Nuclear-grade SLM equipment fault diagnosis result credibility quantification method based on conformal prediction

Abstract

The invention relates to the technical field of nuclear-grade SLM equipment fault diagnosis, and discloses a nuclear-grade SLM equipment fault diagnosis result credibility quantification method based on conformal prediction, and the method comprises the steps: carrying out the preprocessing of a nuclear-grade SLM equipment vibration signal, and constructing a training set, a calibration set and a test set which comprise distributed inner and outer samples; finely adjusting the fault diagnosis model through a feature separation loss function containing a separation loss item and an aggregation loss item; extracting a high-dimensional feature vector of the test sample and calculating a non-consistency measurement score; determining a class condition quantile threshold value and a label ranking threshold value of each class based on the calibration set; screening fault category labels meeting conditions to generate a prediction set; and outputting a diagnosis result credibility quantitative index. According to the method, the out-of-distribution samples can be accurately identified, the size of the rare sample prediction set under the unbalanced data is optimized, the coverage rate and the quantitative practicability are considered, and a reliable basis is provided for equipment maintenance decision making.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for nuclear-grade SLM equipment, specifically to a method for quantifying the reliability of fault diagnosis results for nuclear-grade SLM equipment based on conformal prediction. Background Technology

[0002] In the production and operation of nuclear-grade SLM (Selective Laser Melting) equipment, the chuck, as a core and critical component, directly impacts the accuracy of its fault diagnosis results, affecting equipment production efficiency and the scientific basis of maintenance decisions. With the development of deep learning technology, fault diagnosis methods for nuclear-grade SLM equipment based on this technology have achieved end-to-end diagnosis, improving diagnostic efficiency and accuracy to some extent. However, limited by the black-box nature of deep learning models, existing methods cannot effectively quantify the uncertainty of diagnostic results. Even when faced with samples that are difficult to classify accurately, the model may still output blindly confident diagnostic results, making it impossible for users to judge the reliability of the results, thus affecting the rationality of maintenance planning and production scheduling.

[0003] Nuclear-grade SLM equipment operates under complex conditions with large variations in speed and load, making it highly susceptible to off-distribution fault samples not included in the training data in practical applications. Existing conformal prediction methods rely solely on model output for calibration due to inconsistency metrics, failing to capture the characteristic features of off-distribution samples and leading to a failure of the prediction set, thus hindering the reliable quantification of diagnostic uncertainty for such samples. Furthermore, nuclear-grade SLM equipment vibration signal data exhibits a significant imbalance in the number of fault types, with the vast majority being normal data and some faults, such as missing rolling elements or eccentric wear, being extremely rare. Existing uncertainty quantification methods often prioritize global marginal coverage, tending to assign excessively large prediction sets to rare samples. While this ensures coverage, it negates the practical application significance of uncertainty quantification.

[0004] Existing conformal prediction techniques such as APS, RAPS, and CCP methods have significant limitations when dealing with out-of-distribution sample detection and imbalanced data scenarios. They lack accuracy in identifying out-of-distribution samples, have limited optimization effects on rare sample prediction sets, and struggle to improve prediction efficiency while ensuring coverage. They also fail to meet the actual need for accurate quantification of result reliability in fault diagnosis of nuclear-grade SLM equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a reliable quantification method for fault diagnosis results of nuclear-level SLM equipment based on conformal prediction, so as to solve the problems mentioned in the background art.

[0006] According to one aspect of this application, a method for quantifying the reliability of fault diagnosis results for nuclear-grade SLM equipment based on conformal prediction is provided, comprising the following steps: S1, preprocess the collected vibration signals of the nuclear-grade SLM equipment to construct a sample set including a training set, a calibration set and a test set, wherein the sample set includes in-distribution samples and out-of-distribution samples; S2, a fault diagnosis model is trained based on the training set, and the fault diagnosis model is fine-tuned using a feature separation loss function; the feature separation loss function includes a separation loss term and a clustering loss term, the separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model, and the clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class; S3. Using the fine-tuned fault diagnosis model, extract the high-dimensional feature vector of the test sample; S4. Based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model, calculate the inconsistency metric score of the test sample; the inconsistency metric score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector; S5, Calculate the class conditional quantile threshold and label ranking threshold for each fault category based on the calibration set; S6. Based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold, generate a prediction set for the test sample. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. S7. Based on the prediction set, output the reliability metrics of the diagnostic results of the test sample.

[0007] Preferably, in step S2, for out-of-distribution samples, the average value of the absolute cosine similarity between their normalized feature vector and all normalized class weight vectors is calculated as the separation loss term; for in-distribution samples, the negative value of the cosine similarity between their normalized feature vector and the normalized class weight vector corresponding to their real label is calculated as the aggregation loss term; the feature separation loss function is the weighted sum of the separation loss term and the aggregation loss term.

[0008] Preferably, in step S2, when fine-tuning the fault diagnosis model, the weight coefficients of the separation loss term and the aggregation loss term in the feature separation loss function are dynamically adjusted, and the adjustment is based on the out-of-distribution sample recognition performance and intra-class feature aggregation degree on the calibration set.

[0009] Preferably, in step S5, the method for calculating the class conditional quantile threshold includes: for each fault category, filtering calibration samples belonging to that fault category based on the true label of the calibration sample; calculating the inconsistency metric score corresponding to the true label for each calibration sample based on the fault diagnosis model; and calculating the quantile of the inconsistency metric scores of all calibration samples in that category as the class conditional quantile threshold for that category.

[0010] Preferably, in step S5, the method for determining the label ranking threshold includes, for each fault category, based on the prediction results of the fault diagnosis model on the calibration set, calculating the probability that the true label of the sample in that category appears in the top k positions of the model's predicted probability; and determining the minimum k value that meets the preset coverage requirement as the label ranking threshold for that category.

[0011] Preferably, in step S6, generating the prediction set specifically includes: for the test sample, obtaining the predicted probability ranking of all fault categories output by the fault diagnosis model; for each fault category label in the ranking, determining whether it satisfies the two conditions in step S6; and taking the set of all fault category labels that satisfy the two conditions as the prediction set.

[0012] In another aspect, this application also provides a system for quantifying the reliability of fault diagnosis results for nuclear-level SLM equipment based on conformal prediction, comprising: The preprocessing module is used to preprocess the acquired vibration signals of the nuclear-grade SLM equipment and construct a sample set including a training set, a calibration set, and a test set. The sample set includes in-distribution samples and out-of-distribution samples. The model fine-tuning module is used to train the fault diagnosis model based on the training set and fine-tune the fault diagnosis model through the feature separation loss function. The feature separation loss function includes a separation loss term and a clustering loss term. The separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model. The clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class. The feature extraction module is used to extract high-dimensional feature vectors of test samples using the fine-tuned fault diagnosis model. The inconsistency measurement module is used to calculate the inconsistency measurement score of the test sample based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model. The inconsistency measurement score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector. The threshold calculation module is used to calculate the class conditional quantile threshold and the label ranking threshold for each fault category based on the calibration set. The prediction set generation module is used to generate a prediction set for the test sample based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. The result output module is used to output a reliability metric for the diagnostic results of the test sample based on the prediction set.

[0013] This application also provides a computer device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the conformal prediction-based core-level SLM device fault diagnosis result reliability quantification method as described above.

[0014] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement the method for quantifying the reliability of fault diagnosis results for core-level SLM devices based on conformal prediction as described above.

[0015] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for quantifying the reliability of fault diagnosis results for core-level SLM devices based on conformal prediction.

[0016] This application employs a dual-loss collaborative design of feature separation loss function to constrain out-of-distribution sample features to an orthogonal subspace of class weight vectors, while simultaneously strengthening the clustering of features of similar samples within the distribution. This significantly improves the recognition accuracy of out-of-distribution samples, achieving an AUROC of up to 95.77%, effectively solving the problem of inaccurate quantification of uncertainty in out-of-distribution samples in existing methods. By combining dual screening with class conditional quantile thresholds and label ranking thresholds, the prediction set size for rare samples in imbalanced data is significantly reduced while ensuring the coverage requirements of target confidence for each category. The prediction set size is as low as 1.65 when ρ=0.1, overcoming the shortcomings of traditional methods in terms of quantification practicality. The overall method takes into account both global and class conditional coverage, exhibiting strong robustness to background noise and diverse fault characteristics under complex operating conditions of nuclear-grade SLM equipment. The quantification results are accurate and reliable, providing a scientific basis for equipment maintenance decisions and significantly improving production efficiency and maintenance rationality. Attached Figure Description

[0017] Figure 1 A schematic diagram illustrating a method for quantifying the reliability of fault diagnosis results for core-level SLM equipment based on conformal prediction, provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the model fine-tuning curve provided in the embodiments of this disclosure; Figure 3 A schematic diagram of the class-conditional quantile threshold calculation process provided in the embodiments of this disclosure; Figure 4 This is a schematic diagram of the tag ranking threshold calculation process provided in the embodiments of this disclosure; Figure 5 A comparison chart of coverage and prediction set size provided for embodiments of this disclosure; Figure 6 A schematic diagram of the process for generating a prediction set provided in an embodiment of this disclosure; Figure 7 A schematic diagram of a system structure for quantifying the reliability of fault diagnosis results of core-level SLM equipment based on conformal prediction, provided in an embodiment of this disclosure; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this disclosure are information and data authorized by the user or fully authorized by all parties.

[0020] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for quantifying the reliability of fault diagnosis results of core-level SLM equipment based on conformal prediction, comprising the following method steps: S1, preprocess the collected vibration signals of the nuclear-grade SLM equipment to construct a sample set including a training set, a calibration set and a test set, wherein the sample set includes in-distribution samples and out-of-distribution samples; S2, a fault diagnosis model is trained based on the training set, and the fault diagnosis model is fine-tuned using a feature separation loss function; the feature separation loss function includes a separation loss term and a clustering loss term, the separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model, and the clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class; S3. Using the fine-tuned fault diagnosis model, extract the high-dimensional feature vector of the test sample; S4. Based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model, calculate the inconsistency metric score of the test sample; the inconsistency metric score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector; S5, Calculate the class conditional quantile threshold and label ranking threshold for each fault category based on the calibration set; S6. Based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold, generate a prediction set for the test sample. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. S7. Based on the prediction set, output the reliability metrics of the diagnostic results of the test sample.

[0021] This method is applicable to fault diagnosis scenarios for nuclear-grade SLM equipment, and is particularly suitable for situations where complex operating conditions, such as large variations in speed and load, lead to frequent off-site fault samples, and where the number of fault types in vibration signal data is unbalanced, with some fault samples being extremely rare. It is understood that this method is suitable for equipment equipped with nuclear-grade SLM equipment vibration signal acquisition devices capable of acquiring vibration signals under different operating conditions, including normal and various fault types such as outer race faults, rolling element faults, cage faults, inner race faults, and combined faults; and equipped with computing equipment that meets the model training requirements for model pre-training and fine-tuning.

[0022] In some embodiments, for step S1, the vibration signal of the nuclear-grade SLM device is preprocessed. The core purpose is to convert the original signal into sample data that meets the model input requirements and to complete the sample set division.

[0023] Specifically, an overlapping sliding window is used to segment the original vibration signal. For example, the window size is set to 1024 and the step size to 400. This parameter setting ensures sufficient signal coverage and effective segmentation, avoiding information omission and redundancy. The generated sample data after segmentation includes in-distribution samples and out-of-distribution samples. The in-distribution samples are normal and known fault type samples that have been included in the training scope, while the out-of-distribution samples are fault samples that have not been included in the training data.

[0024] The sample set was then divided into a training set, a calibration set, and a test set. The calibration set comprised 10% of the labeled data. The training set was used for model pre-training and fine-tuning, the calibration set was used to calculate the class conditional quantile threshold and the label ranking threshold, and the test set was used to verify the effectiveness of the method. Simultaneously, based on the divided samples, an out-of-distribution sample set and an imbalanced sample set were constructed. The out-of-distribution sample set included auxiliary out-of-distribution training samples and completely out-of-distribution test samples. The imbalanced sample set was configured with a predetermined ratio of rare to majority samples.

[0025] In some embodiments, for step S2, the fault diagnosis model in this embodiment is built based on contrastive learning, and its core structure includes a feature extraction network, a fully connected layer, and an output layer.

[0026] The feature extraction network employs a deep convolutional neural network architecture, consisting of multiple convolutional layers, pooling layers, and activation functions. Convolutional layers extract local features from the vibration signal, pooling layers reduce feature dimensionality while retaining key information, and the ReLU activation function enhances the model's nonlinear fitting capability. The output of the feature extraction network is connected to a fully connected layer, which maps the extracted high-dimensional features to the category space. The weight vector of the fully connected layer provides the basis for subsequent feature separation and inconsistency measurement. The output layer uses the SoftMax activation function to transform the output of the fully connected layer into predicted probability distributions for each category, achieving a preliminary judgment of the fault category.

[0027] According to embodiments of this disclosure, a fault diagnosis model is pre-trained based on contrastive learning. Pre-training allows the model to learn general feature representations of vibration signals under unsupervised or weakly supervised conditions, thereby improving the model's ability to capture fault features. During pre-training, a large number of unlabeled vibration signal samples are used, and training is performed by constructing sample pairs.

[0028] Specifically, data augmentation is performed on each original sample to generate positive sample pairs, while other samples are randomly selected as negative sample pairs. The model's pre-training loss function uses contrastive loss, which learns discriminative feature representations by minimizing the distance between positive sample pairs and maximizing the distance between negative sample pairs. For example, the pre-training cycle is set to 400 cycles. In each cycle, a batch of sample pairs is input into the model, and the model parameters are updated through backpropagation until the pre-training loss converges, completing the model's pre-training process.

[0029] According to a preferred embodiment of this disclosure, the feature separation loss function includes a separation loss term and a clustering loss term, and its construction process is as follows: First, define the relevant parameters and vectors. Let the normalized feature vector of the out-of-distribution samples be... This vector is extracted from the pre-trained model and then normalized. L2 normalization is used to ensure that the magnitude of the feature vector is 1. The fully connected layer of the fault diagnosis model contains Each category weight vector corresponds to... The fault categories include normal categories. The weight vector of each category is normalized to obtain a normalized category weight vector. ( (and also satisfy) .

[0030] For the separation loss term Its function is to make the feature vectors of out-of-distribution samples orthogonal to the weight vectors of all classes in the fully connected layer, thereby achieving effective separation of out-of-distribution samples from in-distribution samples in the feature space. The normalized feature vectors of out-of-distribution samples are calculated. With each normalized class weight vector The absolute value of the cosine similarity between the categories is taken, and then the average of the absolute values ​​of the cosine similarity across all categories is taken to obtain the separation loss term, which is expressed as:

[0031] in, Representing the eigenvector transpose, For vectors and The dot product of the two vectors is equal to the cosine of the angle between them. By taking the absolute value and averaging, it can reflect the overall orthogonality between the out-of-distribution feature vectors and the weight vectors of all categories. The smaller the value, the more the features of the out-of-distribution samples tend to be distributed in a dimension orthogonal to the class weight vector, and the better the separation effect.

[0032] For aggregation loss term Its function is to closely align the feature vectors of samples within the distribution with the class weight vectors of their corresponding categories, thereby increasing the clustering of features among samples of the same category. Let the true labels of the samples within the distribution be... Its normalized eigenvector is ( The corresponding normalized class weight vector is ,calculate and The negative value of the cosine similarity between the terms yields the clustering loss term, expressed as:

[0033] The larger the value, the higher the alignment between the sample features and the corresponding class weight vector within the distribution. The smaller the value of , the more likely it is that by minimizing this loss term, feature vectors within the same category can be encouraged to cluster towards the corresponding category weight vector, thus enhancing the consistency of intra-class features.

[0034] The feature separation loss function is a weighted sum of the separation loss term and the clustering loss term, and its expression is:

[0035] in, The mapping function representing the fault diagnosis model. This indicates the expected calculation. Indicates in-distribution samples Follow the data distribution , Indicates samples outside the auxiliary distribution Follow the data distribution , and These are weighting coefficients used to balance the importance of the aggregation loss term and the separation loss term. In the initial experiments, and Set all to 1.

[0036] According to an embodiment of this disclosure, the pre-trained fault diagnosis model is fine-tuned based on the constructed feature separation loss function. For example, the fine-tuning period is set to 200 periods, and the specific process is as follows: The training set and the auxiliary out-of-distribution sample set are input into the pre-trained model. The model extracts feature vectors from the in-distribution and out-of-distribution samples using a feature extraction network. After normalization, the feature vectors are obtained. and The fully connected layer outputs the corresponding class weight vector. Calculate the aggregation loss term according to the above formula. and separation loss term The total loss function value is then obtained. A stochastic gradient descent (SGD) optimizer is used, with a learning rate of 0.001, momentum of 0.9, and weight decay coefficient of 0.0001. The gradient of the loss function with respect to the parameters of each layer of the model is calculated using the backpropagation algorithm, and the parameters of the convolutional layers, fully connected layers, etc., are updated accordingly. After each fine-tuning cycle, the loss value of the model on the validation set is calculated. When the loss value on the validation set does not decrease significantly for 10 consecutive cycles (e.g., the decrease is less than 1e-4), the model fine-tuning is considered to have converged, and the fine-tuning process is stopped.

[0037] Please see Figure 2 , Figure 2 This is a schematic diagram of the model fine-tuning curve provided in the embodiments of this disclosure. The horizontal axis represents the fine-tuning period, which is a total of 200 periods, and the vertical axis represents the loss value. Based on the pre-trained contrastive learning diagnostic model, pre-divided out-of-distribution sample data is added to the test set. Due to the scarcity of labeled data, a small amount of auxiliary out-of-distribution data is added only in the supervised fine-tuning stage, and the designed separation loss is added to the training loss. The curve shows the trend of the model loss after adding the feature separation loss, reflecting the convergence effect of model fine-tuning.

[0038] Preferably, during model fine-tuning, the weight coefficients can be dynamically adjusted based on the out-of-distribution sample recognition performance and intra-class feature clustering on the calibration set. and This further enhances the adaptability of the model.

[0039] Specifically, meta-learning or Bayesian optimization methods are selected as the tuning algorithm, with the AUROC value of out-of-distribution sample identification and intra-class feature clustering as the optimization objectives. Intra-class feature clustering is characterized by calculating the average distance between the feature vectors of all samples within the same class; the smaller the average distance, the higher the clustering. During fine-tuning, the weight coefficients are adjusted every 20 cycles, based on the current... and The optimal AUROC value and intra-class feature clustering of the model on the calibration set are used to search for the optimal model using a Bayesian optimization method. and Values, The search range is [0.5, 2.0]. The search range is [0.5, 2.0]. After adjustment, the new... and Substitute the loss function and continue fine-tuning the model until the fine-tuning cycle ends. AUROC (Area Under the Receiver Operating Characteristic Curve) is a metric for measuring model classification performance, particularly suitable for identifying abnormal / target samples in imbalanced data, binary classification, or multi-class classification tasks. Its value ranges from 0 to 1, with values ​​closer to 1 indicating better model performance.

[0040] This solves the technical problem that existing conformal prediction methods cannot effectively capture the characteristics of out-of-distribution samples, leading to inaccurate quantification of the uncertainty of out-of-distribution samples. Through the design of the feature separation loss function and model fine-tuning, the features of out-of-distribution samples are orthogonal to the class weight vector, while the features of in-distribution samples are clustered in the same category. This significantly improves the distinguishing ability between out-of-distribution and in-distribution samples. Experimental verification shows that the recognition precision of out-of-distribution samples reaches 93.01%, the recall rate reaches 94.58%, and the AUROC index reaches a maximum of 95.77%, outperforming existing methods such as RAPS. RAPS, short for Randomized Adaptive Prediction Sets, is an existing conformal prediction technique used to quantify the uncertainty of diagnostic results in classification tasks. It achieves marginal coverage of new data points by constructing a prediction set, thus ensuring the target confidence level 1-α of the overall data.

[0041] In some embodiments, for step S3, the high-dimensional feature vector of the test sample is extracted using the fine-tuned fault diagnosis model. Specifically, the test sample is input into the fine-tuned model, and the sample is processed through convolutional layers, pooling layers, etc., of the feature extraction network to extract a high-dimensional feature representation containing fault features. This feature vector is the output of the penultimate layer of the model, preserving the geometric structure information of the feature space and providing data support for the subsequent calculation of the inconsistency metric score. The dimension of the extracted high-dimensional feature vector is consistent with the input dimension of the fully connected layer of the model, typically set to 1024 dimensions. After feature vector extraction, no further processing is required; it is directly used for the calculation in step S4.

[0042] In some embodiments, for step S4, based on the high-dimensional feature vector of the test sample and the category weight vector of the fault diagnosis model, the non-consistency metric score of the test sample is calculated. This score is used to characterize the degree of deviation between the test sample and the samples in the distribution, providing a basis for the subsequent generation of the prediction set.

[0043] Specifically, we obtain the predicted probability distribution of the test samples output by the fine-tuned model, assuming that the predicted probabilities of each category output by the model are... ,in Indicates that the test sample belongs to the first... The probability of a fault class. Simultaneously, the high-dimensional feature vector of the test sample extracted in step S3 is obtained, and after L2 normalization, it is obtained... ( ), and the normalized class weight vector of the fully connected layer of the model. ( ).

[0044] The inconsistency metric score is obtained by adding the maximum value of the probability distribution output by the model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and the weight vectors of each class. Its expression is:

[0045] in, Indicates the test sample. The mapping function representing the fault diagnosis model. The maximum SoftMax probability output by the model reflects the model's confidence in predicting the class of the test sample. The mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector of the test sample and the weight vectors of all categories reflects the degree of deviation between the features of the test sample and the features of the samples within the distribution.

[0046] For example, the specific calculation includes, firstly, calculating the predicted probabilities for each category output by the model. Perform SoftMax normalization and calculate the SoftMax probability value for each category, i.e. The maximum value among them is selected as the first part of the score. Next, the normalized feature vector of the test samples is calculated. With each normalized class weight vector dot product Sum the absolute values ​​and then divide by the number of categories. This yields the second part of the score. Finally, the two scores are added together to obtain the inconsistency measure score of the test sample. .

[0047] Therefore, the construction of the inconsistency metric score is shifted from the model output space to the feature space, making full use of the rich information within the feature space. This solves the technical problem of existing conformal prediction methods that rely solely on model output calibration for the inconsistency metric function, failing to effectively distinguish between out-of-distribution and in-distribution samples. By combining the model prediction confidence with the degree of deviation from the feature space, the inconsistency metric score can more accurately reflect the uncertainty of the test samples, providing a reliable basis for the accurate generation of subsequent prediction sets and ensuring that the uncertainty of out-of-distribution samples can be effectively quantified.

[0048] In some embodiments, for step S5, the calculation of the class conditional quantile threshold is based on a calibration set, with the aim of setting a specific threshold for each fault category to ensure that each category meets the target confidence level. The conditions for coverage requirements.

[0049] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the process of calculating the conditional quantile threshold provided in an embodiment of this disclosure. Specifically, In S301, category filtering is performed. This is done for each fault category. ( Based on the true labels of the samples in the calibration set, all calibration samples belonging to this fault category are selected to form a dedicated calibration sample set for this category. ,set up Includes One sample.

[0050] In S302, inconsistency metric scores are calculated using a dedicated calibration sample set. For each sample input into the fine-tuned fault diagnosis model, following steps S3 and S4, the high-dimensional feature vector of each sample is extracted and its inconsistency metric score is calculated. ( ),in For the first The inconsistency metric score corresponding to the true label of each calibration sample.

[0051] In S303, quantile calculation is performed using a category-specific calibration sample set. All inconsistent measure scores Sort in ascending order to obtain the sorted fraction sequence. Based on the preset target confidence level. (generally (i.e., target confidence level of 90%), calculate quantile positions. ,in This indicates rounding up. It takes the integer part of the sorted score sequence. The score of the digits is used as the class-conditional quantile threshold for that fault category. ,Right now .

[0052] The mathematical expression for class conditional quantiles is:

[0053] in, For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. This formula means finding the minimum... This ensures that there are at least 10 ... The proportion of the calibration sample's inconsistency measure score is no greater than This ensures that the condition coverage requirements for this category are met.

[0054] For example, if a specific calibration sample set for a certain fault category contains 100 samples and the target confidence level is 90%, then the quantile position... The inconsistency measure score ranked 91st after sorting is taken as the class condition quantile threshold for that category. This threshold ensures that at least 90% of the calibration samples in that category have an inconsistency measure score no greater than this threshold, thus satisfying the class condition coverage requirement.

[0055] In some embodiments, the label ranking threshold is also calculated based on a calibration set, with the aim of determining whether the true label appears before the model's predicted probability in each fault category. The smallest bit The value is used to filter out low-probability tags with lower rankings and optimize the size of the prediction set.

[0056] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating the tag ranking threshold calculation process provided in an embodiment of this disclosure. Specifically, In S401, the predicted probabilities are ranked. The calibration set is input into the fine-tuned fault diagnosis model, and the predicted probability distribution for each category of each calibration sample output by the model is obtained. The predicted probabilities of each sample are then sorted in descending order to obtain the ranking of the predicted probabilities. ,in Indicates the first Large predicted probability values.

[0057] In S402, top-k coverage probability statistics are used for each fault category. The true labels of all calibration samples in this category appear before the predicted probabilities. Bit( The number of times ( ). For each Value, calculate the actual label appearing first The number of bits and the total number of calibration samples in that category The ratio of the two values ​​is used to obtain the value of the ratio. Coverage probability corresponding to the value .

[0058] In S403, the label ranking threshold is determined. Based on preset coverage requirements, it is typically related to the target confidence level. Consistency, i.e., coverage probability Find satisfaction The smallest The value, will The value serves as the label ranking threshold for this fault category. .

[0059] The core principle of this process is the top-k error rate of the classifier, which ensures that the true label is at the top. The coverage probability of a bit meets the target requirement, providing a reasonable ranking constraint for subsequent prediction set selection. Essentially, this is achieved through the formula:

[0060] in, Indicates real label In predicting probability ranking, this formula ensures that in the category... Among them, at least one The proportion of samples whose true label ranking does not exceed This provides a statistical guarantee for the subsequent selection of the prediction set.

[0061] For example, a certain fault category contains 100 calibration samples, with a preset coverage requirement of 90%, and the statistics show... The coverage probability is 75%. If the coverage probability is 92%, then the minimum value that satisfies the coverage requirement is... A value of 2 represents the tag ranking threshold for this category. By setting a label ranking threshold, we ensure that the true labels of more than 90% of the calibration samples in this category appear in the top two of the model's predicted probabilities, providing a basis for subsequent prediction set selection.

[0062] This solves the technical problem of existing uncertainty quantification methods pursuing global marginal coverage, which leads to an excessively large prediction set for rare samples and loses its practical significance. By setting specific class-conditional quantile thresholds and label ranking thresholds for each category, class-conditional coverage is achieved, mathematically expressed as:

[0063] While ensuring that each category meets the target confidence level, it effectively filters low-probability labels, laying the foundation for subsequently reducing the size of the rare sample prediction set. Experiments have verified that... In cases of extreme imbalance, the prediction set size is as low as 1.65, which is much smaller than existing methods.

[0064] Please see Figure 5 , Figure 5 This is a comparison chart of coverage and prediction set size provided in the embodiments of this disclosure. Figure 5 Taking the target coverage with a 90% confidence level as an example, it can be seen that this method basically ensures the guarantee of target coverage. Figure 5As can be seen on the right, under standard conditions, as the target coverage increases, the prediction set needs to include more candidate values ​​to cover potentially correct labels in order to meet higher confidence levels. At different confidence levels, the proposed method achieves a smaller prediction set while maintaining coverage, thus achieving an effective balance between prediction efficiency and accuracy. OURS is the method provided in this disclosure, while APS, RAPS, and CCP are existing conformal prediction algorithms.

[0065] In some embodiments, for step S6, a prediction set of the test samples is generated based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold. The size of the prediction set is used to characterize the degree of uncertainty of the diagnostic results.

[0066] Please see Figure 6 , Figure 6 This is a schematic diagram of the process for generating a prediction set according to an embodiment of this disclosure. Specifically, in S601, the predicted probability ranking is obtained. For each test sample... Input it into the fine-tuned fault diagnosis model and obtain the predicted probability distribution of each category output by the model. The predicted probability distribution is sorted in descending order to obtain the ranking of predicted probabilities for each category label, and the ranking corresponding to each fault category label is recorded. ( ),in Indicate category Rank in the predicted probability sorting.

[0067] In S602, filter the category labels. For each fault category label in the predicted probability ranking. Perform two conditional judgments: The first conditional judgment: the inconsistency metric score corresponding to the fault category label. Is it not greater than its corresponding class conditional quantile threshold? That is, to judge Is it true? If it is true, it means that the category label meets the inconsistency measurement requirements and has the potential to become a true label; if it is not true, then the category label is excluded.

[0068] The second conditional judgment: the ranking of the fault category label in the predicted probability sorting. Is it not greater than its corresponding tag ranking threshold? That is, to judge Is it true? If it is true, it means that the category label has a high confidence ranking in the model prediction; if it is not true, then the category label is excluded.

[0069] In S603, the prediction set is determined. All fault category labels that simultaneously meet the above two conditions are selected to form the prediction set for this test sample. Its mathematical expression is:

[0070] in This represents the set of all fault categories. This formula is an improvement on the existing CCP method's prediction set generation formula, which is:

[0071] Therefore, by joining The ranking constraint effectively filters out low-probability labels with lower rankings, solving the problem of an excessively large prediction set for rare samples.

[0072] The size of the prediction set directly reflects the degree of uncertainty in the diagnostic results of the test samples. The fewer labels the prediction set contains, the clearer the model's judgment on the category of the test samples, and the lower the uncertainty. Conversely, the more labels the prediction set contains, the more ambiguous the model's judgment on the category of the test samples, and the higher the uncertainty.

[0073] For example, a test sample has a non-consistency metric score of 0.35, and a certain fault category... If the class conditional quantile threshold is 0.4 and the label ranking threshold is 3, and the category label ranks 2 in the predicted probability ranking, then the category label meets both conditions and is included in the prediction set; if the class conditional quantile threshold of another category label is 0.3, even if its ranking is 1, it is excluded from the prediction set because the inconsistency measure score of 0.35 is greater than 0.3.

[0074] Therefore, by synergistically selecting based on two conditions, the technical problem of generating excessively large prediction sets for rare samples in existing CCP methods is solved. While ensuring that each category meets the target confidence coverage requirement, low-probability labels with low rankings and excessive inconsistency scores are filtered out, significantly reducing the size of the prediction set for rare samples, improving the practicality of uncertainty quantification, making the quantification results more valuable, and providing a reliable basis for decision maintenance.

[0075] In some embodiments, for step S7, based on the generated prediction set, a reliability metric for the diagnostic results of the test sample is output, and the metric includes at least one of the prediction set, coverage, and prediction set size.

[0076] Specifically, the output prediction set contains the specific content of the set of fault category labels; the coverage ratio is calculated, which is the ratio of the number of test samples with true labels included in the prediction set to the total number of test samples. Its mathematical expression is consistent with the class condition coverage requirement. This reflects the coverage performance of the method; it records the prediction set size, i.e., the number of fault category labels included in the prediction set, characterizing the degree of uncertainty. Simultaneously, it can output auxiliary indicators such as AUROC, precision, and recall, intuitively presenting the reliability of the diagnostic results, depending on actual needs. All quantitative indicators are output in tabular or text format for easy viewing and analysis by users, providing data support for maintenance planning and production scheduling of nuclear-grade SLM equipment.

[0077] For example, in the scenario of quantifying the uncertainty of out-of-distribution samples, the training set includes normal, outer race, rolling element, and cage fault samples under operating conditions C1 (35Hz / 12kN), C2 (37.5Hz / 11kN), and C3 (40Hz / 10kN). Inner race and composite faults in C3 serve as auxiliary out-of-distribution training samples, while inner race faults in C1 and C2 serve as fully out-of-distribution test samples. Following the above implementation process, after model pre-training (400 cycles) and fine-tuning (200 cycles), the feature separation loss function is used... and Collaborative optimization and inconsistency metric scores The calculation showed that the recognition precision of the out-of-distribution test samples reached 93.01%, the recall rate reached 94.58%, and the AUROC index reached 95.77%. The prediction set can accurately reflect the high uncertainty of out-of-distribution samples, which verifies the effectiveness of the method in the out-of-distribution sample quantification scenario.

[0078] For example, in the scenario of quantifying uncertainty in class-imbalanced samples, cage failure samples under operating conditions C1 and C2 are set as rare samples, and then... The configuration data is based on the proportions of 0.2 and 0.3. This is achieved through class-conditional quantile thresholds. With tag ranking threshold The calculation, in At that time, the label ranking threshold for cage failure categories Class conditional quantile threshold According to the prediction set generation formula The generated prediction set size is only 1.65, which is significantly smaller than the 2.11 of the existing CCP method, and the coverage is maintained at over 90%, which fully demonstrates the advantages of this method in class imbalance scenarios and achieves accurate quantification of the uncertainty of rare samples.

[0079] Please see Figure 7 , Figure 7 This invention provides a system 700 for quantifying the reliability of fault diagnosis results for core-level SLM equipment based on conformal prediction. This system embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiments, this system can be specifically applied to various computer devices. The system specifically includes: The preprocessing module 701 is used to preprocess the vibration signals of the acquired nuclear-grade SLM equipment to construct a sample set including a training set, a calibration set, and a test set. The sample set includes in-distribution samples and out-of-distribution samples. The model fine-tuning module 702 is used to train the fault diagnosis model based on the training set and fine-tune the fault diagnosis model through the feature separation loss function; the feature separation loss function includes a separation loss term and a clustering loss term, the separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model, and the clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class; The feature extraction module 703 is used to extract high-dimensional feature vectors of test samples using the fine-tuned fault diagnosis model. The inconsistency measurement module 704 is used to calculate the inconsistency measurement score of the test sample based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model; the inconsistency measurement score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector; The threshold calculation module 705 is used to calculate the class conditional quantile threshold and the label ranking threshold for each fault category based on the calibration set. The prediction set generation module 706 is used to generate a prediction set for the test sample based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. The result output module 707 is used to output a reliability quantification index of the diagnostic results of the test sample based on the prediction set.

[0080] Based on the same inventive concept, this application also provides a computer device, the method corresponding to which can be the method in the foregoing embodiments, and the principle of solving the problem is similar to that method. The computer device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0081] The computer device can be a user device, or a device formed by integrating user devices and network devices through a network, or it can be an application running on the aforementioned devices. The user device includes, but is not limited to, various terminal devices such as computers, mobile phones, tablets, smartwatches, and smart bands. The network device includes, but is not limited to, network hosts, single network servers, multiple network server sets, or cloud computing-based computer sets, and can be used to implement some processing functions when setting an alarm clock. Here, the cloud consists of a large number of hosts or network servers based on cloud computing. Cloud computing is a type of distributed computing, consisting of a virtual computer composed of a group of loosely coupled computer sets.

[0082] Figure 8 The diagram illustrates the structure of an apparatus suitable for implementing the methods and / or technical solutions in the embodiments of this application. The apparatus 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for system operation. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0083] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, touchscreen, microphone, infrared sensor, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), LED display, OLED display, etc., and speakers, etc.; a storage section 808 including one or more computer-readable media such as hard disk, optical disk, magnetic disk, semiconductor memory, etc.; and a communication section 809 including a network interface card such as a LAN (local area network) card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet.

[0084] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a central processing unit (CPU) 801, it performs the functions defined in the methods of this application.

[0085] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0086] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0087] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0088] Furthermore, the inclusion of a single word does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for quantifying the reliability of fault diagnosis results for core-level SLM equipment based on conformal prediction, characterized in that, Includes the following steps: S1, preprocess the collected vibration signals of the nuclear-grade SLM equipment to construct a sample set including a training set, a calibration set and a test set, wherein the sample set includes in-distribution samples and out-of-distribution samples; S2, a fault diagnosis model is trained based on the training set, and the fault diagnosis model is fine-tuned using a feature separation loss function; the feature separation loss function includes a separation loss term and a clustering loss term, the separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model, and the clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class; S3. Using the fine-tuned fault diagnosis model, extract the high-dimensional feature vector of the test sample; S4. Based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model, calculate the inconsistency metric score of the test sample; the inconsistency metric score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector; S5, Calculate the class conditional quantile threshold and label ranking threshold for each fault category based on the calibration set; S6. Based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold, generate a prediction set for the test sample. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. S7. Based on the prediction set, output the reliability metrics of the diagnostic results of the test sample.

2. The method for quantifying the reliability of fault diagnosis results for nuclear-level SLM equipment based on conformal prediction as described in claim 1, characterized in that, In step S2, for out-of-distribution samples, the average value of the absolute cosine similarity between their normalized feature vector and all normalized class weight vectors is calculated as the separation loss term; for in-distribution samples, the negative value of the cosine similarity between their normalized feature vector and the normalized class weight vector corresponding to their real label is calculated as the clustering loss term; the feature separation loss function is the weighted sum of the separation loss term and the clustering loss term.

3. The method for quantifying the reliability of fault diagnosis results for nuclear-level SLM equipment based on conformal prediction as described in claim 1, characterized in that, In step S2, when fine-tuning the fault diagnosis model, the weight coefficients of the separation loss term and the aggregation loss term in the feature separation loss function are dynamically adjusted. The adjustment is based on the out-of-distribution sample recognition performance and intra-class feature aggregation degree on the calibration set.

4. The method for quantifying the reliability of fault diagnosis results for core-level SLM equipment based on conformal prediction according to claim 1, characterized in that, In step S5, the method for calculating the class conditional quantile threshold includes, for each fault category, filtering calibration samples belonging to that fault category based on the true label of the calibration sample; Based on the fault diagnosis model, calculate the inconsistency metric score corresponding to the true label for each calibration sample; Calculate the quantiles of the inconsistency metric scores for all calibrated samples in that class, and use them as the class-conditional quantile threshold for that class.

5. A method for quantifying the reliability of fault diagnosis results for core-level SLM equipment based on conformal prediction, as described in claim 1, is characterized in that... In step S5, the method for determining the label ranking threshold includes, for each fault category, based on the prediction results of the fault diagnosis model on the calibration set, calculating the probability that the true label of the sample in that category appears in the top k positions of the model's predicted probability; and determining the minimum k value that meets the preset coverage requirement as the label ranking threshold for that category.

6. The method for quantifying the reliability of fault diagnosis results for nuclear-level SLM equipment based on conformal prediction as described in claim 1, characterized in that, In step S6, generating the prediction set specifically includes: for the test sample, obtaining the predicted probability ranking of all fault categories output by the fault diagnosis model; for each fault category label in the ranking, determining whether it satisfies the two conditions in step S6; and taking the set of all fault category labels that satisfy the two conditions as the prediction set.

7. A system for quantifying the reliability of fault diagnosis results for nuclear-level SLM equipment based on conformal prediction, characterized in that, include: The preprocessing module is used to preprocess the acquired vibration signals of the nuclear-grade SLM equipment and construct a sample set including a training set, a calibration set, and a test set. The sample set includes in-distribution samples and out-of-distribution samples. The model fine-tuning module is used to train the fault diagnosis model based on the training set and fine-tune the fault diagnosis model through the feature separation loss function. The feature separation loss function includes a separation loss term and a clustering loss term. The separation loss term is used to make the feature vector of out-of-distribution samples orthogonal to the class weight vector of the fully connected layer of the fault diagnosis model. The clustering loss term is used to align the feature vector of in-distribution samples with the class weight vector of the corresponding class. The feature extraction module is used to extract high-dimensional feature vectors of test samples using the fine-tuned fault diagnosis model. The inconsistency measurement module is used to calculate the inconsistency measurement score of the test sample based on the high-dimensional feature vector and the category weight vector of the fault diagnosis model. The inconsistency measurement score is obtained by adding the maximum value of the probability distribution output by the fault diagnosis model to the mean of the absolute values ​​of the cosine similarity between the high-dimensional feature vector and each category weight vector. The threshold calculation module is used to calculate the class conditional quantile threshold and the label ranking threshold for each fault category based on the calibration set. The prediction set generation module is used to generate a prediction set for the test sample based on the inconsistency metric score, the class conditional quantile threshold, and the label ranking threshold. The prediction set includes fault category labels that meet the following two conditions: the inconsistency metric score corresponding to the fault category label is not greater than its corresponding class conditional quantile threshold, and the ranking of the fault category label in the output probability ranking of the fault diagnosis model is not greater than its corresponding label ranking threshold. The result output module is used to output a reliability metric for the diagnostic results of the test sample based on the prediction set.

8. A computer device, wherein the computer device is characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.

9. A computer-readable medium having computer program instructions stored thereon, characterized in that, The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images