An industrial gas concentration analysis system based on infrared light

Abstract

The present invention relates to the technical field of industrial gas analysis, and specifically discloses an infrared light-based industrial gas concentration analysis system, which is used to solve the problems of low wavelength drift control accuracy, poor disturbance adaptability, insufficient multi-gas regulation compatibility, and lack of an active feedback mechanism in existing industrial gas detection. The system includes an environmental parameter acquisition module and an infrared light gas analysis and control module, which collect temperature, humidity, and pressure disturbance information in real time and perform standardized processing. The system constructs human-machine collaborative embedding features by combining the physical characteristics of the three gases CO, N2O, and H2O, the state characteristics of the heating furnace, and the behavior characteristics of the laser control. The system realizes disturbance scenario modeling and adjustment strategy prediction through word vector and knowledge graph embedding fusion representation, thereby improving the accuracy, responsiveness, and system generalization capability of wavelength control, and is suitable for high-reliability wavelength control in multi-gas concentration collaborative detection scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial gas analysis, and more particularly to an industrial gas concentration analysis system based on infrared light. Background Art

[0002] Industrial gas concentration analysis is widely used in a variety of key areas, including safety monitoring, environmental management, and process control. Tunable Diode Laser Absorption Spectroscopy (TDLAS) has been widely used in various gas detection scenarios due to its high selectivity, high sensitivity, and rapid responsiveness. Existing industrial systems generally use mid-infrared quantum cascade lasers as core components for gas detection. By controlling the laser's temperature and current to adjust its emission wavelength, it covers the specific absorption lines of the gas being measured, thereby achieving high-precision inversion of gas concentration. However, actual industrial environments are often accompanied by complex dynamic interferences such as temperature, humidity, and pressure disturbances, and fluctuations in the start and stop of heating furnaces. These factors can cause nonlinear drift in the laser wavelength, resulting in photoelectric signal fluctuations and concentration inversion errors.

[0003] Although existing systems have introduced dynamic environmental compensation modules, which collect temperature, humidity, pressure, and other data through environmental sensing modules, perform normalization processing, construct environmental mapping terms and time integral terms, and correct signals based on compensation factors, these compensation methods are still passive response mechanisms and lack deep modeling of disturbance structures and semantic interpretation of behavioral patterns. Especially in scenarios where multiple gases are measured simultaneously, the control history is complex, and the operating behavior has temporal characteristics, traditional PID or decoupling control schemes face problems of insufficient control and weak generalization capabilities. They lack structured modeling and deep correlation analysis of environmental states, and fail to fully utilize the historical control behavior of the laser and the physical response laws of the gases to form a universal control experience model. Especially in the scenario of simultaneous multi-gas detection, different gases have different sensitivities to disturbances, making laser control strategies incompatible and resulting in poor response accuracy. In addition, the feedback paths of existing systems are mostly limited to one-way closed-loop regulation, lacking active intervention channels from engineering experts or user experience, and cannot form a highly responsive, high-precision, and high-generalization control system under complex operating conditions. Summary of the Invention

[0004] To overcome the above-mentioned shortcomings of the prior art, the present invention provides an infrared-based industrial gas concentration analysis system. By introducing a human-machine collaborative feedback control mechanism, integrating gas characteristics, heating furnace status and historical control behavior, and combining environmental disturbance modeling, it can achieve high-precision wavelength adjustment and concentration detection of CO, N2O, and H2O under dynamic working conditions.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An infrared-based industrial gas concentration analysis system enables laser light to pass through the gas being measured twice to enhance the signal strength of the three gases, CO, N2O, and H2O, while simultaneously detecting their concentrations. It also introduces dynamic environmental compensation correction signals. The system includes a transceiver and a reflector. The transceiver includes a laser, which is an infrared quantum cascade laser with an output wavelength range of 2202.8 to 2205.6 cm. -1 An off-axis parabolic mirror with a hole in the middle is provided on one side of the laser. A window piece 1 is provided on the side of the off-axis parabolic mirror away from the laser. The window piece 1 is used for the first emission and re-incidence of the infrared laser after reflection. A filter is provided above the off-axis parabolic mirror, a photoelectric detector is provided above the filter, a window piece 2 and a full-angle reflector are provided at the reflecting end, a red light generator is provided in the laser, and the red light generator is used for combining red light and infrared light as an emission path indicator. The photoelectric detector is connected to a data inversion and analysis module, the data inversion and analysis module is connected to a dynamic environment compensation module, and the dynamic environment compensation module is connected to an environmental parameter acquisition module. The module is connected to the infrared light gas analysis and control module. The environmental parameter acquisition module collects temperature, humidity and pressure within a range of 0.5 cm from the laser in real time and performs normalization processing to construct a normalized environmental state vector. The infrared light gas analysis and control module is connected to the laser. Based on the characteristics of the measured gas, the state characteristics of the heating furnace, and the characteristics of the laser control behavior, a collaborative feature representation for simultaneous detection of the measured gas concentration is constructed to predict the infrared light adjustment parameters output by the tunable laser based on artificial collaborative feedback. The measured gases include CO, N2O, and H2O.

[0007] As a further solution of the present invention, the infrared light gas analysis and control module includes a human-machine collaborative feature construction submodule, a human-machine collaborative feature embedding submodule, and an infrared light control parameter prediction module. The human-machine collaborative feature construction submodule is connected to the human-machine collaborative feature embedding submodule, the human-machine collaborative feature embedding submodule is connected to the infrared light control parameter prediction module, and the infrared light control parameter prediction module is connected to the environmental parameter acquisition module and the laser.

[0008] In order to extract the measured gas characteristics, heating furnace status characteristics, and laser control behavior characteristics from the existing data:

[0009] As a further solution of the present invention, in the human-machine collaborative feature construction submodule, the method for obtaining the characteristics of the measured gas is as follows: first, the label is initialized according to the physical properties of the three gases CO, N2O, and H2O and the external knowledge graph. After the laser enters the operation stage, the control response behavior label is extracted based on the historical wavelength offset performance and control response of the three gases under different environmental states. The control response behavior label is time-attenuated and weighted according to the occurrence time to obtain the weighted control response behavior label, which is used as the characteristic label of the measured gas. The external knowledge graph is a structured target detection gas physical response knowledge graph, including the semantic relationship between the target detection gas type, environmental state, laser control method and control behavior result. It is constructed in the form of a triple and is used to provide an initial label set summarized based on external experience and experimental data. The external knowledge graph covers the typical absorption bands, interference sensitivity, response laws under environmental disturbances and recommended control strategies of CO, N2O, and H2O gases. In the initialization stage, the external knowledge graph is called to assign a label vocabulary set to each gas, which is used as the characteristic label of the measured gas.

[0010] As a further solution of the present invention, in the human-machine collaborative feature construction submodule, the method of obtaining the state characteristics of the heating furnace is: using the structural parameters and control mode of the heating furnace to initialize the label list, constructing the initialization label, during the operation of the heating furnace, according to the dynamic changes of the temperature, pressure and airflow monitoring data in the furnace, identify the current operating behavior state of the heating furnace, construct a dynamic behavior label, and assign different weights to the operating behavior labels in different time periods according to the preset time attenuation function to form an operating behavior label set that reflects the current operating trend of the heating furnace, which is used as the heating furnace state feature label.

[0011] As a further solution of the present invention, the time decay function formula is:

[0012] ;

[0013] Where: To run the behavior tag index, For the The timestamp when the running behavior tag is identified, is the memory retention attenuation coefficient, For the current moment, For the The weight of a running behavior label.

[0014] As a further solution of the present invention, in the human-machine collaborative feature construction submodule, the method for obtaining the laser control behavior characteristics is: the control instruction log recorded during the operation of the laser is used as the human-machine collaborative behavior text set, each log includes the control action, adjustment parameters, and feedback response words of the laser, and by performing word segmentation and TF-TDF weight calculation on all words in the behavior text, the weighted key behavior labels with TF-TDF weights within the set threshold range are extracted to obtain a list of key behavior weighted labels for each log, which are used as the laser control behavior feature labels.

[0015] In order to embed the measured gas characteristics, heating furnace status characteristics, and laser control behavior characteristics as labels to obtain human-machine collaborative embedding features:

[0016] As a further solution of the present invention, in the human-machine collaborative feature embedding submodule, all labels are embedded using the pre-trained word vector model and then weighted summed to obtain the WE feature of the label group. ;

[0017] The TransE model is used to quantify each node of the external knowledge graph, and the objective function is:

[0018] ;

[0019] Where: is the embedded objective function value, 、 are respectively the positive sample set in the external knowledge graph and the randomly sampled negative sample set, 、 、 are the head entity, relation, and tail entity in the triple respectively. 、 They are used to generate the head entity and tail entity that are randomly replaced by negative samples, 、 、 They are KGE (label embedding based on external knowledge graph) representations of head entity, relationship, and tail entity respectively. 、 are the KGE representations of the head entity and tail entity of the negative sample, is the interval coefficient, where The measured gas feature labels, heating furnace state feature labels and laser control behavior feature labels are input into the TransE model and then weighted summed to obtain the KGE feature of the label group ;

[0020] The WE features and KGE features of the label group are spliced ​​together to obtain the human-machine collaborative embedding features:

[0021] ;

[0022] in, To embed features for human-machine collaboration, , Dimensions for embedding features for human-machine collaboration.

[0023] In order to combine the normalized environment state vector with the human-machine collaborative embedding representation:

[0024] As a further solution of the present invention, the infrared light control parameter prediction module includes a feature recall submodule and an infrared light adjustment amount sorting submodule.

[0025] As a further solution of the present invention, the feature recall submodule uses the current human-machine collaborative embedded features and the normalized environmental state vector to jointly construct the current disturbance scene vector, calculates the cosine similarity between the current disturbance scene vector and the historical disturbance scene vector in the training set, selects the disturbance scenes whose cosine similarity is within the set threshold range as the recall result, and extracts the corresponding laser output infrared light control response behavior as the candidate adjustment strategy set.

[0026] As a further solution of the present invention, the infrared light adjustment amount sorting submodule uses a regression model to score each group of historical adjustment strategies in the recall candidate adjustment strategy set. The input features include the current disturbance scene vector and the recall strategy features. The predicted score of each group of adjustment strategies under the current scene is output, and the predicted scores are arranged in descending order. The one with the highest score is selected as the target adjustment parameter of the current laser.

[0027] In order to solve the problem of no historical response data reference in the initial operation state of the laser:

[0028] As a further solution of the present invention, the infrared light control parameter prediction module also includes an initial parameter analysis submodule. During the laser's initial operation phase, the initial parameter analysis submodule constructs an initial input feature vector based on the embedded features of the normalized environmental state vector, the measured gas characteristic label, and the heating furnace state label. It then calls a pre-trained initial regression model to predict the laser's infrared light target control parameters. Initial laser wavelength control strategies for the three target gases (CO, N2O, and H2O) under current environmental parameters (temperature, humidity, and pressure) are first extracted from an external knowledge graph. These strategy labels are input into TransE embedding and then, combined with the current normalized environmental state vector, are spliced ​​into an initial input feature vector. The initial input feature vector is then inferred using a pre-trained regression model to output the initial laser target control parameters. This regression model is constructed based on existing experimental data before system deployment. Once the laser enters stable operation, the system gradually transitions to a tunable laser wavelength control method based on human-machine collaborative feedback, predicting the laser target control parameters based on the measured gas characteristics, the heating furnace operating status, and the laser's historical control behavior.

[0029] The technical effects of the infrared light-based industrial gas concentration analysis system of the present invention are as follows:

[0030] The present invention combines an infrared quantum cascade laser with a two-way absorption path to achieve simultaneous detection of CO, N2O, and H2O gases. By collecting temperature, humidity, and pressure disturbances around the laser, a normalized environmental state vector is constructed to achieve standardized modeling of the laser operating environment. A human-machine collaborative feedback control mechanism is introduced, which integrates the physical properties of the measured gas, the operating status of the heating furnace, and the historical control behavior of the laser to construct a human-machine collaborative feature representation. The infrared light adjustment parameters of the laser are further predicted based on the human-machine collaborative feature, achieving a priori identification of the laser wavelength offset trend and improving the active control capability of the laser output infrared light. In the joint measurement of CO, N2O, and H2O concentrations and the dynamic and frequent gas concentration detection in the heating furnace, the accuracy, response speed, and control stability of the laser wavelength adjustment are improved, the inversion error of the measured gas concentration is effectively reduced, and the robustness and intelligent self-adaptation capability of the joint measurement of CO, N2O, and H2O concentrations in a strong interference environment are enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A schematic diagram of the optical path of the prior art of the present invention;

[0032] Figure 2 is a system block diagram of the present invention;

[0033] In the figure: 01, laser; 02, off-axis parabolic mirror; 03, window 1; 04, filter; 05, photodetector; 06, window 2; 07, full-angle reflector; 101, transceiver; 102, reflector. DETAILED DESCRIPTION

[0034] 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 only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] Example 1

[0036] like Figure 1As shown, the present invention proposes an infrared light-based industrial gas concentration analysis system, the existing system established by the present invention can make the laser pass through the measured gas twice to enhance the signal intensity of the three gases, CO, N2O, and H2O, to detect the concentration simultaneously, and introduce a dynamic environment compensation correction signal, including a transceiver 101 and a reflector 102, the transceiver 101 is provided with a laser 01, the laser 01 is equipped with a red light generator, one side of the laser 01 is provided with an off-axis parabolic mirror 02 with a middle opening, a filter 04 is provided above the off-axis parabolic mirror 02, a photodetector 05 is provided above the filter 04, and the side of the off-axis parabolic mirror 02 away from the laser 01 is provided with Window 1 03, the reflective end 102 is equipped with a window 2 06, a full-angle reflector 07 is provided on one side of the window 2 06, the photoelectric detector 05 is connected to the data inversion analysis module, the data inversion analysis module is connected to the dynamic environment compensation module, the dynamic environment compensation module is connected to the environment sensing module, the dynamic environment compensation module uses the environmental interference data to construct the normalized vector and normalized deviation of the environmental variable, obtain the environmental parameter mapping item, weightedly accumulate the historical gas concentration inversion error within the set time window, obtain the environmental parameter time integral item, input the two into the dynamic environment compensation formula, and dynamically compensate for the environmental interference. The optical path between the transceiver 101 and the reflective end 102 includes:

[0037] The red indicator light emitted by the red light generator is combined with the infrared light emitted by the laser 01 to form an outgoing light beam, which first passes through the middle opening of the off-axis parabolic mirror 02 and the window 1 03, enters the gas area to be measured, and completes the first absorption; then the laser beam passes through the window 2 06 and is emitted to the full-angle reflector 07. After reflection along the original path, it passes through the window 2 06 again to enter the gas area to be measured and undergoes a second absorption. The reflected light beam passes through the window 1 03 and is incident on the parabola of the off-axis parabolic mirror 02 again. Its reflected light passes through the filter 04 and is focused onto the light-sensitive surface of the photodetector 05. The light signal collected by the photodetector 05 is converted into an electrical signal through photoelectric conversion, and then the data inversion analysis module performs detection electrical signal processing and gas concentration inversion calculation, performs dynamic environment correction based on the dynamic environment compensation algorithm, and feedback controls the laser 01.

[0038] It should be further explained that in order to further solve the impact of the current scene disturbance on the wavelength output of Laser 01, improve the accuracy of data inversion, and conduct structured modeling and deep correlation analysis of environmental states, make full use of the historical control behavior of Laser 01 and the physical response law of gases to form a universal control experience model. Especially in the scenario of simultaneous detection of multiple gases, different gases have different sensitivities to disturbances. The control strategy of Laser 01 is integrated and compatible, and the response accuracy is improved. Combined with the active intervention channel from engineering experts or user experience, a control system with high responsiveness, high precision and high generalization ability is formed under complex working conditions. Figure 2As shown, the improvement of the present invention is that the analysis module is connected to the dynamic environment compensation module, the dynamic environment compensation module is connected to the environmental parameter acquisition module and the infrared light gas analysis and control module, the environmental parameter acquisition module collects the temperature, humidity and pressure within 0.5 cm range from the laser 01 in real time and performs normalization processing to construct a normalized environmental state vector, the infrared light gas analysis and control module is connected to the laser 01, and based on the characteristics of the measured gas, the state characteristics of the heating furnace, and the control behavior characteristics of the laser 01, a collaborative feature representation for simultaneous detection of the measured gas concentration is constructed to predict the infrared light adjustment parameters output by the tunable laser 01 based on artificial collaborative feedback. The measured gases include CO, N2O, and H2O.

[0039] For example, laser 01 was first used in an industrial heating furnace with a temperature of 38°C, a humidity of 70%, and a pressure of 101.2kPa. The goal was to detect the N2O concentration. By calling the external knowledge graph, it was found that N2O was most sensitive to the regulation of the laser 01 wavelength range of 1280-1320nm in an environment with a humidity greater than 65%. After checking the historical data, it was found that the N2O response wavelength had a +1.3nm offset under this condition. The system used this offset behavior as a label for the wavelength to shift toward long waves, and assigned a weight to the set threshold (because it was close to the current state). The final output feature label was long-wave offset, high-humidity interference sensitivity, and medium absorption intensity. The heating furnace ran continuously for 60 minutes, the temperature rose from 400°C to 600°C, and the pressure and airflow rose steadily. The operating data was detected to identify the heating stage and continuous heating behavior. Multiple temperature rise labels were recorded in the past 40 minutes, but the weight was lower the earlier the time. The weight of the label 40 minutes ago was reduced to 0.13, and the weight of the label 10 minutes ago was 0.61. The output characteristic label of the current heating furnace status was a significant temperature rise trend and a high thermal disturbance risk warning. By inputting the control instruction "adjust the gain +1.5 and observe the wavelength drift", the control instruction text was recorded in the instruction log. After word segmentation, the keywords "adjustment", "gain", "wavelength", and "drift" were counted. Combined with the TF-TDF weight calculation, the weights of "gain" and "wavelength" were significant, and they were extracted as key behavior labels. The "gain + adjustment" behavior that frequently appeared in multiple consecutive instructions was further weighted. By extracting the high-frequency operation behavior in the control instruction, the human-machine control intention was quickly understood, providing a data basis for behavior prediction and adjustment optimization.

[0040] It should be noted that the infrared light gas analysis and control module includes a human-machine collaborative feature construction submodule, a human-machine collaborative feature embedding submodule, and an infrared light control parameter prediction module. The human-machine collaborative feature construction submodule is connected to the human-machine collaborative feature embedding submodule, the human-machine collaborative feature embedding submodule is connected to the infrared light control parameter prediction module, and the infrared light control parameter prediction module is connected to the environmental parameter acquisition module and laser 01.

[0041] In order to extract the measured gas characteristics, heating furnace status characteristics, and laser 01 control behavior characteristics from the existing data:

[0042] Specifically, the human-machine collaborative feature construction submodule obtains the characteristics of the measured gas. More specifically, the labels are first initialized according to the physical properties of the three gases CO, N2O, and H2O and the external knowledge graph. After the laser 01 enters the operation stage, the control response behavior labels are extracted based on the historical wavelength offset performance and control response of the three gases under different environmental states. The control response behavior labels are time-attenuated and weighted according to the occurrence time to obtain the weighted control response behavior labels, which are used as the characteristic labels of the measured gases. The external knowledge graph is a structured target detection gas physical response knowledge graph, which includes the semantic relationship between the target detection gas type, environmental state, laser 01 control method and control behavior results. It is constructed in the form of triples and is used to provide an initial label set based on external experience and experimental data. The external knowledge graph covers the typical absorption bands, interference sensitivity, response laws under environmental disturbances, and recommended control strategies of CO, N2O, and H2O gases. In the initialization stage, the external knowledge graph is called to assign a label vocabulary set to each gas, which is used as the characteristic label of the measured gas.

[0043] Specifically, the human-machine collaborative feature construction submodule obtains the state characteristics of the heating furnace. More specifically, the label list is initialized using the structural parameters and control mode of the heating furnace to construct the initialization label. During the operation of the heating furnace, the current operating behavior state of the heating furnace is identified based on the dynamic changes of the temperature, pressure, and airflow monitoring data in the furnace, and a dynamic behavior label is constructed. The operating behavior labels in different time periods are assigned different weights according to the preset time decay function to form an operating behavior label set that reflects the current operating trend of the heating furnace. This is used as the heating furnace state feature label. The time decay function formula is:

[0044] ;

[0045] Where: To run the behavior tag index, For the The timestamp when the running behavior tag is identified, is the memory retention attenuation coefficient, For the current moment, For the The weight of a running behavior label;

[0046] Specifically, the human-machine collaborative feature construction submodule obtains the control behavior characteristics of laser 01. More specifically, the control instruction log recorded during the operation of laser 01 is used as the human-machine collaborative behavior text set. Each log includes the control actions, adjustment parameters, and feedback response words of laser 01. By performing word segmentation and TF-TDF weight calculation on all words in the behavior text, the weighted key behavior labels with TF-TDF weights within the set threshold range are extracted to obtain a list of key behavior weighted labels for each log, which are used as the control behavior feature labels of laser 01.

[0047] In order to clearly illustrate the extraction of the above three features, an example is given below:

[0048] For example, a system deployed inside an industrial heating furnace starts laser 01 for the first time. The current detection object is a mixed gas of CO and N2O. The environmental parameters collected by the system are temperature of 35°C, humidity of 68%, and pressure of 100.8 kPa. The goal is to determine the wavelength control characteristics of laser 01 for CO and N2O in the initial stage to assist in subsequent wavelength adjustment. First, the structured target detection gas physical response knowledge graph is loaded, which contains three-dimensional group information: (CO, high humidity, negative wavelength offset), (N2O, high temperature, increased response bandwidth), (CO, high pressure, recommended strategy: fine step adjustment), (N2O, strong interference, recommended strategy: heating furnace heating). According to the gas types of CO and N2O, their typical labels under high humidity and high temperature conditions are extracted, such as the label vocabulary set assigned to CO ( The label vocabulary set assigned to N2O ("wide absorption range, covering the band 1285–1325 nm", "enhanced nonlinear response at high temperatures", and "recommended strategy: adapt to rising temperature conditions") was selected. A search of past N2O control records in environments above 35°C revealed a historical offset of +1.4 nm. A search of CO control behavior in high-humidity environments revealed a -0.6 nm drift. These control response behaviors were extracted and assigned time-attenuation weights, such as a weight of 0.92 for data within a week, 0.41 for data from a month ago, and 0.07 for data from six months ago. The initial labels were fused with the historical control behavior labels to output a weighted characteristic label set for the measured gas, as shown in Table 1.

[0049] Table 1 Weighted characteristic label set of the measured gas

[0050]

[0051] Even if the system is running for the first time and there is no current on-site feedback behavior, it can reasonably construct the current state characteristics of CO and N2O based on external experience (knowledge graph) and historical response records. The recall and sorting of subsequent adjustment strategies provide high-value feature inputs, improving the accuracy and stability of initial adjustment. The integration of knowledge labels and experience labels makes the system have scalable, explainable, and generalizable initial control capabilities.

[0052] By introducing a mechanism based on external knowledge graph initialization, historical behavior label extraction, time weighting and semantic text modeling in the human-machine collaborative construction submodule, the characteristics of the measured gas, the state characteristics of the heating furnace and the control behavior characteristics of the laser 01 are extracted respectively. The physical properties of the measured gas and the external knowledge graph can be used to construct the characteristic labels of the measured gas. In the initial stage where there is a lack of real-time data, the interference type and the wavelength response law of the laser 01 can be quickly identified with the help of known experience, forming a universal control experience model. Especially in the scenario of simultaneous detection of multiple gases, the sensitivity of different gases to disturbances is different, and the laser 01 is controlled by the laser 01. 1. The control strategy achieves integrated compatibility and improved response accuracy. Combined with the active intervention channel of external knowledge from engineering experts or user experience, a highly responsive, high-precision and high-generalization control system is formed under complex working conditions. By weighting the dynamic labels of the heating furnace's operating behavior with time decay, a behavioral trend representation is constructed to improve the perception and response capabilities of the complex dynamic changes of the heating furnace. The control instruction log is used as semantic text input, and the key control behaviors are extracted through TF-TDF weights. It is possible to extract significantly representative behavior labels from unstructured control behaviors to facilitate the understanding and extraction of control strategies.

[0053] In order to embed the measured gas characteristics, heating furnace status characteristics, and laser 01 control behavior characteristics as labels to obtain human-machine collaborative embedding features:

[0054] It should be noted that in the human-machine collaborative feature embedding submodule, all labels are embedded using the pre-trained word vector model and then weighted summed to obtain the WE feature of the label group. ;

[0055] The TransE model is used to quantify each node of the external knowledge graph, and the objective function is:

[0056] ;

[0057] Where: is the embedded objective function value, 、 are respectively the positive sample set in the external knowledge graph and the randomly sampled negative sample set, 、 、 are the head entity, relation, and tail entity in the triple respectively. 、 They are used to generate the head entity and tail entity that are randomly replaced by negative samples, 、 、 They are KGE (label embedding based on external knowledge graph) representations of head entity, relationship, and tail entity respectively. 、 are the KGE representations of the head entity and tail entity of the negative sample, is the interval coefficient, where The measured gas feature label, heating furnace state feature label and laser 01 control behavior feature label are input into the TransE model and then weighted summed to obtain the KGE feature of the label group ;

[0058] The WE features and KGE features of the label group are spliced ​​together to obtain the human-machine collaborative embedding features:

[0059] ;

[0060] in, To embed features for human-machine collaboration, , Dimensions for embedding features for human-machine collaboration.

[0061] For example, when laser 01 is currently running, the status detected by the heating furnace is the current temperature of 620°C, the current pressure of 0.8 MPa, the current airflow is steadily rising, the continuous operation time of the heating furnace is 70 minutes, and the status labels in the last 30 minutes (temperature increase, high pressure fluctuation, thermal flow disturbance). In the human-machine collaborative feature construction submodule, the weights are calculated based on the acquired monitoring data + time decay function, and the following labels are extracted and quantified as shown in Table 2:

[0062] Table 2 Heating furnace status label display table

[0063]

[0064] Use the pre-trained word vector model to encode the above label words, and the warming state is , and similarly, the thermal disturbance trend is , pressure instability , sum each word vector with the weight to get the WE feature of the heating furnace , then read the triples (heating furnace, state is, heating), (heating, leads to, thermal disturbance), (thermal disturbance, associated, wavelength drift), (high pressure, recommended strategy, steady-state regulation) from the structured external knowledge graph, and map the entities and relationships in these triples to the corresponding TransE embedding model 、 、 , calculated by the trained TransE embedding model , get the positive sample scores of all triplets, randomly sample to generate negative samples, and optimize the total loss function , output the corresponding entity vector embedding, and obtain the weighted sum of the representations of these labels in the TransE embedding space Similarly, the embedded feature representations of the historical control behavior of the measured gas laser 01 are 、 ,final , , and then splice to get the human-machine collaborative embedding feature vector , this vector is the input core feature of the subsequent recall and adjustment strategy ranking model.

[0065] In order to combine the normalized environment state vector with the human-machine collaborative embedding representation:

[0066] It should be noted that the infrared light control parameter prediction module includes a feature recall submodule and an infrared light adjustment amount sorting submodule.

[0067] It should be further explained that the feature recall submodule uses the current human-machine collaborative embedding features and the normalized environment state vector to jointly construct the current disturbance scene vector, calculates the cosine similarity between the current disturbance scene vector and the historical disturbance scene vector in the training set, selects the disturbance scenes whose cosine similarity is within the set threshold range as the recall result, and extracts the corresponding laser 01 output wavelength control response behavior as the candidate adjustment strategy set.

[0068] It should be further explained that the infrared light adjustment amount sorting submodule uses a regression model to score each group of historical adjustment strategies in the recall candidate adjustment strategy set. The input features include the current disturbance scene vector and the recall strategy features. The predicted score of each group of adjustment strategies in the current scene is output and sorted in descending order according to the predicted score. The one with the highest score is selected as the target adjustment parameter of the current laser 01.

[0069] The current system not only considers the control behavior and knowledge labels of laser 01, but also synchronously integrates environmental conditions (temperature, humidity, and pressure) to construct a high-dimensional disturbance scenario vector. This scenario vector combines environmental characteristics with semantic embedding features, and can effectively express the combined impact of current external disturbances and device status. By matching the cosine similarity with historical disturbance scenarios, successful strategies in highly similar scenarios can be recalled from the "experience library," making up for the defect that traditional adjustment algorithms only make local optimal predictions based on current data, enabling the system to have scene recognition and memory capabilities.

[0070] Example 2

[0071] The difference between Example 2 of the present invention and Example 1 is that this example introduces the use of the tunable laser 01 wavelength control system based on human-machine collaborative feedback when there is no historical control behavior data of the laser 01 for the first time.

[0072] In order to address the problem of no historical response data reference in the initial operation state of laser 01, the infrared light control parameter prediction module also includes an initial parameter analysis submodule. During the first operation phase of laser 01, the initial parameter analysis submodule constructs an initial input feature vector based on the normalized environmental state vector, the measured gas feature label, and the embedded features of the heating furnace state label, and calls the pre-trained initial regression model to predict the target adjustment parameters of laser 01. First, the initial laser 01 parameters of the three target gases (CO, N2O, H2O) under the current environmental parameters (temperature, humidity, pressure) are extracted from the external knowledge graph. 1. Wavelength control strategy. After these strategy labels are input into TransE embedding, they are combined with the current normalized environmental state vector to form an initial input feature vector. The pre-trained regression model is used to infer the initial input feature vector and output the initial target adjustment parameters of Laser 01. The regression model is built based on existing experimental data before system deployment. When Laser 01 enters stable operation, the system gradually transitions to a tunable Laser 01 wavelength control method based on human-machine collaborative feedback. The target adjustment parameters of Laser 01 are predicted based on the measured gas characteristics, the operating status of the heating furnace, and the historical adjustment behavior of Laser 01.

[0073] Specifically, in an industrial heat treatment scenario, the system was deployed in an industrial furnace emissions monitoring device, initially used to detect the concentrations of two target gases, CO and N2O, in a high-temperature environment. During the initial startup of laser 01, the normalized environmental state vector collected by the environmental parameter acquisition module was [0.92, 0.30, 0.88], corresponding to a furnace temperature of 480°C, a humidity of 35%, and a pressure of 0.95 MPa. The target gases to be detected were CO and N2O, and the furnace was operating in the initial constant temperature stage.

[0074] The initial parameter analysis submodule first calls upon the structured target gas physical response knowledge graph to extract label triples that match the current gas and environmental conditions. These include information such as "CO wavelength fine-tuning is recommended under high temperature and high pressure conditions, with the adjustment step size halved," "N2O response band shifts upward in a dry atmosphere," and "Moderate thermal interference risk exists in the initial constant temperature stage." Based on this information, the system generates a set of labels, including "high temperature and high pressure fine-tuning," "band shift upward," "moderate interference," and "adjustment step size halved."

[0075] The human-machine collaborative feature embedding submodule then inputs the aforementioned labels into a pre-trained word embedding model and a knowledge graph embedding model. The word embedding model extracts the semantic embeddings of the labels, forming a weighted sum to produce a WE feature vector. The knowledge graph embedding model (using the TransE method) quantifies the relevance of the labels in structured triples, producing a weighted KGE feature vector. The normalized environment state vector is concatenated with the two label embedding vectors to construct the initial input feature vector.

[0076] The initial input feature vector is fed into an initial regression model for inference. This model, trained on extensive historical industrial experimental data collected before system deployment, is capable of predicting the initial adjustment strategy. Ultimately, the system outputs the initial target wavelength adjustment parameters for Laser 01 under the current environmental conditions, with a predicted value of 1573.6 nm.

[0077] This adjustment parameter is used to set the wavelength during the initial operation of Laser 01. It automatically generates a relatively accurate adjustment strategy based on the target gas characteristics, furnace status, and environmental disturbance conditions, without the involvement of historical behavioral data. As feedback data from Laser 01's operation gradually accumulates, the system can smoothly transition to a wavelength control mode based on human-machine collaborative feedback, feature recall, and strategy sorting mechanisms. This implementation effectively achieves the rapid cold start capability of Laser 01, improves the stability and prediction accuracy of the system's initial operation, and has good engineering adaptability and scalability.

[0078] In summary, in combination with Example 1 and Example 2, the present invention combines an infrared quantum cascade laser 01 with a double-pass absorption path to achieve simultaneous detection of CO, N2O, and H2O gases. By collecting the temperature, humidity, and pressure disturbances around the laser 01, a normalized environmental state vector is constructed to achieve standardized modeling of the operating environment of the laser 01. A human-machine collaborative feedback control mechanism is introduced to integrate the physical properties of the measured gas, the operating status of the heating furnace, and the historical control behavior of the laser 01 to construct a human-machine collaborative feature representation. The wavelength adjustment parameters of the laser 01 are further predicted based on the human-machine collaborative feature to achieve a priori identification of the wavelength deviation trend of the laser 01, thereby improving the active control capability of the output wavelength of the laser 01. In the co-measurement of CO, N2O, and H2O concentrations and the dynamic and frequent gas concentration detection in the heating furnace, the accuracy, response speed, and control stability of the wavelength adjustment of the laser 01 are improved, the inversion error of the measured gas concentration is effectively reduced, and the robustness and intelligent self-adaptation capability of the co-measurement of CO, N2O, and H2O concentrations in a strong interference environment are enhanced.

[0079] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0080] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An industrial gas concentration analysis system based on infrared light, comprising a transceiver (101) and a reflector (102), wherein the transceiver (101) comprises a laser (01), a photoelectric detector (05) on one side of the laser (01), the photoelectric detector (05) being connected to a data inversion analysis module, the data inversion analysis module being connected to a dynamic environment compensation module, and characterized in that: The dynamic environment compensation module is connected to the environment parameter acquisition module, the environment parameter acquisition module is connected to the infrared light gas analysis and control module, the infrared light gas analysis and control module is connected to the laser, the environment parameter acquisition module collects the temperature, humidity and pressure within a range of 0.5 cm from the laser in real time, and performs normalization processing to construct a normalized environment state vector, the infrared light gas analysis and control module constructs a collaborative feature representation for simultaneous detection of the concentrations of three measured gases, CO, N2O and H2O, based on the measured gas characteristics, the heating furnace state characteristics and the laser control behavior characteristics, so as to predict the infrared light adjustment parameters output by the tunable laser (01) based on human-machine collaborative feedback; The infrared light gas analysis and control module includes a human-machine collaborative feature construction submodule, a human-machine collaborative feature embedding submodule, and an infrared light control parameter prediction module. The human-machine collaborative feature construction submodule is connected to the human-machine collaborative feature embedding submodule, the human-machine collaborative feature embedding submodule is connected to the infrared light control parameter prediction module, and the infrared light control parameter prediction module is connected to the environmental parameter acquisition module and the laser (01); In the human-machine collaborative feature construction submodule, the method for obtaining the characteristics of the measured gas is as follows: first, the label is initialized according to the physical properties of the three gases CO, N2O, and H2O and the external knowledge graph. After the laser (01) enters the operation stage, the control response behavior label is extracted according to the historical wavelength offset performance and control response of the three gases under different environmental conditions, and the control response behavior label is weighted by time decay according to the occurrence time to obtain the weighted control response behavior label, which is used as the characteristic label of the measured gas. The external knowledge graph is a structured target detection gas physical response knowledge graph, including the semantic relationship between the target detection gas type, environmental state, laser (01) control method and control behavior result. The construction form is a triple, which is used to provide an initial label set based on external experience and experimental data. The external knowledge graph covers the typical absorption bands, interference sensitivity, response rules under environmental disturbances and recommended control strategies of CO, N2O, and H2O gases. In the initialization stage, the external knowledge graph is called to assign a label vocabulary set to each gas and use it as the characteristic label of the measured gas. In the human-machine collaborative feature construction submodule, the method for obtaining the state characteristics of the heating furnace is as follows: the label list is initialized using the structural parameters and control mode of the heating furnace to construct the initialization label; during the operation of the heating furnace, the current operating behavior state of the heating furnace is identified based on the dynamic changes of the temperature, pressure and airflow monitoring data in the furnace, and a dynamic behavior label is constructed; the operating behavior labels in different time periods are assigned different weights according to a preset time decay function, forming an operating behavior label set reflecting the current operating trend of the heating furnace, which is used as the heating furnace state feature label; the heating furnace state feature label is one or more of a heating state, a thermal disturbance trend or an unstable pressure; In the human-machine collaborative feature construction submodule, the method of obtaining the laser control behavior feature is as follows: the control instruction log recorded during the operation of the laser (01) is used as the human-machine collaborative behavior text set, each log includes the control action, adjustment parameters, and feedback response words of the laser (01), and by performing word segmentation and TF-TDF weight calculation on all words in the behavior text, the weighted key behavior labels with TF-TDF weights within the set threshold range are extracted to obtain the key behavior weighted label list of each log, which is used as the laser control behavior feature label.

2. The infrared light-based industrial gas concentration analysis system according to claim 1, characterized in that: In the human-machine collaborative feature embedding submodule, all labels are embedded using the pre-trained word vector model and then weighted summed to obtain the WE feature of the label group. ; The TransE model is used to quantify each node of the external knowledge graph, and the objective function is: ； Where: is the embedded objective function value, 、 are respectively the positive sample set in the external knowledge graph and the randomly sampled negative sample set, 、 、 are the head entity, relation, and tail entity in the triple respectively. 、 They are used to generate the head entity and tail entity that are randomly replaced by negative samples, 、 、 They are KGE representations of head entity, relationship, and tail entity respectively, and the labels are embedded based on the external knowledge graph. 、 are the KGE representations of the head entity and tail entity of the negative sample, is the interval coefficient, where The measured gas feature labels, heating furnace state feature labels and laser control behavior feature labels are input into the TransE model and then weighted summed to obtain the KGE feature of the label group ; The WE features and KGE features of the label group are spliced ​​together to obtain the human-machine collaborative embedding features: ； in, To embed features for human-machine collaboration, , Dimensions for embedding features for human-machine collaboration.

3. The infrared light-based industrial gas concentration analysis system according to claim 2, characterized in that: The infrared light control parameter prediction module includes a feature recall submodule and an infrared light adjustment amount sorting submodule.

4. The infrared light-based industrial gas concentration analysis system according to claim 3, characterized in that: The feature recall submodule uses the current human-machine collaborative embedding feature and the normalized environment state vector to jointly construct the current disturbance scene vector, calculates the cosine similarity between the current disturbance scene vector and the historical disturbance scene vector in the training set, selects the disturbance scene whose cosine similarity is within the set threshold range as the recall result, and extracts the corresponding laser (01) output infrared light control response behavior as the candidate adjustment strategy set.

5. The infrared light-based industrial gas concentration analysis system according to claim 4, characterized in that: The infrared light adjustment amount sorting submodule uses a regression model to score each group of historical adjustment strategies in the recall candidate adjustment strategy set. The input features include the current disturbance scene vector and the recall strategy features. The predicted score of each group of adjustment strategies in the current scene is output and sorted in descending order according to the predicted score. The one with the highest score is selected as the target adjustment parameter of the current laser (01).

6. The infrared light-based industrial gas concentration analysis system according to claim 5, characterized in that: The infrared light control parameter prediction module also includes an initial parameter analysis submodule. During the first operation phase of the laser (01), the initial parameter analysis submodule constructs an initial input feature vector based on the normalized environmental state vector, the measured gas feature label, and the embedded features of the heating furnace state label, and calls a pre-trained initial regression model to predict the infrared light target control parameters of the laser (01).

Images