A method for detecting abnormal industrial waste gas emissions
By establishing a baseline spectrum library and an operating condition emission constraint rule library, and combining sliding window and dynamic time warping algorithms, the problem of high false alarm rate in existing technologies has been solved, and accurate anomaly detection and source tracing of industrial waste gas emissions has been achieved.
Patent Information
- Application Number
- CN202611360818.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, abnormal industrial waste gas emission detection methods rely on single-point concentration threshold judgments, which cannot be combined with production conditions for identification and evaluation, resulting in a high false alarm rate and affecting the accuracy and reliability of alarm results.
By collecting historical exhaust emissions and operating condition data, a baseline map library and an operating condition emission constraint rule library are established. The sliding window algorithm and dynamic time warping algorithm are used for feature extraction and comparison to generate emission anomaly detection results, thereby realizing dynamic monitoring and anomaly identification of emission behavior.
It improves the accuracy and reliability of emission anomaly detection, enabling the identification of early, slowly changing, or more concealed process anomalies, reducing false alarms, and providing accurate anomaly level classification and source location.
Smart Images

Figure CN123116091A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of anomaly detection technology, specifically to a method for detecting anomalies in industrial waste gas emissions. Background Technology
[0002] With the large-scale and continuous development of industrial manufacturing, many industries such as chemical, metallurgical, and power generally generate multi-component complex industrial waste gases during production. The emission behavior of such waste gases is characterized by strong concentration fluctuations and complex component coupling. Furthermore, its emission status is closely related to operating conditions such as production load, equipment start-up and shutdown, and raw material ratio, exhibiting significant operating condition-driven dynamic characteristics.
[0003] In existing technologies, the detection of abnormal industrial waste gas emissions generally adopts a single-point concentration threshold determination method. This method sets a fixed alarm upper limit for the concentration of one or more pollutants, and triggers an alarm signal when the real-time monitoring data exceeds this set value.
[0004] However, single-point concentration threshold determination methods rely solely on instantaneous concentration values and cannot identify and assess the changes and patterns of emission data in conjunction with specific production conditions. Consequently, when normal adjustments to production load or equipment switching cause short-term fluctuations in pollutant concentrations within a reasonable range, single-point concentration threshold determination methods are highly prone to generating false alarms. Ultimately, this results in a persistently high false alarm rate for anomaly detection, severely impacting the accuracy and reliability of alarm results. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a method for detecting abnormal industrial waste gas emissions, thus solving the problems described in the background technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting abnormal industrial waste gas emissions, comprising the following steps: Step S1: Collect historical exhaust gas emission monitoring data and historical operating condition data, and establish a baseline spectrum library based on the historical exhaust gas emission monitoring data and historical operating condition data; Step S2: Collect historical normal fluctuation data, and construct a working condition emission constraint rule library based on the baseline spectrum library and historical normal fluctuation data; Step S3: Collect real-time exhaust emission monitoring data and real-time operating condition data. Based on the real-time operating condition data, retrieve the matching baseline behavior feature vector and the allowable offset range from the operating condition emission constraint rule library. Extract features from the real-time exhaust emission monitoring data to generate a real-time behavior feature vector. Based on the allowable offset range, compare the real-time behavior feature vector with the matching baseline behavior feature vector to determine the emission anomaly event. Step S4: Classify and locate the abnormal emission events according to their abnormality level, and generate abnormal emission detection results.
[0007] Preferably, the feature is that collecting historical exhaust emission monitoring data and historical operating condition data includes: Historical exhaust gas emission monitoring data generated during the target industrial facility's normal and stable operation period are collected. This data is obtained through online monitoring instruments installed on the exhaust gas treatment unit's outlet manifold, specifically including concentration signals output from a CO concentration analyzer employing non-dispersive infrared principles. The temperature signal output by the Pt100 platinum resistance temperature sensor and the flow signal output by the thermal gas mass flow meter. All signals are sampled at a frequency of 1Hz and aligned according to a unified time base; simultaneously, historical operating condition data, including production load levels, are obtained from the distributed control unit. Key equipment start / stop status set and the operating mode indicator of the exhaust gas treatment unit. .
[0008] Preferably, the baseline map library is established based on the historical exhaust gas emission monitoring data and historical operating condition data, including: Based on historical operating data, the continuous production process is divided into multiple comparable operating periods. The partitioning process is implemented by a state machine logic that continuously monitors the operating condition state vector. When the value of any element in this vector changes, the current segment is considered to have ended and a new segment is considered to have started. ; For each comparable runtime segment A set of quantitative features is extracted from the monitoring data it contains to form a digital feature vector. This vector represents the normal emission behavior pattern under this operating condition and becomes a baseline entry. For each segment Assign a running status identifier Encapsulate the information of each segment into a baseline record. All baseline records are stored in a relational database, forming a baseline map library.
[0009] Preferably, historical normal fluctuation data is collected, and based on the baseline spectrum library and historical normal fluctuation data, a base of emission constraint rules for operating conditions is constructed, including: Collect historical normal fluctuation data, and for each group in the baseline map library with the same operating condition label and running status identifier Baseline set , m is the number of corresponding baseline entries in the baseline map library, and the feature vectors of all entries in the analysis set are used. The normal fluctuation range of each feature dimension is determined statistically, which serves as the allowable offset range. Construct deviation type-failure cause mapping rules, based on process mechanism knowledge and historical failure maintenance records, summarize common emission behavior deviation patterns, and associate them with the most likely failure causes; For each working condition-state combination Generate a complete constraint rule record. All constraint rules are recorded. The set of rules constitutes the emission constraint rule library for operating conditions.
[0010] Preferably, the collection of real-time exhaust emission monitoring data and real-time operating condition data includes: Two types of real-time data streams are collected simultaneously. The first type is real-time operating condition data, which is acquired from the production distributed control unit via the OPCDA protocol interface at a frequency of once per second, including the current production load level. Current start / stop status set of key equipment and the current operating mode indicator of the exhaust gas treatment unit. The second category is real-time exhaust emission monitoring data, which is synchronously acquired from online monitoring instruments installed at the outlet of the treatment unit, collecting instantaneous readings of pollutant concentrations at the same 1Hz frequency. Exhaust gas temperature reading and volumetric flow rate readings .
[0011] Preferably, based on real-time operating condition data, the matching baseline behavior feature vector and allowable offset range are determined by retrieving data from the operating condition emission constraint rule base, including: For the current time window Based on the dominant real-time operating condition data within the window, generate current operating condition labels. Meanwhile, based on the concentration sequence within the window Calculate its variance With trend slope Determine the running status indicator of the current window based on the threshold. It can be classified as a stable segment, a transitional segment, or a fluctuating segment; With binary Using the joint search key, a query is performed in the emission constraint rule base for operating conditions. This query yields two key outputs: the matching behavioral baseline feature vector and the allowed offset range.
[0012] Preferably, feature extraction is performed on real-time exhaust emission monitoring data to generate real-time behavioral feature vectors, including: For the current window The pre-processed real-time monitoring data is used to perform feature extraction calculations; specifically, the mean of the concentration sequence within the window is calculated. with standard deviation The trend slope was fitted using the least squares method. And calculate the Pearson correlation coefficient between the temperature and flow rate series. Together, they constitute the real-time behavior feature vector. .
[0013] Preferably, based on an allowed offset range, the real-time behavioral feature vector is numerically compared with the matched baseline behavioral feature vector to determine the emission anomaly event, including: To make a preliminary judgment, i.e., to examine the real-time behavioral feature vector. Whether each dimension exceeds the allowed offset range obtained from the retrieval; Next, a secondary decision is made to calculate the normalized morphological difference degree. ,Will The threshold was obtained by statistically analyzing a large amount of historical normal data. Compare; If the current window If the deviation flag is triggered, a structured candidate exception event record will be generated. Consecutive candidate events with a time interval of less than 1 minute and deviating from the same set of label types are merged into a single emission anomaly event. .
[0014] Preferably, the abnormal emission events are classified into abnormal levels and their sources are located, generating abnormal exhaust emission detection results including: The input list of emission anomalies is aggregated to identify multiple emission anomalies that are consecutive or overlap in time. If the matching operating conditions are the same and the sets of triggering deviation marker types are highly similar, then these events are merged into a single comprehensive abnormal operating condition segment. ; For each abnormal operating condition segment Anomalies are classified into different levels, and the classification is based on a comprehensive score of three quantitative dimensions: persistence, deviation, and risk impact. After the classification is completed, the source tracing and location are performed. This process directly calls the constructed deviation type-fault cause mapping rule set, based on the abnormal operating condition segment. Operating condition label Status indicators The set of deviation marker types that are actually triggered is used as input to perform matching queries in the rule base, generating structured exhaust gas emission anomaly detection results.
[0015] Preferably, the abnormal exhaust emission detection results include: Error number Time of occurrence; Description of operating condition, operating condition label at the time of the anomaly. With running status identifier ; Abnormal level Main deviation features; source tracing and localization results.
[0016] This invention provides a method for detecting abnormal industrial waste gas emissions, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) By using the sliding window algorithm to process real-time data streams, continuous monitoring data can be divided into a series of fixed-length or variable-length analysis windows. This processing method enables subsequent feature extraction and comparison operations to focus on specific and limited time periods, effectively avoiding feature ambiguity caused by excessively long data or gradual changes in operating conditions, thus ensuring the timeliness and relevance of feature extraction and laying the foundation for real-time and accurate anomaly detection.
[0017] (2) By introducing a dynamic time warping algorithm for sequence morphology comparison, the nonlinear scaling, translation, or local deformation that may exist between real-time monitoring sequences and historical baseline sequences on the time axis can be overcome. This method does not require strict alignment of data points, but rather seeks the optimal matching path between sequences, thereby more accurately assessing the similarity of the overall morphology and significantly improving the ability to identify anomalies in complex emission patterns with time-varying characteristics.
[0018] (3) By improving the dynamic time warping algorithm to derivative-constrained dynamic time warping, the original numerical sequence comparison is transformed into a comparison of the first derivative of the sequence. This improvement can effectively filter out the absolute numerical deviation of the sequence and focus on capturing deeper morphological anomalies such as rate of change and fluctuation patterns. Therefore, even if the monitored value is close to the baseline in numerical value, but there are differences in its change process such as rapid, sluggish or jittery changes, it can be detected sensitively, enhancing the ability to detect early, slow-changing anomalies or more concealed process anomalies. Attached Figure Description
[0019] Figure 1 This is a flowchart of an industrial waste gas emission anomaly detection method proposed in this invention.
[0020] Figure 2 This invention provides a hierarchical diagram of the emission constraint rule library for an industrial waste gas emission anomaly detection method.
[0021] Figure 3 This is a hierarchy diagram of the abnormal emission detection results obtained in the industrial waste gas emission anomaly detection method proposed in this invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Please see Figures 1-3 This invention provides a technical solution: a method for detecting abnormal industrial waste gas emissions. Specifically, the method for detecting abnormal industrial waste gas emissions is provided below; please refer to [link / reference needed]. Figure 1 The method includes the following steps: Step S1: Collect historical exhaust gas emission monitoring data and historical operating condition data, and establish a baseline spectrum library based on the historical exhaust gas emission monitoring data and historical operating condition data.
[0024] First, historical exhaust gas emission monitoring data generated during the target industrial facility's normal and stable operation period are collected through online monitoring instruments installed at the exhaust gas treatment unit's outlet manifold. Specifically, this includes the concentration signal output from a CO concentration analyzer employing non-dispersive infrared (NDIR) principles. The temperature signal output by the Pt100 platinum resistance temperature sensor and the flow signal output by the thermal gas mass flow meter. All signals are sampled at a frequency of 1Hz and aligned to a unified time base. Historical operating data, including production load levels, is also acquired from the distributed control unit. Key equipment start / stop status set and the operating mode indicator of the exhaust gas treatment unit. .
[0025] Subsequently, the collected raw time-series data were preprocessed using Python's Pandas library. The preprocessing process first resampled and aligned all data streams based on a unified timestamp, then applied a sliding window outlier removal method based on the 3σ principle for concentration sequences. The mean and standard deviation of the data were calculated within a 5-minute sliding window. Readings that deviated from the mean by more than 3 times the standard deviation were removed, and missing points were filled using linear interpolation.
[0026] After preprocessing, the continuous production process is divided into multiple comparable operating periods based on historical operating data. The partitioning process is implemented by a state machine logic that continuously monitors the operating condition state vector. When the value of any element in this vector changes, the current segment is considered to have ended and a new segment is considered to have started. Each segment The internal operating context remains strictly consistent.
[0027] Next, for each comparable runtime segment... A set of quantitative features is extracted from the monitoring data it contains to form a digital feature vector. This vector represents the normal emission behavior pattern under this operating condition and becomes a baseline entry. Feature extraction covers the core morphological dimension of the behavior, specifically calculating the following features: For concentration levels and fluctuation characteristics, calculate the mean and standard deviation of the concentration sequence within the fragment:
[0028]
[0029] in, For fragments The total number of data points within; This is the arithmetic mean of the concentrations, used to characterize the average level of emission concentrations in subsequent steps; denoted as the sample standard deviation of the concentration, used to characterize the intensity of fluctuations in emission concentrations in subsequent steps. The purpose of the formula is to quantify the central tendency and dispersion of emission concentrations.
[0030] To determine the trend of concentration changes, the slope of the trend is obtained through linear fitting:
[0031] in, The average of the time indices; This represents the slope of the concentration trend. The purpose of the formula is to quantify the average rate of change of concentration over time, used to identify rapid increases or decreases in concentration.
[0032] Flow-temperature correlation characteristics: Calculate the Pearson correlation coefficient between flow and temperature series to quantify their co-variation relationship.
[0033]
[0034] in, Temperature With traffic The mean within the segment; This is the Pearson correlation coefficient between the two parameters. The purpose of the formula is to quantify the degree of linear correlation between temperature and flow rate, and to assess their synergistic relationship.
[0035] Then, for each segment Assign a running status identifier This is to allow for more accurate baseline matching in subsequent step S3. Identification is based on the inherent characteristics of concentration data within a fragment: the concentration variance is calculated. with the absolute value of the trend slope By statistically analyzing all stable periods in historical data from the past three months, the 95th percentile of the historical distribution of concentration variance was set as the threshold. Similarly, the lower bound of the trend slope that clearly indicates the switching of operating conditions is set to... The judgment logic adopts a hierarchical and partitioned approach. First, based on the absolute value of the trend slope, segments are classified into two categories: trend-dominant or non-trend-dominant. Within the non-trend-dominant segments, stable segments and fluctuating segments are further distinguished based on concentration variance, as detailed below: If the absolute value of the trend slope If so, the current segment is determined to be a transition segment, indicating that the concentration within the segment shows a significant unidirectional change trend, and the dominant driving force is the switching of operating conditions or load adjustment; If the absolute value of the trend slope Then, further determine the concentration variance. If concentration variance If the concentration within the segment is within a certain range, it is considered a stable segment, indicating that the concentration within the segment has no obvious trend and the fluctuation range is within the normal allowable range; otherwise, it is considered a fluctuating segment, indicating that the concentration within the segment has no obvious trend but there is a certain degree of random fluctuation or periodic disturbance.
[0036] in, This is the concentration variance threshold. The two thresholds are set as the lower bound of the trend slope, respectively, by taking the 95th and 90th percentiles of the statistical distribution of all normal operation segments in the historical data of the past three months.
[0037] Finally, the information from each segment is encapsulated into a baseline record. :
[0038] in, This is the set of operating condition labels corresponding to this segment; This serves as an indicator of the running status. This is its behavioral feature vector. All baseline records are stored in the relational database MySQL, forming a structure that can be ordered by... Baseline map libraries for joint retrieval.
[0039] In subsequent steps, step S2 will be based on all similar fragments in this library. Calculate the normal fluctuation range for each feature to construct rules; step S3 will be based on real-time operating conditions. Retrieve matching data from this database. and To perform feature-by-feature comparison; the source analysis in step S4 will point to possible causes based on the specific features of the deviation from the anomaly.
[0040] Step S2: Collect historical normal fluctuation data, and construct a working condition emission constraint rule library based on the baseline spectrum library and historical normal fluctuation data.
[0041] Building upon the baseline map library established in step S1, this step aims to quantify and define the permissible normal fluctuation boundaries under corresponding operating conditions for the normal behavior pattern represented by each baseline entry, and to pre-establish the mapping relationship between different deviation patterns and potential fault causes. Together, these constitute the rule knowledge base for subsequent real-time judgment and automated tracing.
[0042] First, historical normal fluctuation data are collected. This data comes from the same data source as in step S1, but the selection scope focuses on comparable historical operating time segments that have been jointly determined to be normal by long-term operating records and process experts, and whose emission monitoring data show reasonable fluctuations in the feature dimensions extracted in step S1. These segments constitute the statistical basis for constructing behavioral fluctuation boundaries and verifying the mapping relationship.
[0043] Subsequently, based on the structured baseline map library output in step S1, the following core operations are performed to construct constraint rules: For each group in the baseline map library with the same operating condition labels and running status identifier Baseline set , m is the number of corresponding baseline entries in the baseline map library, and the feature vectors of all entries in the analysis set are used. The normal fluctuation range for each feature dimension is determined statistically, serving as the allowable offset range. The specific calculation is as follows: For the allowable interval of concentration mean, calculate the concentration mean of all values in the set. 2.5 percentile and the 97.5th percentile This forms an allowable range:
[0044] in, This indicates that 2.5% of historical normal samples had a mean concentration lower than or equal to this value; This indicates that 97.5% of historical normal samples had a concentration mean lower than or equal to this value. The purpose of this range is to define the acceptable normal fluctuation range of the average emission concentration under specific operating conditions and states, covering 95% of historical scenarios, and providing a statistical basis for subsequent judgment on whether the real-time concentration level deviates significantly.
[0045] For the allowed range of trend slopes, calculate all trend slopes in this set. 2.5 percentile and the 97.5th percentile This forms an allowable range:
[0046] in, and The lower and upper bounds of the historical normal trend rate of change were defined. The purpose of this interval is to quantify the acceptable normal range of concentration change trend under specific operating conditions, for subsequent identification of atypical excessively fast or slow changes.
[0047] Regarding the allowable boundaries of the correlation coefficient, the Pearson correlation coefficient This is a two-sided indicator with upper and lower bounds; its normal fluctuation range should consider both positive and negative correlations. Calculate all temperature-flow correlation coefficients in this set. 5th percentile and the 95th percentile Form the allowable interval for the correlation coefficient:
[0048] in, and The lower and upper bounds of the intensity of the coordinated changes in temperature and flow rate in historical normal data were defined. The lower bound is used to identify a weakening or loss of the coordinated relationship, while the upper bound is used to identify an abnormally strong coordinated relationship. This interval is directly used for anomaly detection of the real-time correlation coefficient in step S3. When the correlation coefficient calculated from the real-time data is lower than... or higher At that time, it was determined that the normal linkage between temperature and flow rate was disrupted.
[0049] Secondly, deviation type-failure cause mapping rules are constructed. This step aims to provide a predefined reasoning path for the structured tracing in step S4. Based on process mechanism knowledge and historical failure maintenance records, common emission behavior deviation patterns are summarized and associated with the most likely failure causes. Each mapping rule consists of two parts: a triggering condition and a pointed cause.
[0050] Take several typical mapping rules as examples: Mapping rule R1 is if the real-time behavior exhibits a trend slope. Exceeding the upper limit of the normal range, meaning the rate of change is too rapid, and the correlation coefficient... If the correlation coefficient falls below the allowable lower bound, indicating a loss of directional coordination, it is mapped to a fault path of valve jamming or sudden leakage in the process pipeline. The mechanism behind this rule is that valve malfunction or pipeline leakage can lead to uncontrolled and drastic changes in flow or concentration, thereby disrupting the originally stable temperature-flow coordination relationship.
[0051] The mapping rule R² is if the real-time behavior is represented by the concentration mean. Exceeding the upper limit of the allowable range for average concentration, meaning the concentration level remains consistently high, and the trend slope is... If the trend slope is within the allowable range, meaning the trend is within the normal range, then the fault path is attributed to a decrease in the efficiency of the exhaust gas treatment unit or a breakthrough. This fault may specifically involve equipment such as adsorption towers or catalysts. The mechanism behind this rule is that the performance degradation of the treatment equipment leads to a decrease in its baseline removal rate of pollutants, resulting in an overall increase in the baseline level of the outlet concentration, while the process may be relatively stable.
[0052] Mapping rule R3 is if the real-time behavior is represented by the concentration standard deviation. Higher than the corresponding operating condition in the baseline map library If the fluctuation is abnormally aggravated and the fluctuation is related to the operating frequency of rotating equipment such as fans, then the cause of the failure is mapped to abnormal fan operating conditions or drastic fluctuations in pipeline pressure.
[0053] Finally, the outputs from the two parts are integrated and stored in a structured format. For each working condition-state combination... Generate a complete constraint rule record. The record contains: Behavioral characteristics allowable range And deviation type-fault cause mapping rule set, .
[0054] Meanwhile, for each working condition-state combination, the baseline entry whose eigenvector norm is closest to the set mean is selected from the baseline entry set corresponding to the combination, and its complete concentration time series is extracted and stored as the standard comparison template sequence for the combination, for use in the dynamic time warping comparison step based on the first derivative in step S3.
[0055] All constraint rules record The collection of these rules constitutes the emission constraint rule base for operating conditions. This rule base is linked to the baseline map library from step S1 via key-value pairs. Perform correlation indexing. Its core function is to provide a dual interpretation system for subsequent steps, which includes quantified normal boundaries and abnormal interpretation knowledge: Step S3 will call the allowed interval of behavioral features for real-time comparison and initial screening of anomalies; Step S4 will directly call the deviation type-fault cause mapping rule set to automatically trace and locate the source of the selected abnormal events.
[0056] Step S3: Collect real-time exhaust emission monitoring data and real-time operating condition data. Based on the real-time operating condition data, retrieve the matching baseline behavior feature vector and the allowable offset range from the operating condition emission constraint rule library. Extract features from the real-time exhaust emission monitoring data to generate a real-time behavior feature vector. Based on the allowable offset range, compare the real-time behavior feature vector with the matching baseline behavior feature vector to determine the emission anomaly event.
[0057] Following the baseline map library established in step S1 and the operating condition emission constraint rule library constructed in step S2, this step performs online real-time monitoring. Its purpose is to utilize the established normal behavior knowledge system to continuously monitor and preliminarily identify anomalies in the currently operating exhaust emissions. First, two types of real-time data streams are collected simultaneously. The first type is real-time operating condition data, acquired from the production distributed control unit (DCS) via the OPCDA protocol interface at a frequency of once per second, including the current production load level. Current start / stop status set of key equipment and the current operating mode indicator of the exhaust gas treatment unit. The second category is real-time exhaust emission monitoring data, which is synchronously acquired from an online monitoring instrument installed at the outlet of the treatment unit. This online monitoring instrument is consistent with the model, accuracy, and installation location selected in step S1, and collects instantaneous readings of pollutant concentrations at the same 1Hz frequency. Exhaust gas temperature reading and volumetric flow rate readings .
[0058] After acquiring the raw data, a data preprocessing procedure identical to step S1 is immediately executed to ensure comparable data quality. This procedure is implemented using Python's Pandas library. Specifically, it involves resampling and aligning multiple data streams based on a unified timestamp, and applying a sliding window outlier removal method based on the 3σ principle. That is, for concentration data within a 5-minute window, the mean and standard deviation are calculated, transient outlier readings deviating from the mean by more than 3 times the standard deviation are removed, and missing points are filled using linear interpolation.
[0059] Subsequently, a sliding time window with a length of 5 minutes and a sliding step size of 30 seconds was used. The following core operation process is executed cyclically, unit by unit: First, real-time operating condition matching and rule retrieval, targeting the current time window. Based on the dominant real-time operating condition data within the window, generate current operating condition labels. Meanwhile, based on the concentration sequence within the window... Calculate its variance With trend slope Based on the same threshold used in step S1 and Determine the running status indicator of the current window. It can be classified as a stable segment, a transitional segment, or a fluctuating segment.
[0060] Next, using binary pairs As the joint retrieval key, a query is performed in the emission constraint rule base for operating conditions constructed in step S2. This query yields two key outputs: the matching behavioral baseline feature vector and the allowed offset range. Matching behavioral baseline feature vector This vector originates from the baseline map library in step S1, which corresponds exactly to this operating condition, and represents the historical normal emission behavior pattern under this condition.
[0061] Allowable offset range, including the allowable interval of average concentration. Allowable range of trend slope and the lower bound of the correlation coefficient These boundary values come directly from the records made for the corresponding rules in step S2. The quantified description.
[0062] Second, real-time trajectory reconstruction and feature extraction, for the current window The pre-processed real-time monitoring data is used to perform the same feature extraction calculation as in step S1. Specifically, the mean of the concentration sequence within the window is calculated. with standard deviation The trend slope was fitted using the least squares method. And calculate the Pearson correlation coefficient between the temperature and flow rate series. These calculated features collectively constitute the real-time behavior feature vector. This vector digitally and comprehensively describes the actual emission behavior within the current window. Simultaneously, it provides a preprocessed, complete window concentration sequence. This will be retained for subsequent morphological comparison. Third, based on improved dynamic time warping and refined morphological comparison, this step introduces an improved direct morphological similarity comparison step to more sensitively capture early anomalies related to the change process that are difficult to detect with traditional feature comparisons. For each operating condition-state combination, when constructing the operating condition emission constraint rule base in step S2, the entry with the feature vector norm closest to the set mean is selected from the baseline entry set corresponding to that combination, and its complete concentration time series is used as the benchmark concentration series for that combination. This data is stored for later retrieval during real-time monitoring. This step involves storing the complete concentration time series for the current window. Compared with the baseline concentration time series stored under the same operating conditions and retrieved from the baseline spectrum library. A comparison is performed. The comparison uses a dynamic time warping algorithm based on the first derivative. First, the first-order approximate derivative sequences of the two sequences are estimated to highlight their changing trends. To more robustly reflect local changing trends in noisy environments, the following comprehensive estimation method is adopted:
[0063] in, The value range is 2 to , The length of the sequence; Represents a sequence At the point of time The estimated first derivative at that point; Represents the sequence at time points The value of is determined by this estimation method, which, based on the standard central difference scheme, additionally introduces the change of the current point relative to the previous time point. As a weighted term, and with the backward difference term By integrating the data, on the one hand, high-frequency measurement noise is suppressed from interfering with derivative estimation through weighted averaging; on the other hand, the instantaneous change trend at the current point is given appropriate weight in the estimation result, thus more sensitively reflecting the true trend of emission concentration within the local time window.
[0064] Subsequently, in the derivative sequence and The standard dynamic time warping algorithm is applied to find the optimal nonlinear alignment path and calculate the cumulative distance along this path. The process is as follows: Let the real-time derivative sequence be ,in Let the baseline derivative sequence be... ,in , Represents the real-time derivative sequence Length; Represents the baseline derivative sequence The length. First, construct a... Local distance matrix Its elements Let represent the squared Euclidean distance between two points. Then, the cumulative distance matrix is calculated recursively using dynamic programming. Its elements Ultimately, the cumulative distance... .
[0065] in, and The real-time and baseline derivative sequences are respectively at the 1st... and Derivative estimates at each time point; Indicates the local difference between two points; Indicates starting from the origin Time The minimum cumulative distance. Formula The goal is to find the path with the lowest cost that aligns the two sequences in terms of overall trend of change through dynamic programming. The total cost of this path is represented by a value that indicates a greater similarity in the overall shape of the two derivative sequences. A larger value indicates a greater difference in their rhythm, phase, or shape.
[0066] The purpose of this comparison is to show that even if the final statistical characteristics of real-time concentrations, such as the mean, are different... Falling within the allowed range If the rhythm of its rise or fall, such as rapid, sluggish, or trembling, differs from the historical normal pattern, then... There are differences, DDTW distance This will also significantly increase the ability to effectively identify such early or slowly changing anomalies hidden at the level of change dynamics.
[0067] Fourth, multi-criteria fusion and abnormal event generation: combining feature boundary criteria and morphological similarity criteria to generate candidate abnormal events.
[0068] The first step is to make a preliminary judgment, which involves examining the real-time behavioral feature vector. Are any of the dimensions outside the allowed offset range obtained from the retrieval? like If the concentration level deviates; If the rate of change of the marker deviates; if If this happens, the consistency of the marking direction will deviate.
[0069] Next, a secondary decision is made to calculate the normalized morphological difference degree. :
[0070] in, For window The length is measured in terms of the number of data points within the window. With threshold The threshold is obtained through statistical analysis of a large amount of historical normal data. The specific value can be historical. The 99th percentile of the value. If If so, the marked trajectory shape deviates.
[0071] Finally, if the current window If any of the above deviation flags is triggered, a structured candidate exception event record will be generated. This record contains the event start time, end time, the set of all deviation marker types triggered, and real-time feature vectors. Matching baseline feature vectors And its corresponding allowable range boundaries. To reduce event fragmentation, consecutive candidate events with a time interval of less than 1 minute and deviating from the same tag type set are merged into a more representative emission anomaly event. .
[0072] The final output of this step is a list of emission anomalies sorted by time. Each event All clearly include the time interval of the anomaly, the specific normal behavior rules violated, and quantitative evidence of deviation. The normal behavior rules are associated with the matching working condition labels and operating status identifiers. The deviation evidence refers to the specific characteristic value deviation or morphological distance exceeding the standard, thus providing clear, reliable, and complete information input for step S4 to perform anomaly level classification and accurate source tracing.
[0073] Step S4: Classify and locate the abnormal emission events according to their abnormality level, and generate abnormal emission detection results.
[0074] Following the list of abnormal emission events output in step S3, this step is the final decision-making stage. Its core task is to assess the severity of each abnormal event, infer its potential root causes, and generate a diagnostic report that can be directly used by engineers.
[0075] First, the input list of emission anomaly events is aggregated. This is done for multiple emission anomaly events that are consecutive or overlapping in time. If its matching operating conditions If the sets of identical and triggering deviation marker types are highly similar, then these events will be merged into a single comprehensive abnormal condition segment. Aggregation operations reduce alarm fragmentation, allowing subsequent comprehensive analysis to focus on consistently occurring periods of anomaly. Includes its start and end times and duration. Standardized operating condition indicators Status indicators , and a complete set of feature deviations containing evidence of deviations from all sub-events.
[0076] Subsequently, for each abnormal operating condition section Anomaly severity is classified. The classification is based on a comprehensive score using three quantitative dimensions, each dimension reflecting one aspect of the severity of the anomaly.
[0077] For the persistence dimension score The score is based on the duration of the anomaly. Set time threshold and ,in The preset time is 5 minutes. The preset time is 30 minutes. If... ,but , indicates a transient anomaly; if ,but , indicates a persistent abnormality; if ,but This indicates a long-term abnormality. Among them, abnormal operating condition section Total duration, and This is a pre-defined time boundary point based on process tolerance and alarm response requirements. The purpose of this dimension is to distinguish between long-term anomalies requiring emergency handling and observable short-term disturbances.
[0078] about =5 minutes =Basis for the 30-minute value: =5 minutes. In industrial waste gas continuous emission monitoring systems, the conventional time scale for single data acquisition and sliding window analysis is in the minute range. Concentration fluctuations lasting no more than 5 minutes usually originate from normal process disturbances such as raw material batch differences, purging switching, and load fine-tuning. Such fluctuations account for a large proportion of historical normal operation data and do not have continuous deviation characteristics. Classifying them as transient anomalies can effectively avoid excessive alarms.
[0079] =30 minutes. Based on the process response time of typical waste gas treatment facilities such as regenerative thermal oxidizers, catalytic combustion devices, and adsorption-desorption systems, the process from a substantial deviation from normal operating conditions to a sustained and significant change in outlet concentration typically takes 15 to 30 minutes. Using 30 minutes as a threshold, anomalies exceeding this duration indicate that the treatment unit or upstream process has been in a continuously abnormal state, requiring manual intervention. This timescale also aligns with the typical response cycle for industrial operators from alarm confirmation to on-site inspection.
[0080] For deviation dimension scoring The score is based on the magnitude of feature value deviation. The allowable interval for the key feature is calculated relative to the value retrieved in step S3. Boundary exceeding proportion:
[0081] in, Abnormal period Average observed concentration within the body; and Defined for step S2, and related to the current operating condition. The corresponding average concentration fluctuates normally within the upper and lower bounds. This is the normalized deviation ratio; a value greater than 0 indicates a deviation, and a larger value indicates a more significant deviation. Similarly, the trend slope can be calculated. Correlation coefficient deviation ratio and Take the largest deviation ratio among all features. And mapped to discrete scores: if ,but like ,but like ,but The purpose of this dimension is to quantify the extent to which anomalous behavior disrupts the normal baseline; Scoring based on risk impact dimension The rating is based on the potential risk implied by the type of deviation. Different deviation indicators are assigned different base risk weights. For example, deviations in concentration levels are directly associated with the risk of exceeding environmental limits and are assigned a higher base weight. A deviation in the rate of change indicates an impending equipment failure; this should be assigned a moderate weight. Trajectory morphology deviation indicates early, weak anomalies, which are then assigned basic weights. Abnormal operating condition section Risk impact score This is the maximum value of the base weights corresponding to all deviation identifiers it contains. The purpose of this dimension is to assess the severity of the anomaly from an engineering consequence perspective.
[0082] Finally, the abnormal operating condition section The anomaly level is determined by the following formula:
[0083] in, The scores are based on three dimensions: persistence, deviation, and risk impact. This is a rounding function. The calculated Level value is 1, 2, or 3, corresponding to the three levels of warning, alarm, and severe alarm, respectively.
[0084] After completing the classification, source tracing is performed. This process directly calls the pre-built deviation type-fault cause mapping rule set in step S2. This is based on the abnormal operating condition segment. Operating condition label Status indicators The set of deviation marker types that actually trigger the error is used as input to perform a matching query in the rule base. For example, if If the deviation marker set is a deviation in rate of change or a deviation in direction consistency, then it matches rule R1 from the example in step S2, thereby outputting the mapped fault cause path: valve jamming or sudden leakage in the process pipeline. Each mapping rule provides a logical chain from the abnormal phenomenon to the possible root cause, making the tracing results interpretable.
[0085] It should be noted that during the query process, the set of deviation marker types actually triggered by the abnormal operating condition segment is compared one by one with the triggering conditions of each mapping rule. The condition matching degree of each rule is calculated, and the matching degree is defined as the proportion of the number of deviation marker types that are matched in the rule's triggering conditions to the total number of triggering conditions for that rule. If there is a rule with a matching degree of 100%, the fault cause path corresponding to that rule is directly output. If there is no rule that matches completely, the rule with the highest matching degree is selected as the most likely cause and output, and the detection results are marked with deviation marker types that are partially matched and not matched, for reference during manual review. If there are multiple rules with the highest matching degree and the matching degree is the same, the fault cause paths corresponding to these rules are output together to ensure that no abnormal information is missed.
[0086] Finally, structured exhaust emission anomaly detection results are generated. These results are presented in report form, with each report record corresponding to an abnormal operating condition segment. It must contain at least the following fields: Error number Time of occurrence to ; Operating condition description, operating condition label when the abnormality occurs With running status identifier ; Abnormal level, The value is set to early warning, alarm, or severe alarm; the main deviation characteristics include core deviation markers and a quantitative description of the deviation of key characteristic values, such as the average concentration. Exceeding the allowed limit Exceeding the standard The source tracing and location results are one or more suspected fault sources and preliminary handling suggestions obtained by mapping from the rule base. For example, if the matching rule R1 is suspected to be valve jamming or pipeline leakage, it is recommended to first check the sealing of the valve and related pipelines.
[0087] This step achieves focused analysis through anomaly event aggregation, completes scientific classification based on a three-dimensional quantitative system of continuity, deviation, and risk impact, and realizes automated source tracing and positioning by combining a preset rule base. Finally, it outputs a structured and engineered anomaly detection report. This is the core link in transforming real-time detected anomaly data into actionable decisions for industrial sites, realizing a closed loop for the entire waste gas emission anomaly detection method from technical analysis to practical application.
[0088] This invention provides a method for detecting abnormal industrial waste gas emissions. The method first learns normal emission behavior patterns under different operating conditions from historical data, establishing a structured baseline spectrum library. Then, by analyzing historical normal fluctuation data, it quantifies and defines the allowable fluctuation boundaries for each operating condition pattern, and pre-defines the mapping relationship between different deviation patterns and potential fault causes, constructing an interpretable rule knowledge base.
[0089] During the real-time monitoring phase, this method employs a sliding window to process the real-time data stream and precisely matches the corresponding behavioral baseline and rules based on the current operating conditions. By combining eigenvalue comparison with an improved derivative dynamic time warping algorithm, it not only checks whether key parameters exceed the allowable range but also meticulously compares the morphological differences in the changing trends between the real-time data trajectory and the historical baseline, thereby identifying various anomalies such as concentration levels, rate of change, multi-parameter synergy, and trajectory morphology.
[0090] For the identified abnormal events, the method further conducts a comprehensive evaluation from three dimensions: persistence, deviation, and risk impact, automatically classifies the abnormality level, and infers the most likely source of failure based on preset mapping rules, and finally generates a structured detection report that includes the level, source location, and handling suggestions.
[0091] This method has evolved from traditional single-point threshold alarms into a technical system based on operational condition perception, behavioral pattern recognition, and interpretable tracing. It effectively solves the problems of high false alarm rate, lack of foresight, and weak interpretability in complex industrial environments, and achieves more accurate, stable, and engineering-guiding intelligent emission monitoring.
[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 statement "including a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting abnormal industrial waste gas emissions, characterized in that, Includes the following steps: Step S1: Collect historical exhaust gas emission monitoring data and historical operating condition data, and establish a baseline spectrum library based on the historical exhaust gas emission monitoring data and historical operating condition data; Step S2: Collect historical normal fluctuation data, and construct a working condition emission constraint rule library based on the baseline spectrum library and historical normal fluctuation data; Step S3: Collect real-time exhaust emission monitoring data and real-time operating condition data. Based on the real-time operating condition data, retrieve the matching baseline behavior feature vector and the allowable offset range from the operating condition emission constraint rule library. Extract features from the real-time exhaust emission monitoring data to generate a real-time behavior feature vector. Based on the allowable offset range, compare the real-time behavior feature vector with the matching baseline behavior feature vector to determine the emission anomaly event. Step S4: Classify and locate the abnormal emission events according to their abnormality level, and generate abnormal emission detection results.
2. The method for detecting abnormal industrial waste gas emissions according to claim 1, characterized in that, Collect historical exhaust emission monitoring data and historical operating condition data, including: Historical exhaust gas emission monitoring data generated during the target industrial facility's normal and stable operation period are collected. This data is obtained through online monitoring instruments installed on the exhaust gas treatment unit's outlet manifold, specifically including concentration signals output from a CO concentration analyzer employing non-dispersive infrared principles. The temperature signal output by the Pt100 platinum resistance temperature sensor and the flow signal output by the thermal gas mass flow meter. All signals are sampled at a frequency of 1Hz and aligned according to a unified time base; simultaneously, historical operating condition data, including production load levels, are obtained from the distributed control unit. Key equipment start / stop status set and the operating mode indicator of the exhaust gas treatment unit. .
3. The method for detecting abnormal industrial waste gas emissions according to claim 2, characterized in that, Based on the aforementioned historical exhaust emission monitoring data and historical operating condition data, a baseline spectrum library is established, including: Based on historical operating data, the continuous production process is divided into multiple comparable operating periods. The partitioning process is implemented by a state machine logic that continuously monitors the operating condition state vector. When the value of any element in this vector changes, the current segment is considered to have ended and a new segment is considered to have started. ; For each comparable runtime segment A set of quantitative features is extracted from the monitoring data it contains to form a digital feature vector. This vector represents the normal emission behavior pattern under this operating condition and becomes a baseline entry. For each segment Assign a running status identifier Encapsulate the information of each segment into a baseline record. All baseline records are stored in a relational database, forming a baseline map library.
4. The method for detecting abnormal industrial waste gas emissions according to claim 3, characterized in that, Historical normal fluctuation data are collected, and based on the baseline spectrum library and historical normal fluctuation data, a base of emission constraint rules for operating conditions is constructed, including: Collect historical normal fluctuation data, and for each group in the baseline map library with the same operating condition label and running status identifier Baseline set , m is the number of corresponding baseline entries in the baseline map library, and the feature vectors of all entries in the analysis set are used. The normal fluctuation range of each feature dimension is determined statistically, which serves as the allowable offset range. Construct deviation type-failure cause mapping rules, based on process mechanism knowledge and historical failure maintenance records, summarize common emission behavior deviation patterns, and associate them with the most likely failure causes; For each working condition-state combination Generate a complete constraint rule record. All constraint rules are recorded. The set of rules constitutes the emission constraint rule library for operating conditions.
5. The method for detecting abnormal industrial waste gas emissions according to claim 4, characterized in that, Collect real-time exhaust emission monitoring data and real-time operating condition data, including: Two types of real-time data streams are collected simultaneously. The first type is real-time operating condition data, which is acquired from the production distributed control unit via the OPCDA protocol interface at a frequency of once per second, including the current production load level. Current start / stop status set of key equipment and the current operating mode indicator of the exhaust gas treatment unit. The second category is real-time exhaust emission monitoring data, which is synchronously acquired from online monitoring instruments installed at the outlet of the treatment unit, collecting instantaneous readings of pollutant concentrations at the same 1Hz frequency. Exhaust gas temperature reading and volumetric flow rate readings .
6. The method for detecting abnormal industrial waste gas emissions according to claim 5, characterized in that, Based on real-time operating condition data, the matching baseline behavior feature vector and allowable offset range are determined by retrieving data from the operating condition emission constraint rule base, including: For the current time window Based on the dominant real-time operating condition data within the window, generate current operating condition labels. Meanwhile, based on the concentration sequence within the window Calculate its variance With trend slope Determine the running status indicator of the current window based on the threshold. It can be classified as a stable segment, a transitional segment, or a fluctuating segment; With binary Using the joint search key, a query is performed in the emission constraint rule base for operating conditions. This query yields two key outputs: the matching behavioral baseline feature vector and the allowed offset range.
7. The method for detecting abnormal industrial waste gas emissions according to claim 6, characterized in that, Feature extraction is performed on real-time exhaust emission monitoring data to generate real-time behavioral feature vectors, including: For the current window The pre-processed real-time monitoring data is used to perform feature extraction calculations, and the mean of the concentration sequence within the window is calculated. with standard deviation The trend slope was fitted using the least squares method. And calculate the Pearson correlation coefficient between the temperature and flow rate series. Together, they constitute the real-time behavior feature vector. .
8. The method for detecting abnormal industrial waste gas emissions according to claim 7, characterized in that, Based on the allowed offset range, the real-time behavioral feature vector is numerically compared with the matched baseline behavioral feature vector to determine emission anomaly events, including: To make a preliminary judgment, i.e., to examine the real-time behavioral feature vector. Whether each dimension exceeds the allowed offset range obtained from the retrieval; Perform secondary decision-making and calculate the normalized morphological difference degree. ,Will The threshold was obtained by statistically analyzing a large amount of historical normal data. Compare; If the current window If the deviation flag is triggered, a structured candidate exception event record will be generated. Consecutive candidate events with a time interval of less than 1 minute and deviating from the same set of label types are merged into a single emission anomaly event. .
9. The method for detecting abnormal industrial waste gas emissions according to claim 8, characterized in that, Anomaly events are classified into anomaly levels and their sources are located to generate abnormal exhaust emission detection results, including: The input list of emission anomalies is aggregated to identify multiple emission anomalies that are consecutive or overlap in time. If the matching operating conditions are the same and the sets of triggering deviation marker types are highly similar, then these events are merged into a single comprehensive abnormal operating condition segment. ; For each abnormal operating condition segment Anomalies are classified into different levels, and the classification is based on a comprehensive score of three quantitative dimensions: persistence, deviation, and risk impact. After the classification is completed, the source tracing and location are performed. This process directly calls the constructed deviation type-fault cause mapping rule set, based on the abnormal operating condition segment. Operating condition label Status indicators The set of deviation marker types that are actually triggered is used as input to perform matching queries in the rule base, generating structured exhaust gas emission anomaly detection results.
10. The method for detecting abnormal industrial waste gas emissions according to claim 9, characterized in that, The abnormal exhaust emission detection results include: Error number Time of occurrence; Description of operating condition, operating condition label at the time of the anomaly. With running status identifier ; Abnormal level Main deviation features; source tracing and localization results.
