Data-Driven Approaches With Mechanistic Analysis For Fault Diagnosis Of Cold Rolling Continuous Annealing Processes

The continuous annealing process line (CAPL) is an important unit for producing high-quality cold-rolling strip products. The line is composed of payoff reel, welding machine, entry looper, furnace zone, delivery looper, temper mill, etc. The raw material is cold-rolling steel strip, while the products are tinning black/galvanized sheets. The strip is heated up and cooled down when it is transferred by carrying rolls to goes through each furnace section in order to improve its microstructure, plasticity, and metal-forming property. For stable operation of the CAPL at a high speed, it is desired to maintain thousands of meters of strip tensions in the furnace zone in normal control limits. To achieve this, more than ten tension meters are installed and hundreds of motors are partitioned into several groups to achieve closed-loop tension control. However, unmeasured tensions between the tension meters may deviate from the normal range, i.e. faulty tension situation occurs when there is unexpected speed variation caused by large disturbances induced by transportation of defective material, or when there is roll slippage which makes it difficult to regulate the speed of strip around the slipped roll. The strip is stressed to break when the strip tension exceeds the ultimate tensile stress, and roll slippage occurs when the strip tension is too low, or the tension difference between two sides of one roll is too large. The faulty tension situation may make the strip fold or transverse buckling. A strip break typically results in more than 24 hours loss of production time. Roll-slippage can make the strip be scratched on the surface. Accordingly, it is significant and necessary to monitor the faulty tension situation, and diagnose the strip-break and roll-slippage faults from an engineering application point of view. Furthermore, the study contributes to the academic research of fault diagnosis for complex large-scale industrial processes.The objectives of this dissertation are to find feasible solutions of monitoring of tension-relevant faulty situation and diagnosis of typical process faults (i.e., strip-break and roll-slippage faults) in the CAPL. Several novel data-driven approaches with mechanistic analysis for fault diagnosis of the CAPL are proposed and application research is carried out in the work. The main technical contributions can be summarized as follows:i) The tension model of the CAPL is established considering the process characteristics. The inertia parameter is composed of the inertias of motor and roll, and inertia of strip as well. Tension variations are induced due to thermal expansion and contraction of the strip when it is heated up and cooled down in the furnace. The frictions are not only related to bearing friction but also related to gas flow. By detailed analyses, the tension variations and frictions are difficult to be described by precise mathematical model. Thus a hybrid tension estimation and fault diagnosis method is proposed. The tension estimation model is composed of two parts, i.e. main model and error compensation model. The main model is built by applying an observer-based method. The error compensation model is achieved by applying. neural networks principal component regression (NNPCR) algorithm. Based on the estimated tensions, a fault diagnosis method is then designed to diagnose faulty tension situations by applying rule-based method.ii) The large scale sequential processes in three-level structure achieve control objective by partitioning a couple of control loops to regulate' some process variables collected from the same sampling locations. This makes it difficult to evaluate the fault effects on different locations within one control loop by applying traditional two-level consensus principal component analysis (CPCA) approach. Thus, a novel three-level principal component analysis (TLPCA) based fault diagnosis methodology is proposed. Theoretical analysis result demonstrates the equivalences of the loading, scores and residuals between the TLPCA and principal component analysis (PCA). The TLPCA algorithm is not simply an ad hoc partition of the single CPCA block loadings, although they are shown to be equivalent under a specific scaling rule. This feature makes it enhance the interpretation performance of the TLPCA compared to CPCA. As the CAPL arranges the process variables in three levels, i.e. section level, loop level and roll level, the proposed TLPCA-based fault diagnosis method is applied to analyze the strip-break fault in the large-scale CAPL using the historical normal and historical strip-break fault data sets. First, three statistical indices (i.e. overall section index, block index and sub-block index) are defined for overall section, each loop and each roll respectively, with the corresponding control limits of them derived from the historical normal data. Secondly, three-dimensional block index plot and three-dimensional sub-block index plot are obtained using the block indices and sub-block indices computed from the historical strip-break fault data. By detailed analyses from the figures, the strip-break feature which cannot be obtained using the CPCA-based method is observed, i.e. the propagation of abnormalities along the sampling location appears for the sub-block plot when the materiel travels to downstream.iii) Considering the fault smearing problem for traditional CPCA method, the univariate reconstruction-based contribution is extended to reconstruction based block contribution (RBBC) to diagnose faulty block of large-scale processes. Block reconstructability and diagnosability conditions are derived, and simulation results demonstrate that the RBBC method displays improved correct diagnosis rate compared to CPCA. The proposed RBBC based method is combined with PCA to diagnose the faulty block for the CAPL. For the diagnosed faulty block, the strip-break and roll-slippage faults are diagnosed respectively:a) finite state machine and strip-break feature deduced from the application result of TLPCA-based method discussed in Section ii) are combined to diagnose the strip-break fault, and reconstruction based variable contribution is applied to diagnose the location where the strip breaks; b) the reconstruction based fault diagnosis method for the block indices is applied to diagnose the roll-slippage fault and location of the slipped roll within the faulty block based on roll-slippage feature which is represented by fault direction and exacted from historical roll-slippage data.iv) Experimental study is carried out using the operation data collected from a real CAPL with 11 furnace sections and 133 carrying motors in a large-scale iron and steel company. The application results demonstrate that our proposed methods can effectively diagnose faulty tension situations, and strip-break and roll-slippage faults. The locations of the strip-break and the slipped roll which are difficult to be determined by the practitioners are also diagnosed successfully.