1 Introduction Since the 1990s, modern automatic control systems represented by DCS (Distributed Control System) have been widely used in thermal power units. Thousands of sensor data provide a large amount of information reflecting the operating status of power plant equipment. However, due to sensor failure, drift and various interferences, some bad values ​​may be generated in the measurement data, which will reduce the performance of the system developed based on these data, or even cause the system to fail to work. The inspection of bad values ​​in the measurement data has attracted the attention of scholars at home and abroad [1]. The optimization of measurement data estimates is crucial for the detection of defective data. Currently, there are several methods for obtaining these estimates: 1) Mutual judgment based on hardware redundancy. This method is simple and practical, but requires additional hardware investment and is only applicable to the measurement of some key parameters. Moreover, it cannot accurately locate faulty sensors when the number of redundant sensors is small. 2) Time series prediction models such as AR, ARMA, and Kalman based on the time series relationship of sampled data. However, these models can misjudge sudden changes in process data, and their model parameters require large sampling, which is often not met in actual processes. 3) Metabolic methods with time windows. These methods use finite samples for prediction, avoiding the requirement of large sampling. However, this method can only question defective data and cannot completely verify it, nor can it achieve error correction. 4) Parameter prediction based on mechanistic models using analytical redundancy techniques. However, this requires a sufficiently accurate mechanistic model, and the computational load is large for nonlinear complex systems, sometimes failing to meet real-time requirements [2-4]. [align=left]This paper presents a method for self-correction of bad data based on Auto-associate Neural Networks (AANN). AANN is used to identify the main features and estimate the correlation between process parameters. A parameter prediction model is adopted and the network input data is selected through residual decision logic, effectively avoiding the "residual contamination" problem caused by "outliers" so as to correctly estimate the measurement parameters. [b]2 Auto-associate Neural Networks (AANN) 2.1 AANN Network Structure[/b] Auto-associate neural networks were first proposed by Ballard in 1987 for the coding/decoding problem [5]. Its network prototype is a five-layer feedforward network with a symmetrical topology, as shown in Figure 1. When AANN is applied to the data verification problem, it has obvious physical significance. First, the input data information is compressed through the input layer, mapping layer, and bottleneck layer. The most representative low-dimensional subspace reflecting the system structure is extracted from the high-dimensional parameter space of the network input. At the same time, noise and measurement errors in the measurement data are effectively filtered out. Then, the data is decompressed through the bottleneck layer, demapping layer, and output layer, and the compressed information is restored to the parameter values, thereby realizing the reconstruction of each measurement data. To achieve information compression, the number of nodes in the bottleneck layer of the AANN network is significantly smaller than that in the input layer. Furthermore, to prevent a simple, singular mapping between input and output, all layers except the output layer use nonlinear activation functions. [b]2.2 AANN Sample Selection and Learning Algorithm 2.2.1 Sample Selection[/b] AANN learns the correlation between input parameters. Generally, noise in the measurement data can be considered uncorrelated. Therefore, noisy measurement data can be directly used as both the network input and the training target value, and equation (1) can be used as the target function for network training.[/align][img=272,47]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/153-1.jpg[/img][align=left] Where n is the number of training samples; xi and yi are the network's input and output data, respectively. Due to the presence of noise, the condition for ending network training is not the minimization of E, but the training stops when E converges to a small constant value and tends to remain unchanged. Otherwise, the network will try to learn from the noise, which is called "overlearning", thereby reducing the "generalization" ability of the network. 2.2.2 Improved BP algorithm In this paper, the improved BP algorithm is applied during AANN training. An additional momentum factor is added and an adaptive learning rate is adopted during training [6]. The formula for adjusting node weights is as follows: [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 6] [Image 7] [Image 8] [Image 9] [Image 10] [Image 11] [Image 12] [Image 13] [Image 14] [Image 15 ... [img=346,101]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/153-4.jpg[/img] AANN networks can extract the main information from measurement data and filter out secondary information such as noise. Through data reconstruction, measurement data can be estimated, and therefore can be directly used for measurement data verification, as shown in Figure 2. The process measurement vector x(k) is used as the network input, and the network output vector y(k) is the estimated value of the measurement data. Data verification is then performed based on the estimated residual vector e(k). However, when there are large outliers in the network input parameters, such as large instrument drift or malfunction, the measurement data is obviously unreliable. This will greatly damage the integrity of the network input data. If this measurement parameter is still used as the input value, it will be incorrectly propagated to other parameters, causing so-called "residual contamination," and the correct results cannot be obtained. 3.2 Self-Correction Verification Method for Measurement Data Based on AANN To overcome the shortcomings of conventional verification methods, this paper proposes a self-correction verification method for measurement data based on AANN, as shown in Figure 3. [img=338,190]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/153-5.jpg[/img] In the figure, A is the residual generation module, which generates residuals using the process measurement parameter x(k), the AANN output parameter y(k), and the single-parameter prediction model output parameter [img=26,15]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/153-6.jpg[/img]; T is the network input parameter selection module. During data verification, the measurement data x at the current time is first... (k) is used as input to AANN to obtain the corresponding output data y(k) of the network. Then, based on the historical output data of the network, the predicted values ​​of each parameter are obtained through a single-parameter prediction model [img=26,15]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/153-6.jpg[/img]. Then, the residual decision logic is used to determine whether there is large abnormal data. If there is large abnormal data in the i-th measurement parameter, the corresponding measurement data x is transferred through the T module. Replacing i(k) with [img=36,25]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-3.jpg[/img], the network's input parameters are re-formed [img=34,22]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-1.jpg[/img], resulting in a new network output. [b]3.2.1 Single-Parameter Prediction Model[/b] In a continuously sampled system, the measurement parameter sequence itself is also a discrete random time series. Similarly, the output after AANN processing also has certain temporal characteristics; therefore, time series prediction methods can be used to make preliminary predictions for single parameters. Based on the predicted value, large outliers of a single parameter can be questioned. If the measured parameter and the predicted parameter are not significantly different, the measured parameter can be used as the input parameter of the network. Conversely, a more reliable predicted value than the measured parameter can be used as the network input to reduce the impact of large outliers on the network output data. 3.2.2 Residual Decision Logic For the same parameter, in addition to the measured data, there are also the estimated value of AANN and the estimated value of the single-parameter prediction model. There are certain residuals between these data. (1) When the residual between all the measured data and the output data corresponding to AANN is less than a certain threshold, that is, when equation (6) holds, it indicates that there is no bad data. [img=328,27]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-2.jpg[/img] where n is the number of measured parameters; δi is the corresponding threshold. (2) When equation (6) does not hold, take [sub][img=104,30]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-ss.jpg[/img][/sub] , j=1,2,...,n, where σj is the variance of the j-th measurement parameter. If equation (7) does not hold, then there is a large deviation in the i-th measurement parameter. Replace xi(k) with [img=36,25]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-3.jpg[/img] to form new network input data [img=34,22]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-1.jpg[/img], and recalculate the network. If equation (7) holds, it indicates that the deviation in the measurement parameter is not significant, and the AANN output can be considered as an estimate of the measurement data. [img=380,56]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-5.jpg[/img] In the formula, αi is the detection threshold, and αi>δi. 4. Example [img=380,130]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-6.jpg[/img] Taking the regenerative system of a 200 MW unit in a power plant as shown in Figure 4 as an example, based on the specific layout of the measurement points in the on-site DCS system, the measurement data of the measurement points listed in Table 1 are used as the research object. Within the normal operating range of the unit, the operating parameters corresponding to each measurement point were collected through DCS, with a sampling interval of 30s, for a total of 400 samples, and they were standardized. The first 200 samples were used as network training samples, and the last 200 samples were used as test samples. During data verification, the AANN network structure was 5-8-3-8-5, and the improved BP algorithm was used for training in this processing mode. The error curve is shown in Figure 5. [img=357,134]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-7.jpg[/img][img=342,211]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/zgdjgcxb/2002-6/154-8.jpg[/img][align=left] To verify the effectiveness of the algorithm in this paper, taking the flow measurement point F2 as an example, bad data with constant slope drift as shown in Equation (8) were superimposed on the test data. Figure 6 shows the results of data testing using conventional methods, and Figure 7 shows the test results of the method proposed in this paper. In both figures, (a) represents the measured, estimated, and true values ​​of F2, and (b) represents the squared values ​​of the estimated residuals of each parameter, that is, the squared difference between the measured and estimated values ​​of each parameter. [Image 1] [Image 2] [Image 3] [Image 4] [Image 5] [Image 6] [Image 7] [Image 8] It is evident that the AANN network can accurately reflect the poor data present in the measurement data. However, Figure 6 clearly shows the existence of "residual contamination," which causes the estimated values ​​of other measurement data to also have relatively large reconstruction residuals. The method presented in this paper (as shown in Figure 7) can effectively suppress this phenomenon. Figure 8(b) shows the changes in the sum of squares of the estimated residuals of each parameter before and after applying the self-calibration method, while Figure 8(a) shows the test results without using the self-calibration method. Although the sum of squares of the residuals corresponding to F2 increases significantly, the residuals of other measurement data also increase unreasonably, which obviously cannot guarantee the estimation accuracy of these measurement parameters. After applying the self-calibration method in this section, as shown in Figure 8(b), only the residuals of the data corresponding to the faulty instrument change significantly, while the residuals of other measurement data do not change much. 5 Conclusion The accuracy of measurement data is crucial to the power plant monitoring and optimization system. Without the accuracy of the data, any research and monitoring algorithm is impractical. This paper proposes a self-calibrating data verification method based on AANN networks. By processing and selecting the input data of the neural network, it improves the accuracy of online applications of neural network methods and avoids "residual contamination" during data verification. This allows for the correct detection and location of defective values ​​in the measurement data, as well as the accurate reconstruction and estimation of defective data. The effectiveness of the method is verified through actual field data collection, demonstrating its significant practical value.