Roller kiln temperature correction method based on infrared temperature compensation hybrid model

Through the infrared temperature compensation mixing model and the RNN-DNN tandem neural network, the problem of inaccurate infrared temperature detection of roller kilns is solved, and the accurate detection of the internal temperature field of roller kilns is achieved, and the temperature calibration accuracy and product quality are improved.

CN118936646APending Publication Date: 2024-11-12CENT SOUTH UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410998703.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In the prior art, the infrared temperature detection of roller kilns is not accurate enough to fully reflect the internal temperature field of the kiln. Thermocouple is prone to aging, resulting in a decrease in temperature detection accuracy, affecting the product firing quality.

Method used

Using the calibration method based on the infrared temperature compensation mixing model, the infrared temperature temperature measurement compensation mixing model and the RNN-DNN series neural network are combined with production data and infrared image features to establish a data-driven temperature calibration model to overcome high-temperature radiation interference and lens attenuation. , realize accurate detection of infrared temperature field.

Benefits of technology

Accurate calibration of the infrared temperature of the roller kiln is achieved, high-temperature radiation interference and thermal detector temperature drift is overcome, temperature detection accuracy is improved, and product firing quality and energy consumption efficiency are ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118936646A_ABST
    Figure CN118936646A_ABST
Patent Text Reader

Abstract

The invention discloses a roller kiln temperature correction method based on an infrared temperature compensation hybrid model, and belongs to the technical field of roller kilns. According to the roller kiln temperature correction method based on the infrared temperature compensation hybrid model, an infrared temperature measurement compensation hybrid model is used for calibration, and the formula of the infrared temperature measurement compensation hybrid model is # imgabs0 #, t0, T0'and Th respectively represent the real surface temperature of the measured object, the temperature measured by the thermal infrared imager under interference and the temperature of an adjacent high-temperature object, Fho represents the equivalent angle coefficient of the corresponding measured object to the high-temperature object, and fRNN-DNN is the output of the data driving model RNN-DNN. According to the correction method provided by the invention, the accurate calibration of the infrared temperature is realized, and the accurate detection of the infrared temperature field is further realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of roller kilns, and in particular to a roller kiln temperature correction method based on an infrared temperature compensation hybrid model. Background Art

[0002] The roller kiln is a key place for sintering battery positive electrode materials. Its internal temperature distribution directly affects the sintering quality and energy efficiency of the product. Due to the characteristics of high temperature, high airtightness and complex structure of the roller kiln, it is difficult to accurately control the temperature to meet the uniformity required by the process. Temperature field information is key information for reaction operating conditions, energy consumption levels, production quality and other indicators, and the control basis is to accurately measure and build the temperature field. How to accurately measure the temperature in the roller kiln has become a problem that needs to be solved urgently.

[0003] At present, the temperature is mainly detected by thermocouples inside the kiln in industrial production. However, due to the distribution limitations of the temperature measurement characteristics of thermocouples according to their installation positions, the temperature data only includes the temperature at the end of the thermocouple. When the temperature in the heating zone changes drastically, it cannot fully reflect the temperature field inside the kiln, and it has the characteristic of time lag. This shows that the temperature obtained by single-point temperature measurement has certain limitations. At the same time, since thermocouples are prone to aging, the temperature detection accuracy will gradually decrease, which will affect the firing quality of the product. How to accurately detect the temperature distribution inside the roller kiln is an urgent problem to be solved.

[0004] The existing literature 1 [Luo A, Liu B. Temperature prediction of roller kiln based on mechanism and SSA-ELM data-driven integrated model [C] / / 2022 China Automation Congress (CAC). IEEE, 2022: 5598-5603.] proposed a comprehensive modeling method for roller kiln temperature prediction. In the structure of the literature, firstly, based on the heat transfer mechanism and thermal balance theory, a mechanism model of roller kiln sagger temperature was established; then, in view of the estimation error caused by industrial process complexity and industrial noise, a data-driven error compensation model based on real-time process operation data was constructed. In view of the large lag in roller kiln temperature change, high correlation of process variables, and strong nonlinearity, an error compensation model of the extreme learning machine SSA-ELM optimized by the sparrow search algorithm was established. Then, the mechanism model and the error compensation model were combined to obtain a comprehensive temperature prediction model.

[0005] Reference 2 [Li Ailian, Cao Mengzhen, Yue Wei, et al. Design of roller kiln temperature measurement system based on colorimetry [J]. Control Engineering, 2017, 24(01): 215-221.] proposed a roller kiln temperature measurement system based on colorimetry, which can accurately establish the temperature field inside the kiln. According to the principle of colorimetric temperature measurement and combined with image preprocessing method, a temperature measurement system based on CCD (charge coupled device) image sensor was designed. The system can monitor and measure the internal temperature of the roller kiln and obtain the distribution of the global temperature field. According to the standard value method, the system was calibrated using a blackbody furnace to improve the temperature detection precision and accuracy of the system.

[0006] The roller kiln temperature prediction method proposed in document 1 is a mechanism model established based on the heat transfer mechanism. It is to obtain the temperature distribution of the object by modeling the heat conduction, convection, radiation and other processes on the surface and inside of the object. This mechanism modeling requires certain assumptions about the model boundary conditions, is highly complex, requires many parameters to be determined, and does not use sensing equipment. The effect of the model may not match the actual situation. Document 2 uses optics to obtain internal images, but does not consider the closed characteristics of the roller kiln and the interference of the radiation intensity of high-temperature objects in the process of mechanism modeling. It only performs simple calibration. However, there is currently a lack of a method to establish a temperature compensation model for the roller kiln from the perspective of infrared mechanism and thus realize accurate calibration of the infrared temperature of the roller kiln. Summary of the invention

[0007] The purpose of the present invention is to overcome the above technical deficiencies, provide a roller kiln temperature correction method based on an infrared temperature compensation hybrid model, and solve the technical problem that the infrared temperature of the roller kiln detected in the prior art is not accurate enough.

[0008] In order to achieve the above technical purpose, the technical solution of the present invention provides a method for correcting the temperature of a roller kiln based on an infrared temperature compensation hybrid model, which is calibrated by an infrared temperature measurement compensation hybrid model. The infrared temperature measurement compensation hybrid model formula is:

[0009]

[0010] In the formula, T0, T0′ and T h They represent the true surface temperature of the object being measured, the temperature measured by the infrared thermal imager under interference, and the temperature of the adjacent high-temperature object. The length of the endoscopic long lens is d, the corresponding equivalent refractive index of the lens is τ, and the monochromatic attenuation coefficient is k λ , n is the fitting coefficient in Planck integral; ε0 is the emissivity of the object being measured, F h_o Indicates the equivalent angular coefficient of the object being measured to the high-temperature object, f RNN-DNN It is the output of the data-driven model RNN-DNN network.

[0011] In any embodiment, when the temperature measurement object is the sagger area, the equivalent angular coefficient F of the sagger area is h_o for:

[0012]

[0013] In the formula, is the angular coefficient of the ith silicon carbon rod to the sagger plane where the upper surface of the sagger is located; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0014] In any embodiment, the The calculation formula is:

[0015]

[0016] In the formula, is the angular coefficient of the i-th silicon carbon rod to the yellow plane on the upper surface of the sagger; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0017] In any embodiment, when the temperature measurement object is the second left thermocouple, the corresponding angle coefficient is:

[0018] Wherein, L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, and θ1 is the angle between the connecting line corresponding to the infinitesimal area in surface 1 and surface 1.

[0019] In any embodiment, when the temperature measurement object is the upper thermocouple, the corresponding angle coefficient is:

[0020] Among them, L1 is the effective length of the silicon carbon rod in the roller kiln area; θ1 is the angle between the connecting line corresponding to the microelement area in surface 1 and surface 1, and R is the distance between the two microelement surfaces.

[0021] In any embodiment, the data-driven model is composed of a RNN-DNN series network, the input of the RNN network is composed of production data and infrared temperature measurement data of silicon carbon rods, and the input of the DNN network is composed of the output of the RNN network and infrared image features. Then, the RNN and DNN form a series network structure, and the outputs of the RNN and DNN networks together constitute the f RNN-DNN ;

[0022] The implementation steps of the RNN-DNN tandem neural network are as follows:

[0023] S1: Define the network and set network hyperparameters; set corresponding parameters including time step, network parameters and weighting factors, and initialize the model;

[0024] S2: Input normalization; the exhaust temperature, rear temperature measurement, upper SCR power, lower SCR power, meter active power and meter power production data in the production data; the average temperature value, skewness, kurtosis and infrared temperature measurement data of silicon carbon rods in the image features are used as network input and the unmodeled dynamic target data corresponding to the silicon carbon rod object are normalized;

[0025] S3: Load data; randomly select input sequences and target sequences of the same length as the time step in the temperature data of any pixel point, obtain the hidden layer vector of the RNN network, input all input sequence data into the RNN network to obtain the output value of the RNN network, normalize the output value of the RNN network and the image features as the input of the DNN network, and obtain the output of the DNN network;

[0026] S4: Calculate loss and back propagate; calculate the loss value of the RNN network and the DNN network, thereby obtaining the loss value of the entire network, and back propagate the gradient to update the network parameters;

[0027] S5: Verify the model with the validation set and save the network parameters; select the upper thermocouple data as the validation set, bring it into the network to obtain the network output and bring it into the calibration formula to obtain the calibration temperature, calculate the maximum error, and save the network parameters;

[0028] S6: Train the network; repeat S3 to S5. When the number of iterations is within the set range, the network parameters with the smallest maximum error of the validation set are taken as the optimal network parameters.

[0029] S7: Model verification; select the second thermocouple data from the left as the test set, input it into the network to obtain the network output and input it into the calibration formula to obtain the calibration temperature, draw the calibration curve and error curve, and verify the accuracy of the network.

[0030] In any embodiment, in step S1, the structure of the RNN network consists of an input layer, a hidden layer and an output layer. The network receives an input x at time t. t , after hidden layer s t , the output value is o t ; According to the network calculation formula:

[0031] o t =g(Vs t )

[0032] s t =f(Ux t +Ws t-1 )

[0033] Where V, U and W are weight matrices, g(g) and f(g) are activation functions;

[0034] o t =g(Vs t ) Repeatedly insert formula s t =f(Ux t +Ws t-1 ), we can get

[0035] o t =g(Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +...)))).

[0036] In any embodiment, in step S2, the input of the image features is to perform feature selection on the infrared image in terms of statistical features and texture features extracted;

[0037] Statistical features include: average temperature value

[0038] Where n p is the number of infrared image pixels, T i is the surface source temperature of the infrared image;

[0039] Average Deviation Standard Deviation Skewness Kurtosis

[0040] The texture features are extracted from the spatial temperature co-occurrence matrix TLCM;

[0041] For two pixels (j, k) and (p, q) with temperature values ​​u and v, the angle between them is θ, then the number of times they appear in any direction of the image can be expressed as:

[0042] P(u,v,d,θ)=number{T(j,k)=u,T(p,q)=v,θ}

[0043] Where P is TLCM; parameter d is usually 1, indicating that the distance between two pixels is 1;

[0044] The extracted texture features include: energy

[0045] Where P(u,v) represents the element at temperature u and v in TLCM, G represents the temperature level; entropy Moment of inertia Relevance Where μ x , μy , σ x and σ y is defined as follows: Anti-differentiation According to the definition of TLCM, the above texture features can be extracted from different directions. In order to reduce the fluctuation caused by the single-direction feature extraction, in order to make full use of the features in the four directions of 0°, 45°, 90° and 135°, the average value is often taken as the extracted stable feature:

[0046] In the formula, f i is the eigenvalue of a feature in the i-th calculation direction, is the average value of a feature in four directions.

[0047] In any implementation, in step S4, the calculation function of the loss value of the RNN network is: loss RNN =mean{(y RNN -y) 2};

[0048] where y RNN is the predicted output of the RNN network, T0, T0′ and T h are the actual temperature of the temperature measurement object, the infrared measurement temperature and the set temperature of the silicon carbon rod, τ, k λ and n are the equivalent refractive index of the endoscope long lens, the monochromatic attenuation coefficient of the internal atmosphere and the fitting coefficient, respectively. h_o is the equivalent angular coefficient of the high temperature object to the object being measured, and ε0 is the emissivity of the object being measured.

[0049] In any implementation, in step S4, the calculation function of the loss value of the DNN network is: loss DNN =mean{[y DNN -(yy RNN )] 2}; where y RNN and DNN are the predicted outputs of the RNN network and the DNN network respectively; and / or, in step S4, the calculation function of the loss value of the entire network is: loss RNN-DNN = a·loss RNN +b·loss DNN ;

[0050] In the formula, loss RNN-DNN 、loss RNN and loss DNNThey are the loss function values ​​of the RNN-DNN network, the RNN network, and the DNN network respectively; the weighting factors of the calculation functions of the loss values ​​of RNN and DNN are a and b respectively.

[0051] Compared with the prior art, the beneficial effects of the present invention include: overcoming high temperature radiation interference, lens attenuation and thermal detector temperature drift by the infrared temperature measurement compensation mechanism model, establishing a hybrid model of data-driven mechanism modeling, thereby achieving accurate calibration of infrared temperature, and further achieving accurate detection of infrared temperature field. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a structural block diagram of a roller kiln temperature correction method based on an infrared temperature compensation hybrid model proposed in Example 1 of the present invention.

[0053] Figure 2 It is a diagram showing the internal structure of a roller kiln according to Embodiment 1 of the present invention.

[0054] Figure 3 It is a longitudinal cross-sectional view of a roller kiln according to embodiment 1 of the present invention.

[0055] Figure 4 Schematic diagram of the position of the silicon carbon rod in Example 1 of the present invention; Figure 4 (a) is the position diagram of the silicon carbon rod relative to the second thermocouple on the left. Figure 4 (b) is the position diagram of the silicon carbon rod relative to the upper thermocouple.

[0056] Figure 5 It is a model structure block diagram of the RNN-DNN tandem neural network of Example 1 of the present invention.

[0057] Figure 6 This is a partial RBG infrared image of the box furnace collected by the infrared thermal imager of Example 1 of the present invention from September 25, 2023 to September 26, 2023.

[0058] Figure 7 It is a model structure diagram of the RNN network of Example 1 of the present invention.

[0059] Figure 8 It is the loss function decline curve of the RNN of Example 1 of the present invention.

[0060] Fig. 9 It is the loss function decline curve of the DNN of Example 1 of the present invention.

[0061] Fig.10 This is the verification result of the mechanism model driven by the RNN-DNN network of Example 1 of the present invention. DETAILED DESCRIPTION

[0062] This specific embodiment provides a roller kiln temperature correction method based on an infrared temperature compensation hybrid model, which is calibrated by an infrared temperature measurement compensation mechanism model, and the infrared temperature measurement compensation mechanism model formula is:

[0063]

[0064] In the formula, T0, T0′ and T h They represent the true surface temperature of the object being measured, the temperature measured by the infrared thermal imager under interference, and the temperature of the adjacent high-temperature object. The length of the endoscopic long lens is d, the corresponding equivalent refractive index of the lens is τ, and the monochromatic attenuation coefficient is k λ , n is the fitting coefficient in Planck integral; ε0 is the emissivity of the object being measured, F h_o Indicates the equivalent angular coefficient of the object being measured to the high-temperature object, f RNN-DNN It is the output of the data-driven model RNN-DNN network.

[0065] In some embodiments, when the temperature measurement object is selected as the sagger area, the equivalent angular coefficient F of the sagger area is h_o for:

[0066] In the formula, is the angular coefficient of the ith silicon carbon rod to the sagger plane where the upper surface of the sagger is located; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0067] In some embodiments, The calculation formula is:

[0068]

[0069] In the formula, is the angular coefficient of the i-th silicon carbon rod to the yellow plane on the upper surface of the sagger; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0070] In some embodiments, when the temperature measurement object selects the sagger area as the second left thermocouple, the corresponding angular coefficient is:

[0071] Wherein, L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, and θ1 is the angle between the connecting line corresponding to the infinitesimal area in surface 1 and surface 1.

[0072] In some embodiments, when the temperature measurement object selects the sagger area as the upper thermocouple, the corresponding angular coefficient is:

[0073] Among them, L1 is the effective length of the silicon carbon rod in the roller kiln area; θ1 is the angle between the connecting line corresponding to the microelement area in surface 1 and surface 1, and R is the distance between the two microelement surfaces.

[0074] In some embodiments, the data-driven model is composed of a RNN-DNN series network, the input of the RNN network is composed of production data and infrared temperature measurement data of silicon carbon rods, and the input of the DNN network is composed of the output of the RNN network and infrared image features. Then, the RNN and the DNN form a series network structure, and the outputs of the RNN and the DNN networks together constitute the f RNN-DNN ;

[0075] The implementation steps of the RNN-DNN tandem neural network are as follows:

[0076] S1: Define the network and set network hyperparameters; set corresponding parameters including time step, network parameters and weighting factors, and initialize the model;

[0077] S2: Input normalization; the exhaust temperature, rear temperature measurement, upper SCR power, lower SCR power, meter active power and meter power production data in the production data; the average temperature value, skewness, kurtosis and infrared temperature measurement data of silicon carbon rods in the image features are used as network input and the unmodeled dynamic target data corresponding to the silicon carbon rod object are normalized;

[0078] S3: Load data; randomly select input sequences and target sequences of the same length as the time step in the temperature data of any pixel point, obtain the hidden layer vector of the RNN network, input all input sequence data into the RNN network to obtain the output value of the RNN network, normalize the output value of the RNN network and the image features as the input of the DNN network, and obtain the output of the DNN network;

[0079] S4: Calculate loss and back propagate; calculate the loss value of the RNN network and the DNN network, thereby obtaining the loss value of the entire network, and back propagate the gradient to update the network parameters;

[0080] S5: Verify the model with the validation set and save the network parameters; select the upper thermocouple data as the validation set, bring it into the network to obtain the network output and bring it into the calibration formula to obtain the calibration temperature, calculate the maximum error, and save the network parameters;

[0081] S6: Train the network; repeat S3 to S5. When the number of iterations is within the set range, the network parameters with the smallest maximum error of the validation set are taken as the optimal network parameters.

[0082] S7: Model verification; select the second thermocouple data from the left as the test set, input it into the network to obtain the network output and input it into the calibration formula to obtain the calibration temperature, draw the calibration curve and error curve, and verify the accuracy of the network.

[0083] In some embodiments, in step S1, the structure of the RNN network consists of an input layer, a hidden layer and an output layer. The network receives an input x at time t. t , after hidden layer s t , the output value is o t ; According to the network calculation formula:

[0084] o t =g(Vs t )

[0085] s t =f(Ux t +Ws t-1 )

[0086] Where V, U and W are weight matrices, g(g) and f(g) are activation functions;

[0087] o t =g(Vs t ) Repeatedly insert formula s t =f(Ux t +Ws t-1 ), we can get

[0088] o t =g(Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +...)))).

[0089] In some embodiments, in step S2, the input of image features is to perform feature selection on statistical features and texture features extracted from the infrared image.

[0090] Statistical features include: average temperature value

[0091] Where n p is the number of infrared image pixels, T i is the surface source temperature of the infrared image;

[0092] Average Deviation Standard Deviation Skewness Kurtosis

[0093] The texture features are extracted from the spatial temperature co-occurrence matrix TLCM;

[0094] For two pixels (j, k) and (p, q) with temperature values ​​u and v, the angle between them is θ, then the number of times they appear in any direction of the image can be expressed as:

[0095] P(u,v,d,θ)=number{T(j,k)=u,T(p,q)=v,θ}

[0096] Where P is TLCM; parameter d is usually 1, indicating that the distance between two pixels is 1;

[0097] The extracted texture features include: energy

[0098] Where P(u,v) represents the element at temperature u and v in TLCM, G represents the temperature level; entropy Moment of inertia Relevance Where μ x , μ y , σ x and σ y is defined as follows: Anti-differentiation According to the definition of TLCM, the above texture features can be extracted from different directions. In order to reduce the fluctuation caused by the extraction of single-direction features, and to make full use of the features in the four directions of 0°, 45°, 90° and 135°, the average value is often taken as the extracted stable feature.

[0099] In the formula, f i is the eigenvalue of a feature in the i-th calculation direction, is the average value of a feature in four directions.

[0100] In some embodiments, in step S4, the calculation function of the loss value of the RNN network is: loss RNN =mean{(y RNN -y) 2}; where y RNN is the predicted output of the RNN network, T0, T0′ and T h are the actual temperature of the temperature measurement object, the infrared measurement temperature and the set temperature of the silicon carbon rod, τ, k λ and n are the equivalent refractive index of the endoscope long lens, the monochromatic attenuation coefficient of the internal atmosphere and the fitting coefficient, respectively. h_o is the equivalent angular coefficient of the high temperature object to the object being measured, and ε0 is the emissivity of the object being measured.

[0101] In some embodiments, in step S4, the calculation function of the loss value of the DNN network is: lossDNN =mean{[y DNN -(yy RNN )] 2};

[0102] In the formula, y RNN and DNN They are the predicted outputs of the RNN network and the DNN network respectively.

[0103] In some embodiments, in step S4, the calculation function of the loss value of the entire network is: loss RNN-DNN = a·loss RNN +b·loss DNN ; In the formula, loss RNN-DNN 、loss RNN and loss DNN They are the loss function values ​​of the RNN-DNN network, RNN network, and DNN network respectively; the weighting factors of the calculation functions of the loss values ​​of RNN and DNN are a and b respectively.

[0104] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0105] The following are examples of the present application. The examples described below are exemplary and are only used to explain the present application, and should not be construed as limiting the present application. If no specific techniques or conditions are specified in the examples, the techniques or conditions described in the literature in the art shall be followed.

[0106] Example 1

[0107] This embodiment proposes a roller kiln temperature correction method based on an infrared temperature compensation hybrid model, which is calibrated by an infrared temperature measurement compensation hybrid model. Figure 1 As shown, F h_o is the angular coefficient, and the parameters τ and k are obtained from the blackbody experimental data parameter identification. λ and n are the equivalent refractive index of the endoscope long lens and the monochromatic attenuation coefficient k of the internal atmosphere, respectively. λ and the fitting constant n in Planck's law, T0 and T0′ are the real temperature of the temperature measurement object and the infrared measurement temperature respectively. x1 and x2 are the highly correlated data obtained by correlation analysis of the production process data and the characteristic values ​​after extracting features from the infrared image data respectively. ΔG is the dynamics not modeled in the mechanism model. f(g) represents the temperature compensation model relationship, f RNN-DNN(g) represents the relationship between the RNN-DNN cascade neural network. The infrared temperature data is the infrared temperature of the object being measured, the production process data includes furnace pressure, electric energy and process temperature data, and the infrared image data is the infrared image measured by the infrared thermal imager.

[0108] Among them, the construction process and mechanism of the infrared temperature measurement compensation hybrid model are as follows:

[0109] The mechanism model is constructed based on the infrared temperature measurement mechanism. According to the infrared temperature measurement mechanism, when there are high-temperature objects around the object being measured, the effective infrared radiation brightness received by the endoscope long lens includes three parts: the target's own radiation, the high-temperature object's reflected radiation, and the atmospheric radiation. The formula is shown in formula (1):

[0110] W r =ε0τ a W0+F h_o ρ0τ a W h +ε a W a (1)

[0111] Where W r , W0, W h and W a They are respectively represented by the infrared radiation brightness received at the endoscope lens port, the infrared radiation brightness emitted by the temperature measurement object itself, the infrared radiation brightness of the adjacent high-temperature object, and the infrared radiation brightness of the atmosphere surrounding the measured object. a They are respectively expressed as the emissivity of the object being measured and the emissivity of the atmosphere, τ a represents the transmittance of the atmosphere, ρ0 represents the reflectivity of the object being measured, and F h_o It indicates the equivalent angle coefficient of the measured object to the high temperature object.

[0112] Due to the attenuation effect of the endoscopic long lens, the infrared radiation brightness n passing through the endoscopic long lens mouth is attenuated after passing through the long lens to obtain the infrared radiation brightness W received by the infrared thermal imager rd The relationship between the two is shown in formula (2).

[0113] W rd =τ c W r (2)

[0114] In the formula, τ c It is the attenuation factor of the infrared radiation brightness from the endoscope lens port to the infrared thermal imager receiving point, which depends on how the structural composition of the endoscope lens is defined.

[0115] In an endoscopic long lens, it is difficult to model the internal structure precisely due to its complex internal optical path structure. In order to reduce the complexity of modeling, the endoscopic long lens is regarded as an equivalent lens and an atmosphere.

[0116] After the long lens equivalent, τ c As the attenuation factor of the infrared radiation brightness from the endoscope lens port to the infrared thermal imager receiving point, it can be characterized by the transmittance of the equivalent lens and the attenuation effect of the infrared radiation passing through the internal atmosphere.

[0117] When infrared radiation passes through an absorbing gas, its energy is absorbed by gas molecules along the way. According to Bell's law, the gas attenuation coefficient τ λ,L for

[0118]

[0119] Where W λ,0 is the initial radiation intensity, L is the optical path length, W λ,L is the radiation intensity after atmospheric attenuation with an optical path length of L, k λ is the monochromatic attenuation coefficient of the atmosphere.

[0120] Knowing that the length of the endoscope lens is d and the corresponding equivalent refractive index of the lens is τ, its attenuation factor τ can be obtained c The expression of is shown in formula (4).

[0121]

[0122] According to equations (1), (2) and (4), the infrared radiation brightness W received by the infrared thermal imager under the interference of high temperature objects and endoscopic long lens can be obtained: rd for

[0123]

[0124] Often, infrared detectors work at a certain infrared wavelength. By integrating the received infrared radiation energy, they can output a voltage signal that is positively correlated with the infrared radiation brightness. The thermal detector output voltage can be used to replace the corresponding infrared radiation brightness to express the relationship between infrared radiation brightness and temperature, as shown in formula (6):

[0125] V(T)=K·W (6)

[0126] Wherein, V(T) represents the output voltage signal of the infrared detector when the infrared radiation temperature is T, W is the infrared radiation brightness when the infrared radiation temperature is T, and K is a constant value, reflecting that the output voltage signal is positively correlated with the infrared radiation brightness.

[0127] Substituting equation (6) into equation (5), we can obtain the relationship between the infrared radiation brightness and the output voltage signal of the infrared detector:

[0128]

[0129] Where V(g) represents the output voltage signal of the infrared thermal imager corresponding to the temperature. rd , T0, T h and T a They are respectively expressed as the radiation temperature of the object being measured, the surface temperature of the object being measured, the temperature of the adjacent high-temperature object and the temperature of the pure oxygen atmosphere inside the roller kiln.

[0130] According to Kirchhoff's law, at any given temperature and wavelength, for any material, its emissivity and absorptivity are equal. For an opaque object, its transmittance can be considered as 0. For the atmosphere, the spectral reflectance of the atmosphere can be considered as 0. Therefore, according to this law, equations (8) and (9) can be obtained.

[0131] ρ0=1-ε0 (8)

[0132] ε a =1-τ a (9)

[0133] Substituting equation (8) and equation (9) into equation (7), equation (7) can be expressed as

[0134]

[0135] When the temperature of the thermal detector changes due to the high temperature environment, the responsiveness of the thermal imager will change accordingly. Assume that the change in the output voltage signal V(T) caused by the temperature change of the thermal imager probe is ΔV. When the influence of the working temperature of the thermal detector is taken into account, the output voltage signal relationship will be corrected. The output voltage signal relationship of the endoscopic long-lens uncooled thermal imager for temperature measurement in the roller kiln environment is:

[0136]

[0137] When the thermal imager is used on a roller kiln, it still measures the temperature with the infrared temperature measurement model calibrated without interference. In fact, the thermal imager believes that there is no high-temperature object interfering with the infrared radiation of the measured object. The effective infrared radiation brightness received by the infrared thermal imager mainly includes two parts.

[0138] W rd =ε0τ a W0′+ε a W a (12)

[0139] Where W rd, W0′ and W a They are respectively represented by the infrared radiation brightness received by the infrared detector, the infrared radiation brightness emitted by the temperature measurement object itself under interference, and the infrared radiation brightness of the atmosphere around the object being measured. a They are respectively expressed as the emissivity of the object being measured and the emissivity of the atmosphere, τ a Indicates the transmittance of the atmosphere.

[0140] Substituting formula (6) into formula (12), we get

[0141]

[0142] Where T0′ represents the temperature measured by the infrared thermal imager under interference, T rd It is the radiation temperature of the object being measured.

[0143] Although the infrared temperature measured by the thermal imager under interference contains errors, a temperature compensation model can be established in turn.

[0144] Within a certain wavelength range, assuming that the spectral response rate of the infrared detector remains unchanged, the relationship between the voltage signal and the temperature of the object under test can be obtained by integrating Planck's law, as shown in formula (14).

[0145] V(T)=CT n (14)

[0146] In the formula, C and n are fitting coefficients and are constant values, which are related to the optical properties of the infrared thermal imager, the model and material of the thermal detector, etc., which are determined by the infrared thermal imager itself.

[0147] By combining equations (11), (13) and (14), we can derive the relationship between the actual temperature under interference and the temperature measured by the infrared thermal imager.

[0148]

[0149] Let ΔG = ΔV / C, C is the fitting constant in equation (14), and since the atmosphere inside the kiln is pure oxygen, then ε a =0, then the above formula becomes

[0150]

[0151] In the formula, T0, T0′ and T h They represent the true surface temperature of the object being measured, the temperature measured by the infrared thermal imager under interference, and the temperature of the adjacent high-temperature object. The length of the endoscopic long lens is d, the corresponding equivalent refractive index of the lens is τ, and the monochromatic attenuation coefficient is k λ , n is the fitting coefficient in Planck integral. ε0 and ε aare the emissivity of the object being measured and the emissivity of the pure oxygen atmosphere inside, τ a Indicates the transmittance of the atmosphere, F h_o It represents the equivalent angular coefficient of the object under test to the high-temperature object, and ΔG is the compensation term introduced by the temperature drift caused by the thermal detector being affected by the high-temperature environment.

[0152] The infrared temperature measurement compensation mechanism model (16) has already described the corresponding relationship between the infrared measurement temperature and the real surface temperature under interference. When the real spatial position corresponding to the temperature measurement point on the infrared image is determined, the corresponding correction temperature can be obtained according to (16). Then, the equivalent angular coefficient F of the high temperature object at the corresponding point needs to be calculated. h_o calculate.

[0153] The equivalent angular coefficient is determined by the spatial position between the object being measured and the high-temperature object, so the equivalent angular coefficient is related to the spatial position coordinates. When a pixel point on the infrared image is selected as the temperature measurement object, the corresponding angular coefficient can be calculated based on its spatial position distribution. Here, the equivalent angular coefficients of the high-temperature object are calculated for each point in the upper surface area of ​​the sagger, the second left thermocouple, and the upper thermocouple. Since the silicon carbon rod is the main heat source in the closed space of the roller kiln, the adjacent high-temperature object is regarded as a silicon carbon rod to calculate the angular coefficient of the silicon carbon rod to the object being measured.

[0154] When the medium between surface 1 and surface 2 of two objects is a transparent medium, the surfaces are diffusers and are isothermal, and the radiating surfaces have uniform physical properties and projected radiation, the radiation angle coefficient is only related to the geometric factors of the objects.

[0155]

[0156] In the formula, Q 12 is the projected radiation from surface 1 to surface 2, and Q1 is the effective radiation of surface 1.

[0157] Figure 2 is a schematic diagram of the roller kiln structure. This schematic diagram is a diagram of the internal structure of the roller kiln when the length of the sagger area is dy microelement and the sagger width is L2. Since the upper surface area of ​​the sagger with a length of dy microelement is selected, its physical properties are uniform and since this area is parallel to the silicon carbon rod, the projected radiation is uniform. The corresponding angular coefficient can be calculated by the projection method. The angular coefficient of a single silicon carbon rod to the plane where the upper surface area of ​​the sagger with a length of dy microelement is located is

[0158]

[0159] In the formula, is the angular coefficient of the i-th silicon carbon rod to the yellow plane on the upper surface of the sagger, θ i It is the angle between the plane where the silicon carbon rod is located and the sagger area.

[0160] In fact, the energy projected by the silicon carbon rod to the yellow plane on the upper surface of the sagger is not completely projected to the sagger. According to the definition of the angular coefficient

[0161]

[0162] In the formula, F 12 is the angular coefficient from surface 1 to surface 2, A1 and A2 are the areas of surface 1 and surface 2, θ1 and θ2 are the angles between the corresponding connecting line of the infinitesimal area of ​​the two surfaces and the two surfaces respectively, and R is the distance between the two infinitesimal surfaces.

[0163] exist Figure 2 For the yellow plane, dA2 = dygL1:

[0164]

[0165] For the sagger plane, dA2 = dygL2:

[0166]

[0167] Where A1 and A2 are the areas of the silicon carbon rod and the sagger, θ1 and θ2 are the angles between the corresponding connecting line of the microelement area of ​​the two surfaces and the two surfaces respectively, R is the distance between the two microelement surfaces, and y1 is the integral length of the longitudinal section of the sagger.

[0168] Combining equations (18), (20) and (21)

[0169]

[0170] In the formula, is the angular coefficient of the ith silicon carbon rod to the sagger plane where the upper surface of the sagger is located, L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0171] In the temperature compensation mechanism model (16), when the temperature measurement object is selected as the sagger area, the equivalent angle coefficient of the sagger area is

[0172]

[0173] In the formula, is the angular coefficient of the ith silicon carbon rod to the sagger plane where the upper surface of the sagger is located; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

[0174] like Figure 3As shown, the Oxy coordinate system is established in the longitudinal section of the roller kiln. The coordinate origin is selected to be the horizontal left side of the dy micro-element sagger area. The horizontal distance between the i-th silicon carbon rod and the coordinate origin is d i , the vertical distance is H, then according to the longitudinal section of the roller kiln, θ can be obtained 1_i The calculation formula is

[0175]

[0176] Similarly, according to the projection method, the projection process is as follows Figure 4 In (a) and (b), the thermocouples can be divided into the upper thermocouple and the second left thermocouple according to their spatial position. The second left thermocouple is parallel to the direction of the silicon carbon rod, and the upper thermocouple is perpendicular to the direction of the silicon carbon rod.

[0177] The angle coefficient calculation formula corresponding to the thermocouple can be obtained

[0178]

[0179] There is a compensation term ΔG in the mechanism model, which can be modeled using a data-driven approach. Then the infrared temperature compensation hybrid model (27) can be obtained.

[0180]

[0181] In the formula, T0, T0′ and T h They represent the true surface temperature of the object being measured, the temperature measured by the infrared thermal imager under interference, and the temperature of the adjacent high-temperature object. The length of the endoscopic long lens is d, the corresponding equivalent refractive index of the lens is τ, and the monochromatic attenuation coefficient is k λ , n is the fitting coefficient in Planck integral; ε0 is the emissivity of the object being measured, F h_o Indicates the equivalent angular coefficient of the object being measured to the high-temperature object, f RNN-DNN It is the output of the data-driven model RNN-DNN network.

[0182] In fact, f RNN-DNNIt is to model the unmodeled dynamics in the mechanism model. The unmodeled dynamics can be divided into two categories: the first category is the time series prediction item. The compensation item is introduced due to the temperature drift caused by the thermal detector being affected by the high temperature environment. The compensation item is related to the change of observation time and has obvious time-dependent characteristics. Therefore, the RNN network is introduced to extract the time-dependent characteristics before and after the data. In order to diversify the input information of the model, the training set selects the upper thermocouple and silicon carbon rod as the temperature measurement objects. The production process data and infrared temperature measurement data are used as the model input. The error of the mechanism model is used as the target to train the compensation item. The second category belongs to the multivariate prediction item, and this part of the unmodeled dynamics is related to the image features. Therefore, the difference between the error of the temperature measurement object and the output of the RNN network is selected as the target of the DNN model, and the RNN network output and infrared image features are used as input. Finally, the output of the RNN network and the output of the DNN network together constitute the unmodeled dynamics. The network structure diagram of the data model is shown as follows: Figure 5 shown.

[0183] The input of the network model includes production process data and infrared image features. Production process data includes electric energy data, temperature data and furnace pressure data. Among them, the electric energy data includes 16-dimensional data such as upper SCR voltage and current, meter A item voltage and current, Q meter active power, etc.; the temperature data includes 15-dimensional data such as upper set temperature, heating output, left temperature measurement, etc.; the furnace pressure data includes 6-dimensional data such as actual furnace pressure, oxygen concentration flow, etc. The production data includes a total of 37-dimensional data. According to the correlation analysis, data with a correlation greater than 0.5 are selected. Therefore, this embodiment considers the exhaust temperature and rear temperature measurement in the production temperature data; the production process data of the upper SCR power, lower SCR power, meter active power and meter power in the electric energy data.

[0184] Figure 6 The infrared thermal imager collected some RBG infrared images of the box furnace from September 25, 2023 to September 26, 2023, and used this data to extract and select infrared image features. The features are fully extracted from the two aspects of image statistical features and texture features to quantify the infrared temperature measurement error caused by the high-temperature objects on the furnace wall and sagger.

[0185] (1) Statistical characteristics

[0186] When the temperature of the furnace wall and sagger changes due to the high temperature environment, the surface source information of the infrared thermal image will change. Therefore, the statistical features are extracted based on the two-dimensional temperature in the thermal image to describe the changes in the temperature of the furnace wall and sagger.

[0187] The average temperature value is a description of the temperature level of the entire infrared image. During the heating process of the silicon carbon rod, the average value will change as the temperature of the infrared image changes.

[0188]

[0189] Where n p is the number of infrared image pixels, T i is the surface source temperature of the infrared image.

[0190] The average deviation reflects the deviation between the surface source temperature and the average temperature. During the heating process of the silicon carbon rod, the average deviation will change with the change of the infrared image temperature.

[0191]

[0192] The standard deviation is a measure of the dispersion between the pixels in the surface source temperature. During the heating process of the silicon carbon rod, the standard deviation will change.

[0193]

[0194] The skewness measures the symmetry of the distribution of the surface source temperature. When there is no high-temperature object, the temperature of the measured object will be lower on the image, otherwise it will be higher.

[0195]

[0196] Kurtosis reflects the flatness of the distribution of surface temperature data. Kurtosis will change depending on the degree of interference from high-temperature objects.

[0197]

[0198] (2) Texture features

[0199] Since the high-temperature object increases the infrared radiation received by the infrared thermal imager, the infrared temperature changes and the texture of the infrared thermal image changes. Therefore, it is necessary to extract the corresponding texture features from the infrared image to characterize the infrared temperature measurement error caused by the high-temperature objects in the furnace wall and sagger. The present invention uses the temperature-level co-occurrence matrix (TLCM) for the infrared image to describe the relationship between the temperature values ​​of the infrared thermal image pixel pairs.

[0200] For two pixels (j, k) and (p, q) with temperature values ​​u and v, the angle between them is θ, then the number of times they appear in any direction of the image can be expressed as:

[0201] P(u,v,d,θ)=number{T(j,k)=u,T(p,q)=v,θ} (33)

[0202] P is TLCM. The parameter d is usually 1, indicating that the distance between two pixels is 1.

[0203] In order to avoid the problem that the spatial temperature co-occurrence matrix cannot represent the texture well due to different image sizes and different temperature levels of infrared images, temperature mapping is usually used to map the temperature data to an appropriate temperature level. The temperature mapping formula is shown in formula (34).

[0204]

[0205] Where T(i,j) represents the temperature value corresponding to the pixel point (i,j) on the two-dimensional plane of the infrared thermal image, and T max and T min They represent the maximum and minimum temperature values ​​in the image respectively, and G represents the temperature level.

[0206] To reduce the amount of calculation, the temperature level is selected as 16, and the temperature data in the infrared image is mapped to the corresponding temperature level using temperature mapping. Five texture features, including energy, entropy, moment of inertia, correlation, and anti-difference, are extracted from TLCM.

[0207] Energy represents the uniformity of temperature in the infrared image. When the temperature value in the infrared temperature measurement area changes, the energy will change.

[0208]

[0209] Where P(u,v) represents the element at temperature u and v in the TLCM, and G represents the temperature level.

[0210] Entropy describes the complexity of the texture within an infrared image. If the internal oxygen atmosphere flow and temperature change, and the image uniformity changes, the entropy value will change.

[0211]

[0212] The moment of inertia can be used to measure the total amount of temperature change in an infrared image. The more uneven the temperature distribution of an infrared image is, the greater the moment of inertia is.

[0213]

[0214] The correlation can reflect the similarity of the image in the row and column directions, and the correlation reflects the local temperature correlation in the infrared temperature measurement area. The larger the value, the stronger the correlation.

[0215]

[0216] Among them, μ x , μ y , σ x and σ y is defined as follows:

[0217]

[0218] The difference reflects the clarity and regularity of the image texture. The larger the difference, the clearer and more regular the image texture is.

[0219]

[0220] According to the definition of TLCM, the above texture features can be extracted from different directions. In order to reduce the fluctuation caused by the single-direction feature extraction, in order to make full use of the features in the four directions of 0°, 45°, 90° and 135°, the average value is often taken as the extracted stable feature.

[0221]

[0222] In the formula, f i is the eigenvalue of a feature in the i-th calculation direction, is the average value of a feature in four directions.

[0223] Although 10 image features are extracted in this embodiment, not all extracted image features are effective for temperature compensation. In order to reduce the model complexity caused by feature inputs with low correlation, it is necessary to perform correlation analysis on all features and select features with strong correlation as inputs to the DNN.

[0224] Correlation analysis of image features and unmodeled dynamics in ΔG excluding compensation terms is performed according to Pearson's formula (41).

[0225]

[0226] In the formula, r is the Pearson correlation coefficient of the image feature data, x i and i are the i-th sample values ​​of image feature data and compensation item, respectively. and is the mean value corresponding to the image feature data and the compensation item, and n corresponds to the total number of samples.

[0227] It can be seen from the correlation coefficient Table 1 that not all features have a great correlation with the unmodeled dynamics except the compensation term. In order to reduce the complexity of the model, image features with a correlation greater than 0.5 are selected, i.e. S k A total of 3 dimensions, , and K, are used as the input of the DNN model.

[0228] Table 1 Correlation coefficient table

[0229]

[0230] The series network is first fitted by RNN to obtain the time series prediction item. Specifically, the nodes between the hidden layers in RNN are interconnected. The input of the hidden layer includes not only the output of the input layer at the current moment, but also the output of the hidden layer at the previous moment. This design enables RNN to memorize the data at the previous moment and use it in the calculation of the output at the current moment.

[0231] like Figure 7 As shown, a simple RNN consists of an input layer, a hidden layer, and an output layer. The network receives the input x at time t. t , after the hidden layer st, the output value is o t .

[0232] According to the network calculation formula

[0233] o t =g(Vs t ) (42)

[0234] s t =f(Ux t +Ws t-1 ) (43)

[0235] Where V, U and W are weight matrices. g(g) and f(g) are activation functions.

[0236] Substituting formula (43) into formula (42) repeatedly, we can get

[0237] o t =g(Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +...)))) (44)

[0238] It can be seen that the output of the RNN network in the cascade neural network is affected by the historical input value x t 、x t-1 、x t-2 , ... From the above analysis, it can be seen that the RNN network can extract the temporal features of the input sequence and effectively extract the time dependency of the data, thereby fully extracting the temporal prediction items in the dynamics that are not modeled in this paper.

[0239] The loss function of the RNN network (i.e., the calculation function of the loss value of the RNN network) is the mean square error (MSE)

[0240] loss RNN =mean{(y RNN -y) 2} (45)

[0241] In the formula, y RNNis the predicted output of the RNN network, T0, T0′ and T h are the actual temperature of the temperature measurement object, the infrared measurement temperature and the set temperature of the silicon carbon rod, τ, k λ and n are the equivalent refractive index of the endoscope long lens, the monochromatic attenuation coefficient of the internal atmosphere and the fitting coefficient, respectively. h_o is the equivalent angular coefficient of the high temperature object to the object being measured, and ε0 is the emissivity of the object being measured.

[0242] Since it is a series neural network, the DNN network is used to fit the second type of unmodeled dynamics. The target value of the loss function of the DNN network (i.e., the calculation function of the loss value of the DNN network) depends not only on the selected temperature measurement object, but also on the output value of the RNN network, that is,

[0243] loss DNN =mean{[y DNN -(yy RNN )] 2} (46)

[0244] In the formula, y RNN and DNN They are the predicted outputs of the RNN network and the DNN network respectively.

[0245] In order to calculate the gradients of the series network synchronously, it is necessary to weightedly fuse the loss functions of RNN and DNN, with the weighting factors being a and b respectively.

[0246] loss RNN-DNN = a·loss RNN +b·loss DNN (47)

[0247] In the formula, loss RNN-DNN (That is, the calculation function of the loss value of the RNN-DNN network as a whole is), loss RNN and loss DNN They are the loss function values ​​of RNN-DNN network, RNN network, and DNN network respectively.

[0248] The implementation steps of the RNN-DNN tandem neural network are summarized as follows:

[0249] Step 1: Define the network and set network hyperparameters. Set the corresponding parameters including time step, network parameters, weighting factors, etc., and initialize the model;

[0250] Step 2: Input normalization. The six-dimensional production data including exhaust temperature, rear temperature measurement 1, upper SCR power, lower SCR power, meter active power and meter power in the production data, the three features including average temperature value, skewness and kurtosis in the image features and the infrared temperature measurement data of the silicon carbon rod are used as the network input and the unmodeled dynamic target data corresponding to the silicon carbon rod object are normalized;

[0251] Step 3: Load data. Randomly select input sequences and target sequences of the same length as the time step from the temperature data of any pixel point to obtain the hidden layer vector of the RNN network. Input all input sequence data into the RNN network to obtain the output value of the RNN network. Normalize the output value of the RNN network and the image features as the input of the DNN network to obtain the output of the DNN network.

[0252] Step 4: Calculate loss and back propagate. Calculate the loss values ​​of the RNN network and the DNN network, thereby obtaining the loss value of the entire network, and back propagate the gradient to update the network parameters;

[0253] Step 5: Verify the model with the validation set and save the network parameters. Select the upper layer thermocouple data as the validation set, bring it into the network to get the network output and bring it into the calibration formula to get the calibration temperature, calculate the maximum error, and save the network parameters;

[0254] Step 6: Train the network. Repeat steps 3 to 5. When the number of iterations is within a certain range, the network parameters with the smallest maximum error of the validation set are taken as the optimal network parameters.

[0255] Step 7: Model verification. Select the second thermocouple data from the left as the test set, insert it into the network to get the network output and then insert it into the calibration formula to get the calibration temperature, draw the calibration curve and error curve, and verify the accuracy of the network.

[0256] In order to verify the accuracy of the model, the production sintering data of the box furnace from September 25, 2023 to September 26, 2023 were used. The data includes infrared image RBG map and temperature matrix, production process data and thermocouple measurement temperature, a total of 113 sets of data. 6-dimensional data are selected from the production process data, including exhaust temperature and rear temperature measurement 1 in the production temperature data; upper SCR power, lower SCR power, meter active power and meter power in the electric energy data. 3-dimensional data are selected from the image feature extraction, including three features: mean deviation, skewness and kurtosis. The network input includes 6 dimensions of production data, 3 dimensions of image features and 1 dimension of infrared temperature measurement data of silicon carbon rods, a total of 10 dimensions as the training set. The validation set selects 113 groups of 10-dimensional data of the upper thermocouple. The test set selects 113 groups of 10-dimensional data of the left second thermocouple data.

[0257] The hyperparameters of the RNN network include the number of hidden layers, the hidden layer dimensions, and the training time step. The hyperparameters of the DNN network are the number of hidden layers and the hidden layer dimensions. The key performance parameters are the RNN hidden layer dimensions and the training time step. The RNN-DNN as a whole has weighted fusion parameters a and b. This embodiment uses a cross-validation method to adjust the value of each parameter to select the optimal value.

[0258] First, fix other parameters, including the number of RNN hidden layers to 1, the number of DNN hidden layers to 3, and the DNN hidden layer dimension to 16, and adjust the hidden layer dimension N of the RNN network in the model to multiples of 4. The model performance evaluation indicators are shown in Table 2. From Table 2, it can be seen that when the hidden layer dimension is small, the model prediction MSE is small, the maximum temperature prediction error is small, and the minimum prediction error is the largest. As the hidden layer dimension increases, due to the increase in model complexity and the increase in effective information in the extracted data, the MSE will gradually decrease, and the maximum temperature prediction error will show an overall downward trend. However, when the hidden layer dimension is too large, the sample information is over-extracted, which will cause the maximum temperature prediction error to increase and become overfitting. Therefore, based on the above considerations, when the hidden layer dimension is 16, the model output result is already good. In order not to increase the complexity of the model, the hidden layer dimension N is selected to be 16.

[0259] When N=16, adjust the training time step T of the RNN network, and the model prediction indicators are shown in Table 4-2. The length of the time step determines how much the current output state of the model is related to the previous input. From Table 3, it can be found that the effect is best when T=6, which means that the output of the current state of the RNN model is most related to the input of the previous 6 states, which shows the time-dependent characteristics of the data, and the model output result is already good. Based on the above considerations, the time step of T=6 is selected as the optimal time step.

[0260] Table 2 Performance evaluation indicators of RNN-DNN model when the dimension of RNN hidden layer changes

[0261]

[0262] Table 3 Performance evaluation indicators of RNN-DNN model when the time step changes

[0263]

[0264] According to the overall loss function formula of the RNN-DNN tandem neural network, the weighted fusion parameters a and b are selected. The selection of this part is generally determined by the final loss function value of the back propagation of the RNN and DNN loss functions respectively.

[0265] from Figure 8 and Fig. 9It can be seen that the loss values ​​of RNN and DNN are close to 9:1, so a=0.1 and b=0.9 are selected to ensure the same dimensions in the total loss function of RNN-DNN.

[0266] Based on the above model parameter selection, the number of RNN hidden layers is 1, the dimension is 16, the time step is 6, the number of DNN hidden layers is 3, and the dimension of DNN hidden layers is 16. The RNN-DNN weighted fusion parameters and are 0.1 and 0.9 respectively.

[0267] According to the setting of hyperparameters, the training set is used for training and the test set is used for verification, and the temperature compensation effect of the left second thermocouple under the hybrid drive of the RNN-DNN data-driven model and the mechanism model is obtained. The model verification effect is as follows Fig.10 As shown, Fig.10 The left figure is a temperature correction curve, and the right figure is an error graph. Table 4 is a performance evaluation index table under different data-driven models, from which it can be seen that the accuracy of the model proposed by the present invention is higher than that of other methods.

[0268] Table 4 Performance evaluation indicators driven by different data models

[0269]

[0270] Other beneficial effects of the present invention include:

[0271] 1. The present invention introduces the interference of adjacent high-temperature objects on infrared temperature measurement into the roller kiln temperature measurement model, thereby improving the accuracy of the model.

[0272] 2. The present invention proposes a RNN-DNN series neural network structure. The use of this series structure can simultaneously consider time dependence and multi-factor influences, thereby improving the model's ability to model unmodeled dynamics.

[0273] 3. A temperature compensation model is established from the perspective of infrared temperature measurement mechanism, which improves the interpretability compared with a single data-driven model. At the same time, it takes into account the influence of high-temperature objects caused by the internal structure of the roller kiln and overcomes the interference of the endoscopic long lens.

[0274] 4. The proposed RNN-DNN series network structure can compensate for the unmodeled dynamics in the model. This series structure can extract the time-dependent terms of the temperature drift caused by the thermal detector being affected by the high temperature environment and the multivariate prediction terms related to the current moment due to the influence of the internal atmosphere and furnace wall radiation of the roller kiln, thereby improving the accuracy of the model.

[0275] The specific implementation of the present invention described above does not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A roller kiln temperature correction method based on an infrared temperature compensation hybrid model, characterized in that: Calibration is performed using an infrared temperature measurement compensation hybrid model, the infrared temperature measurement compensation hybrid model formula is: In the formula, T0, T′0 and T h They represent the true surface temperature of the object being measured, the temperature measured by the infrared thermal imager under interference, and the temperature of the adjacent high-temperature object. The length of the endoscopic long lens is d, the corresponding equivalent refractive index of the lens is τ, and the monochromatic attenuation coefficient is k λ , n is the fitting coefficient in Planck integral; ε0 is the emissivity of the object being measured, F h_o Indicates the equivalent angular coefficient of the object being measured to the high-temperature object, f RNN-DNN It is the output of the data-driven model RNN-DNN network.

2. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 1 is characterized in that: When the temperature measurement object is the sagger area, the equivalent angular coefficient F of the corresponding sagger area h_o for: In the formula, is the angular coefficient of the ith silicon carbon rod to the sagger plane where the upper surface of the sagger is located; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

3. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 2 is characterized in that: Said The calculation formula is: In the formula, is the angular coefficient of the i-th silicon carbon rod to the yellow plane on the upper surface of the sagger; L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, θ 1_i is the angle between the ith silicon carbon rod and the sagger area.

4. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 1 is characterized in that: When the temperature measurement object is the second thermocouple on the left, the corresponding angular coefficient is: Wherein, L1 is the effective length of the silicon carbon rod in the roller kiln area, L2 is the width of the sagger, and θ1 is the angle between the connecting line corresponding to the infinitesimal area in surface 1 and surface 1.

5. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 1 is characterized in that: When the temperature measurement object is the upper thermocouple, the corresponding angle coefficient is: Among them, L1 is the effective length of the silicon carbon rod in the roller kiln area; θ1 is the angle between the connecting line corresponding to the microelement area in surface 1 and surface 1, and R is the distance between the two microelement surfaces.

6. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 1 is characterized in that: The data-driven model consists of a RNN-DNN series network. The input of the RNN network consists of production data and infrared temperature measurement data of silicon carbon rods. The input of the DNN network consists of the output of the RNN network and infrared image features. The RNN and DNN form a series network structure. The outputs of the RNN and DNN networks together constitute the f RNN-DNN ; The implementation steps of the RNN-DNN tandem neural network are as follows: S1: Define the network and set network hyperparameters; set corresponding parameters including time step, network parameters and weighting factors, and initialize the model; S2: Input normalization; the exhaust temperature, rear temperature measurement, upper SCR power, lower SCR power, meter active power and meter power production data in the production data; the average temperature value, skewness, kurtosis and infrared temperature measurement data of silicon carbon rods in the image features are used as network input and the unmodeled dynamic target data corresponding to the silicon carbon rod object are normalized; S3: Load data; randomly select input sequences and target sequences of the same length as the time step in the temperature data of any pixel point, obtain the hidden layer vector of the RNN network, input all input sequence data into the RNN network to obtain the output value of the RNN network, normalize the output value of the RNN network and the image features as the input of the DNN network, and obtain the output of the DNN network; S4: Calculate loss and back propagate; calculate the loss value of the RNN network and the DNN network, thereby obtaining the loss value of the entire network, and back propagate the gradient to update the network parameters; S5: Verify the model with the validation set and save the network parameters; select the upper thermocouple data as the validation set, bring it into the network to obtain the network output and bring it into the calibration formula to obtain the calibration temperature, calculate the maximum error, and save the network parameters; S6: Train the network; repeat S3 to S5. When the number of iterations is within the set range, the network parameters with the smallest maximum error of the validation set are taken as the optimal network parameters. S7: Model verification; select the second thermocouple data from the left as the test set, input it into the network to obtain the network output and input it into the calibration formula to obtain the calibration temperature, draw the calibration curve and error curve, and verify the accuracy of the network.

7. The method for correcting the infrared temperature of a roller kiln according to claim 6, characterized in that: In step S1, the structure of the RNN network consists of an input layer, a hidden layer and an output layer. The network receives the input x at time t. t , after hidden layer s t , the output value is o t ; According to the network calculation formula: o t =g(Vs t ) s t =f(Ux t +Ws t-1 ) Where V, U and W are weight matrices, g(g) and f(g) are activation functions; o t =g(Vs t ) Repeatedly insert formula s t =f(Ux t +Ws t-1 ), we can get o t =g(Vf(Ux t +Wf(Ux t-1 +Wf(Ux t-2 +...))))).

8. The roller kiln temperature correction method based on the infrared temperature compensation hybrid model according to claim 6 is characterized in that: In step S2, the input of image features is feature selection in terms of statistical features and texture features extracted from the infrared image; Statistical features include: average temperature value Where n p is the number of infrared image pixels, T i is the surface source temperature of the infrared image; Average Deviation Standard Deviation Skewness Kurtosis The texture features are extracted from the spatial temperature co-occurrence matrix TLCM; For two pixels (j, k) and (p, q) with temperature values ​​u and v, the angle between them is θ, then the number of times they appear in any direction of the image can be expressed as: P(u,v,d,θ)=number{T(j,k)=u,T(p,q)=v,θ} Where, P is TLCM; parameter d It is usually set to 1, indicating that the distance between two pixels is 1; The extracted texture features include: energy Where P(u,v) represents the element at temperature u and v in TLCM, G represents the temperature level; entropy Moment of inertia Relevance Where μ x , μ y , σ x and σ y is defined as follows: Anti-differentiation According to the definition of TLCM, the above texture features can be extracted from different directions. In order to reduce the fluctuation caused by the extraction of single-direction features, and to make full use of the features in the four directions of 0°, 45°, 90° and 135°, the average value is often taken as the extracted stable feature: In the formula, f i is the eigenvalue of a feature in the i-th calculation direction, is the average value of a feature in four directions.

9. The method for calibrating the infrared temperature of a roller kiln according to claim 6, characterized in that: In step S4, the calculation function of the loss value of the RNN network is: loss RNN =mean{(y RNN -y) 2 }; where y RNN is the predicted output of the RNN network, T0, T0′ and T h are the actual temperature of the temperature measurement object, the infrared measurement temperature and the set temperature of the silicon carbon rod, τ, k λ and n are the equivalent refractive index of the endoscope long lens, the monochromatic attenuation coefficient of the internal atmosphere and the fitting coefficient, respectively. h_o is the equivalent angular coefficient of the high temperature object to the object being measured, and ε0 is the emissivity of the object being measured.

10. The method for constructing an infrared compensation hybrid model of a roller kiln according to claim 6, characterized in that: In step S4, the calculation function of the loss value of the DNN network is: loss DNN =mean{[y DNN -(yy RNN )] 2 }; where y RNN and DNN are the predicted outputs of the RNN network and the DNN network respectively; and / or, in step S4, the calculation function of the loss value of the entire network is: loss RNN-DNN = a·loss RNN +b·loss DNN ; In the formula, loss RNN-DNN 、loss RNN and loss DNN They are the loss function values ​​of the RNN-DNN network, the RNN network, and the DNN network respectively; the weighting factors of the calculation functions of the loss values ​​of RNN and DNN are a and b respectively.

Citation Information

Cited By

  • Multi-dimensional parameter calibration method for infrared temperature measurement of composite insulator

    CN119756606A

  • Temperature measurement method and device of thermal imager and storage medium

    CN120846509A

  • Temperature measurement error estimation method and device of thermal infrared imager

    CN121834153A