2021 Volume 61 Issue 10 Pages 2580-2586
Determining appropriate guide vane openings (GVOs) of axial blowers under varying industrial conditions is vital for smooth operations in blast furnace ironmaking. This work analyses the influence GVO variations have on outlet air flow rate and pressure by using data taken from operating, industrial blast furnaces, based on which a support vector machine (SVM)-based GVO prediction model is developed. Outcomes reveal that the change status of GVOs, i.e., whether the GVO angle increases or decreases, is critical in determining the relationship between air flow and pressure. By introducing the change status and removing the transition outliers, predictions for the optimal GVOs required to meet the desired air flow rate and pressure in real time can be more accurately determined. The measured values of GVOs range from 0% to 100%, and the SVM-based model developed in this work predicted the GVOs with an RMSE of 0.2480%, significantly improving upon the baseline model which had an RMSE of 0.6047%. The resulting method can provide insights into the operation of complex ironmaking processes, enabling a more efficient adjustment of GVOs.
Axial blowers, or compressors, are integral components of blast furnace ironmaking processes. In order to ensure process stability and energy efficiency, the blowers are required to be adjusted to meet the air flow conditions, which in turn are determined by the operating conditions of the blast furnaces. The blowers also need to have a very high level of robustness since the production efficiency of blast furnaces depends on their performance.
Synchronous motor-driven constant speed axial blowers, which use variable guide vanes to control the volume of air flows and the air pressure, are commonly installed in steel plants to reduce iron oxide to metallic iron for large blast furnaces.1)
A constant, fixed air flow is supplied to the blast furnace in order to maintain stable and smooth operating conditions. In addition to this, blast furnace operations often need to increase or decrease air volume or use different top pressure, which all require the air flow and pressure of the blower can vary in a large range.2,3,4) Adjusting the opening of the guide vane can change the inlet angle of airflow, thus changing the flow rate and controlling the outlet air pressure. The blower guide vanes are designed based on the pre-rotation mechanism of the air flow.5,6) When the air flow rotates along or against the rotating direction of the impeller, the pressure can be reduced or increased accordingly, so the working range of blowers can be expanded to meet the needs of changing working conditions.
In many ironmaking organisations, the guide vane openings (GVOs) are adjusted manually by skilled workers relying on previous, personal experience, which is highly dependent on the individual’s experience and expertise, and is sometimes inaccurate.7) In recent years, under the various national advanced manufacturing strategic plans,8,9) a few large iron and steel manufacturers have begun to implement condition monitoring systems which automatically adjust the opening angles of guide vanes. Nevertheless, these systems are still very basic, and the varying industrial conditions in ironmaking pose technical challenges in determining the appropriate guide vane angles needed to provide the necessary air flow and pressure in real time. There is therefore a need to better understand the relationship GVOs have with air flow and pressure for constant speed blowers in practical ironmaking.
With the rapid development of artificial intelligent techniques, the application of machine learning (ML) algorithms in industrial production without understanding the underlying nature of phenomena involved have been widely used to model highly non-linear and complex systems.10,11) Among machine learning algorithms, support-vector machines (SVM) are prevalent, and have proved to be a powerful tool for solving problems in classification and pattern recognition.12,13,14) SVM aims to minimise the number of sample points by constructing a hyperplane in a high dimensional space. Despite its success in other fields and applications, the method has not received much attention in the field of GVO adjustment.
The conventional simulation methods, such as the Greitzer model or compressor maps, use equations of fluid flow dynamics to describe system performance.15,16,17,18) However, the ML-based models have been proved to be more accurate and faster than the computational fluid dynamics-based simulations.12) Moreover, the air blower operated in practice makes it much more difficult to determine the suitable quantitative expressions as, for a generic value of GVO, there could be various combinations of air flow rate and pressure. A method based on SVM is thus introduced to predict the optimal GVOs.
Furthermore, speed variations have been shown to threaten the stability of non-constant speed blowers by causing temporary stall harmonics.19,20,21) Therefore, in this work we aim to solve the challenge of GVO optimisation for constant speed blowers. The effect of variations of GVOs on the interaction of flow rate and pressure is firstly investigated using the data collected from a real ironmaking process and, based on which, an SVM-based model is developed to select the optimal GVOs in the process of blast furnace ironmaking. The results can provide an effective guide for selecting a suitable GVOs according to the desired air flow rate and pressure performance, so as to improve the efficiency of energy conversion in blast furnace ironmaking process.
Due to the lack of data from practical iron and steel manufacturing processes, most previous work used experimental or simulation data to construct models, which could deliver desired performances only when comprehensive rules and sufficient historical information were available.22)
Collecting data in real manufacturing production processes is a challenging task, as public network access is not allowed for security purposes. Figure 1 illustrates the configuration of the constant speed axial blast furnace blower, in which the location of the sensors for outlet flow rate, pressure and GVO detection is designated (blue arrows). The ironmaking data collected is then transmitted and stored on our on-site private cloud computing platform.
Constant speed axial blower configuration with location of outlet flow rate, pressure and GVO sensors. (Online version in color.)
The architecture of data collection is divided into four main components: sensors, programmable logic controllers (PLCs), a relay server, and the Kubernetes (K8S)-based cluster, as shown in Fig. 2.
Data collection architecture. (Online version in color.)
Analogue signals from sensors on the blowers are transferred to PLCs, which have been ruggedised and adapted for the control of ironmaking processes. Based on the Modbus in PLCs, sensor data is transferred to a relay server using Open Platform Communications Unified Architecture (OPC UA), which is used to enhance compatibility of our architecture with the machinery. An OPC server is installed in the relay server, which transforms the data from the OPC router into a unified format. The standardised data is then pushed through the private cloud computing infrastructure via the Message Queuing Telemetry Transport (MQTT) protocol via an internal switch (IN SW).
The infrastructure of cloud computing is based on containerised DevOps and managed through K8S. A data structure is designed to use MongoDB that stored the sensor data and analyses results. Apache Spark and Kafka are installed to support, extract, transform, and load operations of large-scale data processes in parallel.
In this work, seven days’ worth of monitoring and process data on practical ironmaking is collected from the No. 1 blast furnace blower in Wuhan Iron and Steel (Group) Corporation of China from 25-05-2020 to 31-05-2020. GVOs (%), outlet flow rate (m3/min), and pressure (kPa) of the constant speed axial blower are the primary monitored process variables. The sampling interval is 10 seconds, giving a total of 58420 samples. During the sample period, no serious abnormalities or faults of the blast furnace are detected through analysing the operations and maintenance logs.
GVOs and their counts of samples are shown in Table 1. There are 15 different values of GVO, with range 34.5% to 41.5%, observed.
| No. | GVO (%) | No. of samples |
|---|---|---|
| 1 | 34.5 | 3654 (1574) |
| 2 | 35.0 | 713 (109) |
| 3 | 35.5 | 12451 (3578) |
| 4 | 36.0 | 1439 (594) |
| 5 | 36.5 | 3424 (682) |
| 6 | 37.0 | 1691 (601) |
| 7 | 37.5 | 18168 (5129) |
| 8 | 38.0 | 5767 (478) |
| 9 | 38.5 | 663 (121) |
| 10 | 39.0 | 2369 (705) |
| 11 | 39.5 | 1023 (336) |
| 12 | 40.0 | 363 (142) |
| 13 | 40.5 | 2707 (242) |
| 14 | 41.0 | 2754 (256) |
| 15 | 41.5 | 1234 (449) |
For variable speed axial blowers, rotational speeds are the key component parameters for estimating the corrected mass flow rate and the pressure ratio, and their relations are depicted using performance maps.15)
Motivated by generating the compressor performance map,15,23) the relationship of outlet flow rate and pressure at different GVOs is investigated. Correlations for all study variables are reported in Table 2. The correlation between outlet flow rate and GVO is 0.9607 (p<0.01), much stronger than the correlation of pressure and GVO, and all correlations among the three variables studied are significant. Figure 3 shows a recorded sample of two hours that contains the interaction between air flow rate and pressure at 10 values of GVO ranging from 34.5% to 41.5%. It can be seen that the clusters of the sample points shifted to the right as the opening angle of the guide vane increased, which is consistent with the performance maps of variable speed blowers generated at different rotational speeds.24,25)
| Variable | Outlet flow rate | Outlet pressure | GVO |
|---|---|---|---|
| Outlet flow rate | 1 | 0.4709* | 0.9607** |
| Outlet pressure | 1 | 0.4481* | |
| GVO | 1 |
** p<0.01, * p<0.05
Interaction between outlet flow rate and pressure at different GVOs within two hours. (Online version in color.)
The GVOs do not change substantially over the 10-second intervals and are only recorded with an accuracy rounded to the nearest 0.5%, and so the GVOs typically transition to a value within 1.5% of their current value. The two-hour record of the GVOs from our monitoring system is shown in Fig. 4, and the corresponding variation information is summarised in Table 3. For example, if the GVO changes from 34.5% to 36.5%, the values of current and previous opening are 36.5% and 34.5%, respectively, and the changing status is recorded as “Increased”. Since 34.5% is the initial value of the GVO, there is no previous opening recorded for it.
GVO record sample for two hours.
| Current opening (%) | Previous opening (%) | Changing status |
|---|---|---|
| 34.5 | — | — |
| 35.5 | 36.5 | Decreased |
| 37.0 | Decreased | |
| 36.0 | 37.5 | Decreased |
| 36.5 | 34.5 | Increased |
| 37.0* | 35.5 | Increased |
| 38.5 | Decreased | |
| 37.5* | 35.5 | Increased |
| 36.0 | Increased | |
| 39.0 | Decreased | |
| 38.5* | 37.0 | Increased |
| 40.0 | Decreased | |
| 39.0 | 37.5 | Increased |
| 40.0* | 38.5 | Increased |
| 41.5 | Decreased | |
| 41.5 | 40.0 | Increased |
The changing status of 37%, 37.5%, 38.5% and 40% contained both “Decreased” and “Increased”, and the other five openings only had one kind of changing status. These variations are in accordance with the relationship shown in Fig. 3.
For each GVO from 35.5% to 41.5%, Fig. 5 presents the air flow rate and pressure of “Decreased” (blue circles) and “Increased” (red triangles) status, respectively. It can be observed from Figs. 5(d), 5(e), 5(f) and 5(h), the samples points, which represent the increase or decrease in the openings, are partitioned in two independent clusters. For a specific value of the GVO, the cluster of the “Increased” status is located to the left of the “Decreased” cluster, which indicates that a larger opening should be employed when the guide vane angle is increased to meet the requirement of the flow rate and pressure, and vice versa.
Outlet flow rate and pressure of “Decreased” and “Increased” status at different GVOs: (a) 35.5%, (b) 36.0%, (c) 36.5%, (d) 37.0%, (e) 37.5%, (f) 38.5%, (g) 39.0%, (h) 40.0%, (i) 41.5%. (Online version in color.)
The outliers of each cluster (indicated by dotted ovals) proceed the transitions of GVOs, and since the clusters move to the right as the GVO increased, cf. Fig. 3, it is worth noting that the direction of these points indicate the changing status of the openings. For instance, if the opening angle is reduced, there would be several samples left on the right side of the cluster. While if the opening is increased, the outliers would appear on the left.
The SVM-based multi-class classifier is developed using Scikit-Learn library in Python. The One-to-Rest approach, which splits the multi-class dataset into multiple binary classification problems, is utilised, and the radial basis function (RBF) kernel is selected for mapping the input space to a high-dimensional space based on comparison of the adaptability. The values of cost parameter, c, and gamma, g, are tuned using 10-fold cross validation.26)
As demonstrated in the last section, the previous status of GVOs has a significant impact on the future air flow rate and pressure, and so to improve model performance we add the GVO variation information to the input of the model.
The input parameters, i.e., outlet flow rate and outlet pressure, are normalised before entering the model, while the GVO change statuses, i.e., “Decreased” and “Increased”, are taken to be 0 and 1, respectively. In order to reduce the influence of opening transitions, the sample points collected in the 60 seconds after each change of the GVOs are removed.
After the pair of c and g is identified using 10-fold cross validation conducted on the complete dataset, the five-fold cross validation (i.e., off-line training and off-line testing) is first applied to evaluate the model performance, which means that 80% of data is used for training, and the remaining 20% for independent validation. This process is repeated with randomly selected data splits five times, and the mean accuracy and root-mean-square error (RMSE) are used as the performance indicator.
Then, in order to simulate the real industrial process, the off-line trained model using the data from the first five days is used to predict the GVO for the following two days (i.e., off-line training and online testing), the schema of which is shown in Fig. 6. The two parameters for the RBF kernel are optimised through the 10-fold cross validation on the training set. In the phase of off-line training, the outlet flow rate, pressure, and GVO change status from the historical data are used to train the SVM classifier. For online testing, we cannot obtain the variation information of guide vanes in advance. Therefore, the input GVO change status is estimated by the variation of outlet flow rate, since these two variables are strongly, positively correlated. If the predicted GVO is consistent with the pre-estimated change status the result is applied. Otherwise, the newly predicted value is discarded, and the previous GVO is retained.
Scheme of online GVO prediction OFR: Outlet flow rate; OP: Outlet pressure; CS: Change status.
The GVO prediction results are shown in Table 4. It can be seen that, in comparison with the baseline method, adding the change status, i.e., “Decreased” or “Increased”, as well as removing the transition outliers, make the prediction accuracy increase from 0.7739 to 0.9325 and from 0.7221 to 0.8917 for off-line and online testing, respectively. Besides, the values of RMSE are reduced from 0.5507% to 0.2194%, and from 0.6047% to 0.2480%, respectively.
| Method | Off-line training and off-line testing | Off-line training and online testing | ||
|---|---|---|---|---|
| Accuracy | RMSE (%) | Accuracy | RMSE (%) | |
| Baseline | 0.7739 | 0.5507 | 0.7221 | 0.6047 |
| Removing transition points | 0.7827 | 0.5176 | 0.7349 | 0.5802 |
| Adding change status | 0.9325 | 0.2194 | 0.8917 | 0.2480 |
As indicated in Fig. 5, the change status is a key parameter to significantly improve the GVO prediction performance, since the samples of “Decreased” and “Increased” are seen to be divided into two clusters under the same GVO. Removing the transition points within 60 seconds has subtle influence on the predicted results as the number of outliers is relatively small, and there is no obvious boundary observed between the transitional and normal samples.
As expected, the result of off-line testing is better than that of online testing. For the online testing, the GVO change status is estimated using air flow rate, and some GVO values appearing in the test set may not found in the training set. In our case, the training data from the first five days did not contain the GVO of 35.0% and 38.5%, both of which appeared in the last two days’ record. Hence these samples with unknown GVOs are all assigned the wrong label. Moreover, changing ambient parameters, such as atmospheric temperatures and snort valves openings, could also affect the model performance.27)
Table 5 presents the comparison of the model predicted results and the real value of GVOs using off-line training and online testing. For the baseline, the biggest difference between the predicted and the real GVOs is 3.0%, and 88.64% of the results are within an error of ±1.5%. By adding the change status, the number of misclassified GVOs is significantly reduced. The gap between the predicted and real value narrowed to within ±1.5%, and 67.69% of the misclassified samples are just 0.5% differing from the real value. The corresponding normalised classification confusion matrix is shown in Fig. 7. The proposed improved model achieves accuracies higher than 0.90 for all the GVOs except for 35.0% and 38.5%, the classification accuracies of which are zero as the training set does not contain these two types of GVOs. The predicted GVOs are then displayed in our real-time monitoring system to support decisions on adjusting the real guide vanes in blast furnaces.
| GVO difference (%) | No. of samples | |
|---|---|---|
| Baseline | Removing transition points and adding change status | |
| ±0.5 | 1972 (41.41%) | 1257 (67.69%) |
| ±0.5 to ±1.0 | 1293 (27.15%) | 488 (26.28%) |
| ±1.0 to ±1.5 | 956 (20.08%) | 112 (6.03%) |
| ±1.5 to ±2.0 | 481 (10.10%) | 0 |
| ±2.0 to ±2.5 | 49 (1.03%) | 0 |
| ±2.5 to ±3.0 | 11 (0.23%) | 0 |
Normalised classification confusion matrix of off-line training and online testing: (a) Baseline, (b) Removing transition points and adding change status. (Online version in color.)
In this paper, we elucidate our proposed GVO prediction method for the constant speed axial blowers in the process of blast furnace ironmaking. The present work performs an investigation on the impact of GVO variations on the relationship between outlet flow rate and pressure, and develop a GVO prediction model using a SVM framework.
It has been found that for the same value for GVO, different change directions, i.e., rising from a lower value or falling from a higher value, in conjunction with the transition outliers, are parameters in determining the relationship between air flow and pressure. Accordingly, by adding the change status of GVO and removing the transition points, the performance of the GVO prediction model is significantly improved. The appropriate GVOs are predicted with an accuracy of 0.8917 with an RMSE of 0.2480% for off-line training and online testing. It needs to be noted that for real-time monitoring, it is not known beforehand which the current guide vane change status is (i.e., Increased or Decreased). Consequently, the estimated change status according to the variation of outlet flow rate is employed.
A more sizable data set for training the model and ambient variables can be introduced to further improve the classification performance. Currently, our engineering collaborators are working hard to obtain more production data from their plant, which will then be used to augment our dataset. This work will then be integrated into the ongoing digital-twin assisted system in order to reinforce maintenance of the blast furnace ironmaking process in real time.
The authors gratefully acknowledge the engineers and managers of our partner iron and steel companies for providing the database and the cooperation. This work has been partially supported by the Foundation for the Training of Young Teachers from Shanghai Universities.
