Abnormality identification method and system based on automatic monitoring data of stationary pollution sources

Abstract

The application belongs to the technical field of pollution source automatic monitoring and supervision, and specifically discloses an abnormality identification method and system based on fixed pollution source automatic monitoring data. The method comprises the following steps: extracting data at a preset frequency or period to form a time-stamped data sequence; performing differential statistical test in the data sequence to generate a suspected abnormality sub-sequence; taking a historical segment that continuously meets the standard and has stable working conditions as a positive sample to perform double Monte Carlo prediction to obtain a mode coincidence rate sequence; performing robust trend estimation to obtain a dynamic baseline that changes over time; calculating the residual of the credibility of the emission mode relative to the dynamic baseline, and outputting an abnormality alarm signal when the residual exceeds an adaptive discrete threshold. The application completes one-dimensional abnormal boundary definition and high-dimensional emission mode drift test in the same system framework, forms a complementary supervision framework combining "short-time-long-time" and "point-surface", and improves the abnormality identification precision and the intelligent level of supervision.

Description

Technical Field

[0001] This application relates to the field of automatic monitoring and supervision of pollution sources, and in particular to an anomaly identification method and system based on automatic monitoring data of stationary pollution sources. Background Technology

[0002] Automatic monitoring of pollutant emissions is a key technological means for real-time monitoring of pollutant emissions from industrial enterprises, and it plays an irreplaceable role in supervising enterprises' compliant emissions.

[0003] However, in practical applications, the quality of automated monitoring data is often compromised. Some companies engage in data falsification and abnormal operations, leading to distorted monitoring data. These behaviors not only conceal the true emissions situation but also increase the difficulty for regulatory authorities to identify illegal emissions. Traditional anomaly detection methods often rely on empirical thresholds, such as setting fixed upper or lower limits for data screening. These methods lack statistical basis, are difficult to adapt to complex and ever-changing emission characteristics, and are prone to missed detections or false alarms.

[0004] In recent years, some studies have attempted to introduce machine learning methods to model and analyze automated monitoring data in order to identify changes in emission patterns and potential abnormal behavior. Meanwhile, significant differences exist between emission pattern change events identified through time series analysis and traditional regulatory records, suggesting blind spots in the current regulatory methods for anomaly detection.

[0005] However, despite the above studies validating the potential of machine learning in emissions data analysis, it still faces the following challenges in practical applications: First, traditional methods lack scientific statistical definition standards for typical anomalies such as "excessive emissions," "production within limits," and "zero-value anomalies." Second, existing methods lack systematic comprehensive analysis capabilities when faced with multiple types of anomalies coexisting. Third, the scarcity of anomaly samples makes it difficult to train machine learning models, and single prediction results are highly random and lack stability. Fourth, static thresholds cannot adapt to the long-term, slow changes in emission patterns caused by equipment aging, process changes, etc.

[0006] Therefore, there is an urgent need for a method that integrates the advantages of statistical analysis and machine learning to complete the one-dimensional anomaly boundary definition and high-dimensional emission pattern drift detection within the same system framework, so as to achieve systematic, scientific and automated anomaly screening of monitoring data and improve the accuracy and intelligence of environmental supervision. Summary of the Invention

[0007] To address the aforementioned problems in the prior art, this invention provides an anomaly identification method and system based on automatic monitoring data from stationary pollution sources, enabling systematic, scientific, and automated anomaly screening of monitoring data, thereby improving the accuracy and intelligence of environmental supervision.

[0008] This application first discloses an anomaly identification method based on automatic monitoring data of stationary pollution sources, including the following steps:

[0009] The data preparation steps involve continuously extracting online monitoring data from each emission outlet and the operating condition markers for each data point at a preset frequency or cycle to form a timestamped data sequence.

[0010] The multi-dimensional anomaly screening step involves performing differential statistical tests on normal emission conditions, maintenance and calibration conditions, and start-up and shutdown conditions within the data sequence based on the emission limit standards of each emission outlet, in order to generate suspected anomaly subsequences.

[0011] The emission pattern self-learning step uses historical data to train the emission pattern recognition model, and uses segments that consistently meet standards and operate under stable conditions as positive samples to perform double Monte Carlo prediction in order to obtain the pattern compliance rate sequence.

[0012] The dynamic baseline construction step involves performing robust trend estimation based on the pattern compliance rate sequence to obtain a dynamic baseline that changes over time, and determining a reliability reference for the emission patterns that the pollution source should meet under normal operating conditions based on the dynamic baseline.

[0013] The adaptive discrete point identification step calculates the residual of the confidence level of the emission pattern relative to the dynamic baseline, and outputs an abnormal alarm signal when the residual exceeds the adaptive discrete threshold.

[0014] In the above-described anomaly identification method, the differential statistical test in the multi-dimensional anomaly initial screening step includes:

[0015] Perform data exceedance anomaly analysis, limit-adhesion operation analysis, zero value anomaly analysis, and cross-sectional area anomaly analysis on data under normal emission conditions;

[0016] Perform anomaly analysis on maintenance / calibration time under maintenance / calibration conditions; and

[0017] Statistics on data mutations / constancy, statistics on differences between normal and abnormal work processes, and statistics on diurnal differences.

[0018] In the above-mentioned anomaly identification method, the data exceeding the standard anomaly analysis includes:

[0019] Data filtering involves selecting records with concentrations exceeding the corresponding emission limits from data marked as normal emission conditions and arranging them by time as a sequence of time points exceeding the limits.

[0020] The interval is divided, and a first-order time difference is performed on the sequence of time points exceeding the standard. If the difference of the first time point is greater than 1.5 times the sampling period, then the first time point is the starting point of the new event; if the difference between the second time point and the next point is greater than 1.5 times the sampling period, then the second time point is the ending point of the current event.

[0021] For anomaly detection, the cumulative duration of all events within the same analysis period is summed, and a discrete upper bound for the cumulative duration of multiple analysis periods is calculated. When a cumulative duration exceeds the discrete upper bound, the corresponding analysis period is considered to have an abnormal data exceeding the standard.

[0022] In the above-mentioned anomaly identification method, the data mutation / constancy analysis includes:

[0023] Calculate the absolute value of the difference between monitoring data at adjacent sampling times to form a change matrix;

[0024] For each monitoring factor, the non-empty change values ​​in the change matrix are counted, and the discrete upper bound of the change is calculated using the 10-times interquartile range method.

[0025] The time point when the change exceeds the discrete upper bound is identified as a data mutation event;

[0026] Data points with changes less than a preset minimum threshold are selected from the change matrix to form a constant data sequence;

[0027] Perform fast vectorization grouping on the constant data sequence and count the duration of constant data in each group;

[0028] The discrete upper bound of the time is calculated using the 4-interquartile range method, and start and end times exceeding the upper bound are identified as data constant anomalies.

[0029] In the above-described anomaly identification method, the cross-sectional area analysis includes:

[0030] The real-time cross-sectional area is calculated using the exhaust gas velocity and corresponding flow rate monitored at the outlet.

[0031] Using the registered value of the discharge permit or the median of long-term observation as the benchmark cross-sectional area, time points where the difference between the real-time area and the benchmark value exceeds 30% or exceeds 1.5 times the interquartile range are identified as cross-sectional area anomalies.

[0032] In the above-mentioned anomaly identification method, the multi-dimensional anomaly initial screening step further includes day-night difference statistics:

[0033] For the normal emission data of each monitoring factor, the data from 5:00 to 20:00 each day is classified as daytime data, and the data from 21:00 to 4:00 the next day is classified as nighttime data.

[0034] Independent samples t-tests were performed on daytime and nighttime data separately to determine whether nighttime data was significantly higher than daytime data, with the significance level set at p < 0.05.

[0035] When the test results are significant, the corresponding dates are marked as having anomalies in diurnal differences.

[0036] In the above-described anomaly identification method, the dual Monte Carlo prediction of the emission pattern self-learning step includes:

[0037] Select positive samples: Select a set of stable and compliant historical data, normalize the historical data, and use it as the benchmark positive sample training set;

[0038] Obtain negative samples and randomly generate a negative sample training set of the same size as the benchmark positive sample training set;

[0039] Training and prediction: The positive sample training set and the negative sample training set are merged to train a random forest classifier. The trained random forest classifier is used to predict the evaluation set, and the predicted current pattern is compared with the baseline pattern to obtain the prediction accuracy.

[0040] Repeat the training and prediction steps, and take the average of the prediction accuracies obtained multiple times to obtain the pattern conformity rate sequence.

[0041] With time as the independent variable and the aforementioned compliance rate as the dependent variable, the slope was estimated using Thiel-Sen robust linear regression.

[0042] If the slope is negative, the regression line is used as the dynamic baseline; if the slope is non-negative, the median of the compliance rate of all cycles is used as the fixed baseline.

[0043] In the above-mentioned anomaly identification method, the adaptive discrete point identification step further includes an emission pattern discrete point identification step, specifically including:

[0044] Calculate the residuals by calculating the difference between the current long-term baseline value and the current compliance rate, and obtain the residual sequence.

[0045] Analyze the dispersion of the residuals and calculate the interquartile range (IQR) of the residual sequence. If IQR = 0, recalculate the IQR for the subset of residual sequences that are greater than the median.

[0046] Calculate the discrete factor. If the discrete factor is greater than 3, trigger an abnormal alarm.

[0047] Accordingly, this application also proposes an anomaly identification system based on automatic monitoring data of stationary pollution sources, including: a data preparation module, used to continuously extract online monitoring data of each emission outlet and the operating condition marker of each data point according to a preset frequency or cycle, so as to form a data sequence with timestamps;

[0048] The multi-dimensional anomaly screening module is used to perform differential statistical tests on normal emission conditions, maintenance and calibration conditions and start-up and shutdown conditions in the data sequence based on the emission limit standards of each emission outlet, so as to generate suspected anomaly subsequences.

[0049] The emission pattern self-learning module is used to train the emission pattern recognition model with historical data and use segments that consistently meet standards and have stable operating conditions as positive samples to perform double Monte Carlo prediction in order to obtain the pattern compliance rate sequence.

[0050] The dynamic baseline construction and discrete point identification module is used to perform robust trend estimation based on the pattern conformity rate sequence to obtain a dynamic baseline that changes over time. Based on the dynamic baseline, the module determines the credibility reference of the emission pattern that the pollution source should meet under normal operating conditions and calculates the residual of the credibility of the emission pattern relative to the dynamic baseline. When the residual exceeds the adaptive discrete threshold, an abnormal alarm signal is output.

[0051] The alarm output module is used to push abnormal information reports.

[0052] This application also proposes a computer program product comprising computer instructions that, when executed, implement the method described above.

[0053] Compared with the prior art, the present invention has the following significant advantages:

[0054] 1. Unified statistical framework: For the first time, eight types of one-dimensional data anomalies (exceeding the standard, touching the limit, zero value, mutation, constant, maintenance timeout, cross-sectional area anomaly, and diurnal difference) and one type of high-dimensional pattern drift are incorporated into the same "discrete boundary" theory, avoiding subjective differences caused by empirical thresholds.

[0055] 2. Robustness to Few-Sample Learning: Utilizing a dual Monte Carlo positive label-free learning framework, negative samples are generated using random perturbations without the need for real negative samples, thus solving the industry problem of "missing outlier samples." Repeated simulations reduce prediction variance, resulting in stable and reproducible results.

[0056] 3. Baseline Dynamic Adaptation: A dynamic baseline is constructed using Thiel-Sen robust regression, which can identify long-term trends of deterioration in emission patterns. Compared with the static threshold method, the false alarm rate decreases over time.

[0057] 4. Intelligent and adjustable threshold: The innovative approach of selectively using variable IQR can adaptively handle the extreme case of "residual concentration", ensuring the robustness of discrete point identification, tolerating small fluctuations and being sensitive to large offsets, thus making it more robust.

[0058] 5. Highly efficient and deployable: The entire process is automated, from raw automatic monitoring data to visual reports, requiring no manual intervention. Single discharge outlet analysis takes less than 2 minutes, meeting the real-time monitoring needs of thousands of discharge outlets in the region.

[0059] 6. Highly interpretable, each anomaly is given with statistical boundaries and discrete factors, directly supporting on-site verification and law enforcement evidence collection. Attached Figure Description

[0060] Figure 1 A flowchart is shown below illustrating an anomaly identification method based on automatic monitoring data of stationary pollution sources according to some embodiments of this application;

[0061] Figure 2 This illustrates the various modules specifically included in anomaly analysis;

[0062] Figure 3 The specific process of the emission mode self-learning step is shown;

[0063] Figure 4 A block diagram of an anomaly identification system based on automatic monitoring data of stationary pollution sources, according to some embodiments of this application, is shown. Detailed Implementation

[0064] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0065] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items. Words such as “comprising” or “including” mean that the element or object preceding “comprising” or “including” covers the element or object listed following “comprising” or “including” and its equivalents, and does not exclude other elements or objects. Words such as “connected” or “linked” are not limited to physical or mechanical connections and can include electrical connections, whether direct or indirect.

[0066] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0067] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0068] Existing automatic monitoring data anomaly identification technology for discharge outlets is still at the level of "single-point threshold comparison" and lacks a unified statistical framework for multiple anomalies. This paper attempts to provide a statistical analysis method for automatic monitoring data of discharge outlets driven by "statistics + AI" to complete the one-dimensional anomaly boundary definition and high-dimensional emission pattern drift detection within the same software framework.

[0069] Figure 1 A flowchart is shown of an anomaly identification method based on automatic monitoring data of stationary pollution sources, according to some embodiments of this application. The anomaly identification method includes the following steps:

[0070] S1, Data preparation step: continuously extract online monitoring data, corresponding emission limit standards, and operating condition markers for each emission point according to a preset frequency or cycle to form a data sequence with timestamps.

[0071] The preset frequency or period is the shortest time period (frequency) for studying pollution emissions. It can be based on years, months, or days (including multiple days). First, analyze the changing patterns and anomalies of pollution data within one period. After accumulating multiple analysis periods, compare the periods to discover long-term emission patterns and anomalies (especially time-related anomalies), thus enabling a comprehensive analysis of short-term and long-term emission patterns. The sampling frequency is generally once per hour, but it can also be combined with the analysis period. If the analysis period is monthly, the frequency can be modified to once per day.

[0072] Each monitoring station can monitor different pollutants. For example, for chemical plants, it is typically necessary to monitor the measured concentration of non-methane total hydrocarbons (unit: mg / m³). 3 Non-methane total hydrocarbon baseline oxygen content concentration (unit: mg / m³) 3 NO x Measured concentration (unit: mg / m³) 3 NO x Reference oxygen concentration (unit: mg / m³) 3Flue gas velocity (m / s), flue gas temperature (°C), flue gas humidity (%), static pressure (kPa), and hot air volume (m³ / s). 3 If the analysis period is in weeks and the sampling frequency is once per hour, then 26,304 records will be generated from January 1, 2022 to December 31, 2024. These 26,304 records will then be used as the dataset for anomaly analysis and as the training set for machine learning.

[0073] The emission limit standards for each monitoring point are different. The corresponding discharge permit limits can be downloaded from the National Discharge Permit Management Information Platform and used as a benchmark value for comparison when screening abnormal data.

[0074] Furthermore, to improve data quality from the source, data cleaning can be performed during data preparation to eliminate the impact of obviously unqualified data on the final analysis results. This specifically includes handling negative values, extreme values, and missing values. In big data scenarios, the simplest method of deletion can be used. Alternatively, data can be repaired based on its historical performance. For example, missing values ​​can be filled using median or interpolation methods, and extreme values ​​can be analyzed to determine their causes before applying appropriate handling methods.

[0075] S2, the multi-dimensional anomaly screening step, involves performing differential statistical tests on normal emission conditions, maintenance and calibration conditions, and boiler start-up and shutdown conditions within the data sequence based on the emission limit standards for each emission outlet, to generate suspected anomaly subsequences. Because in practice, the normal emission concentration of enterprises is far below the limit, if there are fewer than 5 samples with measured values ​​exceeding the limit or exceeding 0.8 times the limit within 3 years, meaning that there is very little data exceeding the limit, anomaly analysis events are generally not triggered.

[0076] Figure 2 This illustrates the various modules involved in anomaly analysis. The differential statistical test in the aforementioned multidimensional anomaly screening step includes the following functional modules:

[0077] a) Perform the following analysis on data under normal emission conditions: S21 for data exceeding standards, S22 for operation under emission limits, S24 for zero value anomaly analysis, and S27 for cross-sectional area anomaly analysis.

[0078] b) Perform maintenance / calibration time anomaly analysis S23 on data under maintenance / calibration conditions; and

[0079] c) Statistics on differences between normal and abnormal work processes (S25), statistics on data mutations / constancy (S26), and statistics on diurnal differences (S28).

[0080] The following examples illustrate the analysis process: data exceeding the standard anomaly analysis S21, data mutation / constancy statistics S26, cross-sectional area anomaly analysis S27, and diurnal difference statistics S28.

[0081] The Data Exceeding Standard Anomaly Analysis S21 identifies abnormal periods in online hourly data marked as "normal emission conditions" where pollutant concentrations have exceeded the corresponding emission limits but have not yet been marked as "exceeding standards" by the system. It performs periodic statistics on "events" to discover potential data distortion or illegal discharge risks.

[0082] The data exceeding the standard anomaly analysis S21 specifically includes:

[0083] 1) Data filtering: Select records with concentrations exceeding the corresponding emission limits from the data marked as normal emission conditions, and arrange them by time as a sequence of time points exceeding the standard;

[0084] 2) Interval division: Perform first-order time difference on the sequence of time points exceeding the standard. If the difference at the first time point is greater than 1.5 times the sampling period, then the first time point is the starting point of the new event; if the difference between the second time point and the subsequent point is greater than 1.5 times the sampling period, then the second time point is the ending point of the current event.

[0085] 3) Anomaly detection: Accumulate the cumulative duration of all events within the same analysis period, calculate the discrete upper bound of the cumulative duration of multiple analysis periods, and when a cumulative duration exceeds the discrete upper bound, it is determined that there is an abnormal data exceeding the standard in the corresponding analysis period.

[0086] For each factor of the change matrix, the data mutation / constant statistics S26 uses the 10×IQR method to determine the "mutation statistical boundary"; the time point that exceeds the boundary is recorded as the mutation event, and it supports fast vectorized grouping output of start and end times.

[0087] Data mutation / constant statistics S26 specifically include:

[0088] 1) Calculate the absolute value of the difference between monitoring data at adjacent sampling times to form a change matrix;

[0089] 2) For each monitoring factor, count the non-empty change values ​​in the change matrix, and use the interquartile range method (10 times) to calculate the discrete upper bound of the change.

[0090] 3) Identify time points where the change exceeds the discrete upper bound as data mutation events;

[0091] 4) Select data points whose change is less than a preset minimum threshold from the change matrix to form a constant data sequence;

[0092] 5) Perform fast vectorization grouping on the constant data sequence and count the duration of constant data in each group;

[0093] 6) Calculate the discrete upper bound of the time using the 4-interquartile range method, and identify start and end times exceeding the upper bound as data constant anomalies.

[0094] Cross-sectional area anomaly analysis S27 specifically includes:

[0095] 1) Calculate the real-time cross-sectional area using the exhaust gas velocity and corresponding flow rate monitored at the outlet;

[0096] 2) Using the registered value of the discharge permit or the median of long-term observation as the benchmark cross-sectional area, select time points where the difference between the real-time area and the benchmark value exceeds 30% or exceeds 1.5 times the interquartile range, and determine them as cross-sectional area anomalies.

[0097] The diurnal variation statistics S28 include:

[0098] 1) For the normal emission data of each monitoring factor, the data from 5:00 to 20:00 each day is classified as daytime data, and the data from 21:00 to 4:00 the next day is classified as nighttime data;

[0099] 2) Perform independent samples t-tests on daytime and nighttime data separately to determine whether nighttime data is significantly higher than daytime data, with the significance level set at p < 0.05;

[0100] 3) When the test results are significant, the corresponding dates are marked as having anomalies in diurnal differences.

[0101] The aforementioned multi-dimensional anomaly screening process is fully vectorized and can be completed in one go within a database or array language, without the need for iterative analysis. The 4×IQR (4 times interquartile range) used is more stringent than the conventional 1.5×IQR (1.5 times interquartile range), meeting the environmental "zero tolerance" regulatory requirements. Furthermore, using "cycle" (analysis cycle) rather than "hourly" (sampling frequency) as the assessment granularity can suppress false positives caused by occasional peaks.

[0102] The same analytical principles apply to other anomaly analyses: anomalies within a short period are analyzed on an analysis cycle basis, data characteristics from multiple analysis cycles are aggregated, and long-term emission pattern changes are compared and analyzed.

[0103] S3, the emission pattern self-learning step, uses historical data to train the emission pattern recognition model, and uses segments that consistently meet standards and have stable operating conditions as positive samples to perform dual Monte Carlo prediction in order to obtain the pattern compliance rate sequence.

[0104] Combination Figure 3 The dual Monte Carlo predictions of the emission model self-learning steps include:

[0105] 1) Select positive samples: Select a set of stable and compliant historical data, and normalize the historical data to use as the benchmark positive sample training set S31. For example, select normal data from January 1, 2023 to December 31, 2023, normalize the historical data, and use it as the benchmark emission model positive sample training set S31.

[0106] 2) Obtain negative samples by randomly generating a negative sample training set of the same size as the baseline positive sample training set. As mentioned earlier, modern enterprises have strict management of pollutant emissions, making it difficult to obtain real-world negative samples. A random function can be used to randomly generate a negative sample training set S32 of the same size as the baseline positive sample training set S31. The elements of the random function are limited to [0,1], and each negative sample is marked to indicate that it is a negative sample. Similarly, positive samples can also be marked to indicate that they are positive samples.

[0107] 3) Training and Prediction: The positive sample training set S31 and the negative sample training set S32 are merged to form a training set containing both positive and negative samples. This training set is used to perform machine training (S34) on the random forest classifier, allowing the classifier to autonomously learn the features of positive and negative samples. The trained random forest classifier is used to predict the data in each period of the evaluation set S33, and the predicted current pattern is compared with the baseline pattern to calculate the prediction accuracy (S35).

[0108] 4) Repeat the training and prediction steps. The more repetitions, the better the machine learning effect of the random forest classifier. It is preferable to repeat more than 1000 times. Take the average of the accuracy of the predictions obtained from multiple repetitions to obtain a stable pattern conformity rate sequence S36.

[0109] 5) Using time as the independent variable and the compliance rate as the dependent variable, the slope is estimated using Thiel-Sen robust linear regression. If the slope is negative, the regression line is used as the dynamic baseline; if the slope is non-negative, the median of the compliance rate over all periods is used as the fixed baseline.

[0110] In this embodiment, no real negative samples are required, which can solve the pain point of "easy to find qualified but difficult to find unqualified" in environmental protection scenarios; double random averaging can eliminate the random bias caused by single sampling, resulting in small variance and high reproducibility; random forest can naturally output feature importance, which is convenient for subsequent location of key pollutants that cause pattern deviation.

[0111] S4, the dynamic baseline construction step, involves performing robust trend estimation based on the pattern conformity rate sequence to obtain a dynamic baseline that changes over time. Based on this dynamic baseline, a reliability reference is determined for the emission patterns that the pollution source should meet under normal operating conditions. Whether the pattern conformity rate sequence obtained in step S3 shows a downward trend or remains stable over time is the basis for managing the long-term behavior of polluting enterprises. This allows for the establishment of a reasonable long-term baseline for emission patterns, providing a reference for discrete point detection.

[0112] In the above embodiments, the Thiel-Sen estimation is insensitive to outliers, thus avoiding the influence of a few extreme periods on the baseline.

[0113] S5: Adaptive discrete point identification step, calculate the residual of the credibility of the emission pattern relative to the dynamic baseline, and when the residual exceeds the adaptive discrete threshold, output an abnormal alarm signal and give a grade of the severity of the jump, so as to give the polluting enterprise clear pollution source information and help the enterprise to correct it.

[0114] Specifically, it includes:

[0115] 1) Calculate the difference between the current long-term baseline value and the current compliance rate to obtain the residual sequence.

[0116] 2) Analyze the dispersion of the residuals, including:

[0117] Calculate the interquartile range (IQR) of the residual difference sequence, i.e., IQR = Q3 - Q1;

[0118] If IQR=0, it indicates that the data is very concentrated. Then, switch to the improved IQR, that is, recalculate the interquartile range IQR for the subset of residual sequences that are greater than the median. If the improved IQR is still 0, it is determined that there are no discrete points in this sequence, and this program terminates. The residuals of another parameter or another period can be recalculated.

[0119] If IQR > 0, the current IQR can be denoted as variable IQR. Dividing the residuals of each period by this value yields the "discrete factor".

[0120] Based on the dispersion factor classification, it is determined whether manual intervention in emissions is required. If 1.5 < dispersion factor ≤ 3, it is considered mild dispersion, and the manufacturer can be alerted to pay attention; if the dispersion factor > 3, it is considered high drift, that is, the emission pattern changes significantly in the corresponding analysis period, which can trigger an anomaly alarm and indicate that the manufacturer's emission pattern has changed significantly.

[0121] The following example uses data from the entire year of 2022 with the status "normal" as a positive sample training set to further illustrate the process of emission pattern self-learning.

[0122] First, take the data from the entire year of 2022 that are in the "normal" state as the positive sample training set and normalize it to the interval [0,1].

[0123] Then initialize Random-Forest (using 500 trees, with no fixed random seed). Start training the Random Forest:

[0124] a) First Monte Carlo: Generate 1000 sets of uniform random matrices of the same size as the training set and label the negative samples;

[0125] b) Training a classifier using a mixture of positive and negative results;

[0126] c) Forecast the weekly data for 2023-2024 and record the average accuracy (acc). t It is preferable to perform a second Monte Carlo calculation, repeated at least 1000 times, to obtain a stable accuracy. t sequence.

[0127] d) Regarding acc t Performing a Theil-Sen robust linear regression yielded a slope of -0.0026 (p<0.01), indicating the presence of a trend drift.

[0128] e) Using the regression line as the baseline, calculate the residual ε t =acc t - Fitted value;

[0129] f) Using the variable IQR method: Since the residual IQR > 0, the discrete factor δ is obtained directly. t =ε t / IQR;

[0130] g) Define δ t >1.5 indicates a slight change, δ t A value >3 indicates a significant change. Therefore, this case showed mild changes for 11 weeks and significant changes for 3 weeks.

[0131] Finally, the automatic reporting module can automatically generate various line charts, trend charts, 1.5 / 3 discrete boundary charts, etc.

[0132] This embodiment quantitatively identifies the time points where significant changes in emission patterns occur in the residual sequence and provides a severity classification for these changes, distinguishing between "deteriorating trends" and "random fluctuations," thus enhancing the reliability of alarms. Furthermore, the calculated discrete factor is dimensionless, facilitating horizontal comparisons between different emission outlets and different pollutants. The selection of 1.5 and 3 as two-level thresholds for the discrete factor aligns with statistical conventions, ensuring industry-wide applicability and strong interpretability.

[0133] Finally, the aforementioned abnormal discharge data and changes in discharge patterns can be summarized and automatically generated into reports and charts, making the conclusions visual and facilitating reference or use by enterprises and law enforcement platforms.

[0134] Step S2 of this application addresses the issue of "hourly-level emissions exceeding limits without being flagged," enabling the detection of instantaneous illegal emissions or data masking. Steps S3-S4, on the other hand, focus on "periodic emission characteristic patterns," identifying slow drifts or sudden changes. These two perspectives complement each other, forming a comprehensive regulatory framework that combines "short-term and long-term" and "point-to-surface" approaches. All algorithms support vectorization or Monte Carlo parallelism, making them suitable for deployment on big data platforms or edge computing nodes.

[0135] Figure 4 A block diagram of an anomaly identification system based on automatic monitoring data of stationary pollution sources, according to some embodiments of this application, is shown. The anomaly identification system is related to… Figure 1-3 The anomaly identification method shown corresponds to the specific working process of the system, which can be referred to in the previous explanation.

[0136] Specifically, the anomaly identification system includes: a data preparation module 41, a multi-dimensional anomaly screening module 42, an emission pattern self-learning module 43, a dynamic baseline construction and discrete point identification module 44, and an alarm output module 45.

[0137] The data preparation module 41 is used to continuously extract online monitoring data from each emission outlet and the operating condition marker of each data point according to a preset frequency or cycle, so as to form a data sequence with timestamps.

[0138] The multi-dimensional anomaly screening module 42 is used to perform differential statistical tests on normal emission conditions, maintenance and calibration conditions and start-up and shutdown conditions in the data sequence based on the emission limit standards of each emission outlet, so as to generate suspected anomaly subsequences.

[0139] The emission pattern self-learning module 43 is used to train the emission pattern recognition model with historical data, and to perform double Monte Carlo prediction using segments that consistently meet standards and have stable operating conditions as positive samples to obtain the pattern compliance rate sequence.

[0140] The dynamic baseline construction and discrete point identification module 44 is used to perform robust trend estimation based on the pattern conformity rate sequence to obtain a dynamic baseline that changes over time. Based on the dynamic baseline, the module determines the credibility reference of the emission pattern that the pollution source should meet under normal operating conditions and calculates the residual of the credibility of the emission pattern relative to the dynamic baseline. When the residual exceeds the adaptive discrete threshold, an abnormal alarm signal is output.

[0141] The alarm output module 45 is used to push abnormal information reports.

[0142] The embodiments of the present invention are not limited to the type of application; they can be system applications or software applications.

[0143] It should be noted that in the examples and description of this patent, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0144] Although this application has been illustrated and described with reference to certain preferred embodiments thereof, those skilled in the art should understand that various changes in form and detail may be made thereto without departing from the spirit and scope of this application.

Claims

1. A method for anomaly identification based on automatic monitoring data of stationary pollution sources, characterized in that, Includes the following steps: The data preparation steps involve continuously extracting online monitoring data from each emission outlet and the operating condition markers for each data point at a preset frequency or cycle to form a timestamped data sequence. The multi-dimensional anomaly screening step involves performing differential statistical tests on normal emission conditions, maintenance and calibration conditions, and start-up and shutdown conditions within the data sequence based on the emission limit standards of each emission outlet, in order to generate suspected anomaly subsequences. The emission pattern self-learning step uses historical data to train the emission pattern recognition model, and uses segments that consistently meet standards and operate under stable conditions as positive samples to perform double Monte Carlo prediction in order to obtain the pattern compliance rate sequence. The dynamic baseline construction step involves performing robust trend estimation based on the pattern compliance rate sequence to obtain a dynamic baseline that changes over time, and determining a reliability reference for the emission patterns that the pollution source should meet under normal operating conditions based on the dynamic baseline. The adaptive discrete point identification step calculates the residual of the confidence level of the emission pattern relative to the dynamic baseline, and outputs an abnormal alarm signal when the residual exceeds the adaptive discrete threshold.

2. The method as described in claim 1, characterized in that, The differential statistical tests in the multidimensional anomaly screening step include: Perform data exceedance anomaly analysis, limit-adhesion operation analysis, zero value anomaly analysis, and cross-sectional area anomaly analysis on data under normal emission conditions; Perform anomaly analysis on maintenance / calibration time under maintenance / calibration conditions; and Statistics on data mutations / constancy, statistics on differences between normal and abnormal work processes, and statistics on diurnal differences.

3. The method as described in claim 2, characterized in that, The data anomaly analysis includes: Data filtering involves selecting records with concentrations exceeding the corresponding emission limits from data marked as normal emission conditions and arranging them by time as a sequence of time points exceeding the limits. The interval is divided, and a first-order time difference is performed on the sequence of time points exceeding the standard. If the difference of the first time point is greater than 1.5 times the sampling period, then the first time point is the starting point of the new event; if the difference between the second time point and the next point is greater than 1.5 times the sampling period, then the second time point is the ending point of the current event. For anomaly detection, the cumulative duration of all events within the same analysis period is summed, and a discrete upper bound for the cumulative duration of multiple analysis periods is calculated. When a cumulative duration exceeds the discrete upper bound, the corresponding analysis period is considered to have an abnormal data exceeding the standard.

4. The method as described in claim 2, characterized in that, The data mutation / constancy analysis includes: Calculate the absolute value of the difference between monitoring data at adjacent sampling times to form a change matrix; For each monitoring factor, the non-empty change values ​​in the change matrix are counted, and the discrete upper bound of the change is calculated using the 10-times interquartile range method. The time point when the change exceeds the discrete upper bound is identified as a data mutation event; Data points with changes less than a preset minimum threshold are selected from the change matrix to form a constant data sequence; Perform fast vectorization grouping on the constant data sequence and count the duration of constant data in each group; The discrete upper bound of the time is calculated using the 4-interquartile range method, and start and end times exceeding the upper bound are identified as data constant anomalies.

5. The method as described in claim 2, characterized in that, The cross-sectional area analysis includes: The real-time cross-sectional area is calculated using the exhaust gas velocity and corresponding flow rate monitored at the outlet. Using the registered value of the discharge permit or the median of long-term observation as the benchmark cross-sectional area, time points where the real-time area differs from the benchmark value by more than 30% or more than 1.5 times the interquartile range are identified as cross-sectional area anomalies.

6. The method as described in claim 1, characterized in that, The multi-dimensional anomaly screening step further includes day-night difference statistics: For the normal emission data of each monitoring factor, the data from 5:00 to 20:00 each day is classified as daytime data, and the data from 21:00 to 4:00 the next day is classified as nighttime data. Independent samples t-tests were performed on daytime and nighttime data separately to determine whether nighttime data was significantly higher than daytime data, with the significance level set at p < 0.

05. When the test results are significant, the corresponding dates are marked as having anomalies in diurnal differences.

7. The method as described in claim 2, characterized in that, The dual Monte Carlo predictions of the emission pattern self-learning step include: Select positive samples: Select a set of stable and compliant historical data, normalize the historical data, and use it as the benchmark positive sample training set; Obtain negative samples and randomly generate a negative sample training set of the same size as the benchmark positive sample training set; Training and prediction: The positive sample training set and the negative sample training set are merged to train a random forest classifier. The trained random forest classifier is used to predict the evaluation set, and the predicted current pattern is compared with the baseline pattern to obtain the prediction accuracy. Repeat the training and prediction steps, and take the average of the prediction accuracies obtained multiple times to obtain the pattern conformity rate sequence. With time as the independent variable and the aforementioned compliance rate as the dependent variable, the slope was estimated using Thiel-Sen robust linear regression. If the slope is negative, the regression line is used as the dynamic baseline; if the slope is non-negative, the median of the compliance rate of all cycles is used as the fixed baseline.

8. The method as described in claim 1, characterized in that, The adaptive discrete point identification step also includes an emission pattern discrete point identification step, specifically including: Calculate the residuals by calculating the difference between the current long-term baseline value and the current compliance rate, and obtain the residual sequence. Analyze the dispersion of the residuals and calculate the interquartile range (IQR) of the residual sequence. If IQR = 0, recalculate the IQR for the subset of residual sequences that are greater than the median. Calculate the discrete factor. If the discrete factor is greater than 3, trigger an abnormal alarm.

9. An anomaly identification system based on automatic monitoring data of stationary pollution sources, characterized in that, include: The data preparation module is used to continuously extract online monitoring data from each emission outlet and the operating condition markers of each data point according to a preset frequency or cycle, so as to form a data sequence with timestamps. The multi-dimensional anomaly screening module is used to perform differential statistical tests on normal emission conditions, maintenance and calibration conditions and start-up and shutdown conditions in the data sequence based on the emission limit standards of each emission outlet, so as to generate suspected anomaly subsequences. The emission pattern self-learning module is used to train the emission pattern recognition model with historical data and use segments that consistently meet standards and operate under stable conditions as positive samples to perform double Monte Carlo prediction in order to obtain the pattern compliance rate sequence. The dynamic baseline construction and discrete point identification module is used to perform robust trend estimation based on the pattern conformity rate sequence to obtain a dynamic baseline that changes over time. Based on the dynamic baseline, the module determines the credibility reference of the emission pattern that the pollution source should meet under normal operating conditions and calculates the residual of the credibility of the emission pattern relative to the dynamic baseline. When the residual exceeds the adaptive discrete threshold, an abnormal alarm signal is output. The alarm output module is used to push abnormal information reports.

10. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed, implement the method as described in any one of claims 1-8.

Images