CO emission prediction method and system based on robust kernel feature space semi-supervised drift detection

Abstract

The invention provides a CO emission prediction method based on robust kernel feature space semi-supervised drift detection. The method comprises the following steps: calculating a first-order difference component of a historical sample through a pre-constructed historical model to construct a historical data set; calculating feature space statistics of the real-time data, and performing drift detection through the feature space statistics to obtain a drift sample; union set taking operation is carried out on the historical samples and the drift samples, and a new training set is obtained; and performing data updating on the historical model through the new training set, and predicting the CO emission through the updated prediction model. According to the method, through semi-supervised drift detection, the problems of concept drift identification and model dynamic updating under truth value deficiency during CEMS fault are solved, and the CO emission prediction precision and the robustness in a complex industrial scene are improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of urban solid waste combustion, in particular to a CO emission prediction method and system based on robust kernel feature space semi-supervised drift detection. BACKGROUND

[0002] With the acceleration of urbanization, the production of municipal solid waste (MSW) is continuously increasing, and the waste incineration (MSWI) technology has become the mainstream treatment method due to its efficient volume reduction and energy recovery capacity. However, the fluctuation of MSW components, equipment aging and seasonal factors cause the incineration process to present strong nonlinearity and time-varying characteristics, making it difficult for traditional static prediction models to adapt to dynamic working conditions.

[0003] At present, the continuous emission monitoring system (CEMS) is usually used in industrial sites to detect the CO concentration in real time, and supervised or unsupervised methods are used for concept drift detection and model updating. The supervised method relies on the true value provided by CEMS to update the model, such as drift detection based on kernel principal component analysis (KPCA) or robust principal component analysis (RPCA); the unsupervised method only uses the change of the feature space of the process variables to identify drift. Both methods need to build a historical prediction model, and after detecting drift, the model parameters are updated using new samples.

[0004] However, the CEMS system has frequent misalignment and faults, which will cause the supervised method to fail when the true value is missing; and the unsupervised method can only detect feature space drift and cannot verify the actual drift of the output space, which is prone to false positives and makes the prediction accuracy insufficient; in addition, both methods do not have a model updating mechanism when the true value is missing, which will make the system unable to adaptively adjust during critical fault periods, resulting in continuous deterioration of prediction accuracy. Therefore, it is necessary to design a CO emission prediction method and system based on robust kernel feature space semi-supervised drift detection. SUMMARY

[0005] The purpose of the present application is to provide a CO emission prediction method and system based on robust kernel feature space semi-supervised drift detection, which solves the problem of concept drift identification and model dynamic updating when the true value is missing due to CEMS failure through semi-supervised drift detection, to improve the prediction accuracy of CO emission and the robustness in complex industrial scenarios.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A CO emission prediction method based on robust kernel feature space semi-supervised drift detection, comprising the following steps:

[0008] Calculate the first-order difference of the historical samples through the pre-constructed historical model to construct a historical data set;

[0009] Calculate the feature space statistics of the real-time data, and perform drift detection through the feature space statistics to obtain a drift sample;

[0010] Take the union of the historical sample and the drift sample to obtain a new training set;

[0011] Update the historical model through the new training set, and predict the CO emission through the updated prediction model.

[0012] Optionally, the first-order difference quantity of the historical sample is calculated through the pre-constructed historical model to construct the historical data set, including:

[0013] Construct a historical model; the historical model includes: an initial prediction model, a kernel principal component analysis model and a robust principal component analysis model;

[0014] Based on the historical model, the first-order difference quantity is calculated according to the output space true value and the feature space of the historical sample to obtain the historical data set.

[0015] Optionally, the feature space statistics of the real-time data are calculated, and drift detection is performed through the feature space statistics to obtain a drift sample, including:

[0016] Calculate the feature space statistics of the real-time data through the historical model;

[0017] Integrate the real-time data whose feature space statistics reach a preset drift condition to obtain the concept drift sample in the feature space and store it in a cache window;

[0018] When the cache window is filled, it is judged whether the CEMS system has failed, and the concept drift sample in the feature space is processed according to the judgment result to obtain a drift sample.

[0019] Optionally, when the cache window is filled, it is judged whether the CEMS system has failed, and the concept drift sample in the feature space is processed according to the judgment result to obtain a drift sample, including:

[0020] If the CEMS system fails, the concept drift sample in the feature space is pseudo-labeled to obtain a pseudo true value;

[0021] If the CEMS system does not fail, the concept drift sample in the feature space is subjected to a PH test to obtain a drift sample.

[0022] Optionally, if the CEMS system fails, the concept drift sample in the feature space is pseudo-labeled to obtain a pseudo true value, including:

[0023] Construct a first-order difference quantity new set according to the first-order difference quantity and the sample true value;

[0024] Calculate the first-order differential component of the last sample in the cache window and the historical sample based on the new set of first-order differential components;

[0025] Select the differential components closest to the first-order differential component of the cache window through the Euclidean distance;

[0026] Construct an LSTM model according to the differential components, and calculate the output spatial differential component through the LSTM model;

[0027] Calculate the pseudo-true value according to the output spatial differential component.

[0028] Optionally, if the CEMS system does not fail, perform PH test on the concept drift samples in the feature space to obtain drift samples, including:

[0029] Calculate the average measurement error of the cache window;

[0030] Calculate the PH statistic according to the average measurement error and the cumulative error;

[0031] According to the comparison result of the PH statistic and the preset drift threshold, sample screening is performed on the concept drift samples in the feature space to obtain drift samples.

[0032] Optionally, update the historical model through the new training set, and predict the CO emission through the updated prediction model, including:

[0033] If the drift detection does not detect drift samples, predict the CO emission through the historical model;

[0034] If the drift detection detects drift samples, update the historical model through the new training set, and repeat the drift detection with the new training set as the historical sample until no drift samples are detected.

[0035] A CO emission prediction system based on robust kernel feature space semi-supervised drift detection, comprising:

[0036] A historical modeling module for calculating the first-order differential component of the historical sample through the pre-constructed historical model to construct a historical data set;

[0037] A real-time detection module for calculating the feature space statistic of real-time data and performing drift detection through the feature space statistic to obtain drift samples;

[0038] A data updating module for performing a union operation on the historical sample and the drift sample to obtain a new training set;

[0039] A model updating module for updating the historical model through the new training set, and predicting the CO emission through the updated prediction model.

[0040] According to the specific embodiments of the present application, the following technical effects are disclosed: the CO emission prediction method based on robust kernel feature space semi-supervised drift detection provided by the present application, the method comprising: calculating the first order difference quantity of the historical sample through the pre-constructed historical model to construct a historical data set; calculating the feature space statistics of real-time data, and performing drift detection through the feature space statistics to obtain a drift sample; performing a set operation on the historical sample and the drift sample to obtain a new training set; updating the historical model through the new training set, and predicting the CO emission through the updated prediction model. The method solves the concept drift recognition and model dynamic updating problem under the missing true value when the CEMS fails through semi-supervised drift detection, and improves the CO emission prediction accuracy and robustness in complex industrial scenes. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0042] Figure 1 The CO emission prediction method flowchart of the present application;

[0043] Figure 2 The comparison curve graph between the pseudo true value labeling result and the true value of the embodiment of the present application;

[0044] Figure 3 The pseudo labeled true value labeling detection result graph of the embodiment of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0046] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0047] As Figure 1 shown, the present application provides a CO emission prediction method based on robust kernel feature space semi-supervised drift detection, comprising the following steps:

[0048] Step 100: Calculate the first-order differential quantity of the historical sample by the pre-constructed historical model to construct the historical data set;

[0049] Specifically, the historical model includes: an initial prediction model KPCA model and RPCA model Then, the first-order differential quantity Δy of all samples in the historical sample Htd and ΔX Htd are calculated. For example, the calculation formula is:

[0050]

[0051] wherein, Y and Y represent the feature space of the sample, respectively, and represent the output space true value of the sample, and all the first-order differential quantities and the historical sample are integrated into the historical data set.

[0052] Step 200: Calculate the feature space statistics of real-time data, and perform drift detection by the feature space statistics to obtain the drift sample;

[0053] Specifically, the feature space statistics of real-time data are calculated according to the KPCA model and (SPE k ) KPCA and it is determined whether the following criterion is met to determine whether the real-time data has drifted, and the expression of the determination criterion is:

[0054]

[0055] wherein, ζ KPCA represents the KPCA-based discrimination criterion, that is, the value of 1 indicates that the drift has occurred, and the value of 0 indicates that the drift has not occurred, and represent the threshold values for judging whether the values of the feature space statistics and (SPE k ) KPCA exceed the limit value. Meanwhile, the feature space statistics of real-time data are calculated according to the RPCA model and (SPE k ) RPCA and it is determined whether the following criterion is met to determine whether the real-time data has drifted, and the expression of the determination criterion is:

[0056]

[0057] wherein, ζ RPCArepresents the RPCA-based discrimination criterion, where a value of 1 indicates drift, and a value of 0 indicates no drift. The drift samples calculated by the two models are then combined to obtain concept drift samples in the feature space and stored in a cache window. When the cache window is full, a CEMS system failure is determined. If a CEMS system failure occurs, the concept drift samples in the feature space are pseudo-labeled to obtain pseudo-true values. If the CEMS system is not faulty, a PH test is performed on the concept drift samples in the feature space to obtain drift samples.

[0058] Specifically, when the cache window to be marked is filled, When the CEMS system fails, a new set of first-order difference components is first constructed, and the expression is:

[0059]

[0060] in, and for The first-order difference with the last sample in the historical sample, and are the true value of the sample and the input feature vector respectively. Next, calculate the cache window and The first-order difference of is calculated as follows:

[0061]

[0062] Then select the cache window and The set Ω of the nearest ε difference components nearest The expression is:

[0063]

[0064] in, and Represents the εth difference between the input feature and the true value of the sample. Then, the new LSTM model f is established using the above data LSTM(1) , and calculate The output spatial difference component The calculation formula is:

[0065]

[0066] The final calculation The pseudo-truth value of The calculation formula is:

[0067]

[0068] Repeat the above process until all samples in the cache window are labeled with true and false values.

[0069] Specifically, if the CEMS system does not fail, the cumulative deviation between the true value and the false true value of the sample is analyzed to determine whether the sample has concept drift. First, the average measurement error Aveero of the samples in the current cache window is calculated. d , the calculation formula is:

[0070]

[0071] in, and are the true value and false value of the mth sample in the window respectively. Next, calculate the cumulative error when the window is filled for the dth time, and the calculation formula is:

[0072]

[0073] Among them, φ t is the smallest The value is the difference between the current moment and the historical accumulation. is the mean error at the previous t-1 moment, is the cumulative error of the prediction value error Obs(t) up to the current moment. It should be noted that when the cache window is filled for the first time, According to the missing drift judgment basis, φ t Expressed as:

[0074]

[0075] Where φ0 is the cumulative average measurement error of the benchmark, λ=0, when The samples within the judgment window have concept drift.

[0076] Next, by calculating the PH statistic PH t To determine whether Obs(t) is abnormal, the expression is:

[0077]

[0078]

[0079] Among them, φ t is the minimum value of historical accumulation, PH t It is the current moment and φ t Then according to PH t Whether the sample drift occurs is determined by whether the following criteria are met. The expression of the judgment criteria is:

[0080]

[0081] Among them, PHdenotes a PH-based discriminant criterion, is a preset drift threshold, and a value of 1 indicates that drift occurs, and a value of 0 indicates that drift does not occur. Finally, a concept drift sample is output.

[0082] Step 300: Perform a set union operation on the historical samples and the drift samples to obtain a new training set;

[0083] Specifically, when the samples in the cache window are determined as drift samples, the historical samples and the drift samples in the current window are taken in set union to update the historical model as a new training set. For example, the constructed The expression is:

[0084]

[0085] wherein, and are a set of the new training set, and are a feature vector of the sample and a pseudo true value, respectively.

[0086] Step 400: Update the historical model through the new training set, and predict the CO emission through the updated prediction model.

[0087] Specifically, if the drift detection does not detect drift samples, the CO emission is predicted through the initial historical model; if the drift detection detects drift samples, the historical model is updated through the new training set, and the new training set is taken as the historical sample to repeat the drift detection until no drift sample is detected.

[0088] Further, based on The output of the new prediction model of the CO concentration obtained after updating and training the initial prediction model is: The expressions of the updated new KPCA model and the new RPCA model are:

[0089]

[0090] The old model is replaced by the new model, and the historical samples are updated at the same time. The updated sample and model expressions are:

[0091]

[0092] The present application takes the industrial field data of 36 hours of a certain MSWI plant as an example of data source. In this example, the industrial field data is processed by 5-minute scale mean value to obtain 432 samples, and is sorted in time sequence, the first two-thirds of the data is taken as the training set, and the last one-third is taken as the test set. The initial prediction model adopts the LSTM algorithm to construct, the training number is 180, the learning rate is 0.005, and the dropout rate is 0.7. The principal component contribution rate threshold of the KPCA model is 0.89, and 5 kernel principal components are selected as the principal space for drift detection, and the calculated T 2 and the SPE values are 9.4707 and 0.4246 respectively. The principal component contribution rate threshold of the RPCA model is 0.82, and 9 principal components are selected as the principal space for drift detection, and the calculated T 2 and the SPE values are 3.3982 and 38.0613 respectively.

[0093] The capacity of the cache window is set to 3, the number of historical samples selected by the Euclidean distance is 15, and the comparison curve between the pseudo true value label result and the true value is as shown in Figure 2 It can be seen that the change curves of the two are similar, so when the CEMS system fails and the true value cannot be obtained, the pseudo label value is used to detect the concept drift in the embodiment, which can ensure the stability and adaptability of the model in dynamic balance. After detecting the drift samples in the feature space, the true value is labeled by the pseudo label, and then the Page-Hinkley method is used for drift detection, and the detection result is as shown in Figure 3 When the to-be-labeled cache window is filled for 8 times, the error tends to rise when the drift occurs, indicating that the prediction error of the model is increasing in the early stage of drift, but as the model is continuously updated, the error gradually tends to be stable. It shows that the model updating method of the embodiment can adapt the model to the change of data through continuous updating, so as to restore to a stable state, effectively identify the drift in the data, and ensure the stability and accuracy of the prediction model through adaptive updating.

[0094] The present application also provides a CO emission prediction system based on robust kernel feature space semi-supervised drift detection, comprising:

[0095] The historical modeling module is used for calculating the first-order difference of the historical samples by the pre-constructed historical model to construct the historical data set;

[0096] The real-time detection module is used for calculating the feature space statistics of the real-time data, and performing drift detection by the feature space statistics to obtain the drift samples;

[0097] The data updating module is used for taking the union of the historical samples and the drift samples to obtain a new training set;

[0098] A model updating module is configured to update the historical model by using the new training set, and predict the CO emission by using the updated prediction model.

[0099] The beneficial effects of the present application are as follows:

[0100] 1) The pseudo true value is obtained by calculation, and replaces the true CO concentration value when the CEMS system fails, which breaks through the dependence of the traditional supervised method on the true value, and ensures that the drift detection can still be performed during the key failure period;

[0101] 2) The double detection of the feature space and the output space (PH inspection) avoids the misjudgment of the unsupervised method in the output space, and reduces the drift misjudgment rate;

[0102] 3) The method only combines the samples of confirmed drift, reduces redundant calculation, and the adaptive updating iteration of the model makes the error converge quickly, and greatly improves the industrial adaptability of the prediction model.

[0103] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0104] The principle and implementation mode of the present application are described by using specific examples in the present application, and the above embodiment is only used to help understand the method and core idea of the present application; at the same time, for the general technical personnel in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In view of the above, the content of the specification should not be understood as the limitation of the present application.

Claims

1. A CO emission prediction method based on semi-supervised drift detection in robust kernel feature space, characterized in that: The steps include: Calculate the first-order difference component of historical samples through a pre-built historical model to construct a historical data set; Calculating feature space statistics of real-time data, and performing drift detection using the feature space statistics to obtain drift samples; Performing a union operation on the historical samples and the drift samples to obtain a new training set; The historical model is updated with data using the new training set, and CO emissions are predicted using the updated prediction model.

2. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 1 is characterized in that: The first-order difference components of historical samples are calculated using a pre-built historical model to construct a historical dataset, including: Constructing the historical model; the historical model includes: an initial prediction model, a kernel principal component analysis model and a robust principal component analysis model; Based on the historical model, the first-order difference component is calculated according to the output space true value and the feature space of the historical samples to obtain the historical data set.

3. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 1 is characterized in that: Calculating feature space statistics of real-time data and performing drift detection using the feature space statistics to obtain drift samples includes: Calculating the feature space statistics of the real-time data using the historical model; Integrate the real-time data whose feature space statistics meet the preset drift condition to obtain concept drift samples in the feature space and store them in a cache window; When the cache window is full, it is determined whether a CEMS system failure occurs, and sample processing is performed on the concept drift samples in the feature space according to the determination result to obtain the drift samples.

4. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 3 is characterized in that: When the cache window is full, determining whether a CEMS system failure occurs, and performing sample processing on the concept drift samples in the feature space according to the determination result to obtain the drift samples, including: If the CEMS system fails, the concept drift samples in the feature space are pseudo-labeled to obtain pseudo-true values; If the CEMS system does not fail, a PH test is performed on the concept drift samples in the feature space to obtain the drift samples.

5. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 4 is characterized in that: If the CEMS system fails, pseudo-labeling is performed on the concept drift samples in the feature space to obtain pseudo-true values, including: Constructing a new set of first-order difference components according to the first-order difference components and the true value of the sample; Calculating the first-order difference component between the cache window and the last sample in the historical samples based on the new set of first-order difference components; Selecting a plurality of difference components that are closest to the cache window and the first-order difference component by using Euclidean distance; Constructing an LSTM model based on the difference component, and calculating an output spatial difference component through the LSTM model; The pseudo-true value is obtained by calculation according to the output spatial difference component.

6. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 4 is characterized in that: If the CEMS system does not fail, a PH test is performed on the concept drift samples in the feature space to obtain the drift samples, including: Calculating an average measurement error of the cache window; Calculating a pH statistic based on the average measurement error and the cumulative error; The concept drift samples in the feature space are screened according to a comparison result between the PH statistic and a preset drift threshold to obtain the drift samples.

7. The CO emission prediction method based on robust kernel feature space semi-supervised drift detection according to claim 1 is characterized in that: Updating the historical model with data using the new training set, and predicting CO emissions using the updated prediction model, including: If the drift detection fails to detect the drift sample, CO emissions are predicted using the historical model; If the drift detection detects the drift sample, the history model is updated using the new training set, and the drift detection is repeated using the new training set as the history sample until no drift sample is detected.

8. A CO emission prediction system based on robust kernel feature space semi-supervised drift detection, characterized by: include: A historical modeling module is used to calculate the first-order difference component of historical samples through a pre-built historical model to construct a historical data set; A real-time detection module is used to calculate the feature space statistics of real-time data and perform drift detection based on the feature space statistics to obtain drift samples; A data updating module is used to perform a union operation on the historical samples and the drift samples to obtain a new training set; The model updating module is used to update the historical model with the new training set and predict CO emissions with the updated prediction model.

Images