Research On Adaptive Correction Of Soft Sensors And High-temperature Fields Soft Sensing

Industrial processing plants are generally instrumented with a large number of traditional hardware sensors with the primary purpose to collect and deliver data for process modeling, monitoring and control. Although these hardware sensors have been widely used in industry, they have many disadvantages such as time-consuming maintenance, aged deterioration, insufficient accuracy, slow dynamics, large noises and low reproducibility. Meanwhile, some critical variables, such as biomass concentration in bioprocess, cannot be measured online by hardware sensors. As a result, it is crucial to develop soft sensors to infer some critical variables of industrial process.Unfortunately, development of soft sensors is lack of holistic direction by the framework of soft sensor techniques. Furthermore, the estimation performance of soft sensors is deteriorated with the changes of the process characteristics, especially for data-driven modeling approaches. For the time-varying processes, soft sensors need to be updated online to maintain their performance. However, the traditional correction methods update soft sensors whenever new measurements arrive. This blind correction would lead to drastic fluctuations of estimations and large computational load.Temperature is one of the important process parameters which need to be measured and controlled in high-temperature industry. The widely applied technique for temperature measurements is contact thermography technology. But it has slow response and can't online real time measure the temperature distribution for measurement targets. The temperature contactless measurement technology based on colored CCD enables the online real time measurement of temperature fields with a good accuracy. However, the traditional measurement technologies are based on the formulas deduced in terms of idealization and hypothesis which have low accuracy and precision, and are very sensitive to the influence of the locale interference.In the paper, based on the discussion and review of algorithms for soft sensing, the framework of soft sensor techniques has been summarized. The framework proposes the definition, implementation procedure of the soft sensors and the popular algorithms for soft sensing. Then, an approach for accessing, monitoring and maintaining the performance of soft sensors is proposed. Finally, the high-temperature field soft sensing technique based on colored CCD has been developed. The major research contents and results are as follows:(1) A novel performance index of soft sensors is designed to evaluate the performance of soft sensors. Comparing with the traditional performance index based on the prediction errors, the proposed performance index has two main advantages:(i) the influence of offline measurement noises is decreased by weighted mean filter, which could avoid the offline measurement noises to trigger the adaptation mechanism.(ii) the index objectively evaluates the current performance of soft sensors based on their design performance with more clear physical significance. The statistic confidence limit of the performance index is determined by the index series under the normal operation condition. If the index is outside the predefined limit, the soft sensor adaptation mechanism would be triggered. The method can avoid the soft sensor being updated blindly by the traditional soft sensor adaptation method.(2) For diagnosing the types of the process characteristics changes which cause the deterioration of soft sensors, a detection method of the type of process characteristics change is proposed. At first, a series of the performance indexes under the normal operation condition, the gradual change condition and the abrupt change condition are collected, respectively. Then, the amplitude of the Discrete Fourier transforms (DFT) of the series are used as feature vectors to train an operation condition classifier. The state classifier is used to judge the type of process characteristics changes according to the amplitude of the DFT of the indexes series.(3) An adaptive correction of soft sensors using offline measurements and based on the operation condition classifier is proposed. When the adaptation mechanism is triggered, the current operation condition is diagnosed by the classifier. When the process characteristics gradually change, the soft-sensor would be updated recursively. When process characteristics have an abrupt change, the soft sensor would be reconstructed from past data in a neighbor around the query point. The adaptive correction uses the different adaptation methods to the different types of operation condition, which can cope with the changes in process characteristics and achieve high estimation performance with high instantaneity.(4) A new type of two-color thermometry formula is proposed, which contains the coefficients of emissivity change and CCD response bandwidth. Based on the formula, a high-temperature field soft sensor using colored CCD with smog disturbance compensation and adaptive correction mechanism is proposed. The proposed two-color thermometry formula has not any idealization hypothesis, which is deduced by the mechanism of the CCD imaging and radiation temperature measurement theory. The two-color thermometry based on the formula has advantages such as higher precision, reliability and strong adaptability compared to the traditional methods. Furthermore, the measured results interfered by the smog are corrected by a radiant energy attenuation compensator. Finally, the soft sensor would be online updated using the measurement by other thermodetector, which improves the adaptability to different measured targets.(5) In order to decrease the influence of industrial interferences, a high-temperature radiator image recognition method based on the color information of the interference radiator is developed. By segmenting the red and green color images and using the mathematics morphology method, the measured targets can be recognized accurately.(6) An improved mean filter algorithm based on two templates is developed. The edge of the target image and the impulse noises are recognized in the first level of template. Then, the Gaussian noises are filtered in the second level of template. The method could reduce Gaussian noises and Salt and Pepper noises and keep the major details of the original images and clear target image border.(7) Based on the research above, a high-temperature-field soft sensing system is developed. The soft sensing system could online detect the surface high-temperature fields of high-temperature radiators and provide temperature information to operators for analysis and decision. The experimental results show that the measurement range of the soft sensing system could cover the dynamic range of temperature in common high-temperature production. It also has advantages such as high precision, short response time and strong anti-interference ability, which meets the need of the actual industry and has high practicability and promotional value.