Online monitoring method and equipment for CO concentration in incinerator based on deep learning

Abstract

The invention discloses a deep learning-based online monitoring method and equipment for the concentration of CO in an incinerator. The method comprises the following steps: acquiring CO concentration data at a first flue in an incinerator and image data of combustion flame and flue gas in a period of time; the CO concentration value in a period of time is matched with the image in the period of time to serve as a data set for model training; after model training is completed, analysis of continuous images of flames of the first flue can be achieved only through a high-temperature camera and an industrial personal computer which are arranged on the first flue, and the identified CO concentration grade is timely reflected to a fireman. Compared with the existing CO concentration measurement method, the method provided by the invention can more rapidly and accurately monitor the concentration grade of CO. In practical application, the system not only can assist a fireman in regulation, but also can be integrated with an automatic combustion control system to carry out intelligent regulation, so that the combustion stability is improved.

Description

Technical Field

[0001] The present invention belongs to the field of waste incineration pollutant emission control, and relates to a method and device for online monitoring of CO concentration, and in particular to a method and device for online monitoring of CO concentration in an incinerator based on deep learning. Background Art

[0002] Municipal solid waste incineration offers the advantages of minimal land use, significant waste reduction, high levels of harmlessness, and potential source recovery. However, the complex composition of municipal solid waste can lead to significant combustion fluctuations within the incinerator. Currently, incinerators rely primarily on manual control, partly dependent on the operator's experience. This results in unstable operating conditions and excessive pollutant emissions, negatively impacting production operations. Incomplete incineration produces carbon monoxide (CO), and excessive CO production can reduce operational efficiency and harm human health.

[0003] There are two main methods for measuring CO in incinerator systems: continuous emission monitoring system (CEMS) and tunable diode laser absorption spectroscopy (TDLAS). However, the CO concentration in the furnace rises rapidly, and the CO concentration can rise to 100 mg / Nm in 10 to 30 seconds. 3 Above, even 1000mg / Nm 3 The above. The CEMS system is installed at the tail end of the incinerator system, and it takes 2 to 3 minutes to measure the flue gas flow and equipment, and it is impossible to measure the CO exceeding the standard in time. The TDLAS system is expensive, and when it works in the incinerator for a long time, it requires professional maintenance and calibration. In addition, there are some soft measurement technologies based on machine learning that use historical DCS data for modeling (such as long short-term memory neural networks) to predict pollutant concentrations in the short term in the future, but this method ignores the information conveyed by the combustion conditions and has low accuracy. Therefore, there is an urgent need for a low-cost, high-accuracy, and high-timeliness measurement method for measuring CO concentrations in incinerators. Summary of the Invention

[0004] To address these issues, the present invention proposes a deep learning-based method and device for online monitoring of CO concentration in incinerators. This method identifies pollutant concentration levels based on flame images within the furnace. Compared to traditional measurement methods, this method offers advantages such as increased timeliness, lower cost, and higher accuracy. This method can assist boiler operators in timely understanding CO concentration changes within the incinerator. The identification results can also be used as feedback for automated pollutant emission control models, ultimately achieving low incinerator emissions.

[0005] The technical solution adopted in the present invention is:

[0006] A method for online monitoring of CO concentration in an incinerator based on deep learning, comprising the following steps:

[0007] Images of burning flames and flue gases in the incinerator are obtained and input into a pre-trained 3D convolutional neural network model to output the CO concentration level.

[0008] Furthermore, the pre-training process of the 3D convolutional neural network model includes:

[0009] Obtain the video of the burning flame and flue gas in the incinerator and the CO concentration data at the first flue of the incinerator, and preprocess them to obtain a flame and flue gas image sequence and preprocessed CO concentration data;

[0010] Divide the preprocessed CO concentration data into multiple CO concentration-time data segments and perform CO concentration level classification;

[0011] Match the CO concentration-time data segments with the flame smoke image sequence, label the image sequence according to the CO concentration level, and obtain the flame smoke image sequence-CO concentration level dataset;

[0012] The 3D convolutional neural network model is trained using flame smoke image sequence-CO concentration level data.

[0013] Furthermore, the video of the combustion flame and flue gas is obtained by installing a high-temperature camera at the first flue of the incinerator; and the CO concentration data is obtained by installing a high-temperature laser flue gas analyzer at the first flue.

[0014] Furthermore, the preprocessing includes: extracting frames from the video and filtering out abnormal and blurred images; and removing missing values ​​and erroneous values ​​in the CO concentration data.

[0015] Furthermore, the preprocessed CO concentration data is divided into multiple CO concentration-time data segments, specifically comprising the steps of setting a CO concentration threshold, selecting an excessive segment and a normal segment in the CO concentration time series, and recording the start and end time of the segment to obtain multiple CO concentration-time data segments.

[0016] Furthermore, the method for dividing the exceeding-standard segments is as follows: calculating the peak value or mean value of the CO concentration; if the peak value or mean value is greater than the CO concentration threshold value, traversing the CO concentration data in chronological order; if a preset maximum length is reached and the segment does not have a CO concentration exceeding the threshold value, recording the start and end time of the segment as a CO concentration segment; if the CO concentration exceeds the threshold value at a certain moment during the traversal, recording the start time of the CO concentration segment as a specified length of time before the moment when the CO concentration exceeds the threshold value, and recording the end time of the CO concentration segment as a specified length of time after the CO concentration returns to below the threshold value; thereby obtaining multiple CO concentration-time data segments.

[0017] Furthermore, the method for classifying CO concentration levels is: calculating the CO concentration peak value or mean value of each CO concentration-time data segment, and classifying the CO concentration-time data segment into several levels according to the peak value or mean value.

[0018] Furthermore, the CO concentration-time data segments are matched with the image sequence, and the image sequence is labeled according to the CO concentration level. Specifically, the steps include: determining the time delay between the measured CO concentration data and the image based on the flow time of the flue gas in the flue, matching the CO concentration-time data segments with the image based on the shooting time of the image, and labeling the image according to the CO concentration level, thereby obtaining a flame flue gas image sequence-CO concentration level dataset.

[0019] Furthermore, the 3D convolutional neural network model is a spatiotemporal modeling deep learning model for video action recognition.

[0020] Furthermore, the 3D convolutional neural network model takes continuous image segments as input. When the number of frames of the original continuous image segments input is N origin When it is greater than the preset input frame number N, determine the starting frame index f start , where 1≤f start ≤N origin -N+1, input N frames of continuous image sequence; when the number of frames of the input original continuous image segment is N origin When the number of input frames is less than the preset number N, the last frame image is repeatedly input until the preset number N of input frames is reached.

[0021] A computer device, comprising:

[0022] one or more processors;

[0023] a memory for storing one or more programs;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned deep learning-based online monitoring method for CO concentration in the incinerator.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) This method installs a high-temperature industrial camera in the first flue of the incinerator to monitor the CO concentration level by inferring the flame and flue gas image sequence. Compared with the existing CEMS system, it can detect the excessive CO in the furnace earlier, and the early warning time is about 1-4 minutes. It can assist the boiler operator to make timely adjustments and reduce pollutant emissions.

[0027] (2) The economic cost of this method is low. It only requires the deployment of a high-temperature camera and a computer equipped with an online CO concentration level monitoring model or software in the furnace to achieve visualization of the first flue and online monitoring of CO concentration. This ensures rapid and accurate monitoring of CO concentration levels while significantly reducing the cost of measurement hardware. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of model training and deployment in an embodiment of the present invention.

[0029] Figure 2 This is a flow chart of the steps for implementing the method in an embodiment of the present invention.

[0030] Figure 3 These are typical flame and smoke images at a time when CO is normal and a time when CO exceeds the standard in an embodiment of the present invention.

[0031] Figure 4 Schematic diagram of the CO data segment division method in an embodiment of the present invention.

[0032] Figure 5 The flowchart of data set production in the embodiment of the present invention is shown.

[0033] Figure 6 This is a monitoring effect diagram of the model on the garbage incineration grate furnace in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0035] like Figure 1 As shown, a method for online monitoring of CO concentration in an incinerator based on deep learning includes the following steps:

[0036] Images of burning flames and flue gases in the incinerator are obtained and input into a pre-trained 3D convolutional neural network model to output the CO concentration level.

[0037] The pre-training process of the 3D convolutional neural network model includes:

[0038] (1) Obtain the video of the burning flame and flue gas in the incinerator and the CO concentration data at the first flue of the incinerator, and perform preprocessing.

[0039] The combustion flame and flue gas videos are captured by installing a high-temperature camera in the first flue of the incinerator, with the camera pointed downward at a 45-degree angle toward the bottom of the first flue. The CO concentration data is acquired by installing a high-temperature laser flue gas analyzer in the first flue. Data and image acquisition should cover a wide range of timeframes to enhance the robustness of the trained model. Therefore, the data and image collection should span no less than three months.

[0040] The preprocessing includes extracting frames from the video, for example, extracting one frame every 1 second and filtering out abnormal and blurred images to obtain a sequence of flame and smoke images; removing missing and erroneous values ​​from the CO concentration data; and calculating and analyzing the CO concentration mean and histogram distribution to serve as a basis for categorizing concentration levels. For example, the smallest data point in the first 90% of the CO concentration histogram can be used as the dividing line between normal and excessive concentrations.

[0041] (2) Divide the preprocessed CO concentration data into multiple CO concentration-time data segments and perform CO concentration level classification. The specific steps include:

[0042] A CO concentration threshold is set based on actual conditions. The CO concentration data is traversed in chronological order. If a segment reaches the preset maximum length and the CO concentration never exceeds the threshold, the segment is recorded as a CO concentration segment with its start and end times. If the CO concentration exceeds the threshold at any point during the traversal, the specified length of time before the CO concentration exceeds the threshold is used as the CO concentration segment start time, and the specified length of time after the CO concentration returns to below the threshold is used as the CO concentration segment end time. The peak or mean CO concentration of each CO concentration-time segment is then calculated to perform grading.

[0043] When partitioning by mean, the CO concentration data is traversed chronologically. If the maximum length is reached and no CO concentration exceeds the threshold, the segment is considered normal. If the CO concentration threshold is exceeded at any point during the traversal, the segment begins with the specified length of time at the time the threshold was exceeded and ends with the specified length of time below the threshold. The mean of the segment is then calculated, and after all segments are partitioned, each segment is classified according to its mean. (If a value exceeds the threshold but the mean is less than the threshold, it is still classified as normal.) When partitioning by peak, the peak value is used as the partitioning basis, and the remaining methods are similar to those for partitioning by mean. To ensure data accuracy, the selection of normal segments is kept at a certain interval from the exceeding threshold segments.

[0044] Each segment can be set to no more than T seconds according to the actual situation. The CO concentration is set from t1 seconds before exceeding the standard to t2 seconds after returning to normal as a complete exceeding standard segment; the CO concentration segment with a time of less than t0 seconds is not included in the data set. The boundaries of different levels c1, c2..., c n , the concentration segments are classified into different levels according to the mean or peak value 0~c1, c1~c2…, c n Generally, 0 to c1 are fragments with normal concentrations, and the rest are fragments exceeding the standard to varying degrees.

[0045] (3) Match the CO concentration-time data segments with the image sequence, label the image sequence according to the CO concentration level, and obtain the image sequence-CO concentration level dataset. The specific steps include:

[0046] The time delay τ (generally 1 to 5 seconds) between the CO concentration data measured by the laser instrument and the image is determined based on the flow time of the flue gas in the first flue. The time corresponding to each image is obtained based on the timestamp of the camera image. Combined with the delay time between the image and the CO concentration data, the CO concentration-time data segment is matched with the image. The image is then labeled according to the CO concentration level to obtain the image sequence-CO concentration level dataset, which is then divided into a training set and a validation set.

[0047] (4) The 3D convolutional neural network model is trained using the training set of the image sequence-CO concentration level dataset, so that the model can be applied to the online monitoring of the CO concentration level in the grate furnace of domestic waste incineration, and the hyperparameters are adjusted using Bayesian optimization.

[0048] The 3D convolutional neural network model is a spatiotemporal modeling deep learning model for video action recognition. For example, a video recognition model such as a Slow-Fast model or an R(2+1)D model can be selected.

[0049] The loss function for model training is cross entropy loss. y j is the one-hot encoded vector of the actual label, Is the probability distribution vector predicted by the model, q is the number of categories. Using the stochastic gradient descent algorithm (sgdm), the parameter update formula is θ t+1 =θ t +v t , where θ t+1 is the parameter of the current model, θ t is the updated model parameter, v t is the current momentum; the momentum update formula is Where: v t is the momentum term at step t; β is the momentum coefficient; is the objective function J θ In the previous step, the parameter θ t-1 The gradient at θ t is the updated parameter after step t; θ t-1 is the parameter of the t-1th step; η is the learning rate.

[0050] The 3D convolutional neural network model takes continuous image segments as input, and the preset number of continuous input frames is N, which is usually 8, 16, 32, or 64. origin When the number of input frames is greater than the preset number N, the starting frame index f is randomly determined start (1≤f start ≤N origin -N+1), input N frames of continuous image sequence; when the number of continuous image segments input is N origin When the number of input frames is less than the preset number N, the last frame image is repeatedly input until the preset number N of input frames is reached.

[0051] The trained 3D convolutional neural network model was used to perform inference analysis on samples from the validation set, and the model's accuracy on the validation set was calculated. The trained model was then deployed on the project's municipal waste incineration grate furnace system, and its accuracy and timeliness were evaluated by comparing it with measurement results from a laser flue gas analyzer.

[0052] Based on the evaluation results and combined with the actual project, the model recognition results within a fixed time are integrated to formulate a CO concentration exceeding standard alarm strategy. For example, within a fixed time window, if the model recognition results exceed the standard more than a specified number of times, an exceeding standard alarm will be issued, thereby reducing the number of false alarms and improving the reliability of the model application. After completion, the high-temperature laser flue gas analyzer can be removed, and the high-temperature camera and model can be used to achieve online monitoring of CO concentration levels. High-temperature laser flue gas analyzers are expensive, ranging from tens of thousands to hundreds of thousands of yuan; the air-cooled high-temperature camera required for this method only costs about 1,000-2,000 yuan, which significantly reduces costs while ensuring effective monitoring of CO concentration in the furnace.

[0053] like Figure 2 As shown in the figure, in a specific embodiment of the present invention, the method was implemented in a domestic waste incineration grate furnace of a waste-to-energy project. The total design capacity of the furnace is 500 tons of waste per day. Before implementation, the incinerator system of the project had excessive CO concentration, and there were many cases of transient CO exceeding the standard during operation. The average hourly CO emission value exceeded 50mg / Nm 3 .

[0054] S1: Image Data Acquisition Phase: A high-temperature industrial camera and a high-temperature flue gas measurement laser were installed in the first flue of the project's waste incineration grate furnace. The CO concentration was measured using an ETG-01OP open-path laser gas detector, based on tunable diode laser absorption spectroscopy (TDLAS) technology, which provides real-time information on changes in target gas concentration over time. The camera features a 45° downward-angled peephole on the top wall of the probe. Cooled with 0.3-0.5 MPa air, it can operate for extended periods at temperatures below 800°C and in environments with mildly corrosive atmospheres. Over 4,800 hours of video footage of the first flue were collected, with CO concentration data collected every 5 seconds, yielding over 3.5 million CO concentration data points.

[0055] S2: Image data preprocessing stage: preliminary screening deleted missing values ​​and erroneous values ​​in the CO concentration data; extracted a frame of image from the video every 1s, and filtered out abnormal and blurred images. By analyzing all DCS data, the average CO concentration was 30.51 mg / Nm 3 .

[0056] S3: Data segment division stage: Divide the CO concentration data of the first flue into multiple CO concentration-time data segments, and perform CO concentration level division. Figure 4 As shown in the figure, the CO concentration-time data segment selection rules are as follows:

[0057] (1) Each clip should not exceed 150 seconds;

[0058] (2) The average value exceeds 60 mg / Nm 3 The fragments with concentration exceeding the standard were defined as the fragments with concentration exceeding the standard;

[0059] (3) The 10 seconds before the CO concentration exceeds the standard and the 10 seconds after it returns to normal are considered a complete exceeding standard segment;

[0060] (4) The segments with normal CO concentration less than 60 seconds were discarded.

[0061] S4: Flame smoke image sequence-CO concentration level dataset production stage: According to the image timestamp, the corresponding time of each image is obtained, and the delay time between the image and the CO concentration data is combined to match the CO concentration-time data segment and the image through time to complete the continuous image sequence dataset production ( Figure 5 ). The average value exceeds 60mg / Nm 3 The segments with excessive CO concentration were defined as segments with excessive CO concentration, and the rest were normal segments. A total of 6683 image sequences with excessive CO concentration and 52775 image sequences with normal CO concentration were obtained. The dataset was divided into training and validation sets in a ratio of 8:2.

[0062] S5: Model Training: The Slow-Fast model was selected. This model combines two feature extraction paths at different rates: one branch processes rapidly changing dynamic information, and the other processes slowly changing static information. This allows the model to better understand both dynamic and static features of the video content. The number of image segments was set to 32, and the model was trained through 32 rounds. After 18,000 iterations, the model's accuracy and loss on the training set reached stability.

[0063] S6: Model Validation: The trained video recognition model was used to perform inference analysis on samples from the validation set. The validation set achieved an accuracy rate exceeding 90%, meeting the requirements for identifying pollutant concentration levels in municipal waste incineration power plants. The trained model was deployed on the project's municipal waste incineration grate system, reading camera footage and performing real-time analysis of continuous images. The recognition results were consistent with the flue gas laser data.

[0064] S7: Model Application Phase: Based on project realities and the model's identification results within a fixed timeframe, a CO concentration over-limit alarm strategy was developed. This strategy triggers an alarm if the model's identification results exceed the limit a specified number of times within a fixed timeframe. In actual application, the model's accuracy in identifying CO over-limits exceeded 90%, with minimal false alarms. The high-temperature laser flue gas analyzer can now be put into operation, and online CO concentration monitoring can be achieved using the high-temperature camera and model.

[0065] The method proposed in the present invention is to identify the pollutant concentration level in the furnace by reasoning and analyzing the continuous images of the flame shape of the first flue. The generation of CO in the garbage incineration grate furnace is mainly due to the incomplete combustion of the fuel. The excessive CO concentration often occurs when the material collapses, deflagrations, etc. occur in the furnace. By installing a high-temperature camera in the first flue of the garbage incineration grate furnace, the flame shape corresponding to different CO concentrations in the furnace can be observed. When the CO concentration in the furnace is normal, the flame is translucent, the flame shape is clearly visible, and there is no obvious smoke in the flue; when CO explodes, the first flue image appears to be filled with white smoke, and the flame shape is difficult to observe clearly. Figure 3As shown in the figure, (a) shows a typical flame and flue gas image at a normal CO concentration, and (b) shows a flame and flue gas image at a CO concentration exceeding the limit. During a CO outbreak, CO concentrations rapidly rise to extremely high values ​​within 5 to 30 seconds and typically persist for approximately 30 to 180 seconds. Correspondingly, the flames within the furnace gradually change from a clear, translucent state to a smoky, dusty state, lasting for a correspondingly long time. Because this phenomenon exhibits characteristics on both spatial and temporal scales, the present invention uses a deep learning algorithm to analyze and model continuous images of the first flue flame. Compared to data-driven modeling methods, this method focuses on the source of CO generation—incomplete fuel combustion. Currently, image recognition within waste incinerators mostly relies on single images. However, during a CO outbreak, the flame images within the flue are dynamic, with dynamic features such as the swaying and flickering of the flames and the rising and spreading of smoke. Single-frame image models can only capture spatial features of the current frame, such as color and shape, and cannot learn these time-varying features. Therefore, the present invention adopts a continuous image recognition algorithm, which can accurately and timely identify the CO exceeding the standard phenomenon.

[0066] When training the video recognition model, a high-temperature laser instrument is installed at the first flue to obtain the CO concentration value at that point, and a high-temperature camera is used to capture images of the first flue. CO concentration data and image data are obtained over a period of time. The CO concentration values ​​over this period of time are matched with the images during that period of time as the training dataset for the model. The training dataset consisting of the first flue image data is input into the model for training, and the hyperparameters of the training model are optimized to obtain a model for identifying carbon monoxide concentration in the furnace.

[0067] When the model is applied, after the video recognition model training is completed, only the high-temperature camera and industrial control computer deployed in the first flue can realize the analysis of the continuous image of the first flue flame and promptly report the identified CO concentration level to the boiler operator. Figure 6 As shown in FIG. 1 , a monitoring effect diagram of the model of the present invention on a waste incineration grate furnace in a specific embodiment is shown. The model identifies the situation when CO explodes and issues an alarm when it predicts that the level exceeds the standard for 29 consecutive seconds. The alarm event is about 70 seconds earlier than that of the flue gas monitoring system.

[0068] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0070] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0072] The above description is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can use the above disclosed methods and technical contents to make many possible changes and modifications to the technical solution of the present invention without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment with equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A method for online monitoring of CO concentration in an incinerator based on deep learning, characterized in that: The following steps are involved: Images of burning flames and flue gases in the incinerator are obtained and input into a pre-trained 3D convolutional neural network model to output the CO concentration level.

2. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 1 is characterized in that: The pre-training process of the 3D convolutional neural network model includes: Obtain the video of the burning flame and flue gas in the incinerator and the CO concentration data at the first flue of the incinerator, and preprocess them to obtain a flame and flue gas image sequence and preprocessed CO concentration data; Divide the preprocessed CO concentration data into multiple CO concentration-time data segments and perform CO concentration level classification; Match the CO concentration-time data segments with the flame smoke image sequence, label the image sequence according to the CO concentration level, and obtain the flame smoke image sequence-CO concentration level dataset; The 3D convolutional neural network model is trained using flame smoke image sequence-CO concentration level data.

3. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 2 is characterized in that: The video of the combustion flame and flue gas is obtained by installing a high-temperature camera at the first flue of the incinerator; the CO concentration data is obtained by installing a high-temperature laser flue gas analyzer at the first flue.

4. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 2 is characterized in that: The preprocessing includes: extracting frames from the video and screening out abnormal and blurred images; and removing missing values ​​and erroneous values ​​in the CO concentration data.

5. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 2 is characterized in that: The preprocessed CO concentration data is divided into multiple CO concentration-time data segments, specifically comprising the following steps: setting a CO concentration threshold, traversing the CO concentration data in chronological order, and if a preset maximum length is reached and the segment does not have a CO concentration exceeding the threshold, recording the start and end time of the segment as a CO concentration segment; if the CO concentration exceeds the threshold at any moment during the traversal, recording the start time of the CO concentration segment as a specified length of time before the moment when the CO concentration exceeds the threshold, and recording the end time of the CO concentration segment as a specified length of time after the CO concentration returns to below the threshold; thereby obtaining multiple CO concentration-time data segments.

6. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 5 is characterized in that: The method for classifying CO concentration levels is as follows: calculating the CO concentration peak value or mean value of each CO concentration-time data segment, and classifying the CO concentration-time data segment into several levels according to the peak value or mean value.

7. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 2 is characterized in that: The matching of CO concentration-time data segments with image sequences and labeling of image sequences according to CO concentration levels specifically includes: determining the time delay between the measured CO concentration data and the image based on the flow time of the flue gas in the flue, matching the CO concentration-time data segments with the image based on the shooting time of the image, and labeling the image according to the CO concentration level, thereby obtaining a flame flue gas image sequence-CO concentration level dataset.

8. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 1 is characterized in that: The 3D convolutional neural network model is a spatiotemporal modeling deep learning model for video action recognition.

9. The method for online monitoring of CO concentration in an incinerator based on deep learning according to claim 1 is characterized in that: The 3D convolutional neural network model takes continuous image segments as input. When the number of frames of the original continuous image segments is N origin When it is greater than the preset input frame number N, determine the starting frame index f start , where 1≤f start ≤N origin -N+1, input N frames of continuous image segments; when the number of frames of the original continuous image segments input is N origin When the number of input frames is less than the preset number N, the last frame image is repeatedly input until the preset number N of input frames is reached.

10. A computer device, characterized in that: The computer device comprises: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the online monitoring method for CO concentration in an incinerator based on deep learning as described in any one of claims 1 to 9.

Images