Method and system for dynamically monitoring heat-conducting property of internal insulation board based on multi-modal data

Abstract

The invention discloses a multi-modal data-based dynamic monitoring method and system for the heat-conducting property of an inner insulation board, the system comprises a first temperature sensor, a second temperature sensor, a heat flow sensor, an infrared thermal imager and a data processing device, and the method comprises the following steps: collecting related data; calculating the heat conductivity coefficient lambda of the inner insulation board by using the collected data; performing data preprocessing, and performing data enhancement on the heat conductivity coefficient lambda and the acquired infrared image data by using an adversarial neural network; model training and testing: carrying out model training by utilizing the preprocessed data, respectively taking the original data as input of the trained model, and respectively carrying out model testing; and finally, performing decision-making layer data fusion based on the D-S evidence theory by using a model test result. The heat conductivity coefficient of the internal thermal insulation wallboard can be accurately, conveniently and efficiently obtained, so that the real-time dynamic monitoring of the heat conductivity of the internal thermal insulation wallboard is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of the construction industry, and particularly relates to a dynamic monitoring method and system for the thermal conductivity of internal insulation boards based on multimodal data. Background Art

[0002] The thermal conductivity of building internal insulation wallboards is a key index for evaluating their thermal insulation performance. By regularly detecting the thermal conductivity, it is possible to understand whether the thermal insulation performance of the thermal insulation material is affected. The existing detection methods for the thermal conductivity of internal insulation wallboards require a long time and a complex operation process, mainly relying on regular or irregular manual inspections, lacking dynamic monitoring of the thermal conductivity of building internal insulation wallboards. Therefore, there is an urgent need to develop a technical solution that can dynamically monitor the thermal insulation performance of internal insulation wallboards during their service life. Summary of the Invention

[0003] To solve the above problems, the present invention provides a dynamic monitoring method and system for the thermal conductivity of internal insulation boards based on multimodal data, which can obtain the thermal conductivity of internal insulation wallboards more accurately, conveniently and efficiently to realize real-time dynamic monitoring of the thermal conductivity of internal insulation wallboards.

[0004] The technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present application discloses a dynamic monitoring system for the thermal conductivity of internal insulation boards based on multimodal data. The internal insulation board is arranged on the inner side of the wall, and a plaster layer is coated on the surface of the internal insulation board. The system includes:

[0006] A first temperature sensor for collecting the temperature T1 between the wall and the internal insulation board;

[0007] A second temperature sensor for collecting the temperature T2 on the surface of the plaster layer;

[0008] A heat flux sensor for collecting the heat flux density q on the surface of the plaster layer;

[0009] An infrared thermal imager for collecting the infrared image on the surface of the plaster layer;

[0010] A data collector, which is communicatively connected to the first temperature sensor, the second temperature sensor and the heat flux sensor, and is used for transmitting the collected temperatures T1, T2 and heat flux density q to a data processing device;

[0011] A data processing device, which is communicatively connected to the data collector and the infrared thermal imager, and is used for receiving the temperature data T1, T2, heat flux density data q and infrared image data, and processing the received data.

[0012] As an alternative technical solution, both the first temperature sensor and the second temperature sensor are platinum resistance temperature sensors; the first temperature sensor is embedded between the wall and the internal insulation board, and the second temperature sensor is attached to the surface of the plaster layer through thermal conductive silicone.

[0013] As an alternative technical solution, the heat flux sensor is attached to the surface of the plaster layer using petroleum jelly as a coupling agent.

[0014] In a second aspect, the present application discloses a dynamic monitoring method for the thermal conductivity of an internal insulation board based on multi-modal data. The method is based on the above system and includes:

[0015] Collect relevant data, where the relevant data includes the temperature T1 between the wall and the internal insulation board, the temperature T2 on the surface of the plaster layer, the heat flux density q on the surface of the plaster layer, and the infrared image data on the surface of the plaster layer;

[0016] Calculate the thermal conductivity λ of the internal insulation board using the collected temperatures T1, T2, and heat flux density q;

[0017] Data preprocessing: Use an adversarial neural network to perform data enhancement processing on the thermal conductivity λ and infrared image data;

[0018] Model training and testing: Use the preprocessed data for model training respectively, and then use the original data as the input of the trained models for model testing respectively; where the original data includes the thermal conductivity λ and infrared image data;

[0019] Perform decision-level data fusion based on the D-S evidence theory using the model test results.

[0020] As an alternative technical solution, the thermal conductivity λ is calculated according to the following formula:

[0021]

[0022] where δ is the thickness of the internal insulation board.

[0023] As an alternative technical solution, the model training and testing specifically include: using the enhanced infrared image data as the input of a convolutional neural network CNN for model training to obtain a trained convolutional neural network CNN; using the enhanced thermal conductivity λ as the input of a basic classifier for model training to obtain a trained basic classifier; and then using the original infrared image data as the input of the trained CNN and using the original thermal conductivity λ as the input of the trained basic classifier for model testing respectively.

[0024] As an alternative technical solution, the convolutional neural network CNN includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a softmax layer. The model training with the enhanced infrared image data as the input of the convolutional neural network CNN includes:

[0025] The input layer receives image data; wherein, the image data is the enhanced infrared image data;

[0026] The convolutional layer performs a convolution operation on the input image data to extract local features and form a feature map;

[0027] The pooling layer performs a downsampling operation on the feature map;

[0028] Through the combined use of several convolutional layers and pooling layers, features are further extracted and compressed;

[0029] After the compressed features are integrated by the fully connected layer, the softmax layer is used to convert the numerical value to a probability, and the final classification result or regression value is output, completing the processing and analysis of the input data by the convolutional neural network.

[0030] As an alternative technical solution, for the l-th convolutional layer of the convolutional neural network CNN, its output can be expressed as:

[0031]

[0032] Wherein, represents the element at the i-th row and j-th column of the k-th output feature map of the l-th layer, represents the element at the (i + q - 1)-th row and (j + p - 1)-th column of the k-th input feature map of the (l - 1)-th layer, represents the element size of the q-th row and p-th column of the k-th convolutional kernel of the l-th layer, which is N U ×N U ,N U represents the height and width of the convolutional kernel, K represents the number of convolutional layers, and g k ,l is the bias of the k-th convolutional kernel of the l-th layer, and * is the convolution operation symbol; f(·) is the activation function, which is used to convert the linear operation into a non-linear one.

[0033] As an alternative technical solution, the downsampling operation of the feature map by the pooling layer is performed through the following formula:

[0034]

[0035] Wherein, represents the corresponding element of the k-th output feature map after the pooling operation, and MaxPooling(·) represents taking the maximum value within the receptive field.

[0036] As an alternative technical solution, the conversion of numerical values to probabilities through the softmax layer is performed by the following formula:

[0037]

[0038] where is a feature vector associated with the input; the softmax(·) function maps the feature set into a C-dimensional vector, where the value of each vector is in the range (0,1) and the sum of all vectors is 1; represents the probabilities of each category; represents the non-linear functional relationship fitted by the convolutional neural network.

[0039] As an alternative technical solution, the decision-level data fusion based on the D-S evidence theory using the model test results includes: analyzing the thermal conductivity and infrared image data samples to be fused, obtaining all possible propositions after data fusion, and constructing the identification framework in the D-S evidence theory. In the identification framework, evidence is assigned to each proposition, and the belief degrees of each proposition in the identification framework are constructed by combining the belief function and the plausibility function in the D-S evidence theory. The information provided by multiple evidences is fused using the D-S evidence theory synthesis method, and the belief degrees of each proposition are synthesized. Each independent model generates a decision output according to its training results, and the output result is a probability value and a classification result. The independent decision results of each modality are fused.

[0040] The beneficial effects of the present invention are as follows:

[0041] 1. A dynamic monitoring method and system for the thermal conductivity of an internal insulation board based on multi-modal data proposed in this application evaluates and monitors the thermal conductivity of the internal insulation wall panel through accurate temperature measurement. Microscopic monitoring is carried out using a temperature sensor. Specifically, the temperature measurement depends on a platinum resistance temperature sensor, which is a high-precision and high-stability temperature measurement device, especially suitable for long-term monitoring and recording of temperature changes. In this application, one of the platinum resistance temperature sensors is embedded between the internal insulation board and the wall to directly contact and measure the temperature inside the internal insulation board. The design of the system allows the sensor to collect temperature data at fixed time intervals. This real-time monitoring can not only help to detect and solve potential insulation problems in a timely manner, but also monitor the deterioration of building insulation materials for timely repair. Overall, this real-time monitoring method and system provide an effective technical means for building energy conservation and performance evaluation of insulation materials.

[0042] 2. During the application, macroscopic monitoring is carried out using an infrared thermal imager, that is, machine vision is used to collect and analyze real-time images of the inner wall surface (i.e., the plaster layer) to achieve non-invasive monitoring of the thermal conductivity of the wall. It uses a high-resolution camera to capture images of the surface of the internal thermal insulation wall panel and analyzes and evaluates the change of the thermal insulation performance of the wall through image processing algorithms.

[0043] 3. In this application, a multi-modal data fusion algorithm is used to identify material degradation. By combining data of two modalities, namely heat flux density data and infrared images, the accuracy and robustness of the identification can be improved, the influence of false alarms and environmental interference can be reduced, the detection speed can be accelerated, the maintenance cost can be lowered, the adaptability and interpretability of the system can be enhanced, the safety can be strengthened, and predictive maintenance can be supported, thereby promoting the development of intelligent technologies and realizing more efficient, safe and economical material management. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 FIG. is an installation schematic diagram of a dynamic monitoring system for the thermal conductivity of an internal thermal insulation board based on multi-modal data.

[0045] Figure 2 FIG. is a flow schematic diagram of a dynamic monitoring method for the thermal conductivity of an internal thermal insulation board based on multi-modal data.

[0046] Figure 3 FIG. is an architecture diagram of a convolutional neural network CNN. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the methods and components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment

[0049] As Figure 1 shown, this application discloses a dynamic monitoring system for the thermal conductivity of an internal thermal insulation board based on multi-modal data. The internal thermal insulation board 7 is arranged inside the wall 6, and a plaster layer 8 is coated on the surface of the internal thermal insulation board 7. The system includes:

[0050] The first temperature sensor 21 is used to collect the temperature T1 between the wall 6 and the internal insulation board 7;

[0051] The second temperature sensor 22 is used to collect the temperature T2 on the surface of the plaster layer 8;

[0052] The heat flux sensor 9 is used to collect the heat flux density q on the surface of the plaster layer 8;

[0053] The infrared thermal imager 5 is used to collect the infrared image on the surface of the plaster layer 8;

[0054] The data collector 4 is communicatively connected to the first temperature sensor 21, the second temperature sensor 22, and the heat flux sensor 9 through a data transmission line 1, and is used to transmit the collected temperatures T1, T2, and heat flux density q to the data processing device 3;

[0055] The data processing device 3 is communicatively connected to the data collector 4 through a data transmission line 1 and is wirelessly communicatively connected to the infrared thermal imager 5, and is used to receive the temperature data T1, T2, the heat flux density data q, and the infrared image data, and process the received data. The processing process can refer to S2 - S5 in the following steps and will not be elaborated here.

[0056] As an optional implementation manner, both the first temperature sensor 21 and the second temperature sensor 22 adopt platinum resistance temperature sensors; the first temperature sensor 21 is buried between the wall 6 and the internal insulation board 7, and the second temperature sensor 22 is attached to the surface of the plaster layer 8 through thermal conductive silicone.

[0057] As an optional implementation manner, the heat flux sensor 9 is attached to the surface of the plaster layer 8 by using vaseline as a coupling agent.

[0058] As Figure 2 shown, the present application discloses a dynamic monitoring method for the thermal conductivity of an internal insulation board based on multi - modal data. This method is based on the above - mentioned Figure 1 system. The collected infrared image is used to determine whether there are deterioration phenomena such as cracks and peeling on the internal insulation board 7 through an image classification algorithm. A decision - level multi - modal data fusion algorithm based on the D - S evidence theory is adopted to effectively fuse the thermal conductivity data and the infrared image data at the decision level, so as to accurately warn of the deterioration of the wallboard insulation performance. Specifically, the method includes:

[0059] S1. Collect relevant data, where the relevant data includes the temperature T1 between the wall 6 and the internal insulation board 7, the temperature T2 on the surface of the plaster layer 8, the heat flux density q on the surface of the plaster layer 8, and the infrared image data on the surface of the plaster layer 8.

[0060] S2. Calculate the thermal conductivity λ of the internal insulation board 7 using the collected temperatures T1, T2 and the heat flux density q.

[0061] Further, the thermal conductivity λ is calculated according to the following formula:

[0062]

[0063] where δ is the thickness of the internal insulation board 7.

[0064] S3. Data preprocessing: Use the adversarial neural network to perform data augmentation on the thermal conductivity λ and the infrared image data.

[0065] S4. Model training and testing: Use the preprocessed data for model training respectively, and then use the original data as the input of the trained model for model testing respectively; where the original data includes the thermal conductivity λ and the infrared image data. Specifically, for the preprocessed infrared image data, a convolutional neural network CNN can be selected to perform classification training on the temperature field uniform images and the temperature field non-uniform images; for the preprocessed thermal conductivity λ, a basic classifier can be selected to perform classification training on the deteriorated and intact data.

[0066] Further, the model training and testing specifically includes: Using the enhanced infrared image data as the input of the convolutional neural network CNN for model training to obtain the trained convolutional neural network CNN; using the enhanced thermal conductivity λ as the input of the basic classifier for model training to obtain the trained basic classifier; and then using the original infrared image data as the input of the trained CNN and using the original thermal conductivity λ as the input of the trained basic classifier for model testing respectively. The basic classifier uses a support vector machine.

[0067] As an optional implementation manner, the convolutional neural network CNN includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and a softmax layer. The using the enhanced infrared image data as the input of the convolutional neural network CNN for model training includes:

[0068] The input layer receives the image data; where the image data is the enhanced infrared image data;

[0069] Use the convolutional layer to perform convolutional operations on the input image data to extract local features to form a feature map;

[0070] Use the pooling layer to perform downsampling operations on the feature map;

[0071] Through the combined use of several convolutional layers and pooling layers, further extract and compress features;

[0072] After the compressed features are integrated by the fully connected layer, the softmax layer is used to convert the numerical values into probabilities, and the final classification result or regression value is output, completing the processing and analysis of the input data by the convolutional neural network.

[0073] Specifically, please refer to Figure 3 , the convolutional neural network CNN mainly consists of an input layer, a convolutional layer, a pooling layer, a Dropout layer, and a fully connected layer. First, the input layer receives image data (i.e., the enhanced infrared image data). Then, the data enters the convolutional layer, where the convolutional kernel slides on the input data, and through the convolutional operation of element-wise multiplication and summation, local features are extracted to form feature maps. Next, the pooling layer performs downsampling on the feature maps. It reduces the spatial dimension of the feature maps, reduces the number of parameters and computational complexity, and retains important features by applying pooling functions such as the maximum value or average value on the feature maps. After the combined use of several convolutional layers and pooling layers, features are further extracted and compressed. Then, the Dropout layer can prevent the network from overfitting and enhance the generalization ability of the model. Finally, the data flows into the fully connected layer. The neurons in the fully connected layer are connected to all neurons in the previous layer, integrating the previously extracted features, and through a series of linear and non-linear transformations, output the final classification result or regression value, etc., completing the processing and analysis process of the input data by the convolutional neural network.

[0074] Furthermore, for the l-th convolutional layer of the convolutional neural network CNN, its output can be expressed as:

[0075]

[0076] where represents the element at the i-th row and j-th column of the k-th output feature map in the l-th layer, represents the element at the (i + q - 1)-th row and (j + p - 1)-th column of the k-th input feature map in the (l - 1)-th layer, represents the element at the q-th row and p-th column of the k-th convolutional kernel in the l-th layer, with a size of N U ×N U , N U represents the height and width of the convolutional kernel, K represents the number of convolutional layers, and g k ,l is the bias of the k-th convolutional kernel in the l-th layer, and * is the convolutional operation symbol; f(·) is the activation function, which is used to convert the linear operation into a non-linear one.

[0077] After the convolutional operation, to further reduce the computational complexity, feature compression is performed in the pooling layer. Taking max pooling as an example:

[0078]

[0079] where Denote the corresponding element of the k-th output feature map after the pooling operation, and MaxPooling(·) represents taking the maximum value within the receptive field.

[0080] After the feature input enters the fully connected layer, the numerical values will be converted into probabilities through the softmax layer:

[0081]

[0082] Among them, is the feature vector associated with the input; the softmax(·) function maps the feature set into a C-dimensional vector, where the value of each vector is in the range (0, 1), and the sum of all vectors is 1; represents the probabilities of each category; represents the non-linear function relationship fitted by the convolutional neural network.

[0083] The original CNN model is prone to overfitting when there is less training data. By adding a Dropout layer, overfitting can be reduced and the generalization ability of the model can be improved.

[0084] Suppose a neural network with L hidden layers, where each hidden layer l contains n [l] neurons. Define a binary mask vector d [l] for each hidden layer l, and use to indicate whether the i-th neuron is retained. During model training, for each training example t, the Dropout layer will randomly apply the binary mask vector d [l] to the output of each hidden layer l, generating a new loss function L [l] :

[0085]

[0086] Among them W [l] and b [l] are the weights and biases of the l layer, is the network output, is the loss function, and Γ is the number of training samples. It should be noted that during model testing, the Dropout layer keeps the expected output unchanged.

[0087] Finally, multi-class probability output is performed in the Softmax layer to obtain the probability in the infrared image modality.

[0088] S5. Use the model test results to perform decision-level data fusion based on the D-S evidence theory, including: analyzing the thermal conductivity and infrared image data samples to be fused, obtaining all possible propositions after data fusion, and constructing the identification framework in the D-S evidence theory. In the identification framework, assign evidence to each proposition, construct the belief degree of each proposition in the identification framework by combining the belief function and the likelihood function in the D-S evidence theory, use the D-S evidence theory synthesis method to fuse the information provided by multiple evidences, synthesize the belief degrees of each proposition, and each independent model generates a decision output according to its training results, and the output results are probability values and classification results. Then fuse the independent decision results of each modality.

[0089] Specifically, assume that there is a set Θ = {θ1, θ2, … θ n}, and each element in the set represents a possible result of the event. Specifically, n = 4, θ1 represents that cracks occur in the inner insulation board, θ2 represents that water seepage occurs in the inner insulation board, θ3 represents that hollowing occurs in the inner insulation board, and θ4 represents that the performance of the inner insulation board deteriorates. The belief assignment function is expressed as:

[0090]

[0091] where represents all subsets in the set. For the set and A ≠ Θ, if there is m(A) > 0, then A is called a focal element, and m(A) represents the belief value of the evidence source for the proposition corresponding to the set A; represents the empty set, that is, it does not belong to any deterioration type, and and means that the function satisfies both condition 1 and condition 2. Condition 1: The probability that the sample does not belong to any deterioration type is 0. Condition 2: The sum of the probabilities that the sample belongs to other deterioration types is 1.

[0092] Based on the Dempster synthesis rule, fuse different evidence sources for the same set. Taking two evidence sources E1 and E2 as an example, the corresponding belief assignment functions are m1 and m2 respectively, and the corresponding focal elements are A i and B j , then there is:

[0093]

[0094] where ζ is the conflict coefficient, which can describe the conflict situation between evidences. Based on the above fusion rule, the D-S evidence theory makes a more reliable decision by fully considering the influence of different evidence sources on the same proposition.

[0095] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. All technical solutions falling within the scope defined by the claims of the present invention are within the protection scope of the present invention.

Claims

1. A dynamic monitoring system for thermal conductivity of an inner insulation board based on multimodal data, wherein the inner insulation board is arranged inside the wall and a plaster layer is coated on the surface of the inner insulation board, characterized in that: The system comprises: The first temperature sensor is used to collect the temperature T1 between the wall and the inner insulation board; The second temperature sensor is used to collect the temperature T2 of the surface of the plaster layer; Heat flux sensor, used to collect heat flux density q on the surface of the plaster layer; Infrared thermal imager, used to collect infrared images of the surface of the plaster layer; A data collector, which is in communication with the first temperature sensor, the second temperature sensor and the heat flux sensor, and is used to transmit the collected temperatures T1, T2 and heat flux density q to a data processing device; The data processing device is connected to the data collector and the infrared thermal imager for receiving the temperature data T1, T2, the heat flux density data q and the infrared image data, and processing the received data.

2. The dynamic monitoring system for thermal conductivity of inner insulation board based on multimodal data according to claim 1 is characterized in that: The first temperature sensor and the second temperature sensor are both platinum resistance temperature sensors; the first temperature sensor is buried between the wall and the inner insulation board, and the second temperature sensor is attached to the surface of the plaster layer through thermal conductive silicone.

3. The dynamic monitoring system for thermal conductivity of inner insulation board based on multimodal data according to claim 1 is characterized in that: The heat flow sensor is attached to the surface of the plaster layer using vaseline as a coupling agent.

4. A dynamic monitoring method for thermal conductivity of an inner insulation board based on multimodal data, wherein the inner insulation board is arranged inside the wall and a plaster layer is coated on the surface of the inner insulation board, characterized in that: The method comprises: Collect relevant data, including the temperature T1 between the wall and the inner insulation board, the temperature T2 of the surface of the plaster layer, the heat flux density q of the surface of the plaster layer, and the infrared image data of the surface of the plaster layer; The thermal conductivity λ of the inner insulation board is calculated using the collected temperatures T1, T2 and heat flux density q; Data preprocessing: Use adversarial neural network to perform data enhancement processing on thermal conductivity λ and infrared image data; Model training and testing: using the preprocessed data to perform model training respectively, and then using the original data as the input of the trained model respectively, and performing model testing respectively; wherein the original data includes the thermal conductivity λ and infrared image data; The model test results are used to perform decision-making layer data fusion based on DS evidence theory.

5. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to claim 4 is characterized in that: The thermal conductivity λ is calculated according to the following formula: Wherein, δ is the thickness of the inner insulation board.

6. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to claim 4 is characterized in that: The model training and testing specifically include: using the enhanced infrared image data as the input of the convolutional neural network CNN for model training to obtain the trained convolutional neural network CNN; using the enhanced thermal conductivity coefficient λ as the input of the basic classifier for model training to obtain the trained basic classifier; and then using the original infrared image data as the input of the trained CNN and using the original thermal conductivity coefficient λ as the input of the trained basic classifier to perform model testing respectively.

7. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to claim 6 is characterized in that: The convolutional neural network CNN includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and a softmax layer. The enhanced infrared image data is used as the input of the convolutional neural network CNN for model training, which includes: The input layer receives image data; wherein the image data is enhanced infrared image data; Use the convolution layer to perform convolution operations on the input image data to extract local features to form a feature map; Use the pooling layer to downsample the feature map; After several layers of convolutional layers and pooling layers are combined, features are further extracted and compressed; After the compressed features are integrated through the fully connected layer, the softmax layer is used to transform the numerical value into probability, and the final classification result or regression value is output to complete the processing and analysis of the input data by the convolutional neural network.

8. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to claim 7 is characterized in that: For the lth convolutional layer of the convolutional neural network CNN, its output can be expressed as: in, represents the element at the i-th row and j-th column of the k-th output feature map of the l-th layer, represents the element located in the i+q-1th row and j+p-1th column of the kth input feature map of the l-1th layer, Indicates that the element size of the qth row and pth column of the kth convolution kernel in the lth layer is N U ×N U , N U represents the height and width of the convolution kernel, K represents the number of convolution layers, and g k,l is the bias of the kth convolution kernel in the lth layer, * is the convolution operator symbol; f(·) is the activation function, which is used to convert linear operations into nonlinear ones.

9. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to claim 7, characterized in that: The downsampling operation of the feature map using the pooling layer is performed by the following formula: in, represents the corresponding element of the k-th output feature map after the pooling operation, and MaxPooling(·) represents the maximum value within the receptive field; The conversion from numerical value to probability through the softmax layer is performed by the following formula: in, is the feature vector associated with the input; the softmax(·) function maps the feature set into a C-dimensional vector, where the value of each vector is in the range (0,1) and the sum of all vectors is 1; Represents the probability of each category; Represents the nonlinear function relationship fitted by the convolutional neural network.

10. The method for dynamically monitoring the thermal conductivity of an inner insulation board based on multimodal data according to any one of claims 4 to 9, characterized in that: The decision-level data fusion based on DS evidence theory using the model test results includes: analyzing the thermal conductivity and infrared image data samples to be fused, obtaining all propositions that may appear after data fusion, and forming a recognition framework in the DS evidence theory. In the recognition framework, evidence is assigned to each proposition, and the trust of each proposition in the recognition framework is constructed by combining the credibility function and likelihood function in the DS evidence theory. The information provided by multiple evidences is fused using the DS evidence theory synthesis method, and the trust of each proposition is synthesized. Each independent model generates a decision output according to its training result, and the output result is a probability value and a classification result, and the independent decision results of each modality are fused.

Images