A Desulfurization Prediction Data Preprocessing Method Based on Box Plot-Knowledge-Experience Dual-Driven Approach

Abstract

This invention relates to the field of artificial intelligence technology, specifically to a desulfurization prediction data preprocessing method based on a box plot-knowledge-experience dual-drive approach, comprising the following steps: S1, collecting and cleaning key parameters such as flue gas flow rate, temperature, pH value, and current, and verifying the data type and numerical rationality; S2, constructing derived features and fusing laboratory and inspection data to form multi-source data; S3, setting multiple anomaly counters and detecting outliers based on box plots; S4, using expert rules to label semantic tags and identify consistency of operating conditions; S5, counting missing, mutation, and reuse times and updating abnormal states; S6, performing Fourier denoising and Z-score standardization. This invention achieves high-precision, robust, and highly intelligent desulfurization data preprocessing, effectively improving data quality, system stability, and engineering applicability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a desulfurization prediction data preprocessing method based on a box plot-knowledge-experience dual-drive approach. Background Technology

[0002] With the widespread application of desulfurization prediction models in coal-fired power plants, data preprocessing, as a fundamental step in model construction and operation, is becoming increasingly important. As my country's energy structure and power production methods are undergoing rapid transformation, coal-fired power units need to simultaneously address tasks such as ensuring power supply and peak shaving, and gradually adapt to complex operating conditions such as deep peak shaving and low-load operation. Against this backdrop, the operation and management of desulfurization systems face higher requirements for intelligence and automation.

[0003] Currently, desulfurization control in coal-fired power plants mainly relies on manual intervention and experience-based judgment, resulting in a low level of automation and difficulty in effectively coping with the complexity and dynamic changes in operating conditions. Taking the limestone-gypsum wet desulfurization system as an example, its key parameters, such as the pH value of gypsum slurry and the SO2 concentration in flue gas, have significant time delay, nonlinearity, and strong coupling characteristics, making it difficult for traditional control strategies to achieve accurate modeling and dynamic adjustment. At the same time, with the increase in sensor deployment density, the large amount of multi-source heterogeneous data generated by the system also has significant deficiencies in terms of quality, structure, and temporal consistency, which seriously restricts the stability and generalization ability of the prediction model. Summary of the Invention

[0004] This invention provides a desulfurization prediction data preprocessing method based on box plot and knowledge experience dual-drive. By constructing a preprocessing process with the capabilities of anomaly detection, data repair, noise suppression and semantic recognition, it ensures that the model input data has a good foundation in terms of accuracy, stability and expression consistency, thereby improving the reliability and applicability of the desulfurization prediction model and providing key data support for the intelligent operation and green transformation of coal-fired power units.

[0005] The desulfurization prediction data preprocessing method based on box plot and knowledge-experience dual-drive includes the following steps:

[0006] S1, Data Acquisition and Cleaning: Real-time acquisition of various key parameters during the operation of the desulfurization system, including flue gas flow rate, flue gas temperature, gypsum slurry pH value, and circulating pump current. At the same time, cleaning operations based on data integrity and logical consistency are performed on the acquired key parameters, including verification of data type validity (e.g., whether it is floating-point data) and verification of numerical rationality (e.g., whether there are abnormal cases where non-negative characteristic values ​​are less than zero).

[0007] S2, Multi-source data construction: Process and reconstruct the key parameters after the cleaning operation, construct derived features through computational fusion, logical combination or statistical derivation, and integrate auxiliary information from laboratory test data and manual inspection data to supplement external feature parameters closely related to the desulfurization process.

[0008] S3, Box Plot Test: For the constructed multi-source data, a missing value counter, a reused value counter, a mutation value counter, and an anomaly status device are set to comprehensively record and identify various abnormal states. At the same time, based on the statistical principle of box plots, quartile intervals of features are constructed according to the working condition category, and potential outliers are identified by using the upper and lower limits, so as to realize the detection of outliers and anomaly identification of multi-source data.

[0009] S4, Knowledge and Experience Verification: Integrating expert experience and prior knowledge to perform multi-dimensional anomaly identification and operating condition consistency verification on multi-source data, the relationship between key operating parameters, equipment status and typical operating conditions is expressed in a structured way, and semantic labeling is performed on real-time multi-source data streams with the help of rule engine or feature combination logic, so as to realize the identification and early warning of abnormal operating conditions, boundary behavior or potential failure modes.

[0010] S5, Anomaly Statistical Validation: Dynamically monitors and identifies the statistical indicators of missing values, mutation values, and reused values ​​in multi-source data;

[0011] S6, Data Denoising and Normalization: Fourier transform noise reduction is performed on multi-source data. At the same time, Z-score normalization strategy is adopted to unify the scale and regularize the distribution of the noise-reduced multi-source data.

[0012] Optionally, the data acquisition and cleaning in S1 includes:

[0013] S11, Sensor Deployment and Data Acquisition: Multiple types of industrial-grade sensors are deployed in the desulfurization tower process, including flow sensors, temperature sensors, pH meters, pressure sensors and current transformers. The sensors are connected to the data acquisition terminal through signal acquisition cards to realize continuous acquisition of real-time data.

[0014] S12, Data type verification mechanism: The key parameters collected are automatically verified using a regular expression matching strategy. If the key parameters collected conform to the floating-point format, they are converted and stored in the database. If they do not conform, they are marked as abnormal, triggering an alarm and recording the error.

[0015] S13, Data Physical Constraint Verification and Anomaly Correction: Combining the engineering attributes and physical laws of variables in the desulfurization process, a characteristic dimension rule base is constructed to determine the reasonableness of the range of key parameter values. For key parameters that should be non-negative, if an outlier value less than zero appears, it is corrected to the minimum value (such as 0).

[0016] Optionally, the multi-source data construction in S2 includes:

[0017] S21, Multi-point data fusion processing: A dynamic fusion algorithm based on variance discrimination is used to process the same key parameter. When the variance of the measured value of the key parameter is less than the preset threshold, the average value is taken as the fusion result. When the variance exceeds the preset threshold, the two sets of data with the smallest difference are selected for weighted average.

[0018] S22, based on the desulfurization process, a variety of derived variables are calculated based on the collected key parameters. Among them, the total sulfur in the flue gas is calculated by multiplying the total air volume and the original flue gas SO2 concentration, and the sulfur content is calculated by the ratio of the total sulfur to the total coal volume.

[0019] S23, Heterogeneous Data Integration and Structured Conversion: Introducing external data from laboratory testing and manual inspection, and performing unified format conversion and structured encoding processing on text, enumeration, and numerical information.

[0020] Optionally, the box plot verification in S3 includes:

[0021] S31, Eigenvalue stability detection: For cases where multi-source data remain unchanged within a continuous time window, determine whether there is a data acquisition failure or malfunction.

[0022] S32, Missing value identification and imputation: For cases where missing values ​​exist in multi-source data, the missing event is considered to trigger the compensation mechanism;

[0023] S33, Temporal Abrupt Change Detection: The change trend of continuous temporal data in multi-source data is judged by the first-order difference jump detection method. If the continuous difference value exceeds the abrupt change threshold, it is regarded as a temporal abnormality. The abrupt change value counter is incremented by 1 and replaced with the value of the previous time step. At the same time, the reuse value counter is incremented by 1.

[0024] S34, Box Plot Outlier Detection: The box plot method is used to identify outlier data points. For multi-source data, the first quartile Q1, the third quartile Q3, and the interquartile range IQR are calculated. Based on the box plot principle, a normal value interval is constructed, namely [Q1-k×IQR, Q3+k×IQR], where k is the anomaly sensitivity coefficient. If a data point in the multi-source data exceeds the range of this interval, it is determined to be an outlier.

[0025] Optionally, the eigenvalue stability detection in S31 includes:

[0026] S311, Set monitoring time window: Set the time window threshold. And extract the sequence of continuous observations from the multi-source data within this window;

[0027] S312, Determine if the feature value is constant: Determine whether the multi-source data remains unchanged within the time window; if the value is equal at all times, it is considered a constant state;

[0028] S313, Mark Abnormal Status: When the multi-source data remains constant within the window, set the first bit of its abnormal status device from 0 to 1 to mark it as having sensor lag or data freeze abnormality.

[0029] Optionally, the missing value identification and imputation in S32 includes:

[0030] S321, Record missing events: When it is detected that the observation value of multi-source data is empty at the current time, increment the corresponding missing value counter by 1 and record the frequency of missing events;

[0031] S322, Perform missing value imputation: Automatically imput the current missing value using the valid observation value from the previous time step;

[0032] S323, Update the reuse counter: After the filling operation is completed, increment the reuse value counter by 1 to record the number of times the compensation action is executed.

[0033] Optionally, the knowledge and experience verification in S4 includes:

[0034] S41, Anomaly identification based on empirical rules: Based on the operation mechanism and engineering practice of the desulfurization system, a physical range is set for the key parameters. If the collected key parameters exceed the range, they are regarded as anomalies and replaced with the valid value of the key parameters at the previous moment. At the same time, the reuse value counter is incremented by 1 to record the correction behavior.

[0035] S42, Operating Condition Semantic Tag Labeling: Utilizing a predefined set of operating condition identification rules, combined with a rule engine, multi-source data is matched in real time. Based on the combination logic of multi-source data, corresponding operating condition semantic tags are automatically assigned to achieve the identification of operating status.

[0036] Optionally, the anomaly check in S5 includes:

[0037] S51, Missing Value Statistics and Anomaly Detection: Set the sliding time window length threshold and the upper limit threshold for missing values, and count the number of missing values ​​in each multi-source data within the window in real time. If the cumulative number of missing values ​​exceeds the upper limit threshold for missing values, it is considered a missing value anomaly. The second bit of the anomaly status device is set from 0 to 1, and an early warning is triggered.

[0038] S52, Mutation value statistics and anomaly detection: Based on the results of the first-order difference, the number of occurrences of mutation events is counted. If the number of mutations exceeds the threshold of the sliding time window length within the set window, it is marked as a fluctuation anomaly, the third bit of the anomaly status device is updated to 1, and the alarm process is initiated.

[0039] S53, Reuse Value Statistics and Anomaly Detection: Monitor the frequency of multi-source data being reused by the previous data. If the number of reuses exceeds the time window and number threshold of the reuse behavior within the window, it is judged as a reuse anomaly. Update the corresponding bit of the anomaly status device and issue an early warning, reflecting that there is a problem with equipment lag or abnormal data processing.

[0040] Optionally, the data denoising and normalization in S6 includes:

[0041] S61, Fourier Transform Noise Reduction: To address high-frequency interference in multi-source data, the Discrete Fourier Transform (DFT) is used to transform the time series to the frequency domain. Based on the amplitude spectrum distribution, a cutoff frequency threshold is set. High-frequency noise components are filtered out and then restored to the denoised time-domain signal through inverse transformation.

[0042] S62, Z-score normalization: Performs Z-score normalization on the denoised multi-source data.

[0043] The beneficial effects of this invention are:

[0044] This invention identifies and removes outliers using box plots, and compensates and corrects data for specific operating conditions using expert experience rules, thereby improving the authenticity and consistency of the data. At the same time, it introduces Fourier transform noise reduction technology to effectively suppress measurement noise and short-term fluctuations, enhance the stability and reliability of the data, and provide high-quality input for subsequent processing.

[0045] This invention, by considering the coupling characteristics between variables and the switching features of operating conditions, and combining historical data and real-time monitoring results, dynamically evaluates data quality and implements anomaly repair and sequence smoothing, thereby improving the stability and adaptability of the preprocessing process under varying operating conditions and ensuring the accuracy of control strategies and adjustment responses.

[0046] This invention reduces manual intervention and improves the intelligence and efficiency of desulfurization data processing by constructing a standardized and automated data preprocessing workflow. The method has good versatility and scalability, and is applicable to various operating loads and conditions, providing effective support for the refined management and low-carbon operation of desulfurization systems in coal-fired power plants. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1This is a schematic diagram of the preprocessing method according to an embodiment of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0050] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0051] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0052] like Figure 1 As shown, the desulfurization prediction data preprocessing method based on box plot-knowledge-experience dual-drive includes the following steps:

[0053] Step 1, Data Acquisition and Cleaning: Key parameters of the desulfurization system are collected in real time, and preliminary cleaning is performed based on data integrity and logical consistency rules to ensure the usability and accuracy of the raw data.

[0054] Specifically, it includes:

[0055] 1. Sensor Deployment and Data Acquisition: Various types of industrial-grade sensors, including but not limited to flow sensors, temperature sensors, pH meters, pressure sensors, and current transformers, are deployed at key process stages of the desulfurization tower. These sensors are connected to a data acquisition terminal via signal acquisition cards to achieve continuous real-time acquisition of operating data. The acquired data is transmitted to the data acquisition module via industrial communication protocols and simultaneously written into the database for storage, serving as the input source for subsequent data processing.

[0056] 2. Data Type Validation Mechanism: To address the issue of data type consistency in the collected data, a regular expression matching strategy is used to automatically validate the data format of each field. The specific process includes: constructing a regular expression template for matching floating-point formats, comparing each input data entry; if a match is successful, performing type conversion and storing the data in the database; if a match fails, it is considered abnormal data, and the system will trigger an alarm mechanism and record the data entry for manual verification, preventing illegal formats from interfering with the calculation model.

[0057] 3. Data Physical Constraint Verification and Anomaly Correction: Combining the engineering attributes and physical laws of variables in the desulfurization process, a characteristic dimension rule base is constructed to determine the reasonableness of the range of key parameter values. For example, variables such as slurry pH value and flue gas flow rate should theoretically not be negative. The system automatically detects parameter values ​​according to preset physical rules. If the detected value is less than the theoretical lower limit, it is automatically corrected to the set minimum value (such as 0) to enhance the physical consistency of the data and the stability of the model input; data that meets the physical constraints is retained to ensure that no effective information is lost during the data cleaning process.

[0058] Step 2, Multi-source data construction: Derived features are constructed from the original features through computational fusion, logical combination and other methods to meet the model input requirements; at the same time, external data such as laboratory testing and manual inspection are introduced to improve the integrity of feature dimensions and the semantic relevance of desulfurization.

[0059] Specifically, it includes:

[0060] 1. Multi-point data fusion processing: For data from multiple sensors at different measurement points on the same physical quantity, a dynamic fusion algorithm is used to improve the robustness of feature values ​​and reduce the impact of single-point measurement errors on the overall model. For example, when processing data from multiple temperature sensors, the real-time variance σ² of the three sensors is first calculated to assess the consistency between the measured values. When σ² is less than or equal to a preset threshold, the data is considered relatively consistent, and the arithmetic mean of the three is directly used as the fusion output. When σ² is greater than the threshold, it indicates a large deviation. In this case, the absolute difference between each pair of the three data sets is further calculated, and the pair with the smallest difference is selected as the reliable data source. The mean of this pair of data is used as the final fusion result.

[0061] 2. Construction of Derivative Features Based on Physical Mechanisms: Combining the desulfurization process flow and its physical logic, derived variables are calculated using the collected basic features to enhance the engineering rationality of the features. For example, the total coal quantity can be obtained by summing the speed feedback values ​​of each coal feeder; the total sulfur content in the flue gas can be calculated by multiplying the total air volume by the original flue gas SO2 concentration; and the sulfur content ratio is expressed as the ratio of total sulfur to total coal quantity. Such features not only enrich the model input but also improve the model's interpretability of key desulfurization indicators.

[0062] 3. Heterogeneous Data Integration and Structured Transformation: To supplement important operational information not covered by the real-time monitoring system, this module further introduces external data sources such as laboratory testing data and manual inspection data. The acquired heterogeneous information (including text, enumeration, and numerical data) undergoes unified format conversion and structured processing. For example, text fields are converted into standard labels using rule extraction and mapping; enumeration data is converted into numerical vectors that can participate in modeling through one-hot encoding. These processes ensure that various types of information can be effectively integrated into the main data framework, providing the model with a more comprehensive description of operational conditions.

[0063] Step 3, Box Plot Test: Set up anomaly counters such as missing values, reused values, and abrupt changes for each feature, establish quartile intervals using the box plot method, and identify outliers according to the working condition type to achieve multidimensional detection and labeling of data anomalies.

[0064] Specifically, it includes:

[0065] 1. For cases where feature values ​​remain unchanged within a continuous time window, determine if there is a data acquisition fault or failure. The specific method is to set a time window threshold. For each feature value in the window Monitor changes within the range; if a certain characteristic value is within... The value remains constant. The system automatically sets the first bit of the anomaly marker corresponding to this feature from 0 to 1, indicating that the feature may be abnormal (such as sensor lag or data freeze).

[0066] 2. For cases where feature values ​​are missing, the system uses this module to monitor in real time and immediately triggers a compensation mechanism when empty data is detected. The compensation process includes the following steps: First, the missing value counter for the corresponding feature is incremented by 1 to record the frequency of missing values; second, to ensure the integrity and continuity of the data sequence, the system automatically uses valid data for that feature from the previous time point to fill in the missing values; finally, after the filling operation is completed, the reuse value counter is simultaneously incremented by 1 to count the number of times the compensation action is executed and to assist in subsequent data quality assessment and model adjustment.

[0067] 3. A first-order difference jump detection method is used to determine the changing trend of continuous time series data. Specifically, it monitors whether there are any sudden and drastic changes in the data by calculating the difference sequence between feature values ​​of adjacent time steps. When the difference value of a certain feature consistently exceeds the mutation threshold within a continuous time period, the system determines that the feature has a potential time series anomaly. For detected mutations, the system increments the corresponding mutation value counter by 1, replaces it with valid data from the previous time step, and simultaneously increments the reuse value counter by 1 to record this compensation behavior, ensuring the stationarity and continuity of the data sequence.

[0068] 4. A box plot-based statistical test method is introduced to identify anomalies in feature data. For any feature dataset, its first quartile (Q1), third quartile (Q3), and interquartile range (IQR = Q3 - Q1) are first calculated. Based on the box plot principle, a normal value interval for this feature is constructed: [Q1 - k × IQR, Q3 + k × IQR], where k is the anomaly sensitivity coefficient, typically 1.5 or 3. If a data point exceeds this interval, it is considered an outlier. For detected outliers, the system replaces them with valid values ​​from the previous time step to maintain data continuity, and simultaneously increments the reuse value counter by 1 for subsequent anomaly frequency statistics and model reliability analysis.

[0069] Step 4, Knowledge and Experience Verification: Based on expert knowledge, a knowledge graph of desulfurization operating conditions is constructed to express the semantic relationship between variables and operating conditions in a structured manner. The real-time data stream is labeled with a rule engine to achieve the identification of complex operating conditions, auxiliary judgment of abnormal behavior, and early warning support.

[0070] Specifically, it includes:

[0071] 1. Based on the operating mechanism and engineering practice of the desulfurization system, expert knowledge is gathered to set reasonable physical ranges for key parameters. For each target variable, upper and lower limits are set. If the collected data exceeds these limits, it is judged as physically unreasonable and marked as an outlier. For such data, the system replaces it with valid data from the previous moment to ensure data continuity and reliability. At the same time, the reuse value counter is incremented by 1 to record this correction operation.

[0072] 2. Utilizing a predefined set of operating condition identification rules, combined with a rule engine, the system performs logical matching and feature combination judgments on the input data to achieve automated labeling of operating condition semantic tags. This process establishes a rule base consisting of multiple Boolean logic expressions, where each rule is composed of a combination of logic consisting of a set of process variables and expert-defined thresholds. The system performs real-time matching on each rule and maps the input n-dimensional feature vector to a corresponding set of semantic tags through a tag mapping function, thereby accurately describing the current operating state and providing a data foundation and decision support for subsequent control strategy adjustments and anomaly warnings.

[0073] Step 5, Anomaly Statistical Validation: The counters for missing, mutation, and reuse are dynamically monitored within a set time window. When the number of occurrences exceeds the limit, the anomaly status flag is automatically updated and an early warning is issued, realizing closed-loop management and multi-source collaborative identification of abnormal behavior.

[0074] Specifically, it includes:

[0075] 1. Missing Value Statistics and Anomaly Detection: A sliding time window length and an upper limit threshold for the number of missing values ​​are preset. During real-time data inflow, the system continuously counts the number of missing values ​​for each feature within the window. If a feature value is missing at a certain moment, the system records the status as 1 using an indicator function; otherwise, it records it as 0. At the end of the window period, the system checks whether the count value exceeds the preset threshold. If it does, it is considered an anomaly. At this time, the system sets the second bit of the anomaly status device for the corresponding feature from 0 to 1, indicating potential problems such as sensor malfunction, data acquisition interruption, or communication abnormalities, and outputs a warning signal.

[0076] 2. Mutation Value Statistics and Anomaly Detection: The system continuously tracks and monitors the mutation behavior of feature data. By setting the time window length and mutation count threshold, it counts the jump points identified by the first-order difference. When a feature value is detected to have mutated at a certain moment and been replaced by data from the previous moment, the system marks it as 1; otherwise, it is marked as 0. If the cumulative number of mutations within the window period exceeds the preset threshold, the feature is considered to have abnormal fluctuations. The third bit of the anomaly status indicator is automatically updated to 1, and the corresponding alarm process is initiated.

[0077] 3. Reused Value Statistics and Anomaly Detection: A time window and frequency threshold are set for feature reuse behavior to determine how frequently a feature value is replaced by data from the previous time step due to anomalies. Within the sliding window, the system counts the cumulative number of times a current value is generated by reusing data from the previous time step. When the number of reuse behaviors exceeds the set threshold, the system determines it as a reuse anomaly, which may reflect equipment response delays, abnormal data processing logic, or external environmental interference. In this case, the third bit of the anomaly status indicator is updated to 1, and a warning is issued.

[0078] Step 6, Data Denoising and Normalization: Fourier transform is used to perform frequency domain denoising on the original data to suppress random disturbances; at the same time, the Z-score normalization method is applied to unify the feature scale and distribution, thereby improving the stability of model training and the standardization of input data.

[0079] Specifically, it includes:

[0080] 1. Fourier Transform Noise Reduction: To address potential high-frequency disturbances in the original time series, the Discrete Fourier Transform is used to transform it to the frequency domain. Let the original time series signal be: The signal is transformed from the time domain to the frequency domain using the Discrete Fourier Transform: ,in, Indicates the first One frequency component, The total number of data points. The imaginary unit. In the frequency domain, based on the signal's amplitude spectrum. Distribution characteristics, setting cutoff frequency threshold For high-frequency noise components (i.e. ) is filtered out, that is: ; Then, the denoised time-domain signal is recovered through inverse transform: ,in, This is a frequency index, indicating the location of a specific frequency component. The absolute value of the frequency index is used for uniform measurement. The distance from the center frequency is used to eliminate negative frequencies in symmetrical frequencies. This indicates the frequency components after removing high frequencies. This represents the time-domain signal after noise reduction. This method effectively reduces high-frequency noise caused by equipment jitter, sampling errors, or environmental interference while preserving the key structural features of the signal, thus improving data stability and the robustness of subsequent model analysis.

[0081] Z-score standardization: To eliminate scale differences between different feature variables and improve the training efficiency and convergence stability of the model, the system performs Z-score standardization on each feature variable. Let the sample sequence of a certain feature variable be... Where n is the total number of samples and its mean is... with standard deviation The calculations are as follows: ; The Z-score standardization formula is then expressed as: ,in, Indicates sample The standardized results show that the mean of the standardized feature data is 0 and the standard deviation is 1, which helps improve the robustness and convergence speed of the model when dealing with data of different scales.

[0082] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0083] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A desulfurization prediction data preprocessing method based on box plot and knowledge-experience dual-drive approach, characterized in that, Includes the following steps: S1, Data Acquisition and Cleaning: Real-time acquisition of various key parameters during the operation of the desulfurization system, including flue gas flow rate, flue gas temperature, gypsum slurry pH value, and circulating pump current. At the same time, cleaning operations based on data integrity and logical consistency are performed on the acquired key parameters, including verification of data type legality and numerical rationality. S2, Multi-source data construction: Process and reconstruct the key parameters after the cleaning operation, construct derived features through computational fusion, logical combination or statistical derivation, and integrate auxiliary information from laboratory test data and manual inspection data to supplement external feature parameters closely related to the desulfurization process. S3, Box Plot Test: For the constructed multi-source data, a missing value counter, a reused value counter, a mutation value counter, and an anomaly status device are set to comprehensively record and identify various abnormal states. At the same time, based on the statistical principle of box plots, quartile intervals of features are constructed according to the working condition category, and potential outliers are identified by using the upper and lower limits, so as to realize the detection of outliers and anomalies in multi-source data. S4, Knowledge and Experience Verification: Integrating expert experience and prior knowledge to perform multi-dimensional anomaly identification and operating condition consistency verification on multi-source data, the relationship between key operating parameters, equipment status and typical operating conditions is expressed in a structured way, and semantic labeling is performed on real-time multi-source data streams with the help of rule engine or feature combination logic, so as to realize the identification and early warning of abnormal operating conditions, boundary behavior or potential failure modes. S5, Anomaly Statistical Validation: Dynamically monitors and identifies the statistical indicators of missing values, mutation values, and reused values ​​in multi-source data; S6, Data Denoising and Normalization: Fourier transform noise reduction is performed on multi-source data. At the same time, Z-score normalization strategy is adopted to unify the scale and regularize the distribution of the noise-reduced multi-source data. The data acquisition and cleaning in S1 includes: S11, Sensor Deployment and Data Acquisition: Multiple types of industrial-grade sensors are deployed in the desulfurization tower process, including flow sensors, temperature sensors, pH meters, pressure sensors and current transformers. The sensors are connected to the data acquisition terminal through signal acquisition cards to realize continuous acquisition of real-time data. S12, Data type verification mechanism: The key parameters collected are automatically verified using a regular expression matching strategy. If the key parameters collected conform to the floating-point format, they are converted and stored in the database. If they do not conform, they are marked as abnormal, triggering an alarm and recording the error. S13, Data Physical Constraint Verification and Anomaly Correction: Combining the engineering attributes and physical laws of variables in the desulfurization process, a characteristic dimension rule base is constructed to determine the reasonableness of the range of key parameter values. For key parameters that should be non-negative, if an outlier value less than zero appears, it is corrected to the minimum value. The multi-source data construction in S2 includes: S21, Multi-point data fusion processing: A dynamic fusion algorithm based on variance discrimination is used to process the same key parameter. When the variance of the measured value of the key parameter is less than the preset threshold, the average value is taken as the fusion result. When the variance exceeds the preset threshold, the two sets of data with the smallest difference are selected for weighted average. S22, based on the desulfurization process, a variety of derived variables are calculated based on the collected key parameters. Among them, the total sulfur in the flue gas is calculated by multiplying the total air volume and the original flue gas SO2 concentration, and the sulfur content is calculated by the ratio of the total sulfur to the total coal volume. S23, Heterogeneous Data Integration and Structured Conversion: Introducing external data from laboratory testing and manual inspection, and performing unified format conversion and structured encoding processing on text, enumeration, and numerical information; The box plot verification in S3 includes: S31, Eigenvalue stability detection: For cases where multi-source data remain unchanged within a continuous time window, determine whether there is a data acquisition failure or malfunction. S32, Missing value identification and imputation: For cases where missing values ​​exist in multi-source data, the missing event is considered to trigger the compensation mechanism; S33, Temporal Abrupt Change Detection: The change trend of continuous temporal data in multi-source data is judged by the first-order difference jump detection method. If the continuous difference value exceeds the abrupt change threshold, it is regarded as a temporal abnormality. The abrupt change value counter is incremented by 1 and replaced with the value of the previous time step. At the same time, the reuse value counter is incremented by 1. S34, Box Plot Outlier Detection: The box plot method is used to identify outlier data points. For multi-source data, the first quartile Q1, the third quartile Q3, and the interquartile range IQR are calculated, and a normal value interval is constructed based on the box plot principle, namely [Q1-k×IQR, Q3+k×IQR], where k is the anomaly sensitivity coefficient. If a data point in the multi-source data exceeds the range of this interval, it is determined to be an outlier. The knowledge and experience verification in S4 includes: S41, Anomaly identification based on empirical rules: Based on the operation mechanism and engineering practice of the desulfurization system, a physical range is set for the key parameters. If the collected key parameters exceed the range, they are regarded as anomalies and replaced with the valid value of the key parameters at the previous moment. At the same time, the reuse value counter is incremented by 1 to record the correction behavior. S42, Operating condition semantic labeling: Using a predefined set of operating condition identification rules, combined with a rule engine, multi-source data is matched in real time. Based on the combination logic of multi-source data, corresponding operating condition semantic labels are automatically assigned to achieve the identification of operating status. The anomaly checker in S5 includes: S51, Missing Value Statistics and Anomaly Detection: Set the sliding time window length threshold and the upper limit threshold for missing values, and count the number of missing values ​​in each multi-source data within the window in real time. If the cumulative number of missing values ​​exceeds the upper limit threshold for missing values, it is considered a missing value anomaly. The second bit of the anomaly status device is set from 0 to 1, and an early warning is triggered. S52, Mutation value statistics and anomaly detection: Based on the results of the first-order difference, the number of occurrences of mutation events is counted. If the number of mutations exceeds the threshold of the sliding time window length within the set window, it is marked as a fluctuation anomaly, the third bit of the anomaly status device is updated to 1, and the alarm process is initiated. S53, Reuse Value Statistics and Anomaly Detection: Monitor the frequency of multi-source data being reused by the previous data. If the number of reuses exceeds the time window and number threshold of the reuse behavior within the window, it is judged as a reuse anomaly. Update the corresponding bit of the anomaly status device and issue an early warning, reflecting that there is a problem with equipment lag or abnormal data processing.

2. The desulfurization prediction data preprocessing method based on box plot-knowledge experience dual-drive as described in claim 1, characterized in that, The eigenvalue stability detection in S31 includes: S311, Set monitoring time window: Set the time window threshold 𝑇, and extract the continuous observation sequence of multi-source data within the window; S312, Determine if the feature value is constant: Determine whether the multi-source data remains unchanged within the time window; if the value is equal at all times, it is considered a constant state; S313, Mark Abnormal State: When multi-source data remains constant within the window, set the first bit of its abnormal state register from 0 to 1 to mark it as an abnormal state.

3. The desulfurization prediction data preprocessing method based on box plot-knowledge experience dual-drive as described in claim 2, characterized in that, The abnormal states include sensor malfunctions or data freezes.

4. The desulfurization prediction data preprocessing method based on box plot-knowledge experience dual-drive as described in claim 3, characterized in that, The missing value identification and imputation in S32 includes: S321, Record missing events: When it is detected that the observation value of multi-source data is empty at the current time, increment the corresponding missing value counter by 1 and record the frequency of missing events; S322, Perform missing value imputation: Automatically imput the current missing value using the valid observation value from the previous time step; S323, Update the reuse counter: After the filling operation is completed, increment the reuse value counter by 1 to record the number of times the compensation action is executed.

5. The desulfurization prediction data preprocessing method based on box plot-knowledge experience dual-drive according to claim 4, characterized in that, The data denoising and normalization in S6 include: S61, Fourier Transform Noise Reduction: To address high-frequency interference in multi-source data, a Discrete Fourier Transform is used to convert the time series to the frequency domain. Based on the amplitude spectrum distribution, a cutoff frequency threshold is set. High-frequency noise components are filtered out and then restored to the denoised time-domain signal through inverse transformation. S62, Z-score normalization: Performs Z-score normalization on the denoised multi-source data.

Images