An intelligent manufacturing anomaly detection method based on intrinsic parameter learning

Abstract

The invention discloses an intelligent manufacturing anomaly detection method based on intrinsic parameter learning, which belongs to the field of artificial intelligence. The method comprises the following steps: inputting a high-dimensional multispectral data set into a machine learning model, wherein the high-dimensional multispectral data set comprises: high-dimensional multispectral data features and label information corresponding to the high-dimensional multispectral data features; classifying model parameters based on the high-dimensional multispectral data features and the label information corresponding to the high-dimensional multispectral data features to obtain intrinsic parameters and extrinsic parameters; optimizing the intrinsic parameters to obtain optimized intrinsic parameters, and optimizing the extrinsic parameters to obtain optimized extrinsic parameters; performing constraint optimization enhancement on the optimized intrinsic parameters, and performing objective function normalization processing to obtain a final objective function, and obtaining an out-of-distribution sample detection result based on the final objective function.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to an intelligent manufacturing anomaly detection method based on intrinsic parameter learning. Background Art

[0002] Traditional machine vision inspection systems mainly rely on visible light imaging, which makes it difficult to fully capture subtle defects and potential safety hazards of products. Multispectral machine vision technology can provide richer surface and material features and achieve more sophisticated quality inspection by collecting image information in multiple wavelength ranges.

[0003] However, when applying machine learning algorithms for defect detection and quality control, the model may face the problem of covariance overfitting, resulting in a decrease in detection performance when processing out-of-distribution (OOD) samples. This may result in the inability to identify new defect types or abnormal conditions in a timely manner during the production process, affecting product quality and production safety.

[0004] In the existing technology, some methods try to alleviate overfitting through data enhancement, regularization and other means, but it is still not enough to fundamentally solve the impact of covariance overfitting on OOD detection. At the same time, traditional methods also have the problems of high computational complexity and poor real-time performance in processing high-dimensional and multi-spectral data.

[0005] Therefore, there is an urgent need for a method that combines intrinsic parameter learning, which can effectively alleviate covariance overfitting and improve the generalization ability of machine learning models in anomaly detection, and is applicable to multispectral machine vision data, thereby improving the quality control level in the intelligent manufacturing process. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes an intelligent manufacturing anomaly detection method based on intrinsic parameter learning to solve the problems existing in the above prior art.

[0007] To achieve the above object, the present invention provides a smart manufacturing anomaly detection method based on intrinsic parameter learning, comprising:

[0008] Inputting a high-dimensional multispectral data set into a machine learning model, wherein the high-dimensional multispectral data set includes: high-dimensional multispectral data features and label information corresponding to the high-dimensional multispectral data features;

[0009] Based on the high-dimensional multispectral data features and the label information corresponding to the high-dimensional multispectral data features, the model parameters are classified to obtain intrinsic parameters and extrinsic parameters; wherein the intrinsic parameters are parameters that are not affected by distribution shift; and the extrinsic parameters are parameters that have a negative impact on OOD detection under the condition of covariance shift;

[0010] Optimizing the intrinsic parameters to obtain optimized intrinsic parameters, and optimizing the extrinsic parameters to obtain optimized extrinsic parameters;

[0011] The optimized intrinsic parameters are subjected to constrained optimization enhancement, and objective function normalization is performed to obtain a final objective function, and an out-of-distribution sample detection result is obtained based on the final objective function.

[0012] Optionally, the classification criteria for classifying the model parameters to obtain intrinsic parameters and extrinsic parameters are:

[0013]

[0014] In the formula, represents the model parameters; express The gradient of D train represents the training data set; Indicates the size of the parameter weight; Indicates the index value used to judge the intrinsic nature of the parameter.

[0015] Optionally, the process of optimizing the intrinsic parameters to obtain optimized intrinsic parameters includes: performing forward gradient update on the intrinsic parameters based on a loss function and weight decay to obtain optimized intrinsic parameters;

[0016] Among them, the expression of the optimized intrinsic parameters is:

[0017] ;

[0018] In the formula, represents the current value of the intrinsic parameter at the tth iteration; Represents the updated internal parameter value at the t+1th iteration; represents the loss function; represents the learning rate; Represents the loss function Relative to internal reference The gradient at iteration t; represents the weight decay factor; represents a symbolic function; Represents the training dataset.

[0019] Optionally, the process of optimizing the extrinsic parameters to obtain optimized extrinsic parameters includes: performing negative gradient update on the extrinsic parameters based on weight decay to obtain optimized extrinsic parameters;

[0020] Among them, the expression of the optimized extrinsic parameters is:

[0021] ;

[0022] In the formula, Represents the external parameters at the tth step; Represents the external parameters updated at step t+1; represents the weight decay factor; Represents the learning rate.

[0023] Optionally, the expression for constrained optimization enhancement of the optimized intrinsic parameters is:

[0024] ;

[0025] In the formula, express The L2 norm of ; α represents a constant.

[0026] Optionally, the expression for normalizing the objective function is:

[0027] ;

[0028] In the formula, represents the average value of cross entropy loss; Represents the model's response to input x i Output: represents the model parameter; τ represents the temperature parameter; λ represents the regularization coefficient; Representation parameters The L1 norm of express The L2 norm of ; n represents the number of training samples.

[0029] Optionally, the final objective function is:

[0030] ;

[0031] In the formula, represents the final objective function.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] The present invention significantly improves the performance of intelligent manufacturing and anomaly detection by inputting high-dimensional multispectral data and performing intrinsic parameter learning. First, by dividing the model parameters into intrinsic parameters and extrinsic parameters and optimizing them separately, the robustness and stability of the model under the condition of covariance shift are enhanced. Secondly, the optimized intrinsic parameters are constrained optimized and the objective function is normalized, which effectively avoids overconfident predictions and improves detection accuracy and reliability. In addition, the method significantly enhances the generalization ability of the model for out-of-distribution samples, enabling it to adapt to a wider range of scenarios. At the same time, the optimized calculation strategy reduces the complexity, improves the training and detection efficiency, and meets the real-time requirements of intelligent manufacturing. In summary, the present invention provides an efficient and reliable solution for quality control and safety detection in industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 This is a flow chart of an intelligent manufacturing anomaly detection method based on intrinsic parameter learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] The present invention aims to solve the shortcomings of traditional quality inspection methods in the intelligent manufacturing process. By combining multispectral machine vision technology and intrinsic parameter learning, an anomaly detection method for alleviating covariance overfitting is proposed, thereby achieving efficient detection and accurate classification of out-of-distribution (OOD) samples in the production process.

[0039] Through intrinsic parameter learning technology, the generalization ability of machine learning models on multispectral machine vision data is optimized, and the impact of covariance overfitting on the detection performance of out-of-distribution samples is reduced.

[0040] Multispectral imaging technology is used to comprehensively capture subtle features (such as surface texture, invisible marks, and material properties) to improve the ability to detect small defects and potential quality problems. Efficient algorithms and system architectures are designed to adapt to the processing needs of high-dimensional multispectral data, and to achieve real-time anomaly detection and rapid feedback in the production process. Through intelligent quality control solutions, production risks and economic losses caused by product defects or anomalies are reduced, while production efficiency and product consistency are improved. Provide a scalable quality inspection method for intelligent manufacturing, suitable for feature extraction and anomaly identification of multispectral data, and help the industry develop in a more intelligent and efficient direction.

[0041] The present invention proposes intrinsic parameter learning to solve the following key problems: How to accurately distinguish intrinsic parameters from extrinsic parameters? What strategies should be adopted to effectively optimize these intrinsic and extrinsic parameters? And, how to improve the reliability of these parameters?

[0042] The goal of identifying intrinsic parameters in the present invention is to divide the parameters of the model into two groups: intrinsic parameters and extrinsic parameters. Inspired by the Lottery Ticket Hypothesis, the present invention believes that in the OOD (Out-of-distribution) detection network, only a part of the key parameters play a role in learning information with covariance robustness (called intrinsic information). This information is not affected by distribution shift and is crucial for accurate classification, so these parameters are called "intrinsic parameters". In contrast, other parameters learn information that is sensitive to covariance changes (called extrinsic information), which will have a negative impact on OOD detection under covariance shift conditions. These parameters are called "extrinsic parameters".

[0043] By designing an accurate criterion to identify the intrinsic parameter set, the negative impact of covariance sensitive information can be minimized during the learning process, thereby improving the OOD detection performance under covariance shift. To achieve this goal, the present invention proposes an effective intrinsic parameter selection criterion.

[0044] Furthermore, the present invention proposes a classification criterion based on the combination of parameter gradient size and weight size for the classification method of intrinsic parameters and extrinsic parameters, which is used to identify intrinsic parameters and extrinsic parameters in the model, wherein the intrinsic parameters are parameters that have a significant contribution to the model performance, and are mainly used to learn covariance robust information and enhance the stability and generalization ability of the model in out-of-distribution (OOD) detection. The information that the extrinsic parameters are sensitive to covariance changes may have a negative impact on the model performance, and they are separately identified and processed through the classification criterion.

[0045] Furthermore, the present invention designs differentiated optimization rules for the parameter category-based grouping optimization strategy according to the different characteristics of intrinsic parameters and extrinsic parameters. For intrinsic parameters, the gradient update rule is adopted to enhance their learning ability of covariance robust information, ensuring that the model can still effectively detect abnormal samples under distribution shift conditions. For extrinsic parameters, their influence is gradually weakened only through weight attenuation to avoid interference with the performance of the model, thereby improving the robustness of the model.

[0046] Furthermore, the present invention enhances the reliability of intrinsic parameters through constrained optimization. During the training process, the present invention constrains the logit value output by the classifier to limit the continuous growth of the logit amplitude and prevent the model from making overconfident predictions on samples outside the distribution; by stabilizing parameter weights and gradient changes, the convergence and reliability of the model optimization process are enhanced.

[0047] Furthermore, the present invention is about the fusion application of multispectral machine vision technology, combining multispectral machine vision technology, and using multi-band imaging data to comprehensively capture subtle defects and quality features. Among them, the present invention improves the feature extraction efficiency and anomaly detection capability in multispectral data processing through intrinsic parameter learning methods, realizes the deep fusion of high-dimensional multispectral data and machine learning models, and meets the needs of high-precision quality inspection in intelligent manufacturing.

[0048] Furthermore, the present invention has wide adaptability to intrinsic parameter learning. Since the intrinsic parameter learning method has good versatility and can be widely adapted to mainstream out-of-distribution detection methods and deep learning frameworks, the present invention can achieve high-precision anomaly detection in quality inspection and can be extended to intelligent manufacturing and quality management scenarios in other industries, thus having significant promotion value.

[0049] like Figure 1 As shown, this embodiment provides a smart manufacturing anomaly detection method based on intrinsic parameter learning, including the following steps:

[0050] Step 1: Data input: input a high-dimensional multispectral data set into a machine learning model. The high-dimensional multispectral data set includes: high-dimensional multispectral data features and label information corresponding to the high-dimensional multispectral data features.

[0051] The data input accepted by the present invention is divided into the following forms:

[0052] Training Dataset : Includes pairs of input x and label y, in the format The input x is high-dimensional multispectral data, which is an image, time series or other forms of multidimensional data; the label information y is a discrete category label with a value range of {1,2,⋯,K}.

[0053] Test data set: contains two parts: ① Test set under covariance shift , used to simulate test samples under distribution change conditions. ② Test set under semantic shift , used to verify the model's ability to detect OOD data.

[0054] The data is input in a standardized tensor format. For example, the shape of image data is (N, C, H, W), where N represents the batch size, C represents the number of channels, and H and W represent the height and width of the image, respectively.

[0055] Step 2: Parameter identification and classification: Based on the high-dimensional multispectral data features and label information corresponding to the high-dimensional multispectral data features, the model parameters are divided into intrinsic parameters and extrinsic parameters.

[0056] Traditional methods are usually directly based on a single indicator of weight value or gradient value, which easily ignores the combined effect of the two, and the classification criteria of traditional methods lack robustness, resulting in inaccurate parameter classification. Therefore, the present invention reduces human subjective intervention through the strict definition of mathematical formulas, and comprehensively measures the importance of parameters by introducing the product of gradient and weight, so that the classification is more accurate. As a specific implementation method of this embodiment, the following steps are included:

[0057] 1. Determine the classification standard: Use the product of parameter weight and gradient size as the classification basis, and the classification standard Defined as:

[0058] .

[0059] represents the model parameters; express The gradient of Indicates the size of the parameter weight; An indicator value used to judge the intrinsic nature of a parameter.

[0060] when When the value of is large, it is considered is an intrinsic parameter; when When the value of is small, it is considered is an external parameter. As a specific implementation of this embodiment, calculate the g of all parameters i Value. i The values ​​are sorted from largest to smallest. A percentage p% is set, in this embodiment, the percentage is 20%, the first p% of the parameters are taken as intrinsic parameters, and the rest are extrinsic parameters.

[0061] 2. Data processing and result classification: The prior art usually requires preset empirical thresholds, which are highly subjective and not suitable for complex models. Therefore, the present invention uses a dynamic calculation method to set thresholds to ensure generalization to different models. As a specific implementation method of this embodiment, the following steps are included: Automatically calculate the threshold of each parameter The parameter category is divided according to the threshold value. The threshold T is the g of the p% position in the sorting. i value.

[0062] Step 3: Optimization of classification results: Optimize intrinsic parameters and extrinsic parameters respectively to reduce the impact of covariance drift sensitive information on model performance. Optimize the intrinsic parameters to obtain optimized intrinsic parameters, and optimize the extrinsic parameters to obtain optimized extrinsic parameters.

[0063] The prior art uniformly updates all parameters during the classification result optimization process without distinguishing the role of intrinsic and extrinsic parameters, which is prone to overfitting under covariance drift conditions, resulting in reduced detection performance. Based on this technical defect, the present invention specifically optimizes intrinsic parameters by distinguishing parameter categories, thereby improving its ability to learn drift robust information.

[0064] Furthermore, the prior art does not specifically suppress the influence of external parameters, resulting in the dominance of covariance drift sensitive information. Therefore, the present invention uses weight attenuation to gradually reduce the influence of external parameters to weaken its learning of drift sensitive information.

[0065] After determining the intrinsic parameters and the extrinsic parameters, different update rules are adopted for the two types of parameters respectively. As a specific implementation of this embodiment, the following steps are included:

[0066] (1) Intrinsic parameter optimization: optimizing the intrinsic parameters to obtain optimized intrinsic parameters

[0067] Based on the loss function and weight decay, the intrinsic parameters Perform a forward gradient update:

[0068] .

[0069] In the formula, is the current value of the intrinsic parameters, which represents the intrinsic parameters at the tth iteration. The intrinsic parameters are usually the parameters used in the camera or sensor model to describe image distortion, focal length, etc. This is the updated internal parameter value at the t+1th iteration. The goal is to optimize and update the internal parameter to reduce the loss function The value of . This is the learning rate, which controls the step size of each update. A larger learning rate may cause the update to be too large or even miss the optimal solution, while a smaller learning rate may cause the convergence to be too slow. This is the loss function Relative to internal reference The gradient at the tth iteration. The gradient represents the impact of the current internal parameter on the loss, and the gradient descent algorithm uses it to adjust the internal parameter to minimize the loss. This is the weight decay factor, which is used to control the influence of the regularization term. Weight decay is a common regularization method to prevent the model from overfitting. Larger A value of can increase the influence of the regularization term, thus making the intrinsic parameter smaller. This is the sign function, which returns the sign of the intrinsic parameter (i.e. 1 for positive values, -1 for negative values, and 0 for zero). The purpose of this term is to adjust the directionality of the intrinsic parameter by weight decay (i.e., for positive intrinsic parameters, the update will decrease, and for negative intrinsic parameters, the update will increase). This is the training dataset, which usually contains images or measurement data captured by a camera or sensor. This dataset is used to train the model and calculate the loss function to guide the update of the intrinsic parameters.

[0070] (2) External parameter optimization

[0071] For external parameters Use negative updates based only on weight decay:

[0072] .

[0073] In the formula, represents the external parameters at the tth step. Represents the external parameters updated at step t+1. Represents the weight decay factor, which is used for regularization. Represents the learning rate, which controls the magnitude of each update.

[0074] Through the above optimization strategy, the learning ability of intrinsic parameters is enhanced, while the negative impact of extrinsic parameters is suppressed, thereby improving the OOD detection performance under covariance shift conditions.

[0075] Step 4, improve the calculation and analysis method: control the weight amplitude of the intrinsic parameters, prevent the problem of overconfident prediction, and improve the reliability of OOD detection. Perform constraint optimization and enhancement on the optimized intrinsic parameters, and perform objective function normalization to obtain the final objective function, and obtain the out-of-distribution sample detection result based on the final objective function.

[0076] Conventional training directly optimizes the loss function, which leads to excessive Logit values ​​and overconfidence in the model's prediction of OOD data. Therefore, the present invention adds constraints on the Logit amplitude, explicitly controls the confidence of the model output, and avoids overconfidence. The specific method includes the following steps:

[0077] (1) Add constraint optimization

[0078] Add to optimization goals Norm constraint:

[0079] .

[0080] In the formula, express The L2 norm (also called the Euclidean norm) of . α is a constant that represents the function The L2 norm of is constrained to be equal to a fixed value of .

[0081] (2) Alternative objective function

[0082] In view of the problem that the traditional method simply uses Lagrange multipliers for constrained optimization, the calculation is unstable and the effect is limited, the present invention introduces normalization processing to keep the Logit amplitude constant, fundamentally solving the overconfidence problem. It improves the stability of optimization and prevents the problem of gradient instability. This embodiment optimizes the Logit amplitude through the objective function of normalization processing, and the specific expression is:

[0083] .

[0084] In the formula, is the average cross entropy loss (LCE), which measures the difference between the model prediction and the true label. is the model's response to input x i The output, are model parameters. τ is a temperature parameter used to scale the output of the model. yes The L2 norm (Euclidean norm) of is used to normalize the output. λ is the regularization coefficient, which controls the strength of the regularization term. is a parameter The L1 norm of , which represents the sum of the absolute values ​​of the parameters. n represents the number of training samples.

[0085] During the optimization process, the magnitude of the intrinsic parameter weight may continue to increase, causing the model to be overconfident in the prediction results of OOD samples. To avoid this phenomenon, this embodiment proposes a method based on constrained optimization by keeping the logit value The norm is constant to solve this problem. The objective function is defined as:

[0086] .

[0087] In the formula, represents the final objective function. is the logit output of the classifier, that is, the model's response to the input x i Output: is the parameter for adjusting the size of the logit hypersphere; For function logit value Norm. This is the average cross entropy loss (LCE), which is used to measure the model's performance on the training data D. train The difference between the prediction and the true label on .

[0088] Through this constrained optimization method, the problem of overconfident predictions is effectively alleviated, the gradients during training are stabilized, the reliability of intrinsic parameters is further improved, and ultimately more efficient OOD detection performance under covariance shift is achieved.

[0089] Step 5: Result verification: Verify the effectiveness of existing methods

[0090] Traditional methods cannot effectively distinguish internal and external parameters and perform poorly on drift data. The present invention combines classification and optimization strategies to systematically improve detection performance. Specific implementation methods of this embodiment include:

[0091] The weight of intrinsic parameters gradually increases with training, and the robustness under covariance drift is significantly enhanced.

[0092] The weight of extrinsic parameters is gradually reduced, and the influence of drift-sensitive information is effectively suppressed.

[0093] Step 6: Data output

[0094] (1) Classification results

[0095] ID data classification results: The model outputs classification probability distribution for ID data: .

[0096] The output is a probability matrix of (N, K), which indicates the probability that each sample belongs to K categories.

[0097] (2) OOD detection results

[0098] Detection score: The model calculates the OOD score S for each input sample OOD (x), the specific form of which depends on the post-processing method (e.g., energy score or contrast feature distance).

[0099] Classification label: By setting the threshold T, the score S OOD (x) is mapped to a binary label:

[0100] .

[0101] (3) Visual output: Classification results and detection results can be presented in tabular or graphical form for further analysis.

[0102] Furthermore, as a preferred solution of this embodiment, during the experiment, this embodiment uses multiple standard data sets for testing, including CIFAR-10, CIFAR-100 and ImageNet-200. For CIFAR-10 and CIFAR-100, this embodiment uses the official training sets as , and use the test sets of CIFAR-10c and CIFAR-100c as , and the OOD group includes MNIST, SVHN, Textures, and Places365 datasets. For ImageNet-200, this embodiment uses the official training set of ImageNet-200 as , and use test data from ImageNet-C, ImageNet-R, and ImageNet-V2 as ,OOD datasets include Ninco, iNaturalist, Textures, and OpenImage-O.

[0103] In terms of algorithm comparison, this embodiment compares the IPL algorithm with a variety of representative methods, including MSP, Mahalanobis, Energy, GradNorm, LogitNorm, React, Vim, VOS, and MIXOE. Among these methods, MSP, Mahalanobis, Energy, GradNorm, React, and Vim are test methods, while LogitNorm, VOS, and MIXOE are training methods.

[0104] The experimental evaluation used two key indicators, AUROC and FPR95. AUROC measures the ability of the model to distinguish between ID samples and OOD samples. The higher the value, the stronger the model's ability to distinguish. FPR95 measures the proportion of ID samples misclassified as OOD when the detection accuracy of OOD samples is 95%, reflecting the frequency of model misclassification.

[0105] In terms of experimental results, IPL performed well in most tests. For example, on the CIFAR-10 dataset, IPL significantly outperformed most baseline methods (such as MSP, Energy, and VOS) in the FPR95 indicator, and also showed high performance (91.31%) in the AUROC indicator. Compared with other methods, IPL significantly reduced the frequency of ID samples being misclassified as OOD while maintaining high detection accuracy, thereby achieving more accurate OOD detection. The experimental results are shown in Table 1.

[0106] Table 1

[0107]

[0108] This paper focuses on "anomaly detection method for alleviating covariance overfitting through intrinsic parameter learning (IPL)". From the perspective of parameters, it innovatively solves the performance bottleneck of out-of-distribution (OOD) detection under covariance shift conditions. The main technical effects include:

[0109] 1. For the first time, the OOD detection problem under covariance shift is solved from a parameter perspective.

[0110] This paper proposes a new method from the perspective of parameters, which divides model parameters into two categories: intrinsic parameters and extrinsic parameters. This innovative classification method effectively alleviates the reliance on sensitive information of covariance changes in traditional methods and significantly improves the robustness and generalization ability of out-of-distribution sample detection.

[0111] 2. Realize group optimization of parameters to improve model performance.

[0112] By designing the division criteria of intrinsic parameters and extrinsic parameters, and using different optimization rules for the two types of parameters: for intrinsic parameters, gradient updates are used to enhance their ability to learn covariance robust information; for extrinsic parameters, their influence is reduced only through weight decay, gradually reducing the dependence on covariance sensitive information. This strategy effectively reduces the negative effects of covariance sensitive information, allowing the model to maintain stable performance in complex environments.

[0113] 3. Enhance the effectiveness of intrinsic parameters through constrained optimization.

[0114] The present invention isolates the influence of the norm of the classifier output logit on the prediction result through the constrained optimization method, thereby alleviating the overconfidence problem of the model on out-of-distribution samples. norm, effectively suppressing the infinite growth of the intrinsic parameter weight amplitude;

[0115] At the same time, this method stabilizes the gradient updates during the training phase, further improving the reliability of intrinsic parameters and model performance.

[0116] 4. Realize the integration of multispectral machine vision technology and intrinsic parameter learning.

[0117] Combined with multispectral machine vision technology, this invention achieves a comprehensive improvement in anomaly detection and quality control by extracting and optimizing the features of high-dimensional multispectral data in the production process. The intrinsic parameter learning method significantly improves the sensitivity of out-of-distribution defect detection; compared with traditional methods, the fusion of multispectral data further enhances the model's refined feature recognition ability, providing strong technical support for intelligent manufacturing.

[0118] 5. Strong versatility and easy integration.

[0119] The present invention has been verified by a large number of experiments, proving the effectiveness and versatility of the intrinsic parameter learning method. IPL can not only adapt to the most popular OOD detection method, but also be easily integrated into different deep learning frameworks and widely used in anomaly detection tasks in different scenarios.

[0120] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for detecting anomalies in intelligent manufacturing based on intrinsic parameter learning, characterized in that: The following steps are involved: Inputting a high-dimensional multispectral data set into a machine learning model, wherein the high-dimensional multispectral data set includes: high-dimensional multispectral data features and label information corresponding to the high-dimensional multispectral data features; Based on the high-dimensional multispectral data features and the label information corresponding to the high-dimensional multispectral data features, the model parameters are classified to obtain intrinsic parameters and extrinsic parameters; wherein the intrinsic parameters are parameters that are not affected by distribution shift; and the extrinsic parameters are parameters that have a negative impact on OOD detection under the condition of covariance shift; Optimizing the intrinsic parameters to obtain optimized intrinsic parameters, and optimizing the extrinsic parameters to obtain optimized extrinsic parameters; Performing constrained optimization enhancement on the optimized intrinsic parameters, and performing objective function normalization processing to obtain a final objective function, and obtaining an out-of-distribution sample detection result based on the final objective function; The classification criteria for classifying model parameters into intrinsic parameters and extrinsic parameters are: ； In the formula, represents the model parameters; express The gradient of D train represents the training data set; Indicates the size of the parameter weight; represents the index value used to judge the intrinsic nature of the parameter; The process of optimizing the intrinsic parameters to obtain optimized intrinsic parameters includes: performing forward gradient update on the intrinsic parameters based on the loss function and weight decay to obtain optimized intrinsic parameters; Among them, the expression of the optimized intrinsic parameters is: ； In the formula, represents the current value of the intrinsic parameter at the tth iteration; Represents the updated internal parameter value at the t+1th iteration; represents the loss function; represents the learning rate; Represents the loss function Relative to internal reference The gradient at iteration t; represents the weight decay factor; represents a symbolic function; represents the training data set; The process of optimizing the extrinsic parameters to obtain optimized extrinsic parameters includes: performing negative gradient update on the extrinsic parameters based on weight decay to obtain optimized extrinsic parameters; Among them, the expression of the optimized extrinsic parameters is: ； In the formula, Represents the external parameters at the tth step; Represents the external parameters updated at step t+1; represents the weight decay factor; Represents the learning rate.

2. The intelligent manufacturing anomaly detection method based on intrinsic parameter learning according to claim 1 is characterized in that: The expression for constrained optimization enhancement of the optimized intrinsic parameters is: ； In the formula, express The L2 norm of ; α represents a constant.

3. The intelligent manufacturing anomaly detection method based on intrinsic parameter learning according to claim 2 is characterized in that: The expression for normalizing the objective function is: ； In the formula, represents the average value of cross entropy loss; Represents the model's response to input x i Output: represents the model parameter; τ represents the temperature parameter; λ represents the regularization coefficient; Representation parameters The L1 norm of express The L2 norm of ; n represents the number of training samples.

4. The intelligent manufacturing anomaly detection method based on intrinsic parameter learning according to claim 3 is characterized in that: The final objective function is: ； In the formula, represents the final objective function.

Images