面向刀具状态监测的多状态诊断与预测框架（特征学习）

In this paper, a multi-state diagnosis and prognosis (MDP) framework is proposed for tool condition monitoring via a deep belief network based multi-state approach (DBNMS).

For fault diagnosis, a cost-sensitive deep belief network (namely ECS-DBN) is applied to deal with the imbalanced data problem for tool state estimation.

An appropriate prognostic degradation model is then applied for tool wear estimation based on the different tool states.

The proposed framework has the advantage of automatic feature representation learning and shows better performance in accuracy and robustness.

The effectiveness of the proposed DBNMS is validated using a real-world dataset obtained from the gun drilling process.

This dataset contains a large amount of measured signals involving different tool geometries under various operating conditions.

The DBNMS is examined for both the tool state estimation and tool wear estimation tasks.

In the experimental studies, the prediction results are evaluated and compared with popular machine learning approaches, which show the superior performance of the proposed DBNMS approach.

Tool condition monitoring (TCM) has become indispensable to smart manufacturing, automated machining, and other industrial processes nowadays. It not only reduces unnecessary machine downtime and maintenance costs, but also improves the quality and precision of the product. The TCM framework provides diagnostics and prognostics to estimate tool states (e.g. fresh, progressive wear, accelerated wear, worn, etc.) and predict tool wear.

The idea of TCM is to monitor the health condition of the tool continuously using data analytics. Signals such as force, torque, vibration and acoustic emission can be collected and monitored using various sensors mounted on the machinery systems. The data-driven approaches have become a mainstream solution to TCM. They make use of computational intelligence, machine learning or deep learning models that learn from run-to-failure historical data from the system. Such approach can learn the knowledge from data without domain knowledge. Since a perfectly defined physical model of tool wear is not available, data-driven approaches are appealing in practice.

While these conventional methods work in many tool condition monitoring applications, they suffer from two shortcomings. Firstly, the features are manually extracted highly relying on prior domain knowledge. Moreover, the hand-crafted features extracted from one application scenario may not be generalized to other scenarios. Secondly, due to their shallow architectures, conventional NNs have a limited ability of learning complex non-linear prediction in diagnostics and prognostics. We consider that deep belief networks (DBNs) [3, 4] have the potential to overcome the aforementioned shortcomings. DBNs with unsupervised generative feature learning could be able to mine the useful information from raw data and approximate complex non-linear mappings between raw data and the tasks.

There are two main tasks, namely diagnosis and prognosis, dichotomized the prediction process in TCM system. The previous studies have mostly focused on either diagnosis or prognosis in TCM [5, 6]. Diagnosis is to estimate what the current health state is. Prognosis is to predict what will happen next. Prognostics is the study as to show how the tool condition degrades and to estimate the remaining useful life (RUL) of the tool. With effective and reliable estimation of RUL, TCM can reduce overall downtime of the manufacturing processes. Although prognostics plays an important role in TCM, it still a lukewarm research area with few reported studies. In a TCM system, the tool wear estimation forms the basis of tool RUL estimation. In this paper, we would like to focus on tool state estimation as the main diagnostics task and tool wear estimation as the main prognostics task. The performance of prognosis can be improved based on more accurate current health state estimation. Because the degradation trends of the system/components may be different based on different current health states, the results of diagnostics and prognostics are tightly related with the overall performance of the TCM system. Since the distribution of data in different health states are naturally multifarious, any single model is quite hard to handle them. We consider that multi-state diagnosis and prognosis framework distinguishes health states in finer details, that allows us to apply different models according to the diagnostic data attributes.

1. A multi-state diagnosis and prognosis framework. In this paper, a multi-state diagnosis and prognosis (MDP) framework is proposed for tool condition monitoring via a deep belief network based multi-state approach (DBNMS). For fault diagnosis, a cost-sensitive deep belief network (namely ECS-DBN) is applied to deal with the imbalanced data problem for tool state estimation.

2. State-conditional prognostic modeling. An appropriate prognostic degradation model is then applied for tool wear estimation based on the different tool states. The proposed framework has the advantage of automatic feature representation learning and shows better performance in accuracy and robustness.

3. Validation on a real-world gun drilling dataset. The effectiveness of the proposed DBNMS is validated using a real-world dataset obtained from the gun drilling process. This dataset contains a large amount of measured signals involving different tool geometries under various operating conditions. The DBNMS is examined for both the tool state estimation and tool wear estimation tasks.

In this paper, a multi-state diagnosis and prognosis (MDP) framework has been proposed for tool condition monitoring using a deep belief network based multi-state approach (DBNMS). The proposed DBNMS is based on the multiple tool states identified by ECS-DBN that can switch to appropriate prognostic degradation models for prediction. The DBNMS has been applied to tool wear prediction on gun drilling and the experimental studies show that the DBNMS outperforms many popular machine learning algorithms in tool condition monitoring. It has also been shown that the DBNMS is able to generate more accurate and robust prognostic predictions and has good generalization ability over various operating conditions.

To elevate the overall performance of TCM, diagnosis and prognosis are tied in one framework. Due to different data attributes in different health states, a multi-state diagnosis and prognosis framework has been proposed. The proposed MDP framework is one step further towards an unified end-to-end diagnosis and prognosis framework for TCM.

We hope to extend the idea to other conventional data-driven frameworks. Our future work includes the application of multi-objective deep belief networks ensemble (MODBNE) [4] as the degradation model to obtain optimal hyper-parameters for better performance. Other deep learning architectures will also be examined based on the gun drilling real-world experimental datasets to achieve better accuracy in TCM.