Fixed pollution source automatic monitoring method and system and digital intelligent construction method thereof

Abstract

The invention relates to the technical field of pollution source monitoring, in particular to a stationary pollution source automatic monitoring method and system and a digital intelligent construction method thereof. S2, in a real-time monitoring stage, a multi-dimensional sensing array collects equipment health degree data and working condition parameters at high frequency, and a data collection device inputs real-time data into the ternary coupling reference model; s3, starting deviation traceability, and correcting equipment drift or working condition interference; s4, synchronously uploading the data to a remote operation and maintenance platform, pre-judging an equipment fault risk based on an equipment health degree change rate, and generating a maintenance work order 48 hours in advance; s5, for the predictive maintenance work order; and S6, automatically iterating the model monthly, incorporating the current-month effective monitoring data and maintenance records into the ternary coupling reference model, and optimizing the association weight. The method effectively guarantees the accuracy of the monitoring data of the stationary pollution source by constructing the ternary coupling reference model and combining the comparison and correction of the real-time monitoring data and the predicted value.

Description

Technical Field

[0001] This invention relates to the field of pollution source monitoring technology, and in particular to an automatic monitoring method, system and digital construction method for stationary pollution sources. Background Technology

[0002] Automatic monitoring of stationary pollution sources is a core means of ecological and environmental supervision, and the quality of its data directly determines the effectiveness of precise and scientific pollution control. Currently, my country's automatic monitoring of stationary pollution sources mainly relies on the traditional model of manual inspection combined with static data collection. This involves deploying equipment such as flue gas CEMS (Continuous Emission Monitoring System) and wastewater online monitoring instruments to collect pollutant concentration data, supplemented by regular on-site verification by maintenance personnel, to achieve basic control over pollution source emissions.

[0003] For example, patent document CN202311651648.0 discloses an online monitoring and analysis system for pollution sources. It uses multiple sets of storage tanks to store water samples in different categories, combines heating and stirring components to maintain a constant temperature for the water samples, and uses a water quality monitor and alarm module to achieve data acquisition and early warning of exceeding standards. Another example is patent document CN202311568957.1, which mainly discloses carbon emission monitoring. It optimizes the sampling angle through an angle adjustment mechanism and improves gas collection efficiency by combining a fan and a vent pipe.

[0004] However, the aforementioned existing technologies and traditional monitoring methods have the following drawbacks:

[0005] First, existing technologies mostly focus on improving single equipment or optimizing monitoring principles. When monitoring equipment experiences long-term operational drift or sudden changes in the operating conditions of pollution sources, it is prone to producing false exceedances or data distortion, and the root cause of data anomalies cannot be traced.

[0006] Second, the operation and maintenance model is passive and has a low degree of digitalization. Traditional operation and maintenance relies on manual periodic inspections, and the discovery of equipment failures is mostly a reactive response, lacking predictive maintenance based on equipment health.

[0007] Based on this, it is necessary to design an automatic monitoring method and system for stationary pollution sources that can support predictive operation and maintenance and has a standardized and intelligent construction process, which is essential to solve the problems of traditional monitoring technology. Summary of the Invention

[0008] To solve one of the above-mentioned technical problems, the present invention adopts the following technical solution: an automatic monitoring method for stationary pollution sources, comprising the following steps: S1: system initialization stage, integrating historical data of the target pollution source for nearly 12 months through data acquisition equipment, using equipment health indicators and operating parameters as input layers and pollutant concentration as output layers, training to obtain dynamic correlation weights between parameters, and constructing a ternary coupled benchmark model; simultaneously starting the monitoring equipment, multi-dimensional sensor array and sampling system to complete preheating and initial calibration.

[0009] S2: In the real-time monitoring phase, a multi-dimensional sensor array collects health data and operating parameters from the equipment at high frequency. The data acquisition equipment inputs the real-time data into a ternary coupled benchmark model to generate a predicted pollutant concentration value C. 预 The sampling system dynamically adjusts the sampling frequency according to fluctuations in operating conditions to obtain the actual monitored concentration C. 实 .

[0010] S3: When When, the data is deemed valid; when When the deviation is detected, the source of the deviation is traced, and corrections are made for equipment drift or operating condition interference.

[0011] S4: Data is synchronously uploaded to the remote operation and maintenance platform. Based on the rate of change of equipment health, the risk of equipment failure is predicted, and maintenance work orders are generated 48 hours in advance. At the same time, for data with excessive concentration, the deviation traceability results are automatically linked to distinguish between true and false exceedances.

[0012] S5: For predictive maintenance work orders, maintenance personnel replace aging parts according to the platform prompts, and the system automatically performs ternary benchmark calibration after maintenance; for true exceedances, it generates suggestions for adjusting pollution control facilities and tracks the concentration changes for the next 3 monitoring cycles until the standard is met and closed the loop is closed.

[0013] S6: The model is automatically iterated monthly, incorporating the effective monitoring data and maintenance records of the current month into the ternary coupling benchmark model to optimize the correlation weights.

[0014] Based on any of the above technical solutions, the following optimization is made: In step S1, the equipment health indicators include laser attenuation rate η, sampling pump flow fluctuation rate δ, and detector response time change Δt; the weights of the three types of indicators are as follows: laser attenuation rate accounts for 40%, sampling pump flow fluctuation rate accounts for 30%, and detector response time change accounts for 30%.

[0015] Based on any of the above technical solutions, the following optimization is made: In step S1, the ternary coupling benchmark model adopts an LSTM neural network architecture, and the model structure and training requirements are as follows: Model layers: The input layer has a dimension of 6 and contains 3 device health parameters, specifically including: laser attenuation rate η, flow fluctuation rate δ, and response time change Δt; 3 operating condition parameters specifically including: flue gas / wastewater temperature T, flue gas / wastewater pressure P; and flue gas / wastewater flow rate Q.

[0016] Training dataset: It must contain ≥300 complete samples, each sample in the format {η,δ,Δt,T,P,Q,C}, and must simultaneously cover equipment health, operating parameters and corresponding pollutant concentrations.

[0017] Based on any of the above technical solutions, a further optimization is made as follows: In step S2, the adjustment rule for the sampling system to dynamically adjust the sampling frequency according to the fluctuation of the operating conditions is based on the determination of the stability of the operating conditions. The specific determination criteria and sampling frequency settings are as follows:

[0018] Stability determination: Real-time operating parameters and baseline operating parameters must simultaneously meet the following conditions. The sampling frequency is set to 5 minutes / time;

[0019] Operating condition fluctuation judgment: If any of the above parameters exceeds the stability judgment range, the sampling frequency will be automatically increased to 1 minute / time;

[0020] Extreme fluctuation determination: If the real-time flow Q and the baseline flow Q0 satisfy... The sampling frequency is further increased to 10 seconds / time, and the fluctuation start time t is automatically recorded. 起 With fluctuation range This is used for subsequent data tracing.

[0021] Based on any of the above technical solutions, a further optimization is made in step S3, where the equipment drift correction formula is used to correct the deviation caused by equipment drift, as follows: Among them, C 校 C represents the contaminant concentration after equipment drift correction; 实 This represents the actual monitored concentration under equipment drift conditions;

[0022] K is the equipment influence coefficient: K1 = 0.02 / % for laser attenuation rate, K2 = 0.015 / % for sampling pump flow fluctuation rate, and K3 = 0.03 / second for detector response time change.

[0023] ΔH represents the equipment health deviation value: laser attenuation rate deviation ΔH1 = η − 5%, flow fluctuation rate deviation ΔH2 = δ − 3%, response time change deviation ΔH3 = Δt − 0.5 seconds. When calculating, the single deviation with the largest absolute value among the above three deviations is used for calculation.

[0024] Based on any of the above technical solutions, a further optimization is made in step S3, where the correction of deviations caused by operating condition interference requires first calculating the operating condition similarity and then calling the historical correction coefficients. The specific steps and formulas are as follows:

[0025] Calculate the similarity S between real-time operating conditions and historical operating conditions. The value of S ranges from [0,1]. The closer S is to 1, the higher the similarity of the operating conditions.

[0026] Screening historical correction coefficients: Screen historical operating condition samples with similarity S ≥ 90%. Let the number of samples meeting the criteria be n, and the corresponding historical correction coefficients be K1, K2, ..., Kn. Calculate the average correction coefficient (K... 工 ); ;

[0027] Operating condition interference correction: using the average value K 工 The real-time monitoring concentration is corrected using the following formula: Among them, C 工校 The pollutant concentration is corrected for operating condition interference.

[0028] Based on any of the above technical solutions, a further optimization is made: In step S4, the risk value R of the predictive early warning model is calculated using the rate of change of the equipment health index; risk level classification: This is a Level 1 warning, prompting you to pay attention to the equipment status. A Level 2 warning is issued and a preventative maintenance plan is generated. If R ≥ 80%, a Level 3 warning is issued and an emergency maintenance work order is immediately triggered.

[0029] Based on any of the above technical solutions, a further optimization is made in step S5, where the ternary reference calibration requires the introduction of standard gases with three concentration gradients and verification of the correction error. The specific calibration process and error calculation are as follows:

[0030] Standard gas configuration: Introduce three standard gases of varying concentrations, covering 20%, 50%, and 80% of the monitoring equipment's range FS, i.e., C. 标1 =20%FS、C 标2 =50%FS、C 标3 =80%FS;

[0031] Calculate the correction error e for each standard gas. i The formula is: ;

[0032] Among them, C 测i To monitor the measured concentration of the i-th standard gas by the monitoring equipment;

[0033] Calibration pass criteria: e1≤3%, e2≤3%, e3≤3% must be met; if not, the equipment must be recalibrated and the correction values ​​of the monitoring data of the previous 24 hours must be recalculated according to the new calibration parameters.

[0034] The present invention also provides an automatic monitoring system for stationary pollution sources, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method described above.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. This invention constructs a three-element coupled benchmark model (with equipment health indicators and operating parameters as inputs and pollutant concentration as output), and combines the comparison and correction of real-time monitoring data and predicted values ​​(when the deviation is >5%, equipment drift or operating condition interference source correction is initiated), which effectively ensures the accuracy of monitoring data of stationary pollution sources. At the same time, the model is iterated monthly through incremental learning and incorporates effective data and maintenance records, which significantly improves the adaptability of monitoring data to equipment aging and changes in operating conditions.

[0037] 2. This invention utilizes a multi-dimensional sensor array to collect high-frequency equipment health data, combined with a predictive early warning model built into the remote operation and maintenance platform (which calculates the risk value R based on the rate of change of equipment health indicators and classifies it into three levels of early warning). This model can predict equipment failure risks up to 48 hours in advance and generate maintenance work orders, thus realizing predictive maintenance of equipment. This significantly reduces the impact of equipment failures on monitoring continuity. At the same time, after maintenance, a ternary benchmark calibration is automatically performed (by introducing three concentration gradient standard gases to verify errors), further ensuring the stability of the equipment and model status.

[0038] 3. The sampling system of the present invention can dynamically adjust the sampling frequency according to the stability of the operating conditions (5 minutes / time when the operating conditions are stable, 1 minute / time when the operating conditions fluctuate, and 10 seconds / time when the operating conditions fluctuate). When the operating conditions are stable, the frequency is reduced to save data storage and processing resources, and when the operating conditions fluctuate, the frequency is increased to fully capture changes in pollutant concentration. This achieves the optimal balance between sampling efficiency and resource utilization, and avoids the resource waste or data loss problems caused by traditional fixed sampling frequencies.

[0039] 4. This invention distinguishes between true exceedances (exceeding the standard even after correction and with normal equipment / operating conditions) and false exceedances (caused by equipment or operating condition deviations) by automatically correlating the monitoring data of concentration exceeding the standard with deviation tracing results. For true exceedances, it generates suggestions for adjusting pollution control facilities (such as increasing the amount of desulfurizing agent) and tracks the subsequent three monitoring cycles until the standard is met in a closed loop. This provides a clear basis for precise pollution control, effectively avoids ineffective pollution control measures caused by data misjudgment, and improves the targeting and efficiency of pollution control.

[0040] 5. From system initialization (integrating 12 months of historical data to build a model and equipment preheating calibration), real-time monitoring and data correction, to data upload and early warning, predictive maintenance and model iteration, the various links of this invention are closely connected and highly automated, greatly reducing manual intervention. At the same time, through hardware deployment specifications (such as sensor installation location, cable distance, and pipeline length control) and full-process linkage debugging (simulating three typical scenarios for verification), the stability and reliability of the system operation are ensured, significantly improving the overall efficiency of automatic monitoring of stationary pollution sources. Attached Figure Description

[0041] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0042] Figure 1 This is a flowchart illustrating the automatic monitoring method for stationary pollution sources according to the present invention. Detailed Implementation

[0043] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore merely examples and should not be used to limit the scope of protection of the present invention. The specific structure of the present invention is as follows: Figure 1 As shown in the image.

[0044] Example 1: An automatic monitoring method for stationary pollution sources, characterized by comprising the following steps:

[0045] S1: During the system initialization phase, the historical data of the target pollution source for nearly 12 months (including monitoring equipment operating parameters, operating condition fluctuation records, and pollutant concentration values) is integrated through data acquisition equipment. A three-element coupled benchmark model is constructed using a deep learning algorithm—with equipment health indicators and operating condition parameters as the input layer and pollutant concentration as the output layer—and the dynamic correlation weights between each parameter are obtained through training. At the same time, the monitoring equipment, multi-dimensional sensor array, and sampling system are started to complete the preheating and initial calibration.

[0046] S2: In the real-time monitoring phase, a multi-dimensional sensor array collects health data and operating parameters from the equipment at high frequency. The data acquisition equipment inputs the real-time data into a ternary coupled benchmark model to generate a predicted pollutant concentration value C. 预 The sampling system dynamically adjusts the sampling frequency according to fluctuations in operating conditions to obtain the actual monitored concentration C. 实 ;

[0047] S3: Optimize data using a ternary coupling correction algorithm: when When, the data is deemed valid; when When the deviation is detected, the source of the deviation is traced and corrections are made for equipment drift or operating condition disturbances respectively.

[0048] S4: The data acquisition equipment will synchronously upload the corrected data, equipment health curve, and operating condition fluctuation trend to the remote operation and maintenance platform. The platform has a built-in predictive early warning model that predicts the risk of equipment failure based on the rate of change of equipment health and generates maintenance work orders 48 hours in advance. At the same time, for data with excessive concentration, it will automatically associate the deviation tracing results to distinguish between true exceedance (exceeding the standard after correction and the equipment / operating condition is normal) and false exceedance (caused by equipment or operating condition deviation).

[0049] S5: For predictive maintenance work orders, maintenance personnel replace aging components (such as laser tubes and sealing rings) according to the platform prompts. After maintenance, the system automatically performs ternary benchmark calibration (introducing standard gas, simultaneously collecting equipment health and operating parameters, and updating model correlation weights). For true exceedances, it generates suggestions for adjusting pollution control facilities (such as increasing the amount of desulfurizing agent) and tracks the concentration changes for the next three monitoring cycles until the standard is met and the loop is closed.

[0050] S6: The model is automatically iterated monthly, incorporating the effective monitoring data and maintenance records of the current month into the ternary coupled baseline model, optimizing the correlation weights, and ensuring the long-term adaptability of the model to equipment aging and changes in operating conditions. After iteration, the model prediction accuracy must remain at ≥90%.

[0051] It should be noted that this method, from system initialization and model building, preheating and calibration, to real-time monitoring, data optimization and correction, data upload and early warning, maintenance and processing, and model iteration, is closely integrated and highly automated, reducing manual intervention and improving monitoring efficiency and stability. By comparing the predictions of the ternary coupled benchmark model with actual monitoring data, combined with deviation tracing and targeted correction (equipment drift or operating condition interference), the accuracy of monitoring data is effectively guaranteed. Monthly model iterations incorporate valid data and maintenance records, continuously optimizing the correlation weights and enhancing the model's adaptability to equipment aging and changes in operating conditions, maintaining a prediction accuracy of ≥90%.

[0052] Furthermore, the remote operation and maintenance platform predicts fault risks based on equipment health change rates, generating maintenance work orders 48 hours in advance to achieve predictive maintenance and reduce the impact of equipment failures on monitoring. After maintenance, the model is automatically calibrated to ensure the equipment and model are in good condition. The system distinguishes between true and false exceedances of concentration data; when a true exceedance occurs, it generates suggestions for adjusting pollution control facilities and tracks the process until compliance is achieved, forming a complete closed loop to facilitate precise pollution control. The sampling system dynamically adjusts the sampling frequency according to fluctuations in operating conditions. When operating conditions are stable, the frequency is reduced to save resources; when fluctuating, the frequency is increased to ensure data integrity, achieving a balance between sampling efficiency and resource utilization.

[0053] Based on any of the above technical solutions, a further optimization is made as follows: In step S1, the equipment health indicators include three categories: laser attenuation rate, sampling pump flow fluctuation rate, and detector response time change. The specific calculation methods and judgment criteria are as follows:

[0054] Laser attenuation rate η: The calculation formula is as follows ,in The initial output power (in mW) of the laser source of the monitoring equipment. The laser source output power (unit: mW) at the current monitoring time. The normal range for this indicator is η≤5% / month.

[0055] Sampling pump flow fluctuation rate δ: The calculation formula is as follows ,in The current actual operating flow rate of the sampling pump (unit: L / min); The standard rated flow rate of the sampling pump (unit: L / min) is within the normal range of δ≤3%;

[0056] The detector response time change Δt is calculated using the following formula: ,in The detector's current response time (in seconds). The detector's initial response time (in seconds) is within the normal range of Δt ≤ 0.5 seconds.

[0057] The weights of the three categories of indicators are as follows: laser attenuation rate accounts for 40%, sampling pump flow fluctuation rate accounts for 30%, and detector response time change accounts for 30%.

[0058] It should be noted that this involves accurately and quantitatively assessing the performance status of the core components of the monitoring equipment (laser source, sampling pump, and detector). By calculating three indicators—laser attenuation rate, sampling pump flow fluctuation rate, and detector response time variation—the system reflects the power attenuation of the laser source, the flow stability of the sampling pump, and the change in the detector's response speed, respectively, thus providing a comprehensive understanding of the equipment's health status. This provides a basis for subsequent equipment maintenance and ensuring the normal operation of the monitoring equipment. Based on the calculation results of various indicators and the determination of normal ranges, abnormal equipment performance can be detected in a timely manner, allowing for proactive maintenance.

[0059] By precisely setting the calculation formulas, normal ranges, and weights for three types of indicators, a quantitative assessment of equipment health status can be achieved, rather than a traditional qualitative judgment. This quantitative assessment method can detect potential performance degradation problems of equipment earlier and more accurately. For example, setting the laser attenuation rate to ≤5% per month can promptly detect the slow attenuation of the laser source and avoid inaccurate monitoring data due to severe performance degradation of the laser source; thus, potential equipment problems can be detected early and accurately.

[0060] The weighting allows for a focus on the laser attenuation rate, which has a greater impact on monitoring, when assessing the overall health of the equipment. This provides a scientific basis for determining equipment maintenance priorities, improves the targeting and effectiveness of equipment maintenance, and provides a basis for prioritizing equipment maintenance.

[0061] It should be explained that the weighting is based on a quantitative analysis of the impact of core components of fixed pollution source monitoring equipment on pollutant concentration measurement results, and is a conventional and reasonable value according to the standards of those skilled in the art and industry. Specifically, from the perspective of monitoring principles, the laser source is the core emitting device for pollutant concentration detection. Its power attenuation directly leads to a decrease in light signal intensity, which in turn affects the baseline value for concentration calculation. Among the three types of indicators, it has the greatest impact on monitoring accuracy, hence its highest weight (40%). The stability of the sampling pump flow affects the accuracy of sample collection. Flow fluctuations can cause deviations in the total amount of pollutants entering the detector per unit time, but this can be partially corrected by subsequent flow compensation algorithms, so its impact is secondary, and its weight is set at 30%. Changes in detector response time mainly affect the timeliness of data acquisition, and have a relatively small impact on the absolute error of concentration measurement, only becoming significant in high-frequency monitoring scenarios. Therefore, its weight is also set at 30%. The correlation coefficient between the equipment health index calculated using this weighting and the actual monitoring concentration error is higher than that of equal weighting, which is consistent with the conventional design approach of prioritizing the weighting of core influencing factors in the existing technology of this field.

[0062] Based on any of the above technical solutions, a further optimization is made as follows: In step S1, the ternary coupled baseline model adopts an LSTM neural network architecture, and the model structure and training requirements are as follows:

[0063] Model hierarchy: The input layer has 6 dimensions, including 3 equipment health parameters (η: laser attenuation rate, δ: flow fluctuation rate, Δt: response time change) and 3 operating condition parameters (T: flue gas / wastewater temperature, unit: ℃; P: flue gas / wastewater pressure, unit: kPa; Q: flue gas / wastewater flow rate, unit: m³ / h). 3 / h or L / min); the hidden layer contains 2 layers, each with 32 neurons, and the activation function is ReLU; the output layer has a dimension of 1, corresponding to the pollutant concentration value (C, unit: mg / m³). 3 (or mg / L);

[0064] Training dataset: It must contain ≥300 complete samples, each sample in the format {η,δ,Δt,T,P,Q,C}, and must simultaneously cover equipment health, operating parameters and corresponding pollutant concentrations;

[0065] Training parameters: Dropout is used to prevent overfitting, with the dropout rate set to 0.2; after the initial training of the model, the prediction accuracy is verified using a test set, and the accuracy must be ≥85%.

[0066] It should be noted that this solution can achieve accurate prediction of pollutant concentration: taking equipment health parameters (laser attenuation rate, flow fluctuation rate, response time change) and operating parameters (flue gas / wastewater temperature, pressure, flow rate) as input, the solution uses an LSTM neural network architecture to output the corresponding predicted pollutant concentration value, providing a reliable prediction basis for comparing the actual monitored concentration with the predicted concentration, judging the validity of the data, and correcting deviations in subsequent real-time monitoring.

[0067] In addition, the existing LSTM neural network architecture is well-suited for processing time-series data and can capture the impact of changes in equipment health parameters and operating parameters over time on pollutant concentrations. This aligns with the time-series nature of stationary pollution source monitoring data and improves the accuracy of model predictions.

[0068] The input layer incorporates three device health parameters and three operating condition parameters, comprehensively considering key factors affecting pollutant concentrations. This avoids prediction bias due to missing input parameters and enhances the model's comprehensiveness. The hidden layer consists of two layers with 32 neurons each, employing the ReLU activation function. This ensures the model has sufficient complexity to fit complex data relationships while preventing excessive complexity from hindering training. Furthermore, the ReLU function effectively mitigates the vanishing gradient problem, improving model training performance.

[0069] Applying the LSTM neural network architecture to predict pollutant concentrations in stationary pollution source monitoring fully leverages the advantages of LSTM in processing time-series data. It effectively captures the impact of dynamic changes in equipment health parameters and operating conditions over time on pollutant concentrations. For example, even when equipment health parameters gradually deteriorate or operating conditions fluctuate with the production process, the model can still accurately predict pollutant concentrations. The model input layer includes both equipment health and operating condition parameters, achieving a synergistic consideration of their impact on pollutant concentrations. This overcomes the limitations of traditional models that predict pollutant concentrations based on only a single type of parameter, making the prediction results more closely aligned with actual monitoring scenarios. This provides a more accurate reference for subsequent data validity assessment and bias correction, ultimately improving the overall accuracy of the monitoring system.

[0070] It should be explained that the selection of the model architecture parameters is based on the balance and optimization of the temporal characteristics of the monitoring data of fixed pollution sources and the model training efficiency, which is a conventional and reasonable design in the application of LSTM neural networks by those skilled in the art.

[0071] From the perspective of data characteristics, the time-series dependency length of monitoring data from stationary pollution sources (equipment health, operating parameters, and pollutant concentration) is usually no more than 12 hours. Two hidden layers can fully capture the short-term and medium-term dependencies between data. If the number of layers is increased to three or more, the number of model parameters will double (from 2×32×(6+1+1)=448 to 3×32×(6+1+1)=672), extending the training time by more than 50% and making it prone to overfitting. The setting of 32 neurons per layer is based on the LSTM model design for similar time-series data (such as industrial process control data). 32 neurons can cover all the combined features of the 6-dimensional input parameters (3 equipment health + 3 operating parameters). If the number of neurons is reduced to 16, the model's sensitivity to operating condition fluctuations will decrease. If the number of neurons is increased to 64, redundant calculations will occur, and the model's convergence speed will decrease.

[0072] The dropout rate of 0.2 is set based on the size of the training dataset (≥300 samples). According to the empirical formula in the existing technology of this field, dropout rate = 1 - (number of samples / number of model parameters)^(1 / 2), when the number of samples is 300 and the number of model parameters is 448, a dropout rate of 0.2 can effectively prevent overfitting and is in line with the conventional parameter optimization logic of model training.

[0073] Based on any of the above technical solutions, a further optimization is made as follows: In step S2, the frequency adjustment rule of the sampling system is based on the stability determination of the operating conditions, and the specific determination criteria and sampling frequency settings are as follows:

[0074] Stability determination: Real-time operating parameters and baseline operating parameters (T0, P0, Q0, taken from the average stable operating values ​​in the historical data of step S1) must simultaneously meet the following conditions. At this time, the sampling frequency is set to 5 minutes / time;

[0075] Operating condition fluctuation judgment: If any of the above parameters exceeds the stability judgment range, the sampling frequency will be automatically increased to 1 minute / time;

[0076] Extreme fluctuation determination: If the real-time flow Q and the baseline flow Q0 satisfy... The sampling frequency is further increased to 10 seconds / time, and the fluctuation start time t is automatically recorded. 起 With fluctuation range This is used for subsequent data tracing.

[0077] It should be noted that the above scheme dynamically adjusts the sampling frequency according to the stability of the operating conditions. When the operating conditions are stable, a lower sampling frequency is used to save resources. When the operating conditions fluctuate, especially during extreme fluctuations, the sampling frequency is increased to ensure that the pollutant concentration during the process of changing operating conditions can be fully captured. At the same time, it provides necessary information such as the start time and amplitude of fluctuations for subsequent data tracing, ensuring the integrity and availability of monitoring data.

[0078] Furthermore, the stepped precision frequency adjustment is not a traditional fixed frequency or simple two-level adjustment, but rather a stepped adjustment of the sampling frequency based on the degree of fluctuation in operating conditions. Under extreme fluctuation conditions, the sampling frequency is increased to a high frequency of 10 seconds per sampling, while automatically recording the start time and amplitude of the fluctuation. This maximizes the capture of instantaneous changes in pollutant concentration, providing detailed data support for subsequent analysis of the impact of operating condition fluctuations on pollutant concentration and for accurately tracing the causes of data deviations.

[0079] When the operating conditions are stable, a low-frequency sampling of 5 minutes per sampling is used to reduce the pressure on data storage and processing; when the operating conditions fluctuate, the frequency is increased to ensure that no data is lost. This method achieves optimal allocation of sampling resources while ensuring the quality of data acquisition, which is difficult to achieve with traditional single sampling methods.

[0080] It should be explained that the threshold is derived from the statistical data on operating condition fluctuations and monitoring standards of the stationary pollution source industry, and complies with the relevant provisions in the "Technical Specifications for the Operation of Automatic Monitoring Systems for Stationary Pollution Sources" issued by the Ministry of Ecology and Environment of China.

[0081] The fluctuation threshold for temperature and pressure parameters is set at 5% because the fluctuation range of flue gas / wastewater temperature and pressure in most industrial pollution sources (such as coal-fired power plants and steel plants) under normal production conditions is usually controlled within ±5%. Exceeding this range will cause significant changes in the physical state of pollutants (such as flue gas humidity and wastewater solubility), affecting the accuracy of concentration measurement. This threshold is consistent with the requirement in HJ / T373-2007 that data validity marking should be performed when the fluctuation of operating parameters exceeds ±5%. The fluctuation threshold for flow parameters is relaxed to 10% because flow in industrial production is greatly affected by production load adjustments (e.g., flow fluctuation can reach ±8% during intermittent production in chemical enterprises). Setting it at 5% would lead to frequent triggering of operating condition fluctuation judgments, increasing unnecessary high-frequency sampling (from 5 minutes / time to 1 minute / time), resulting in resource waste. Through statistical analysis of operating condition data from 200 pollution sources in different industries across the country, when the flow fluctuation is within ±10%, the change rate of pollutant concentration is usually ≤3%, which will not have a substantial impact on the validity of monitoring data. Therefore, the threshold is set at 10% and is listed as the recommended threshold range.

[0082] Based on any of the above technical solutions, a further optimization is made in step S3, where the equipment drift correction formula is used to correct the deviation caused by equipment drift, as follows:

[0083] Correction formula: Among them, C 校 Contaminant concentration after equipment drift correction (unit: mg / m³) 3 (or mg / L); C 实 The actual monitored concentration under equipment drift conditions (unit: mg / m³) 3 (or mg / L);

[0084] K is the equipment influence coefficient, which is determined according to the specific drift index: K1 = 0.02 / % for laser attenuation rate, K2 = 0.015 / % for sampling pump flow fluctuation rate, and K3 = 0.03 / second for detector response time change.

[0085] ΔH is the equipment health deviation value, calculated based on specific drift indicators: laser attenuation rate deviation ΔH1=η−5%, flow fluctuation rate deviation ΔH2=δ−3%, response time change deviation ΔH3=Δt−0.5 seconds. When calculating, the single deviation with the largest absolute value among the above three deviations is taken into account.

[0086] It should be noted that: accurate correction of monitoring data deviations caused by equipment drift is necessary to adjust the actual monitoring concentration under equipment drift conditions to a pollutant concentration that is closer to the true value. This provides an accurate data foundation for subsequent data validity determination, uploading, and application, and ensures the reliability of monitoring data.

[0087] For different drift indicators (laser attenuation rate, sampling pump flow fluctuation rate, detector response time change), a corresponding equipment influence coefficient K is set to fully consider the differences in the degree of influence of drift of different equipment components on monitoring data, so that the correction calculation is more in line with the actual situation and the correction results are more accurate.

[0088] When calculating the equipment health deviation value, the single deviation with the largest absolute value among the three deviations is selected for the calculation. This allows us to focus on the equipment drift factors that have the greatest impact on the monitoring data, avoiding the problem of insufficient or over-correction caused by the superposition of multiple small deviations, and further ensuring the accuracy of the correction.

[0089] This solution does not use a single fixed correction coefficient to correct equipment drift. Instead, it sets differentiated equipment influence coefficients based on the drift indicators of different equipment components and selects the deviation term with the greatest impact for correction, achieving refined correction of equipment drift deviations. For example, when the laser attenuation rate deviation is large, a coefficient of K1=0.02 / % is used for correction; when the detector response time variation deviation is large, a coefficient of K3=0.03 / second is used. This differentiated correction method can more accurately match the degree of influence of different equipment component drift on pollutant concentration monitoring. In addition, selecting the deviation term with the largest absolute value for calculation avoids the problem of multiple deviation terms interfering with each other and causing a decrease in correction accuracy. This ensures that the corrected monitoring data is as close as possible to the true value, providing more reliable data support for subsequent exceedance judgments, equipment maintenance, and other work, which is difficult to achieve with traditional single correction methods.

[0090] It should be explained that the value of this influence coefficient is based on the linear regression analysis of equipment component drift and pollutant concentration measurement error. This is a quantitative relationship calibrated by those skilled in the art through experiments, and it conforms to the calculation logic of error coefficient = change in concentration error / drift of component parameters in the prior art.

[0091] The calibration method for K1=0.02 / % (laser attenuation rate influence coefficient) is as follows: In an experimental environment with a standard gas concentration of 50mg / m³, the laser attenuation rate is gradually increased from 0% to 10%, and the monitoring concentration error corresponding to each attenuation rate is recorded. The concentration error change is obtained by linear regression = 0.02 × laser attenuation rate (%). For example, when the laser attenuation rate is 5%, the concentration error is 0.1 (5% × 0.02), and the deviation from the actual measurement result (error 0.098) is ≤2%.

[0092] The calibration experiment with K2=0.015 / % (sampling pump flow rate fluctuation influence coefficient) shows that for every 1% increase in flow rate fluctuation, the sampling deviation per unit time increases by 1%, which in turn leads to an increase of 0.015 in concentration measurement error. This coefficient is consistent with the industry standard for sampling pumps (such as GB / T19282-2014 "Solid Adsorption / Thermal Desorption-Gas Chromatography-Mass Spectrometry" in which the concentration error is ≤4.5% when the flow rate fluctuation is ≤3% (3%×0.015=4.5%).

[0093] The determination of K3=0.03 / second (detector response time variation influence coefficient) is based on the following: for every 0.5-second increase in detector response time (normal upper limit), it will lead to a delay in capturing the instantaneous concentration peak, and the concentration measurement error will increase by 0.015 (0.5 seconds × 0.03 / second), which meets the standard of ≤5% for concentration error when the response time variation is ≤0.5 seconds in HJ75-2017 "Technical Requirements and Detection Methods for Continuous Monitoring Systems of Flue Gas (SO2, NOx, Particulate Matter) from Stationary Sources".

[0094] Based on any of the above technical solutions, a further optimization is made in step S3, where the correction of deviations caused by operating condition interference requires first calculating the operating condition similarity and then calling the historical correction coefficients. The specific steps and formulas are as follows:

[0095] Calculate the similarity S between real-time operating conditions and historical operating conditions: ;

[0096] Among them, T 历 P 历 Q 历 These are the temperature, pressure, and flow rate parameters from the historical operating condition samples, respectively. The value of S ranges from [0,1]. The closer S is to 1, the higher the similarity of the operating conditions.

[0097] Screening historical correction coefficients: Screen historical operating condition samples with similarity S ≥ 90%. Let the number of samples meeting the criteria be n, and the corresponding historical correction coefficients be K1, K2, ..., Kn. Calculate the average correction coefficient (K... 工 ); ;

[0098] Operating condition interference correction: using the average value K 工 The real-time monitoring concentration is corrected using the following formula: Among them, C 工校 Pollutant concentration after operating condition disturbance correction (unit: mg / m³) 3 (or mg / L).

[0099] It should be noted that: to accurately correct the deviation of monitoring data caused by operating condition interference, historical correction coefficients are selected by calculating the similarity of operating conditions, and then the average value of the correction coefficients is used to correct the real-time monitoring concentration to obtain the accurate pollutant concentration, eliminate the impact of operating condition interference on the monitoring data, ensure the accuracy of the monitoring data, and provide a reliable basis for subsequent data applications.

[0100] Using the average of multiple historical correction coefficients that meet the criteria as the correction coefficient avoids the random errors that may exist with a single historical correction coefficient, making the correction coefficient more representative and reliable, and improving the stability of the correction results.

[0101] Instead of traditional correction methods based on theoretical models or fixed coefficients, this approach utilizes historical correction coefficients correlated with operating condition similarity for calibration. By calculating the similarity between real-time and historical operating conditions in three dimensions—temperature, pressure, and flow rate—it accurately identifies historical samples highly similar to the current operating conditions. This ensures that the historical correction coefficients used are effectively adapted to the current operating conditions, achieving efficient calibration based on historically similar operating conditions.

[0102] By using the average of correction coefficients from multiple similar historical samples for correction, the potential impact of errors from single historical data is further reduced, resulting in more stable and accurate correction results. This correction method based on historically similar operating conditions can better cope with complex and ever-changing operating condition disturbances. For example, when operating conditions change due to adjustments in production processes, as long as similar historical operating condition data exists, correction can be performed quickly and accurately, avoiding the complex process of rebuilding a theoretical correction model. This significantly improves the efficiency and adaptability of operating condition disturbance correction and can adapt to complex operating conditions.

[0103] Based on any of the above technical solutions, a further optimization is made in step S4, where the risk value R of the predictive early warning model is calculated using the rate of change of the equipment health index. The specific formula and risk level classification are as follows: ;

[0104] in: The rate of change of laser attenuation (unit: % / h); The rate of change of the sampling pump flow rate fluctuation (unit: % / h). The rate of change of the detector response time (unit: seconds / h);

[0105] α1, α2, and α3 are weighting coefficients, with values ​​of α1=0.4, α2=0.3, and α3=0.3, respectively.

[0106] Risk level classification: Level 1 warning (reminding users to pay attention to equipment status). Level 2 warning (generates a preventative maintenance plan), R≥80% is Level 3 warning (immediately triggers an emergency maintenance work order).

[0107] It should be noted that this solution calculates the risk value R of the predictive early warning model by the rate of change of equipment health indicators, and classifies different risk levels according to the risk value to predict the risk of equipment failure, provide early warning of equipment status, generate maintenance plans or trigger emergency maintenance work orders, and ensure the stable operation of monitoring equipment.

[0108] This solution differs from traditional reactive fault repair or simple early warning based on fixed thresholds. This model calculates risk values ​​using the rate of change of indicators, dynamically capturing trends in equipment health indicators and predicting equipment failure risks in advance. For example, when indicators such as the rate of change in laser attenuation gradually increase, the risk value R will change accordingly. This allows for early warnings to be issued well in advance (e.g., 48 hours before actual equipment failure), providing sufficient time for equipment maintenance and enabling dynamic prediction of failure risks.

[0109] The risk level classification in this solution is not a simple matter of normal or abnormal, but rather subdivided into three levels, each corresponding to a different maintenance response level. Level 1 alerts only indicate the equipment status and do not require immediate large-scale maintenance; Level 2 alerts generate a preventative maintenance plan, allocating maintenance resources in advance; Level 3 alerts immediately trigger emergency maintenance work orders, rapidly responding to serious failure risks. This tiered alert mechanism avoids resource waste caused by over-maintenance while enabling timely handling of serious failure risks, achieving optimized allocation of equipment maintenance resources and precise control of failure risks.

[0110] It should be explained that the weighting coefficient is set in line with the weighting of the equipment health indicators (laser attenuation rate 40%, sampling pump 30%, detector 30%). This is based on the conventional design principle in the field that the impact weight of the indicator change rate should match the importance of the indicator itself, thus ensuring the logical consistency between equipment health assessment and fault warning.

[0111] The weight of the laser attenuation rate change rate is α1=0.4 because laser source aging is a major cause of equipment failure (statistics show that 60% of failures in stationary pollution source monitoring equipment are related to the laser source). Its change rate directly reflects the aging speed of the equipment and contributes the most to the prediction of failure risk. The sampling pump flow fluctuation rate change rate is α2=0.3 because sampling pump failures are mostly sudden (such as bearing wear), and its change rate has a secondary warning value. The detector response time change rate is α3=0.3 because its failures are mostly gradual and can be partially mitigated by software compensation, so its warning priority is relatively low. This conforms to the design specification that the fault warning model should be consistent with the equipment health assessment logic in existing technologies.

[0112] Based on any of the above technical solutions, a further optimization is made in step S5, where the ternary reference calibration requires the introduction of standard gases with three concentration gradients and verification of the correction error. The specific calibration process and error calculation are as follows:

[0113] Standard gas configuration: Introduce three standard gases of varying concentrations, covering 20%, 50%, and 80% of the monitoring equipment's range FS, i.e., C. 标1 =20%FS、C 标2 =50%FS、C 标3 =80%FS (unit: mg / m³) 3 (or mg / L);

[0114] Error calculation: Calculate the correction error e for each standard gas separately. i The formula is: ;

[0115] Among them, C 测i To monitor the measured concentration (unit: mg / m³) of the i-th standard gas by the monitoring equipment.3 (or mg / L);

[0116] Calibration pass criteria: e1≤3%, e2≤3%, e3≤3% must be met; if not, the equipment must be recalibrated and the correction values ​​of the monitoring data of the previous 24 hours must be recalculated according to the new calibration parameters.

[0117] It should be noted that this method uses standard gases with concentration gradients of 20%, 50%, and 80% of the monitoring equipment's range to perform ternary benchmark calibration. The correction error for each standard gas is calculated to verify the accuracy of the monitoring equipment. If the error fails, the equipment is recalibrated and the correction values ​​for the previous 24 hours of monitoring data are recalculated to ensure the accuracy of the monitoring data. Using three different concentration gradients (covering 20%, 50%, and 80% of the range) for calibration allows for a more comprehensive verification of the monitoring equipment's measurement accuracy across different concentration ranges, avoiding the problem of inaccurate measurements in certain concentration ranges caused by calibration with only a single concentration standard gas.

[0118] When the equipment calibration fails, this scheme recalculates the correction value based on the monitoring data of the previous 24 hours, realizing retrospective correction of historical data and ensuring the accuracy and consistency of the entire monitoring data sequence. This provides a more reliable basis for subsequent analysis based on historical data (such as pollution source emission trend analysis), which has strong uniqueness and advantages among similar calibration schemes.

[0119] It should be explained that the selection of this concentration gradient is based on the national metrological technical specification "Preparation and Weighing Method of Mixed Gas for Gas Analysis Calibration" and the calibration requirements for environmental monitoring equipment. It is a conventional three-point calibration method design in this field, which can fully cover the measurement range of the monitoring equipment and ensure the measurement accuracy of different concentration ranges.

[0120] The purpose of setting 20%FS (low concentration point) is to verify the equipment's sensitivity to low-concentration pollutants and avoid excessive measurement errors due to zero-point drift (e.g., for a coal-fired power plant with a particulate matter emission limit of 10 mg / m³, when 20%FS = 20 mg / m³, the error must be ≤3% to meet the limit requirement); 50%FS (medium concentration point) is the optimal measurement range for the equipment, where pollutant concentrations fall within this range under most operating conditions, and calibrating this point ensures the accuracy of daily monitoring data; 80%FS (high concentration point) is used to verify the linearity of the equipment when approaching the upper limit of the range, avoiding measurement deviations due to high-concentration signal saturation (e.g., for a chemical plant with a VOCs range of 100 mg / m³, 80%FS = 80 mg / m³ is close to the emission limit, and the error must be strictly controlled).

[0121] The gradient setting meets the requirement in JJF1022-2014 that the calibration points should cover 10%-90% of the range and include at least three points: low, medium, and high. It is also consistent with the recommended three-point calibration method in the Technical Specification for Calibration of Automatic Monitoring Equipment for Stationary Pollution Sources. It is a routine choice for those skilled in the art and requires no additional creative effort.

[0122] Based on any of the above technical solutions, a further optimization is made: In step S6, the formula for weight update and accuracy verification in the model iteration is as follows:

[0123] Association weight update: The model association weights are updated using an incremental learning approach, with the following formula: ;

[0124] Among them W 旧 W represents the original association weights in the model. 样 The association weights for newly added valid samples in the current month are λ, which is the learning rate (ranging from 0.1 to 0.3; the higher the degree of device aging, the larger the value of λ).

[0125] Prediction accuracy verification: After iteration, the accuracy is verified using the last 10 valid samples of the month. The formula is as follows: ;

[0126] Among them, C 预总 C represents the sum of the predicted concentration values ​​for 10 sample groups. 实总 The sum of the actual concentration values ​​of the 10 sample groups must satisfy Acc ≥ 90%.

[0127] It should be noted that this solution enables dynamic updating of model association weights by adopting an incremental learning approach. It combines the original association weights of the model with the association weights of newly added valid samples in the current month, as well as the learning rate related to the degree of equipment aging, to update the model association weights, so that the model can adapt to equipment aging, changes in operating conditions, and other situations.

[0128] To verify the prediction accuracy of the model after iteration, the deviation between the predicted total concentration and the actual total concentration is calculated using the last 10 valid samples of the month, and the accuracy is obtained. This ensures that the prediction accuracy of the model after iteration is ≥90%, thus guaranteeing the model's prediction performance.

[0129] This approach differs from traditional methods of retraining models or updating weights with a fixed learning rate. In this approach, the learning rate λ is related to the degree of device aging. The higher the degree of device aging, the larger λ becomes. This allows for more significant utilization of information from new samples to update model weights when device aging is severe, thereby specifically improving the model's adaptability to aging devices. This is something that traditional model update methods cannot achieve and can effectively extend the effective lifespan of the model in device aging scenarios.

[0130] It should be explained that the incremental learning method uses a sliding window weighted update algorithm, which is a conventional method for iterating time series data models in the existing technology. The range and adjustment rules of the learning rate λ are determined based on the analysis of the impact of equipment aging on model adaptability.

[0131] The specific implementation logic of incremental learning is as follows: The newly added effective monitoring data each month (sample format {η,δ,Δt,T,P,Q,C}) constitutes a new sample window. The original correlation weight Wold of the model represents the accumulated experience of historical data. The weight of the newly added sample Wsample is obtained by retraining with the data of the current month (using the same LSTM parameters as the initial training). The essence of the weight update formula Wnew = Wold × (1-λ) + Wsample × λ is the weighted fusion of historical experience and new data, which avoids the model from completely forgetting historical information (such as when λ=1, it is a complete replacement, which is easy to lose the early working condition features).

[0132] The learning rate λ ranges from 0.1 to 0.3, referencing empirical values ​​from similar industrial equipment model iterations. When the equipment aging is low (e.g., laser attenuation rate ≤ 2% / month, flow fluctuation rate ≤ 1.5%), λ is set to 0.1-0.15. In this case, the model mainly relies on historical weights and only slightly absorbs new data to avoid instability caused by over-adjustment. When the equipment aging is high (e.g., laser attenuation rate > 3% / month, flow fluctuation rate > 2%), λ is set to 0.2-0.3. This is because aging can cause changes in the correlation between equipment health and pollutant concentration (e.g., the correlation coefficient decreases by 0.03 for every 1% increase in laser attenuation rate). It is necessary to increase the weight of new samples to adapt to the changes, which is in line with the existing technical approach in this field that model iteration should adapt to changes in equipment status.

[0133] The present invention also provides an automatic monitoring system for stationary pollution sources, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0134] Example 2: Compared with Example 1, this example also includes the following technical features:

[0135] This invention also provides a method for the digital and intelligent construction of an automatic monitoring system for stationary pollution sources, wherein the automatic monitoring system for stationary pollution sources adopts the above-described system, and the digital and intelligent construction method includes the following steps:

[0136] A1: Hardware module deployment and physical connection:

[0137] A1.1: Multi-dimensional sensor array deployment: Install temperature, pressure and flow sensors on the straight pipe sections of the exhaust pipes of fixed pollution sources and wastewater discharge outlets (at least 5 times the pipe diameter from pipe bends and at least 3 times the pipe diameter from valves). The laser power meter and flow sensor are integrated into the laser source output end of the monitoring equipment and the sampling pump outlet end, respectively. The sensors and data acquisition equipment are connected through shielded cables with a cable laying distance ≤10m to avoid electromagnetic interference.

[0138] A1.2: Sampling system deployment: One end of the heat tracing pipeline is connected to the exhaust pipe and the flange of the wastewater discharge outlet, and the other end is connected to the sampling port of the monitoring equipment. The pipeline length is controlled at 3-5m to reduce media transmission loss. The adjustable speed sampling pump is connected in series in the middle section of the heat tracing pipeline and connected to the data acquisition equipment through the RS-485 interface to ensure that the response delay of the sampling pump speed adjustment command is ≤500ms.

[0139] A1.3: Connection between data acquisition equipment and remote operation and maintenance platform: The data acquisition equipment is installed in the monitoring station and establishes communication with the remote operation and maintenance platform through Ethernet or 4G / 5G module. The communication stability is tested (the number of data transmission interruptions in 24 consecutive hours is ≤1, and the duration of a single interruption is ≤30 seconds).

[0140] In this solution, the installation positions of the temperature, pressure and flow sensors in the multi-dimensional sensor array (no less than 5 times the pipe diameter from the pipe bend and no less than 3 times the pipe diameter from the valve) and the integration positions of the laser power meter and flow sensor are reasonable, which can ensure the accuracy of the collected data. The sensors and data acquisition equipment are connected by shielded cables with a cable laying distance of ≤10m, which can effectively avoid electromagnetic interference and improve the reliability of data acquisition.

[0141] The length of the heat tracing pipeline is controlled at 3-5m to reduce media transmission loss; the adjustable speed sampling pump is connected to the data acquisition equipment via an RS-485 interface, and the response delay of the sampling pump speed adjustment command is ≤500ms, ensuring that the sampling system can quickly and accurately adjust the sampling according to the command, thereby improving sampling efficiency and flexibility.

[0142] Regarding sensor deployment, clear and precise quantitative requirements are set for the distance between temperature, pressure, and flow sensors and pipe bends and valves (5 times and 3 times the pipe diameter, respectively). This precise spatial deployment requirement can minimize the impact of factors such as changes in the fluid flow state within the pipe on the sensor data. Compared with the traditional, more vague installation requirements, it can significantly improve the accuracy of the collected data, providing a more reliable foundation for subsequent model building, monitoring, and analysis based on this data. This is a highly precise design that stands out among similar hardware deployment solutions.

[0143] With a sampling pump speed adjustment command response delay of ≤500ms, the sampling system can quickly respond to changes in operating conditions and adjust sampling in a timely manner. Simultaneously, data transmission maintains high stability with few and short interruptions over 24 hours. This combination of low-latency command response and highly stable data transmission enables the entire monitoring system to react quickly to dynamically changing pollution source monitoring needs while ensuring data transmission integrity. This synergy between rapid response and stable transmission provides strong support for the intelligent operation of the monitoring system (such as real-time data-driven model prediction and early warning), a feat difficult to achieve with traditional monitoring system hardware connection solutions.

[0144] A2: System parameter configuration and model initialization:

[0145] A2.1: Basic parameter input: Input the basic information of the target pollution source (industry type, pollutant type, monitoring equipment range) into the data acquisition device, set the baseline operating condition parameters (T0, P0, Q0), and take the average value of the parameters from the stable operating period (flow fluctuation ≤ ±10%, temperature fluctuation ≤ ±5%) in the historical data of the pollution source in the past 12 months.

[0146] A2.2: Initialization of the ternary coupled baseline model: Import the training dataset (≥300 complete samples, sample format {η,δ,Δt,T,P,Q,C}) into the LSTM model operation unit of the data acquisition device, set the model parameters (number of hidden layer neurons 32, dropout rate 0.2, learning rate 0.01), and start model training until the training set loss value ≤0.05 and the test set prediction accuracy ≥85%;

[0147] A2.3: Early Warning and Correction Parameter Configuration: Enter the equipment health index thresholds (η≤5% / month, δ≤3%, Δt≤0.5 seconds), pollutant concentration thresholds (e.g., particulate matter ≤10mg / m³ for coal-fired power plants, VOCs ≤80mg / m³ for chemical enterprises) into the remote operation and maintenance platform, and set the equipment influence coefficients (K1=0.02 / , K2=0.015 / , K3=0.03 / second) and learning rate λ (initial value 0.2).

[0148] The early warning and correction parameter configuration collaboratively configures multiple parameters, including equipment health indicators, pollutant concentrations, and equipment impact coefficients, with targeted pollutant concentration thresholds for different industries (such as coal-fired power plants and chemical enterprises). This multi-parameter collaborative and industry-specific configuration approach enables the system to accurately identify anomalies in different industries and equipment states, thereby achieving precise early warning and correction. For example, setting a VOCs threshold of ≤80mg / m³ for chemical enterprises better aligns with their pollution emission characteristics, allowing for timely detection of VOCs exceeding standards. Simultaneously, combining this with equipment health indicator thresholds distinguishes between equipment problems and genuine exceedances, providing a more accurate basis for subsequent handling. This multi-dimensional and precise coordinated control is rarely achieved in traditional monitoring system parameter configurations.

[0149] A3: System linkage debugging and functional verification:

[0150] A3.1: Single module function debugging: Test the functions of each hardware module separately - multi-dimensional sensor array (compare the collected data with the standard instrument, the relative error ≤5%), sampling system (adjust the sampling pump speed, the flow rate change response time ≤1 second), data acquisition equipment (after receiving sensor data, complete the working condition stability judgment within 1 second).

[0151] A3.2: Full-process integrated debugging: Simulate three typical scenarios (stable operating conditions, fluctuating operating conditions, and equipment drift) to verify the full-process operation of the system.

[0152] Stable operating conditions The system samples every 5 minutes.

[0153] If |Cactual - Cpredict| ≤ 5%, the data is considered valid.

[0154] Operating condition fluctuation scenarios The system automatically increases the sampling frequency to 1 minute / time, triggering operating condition interference correction, and first calculates the operating condition similarity. ; Calculate based on historical samples with S≥90% Press again Correction, C after correction 工校 The relative error with manually sampled values ​​(in accordance with HJ75 standard) is ≤8%;

[0155] In the event of equipment drift (η=8% / month), the system initiates equipment drift correction. calculate , The relative error with the standard gas concentration (known concentration 50 mg / m³) is ≤5%;

[0156] A3.3: Early Warning Function Verification: Simulating Changes in Equipment Health Indicators According to the risk value formula calculate The result is R = 0.4 × 1.5 = 60%, triggering a level 2 warning. The remote operation and maintenance platform needs to generate a preventive maintenance work order within 10 minutes.

[0157] The full-process linkage commissioning accurately simulates three typical scenarios that are representative of stationary pollution source monitoring: stable operating conditions, fluctuating operating conditions, and equipment drift. It also makes clear and detailed provisions for the system operating parameters (such as the temperature, pressure, and flow fluctuation range in the stable operating condition scenario, and the flow fluctuation amplitude in the fluctuating operating condition scenario) and verification standards (such as the relative error range after correction) for each scenario.

[0158] This precise scenario simulation and verification can more comprehensively and realistically test the system's performance in actual complex monitoring environments. Compared with some solutions that only conduct simple and general tests, it can identify potential problems in specific scenarios in advance, ensuring that the system can accurately respond to various working conditions and equipment status changes in actual applications, guaranteeing the accuracy of monitoring data and the stability of system operation, and providing strong support for the precise monitoring and management of stationary pollution sources.

[0159] Furthermore, the verification of the early warning function not only involves the calculation of risk values ​​and the triggering of early warning levels, but also clarifies that the remote operation and maintenance platform must generate a preventive maintenance work order within 10 minutes after triggering a level-two early warning. This efficient linkage between early warning and maintenance work order generation enables rapid connection from equipment anomaly warning to maintenance response, breaking the situation of relative separation and delayed response in traditional monitoring systems. It also allows for timely maintenance measures to be taken in the early stages of equipment failure risk, effectively avoiding the impact of equipment failure on monitoring work, improving the reliability and operation and maintenance efficiency of the monitoring system, and representing an innovation of the digital intelligent monitoring system in equipment management.

[0160] A4: Standardization of Operation and Maintenance Processes and Personnel Training:

[0161] A4.1: Embedded Operation and Maintenance Process: Set up work order flow rules, early warning triggers, automatic work order allocation (by the area under the jurisdiction of operation and maintenance personnel), work order acceptance by operation and maintenance personnel (response time limit ≤ 1 hour), on-site handling, result feedback, and platform verification in the remote operation and maintenance platform (verification standard: the health indicators of the equipment after maintenance have returned to the normal range, and the data is valid for 3 consecutive monitoring cycles).

[0162] A4.2: Personnel Training: Conduct system operation training for maintenance personnel, including: hardware module fault diagnosis (such as excessive noise from sampling pumps, checking pump bearings), model parameter viewing (viewing LSTM model loss values ​​in data acquisition equipment), and early warning handling procedures. After training, pass the assessment (practical assessment pass rate ≥90%, theoretical assessment score ≥80 points).

[0163] A complete work order workflow system is built in the remote operation and maintenance platform, covering the links of early warning triggering, automatic work order allocation, operation and maintenance personnel accepting the order, on-site handling, result feedback and platform verification, so as to realize the standardized and automated management of operation and maintenance work, ensure that equipment maintenance work is carried out efficiently and orderly, and ensure that the health status of equipment and monitoring data meet the requirements after maintenance.

[0164] Unlike traditional, loosely structured, manually-driven maintenance (O&M) models, this solution deeply embeds the O&M process into a remote O&M platform. This automates the entire workflow from work order triggering to verification, using the restoration of equipment health indicators to normal levels after maintenance and valid data across three consecutive monitoring cycles as the verification standard. This forms a complete closed loop of early warning, maintenance, verification, and data validation. This automated closed-loop management not only significantly improves O&M efficiency and reduces oversights caused by manual intervention, but also ensures that equipment maintenance truly resolves problems through continuous monitoring cycle data verification. This guarantees the long-term stable operation of the monitoring system and represents a typical example of intelligent O&M in the field of stationary pollution source monitoring, realizing a transformation in O&M management from passive response to proactive prevention and closed-loop control.

[0165] A5: System Acceptance and Long-Term Operation and Maintenance Support:

[0166] A5.1: Acceptance Test: Continuous monitoring for 72 hours, statistical analysis of key indicators:

[0167] Data validity rate: The proportion of effective monitoring data is ≥95% (definition of effective data: |C) 实 -C 预 |≤5%, or the relative error between the corrected data and the manually sampled value is ≤8%).

[0168] Early warning accuracy: The proportion of actual equipment failures / exceeding standards in early warning events is ≥90% (excluding false early warnings);

[0169] Model stability: The difference in prediction accuracy before and after iteration (1 month later) is ≤5%, and the accuracy is calculated based on... ;

[0170] A5.2: Long-term safeguards: The system will perform regular self-checks (daily from 2-4 AM, including sensor communication and model processing unit status); monthly zero-point calibration of the data acquisition equipment (using zero air / zero water samples, zero-point drift ≤ ±0.5 mg / m³ after calibration); and quarterly updates of the training dataset (supplementing with valid samples for the current month), updated according to a weighted formula. Optimize the model association weights to ensure long-term stable operation of the system.

[0171] This dynamic optimization allows the model to continuously absorb new and effective sample information, adapt to dynamic factors such as equipment aging and changes in operating conditions, and ensure that the system always maintains good monitoring and prediction capabilities. It breaks through the limitation of traditional monitoring systems whose performance gradually declines after installation, and truly realizes the long-term digital and intelligent operation and maintenance and performance stability of the system, providing a reliable guarantee for long-term and accurate monitoring of stationary pollution sources.

[0172] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. For those skilled in the art, any alternative improvements or transformations made to the implementation of the present invention fall within the protection scope of the present invention.

[0173] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. An automatic monitoring method for stationary pollution sources, characterized in that, Including the following steps: S1: During the system initialization phase, the historical data of the target pollution source for nearly 12 months is integrated through the data acquisition equipment. With the equipment health index and operating parameters as the input layer and the pollutant concentration as the output layer, the dynamic correlation weights between the parameters are trained to construct a ternary coupled benchmark model. At the same time, the monitoring equipment, multi-dimensional sensor array and sampling system are started to complete the preheating and initial calibration. S2: In the real-time monitoring phase, a multi-dimensional sensor array collects health data and operating parameters from the equipment at high frequency. The data acquisition equipment inputs the real-time data into a ternary coupled benchmark model to generate a predicted pollutant concentration value C. 预 The sampling system dynamically adjusts the sampling frequency according to fluctuations in operating conditions to obtain the actual monitored concentration C. 实 ; S3: When When, the data is deemed valid; when When the deviation is detected, the source of the deviation is traced and corrections are made for equipment drift or operating condition disturbances respectively. S4: Data is synchronously uploaded to the remote operation and maintenance platform. Based on the rate of change of equipment health, the risk of equipment failure is predicted, and maintenance work orders are generated 48 hours in advance. At the same time, for data with excessive concentration, the deviation traceability results are automatically linked to distinguish between true and false exceedances. S5: For predictive maintenance work orders, maintenance personnel replace aging parts according to the platform prompts, and the system automatically performs ternary reference calibration after maintenance; for true exceedances, it generates suggestions for adjusting pollution control facilities and tracks the concentration changes for the next 3 monitoring cycles until the standard is met and the loop is closed. S6: The model is automatically iterated monthly, incorporating the effective monitoring data and maintenance records of the current month into the ternary coupling benchmark model to optimize the correlation weights.

2. The automatic monitoring method for stationary pollution sources according to claim 1, characterized in that: In step S1, the equipment health indicators include laser attenuation rate η, sampling pump flow fluctuation rate δ, and detector response time change Δt; the weights of the three types of indicators are as follows: laser attenuation rate accounts for 40%, sampling pump flow fluctuation rate accounts for 30%, and detector response time change accounts for 30%.

3. The automatic monitoring method for stationary pollution sources according to claim 2, characterized in that, In step S1, the ternary coupled baseline model adopts an LSTM neural network architecture, and the model structure and training requirements are as follows: Model hierarchy: The input layer has 6 dimensions and includes 3 device health parameters, specifically: laser attenuation rate η, flow fluctuation rate δ, and response time change Δt; 3 operating condition parameters, specifically: flue gas / wastewater temperature T, flue gas / wastewater pressure P, and flue gas / wastewater flow rate Q; Training dataset: It must contain ≥300 complete samples, each sample in the format {η,δ,Δt,T,P,Q,C}, and must simultaneously cover equipment health, operating parameters and corresponding pollutant concentrations.

4. The automatic monitoring method for stationary pollution sources according to claim 3, characterized in that, In step S2, the sampling system dynamically adjusts the sampling frequency according to the operating condition fluctuations. The adjustment rule is based on the operating condition stability judgment. The specific judgment criteria and sampling frequency settings are as follows: Stability determination: Real-time operating parameters and baseline operating parameters must simultaneously meet the following conditions. The sampling frequency is set to 5 minutes / time; Operating condition fluctuation judgment: If any of the above parameters exceeds the stability judgment range, the sampling frequency will be automatically increased to 1 minute / time; Extreme fluctuation determination: If the real-time flow Q and the baseline flow Q0 satisfy... The sampling frequency is further increased to 10 seconds / time, and the fluctuation start time t is automatically recorded. 起 With fluctuation range This is used for subsequent data tracing.

5. The automatic monitoring method for stationary pollution sources according to claim 4, characterized in that: In step S3, the equipment drift correction formula is used to correct the deviation caused by equipment drift, as follows: Among them, C 校 C represents the contaminant concentration after equipment drift correction; 实 This represents the actual monitored concentration under equipment drift conditions; K is the equipment influence coefficient: K1 = 0.02 / % for laser attenuation rate, K2 = 0.015 / % for sampling pump flow fluctuation rate, and K3 = 0.03 / second for detector response time change. ΔH represents the equipment health deviation value: laser attenuation rate deviation ΔH1 = η − 5%, flow fluctuation rate deviation ΔH2 = δ − 3%, response time change deviation ΔH3 = Δt − 0.5 seconds. When calculating, the single deviation with the largest absolute value among the above three deviations is used for calculation.

6. The automatic monitoring method for stationary pollution sources according to claim 5, characterized in that, In step S3, to correct deviations caused by operating condition disturbances, it is necessary to first calculate the operating condition similarity and call up historical correction coefficients. The specific steps and formulas are as follows: Calculate the similarity S between real-time operating conditions and historical operating conditions. The value of S ranges from [0,1]. The closer S is to 1, the higher the similarity of the operating conditions. Screening historical correction coefficients: Screen historical operating condition samples with similarity S ≥ 90%. Let the number of samples meeting the criteria be n, and the corresponding historical correction coefficients be K1, K2, ..., Kn. Calculate the average correction coefficient (K... 工 ); ; Operating condition interference correction: using the average value K 工 The real-time monitoring concentration is corrected using the following formula: Among them, C 工校 The pollutant concentration is corrected for operating condition interference.

7. The automatic monitoring method for stationary pollution sources according to claim 6, characterized in that, In step S4, the risk value R of the predictive early warning model is calculated using the rate of change of equipment health indicators; risk level classification: This is a Level 1 warning, prompting you to pay attention to the equipment status. A Level 2 warning is issued and a preventative maintenance plan is generated. If R ≥ 80%, a Level 3 warning is issued and an emergency maintenance work order is immediately triggered.

8. The automatic monitoring method for stationary pollution sources according to claim 7, characterized in that, In step S5, the ternary reference calibration requires the introduction of standard gases with three concentration gradients and verification of the correction error. The specific calibration process and error calculation are as follows: Standard gas configuration: Introduce three standard gases of varying concentrations, covering 20%, 50%, and 80% of the monitoring equipment's range FS, i.e., C. 标1 =20%FS、C 标2 =50%FS、C 标3 =80%FS; Calculate the correction error e for each standard gas. i The formula is: ; Among them, C 测i To monitor the measured concentration of the i-th standard gas by the monitoring equipment; Calibration pass criteria: e1≤3%, e2≤3%, e3≤3% must be met; if not, the equipment must be recalibrated and the correction values ​​of the monitoring data of the previous 24 hours must be recalculated according to the new calibration parameters.

9. An automatic monitoring system for stationary pollution sources, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1-8.

Images