Intelligent monitoring and early warning method and system for denitration parameters

Abstract

The invention relates to a denitration parameter intelligent monitoring and early warning method and system. The intelligent monitoring and early warning method for the denitration parameters comprises the following steps: firstly, collecting historical NOx data at an outlet of an absorption tower, and segmenting the historical NOx data into an over-range data segment and a non-over-range data segment according to a fixed interval; preprocessing and enhancing the two types of data to form a data set; training an auto-encoder sub-model by using a non-outrange data segment, training an isolated forest sub-model by using an outrange data segment, and constructing a hybrid prediction model; nOx data are collected in real time and input into the model, and the outrange probability and the correction value at the next moment are output; and finally, when the outrange probability is greater than a preset value, triggering an alarm, and outputting a correction value for adjustment reference. Compared with the prior art, the problem that environmental protection parameters exceed the standard due to the range upper limit during denitration parameter monitoring is solved, and the operation stability of a unit is improved.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a smart monitoring and early warning method and system for denitrification parameters. Background Technology

[0002] The main problem with the existing denitrification system operation is that the automatic ammonia injection control value uses the hourly average of the measured NOx value at the absorber outlet. This is limited by the upper limit of the DCS data range (100mg / Nm3). Compared with the data on the Environmental Protection Bureau's website, which has no upper limit, the NOx measurement value rises sharply and exceeds the transmitter's upper limit due to air volume disturbances during operations such as starting and stopping the coal mill. The data collected by the Environmental Protection Bureau and the DCS data are incompatible, which can easily lead to environmental parameters exceeding the standard and cause data exceeding the standard cases.

[0003] The invention patent with publication number CN114225662A discloses an optimized control method for flue gas desulfurization and denitrification based on a hysteresis model. It solves the time delay problem in predicting sulfur dioxide concentration at the desulfurization outlet and nitrogen oxide concentration at the denitrification inlet, but does not solve the problem of data deviation between the environmental protection bureau and DCS caused by the upper limit setting of the measurement range.

[0004] The invention patent with publication number CN114048580A discloses a method for predicting air preheater blockage. By modifying the air preheater soot blowing system and ammonia supply system, it effectively avoids the generation of ammonium bisulfate in the air preheater during deep peak shaving. However, it does not address the impact of the upper limit of the range on environmental protection parameters.

[0005] Therefore, existing technologies have the drawback of insufficient early warning for exceeding environmental parameters when facing the problem of upper limit of measurement range. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method and system for intelligent monitoring and early warning of denitrification parameters.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] According to one aspect of the present invention, a method for intelligent monitoring and early warning of denitrification parameters is provided, the method comprising the following steps:

[0009] S1. Collect historical NOx data at the outlet of the absorption tower, and divide the data into over-range data segments and under-range data segments according to fixed time intervals.

[0010] S2. Preprocess and enhance the over-range and non-over-range data segments, and output the dataset.

[0011] S3. Train the autoencoder sub-model using the non-over-range data segments in the dataset, train the isolated forest sub-model using the over-range data segments in the dataset, and construct a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model.

[0012] S4. Collect NOx data in real time and input the data into the hybrid prediction model to predict and output the over-range probability and NOx data correction value at the next moment.

[0013] S5. Combining the real-time collected NOx data and the prediction results output by the hybrid prediction model, a three-level alarm is output, and the NOx data correction value and hourly average correction value are output as adjustment references.

[0014] As a preferred technical solution, data augmentation in S2 includes: performing linear interpolation, noise injection, time shifting, or scaling on the over-range data segments and the non-over-range data segments to generate new anomalous sample augmentation datasets.

[0015] As a preferred technical solution, the autoencoder sub-model in S3 includes an encoder and a decoder; the encoder is used to perform dimensionality reduction encoding on the data segment that is not out of range, mapping it to a low-dimensional latent space, and the decoder is used to decode the encoded data in the latent space and output the reconstructed data segment.

[0016] As a preferred technical solution, the specific process of training the autoencoder sub-model using the data segment that is not out of range in the dataset in S3 is as follows: First, the data segment that is not out of range is input into the autoencoder sub-model, and the reconstructed data segment is output; then, based on the data segment that is not out of range and the reconstructed data segment, the reconstruction error is established as the loss function, and the weights and biases of the neurons in the autoencoder sub-model are continuously adjusted through the backpropagation algorithm to minimize the reconstruction error until the model converges, thus completing the training of the autoencoder sub-model.

[0017] As a preferred technical solution, the specific formula for the reconstruction error is:

[0018]

[0019] Where E is the reconstruction error; X i This is the input data segment that is not outside the range; is the reconstructed data segment output by the decoder; n is the length of the data segment.

[0020] As a preferred technical solution, the specific process of training the isolated forest sub-model in S3 is as follows: For the training set data of the overrange data segment, a data segment and a split point on the data segment are randomly selected, and the data is divided into left and right subtrees. This process is repeated until each data point is isolated on a node, thus constructing multiple isolated trees. During the training process, the path length of each data point in the tree is recorded. The shorter the path, the greater the probability of overrange. By continuously adjusting the splitting strategy, the isolated forest sub-model is optimized so that it can accurately identify the characteristics of overrange data.

[0021] As a preferred technical solution, when processing NOx data using a hybrid prediction model in S4, the reconstructed data is first obtained using an autoencoder sub-model, and then the data is reconstructed using an isolated forest sub-model, thereby outputting the probability that the NOx data at the next moment belongs to the over-range data; at the same time, the NOx data correction value is calculated by combining the reconstruction error of the autoencoder and the historical non-over-range data segments.

[0022] As a preferred technical solution, the three-level alarm output in S5 includes the processes of first-level alarm, second-level alarm and third-level alarm.

[0023] The first-level alarm is as follows: if any NOx data reaches the preset upper limit, the alarm will be triggered immediately, and the NOx concentration mark at the reactor outlet will be displayed on the interactive interface for 3600 seconds to indicate that the range has been exceeded.

[0024] The level 2 alarm is: calculate the hourly average of the corrected NOx data. If any hourly average exceeds the preset maximum average, the alarm will be triggered immediately and the high hourly average value at the denitrification reactor outlet will be displayed on the interactive interface.

[0025] A level 3 alarm is triggered if the probability of exceeding the range at the next moment is greater than a preset probability value.

[0026] As a preferred technical solution, the preset maximum average value is calculated from the preset range upper limit.

[0027] According to another aspect of the present invention, a smart monitoring and early warning system for denitrification parameters is provided, the system comprising a data acquisition module, a data processing module, a model training module, a real-time early warning module, and a human-computer interaction module;

[0028] The data acquisition module is used to collect historical NOx data at the outlet of the absorption tower. After the data is segmented according to fixed time intervals, it is divided into over-range data segment and under-range data segment.

[0029] The data processing module is used to preprocess and enhance the over-range and under-range data segments, and output the dataset.

[0030] The model training module is used to train an autoencoder sub-model using the non-over-range data segments in the dataset, and to train an isolated forest sub-model using the over-range data segments in the dataset, thereby constructing a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model.

[0031] The real-time early warning module combines real-time collected NOx data with the prediction results output by the hybrid prediction model to output three levels of alarms, and outputs NOx data correction values ​​and hourly average correction values ​​as adjustment references.

[0032] The human-computer interaction module is used to display real-time NOx concentration data, historical trends, alarm information, and NOx data correction values.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. In this invention, by dividing the data into over-range and non-over-range segments, a hybrid model is constructed using an autoencoder and an isolated forest. NOx data is then collected in real-time and input into the model, outputting the over-range probability and correction value for the next time step. Finally, when the over-range probability exceeds a preset value, an alarm is triggered, and the NOx data correction value and hourly average correction value are output as adjustment references. Through targeted modeling, non-over-range data learns the patterns of the low-dimensional feature space through the autoencoder, while over-range data captures outlier features through the isolated forest, achieving differentiated modeling for data in different states and improving prediction accuracy. Furthermore, this method enables the system to have early warning capabilities. The hybrid model simultaneously outputs the over-range probability and correction value, triggering an alarm when the over-range probability exceeds a preset probability value at the next time step. This solves the problem of DCS alarms not being triggered due to the upper limit of the measurement point's range and improves the unit's operational stability.

[0035] 3. Combining real-time collected NOx data with the prediction results output by the hybrid prediction model, a three-level alarm output is implemented. If any NOx data reaches the preset upper limit, an alarm is triggered immediately. If any hourly average exceeds the preset maximum average, an alarm is triggered immediately. If the probability of exceeding the limit at the next moment is greater than the preset probability value, an alarm is triggered. The first-level alarm targets real-time exceedances, the second-level alarm targets trend anomalies, and the third-level alarm predicts in advance. This assists maintenance personnel in timely intervention and adjustment of the ammonia injection system's operation mode to prevent environmental parameters from exceeding limits. It achieves timely and accurate early warning, avoiding the problem of poor control of environmental indicators by maintenance personnel due to the limitations of the sensor element's upper limit, thus improving safety.

[0036] 4. In this invention, when processing NOx data using a hybrid prediction model, the autoencoder sub-model is first used to obtain reconstructed data, and then the isolated forest sub-model is used to reconstruct the data, thereby outputting the probability that the NOx data at the next moment belongs to the out-of-range data. Simultaneously, combining the autoencoder's reconstruction error and historical non-out-of-range data segments, a NOx data correction value is calculated. Combining the autoencoder reconstruction error and the isolated forest probability output, the reconstruction error reflects the degree of deviation of the current data from the normal pattern, while the out-of-range probability reflects the statistical significance of the anomaly. The combination of these two factors improves decision-making reliability. The correction value is calculated by weighting the historical normal data mean and the reconstruction error, providing maintenance personnel with an adjustment reference for the control system.

[0037] 5. In this invention, the data acquisition, processing, modeling, early warning, and interaction modules work together, enabling unmanned operation throughout the entire process from data acquisition to alarm output, reducing the cost of manual monitoring. The human-computer interaction module provides visualized information such as trend curves, alarm logs, and correction suggestions to help maintenance personnel quickly locate problems and improve efficiency. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of a smart monitoring and early warning method for denitrification parameters in this invention;

[0039] Figure 2 This is a schematic diagram of the over-range alarm logic calculation module for the NOx concentration marking at the reactor outlet in the embodiment.

[0040] Figure 3 This is a schematic diagram of the logic calculation module for the hourly average value at the outlet of the denitrification reactor in the embodiment. Detailed Implementation

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0042] This application aims to address the issue of DCS alarms not being triggered due to the upper limit of the measuring point range by using an online alarm system based on a monitoring and early warning model. By setting alarm parameters for the Nebula system, it enables operators to intervene and adjust the operation mode of the ammonia injection system in a timely manner, prevent environmental parameters from exceeding standards, improve the control capability of environmental indicators, enhance the operator's sensitivity to environmental parameters, strengthen the judgment of boiler combustion adjustments, and improve the stability of unit operation.

[0043] Example 1

[0044] like Figure 1As shown, this embodiment employs a smart monitoring and early warning method for denitrification parameters, which includes the following steps:

[0045] S1. Collect historical NOx data at the outlet of the absorption tower, and divide the data into over-range data segments and under-range data segments according to fixed time intervals.

[0046] S2. Preprocess and enhance the over-range and non-over-range data segments, and output the dataset.

[0047] S3. Train the autoencoder sub-model using the non-over-range data segments in the dataset, train the isolated forest sub-model using the over-range data segments in the dataset, and construct a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model.

[0048] S4. Collect NOx data in real time and input the data into the hybrid prediction model to predict and output the over-range probability and NOx data correction value at the next moment.

[0049] S5. Combining the real-time collected NOx data and the prediction results output by the hybrid prediction model, a three-level alarm is output, and the NOx data correction value and hourly average correction value are output as adjustment references.

[0050] The data augmentation in step S2 includes: performing linear interpolation, noise injection, time shifting, or scaling on the over-range data segment and the non-over-range data segment to generate new anomalous sample augmentation datasets; preferably, the autoencoder sub-model in step S3 includes an encoder and a decoder; the encoder is used to perform dimensionality reduction encoding on the non-over-range data segment, mapping it to a low-dimensional latent space, and the decoder is used to decode the encoded data in the latent space and output the reconstructed data segment.

[0051] The specific process of training the autoencoder sub-model using the data segment that is not out of range in step S3 is as follows: First, input the data segment that is not out of range into the autoencoder sub-model and output the reconstructed data segment; then, based on the data segment that is not out of range and the reconstructed data segment, establish the reconstruction error as the loss function, and continuously adjust the weights and biases of neurons in the autoencoder sub-model through the backpropagation algorithm to minimize the reconstruction error until the model converges and the training of the autoencoder sub-model is completed.

[0052] The specific process of training the isolated forest sub-model in step S3 is as follows: For the training set data of the overrange data segment, randomly select a data segment and a split point on the data segment, divide the data into left and right subtrees, and repeat this process until each data point is isolated on a node, thus constructing multiple isolated trees; during the training process, record the path length of each data point in the tree. The shorter the path, the greater the probability of overrange; by continuously adjusting the splitting strategy, optimize the isolated forest sub-model so that it can accurately identify the characteristics of overrange data.

[0053] The three-level alarm output in S5 includes the first-level alarm, second-level alarm, and third-level alarm processes.

[0054] The first-level alarm is as follows: if any NOx data reaches the upper limit of the preset range, the alarm will be triggered immediately, and the NOx concentration mark at the reactor outlet will be displayed on the interactive interface for 3600 seconds to indicate that the range has been exceeded.

[0055] The secondary alarm is as follows: calculate the hourly average of the corrected value of NOx data. If any hourly average exceeds the preset maximum average, an alarm will be triggered immediately, and the high hourly average value at the denitrification reactor outlet will be displayed on the interactive interface.

[0056] The three-level alarm is triggered if the probability of exceeding the range at the next moment is greater than a preset probability value.

[0057] In step S4, when processing NOx data using a hybrid prediction model, the autoencoder sub-model is first used to obtain reconstructed data, and then the isolated forest sub-model is used to reconstruct the data, thereby outputting the probability that the NOx data at the next moment belongs to the over-range data; at the same time, the NOx data correction value is calculated by combining the reconstruction error of the autoencoder and the historical non-over-range data segments.

[0058] Because there is a discrepancy between the export NOx baseline oxygen content DCS data and the data on the Environmental Protection Bureau's website, relevant alarms have been established in the system before the upper limit of the measurement range is re-registered with the Environmental Protection Bureau:

[0059] 1. When NOx concentration reaches the upper limit of 100 mg / Nm3 on either side of the absorber outlet, the system will alarm "NOx concentration mark at reactor outlet exceeds range" for 3600 seconds.

[0060] 2. Simultaneously with exceeding the measurement range, if the hourly average of the "SCR outlet concentration correction value" on the denitrification system screen exceeds 45 mg / Nm3, the system will alarm with "High hourly average value at the outlet of the x side of the denitrification reactor".

[0061] 3. The system adds the hourly average value of NOx correction at the A / B side outlet of the Selective Catalytic Reduction (SCR) measurement point for reference when exceeding the measurement range.

[0062] This solution resolves the issue of DCS alarms not being triggered due to the upper limit of the measuring point's range. Furthermore, by setting system alarm parameters, operators can promptly intervene and adjust the ammonia injection system's operation to prevent environmental parameters from exceeding limits.

[0063] This solution, based on a smart power plant PaaS (Platform as a Service) layer intelligent monitoring and early warning model, can quickly and accurately prevent the limitations of sensor components from affecting operators' control of environmental indicators. The intelligent alarm and early warning model improves operators' understanding of environmental parameters and enhances their sensitivity to these parameters. This optimization method strengthens operators' judgment on boiler combustion adjustments, allowing for early detection of operations that may affect furnace combustion, such as switching coal mills, thus improving unit operational stability.

[0064] Example 2

[0065] In this embodiment, a smart monitoring and early warning system for denitrification parameters is adopted. The system includes a data acquisition module, a data processing module, a model training module, a real-time early warning module, and a human-computer interaction module.

[0066] The data acquisition module is used to collect historical NOx data at the outlet of the absorption tower. After the data is segmented according to fixed time intervals, it is divided into over-range data segment and under-range data segment.

[0067] The data processing module is used to preprocess and enhance the over-range and under-range data segments, and output the dataset.

[0068] The model training module is used to train an autoencoder sub-model using the non-over-range data segments in the dataset, and to train an isolated forest sub-model using the over-range data segments in the dataset, thereby constructing a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model.

[0069] The real-time early warning module combines real-time collected NOx data with the prediction results output by the hybrid prediction model to output three levels of alarms, and outputs NOx data correction values ​​and hourly average correction values ​​as adjustment references.

[0070] The human-computer interaction module is used to display real-time NOx concentration data, historical trends, alarm information, and NOx data correction values.

[0071] In the denitrification system of a 600MW coal-fired power unit, a laser-type NOx concentration analyzer installed at the outlet of the absorption tower continuously collects nearly one year's worth of historical NOx data at one-minute intervals. The collected data includes NOx concentration values ​​under normal operating conditions, load fluctuation conditions, and equipment start-up and shutdown conditions. These data are then segmented into one-hour intervals, with a preset upper limit of 100 mg / Nm³. 3 The data segment was divided into an over-range data segment (NOx concentration ≥ 100 mg / Nm³). 3 ) and the data segment not exceeding the range (NOx concentration < 100 mg / Nm 3 ).

[0072] The two types of data segments were preprocessed. First, data segments with a missing rate exceeding 10% were removed. For the few missing data points, cubic spline interpolation was used to fill them in. Then, Z-score standardization was used to map the data uniformly to a distribution interval with a mean of 0 and a standard deviation of 1, eliminating the influence of units.

[0073] In the data augmentation stage, diverse processing methods are applied to both over-range and under-range data segments:

[0074] Linear interpolation: Randomly inserting new data points into a data sequence to simulate the gradual change of the data;

[0075] Noise injection: Add random noise that conforms to a normal distribution (mean 0, standard deviation 0.1) to the data to simulate sensor measurement error;

[0076] Time shifting or scaling: Shift the data sequence forward or backward by 5-10 time steps, or scale the data sequence by 0.8-1.2 times to generate new anomalous samples. The final result is a complete dataset containing both the original and augmented data, divided into training and test sets in a 7:3 ratio.

[0077] 2. Model Training and Construction Phase

[0078] The autoencoder sub-model was trained using under-range data segments from the dataset. The autoencoder employed a 5-layer neural network structure: the encoder contained 3 layers with 128, 64, and 32 neurons respectively, using the ReLU activation function; the decoder also contained 3 layers with 64, 128 neurons respectively, and the original data dimension. During training, the under-range data segments were input into the autoencoder, and the reconstructed data segments were output. An optimizer was used to iterate for 200 epochs at a learning rate of 0.001 to minimize the reconstruction error until the model converged.

[0079] The isolated forest sub-model is trained using overrange data segments from the dataset, with each tree randomly sampling 50% of the training set. During training, for each overrange data segment, a feature and a split point on that feature are randomly selected, and the data is divided into left and right subtrees. This process is repeated until each data point is isolated on a single node. The path length of each data point in the tree is recorded; the shorter the path, the higher the probability that the data point belongs to the overrange data segment. By adjusting the sampling ratio and the number of trees, the isolated forest sub-model is optimized to accurately identify the features of the overrange data segment. Finally, the trained autoencoder sub-model and the isolated forest sub-model are combined to construct a hybrid prediction model.

[0080] 3. Real-time monitoring and alarm stage

[0081] During real-time NOx data acquisition, the NOx concentration data at the absorber outlet is obtained once per second via the interface between the DCS system and the intelligent monitoring and early warning system. The real-time acquired data is input into a hybrid prediction model. The autoencoder sub-model first encodes and decodes the data to obtain reconstructed data and calculates the reconstruction error. The isolated forest sub-model then calculates the probability that the data belongs to the out-of-range range at the next moment based on the reconstructed data. Simultaneously, combining the autoencoder's reconstruction error and the distribution characteristics of historical non-out-of-range data segments, the Kalman filter algorithm is used to calculate the NOx data correction value.

[0082] It outputs three levels of alarms, including the first-level alarm, second-level alarm, and third-level alarm processes;

[0083] Construct a logic calculation module for over-range alarm of NOx concentration labeling at reactor outlet, such as... Figure 2 As shown; construct the logic calculation module structure for the hourly average value at the outlet of the denitrification reactor, as follows: Figure 3 As shown, the first-level alarm process is performed by the over-range alarm logic calculation module of the NOx concentration marker at the reactor outlet, and the second-level alarm process is performed by the high hourly average value logic calculation module at the denitrification reactor outlet.

[0084] The first-level alarm is as follows: if any NOx data reaches the preset upper limit, the alarm will be triggered immediately, and the NOx concentration mark at the reactor outlet will be displayed on the interactive interface for 3600 seconds to indicate that the range has been exceeded.

[0085] The level 2 alarm is: calculate the hourly average of the corrected NOx data. If any hourly average exceeds the preset maximum average, the alarm will be triggered immediately and the high hourly average value at the denitrification reactor outlet will be displayed on the interactive interface.

[0086] A level 3 alarm is triggered if the probability of exceeding the range at the next moment is greater than a preset probability value.

[0087] 4. System Deployment and Effect Verification

[0088] This solution is deployed based on the PaaS layer of a smart power plant, utilizing the containerized deployment environment and distributed computing resources provided by the platform to achieve efficient operation of the hybrid prediction model. During the trial operation after system deployment, a total of 68 over-range events were detected, including 63 Level 1 alarms and 5 Level 2 alarms. Timely alarm notifications and adjustments to the ammonia injection system effectively prevented multiple instances of exceeding environmental parameter limits. Simultaneously, by referring to the NOx data correction values ​​and alarm information provided by the system, operators can more accurately judge the boiler combustion conditions and perform operations such as coal mill switching in advance, reducing NOx concentration fluctuations caused by improper combustion adjustments and significantly improving unit operational stability.

[0089] This embodiment verifies the effectiveness and reliability of the intelligent monitoring and early warning method for denitrification parameters through practical application. It can effectively solve the problem of untimely alarms caused by the upper limit of the measuring point range, and improve the monitoring level of environmental parameters and the safety of unit operation.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for intelligent monitoring and early warning of denitrification parameters, characterized in that, The method steps include: S1. Collect historical NOx data at the outlet of the absorption tower, and divide the data into over-range data segments and under-range data segments according to fixed time intervals. S2. Preprocess and enhance the over-range and under-range data segments, and output the dataset. S3. Train the autoencoder sub-model using the non-over-range data segments in the dataset, train the isolated forest sub-model using the over-range data segments in the dataset, and construct a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model. S4. Collect NOx data in real time and input the data into the hybrid prediction model to predict and output the over-range probability and NOx data correction value at the next moment. S5. Combining the real-time collected NOx data and the prediction results output by the hybrid prediction model, a three-level alarm is output, and the NOx data correction value and hourly average correction value are output as adjustment references.

2. The intelligent monitoring and early warning method for denitrification parameters according to claim 1, characterized in that, The data augmentation in S2 includes: performing linear interpolation, noise injection, time shifting, or scaling on over-range and under-range data segments to generate new anomalous sample augmentation datasets.

3. The intelligent monitoring and early warning method for denitrification parameters according to claim 1, characterized in that, The autoencoder sub-model in S3 includes an encoder and a decoder; the encoder is used to perform dimensionality reduction encoding on the data segment that is not out of range, mapping it to a low-dimensional latent space, and the decoder is used to decode the encoded data in the latent space and output the reconstructed data segment.

4. The intelligent monitoring and early warning method for denitrification parameters according to claim 3, characterized in that, The specific process of training the autoencoder sub-model using the non-over-range data segment in the dataset in S3 is as follows: First, the non-over-range data segment is input into the autoencoder sub-model, and the reconstructed data segment is output; then, based on the non-over-range data segment and the reconstructed data segment, the reconstruction error is established as the loss function, and the weights and biases of the neurons in the autoencoder sub-model are continuously adjusted through the backpropagation algorithm to minimize the reconstruction error until the model converges, thus completing the training of the autoencoder sub-model.

5. The intelligent monitoring and early warning method for denitrification parameters according to claim 4, characterized in that, The specific formula for the reconstruction error is: Where E is the reconstruction error; X i This is the input data segment that is not outside the range; is the reconstructed data segment output by the decoder; n is the length of the data segment.

6. The intelligent monitoring and early warning method for denitrification parameters according to claim 1, characterized in that, The specific process of training the isolated forest sub-model in S3 is as follows: For the training set data of the overrange data segment, randomly select a data segment and a split point on the data segment, divide the data into left and right subtrees, repeat this process until each data point is isolated on a node, and construct multiple isolated trees. During training, the path length of each data point in the tree is recorded. The shorter the path, the greater the probability of exceeding the range. By continuously adjusting the segmentation strategy, the isolated forest sub-model is optimized so that it can accurately identify the characteristics of the data exceeding the range.

7. The intelligent monitoring and early warning method for denitrification parameters according to claim 1, characterized in that, In S4, when processing NOx data using a hybrid prediction model, the autoencoder sub-model is first used to obtain reconstructed data, and then the isolated forest sub-model is used to reconstruct the data, thereby outputting the probability that the NOx data at the next moment belongs to the over-range data; at the same time, the NOx data correction value is calculated by combining the reconstruction error of the autoencoder and the historical non-over-range data segments.

8. The intelligent monitoring and early warning method for denitrification parameters according to claim 1, characterized in that, The three-level alarm output in S5 includes the processes of level one alarm, level two alarm, and level three alarm. The first-level alarm is as follows: if any NOx data reaches the upper limit of the preset range, the alarm will be triggered immediately, and the NOx concentration mark at the reactor outlet will be displayed on the interactive interface for 3600 seconds to indicate that the range has been exceeded. The secondary alarm is as follows: calculate the hourly average of the corrected value of NOx data. If any hourly average exceeds the preset maximum average, an alarm will be triggered immediately, and the high hourly average value at the denitrification reactor outlet will be displayed on the interactive interface. The three-level alarm is triggered if the probability of exceeding the range at the next moment is greater than a preset probability value.

9. The intelligent monitoring and early warning method for denitrification parameters according to claim 8, characterized in that, The preset maximum mean value is obtained by linear scaling of the preset range.

10. A smart monitoring and early warning system for denitrification parameters, characterized in that, The system operates using a smart monitoring and early warning method for denitrification parameters as described in any one of claims 1-9. The system includes a data acquisition module, a data processing module, a model training module, a real-time early warning module, and a human-computer interaction module. The data acquisition module is used to collect historical NOx data at the outlet of the absorption tower. After the data is segmented according to a fixed time interval, it is divided into over-range data segment and under-range data segment. The data processing module is used to preprocess and enhance the over-range data segment and the non-over-range data segment, and output the dataset; The model training module is used to train an autoencoder sub-model using the non-over-range data segment in the dataset, and to train an isolated forest sub-model using the over-range data segment in the dataset, thereby constructing a hybrid prediction model that includes the trained autoencoder sub-model and the isolated forest sub-model. The real-time early warning module is used to combine the real-time collected NOx data and the prediction results output by the hybrid prediction model to output a three-level alarm, and output the NOx data correction value and the hourly average correction value as an adjustment reference. The human-computer interaction module is used to display real-time NOx concentration data, historical trends, alarm information, and NOx data correction values.

Images