Method and equipment for evaluating credibility of multispectral diagnosis result

Abstract

The invention relates to a multi-spectral diagnosis result credibility evaluation method. The method comprises the following steps: synchronously acquiring three-spectral images and environmental parameters of target equipment; the collected three-spectrum image is preprocessed; inputting the preprocessed image into the improved model to obtain a prediction probability distribution vector of fusion features and three-spectral-band branches; obtaining a prediction category and uncertainty entropy; carrying out confidence score calculation; judging whether to trigger a feedback condition; according to the method, the complementary characteristics of three spectrums of visible light, infrared light and ultraviolet light are utilized, and the CBAM attention module is introduced, so that key characteristics under different spectrums are enhanced, and the sensing ability of the model to an abnormal region is improved; the method achieves the prediction result confidence evaluation based on a Monte Carlo Dropout reasoning mechanism, effectively recognizes a low-confidence result, gives a diagnosis conclusion, can also evaluate the confidence, is suitable for deployment in key scenes such as a high-pressure valve hall, and facilitates the improvement of the overall safety and stability of a system.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and intelligent diagnosis technology, and in particular to a method and device for evaluating the reliability of multispectral diagnostic results. Background Technology

[0002] With the development of power systems towards intelligence and digitalization, online real-time monitoring and diagnostic technologies have been widely used in high-voltage equipment inspection. Traditional visible light or infrared monitoring methods, while capable of capturing visual defects and hotspot information, lack sensitivity to very early signs of insulation degradation and weak signals such as partial discharge. Single ultraviolet corona detection, although fast-responding, is susceptible to environmental interference and cannot provide temperature and morphological information. In recent years, multispectral fusion diagnostic technology has made significant progress in fault feature extraction by comprehensively utilizing visible, infrared, and ultraviolet signals. However, it lacks a unified reliability evaluation and conflict resolution mechanism for the resulting multi-channel diagnostic results, leading to false alarms, missed alarms, and difficulties in quantifying model uncertainty. Therefore, there is an urgent need for a technical solution that can assess the confidence level of multispectral diagnostic results and automatically trigger feedback when there is low confidence or inter-channel diagnostic conflict, continuously improving system robustness and early warning accuracy. Summary of the Invention

[0003] To address the problem that existing multispectral diagnostic methods can only output fault categories and cannot evaluate the reliability of results, the primary objective of this invention is to provide a method for assessing the confidence of prediction results and effectively identifying low-confidence results in multispectral diagnostic results.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for evaluating the reliability of multispectral diagnostic results, the method comprising the following sequential steps:

[0005] (1) Synchronously acquire the three-spectral images and environmental parameters of the target device, wherein the three-spectral images include visible light images. Infrared images and ultraviolet images The environmental parameters include light intensity. Ambient temperature T and relative humidity ;

[0006] (2) The acquired three-spectral images are preprocessed to obtain the preprocessed images;

[0007] (3) Input the preprocessed image into the improved model, extract the primary features of the three spectra, and then stitch and fuse them to obtain the fused features. And the predicted probability distribution vector of the three spectral branches;

[0008] (4) Based on fusion characteristics The predicted probability distribution vectors of the three spectral branches are used to perform multiple forward inferences through the Monte Carlo Dropout inference mechanism to obtain the predicted category R and the uncertainty entropy. ;

[0009] (5) Based on the prediction category R and the uncertainty entropy Confidence scores were calculated based on environmental parameters.

[0010] (6) Perform feedback determination to determine whether the feedback condition is triggered. If it is triggered, return to step (1). If the feedback condition is not triggered, end the current diagnosis, output and store the final diagnosis result. The final diagnosis result includes the predicted category R and the confidence score.

[0011] Step (1) specifically refers to: capturing surface structure, color, and appearance information of the power equipment using a visible light camera; detecting the thermal radiation distribution on the surface of the power equipment using an infrared thermal imager; observing corona discharge and partial discharge phenomena using an ultraviolet image detector; and collecting light intensity data using a light sensor. The ambient temperature (T) is read in real time using a temperature sensor, and the relative humidity is obtained using a humidity sensor. .

[0012] In step (2), the preprocessing includes registration, denoising and normalization.

[0013] In step (3), the improved model refers to the integration of a CBAM attention module into a multi-branch convolutional neural network;

[0014] The multi-branch convolutional neural network includes three parallel spectral branches: visible light branch, infrared branch, and ultraviolet branch. Each spectral branch receives the image of the corresponding spectrum as input and outputs a feature map. Each spectral branch is composed of multiple convolutional blocks connected in series. Each convolutional block includes a concatenated layer of convolution, a batch normalization (BN) layer, a ReLU activation function, and a max pooling operation. A CBAM attention module is introduced after each convolutional block. The CBAM attention module is composed of a channel attention submodule and a spatial attention submodule connected in series.

[0015] The preprocessed image is input into the improved model to obtain the fused attention-enhanced feature map extracted from the visible light branch. Fusion attention-enhanced feature maps extracted from infrared branches Infrared branch extracted fusion attention-enhanced feature map , The three-way spectral primary characteristics are:

[0016]

[0017]

[0018]

[0019] In the formula, For CBAM attention module, Let n be the number of convolutional blocks. Input image for visible light, For infrared input images, Input image for ultraviolet light;

[0020] Primary characteristics of three-way spectra By splicing the channels, a splicing feature map is obtained. :

[0021] ;

[0022] In the formula, To perform splicing operations along the channel dimension;

[0023] Then, the cross-channel spatial collaborative attention fusion module is entered to extract fusion features. :

[0024] ;

[0025] In the formula, For activation function, For one Standard convolution operations; Indicates batch normalization;

[0026] Simultaneously, the predicted probability distribution vector of the visible light branch is obtained. :

[0027] , ;

[0028] In the formula, Let i be the probability that the visible light branch is predicted to be of class i. This represents the total number of diagnostic status categories. It is the logits vector of the visible light branch. A fully connected layer for the visible light branch. To flatten a multidimensional feature map into a one-dimensional vector; For activation functions;

[0029] Obtain the predicted probability distribution vector of the infrared branch. :

[0030] ;

[0031] In the formula, It is the logits vector of the infrared branch. It is a fully connected layer for the infrared branch;

[0032] Obtain the predicted probability distribution vector of the ultraviolet branch. :

[0033] ;

[0034] In the formula, It is the logits vector of the ultraviolet branch. It is a fully connected layer with an ultraviolet branch.

[0035] Step (4) specifically includes the following steps in sequence:

[0036] (4a) The Monte Carlo Dropout inference mechanism is used to fuse features. conduct The next random forward propagation, through The forward pass under different masks yields the following results: The predicted probability distribution vector after group fusion:

[0037] ;

[0038] in, Indicates the first The fused predicted probability distribution vector after the first forward propagation. Indicates the number of categories; The improved model; the inference function after randomly discarding some neurons in the t-th iteration. The parameter subset for each sampling includes the retained weights, retained biases, and moving average statistics of the BN layer. t is the sequence number of a single forward propagation, t=1,2,…,T;

[0039] (4b) Perform average prediction distribution: The average of the predicted probability distribution vectors after multiple fusions is used as the final classification probability. :

[0040] ;

[0041] The final predicted category R is:

[0042] ;

[0043] ;

[0044] in, Indicates the first The average predicted probability of the class For visible light branch prediction categories, For infrared branch prediction categories, For ultraviolet branch prediction categories; This represents the total number of diagnostic status categories. Let i be the probability that the visible light branch is predicted to be of class i. Let be the probability of the infrared branch being predicted as class i. The probability of predicting the ultraviolet branch as the i-th class;

[0045] (4c) Perform uncertainty entropy estimation:

[0046] The uncertainty entropy of the classification probability distribution is defined using the Shannon entropy formula. :

[0047] ;

[0048] when Lower values ​​indicate concentrated and highly certain predictions; when A higher value indicates scattered predictions and a lack of confidence.

[0049] Step (5) specifically includes the following steps in sequence:

[0050] (5a) Calculate the spectral consistency score :

[0051] ;

[0052] in, Indicates the maximum similarity; For visible light branch prediction categories, For infrared branch prediction categories, For UV branch prediction categories; both refer to... , , There are exactly two identical elements among the three; the three are completely different. , , They are all different;

[0053] Define the average Euclidean distance difference between the three spectral branches. for:

[0054] ;

[0055] In the formula, Let be the predicted probability distribution vector of the infrared branch. Let be the predicted probability distribution vector of the visible light branch. This is the predicted probability distribution vector for the ultraviolet branch;

[0056] but The calculation formula is:

[0057] ;

[0058] in: This is an adjustable scaling factor. , When the distributions are closer, Approaching 1; when the distribution differences are large, Approaching 0;

[0059] (5b) Calculate the environmental disturbance score :

[0060] ;

[0061] in, Given the current light intensity, Standard lighting; The current temperature. Standard temperature; The current relative humidity, Standard humidity; , , All are positive real hyperparameters that control the sensitivity to environmental influences;

[0062] The range is The closer the value is to 1, the more ideal the environment.

[0063] (5c) Calculate confidence score :

[0064] ;

[0065] in: , , And all of them are confidence-based fusion weights, satisfying:

[0066] ;

[0067] Confidence score The higher the value, the more reliable the current diagnosis.

[0068] In step (6), the feedback condition specifically refers to the following condition: the feedback mechanism is triggered when any of the following conditions are met:

[0069] (6a) ;

[0070] (6b) ;

[0071] (6c) ;

[0072] in, To calculate the overall confidence score, The minimum confidence threshold is set. Scoring for spectral consistency. For uncertainty entropy, The maximum threshold of uncertainty entropy. This represents the maximum similarity.

[0073] Another object of the present invention is to provide an electronic device comprising:

[0074] Processor; and

[0075] A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the multispectral diagnostic result reliability assessment method as described above.

[0076] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the multispectral diagnostic result reliability assessment method as described above.

[0077] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, multispectral fusion and attention enhancement are carried out in parallel, resulting in more comprehensive diagnostic features: The present invention fully utilizes the complementary characteristics of visible light, infrared light, and ultraviolet light, and by introducing the CBAM attention module, it enhances the key features under different spectra, thereby improving the model's ability to perceive abnormal regions; the three spectral features have been weighted and expressed in terms of channel and spatial dimensions before fusion, ensuring that the fused features have stronger fault characterization capabilities and spatial positioning accuracy, so that diagnosis no longer depends on a single information source, improving the model's generalization ability and adaptability; Second, based on Monte Carlo methods... The Monte Carlo Dropout inference mechanism enables confidence assessment of prediction results and effectively identifies low-confidence results: In the inference stage, this invention uses the Monte Carlo Dropout inference mechanism to perform multiple forward propagations on the same fused feature to construct a set of prediction distributions. From this set, the average prediction result and uncertainty information entropy are extracted. In this way, not only is a diagnostic conclusion provided, but its confidence level can also be assessed, identifying whether the model is "confident" in the current sample. This effectively avoids erroneous high-confidence predictions caused by model overfitting or data bias. It is particularly suitable for deployment in critical scenarios such as high-pressure valve halls, and helps to improve the overall security and stability of the system. Attached Figure Description

[0078] Figure 1 This is a flowchart of the method of the present invention;

[0079] Figure 2 This is a schematic diagram of the improved model in this invention;

[0080] Figure 3 This is a flowchart for confidence assessment. Detailed Implementation

[0081] like Figure 1 As shown, a method for assessing and correcting the reliability of multispectral diagnostic results includes the following sequential steps:

[0082] (1) Synchronously acquire the three-spectral images and environmental parameters of the target device, wherein the three-spectral images include visible light images. Infrared images and ultraviolet images The environmental parameters include light intensity. Ambient temperature T and relative humidity ;

[0083] (2) The acquired three-spectral images are preprocessed to obtain the preprocessed images;

[0084] (3) Input the preprocessed image into the improved model, extract the primary features of the three spectra, and then stitch and fuse them to obtain the fused features. And the predicted probability distribution vector of the three spectral branches;

[0085] (4) Based on fusion characteristics The predicted probability distribution vectors of the three spectral branches are used to perform multiple forward inferences through the Monte Carlo Dropout inference mechanism to obtain the predicted category R and the uncertainty entropy. ;

[0086] (5) Based on the prediction category R and the uncertainty entropy Confidence scores were calculated based on environmental parameters.

[0087] (6) Perform feedback determination to determine whether the feedback condition is triggered. If it is triggered, return to step (1). If the feedback condition is not triggered, end the current diagnosis, output and store the final diagnosis result. The final diagnosis result includes the predicted category R and the confidence score.

[0088] Step (1) specifically refers to: capturing surface structure, color, and appearance information of the power equipment using a visible light camera; detecting the thermal radiation distribution on the surface of the power equipment using an infrared thermal imager; observing corona discharge and partial discharge phenomena using an ultraviolet image detector; and collecting light intensity data using a light sensor. The ambient temperature (T) is read in real time using a temperature sensor, and the relative humidity is obtained using a humidity sensor. .

[0089] In step (2), the preprocessing includes registration, denoising and normalization.

[0090] like Figure 2 As shown, in step (3), the improved model refers to the integration of a CBAM attention module into a multi-branch convolutional neural network;

[0091] The multi-branch convolutional neural network includes three parallel spectral branches: visible light branch, infrared branch, and ultraviolet branch. Each spectral branch receives the image of the corresponding spectrum as input and outputs a feature map. Each spectral branch is composed of multiple convolutional blocks connected in series. Each convolutional block includes a concatenated layer of convolution, a batch normalization (BN) layer, a ReLU activation function, and a max pooling operation. A CBAM attention module is introduced after each convolutional block. The CBAM attention module is composed of a channel attention submodule and a spatial attention submodule connected in series.

[0092] The preprocessed image is input into the improved model to obtain the fused attention-enhanced feature map extracted from the visible light branch. Fusion attention-enhanced feature maps extracted from infrared branches Infrared branch extracted fusion attention-enhanced feature map , The three-way spectral primary characteristics are:

[0093]

[0094]

[0095]

[0096] In the formula, For CBAM attention module, Let n be the number of convolutional blocks. Input image for visible light, For infrared input images, Input image for ultraviolet light;

[0097] Primary characteristics of three-way spectra By splicing the channels, a splicing feature map is obtained. :

[0098] ;

[0099] In the formula, To perform splicing operations along the channel dimension;

[0100] Then, the cross-channel spatial collaborative attention fusion module is entered to extract fusion features. :

[0101] ;

[0102] In the formula, For activation function, For one Standard convolution operations; Indicates batch normalization;

[0103] Simultaneously, the predicted probability distribution vector of the visible light branch is obtained. :

[0104] , ;

[0105] In the formula, Let i be the probability that the visible light branch is predicted to be of class i. This represents the total number of diagnostic status categories. It is the logits vector of the visible light branch. A fully connected layer for the visible light branch. To flatten a multidimensional feature map into a one-dimensional vector; For activation functions;

[0106] Obtain the predicted probability distribution vector of the infrared branch. :

[0107] ;

[0108] In the formula, It is the logits vector of the infrared branch. It is a fully connected layer for the infrared branch;

[0109] Obtain the predicted probability distribution vector of the ultraviolet branch. :

[0110] ;

[0111] In the formula, It is the logits vector of the ultraviolet branch. It is a fully connected layer with an ultraviolet branch.

[0112] Step (4) specifically includes the following steps in sequence:

[0113] (4a) The Monte Carlo Dropout inference mechanism is used to fuse features. conduct The next random forward propagation, through The forward pass under different masks yields the following results: The predicted probability distribution vector after group fusion:

[0114] ;

[0115] in, Indicates the first The fused predicted probability distribution vector after the first forward propagation. Indicates the number of categories; The improved model; the inference function after randomly discarding some neurons in the t-th iteration. The parameter subset for each sampling includes the retained weights, retained biases, and moving average statistics of the BN layer. t is the sequence number of a single forward propagation, t=1,2,…,T;

[0116] (4b) Perform average prediction distribution: The average of the predicted probability distribution vectors after multiple fusions is used as the final classification probability. :

[0117] ;

[0118] The final predicted category R is:

[0119] ;

[0120] ;

[0121] in, Indicates the first The average predicted probability of the class For visible light branch prediction categories, For infrared branch prediction categories, For ultraviolet branch prediction categories; This represents the total number of diagnostic status categories. Let i be the probability that the visible light branch is predicted to be of class i. Let be the probability of the infrared branch being predicted as class i. The probability of predicting the ultraviolet branch as the i-th class;

[0122] (4c) Perform uncertainty entropy estimation:

[0123] The uncertainty entropy of the classification probability distribution is defined using the Shannon entropy formula. :

[0124] ;

[0125] when Lower values ​​indicate concentrated and highly certain predictions; when A higher value indicates scattered predictions and a lack of confidence.

[0126] Unlike conventional single-inference, the Monte Carlo Dropout inference mechanism can capture the changes in the responses of different subnetworks in the model parameter space to the same input, thus truly reflecting the ambiguity of classification, making it suitable for use in important scenarios (such as power equipment fault detection).

[0127] like Figure 3 As shown, step (5) specifically includes the following steps in sequence:

[0128] (5a) Calculate the spectral consistency score :

[0129] ;

[0130] in, Indicates the maximum similarity; For visible light branch prediction categories, For infrared branch prediction categories, For UV branch prediction categories; both refer to... , , There are exactly two identical elements among the three; the three are completely different. , , They are all different;

[0131] Define the average Euclidean distance difference between the three spectral branches. for:

[0132] ;

[0133] In the formula, Let be the predicted probability distribution vector of the infrared branch. Let be the predicted probability distribution vector of the visible light branch. This is the predicted probability distribution vector for the ultraviolet branch;

[0134] but The calculation formula is:

[0135] ;

[0136] in: This is an adjustable scaling factor. , When the distributions are closer, Approaching 1; when the distribution differences are large, Approaching 0;

[0137] (5b) Calculate the environmental disturbance score :

[0138] ;

[0139] in, Given the current light intensity, Standard lighting; The current temperature. Standard temperature; The current relative humidity, Standard humidity; , , All are positive real hyperparameters that control the sensitivity to environmental influences;

[0140] The range is The closer the value is to 1, the more ideal the environment.

[0141] Fluctuations in environmental parameters can affect the quality of infrared and ultraviolet images, especially in cases of excessively high temperatures or insufficient illumination, which can easily lead to false detections.

[0142] (5c) Calculate confidence score :

[0143] ;

[0144] in: , , And all of them are confidence-based fusion weights, satisfying:

[0145] ;

[0146] Confidence score The higher the value, the more reliable the current diagnosis.

[0147] In step (6), the feedback condition specifically refers to the following condition: the feedback mechanism is triggered when any of the following conditions are met:

[0148] (6a) The confidence score is lower than the set threshold;

[0149] (6b) Significant inconsistencies exist in diagnostic results across different spectral bands;

[0150] (6c) This indicates that the forecasts are too scattered and have high uncertainty.

[0151] in, To calculate the overall confidence score, The minimum confidence threshold is set. Scoring for spectral consistency. For uncertainty entropy, The maximum threshold of uncertainty entropy. This represents the maximum similarity.

[0152] If the confidence level is insufficient or there is a spectral conflict, a feedback condition is triggered, thereby forming an intelligent closed loop, reducing the misdiagnosis rate, the missed diagnosis rate and enhancing interpretability.

[0153] In summary, this invention employs multispectral fusion and attention enhancement in parallel, resulting in more comprehensive diagnostic features. It fully utilizes the complementary characteristics of visible, infrared, and ultraviolet light spectra, and by introducing a CBAM attention module, it strengthens key features under different spectra, thereby enhancing the model's ability to perceive abnormal regions. Before fusion, the three spectral features are weighted by channel and spatial dimensions, ensuring that the fused features possess stronger fault characterization capabilities and spatial positioning accuracy. This eliminates reliance on a single information source for diagnosis, improving the model's generalization ability and adaptability. Furthermore, it uses a Monte Carlo Dropout inference mechanism to assess the confidence of prediction results, effectively identifying low-confidence results. During the inference phase, this invention uses the Monte Carlo Dropout inference mechanism to perform multiple forward propagations on the same fused feature, constructing a prediction distribution set from which the average prediction result and uncertainty entropy are extracted. This approach not only provides diagnostic conclusions but also assesses their confidence, identifying whether the model is "confident" in the current sample. It effectively avoids erroneous high-confidence predictions caused by model overfitting or data bias, making it particularly suitable for deployment in critical scenarios such as high-pressure valve rooms, contributing to improved overall system security and stability.

[0154] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for assessing the reliability of multispectral diagnostic results, characterized in that: The method includes the following steps in sequence: (1) Synchronously acquire the three-spectral images and environmental parameters of the target device, wherein the three-spectral images include visible light images. Infrared images and ultraviolet images The environmental parameters include light intensity. Ambient temperature T and relative humidity ; (2) The acquired three-spectral images are preprocessed to obtain the preprocessed images; (3) Input the preprocessed image into the improved model, extract the primary features of the three spectra, and then stitch and fuse them to obtain the fused features. And the predicted probability distribution vector of the three spectral branches; (4) Based on fusion characteristics The predicted probability distribution vectors of the three spectral branches are used to perform multiple forward inferences through the Monte Carlo Dropout inference mechanism to obtain the predicted category R and the uncertainty entropy. ; (5) Based on the prediction category R and the uncertainty entropy Confidence scores were calculated based on environmental parameters; (6) Perform feedback determination to determine whether the feedback condition is triggered. If it is triggered, return to step (1). If the feedback condition is not triggered, end the current diagnosis, output and store the final diagnosis result. The final diagnosis result includes the predicted category R and the confidence score.

2. The method for assessing the reliability of multispectral diagnostic results according to claim 1, characterized in that: Step (1) specifically refers to: capturing surface structure, color, and appearance information of the power equipment using a visible light camera; detecting the thermal radiation distribution on the surface of the power equipment using an infrared thermal imager; observing corona discharge and partial discharge phenomena using an ultraviolet image detector; and collecting light intensity data using a light sensor. The ambient temperature (T) is read in real time using a temperature sensor, and the relative humidity is obtained using a humidity sensor. .

3. The method for assessing the reliability of multispectral diagnostic results according to claim 1, characterized in that: In step (2), the preprocessing includes registration, denoising and normalization.

4. The method for evaluating the reliability of multispectral diagnostic results according to claim 1, characterized in that: In step (3), the improved model refers to the integration of a CBAM attention module into a multi-branch convolutional neural network; The multi-branch convolutional neural network includes three parallel spectral branches: visible light branch, infrared branch, and ultraviolet branch. Each spectral branch receives the image of the corresponding spectrum as input and outputs a feature map. Each spectral branch is composed of multiple convolutional blocks connected in series. Each convolutional block includes a concatenated layer of convolution, a batch normalization (BN) layer, a ReLU activation function, and a max pooling operation. A CBAM attention module is introduced after each convolutional block. The CBAM attention module is composed of a channel attention submodule and a spatial attention submodule connected in series. The preprocessed image is input into the improved model to obtain the fused attention-enhanced feature map extracted from the visible light branch. Fusion attention-enhanced feature maps extracted from infrared branches Infrared branch extracted fusion attention-enhanced feature map , The three-way spectrum consists of primary characteristics: ； ； ； In the formula, For CBAM attention module, Let n be the number of convolutional blocks, and n be the total number of convolutional blocks. Input image for visible light, For infrared input images, Input image for ultraviolet light; Primary characteristics of three-way spectra By splicing the channels, a splicing feature map is obtained. : ； In the formula, To perform splicing operations along the channel dimension; Then, the cross-channel spatial collaborative attention fusion module is entered to extract fusion features. : ； In the formula, For activation function, For one Standard convolution operations; Indicates batch normalization; Simultaneously, the predicted probability distribution vector of the visible light branch is obtained. : ， ； In the formula, Let i be the probability that the visible light branch is predicted to be of class i. This represents the total number of diagnostic status categories. It is the logits vector of the visible light branch. A fully connected layer for the visible light branch. To flatten a multidimensional feature map into a one-dimensional vector; For activation functions; Obtain the predicted probability distribution vector of the infrared branch. : ； In the formula, It is the logits vector of the infrared branch. It is a fully connected layer for the infrared branch; Obtain the predicted probability distribution vector of the ultraviolet branch. : ； In the formula, It is the logits vector of the ultraviolet branch. It is a fully connected layer with an ultraviolet branch.

5. The method for assessing the reliability of multispectral diagnostic results according to claim 1, characterized in that: Step (4) specifically includes the following steps in sequence: (4a) The Monte Carlo Dropout inference mechanism is used to fuse features. conduct The next random forward propagation, through The forward pass under different masks yields the following results: The predicted probability distribution vector after group fusion: ； in, Indicates the first The fused predicted probability distribution vector after the first forward propagation. Indicates the number of categories; The improved model; the inference function after randomly discarding some neurons in the t-th iteration. The parameter subset for each sampling includes the retained weights, retained biases, and moving average statistics of the BN layer. t is the sequence number of a single forward propagation, t=1,2,…,T; (4b) Perform average prediction distribution: The average of the predicted probability distribution vectors after multiple fusions is used as the final classification probability. : ； The final predicted category R is: ； ； in, Indicates the first The average predicted probability of the class For visible light branch prediction categories, For infrared branch prediction categories, For ultraviolet branch prediction categories; This represents the total number of diagnostic status categories. Let i be the probability that the visible light branch is predicted to be of class i. Let be the probability of the infrared branch being predicted as class i. The probability of predicting the ultraviolet branch as the i-th class; (4c) Perform uncertainty entropy estimation: The uncertainty entropy of the classification probability distribution is defined using the Shannon entropy formula. : ； when Lower values ​​indicate concentrated and highly certain predictions; when A higher value indicates scattered predictions and a lack of confidence.

6. The method for evaluating the reliability of multispectral diagnostic results according to claim 1, characterized in that: Step (5) specifically includes the following steps in sequence: (5a) Calculate the spectral consistency score : ； in, Indicates the maximum similarity; For visible light branch prediction categories, For infrared branch prediction categories, For UV branch prediction categories; both refer to... , , There are exactly two identical elements among the three; the three are completely different. , , They are all different; Define the average Euclidean distance difference between the three spectral branches. for: ； In the formula, Let be the predicted probability distribution vector of the infrared branch. Let be the predicted probability distribution vector of the visible light branch. This is the predicted probability distribution vector for the ultraviolet branch; but The calculation formula is: ； in: This is an adjustable scaling factor. , When the distributions are closer, Approaching 1; when the distribution differences are large, Approaching 0; (5b) Calculate the environmental disturbance score : ； in, Given the current light intensity, Standard lighting; The current temperature. Standard temperature; The current relative humidity, Standard humidity; , , All are positive real hyperparameters that control the sensitivity to environmental influences; The range is The closer the value is to 1, the more ideal the environment. (5c) Calculate confidence score : ； in: , , And all of them are confidence-based fusion weights, satisfying: ； Confidence score The higher the value, the more reliable the current diagnosis.

7. The method for assessing the reliability of multispectral diagnostic results according to claim 1, characterized in that: In step (6), the feedback condition specifically refers to the following condition: the feedback mechanism is triggered when any of the following conditions are met: （6a） ； （6b） ； （6c） ； in, To calculate the overall confidence score, The minimum confidence threshold is set. Scoring for spectral consistency. For uncertainty entropy, The maximum threshold of uncertainty entropy. This represents the maximum similarity.

8. An electronic device, comprising: processor; as well as A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the multispectral diagnostic result reliability assessment method as described in any one of claims 1-7.

9. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the multispectral diagnostic result reliability assessment method as described in any one of claims 1-7.

Images