Welding implicit anomaly detection and identification method based on multi-source data

Through multi-source data fusion and feature extraction technology, combined with LDA and LOF methods, the probabilistic neural network and artificial bee colony algorithm are used to solve the problem that it is difficult to detect hidden abnormalities in welding in a single image information, and the influence of multiple factors during the welding process is achieved. comprehensive detection and accurate identification.

CN118965216AActive Publication Date: 2024-11-15BEIJING INST OF TECH

Patent Information

Application Number
CN202410960492.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-15
Estimated Expiration
2044-07-17

AI Technical Summary

Technical Problem

The existing welding quality detection methods mainly rely on single image information, making it difficult to detect hidden abnormal problems during welding, and cannot fully reflect the influence of various factors during welding.

Method used

Using multi-source data fusion method, combining image information, temperature, current, voltage, spectrum and sound information, weld recessive abnormalities are identified through feature extraction, LDA dimensionality reduction, LOF anomaly detection, isolated forest segmentation and probability neural networks, and artificial bees are used to identify welded recessive abnormalities. The group algorithm optimizes the smoothing factor to improve detection accuracy.

Benefits of technology

It realizes comprehensive detection of surface and hidden abnormalities during welding, improves the accuracy and robustness of abnormal detection, can better identify multiple welding failures, and enhances the interpretability of online testing of welding quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965216A_ABST
    Figure CN118965216A_ABST
Patent Text Reader

Abstract

The invention discloses a welding implicit anomaly detection and identification method based on multi-source data, and belongs to the field of welding quality detection. The implementation method comprises the following steps of: acquiring multi-modal data information of welding quality of each object in real time, wherein the multi-modal data information comprises sound, current, voltage, spectrum, temperature information and image information; each information source is collected by a corresponding sensor, and the characteristics of the information sources are extracted. And carrying out normalization processing on the extracted feature values, and carrying out dimensionality reduction on the data based on abnormal categories by adopting an LDA method. And detecting abnormal points by using a local outlier factor (LOF) method. Abnormal points are identified on the whole by using an isolated forest, a detection region is segmented, and a new neighborhood search space is defined in the segmented region. And adopting a PNN (probabilistic neural network) method to identify abnormal types. And optimizing a smoothing factor of the PNN by using an artificial bee colony algorithm ABC to obtain an optimal smoothing factor and establish an optimal PNN model, and realizing welding recessive anomaly detection and identification according to the optimal PNN network optimization model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a welding hidden anomaly detection and identification method based on multi-source data, and belongs to the field of welding quality detection. Background Art

[0002] Welding is a commonly used metal connection process, which connects metal parts by heating and melting metal materials and adding appropriate filler materials. Welding is widely used in manufacturing, construction, automobile manufacturing, aerospace and other fields. The most important thing in the welding process is to ensure the quality and reliability of welding. Popular welding quality inspections generally use single welding quality information, such as welding image information using convolutional neural networks, welding sound information using sound sensors, and welding quality detection methods based on pulse infrared thermal imaging technology. At present, advanced welding quality detection methods are still based on single image information detection. Through the improvement of corresponding image extraction algorithms and extracted features, this welding quality detection method is continuously optimized. The welding quality detection method based on single image information is an effective method because the quality of welding mainly depends on the size and shape of the weld. Although the welding quality detection method based on image information has a good detection effect, the image information can only reflect some surface welding quality problems, and some hidden welding quality problems cannot be detected. Therefore, a welding quality detection method based on multiple data sources is adopted. Since different information can reflect the influence of different factors in the welding process, the use of graphics, sound, current, voltage, spectrum, temperature and other information can more comprehensively reflect the abnormalities in the welding process, especially the sound, current and other information, which can detect the hidden abnormalities in welding and greatly improve the effect of welding abnormality detection. Summary of the invention

[0003] To solve the problem of difficulty in discovering hidden abnormalities in welding quality. The purpose of the present invention is to provide a method for detecting and identifying hidden abnormalities in welding based on multi-source data, which uses sound, current, voltage, spectrum, temperature information, and combines image information. While being able to detect surface welding abnormalities, it can also detect hidden abnormalities hidden under the surface. First, each information source is collected using the corresponding sensor and its features are extracted. The extracted eigenvalues ​​are normalized, and the LDA method is used to reduce the dimension of the data based on the abnormal category. The intrinsic dimension is estimated using a statistical method, and the dimension after dimensionality reduction is equal to the intrinsic dimension. The local outlier factor LOF (Local Outlier Factor) method is used to detect abnormal points. Isolation forest isolation is used to identify abnormal points as a whole, the detection area is segmented, and a new neighborhood search space is defined in the segmented area, which significantly reduces the calculation time of LOF and improves the accuracy of detection. Finally, the probabilistic neural network PNN (Probabilistic Neural Network) method is used to identify the abnormal types. The artificial bee colony algorithm ABC (Artificial Bee Colony Algorithm) is used to optimize the smoothing factor of PNN, obtain the optimal smoothing factor and establish the optimal PNN model, and realize the detection and recognition of welding hidden anomalies based on the optimal probabilistic neural network PNN network optimization model.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] The present invention discloses a method for detecting and identifying welding hidden anomalies based on multi-source data, comprising the following steps:

[0006] Step 1: for n identical welding objects, multimodal data information of welding quality of each object is collected in real time and stored in a database; the multimodal data includes: front image of molten pool, temperature field information on both sides of molten pool, current and voltage information during welding, sound information during welding and arc spectrum information during welding;

[0007] Step 2: Extract features from the multimodal data information collected in step 1 to obtain multi-source data; extract the size information of the molten pool and the grayscale value of the image from the front image of the molten pool; extract the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness of the corresponding signals from the temperature field information on both sides of the molten pool, the current and voltage information during welding, the sound information during welding and the arc spectrum information during welding;

[0008] The length, width, and grayscale value information of the molten pool image are [g 1 , g 2 , g 3 ], the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of temperature are [t1 , t 2 , t 3 , t 4 , t 5 , t 6 ], the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of the current are [c 1 , c 2 , c 3 , c 4 , c 5 , c 6 ], the voltage mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information are [v 1 , v 2 , v 3 , v 4 , v 5 , v 6 ], the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of the sound are [a 1 , a 2 , a 3 , a 4 , a 5 , a 6 ], the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of the spectrum are [l 1 , l 2 , l 3 , l 4 , l 5 , l 6 ];

[0009] Then each data x′ in n welding objects i , expressed as [g i,1 , ..., g i,3 , t i,1 , ..., t i,6 , C i,1 , ..., c i,6 , v i,1 , ..., v i,6 , a i,1 , ..., a i,6 , l i,1 , ..., l i,6 ]; where i∈[1,n];

[0010] Step 3: Normalize the multi-source data extracted in step 2; then use the LDA method to reduce the dimension of the multi-source data;

[0011] Step 3.1: Convert each eigenvalue to the range of [-1, 1], and use the following formula for each multi-source data x′ i Perform normalization:

[0012]

[0013] In the formula, max(x′ i,j ) is the multi-source data x′ i The maximum value in the jth dimension, min(x′ i,j ) is the multi-source data x′ i The minimum value in the jth dimension; where i∈[1,n]; after normalization, each welding object is y i , expressed as

[0014] [G i,1 , ..., G i,3 , T i,1 , ..., T i,6 , C i,1 , ..., C i,6 , V i,1 , ..., V i,6 , A i,1 , ..., A i,6 , L i,1 , ..., L i,6 ]; where i∈[1,n];

[0015] Step 3.2: n normalized data y i Form a matrix Y and estimate the intrinsic dimension of the matrix Y;

[0016] Step 3.2.1, let Y be

[0017]

[0018] Get the matrix Y T The eigenvectors of Y

[0019] [λ g,1 , ..., λ g,3 ,λ t,1 , ..., λ t,6 ,λ c,1 , ..., λ c,6 ,λ v,1 , ..., λ v,6 ,λ a,1 , ..., λ a,6 ,λ l,1 , ..., λ l,6 ];

[0020] Step 3.2.2, generate a matrix X(b) of random data with the same number n and dimension as the Y matrix; where b∈[1,a], m is the dimension of X(b), which is the same as the dimension of X; then X(b) has the following matrix form;

[0021]

[0022] Step 3.2.3. Find the eigenvector [λ(b) 1 ,λ(b) 2 , …, λ(b) m ], and further calculate the average value of all eigenvalues Where j is the data dimension, j∈[1,m]; then the mean eigenvector is [λ(B) 1 ,λ(B) 2 , …, λ(B) m ];

[0023] Step 3.2.4: Compare matrices Y of the same dimension one by one T The eigenvectors of Y and the eigenvalues ​​in the mean eigenvector determine the matrix Y T The eigenvalue of Y is greater than that of matrix X (b) T The number of eigenvalues ​​of X(b) is denoted as d, then the normalized data y i The intrinsic dimension of is d;

[0024] Step 3.3: Use LDA dimensionality reduction method to reduce the dimensionality of multi-source data;

[0025] (1) The welding quality is divided into N categories, the number of welding objects is n, and the jth category set of welding objects is Y j , calculate the intra-class divergence S w :

[0026]

[0027] Where S wj represents the intra-class divergence of the jth class, μ j represents the welding object y of the jth type i The mean vector of , i∈[1,n], j∈[1,N];

[0028] (2) Calculate the inter-class divergence S b :

[0029]

[0030] where n j For the jth type of welding object y i The number of, μ is the welding object y i The mean vector of ;

[0031] (3) Using intra-class divergence S w and the inter-class divergence S b , and the target is:

[0032]

[0033] The matrix W is the matrix of the eigenvectors corresponding to the largest d eigenvalues; further deduction shows that the optimization objective function is:

[0034]

[0035] Where d is the intrinsic dimension of the data; then for any normalized data y i , after dimensionality reduction, we get data x i =W T y i , i∈[1,n], n is the number of samples; then the reduced-dimensional data x i Expressed as [x i,1 , x i,2 , x i,3 , ..., x i,d ];

[0036] Step 4: Use the local outlier factor algorithm LOF to model the anomaly detection of multi-source data to obtain an anomaly detection model; identify new abnormal welding objects based on the anomaly detection model;

[0037] Step 4.1: Use the isolation forest method to select the first c layers of the isolation forest to obtain the data x after n dimensionality reduction in step 3 i Split, i∈[1,n], and get 2 c A sample subset;

[0038] The first layer of segmentation selects a dimension j and then selects a data point x i , according to x i,j The value of divides the data set X into two parts X′ 1 and X′ 2 , X′ 1 Some j-th dimension is greater than x i,j , X′ 2 Some j-th dimension is less than or equal to x i,j ; The second layer further divides X′ 1 and X′ 2 , and so on to the cth layer;

[0039] Determine the new welding object point x n+1 With the segmentation boundary x i,j If x n+1 The distance to the segmentation boundary is close, that is, min[ther(x i,j )]<x n+1,j <max[ther(x i,j )], then change the position of the segmentation boundary; then let x i,j for ther(x i,j );

[0040]

[0041] Among them, a i and b i are two random points in the sample subset with a distance k between them;

[0042] Step 4.2: Use Isolation Forest to extract the 2 c Select the neighborhood search space from the sample subsets;

[0043] Calculate each dimension reduction data x i The average path length The calculation formula is:

[0044]

[0045] N is the number of spanning trees in the isolation forest, p i,j is the i-th dimension reduction data x i The length of the path in the jth tree, i∈[1,n];

[0046] According to the dimension reduction data x i The normal and abnormal data in the When , the point is added to the neighborhood search space, and the point in the sample subset search space is recorded as The number of sample subsets divided is l∈[1,2 c ];

[0047] Step 4.3: Calculate the new welding object point x n+1 LOF value;

[0048] Determine the new welding object x n+1 k-neighborhood of x; first determine x by the following formula n+1 The sample subset it belongs to is l, and the points in this sample subset are

[0049]

[0050] in, is a sample subset, the points in The average path length of

[0051] Then calculate the new welding object x n+1 With all points in the sample subset, The distance between

[0052]

[0053] Among them, j is The dimension of , j∈[1,d]; in the first, subset, the distance point The kth point The distance between

[0054]

[0055] Point x n+1 The k-neighborhood of point x n+1 Centered on The set of points, then point x n+1 The k-neighborhood of Point x n+1 The number of points in the k-neighborhood of Point x n+1 Points in the k-neighborhood of To point x n+1 The k-th reachable distance is:

[0056]

[0057] Calculate point x n+1 The local reachable density of is:

[0058]

[0059] Calculate point x n+1 Local anomaly factor:

[0060]

[0061] in For x n+1 Points in the k-neighborhood of The local reachable density of

[0062] Determine point x n+1 The range of local outlier factors is used to complete the anomaly detection modeling of multi-source data:

[0063]

[0064] Step 5: Construct a probabilistic neural network (PNN) network model for welding hidden anomaly detection and identification, use the ABC algorithm to optimize the smoothing factor σ of the PNN, obtain the probabilistic neural network (PNN) network optimization model, and realize welding hidden anomaly detection and identification based on the probabilistic neural network (PNN) network optimization model.

[0065] Step 5.1: The PNN model consists of an input layer, a hidden layer, a summation layer, and an output layer;

[0066] Input the abnormal data point v after dimensionality reduction i, abnormal data point v i is the data point x after dimensionality reduction i For each abnormal point v i There are d dimensions [v i,1 , v i,2 , v i,3 , …, v i,d ], and has an abnormal type in u, which is input to the input layer; where i < n; the hidden layer and the summation layer are calculated together using the Parzen method to calculate the abnormal point v i The jth data point with the hth type of anomaly , h∈u; and weighted sum the hidden layer outputs of the same abnormal pattern, the relationship is as follows:

[0067]

[0068] Where σ is the smoothing factor, d is the dimension of the input vector, and v i is the input vector, is the jth abnormal data point of the hth type of abnormality; m is the jth abnormal data point v i Number;

[0069] The number of neurons in the output layer is the total number of abnormal categories; for f h (v i ) Perform threshold identification and find the neuron with the largest posterior probability density among all output layer neurons;

[0070] y=argmax(f h (v i ))

[0071] Step 5.2: Use the artificial bee colony algorithm ABC to optimize the smoothing factor σ of PNN; use the optimization model to identify the abnormal cause of the abnormal result of the abnormal detection model in step 4;

[0072] Step 5.2.1, set the number of nectar sources to o, the number of leading bees and the upper limit of the number of nectar source tests; determine the value range of the smoothing factor to [0, 1]; randomly generate the initial smoothing factor, and its expression is

[0073]

[0074] where σ z is the zth nectar source, z∈[1,o]; j is σ z The dimension is the jth smoothing factor in PNN; the number of j is the number of outliers v i Number; A random number between 0 and 1;

[0075] Step 5.2.2: Use the correct number of PNN classifications as the fitness function to evaluate the effect of the smoothing factor; where cor(σ z )[argmax(f(v i ))] means that for all abnormal data points v i , using the smoothing factor σ z The number of correct prediction results;

[0076] Step 5.2.3, Leading Bee Stage, at the initial smoothing factor σ z Find a new smoothing factor σ′ near i , the formula is;

[0077]

[0078] σ′ zj =max{0,min{v zj , 1}}

[0079] in is a random value between -1 and 1, σ kj is zj The random values ​​in; compare cor(σ z )[argmax(f(x i ))] and cor(σ′ z )[argmax(f(x i ))], where cor(σ′ z )[argmax(f(x i ))] is the smoothing factor σ′ z The number of correct prediction results; when cor(σ z )[argmax(f(x i ))]>cor(σ′ z )[argmax(f(x i ))], leaving Cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As a new smoothing factor, when cor(σ z )[argmax(f(x i ))]<Cor(σ′ z )[argmax(f(x i ))], leaving cor(σ′ z )[argmax(f(x i ))]'s smoothing factor σ′ z As a new smoothing factor, when cor(σ z )[argmax(f(x i))]=cor(σ′ z )[argmax(f(x i ))], leaving cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As the new smoothing factor; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1;

[0080] Step 5.2.4, in the bee-following stage, the roulette wheel method is used to select the smoothing factor:

[0081]

[0082] Among them, p z To select nectar source z The probability of z∈[1,o]; the follower bee selects the nectar source σ from the optimal result of step 5.2.3 according to the probability z , then in σ z Nearby, find the new smoothing factor σ′ according to the method in step 5.2.3 i , leaving the better smoothing factor as σ z ; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1;

[0083] Step 5.2.5, Scouting bee stage, search for the number of times each nectar source is updated R (σ z ), when the preset threshold is reached and it is not updated, the honey source is updated using the following formula to search for the global optimal solution;

[0084]

[0085] According to the method of step 5.2.3, compare the optimal σ obtained in step 5.2.4 z The cor(σ z )[argmax(f(x i ))] and σ′ updated from 5.2.5 z cor(σ′ z )[argmax(f(x i ))], leaving the better smoothing factor as σ z ; Repeat steps 5.2.1 to 5.2.5 until the number of iterations is reached and the optimized smoothing factor is obtained.

[0086] Step 5.3: Use the optimized model to identify the abnormal cause of the abnormal result of the abnormal detection model in step 4; use the optimized PNN model to identify the new abnormal welding object v identified in step 4. n+1 Input, calculate the new outlier point v n+1 and the abnormal data point v after dimensionality reduction i The jth data point of the hth type of anomaly in The relationship between the hidden layer outputs of the same abnormal mode is weighted and summed, and the relationship is as follows:

[0087]

[0088] Where σ is the smoothing factor, d is the dimension of the input vector, and v n+1 is the new input vector, is the abnormal data point v after dimensionality reduction i The jth abnormal data point of the hth abnormality in the , h∈u; m is the jth abnormal data point v i Number;

[0089] f h (v n+1 ) to identify the threshold value, and find the neuron with the largest posterior probability density among all the output layer neurons. The type represented by this neuron is v n+1 abnormal category, that is, realizing the detection and identification of hidden abnormalities in welding.

[0090] Beneficial Effects

[0091] 1. The currently popular method of using a single industrial camera to extract welding image information for welding anomaly identification has a good effect on identifying welding appearance defects (such as weld nodules, slag inclusions, welding deformation, surface pores and surface cracks), but it is not ideal for identifying welding defects such as hidden pores, hidden cracks, cold welding, false welding, and virtual welding. Welding is a coupled process affected by multiple factors. The present invention discloses a welding hidden anomaly detection and identification method based on multi-source data. By adding data such as temperature, sound, spectrum, current and voltage data, combined with visual information, it can better characterize welding state changes and help detect and classify more types of welding faults. The comprehensive use of multiple data sources can improve the accuracy, robustness and interpretability of the welding anomaly classification model, thereby achieving more comprehensive and accurate online detection of welding quality and anomaly classification.

[0092] 2. The mutual coupling between welding process parameters makes it difficult to express welding quality with process parameters. The present invention discloses a welding hidden anomaly detection and identification method based on multi-source data, which adopts the LDA dimensionality reduction method. LDA dimensionality reduction can eliminate redundant features of data, reduce the complexity of the model, make the modeling process simple and intuitive, and also eliminate the influence of collinearity between data. In addition, LDA is a supervised dimensionality reduction method, which can perform feature dimensionality reduction according to type, and can better build a model with classification efficiency. The intrinsic dimension is estimated using statistical ideas, which has high accuracy and can be well applied to the LDA dimensionality reduction method.

[0093] 3. Since the normal data stations are the majority in the overall sample, directly using the neural network method to detect and classify outliers may affect the accuracy of the model. The present invention discloses a welding hidden anomaly detection and identification method based on multi-source data, which uses an outlier factor detection method to first detect outliers, and then uses a probabilistic neural network to classify the outliers. The local outlier factor LOF method, as a density-based anomaly detection algorithm, can discover outliers based on changes in neighborhood density, can more comprehensively capture invisible anomalies in the welding process, and detect abnormalities that are difficult to observe with a single information source.

[0094] 4. The LOF method needs to determine the K neighborhood of each point, so it is necessary to calculate the distance between each sample point and all other sample points in the data set to find the K neighborhood, which has high time complexity. The present invention discloses a welding hidden anomaly detection and identification method based on multi-source data, which adopts the isolation forest method to first divide the data into several subsets, and then select a new neighborhood search space based on the isolation forest idea (not the entire subset) in the subset to reduce the time complexity of the LOF search K neighborhood. At the same time, the isolation forest can effectively detect global anomalies. LOF is suitable for discovering local outliers and processing outliers in dense areas of the data set. It is more effective for data sets with uneven density. Therefore, the method of combining isolation forest and LOF can take into account both global and local outlier detection.

[0095] 5. The present invention discloses a method for detecting and identifying welding hidden anomalies based on multi-source data. For PNN, the only thing that needs to be adjusted is the smoothing factor σ. In PNN, the smoothing factor is used to control the smoothness of the model. It is used as a parameter of the Gaussian kernel function when calculating the similarity, affecting the adaptability of the model to the data distribution. When training the PNN model, the smoothing factor is usually adjusted through training data to make the model better adapt to the data. The artificial bee colony algorithm simulates the honey-collecting behavior of bees to solve some multidimensional and multimodal optimization problems in life. It was originally applied to numerical optimization problems. The smoothing factor of PNN determines the smoothness of the decision boundary. A larger smoothing factor will result in a smoother decision boundary, while a smaller smoothing factor will result in a sharper decision boundary. By optimizing the smoothing factor through the artificial bee colony algorithm, a smoothing factor value suitable for the problem can be found to control the complexity of the model and avoid over-smoothing or over-sharpening. The artificial bee colony algorithm has a global search capability, can explore a wider solution space, dynamically adjusts the value of the smoothing factor during the iteration process, can make the smoothing factor adaptively adjusted along with the optimization process, and the optimized smoothing factor has a higher abnormality recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 It is a flow chart of a welding hidden anomaly detection and identification method based on multi-source data provided by an embodiment of the present invention.

[0097] Figure 2 It is a schematic diagram of data abnormal point detection in a welding hidden abnormality detection and identification method based on multi-source data provided by an embodiment of the present invention.

[0098] Figure 3 It is a schematic diagram of abnormality type identification in a welding hidden abnormality detection and identification method based on multi-source data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0099] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0100] Embodiment 1:

[0101] In the field of welding anomaly detection, welding quality inspection generally uses a single welding quality information, such as welding image information using a convolutional neural network, or welding sound information using a sound sensor. This welding quality detection method using a single image information is an effective method, but it can only detect surface welding problems, and some hidden welding quality problems cannot be detected. The method using multi-source welding information can detect welding quality anomalies and abnormalities that are difficult to detect on the surface, such as lack of fusion, internal pores, etc. The present invention takes the welding between the upper and lower covers of the antenna cover in the manufacturing process of communication equipment as an example for theoretical verification.

[0102] like Figure 1 As shown, the present embodiment discloses a method for detecting and identifying welding hidden anomalies based on multi-source data, and the specific implementation steps are as follows:

[0103] 1. Welding information acquisition

[0104] Collect multimodal data of the welding between the upper and lower covers of the antenna cover. Use industrial CCD cameras (image acquisition devices), thermal imagers, current and voltage Hall sensors, microphones and spectrometers. The image acquisition device collects the front image of the molten pool and obtains the shape and size information of the molten pool. The thermal imager collects the temperature field information on both sides of the molten pool. The current and voltage Hall sensors collect the current and voltage information during welding. The microphone collects the sound information during the welding process. The spectrometer collects the arc spectrum information during welding. The controller is used to collect image, temperature, current and other data in real time and store them in the corresponding database. At the same time, during the data transmission process, multi-source data is synchronously transmitted and feature synchronization is processed.

[0105] Based on the original welding signal, features that are more conducive to subsequent real-time information processing and process inspection are extracted. For image information, the length and width of the molten pool and the grayscale value of the image are extracted. For temperature, current, voltage, sound and spectral information, the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness of the corresponding signal are extracted. For the extraction of these characteristic signals, the original signal is processed and extracted using dedicated related software on the market. For each welding object x′ i , both contain the length, width, and grayscale value information of the melt pool image [g 1 , g 2 , g 3 ], mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of temperature [t 1 , t 2 , t 3 , t 4 , t 5 , t 6 ], current mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information [c 1 , c 2 , c 3 , c 4 , c 5 , c 6 ], voltage mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information [v 1 , v 2 , v 3 , v 4 , v 5 , v 6], the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of the sound [a 1 , a 2 , a 3 , a 4 , a 5 , a 6 ], as well as the mean, mean square, variance, peak-to-peak value, kurtosis coefficient and skewness information of the spectrum [l 1 , l 2 , l 3 , l 4 , l 5 , l 6 ]. Then every data x′ i , which can be expressed as [g i,1 , …, g i,3 , t i,1 ,…,t i,6 , c i,1 , …, c i,6 , v i,1 , …, v i,6 , a i,1 , …, a i,6 , l i,1 , ..., l i,6 ]; where i∈[1,n]; in addition to the 33-dimensional process parameters, it also includes the numbers represented by the corresponding abnormal types, as shown in Table 1.

[0106] Table 1 Types of welding abnormalities

[0107]

[0108] 2. Normalization

[0109] The data x′ i Each eigenvalue of is transformed into the range of [-1, 1], and each multi-source data x′ is calculated by the following formula i Perform normalization:

[0110]

[0111] In the formula, max(x′ i,j ) is the multi-source data x′ i The maximum value in the jth dimension, min(x′ i,j ) is the multi-source data x′ i The minimum value in the jth dimension; where i∈[1,n]; after normalization, each welding object is y i , expressed as [G i,1 , ..., G i,3 , T i,1 , ..., T i,6 , C i,1, ..., C i,6 , V i,1 , ..., V i,6 , A i,1 , ..., A i,6 , L i,1 , ..., L i,6 ]; where i∈[1, n]; the normalized feature data is shown in Table 2;

[0112] Table 2. Feature data after normalization

[0113]

[0114] 3. Determine the intrinsic dimension

[0115] Using statistical thinking, if the variation of a set of data is caused by real situations, its eigenvalue should be greater than the average eigenvalue of random data with the same number of subjects and variables; if it is smaller than the average eigenvalue of random variables, it will be impossible to distinguish whether the variation represented by the factor is caused by real situations or random errors, and it will have no value to retain.

[0116] n = 1500 normalized data y i Form the matrix Y, let Y be:

[0117]

[0118] Get the matrix Y T The eigenvectors of Y

[0119] [λ g,1 , ..., λ g,3 ,λ t,1 , ..., λ t,6 ,λ c,1 , ..., λ c,6 ,λ v,1 , ..., λ v,6 ,λ a,1 , ..., λ a,6 ,λ l,1 , ..., λ l,6 ].

[0120] Generate a matrix X(b) of random data with the same number n and dimension as the Y matrix. Where b∈[1,a], m is the dimension of X(b), which is the same as the dimension of X. Find the eigenvector [λ(b) 1 ,λ(b) 2 , …, λ(b) m ], and further calculate the average value of all eigenvalues Where j is the data dimension, j∈[1,m]; then the mean eigenvector is [λ(B) 1 ,λ(B) 2 , …, λ(B) m ]. Compare matrices Y with the same dimension one by one T The eigenvectors of Y and the eigenvalues ​​in the mean eigenvector determine the matrix Y T The eigenvalue of Y is greater than that of matrix X (b) T The number of eigenvalues ​​of X(b), denoted as d, is used to obtain the normalized data y i The intrinsic dimension of is d=6.

[0121] 4. LDA Dimensionality Reduction

[0122] (1) The welding quality is divided into N = 7 categories, the number of welding objects is n = 1500, and the jth category set of welding objects is Y j , calculate the intra-class divergence S w :

[0123]

[0124] Where S wj represents the intra-class divergence of the jth class, μ j represents the welding object y of the jth type i The mean vector of , i∈[1,n], j∈[1,N];

[0125] (2) Calculate the inter-class divergence S n :

[0126]

[0127] where n j For the jth type of welding object y i The number of, μ is the welding object y i The mean vector of ;

[0128] (3) Using intra-class divergence S w and the inter-class divergence S b , and the target is:

[0129]

[0130] The matrix W is the matrix of the eigenvectors corresponding to the largest d eigenvalues; further deduction shows that the optimization objective function is:

[0131]

[0132] Where d is the intrinsic dimension of the data. Then for any normalized data y i , after dimensionality reduction, we get data x i =WT y i , i∈[1,n], n is the number of samples. Then the reduced-dimensional data x i Expressed as [x i,1 , x i,2 , x i,3 , ..., x i,d ]; the feature data after LDA dimension reduction is shown in Table 3; the weight matrix W of LDA is as follows:

[0133]

[0134] Table 3 Feature data after LDA dimensionality reduction

[0135]

[0136] 5. Data anomaly detection

[0137] The steps of data outlier detection are as follows: Figure 2 As shown;

[0138] 5.1 Using the isolation forest method, the data x is obtained after 1500 dimensionality reduction in step 3 i Perform segmentation, perform segmentation 3 times, i∈[1, 1500], and obtain 8 sample subsets X 1 , X 2 , X 3 , ..., X 8 .

[0139] In the first layer of segmentation, a dimension j is randomly selected, and then a data point x is selected. i , according to x i,j The value of divides the data set X into two parts X′ 1 and X′ 2 , X′ 1 Some j-th dimension is greater than x i,j , X′ 2 Some j-th dimension is less than or equal to x i,j ; The second layer further divides X′ 1 and X′ 2 , and so on to the third layer;

[0140] Determine the new welding object point x n+1 With the segmentation boundary x i,j If x n+1 The distance to the segmentation boundary is close, that is, min[ther(x i,j )]<x n+1,j <max[ther(x i,j )], then change the position of the segmentation boundary; then let x i,j for ther(x i,j );

[0141]

[0142] Among them, a i and b i are two random points in the sample subset with a distance k between them;

[0143] 5.2 Use Isolation Forest to select the neighborhood search space from the 8 sample subsets obtained in 5.1;

[0144] Calculate each dimension reduction data x i The average path length The calculation formula is:

[0145]

[0146] N = 100 is the number of spanning trees in the isolation forest, p i,j is the i-th dimension reduction data x i The length of the path in the jth tree, i∈[1,n];

[0147] According to the dimension reduction data x i The normal and abnormal data in the When , the point is added to the neighborhood search space, and the point in the sample subset search space is recorded as The number of divided sample subsets l∈[1,8];

[0148] 5.3 Calculate the new welding object point x n+1 LOF value;

[0149] Determine the new welding object x n+1 k-neighborhood of x; first determine x by the following formula n+1 The sample subset it belongs to is l, and the points in this sample subset are

[0150]

[0151] in, is a sample subset, the points in The average path length of

[0152] Then calculate the new welding object x n+1 With all points in the sample subset, The distance between

[0153]

[0154] Among them, j is The dimension of , j∈[1,d]; in the first, subset, the distance point The kth point The distance between

[0155]

[0156] Point x n+1 The k-neighborhood of point x n+1 Centered on The set of points, then point x n+1 The k-neighborhood of Point x n+1 The number of points in the k-neighborhood of Point x n+1 Points in the k-neighborhood of To point x n+1 The k-th reachable distance is:

[0157]

[0158] Calculate point x n+1 The local reachable density of is:

[0159]

[0160] Calculate point x n+1 Local anomaly factor:

[0161]

[0162] Calculate the training data x i and the test data point x n+i The local anomaly factors of x are shown in Table 4. n+i is a new welding object point; the local anomaly factor values ​​of the training data and the local anomaly factor values ​​of the test data are shown in Tables 4 and 5;

[0163] Table 4 Local anomaly factors of training data

[0164]

[0165] Table 5 Local anomaly factors of test data

[0166]

[0167] Determine whether a data point is abnormal based on the value range of the local anomaly factor.

[0168]

[0169] According to the test data point x n+iThe range of local outlier factors is used to identify abnormalities in the test data, where 0 represents a normal point and 1 represents an abnormal point. The abnormality identification of the test data points is shown in Table 6.

[0170] Table 6 Abnormal discrimination output of test data points

[0171]

[0172] 6. Identification of abnormal types

[0173] The steps for identifying abnormal types are as follows: Figure 3 As shown;

[0174] 6.1PNN model consists of input layer, hidden layer, summation layer and output layer;

[0175] Input the abnormal data point v after dimensionality reduction i , abnormal data point v i is the data point x after dimensionality reduction i For each abnormal point v i There are d dimensions [v i,1 , v i,2 , v i,3, …, v i,d ], d = 6, and there are 6 types of anomalies, which are input to the input layer; where i < n; the hidden layer and the summation layer are calculated together using the Parzen method to calculate the anomaly point v i The jth data point with the hth type of anomaly , h∈u; and weighted sum the hidden layer outputs of the same abnormal pattern, the relationship is as follows:

[0176]

[0177] Where σ is the smoothing factor, d is the dimension of the input vector, and v i is the input vector, is the jth abnormal data point of the hth type of abnormality; m is the jth abnormal data point v i Number;

[0178] The number of neurons in the output layer is the total number of abnormal categories; for f h (v i ) Perform threshold identification and find the neuron with the largest posterior probability density among all output layer neurons;

[0179] y=argmax(f h (v i ))

[0180] 6.2: Use the artificial bee colony algorithm ABC to optimize the smoothing factor σ of PNN; use the optimization model to identify the abnormal causes of the abnormal results of the abnormal detection model in step 4;

[0181] 6.2.1. Set the number of nectar sources to o, the number of leading bees, and the upper limit of the number of nectar source tests; determine the value range of the smoothing factor to be [0, 1]; randomly generate the initial smoothing factor, and its expression is

[0182]

[0183] where σ z is the zth nectar source, z∈[1,o]; j is σ z The dimension is the jth smoothing factor in PNN; the number of j is the number of outliers v i Number; A random number between 0 and 1;

[0184] 6.2.2. The correct number of PNN classifications is used as the fitness function to evaluate the effect of the smoothing factor; where Cor(σ z )[argmax(f(v i ))] means that for all abnormal data points v i , using the smoothing factor σ z The number of correct prediction results;

[0185] 6.2.3. Leading bee stage, at the initial smoothing factor σ z Find a new smoothing factor σ′ near i , the formula is;

[0186]

[0187] σ′ zj =max{0,min{v zj , 1}}

[0188] in is a random value between -1 and 1, σ kj is zj The random values ​​in; compare cor(σ z )[argmax(f(x i ))] and cor(σ′ z )[argmax(f(x i ))], where cor(σ′ z )[argmax(f(x i ))] is the smoothing factor σ′ z The number of correct prediction results; when cor(σ z )[argmax(f(xi ))]>cor(σ′ z )[argmax(f(x i ))], leaving cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As a new smoothing factor, when cor(σ z )[argmax(f(x i ))]<cor(σ′ z )[argmax(f(x i ))], leaving cor(σ′ z )[argmax(f(x i ))]'s smoothing factor σ′ z As a new smoothing factor, when cor(σ z )[argmax(f(x i ))]=cor(σ′ z )[argmax(f(x i ))], leaving cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As the new smoothing factor; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1;

[0189] 6.2.4. In the bee-following stage, the roulette wheel method is used to select the smoothing factor:

[0190]

[0191] Among them, p z To select nectar source z The probability of z∈[1,o]; the follower bee selects the nectar source σ from the optimal result of 6.2.3 according to the probability z , then in σ z Nearby, find the new smoothing factor σ′ according to the method in 6.2.3 i , leaving the better smoothing factor as σ z ; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1;

[0192] 6.2.5. Scouting bee stage, search for each nectar source and update the number of times R(σz ), when the preset threshold is reached and it is not updated, the honey source is updated using the following formula to search for the global optimal solution;

[0193]

[0194] According to the method of 6.2.3, compare the optimal σ obtained by 6.2.4 z The cor(σ z )[argmax(f(x i ))] and σ′ updated from 6.2.5 z cor(σ′ z )[argmax(f(x i ))], leaving the better smoothing factor as σ z ; Repeat steps 6.2.1 to 6.2.5 until the number of iterations is reached and the optimized smoothing factor is obtained. The smoothing factor of the probabilistic neural network is shown in Table 7;

[0195] Table 7 Smoothing factors of probabilistic neural networks

[0196]

[0197] 6.3. Use the optimized model to identify the abnormal cause of the abnormal result of the abnormal detection model in step 4; use the optimized PNN model to identify the new abnormal welding object v identified in step 4 n+1 Input, calculate the new outlier point v n+1 and the abnormal data point v after dimensionality reduction i The jth data point of the hth type of anomaly in The relationship between the hidden layer outputs of the same abnormal mode is weighted and summed, and the relationship is as follows:

[0198]

[0199] Where σ is the smoothing factor, d is the dimension of the input vector, and v n+1 is the new input vector, is the abnormal data point v after dimensionality reduction i The jth abnormal data point of the hth abnormality in the , h∈u; m is the jth abnormal data point v i Number;

[0200] f h (v n+1 ) to identify the threshold value, and find the neuron with the largest posterior probability density among all the output layer neurons. The type represented by this neuron is v n+1 The abnormal categories of the test data are shown in Table 8;

[0201] Table 8 Abnormal category output of test data

[0202]

[0203] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A welding hidden anomaly detection and identification method based on multi-source data, characterized in that: The following steps are included: Step 1: for n identical welding objects, multimodal data information of welding quality of each object is collected in real time and stored in a database; the multimodal data includes: front image of molten pool, temperature field information on both sides of molten pool, current and voltage information during welding, sound information during welding and arc spectrum information during welding; Step 2: Extract features from the multimodal data information collected in step 1 to obtain multi-source data; extract the size information of the molten pool and the grayscale value of the image from the front image of the molten pool; extract the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness of the corresponding signals from the temperature field information on both sides of the molten pool, the current and voltage information during welding, the sound information during welding and the arc spectrum information during welding; Step 3: Normalize the multi-source data extracted in step 2; then use the LDA method to reduce the dimension of the multi-source data; Step 4: Use the local outlier factor algorithm LOF to model the anomaly detection of multi-source data to obtain an anomaly detection model; identify new abnormal welding objects based on the anomaly detection model; Step 5: Construct a probabilistic neural network (PNN) network model for welding hidden anomaly detection and identification, use the ABC algorithm to optimize the smoothing factor σ of the PNN, obtain the probabilistic neural network (PNN) network optimization model, and realize welding hidden anomaly detection and identification based on the probabilistic neural network (PNN) network optimization model.

2. The method for detecting and identifying welding hidden anomalies based on multi-source data according to claim 1, characterized in that: The implementation method of step 2 is: Feature extraction is performed on the multimodal data information collected in step 1 to obtain multi-source data; the length, width, and grayscale value information of the molten pool image are [g1, g2, g3], the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient, and skewness information of the temperature are [t1, t2, t3, t4, t5, t6], and the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient, and skewness information of the current are [c1, c2, c3, c4, c5, c6], the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information of voltage is [v1, v2, v3, v4, v5, v6], the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information of sound is [a1, a2, a3, a4, a5, a6], and the mean, mean square value, variance, peak-to-peak value, kurtosis coefficient and skewness information of spectrum is [l1, l2, l3, l4, l5, l6]; Then each data x′ in n welding objects i , expressed as [g i,1 , ..., g i,3 , t i,1 , ..., t i,6 , C i,1 , ..., C i,6 , v i,1 , ..., v i,6 , a i,1 , ..., a i,6 , l i,1 , ..., l i,6 ]; where i∈[1,n].

3. A welding hidden anomaly detection and identification method based on multi-source data as claimed in claim 2, characterized in that: The implementation method of step three is: Step 3.1: Convert each eigenvalue to the range of [-1, 1], and use the following formula for each multi-source data x′ i Perform normalization: In the formula, max(x′ i,j ) is the multi-source data x′ i The maximum value in the jth dimension, min(x′ i,j ) is the multi-source data x′ i The minimum value in the jth dimension; where i∈[1, n]; each welding object after normalization is y i , expressed as [G i,1 , ..., G i,3 , T i,1 , ..., T i,6 , C i,1 , ..., C i,6 , V i,1 , ..., V i,6 , A i,1 , ..., A i,6 , L i,1 , ..., L i,6 ]; where i∈[1,n]; Step 3.2: n normalized data y i Form a matrix Y and estimate the intrinsic dimension of the matrix Y; Step 3.2.1, let Y be Get the matrix Y T The eigenvectors of Y [l g,1 ,...,l g,3 ,l t,1 ,...,l t,6 ,l c,1 ,...,l c,6 ,l v,1 ,...,l v,6 ,l a,1 ,...,l a,6 ,l l,1 ,...,l l,6 ]; Step 3.2.2, generate a matrix X(b) of random data with the same number n and dimension as the Y matrix; where b∈[1,a], m is the dimension of X(b), which is the same as the dimension of X; then X(b) has the following matrix form; Step 3.2.

3. Find the eigenvectors [λ(b)1, λ(b)2, …, λ(b)m] of the matrix X(b), and further calculate the average of all eigenvalues Where j is the data dimension, j∈[1,m]; then the mean eigenvector is [λ(B)1,λ(B)2,…,λ(B) m ]; Step 3.2.4: Compare matrices Y of the same dimension one by one T The eigenvectors of Y and the eigenvalues ​​in the mean eigenvector determine the matrix Y T The eigenvalue of Y is greater than that of matrix X (b) T The number of eigenvalues ​​of X(b) is denoted as d, then the normalized data y i The intrinsic dimension of is d; Step 3.3: Use LDA dimensionality reduction method to reduce the dimensionality of multi-source data; (1) The welding quality is divided into N categories, the number of welding objects is n, and the jth category set of welding objects is Y j , calculate the intra-class divergence S w : Where S wj represents the intra-class divergence of the jth class, μ j represents the welding object y of the jth type i The mean vector of , i∈[1,n], j∈[1,N]; (2) Calculate the inter-class divergence S b : where n j For the jth type of welding object y i The number of, μ is the welding object y i The mean vector of ; (3) Using intra-class divergence S w and the inter-class divergence S b , and the target is: The matrix W is the matrix of the eigenvectors corresponding to the largest d eigenvalues; further deduction shows that the optimization objective function is: Where d is the intrinsic dimension of the data; then for any normalized data y i , after dimensionality reduction, we get data x i =W T y i , i∈[1,n], n is the number of samples; then the reduced-dimensional data x i Expressed as [x i,1 , x i,2 , x i,3 , ..., x i,d ].

4. A welding hidden anomaly detection and identification method based on multi-source data as claimed in claim 3, characterized in that: Step 4 is implemented as follows: Step 4.1: Use the isolation forest method to select the first c layers of the isolation forest to reduce the n dimensions in step 3 to obtain the data x i Split, i∈[1,n], and get 2 c A sample subset; The first layer of segmentation selects a dimension j and then selects a data point x i , according to x i,j The value of divides the data set X into two parts X1 and X2, and the jth dimension of part X1 is greater than x i,j , the jth dimension of part X2 is less than or equal to x i,j ; The second layer further splits X1 and X2, and so on to the cth layer; Determine the new welding object point x n+1 With the segmentation boundary x i,j If x n+1 The distance to the segmentation boundary is close, that is, min[ther(x i,j )] <x n+1,j <max[ther(x i,j )], then change the position of the segmentation boundary; then let x i,j for ther(x i,j ); Among them, a i and b i are two random points in the sample subset with a distance k between them; Step 4.2: Use Isolation Forest to extract the 2 c Select the neighborhood search space from the sample subsets; Calculate each dimension reduction data x i The average path length The calculation formula is: N is the number of spanning trees in the isolation forest, p i,j is the i-th dimension reduction data x i The length of the path in the jth tree, i∈[1,n]; According to the dimension reduction data x i The normal and abnormal data in the When , the point is added to the neighborhood search space, and the point in the sample subset search space is recorded as The number of sample subsets divided is l∈[1,2 c ]; Step 4.3: Calculate the new welding object point x n+1 LOF value; Determine the new welding object x n+1 k-neighborhood of x; first determine x by the following formula n+1 The sample subset it belongs to is l, and the points in this sample subset are in, is a sample subset, the points in The average path length of Then calculate the new welding object x n+1 With all points in the sample subset, The distance between Among them, j is The dimension of , j∈[1,d]; in the first, subset, the distance point The kth point The distance between Point x n+1 The k-neighborhood of point x n+1 Centered on The set of points, then point x n+1 The k-neighborhood of Point x n+1 The number of points in the k-neighborhood of Point x n+1 Points in the k-neighborhood of To point x n+1 The k-th reachable distance is: Calculate point x n+1 The local reachable density of is: Calculate point x n+1 Local anomaly factor: in For x n+1 Points in the k-neighborhood of The local reachable density of Determine point x n+1 The range of local outlier factors is used to complete the anomaly detection modeling of multi-source data:

5. A welding hidden anomaly detection and identification method based on multi-source data as claimed in claim 4, characterized in that: The implementation method of step five is: Step 5.1: The PNN model consists of an input layer, a hidden layer, a summation layer, and an output layer; Input the abnormal data point v after dimensionality reduction i , abnormal data point v i is the data point x after dimensionality reduction i For each abnormal point v i There are d dimensions [v i,1 , v i,2 , v i,3 , …, v i,d ], and has an exception type in u, which is input to the input layer; where i < n; the Parzen method is used to calculate the hidden layer and the summation layer together to calculate the outlier v i with the j-th data point of the h-th outlier where h ∈ u; and the hidden layer outputs of the same outlier pattern are weighted and summed, and the relational expression is as follows: Where σ is the smoothing factor, d is the dimension of the input vector, and v i is the input vector, is the jth abnormal data point of the hth type of abnormality; m is the jth abnormal data point v i Number; The number of neurons in the output layer is the total number of abnormal categories; for f h (v i ) Perform threshold identification and find the neuron with the largest posterior probability density among all output layer neurons; y=argmax(f h (v i )) Step 5.2: Use the artificial bee colony algorithm ABC to optimize the smoothing factor σ of PNN; use the optimization model to identify the abnormal cause of the abnormal result of the abnormal detection model in step 4; Step 5.2.1, set the number of nectar sources to o, the number of leading bees, and the upper limit of the number of nectar source tests; Determine the value range of the smoothing factor to be [0, 1]; randomly generate the initial smoothing factor, then its expression is where σ z is the zth nectar source, z∈[1,o]; j is σ z The dimension is the jth smoothing factor in PNN; the number of j is the number of outliers v i Number; A random number between 0 and 1; Step 5.2.2: Use the correct number of PNN classifications as the fitness function to evaluate the effect of the smoothing factor; where cor(σ z )[argmax(f(v i ))] means that for all abnormal data points v i , using the smoothing factor σ z The number of correct prediction results; Step 5.2.3, Leading Bee Stage, at the initial smoothing factor σ z Find a new smoothing factor σ′ near i , the formula is; in zj =max{0,min{v zj ,1}} in is a random value between -1 and 1, σ kj is zj The random values ​​in; compare cor(σ z )[argmax(f(x i ))] and cor(σ′ z )[argmax(f(x i ))], where cor(σ′ z )[argmax(f(x i ))] is the smoothing factor σ′ z The number of correct prediction results; when cor(σ z )[argmax(f(x i ))]>cor(σ′ z )[argmax(f(x i ))], leaving cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As a new smoothing factor, when cor(σ z )[argmax(f(x i ))] <cor(σ′ z )[argmax(f(x i ))], leaving cor(σ′z)[argmax(f(x i ))]'s smoothing factor σ′ z As a new smoothing factor, when cor(σ z )[argmax(f(x i ))]=cor(σ′ z )[argmax(f(x i ))], leaving cor(σ z )[argmax(f(x i ))] with a smoothing factor σ z As the new smoothing factor; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1; Step 5.2.4, in the bee-following stage, the roulette wheel method is used to select the smoothing factor: Among them, p z To select nectar source z The probability of z∈[1,o]; the follower bee selects the nectar source σ from the optimal result of step 5.2.3 according to the probability z , in σ z Find the new smoothing factor σ′ according to the method in step 5.2.3 i , leaving the better smoothing factor as σ z ; R(σ z ) is the number of times the zth nectar source is updated, if σ z Transformation, R(σ z )=R(σ z )+1; Step 5.2.5, Scouting bee stage, search for the number of times each nectar source is updated R (σ z ), when the preset threshold is reached and it is not updated, the honey source is updated using the following formula to search for the global optimal solution; According to the method of step 5.2.3, compare the optimal σ obtained in step 5.2.4 z The cor(σ z )[argmax(f(x i ))] and σ′ updated by step 5.2.5 z cor(σ′ z )[argmax(f(x i ))], leaving the better smoothing factor as σ z ; Repeat steps 5.2.1 to 5.2.5 until the number of iterations is reached and the optimized smoothing factor is obtained; Step 5.3: Use the optimized model to identify the abnormal cause of the abnormal result of the abnormal detection model in step 4; use the optimized PNN model to identify the new abnormal welding object v identified in step 4. n+1 Input, calculate the new outlier point v n+1 and the abnormal data point v after dimensionality reduction i The jth data point of the hth type of anomaly in The relationship between the hidden layer outputs of the same abnormal mode is weighted and summed, and the relationship is as follows: Where σ is the smoothing factor, d is the dimension of the input vector, and v n+1 is the new input vector, is the abnormal data point v after dimensionality reduction i The jth abnormal data point of the hth abnormality in the , h∈u; m is the jth abnormal data point v i Number; f h (v n+1 ) to identify the threshold value, and find the neuron with the largest posterior probability density among all the output layer neurons. The type represented by this neuron is v n+1 abnormal category, that is, realizing the detection and identification of hidden abnormalities in welding.

Citation Information

Patent Citations

  • Smart home control system, control method, computer equipment and storage medium

    CN111856958A

  • Laser welding process monitoring method and system based on multi-source data fusion and medium

    CN118051871A

  • Welding method based on artificial intelligence and welding robot

    CN118237825A

  • Systems and methods for data-driven anomaly detection

    US20140379301A1

Cited By

  • Intelligent image data identification and analysis method, system and device based on AI model

    CN119807978A

  • Intelligent Image Data Recognition and Analysis Method, System and Device Based on AI Model

    CN119807978B

  • High-voltage bushing state diagnosis method and system

    CN120495194A

  • A high voltage bushing condition diagnosis method and system

    CN120495194B

  • Automobile injection molding part defect classification method based on database comparison

    CN120747643A