conduct back-propagation. During back-propagation, the error from the output layer is transmitted back, through the hidden layers, to the input layer.
Knowledge acquisition and diagnostic reasoning
Both deep knowledge and shallow knowledge (experiential knowledge) are needed in the diagnosis of machinery faults. The former consists of knowledge about the machine structure, behavior, function, fault tree analysis, and parameter/state estimation. The latter is predominantly obtained from engineers’ and maintenance experts’ experience, fault statistics, process history or extracted from the fault samples by neural networks.Deep knowledge is represented as frames or rules, while shallow knowledge is represented as rules or independent facts.A high-quality diagnostic strategy must guarantee the efficiency of diagnosis, and in the meantime achieve good results. From this point of view, the integrated diagnostic scheme combines neural networks and expert systems into one unit, with the former guaranteeing diagnosis efficiency and the latter guaranteeing good diagnosis results.
The author Rajesh Angadi is Bachelor of Engineering, PMP with a Master in Business Administration and is Hadoop Certified. He possesses more than two decades of experience in Information Technology and worked with industry majors like Unisys, Intel, Satyam, Microsoft, Ford, Hartford, Compaq, and Princeton.
