Sandwich gold purity evaluation method, device, equipment and medium based on mixed model

Abstract

The application discloses a sandwich gold purity evaluation method and device based on a mixed model, equipment and a medium. The method comprises the following steps: collecting original data; inputting the processed original data into a mixed model for training to obtain a trained mixed model; wherein the mixed model comprises a feature extraction module, a weighted fusion unit and a classification module connected in sequence; collecting a second pulse eddy current signal and a second supplementary feature corresponding to the to-be-detected gold; inputting the second pulse eddy current signal and the second supplementary feature into the trained mixed model to output a prediction vector; converting the prediction vector into a one-hot code to obtain the number of the to-be-detected gold through the one-hot code. The problem that the existing classification mixed model cannot be applied to the high-fraud gold purity detection scene is solved, the classification accuracy of the fraud gold purity detection is improved, and the application range of the classification mixed model in the rare sample detection field is widened.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pulsed eddy current metal detection using deep learning methods, and particularly relates to a laminated gold purity evaluation method and device based on a hybrid model, equipment and medium. BACKGROUND

[0002] Currently, gold purity detection mainly relies on traditional methods such as chemical analysis, mass spectrometry and spectral analysis. Although these methods have high precision, they generally have problems such as complex operation, long detection period, dependence on professional equipment and personnel, and most belong to destructive detection, which is difficult to apply to large-scale or on-site rapid detection scenarios. As a non-destructive detection technology, pulsed eddy current detection (PECT) has been applied to the detection of gold purity of laminated gold plates.

[0003] For PEC signal processing, the existing technology mostly uses traditional machine learning methods. Such methods rely on manual feature design and have limited representation ability for complex and high-dimensional PEC signals. In particular, in the case of small sample size, the classification accuracy and generalization ability are insufficient.

[0004] In recent years, deep learning (DL) methods have shown great ability in time series classification. For example, models such as multi-layer perceptron (MLP) can automatically learn high-dimensional features. However, in gold purity evaluation, pure deep learning models are prone to overfitting in small sample cases, and have insufficient ability to distinguish subtle differences in high-purity adulterated gold.

[0005] In the prior art, a hybrid algorithm of feedforward neural network and support vector machine has been used for classification tasks, such as a hybrid model of MLP and SVM. The MLP model extracts features and the SVM model completes classification, improving the recognition effect in some scenarios. Since gold samples are scarce and expensive, if the existing MLP-SVM hybrid model is directly applied to the gold detection scenario, on the one hand, the feature dimension generated by the output layer of the MLP model is too low, the information density is insufficient, and it is difficult to carry the key discriminant features of the gold sample purity; plus the small amount of gold sample data, the SVM model cannot learn effective training information from low-quality features; ultimately causing model overfitting and a significant decline in classification accuracy; on the other hand, the existing model usually directly inputs the features into the SVM model for training after the MLP model output features, and adopts a non-discriminatory processing strategy for the features, which cannot highlight the core features in the gold pulsed eddy current signal that are directly related to purity, ultimately leading to difficulty in adapting to the high-precision requirements of gold purity detection. SUMMARY

[0006] The application aims to provide a sandwich gold purity evaluation method, device, equipment and medium based on a hybrid model, so as to solve the problem that the existing classification hybrid model cannot be applied to high-fraud gold purity detection scenarios due to low feature information density, poor small sample adaptability and indiscriminate processing of features, thereby improving the classification accuracy of fraud gold purity detection and widening the application range of the classification hybrid model in the field of scarce sample detection.

[0007] To achieve the above-mentioned purpose, the technical solution adopted by the application is as follows: a sandwich gold purity evaluation method based on a hybrid model, the method comprising the following steps:

[0008] S1: collecting original data; wherein the original data comprises a first pulsed eddy current signal and a first supplementary feature;

[0009] The first pulsed eddy current signal is a signal returned by the pulsed eddy current detection equipment after the pulsed eddy current detection equipment sends a pulsed waveform to the test gold and the test gold generates eddy current;

[0010] S2: inputting the processed original data into the hybrid model for training to obtain a trained hybrid model; wherein the hybrid model comprises a feature extraction module, a weighted fusion unit and a classification module connected in sequence;

[0011] S3: collecting a second pulsed eddy current signal and a second supplementary feature corresponding to the gold to be tested;

[0012] S4: inputting the second pulsed eddy current signal and the second supplementary feature into the trained hybrid model to output a prediction vector;

[0013] S5: converting the prediction vector into a one-hot code, obtaining a corresponding number of the gold to be tested through the one-hot code, and the number is used to reflect the purity of the gold.

[0014] Further, the specific steps of training the hybrid model comprise:

[0015] S2.1: inputting the processed original data into the initial model corresponding to the feature extraction module for training to obtain a trained initial model, wherein the initial model comprises at least an input layer, a hidden layer and a classification layer;

[0016] S2.2: removing the classification layer of the initial model to obtain the feature extraction module;

[0017] S2.3: inputting the processed original data into the feature extraction module to obtain time series feature data;

[0018] S2.4: inputting the time series feature data into the weighted fusion unit to obtain optimized time series feature data;

[0019] S2.5: input the optimized time series feature data into the classification module for training to obtain a trained classification module;

[0020] S2.6: sequentially connect the feature output end of the feature extraction module, the weighted fusion unit and the feature input end of the trained classification module to obtain a trained hybrid model.

[0021] Further, in the weighted fusion unit, based on the decay physical characteristics of the pulsed eddy current signal, the time series feature data is divided into feature intervals corresponding to the decay stage, and the corresponding feature intervals are weighted and fused.

[0022] Further, in the S2.4, the feature intervals corresponding to the decay stage are specifically: the rising edge section, the rapid decay section, and the slow decay section.

[0023] Further, in the S2.4, the weighted fusion adopts weighted splicing fusion, multiplies the feature intervals by the corresponding weights respectively, and splices them in the original order to obtain the optimized time series feature data.

[0024] Further, the first supplementary feature and the second supplementary feature both include the environmental temperature when collecting the pulsed eddy current signal, and the probe temperature, power supply temperature and heat sink temperature of the pulsed eddy current detection device.

[0025] Further, the initial model includes an MLPM model, and the MLPM model has two data input channels, wherein the first channel is used for inputting the original data, and the second channel is used for inputting the processed original data, wherein the processed original data includes the normalized original data.

[0026] Based on the same concept, the application also provides a hybrid model-based interlayer gold purity evaluation device, which comprises:

[0027] A first acquisition module is configured to acquire original data, wherein the original data includes a first pulsed eddy current signal and a first supplementary feature.

[0028] A hybrid model training module is configured to input the processed original data into a hybrid model for training to obtain a trained hybrid model.

[0029] A second acquisition module is configured to acquire test data of a test gold, wherein the test data includes a second pulsed eddy current signal corresponding to the test gold and a second supplementary feature.

[0030] A gold purity prediction module is configured to predict the purity of the test gold and output a prediction vector.

[0031] The gold purity result identification module is configured to obtain the purity result of the gold to be tested according to the prediction vector.

[0032] Based on the same concept, the present application also provides an electronic device,

[0033] The device comprises:

[0034] A memory is configured to store a computer program.

[0035] A processor is configured to implement the mixed model-based sandwich gold purity evaluation method as described above when the computer program is executed.

[0036] Based on the same concept, the present application also provides a computer-readable storage medium,

[0037] The computer-readable storage medium stores a computer program, and the computer program is configured to implement the mixed model-based sandwich gold purity evaluation method as described above when executed by a processor.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] The present application provides a mixed model, MLPM-weighted fusion unit-SVM model, specifically comprising: removing the classification layer of the MLPM model, based on the MLPM model with the classification layer removed, performing deep feature extraction on the input original data, inputting the extracted high-dimensional features into the SVM model for training, and further constructing a mixed model training method suitable for gold purity detection; the design enlarges the difference of the feature data output by the MLPM model, solves the problem that the SVM model cannot obtain effective training information due to insufficient data dimension of the prediction vector output by the classification layer of the MLPM model in gold purity detection; further based on the decay physical property of the pulse eddy current signal, the data after feature extraction are weighted and fused, the features of the fast decay segment strongly related to the purity in gold detection are strengthened, the SVM model is facilitated to preferentially learn the information directly related to the gold purity, and the recognition accuracy of high-precision fake gold is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only one embodiment of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0041] Figure 1 is a flowchart of the mixed model-based sandwich gold purity evaluation method in the embodiment of the present application;

[0042] Figure 2 is a flowchart of training a mixed model according to an embodiment of the present application;

[0043] Figure 3 is a PEC signal diagram of a same piece of gold sample measured by a pulse eddy current device according to an embodiment of the present application at a probe in a low temperature state and a high temperature state;

[0044] Figure 4 is a typical image of gold PEC signal detected by a device according to an embodiment of the present application;

[0045] Figure 5 is an architecture of an MLPM-weighted fusion unit-SVM model and a training schematic diagram thereof according to an embodiment of the present application, Figure 5 (a) is a schematic diagram of an overall training process of the MLPM-weighted fusion unit-SVM model, Figure 5 (b) is a schematic diagram of a training process of an SVM model according to an embodiment of the present application, Figure 5 (c) is a schematic diagram of a data analysis process of the MLPM-weighted fusion unit-SVM model according to an embodiment of the present application;

[0046] Figure 6 is a confusion matrix diagram of a prediction result of an SVM model according to an embodiment of the present application;

[0047] Figure 7 is a confusion matrix diagram of a prediction result of an MLPM model according to an embodiment of the present application;

[0048] Figure 8 is a confusion matrix diagram of a prediction result of the MLPM-weighted fusion unit-SVM model according to an embodiment of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0050] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0051] Embodiment 1

[0052] Figure 1 shows a flowchart of a sandwich gold purity evaluation method based on a mixed model according to an embodiment of the present application. As shown in the figure, Figure 1As shown, the interlayer gold purity evaluation method comprises the following steps:

[0053] S1: Collecting original data.

[0054] Pulsed eddy current detection has been verified to be effective in gold adulteration detection. Compared with traditional detection methods, PECT has the advantages of fast, low cost and non-contact, and therefore has wide application prospects in the detection and evaluation of gold adulteration.

[0055] A large number of PEC signals and supplementary features of various gold samples are collected by the pulsed eddy current detection device as input data of the mixed model, wherein the PEC signal is the signal returned by the gold sample to the pulsed eddy current detection device after the pulsed eddy current detection device sends a pulse waveform to the gold sample and the gold sample generates eddy current. The supplementary features include environmental temperature, and probe temperature, power supply temperature and heat sink temperature of the pulsed eddy current detection device. The environmental temperature is the environmental temperature at the time of testing, which is collected by a sensor. The probe temperature is the probe temperature of the pulsed eddy current detection device, and the power supply temperature and the heat sink temperature are the temperatures of the mixed power supply module of the pulsed eddy current detection device. The above three temperatures are obtained by software monitoring of the temperature control module of the pulsed eddy current detection device, and the above temperatures are controlled within a certain range by the temperature control module. The probe temperature of the pulsed eddy current detection device will affect the results, Figure 3 The correlation between the pulse eddy current amplitude and the probe temperature of the same gold sample is shown. It can be seen that the amplitude and the probe temperature of the same gold sample show correlation during collection, so the probe temperature is also used as data for model training, and the other three temperatures are also incorporated into the training set as supplementary features. The PEC signal is a 1000-dimensional time series, and the environmental temperature, probe temperature, power supply temperature and heat sink temperature are 1-dimensional time series, and the input features are 1004-dimensional feature vectors.

[0056] The purity and other properties of each piece of gold are shown in Table 1. The PEC data of 13 kinds of gold samples with different purity and thickness are collected by the pulsed eddy current detection device. The training set and the test set both include data of 13 kinds of gold samples. 200 groups of data are collected for each gold sample, and 4 data are collected through rotation and flipping operations, so that 800 data are obtained for each gold sample. In the training set and the test set, the sample data of each gold sample is 400. Therefore, the dimensions of the training set and the test set are both (5200, 1004).

[0057] Table 1: Gold purity attribute table

[0058]

[0059] S2: input the processed original data into the mixed model for training to obtain a trained mixed model, wherein the mixed model comprises a feature extraction module, a weighted fusion unit and a classification module connected in sequence.

[0060] The processing of the original data comprises standardizing the original data, and the standardization method adopts a Z-score normalization method, so that the mean of the data in each dimension is zero and the standard deviation is 1. Then, the obtained training data is used for network training, and the test set is used for performance evaluation of the model.

[0061] The standardization method adopts a Z-score normalization method, which is calculated in the following manner:

[0062] ,

[0063] Wherein, is the collected original data, i is the number of dimensions, j is the data amount of each dimension, x ji is the jth sample data in i dimensions, μ i and σ i are the mean and standard deviation of each dimension X i = (x 1i , x 2i , x 3i, , …, x ji ), z ji represents the normalized data of the jth sample data in i dimensions. Wherein, X i = (x i1 , x i2 , x i3 …x ij ) is a flattened tensor when the dimension is i, the number of dimensions i is 1004, and j is 5200 in the training set and the test set. μ i and σ i are the mean and standard deviation of X j = (x 1j , x 2j , x 3j, , …, x ij ), and z ij represents the normalized data of the jth sample data in i dimensions.

[0064] Wherein, μ i and σ i are calculated by the following formulas:

[0065] ,

[0066] ,

[0067] N = 5200.

[0068] wherein, Figure 4 The figure shows the amplitude of the induced pulse eddy current generated in the sample under test with respect to time, dimension i is obtained by dividing the pulse eddy current decay signal into 0.1 ms intervals for the first 100 ms, obtaining 1000 nodes, taking the amplitude signal data corresponding to the 1000 nodes as the first 1000 dimensions, and taking the environmental temperature, probe temperature, power supply temperature and heat sink temperature data as the 1001st, 1002nd, 1003rd and 1004th dimensions respectively, and j represents the sample data of the gold sample.

[0069] The training process of the hybrid model will be described in detail below.

[0070] Because gold samples are valuable and scarce, it is difficult to obtain a large amount of real and labeled gold sample data. Because gold is expensive and supply is limited, it is costly and difficult to practice to collect a large amount of PEC data with purity labels. Therefore, it is of great significance to introduce deep learning technology to classify gold into multiple categories (e.g. pure gold and adulterated gold), collect and analyze a small amount of labeled data, and then establish a classification model according to the degree of adulteration to identify the purity of gold.

[0071] S2.1: input the processed original data into the initial model corresponding to the feature extraction module for training to obtain the trained initial model, wherein the initial model at least includes an input layer, a hidden layer and a classification layer.

[0072] In the embodiment of the present application, the initial model is a variant model of the multi-layer perceptron (MLP) model, namely the multi-input channel MLP-Mixer (MLPM) model. The MLPM model is suitable for learning feature representation from time series data. It extracts key features of data by stacking multiple network layers that fuse local features, and can accurately capture subtle changes and evolution trends in different time periods. This feature makes the MLPM model have a significant advantage in identifying complex patterns and abnormal events in PEC time series, and shows higher recognition accuracy and robustness.

[0073] The MLPM model comprises a dual-channel input layer, a hidden layer for stacking feature extraction, and a classification layer for outputting a feature vector. One of the two data input channels of the dual-channel input layer is used to input the original data, and the other is used to input the standardized data. Through the dual-channel design, the influence of probe temperature fluctuations can be eliminated. Of course, the initial model of the embodiment of the present application is not limited to a specific model, and any deep learning model with classification function can be applied to the embodiment of the present application, such as the existing MLP model, LSTM model. In the above single-channel input model, the standardized data is input into the test channel.

[0074] In the embodiment of the present application, before training the MLPM model, the hyperparameters of the MLPM model are set, and the training environment of the MLPM model is configured. By setting appropriate parameters, the convergence and stability of the model can be guaranteed, and the generalization ability of the model can be improved. Specifically, the number of blocks is set to 36, the learning rate is initialized to 1 × 10 −3 , the number of training iterations epoch is 200. The Adam optimizer is used to update all weights and biases of the model, and the L2 regularization parameter is 1 × 10 −3 . The cross-entropy loss function is used. The gradient function is used to update the model parameters. The batch size is 120. When the change of Loss tends to be stable during the training process, it is considered that the model has converged. In order to prevent overfitting, early stopping operation is performed at this time to obtain the trained MLPM model.

[0075] The training process is coded using the Pytorch machine learning library and trained on an Nvidia GeForce RTX 4070Ti GPU. The training of MLPM-36 for 200 iterations takes 11 minutes and 3.6 seconds. As the number of blocks increases, the model's calculations will become more complex, and the training time will also become longer.

[0076] S2.2: Remove the classification layer of the initial model to obtain the feature extraction module.

[0077] The neural network architecture after removing the classification layer of the MLPM model is called MLPM-P. The MLPM-P is used to reduce the dimension of the data and enlarge the difference between different feature data. The MLPM-P is responsible for extracting discriminative features from high-dimensional PEC signals to achieve the function of feature extraction.

[0078] S2.3: Input the processed original data into the feature extraction module to obtain time series feature data.

[0079] The dataset for training the MLPM model is input again into the MLPM-P, the feature data is obtained through forward calculation using the weights of the MLPM model and without backward propagation parameter update, and time series feature data output by the MLPM-P, namely P data, is obtained. The P data has a proper dimension size, and the SVM model can better extract useful feature information. The reason why the data after the classification layer of the MLPM model is not used to train the SVM model is that the data type output by the classification layer is a feature vector corresponding to a one-hot code, the data dimension is too low, and the feature that can be represented is too little, so that the subsequent classification model cannot obtain effective training information.

[0080] S2.4: inputting the time series feature data into a weighted fusion unit to obtain optimized time series feature data.

[0081] According to Figure 4 As can be seen from the image of the induced pulse eddy current amplitude change with time, the pulse eddy current signal curve is roughly divided into three stages with time: the rising edge section of 0-5 ms, the rapid decay section of 5-50 ms, and the slow decay section of 50-100 ms. In the rising edge section, eddy current is rapidly generated in the measured gold after the excitation signal is applied; in the rapid decay section, the eddy current rapidly spreads and the energy rapidly dissipates in the gold, in this stage, the signal amplitude decreases at the fastest rate and is strongly related to the electrical conductivity of the gold (directly related to the purity of the gold); in the slow decay section, the eddy current energy has been mostly dissipated, only a small amount of eddy current slowly decays, and the signal decreases significantly in the slow decay section.

[0082] It can be seen that, of the 1000 input nodes, the 500 nodes in the rapid decay section are core features, and the 500 nodes in the rising edge section and the slow decay section are redundant data. If the P data is directly input into the subsequent classification model for equivalent processing, the core features will be diluted by the redundant data, which reduces the classification accuracy of the model in the high-purity small sample detection scenario of gold. Based on this, the embodiment of the application divides the P data into feature intervals corresponding to the decay stages according to the physical law of the pulse eddy current signal change with time, and performs weighted fusion on the corresponding feature intervals.

[0083] The specific steps of the weighted fusion include giving different weights to the P data corresponding to different stages according to the physical prior knowledge of gold detection. For the rapid decay section as the core feature section, the weight is set to 0.6, and for the rising edge section and the slow decay section, the weight is set to 0.2.

[0084] The P data corresponding to the feature intervals is multiplied by the corresponding weight respectively to obtain feature data with physical meaning. The weighted P data is spliced in the original order to obtain optimized P data.

[0085] By multiplying each segment feature by a weight and then using a weighted splicing fusion method, the contribution of the core interval feature is strengthened while the original feature dimension and order of each interval are completely preserved, facilitating subsequent feature backtracking.

[0086] S2.5: input the optimized time series feature data into the classification module for training to obtain the trained classification module.

[0087] The SVM model is a classic machine learning algorithm that has good adaptability to small samples and nonlinear data. The SVM model can efficiently classify with fewer samples and can handle nonlinear patterns through kernel function conversion. Therefore, the SVM model performs well in handling the sparse labels and complex classification problems of PEC time series.

[0088] On high-dimensional data, neural networks generally perform better than SVM models in classification. The reasons are as follows: (1) Neural networks can learn nonlinear relationships through multiple hidden layers and activation functions, so they can handle more complex data distributions. In high-dimensional data, there may be nonlinear relationships, while SVM models can only introduce nonlinearity through kernel functions, which is not applicable in some cases. (2) Neural networks can automatically learn feature representations through backpropagation algorithms, which means they can extract the most discriminative features from raw data. In contrast, SVM models require manual selection and design of features, and feature selection may be more difficult for high-dimensional data. (3) Neural networks have an advantage in handling large-scale data sets. By using deep learning frameworks and parallel computing techniques, neural networks can effectively handle large amounts of high-dimensional data and can improve classification performance by increasing the number of network layers and nodes.

[0089] Therefore, neural networks are suitable for handling high-dimensional data, but SVM models have an advantage in cases where data is linearly separable or has a low feature dimension. Moreover, SVM models can provide an explanation for the classification decision. On low-dimensional data, the classification performance of SVM models is superior to that of neural networks: (1) For low-dimensional data, when the data is linearly separable, the SVM model can quickly determine the optimal hyperplane to separate data of different categories. However, the neural network may learn an unnecessarily more complex model. (2) The SVM model has strong robustness to outliers. On low-dimensional data, scattered points or outliers may easily occur. Neural networks are easily affected by these disturbances, while the SVM model can adjust the fault tolerance of the model by introducing a penalty parameter C, thereby reducing the impact of these disturbances. Therefore, the SVM model is more robust on low-dimensional data. (3) On low-dimensional data, neural networks may require more data for training and require more computing resources and time. In contrast, the SVM model is faster to train and performs well on feature-poor data sets.

[0090] In summary, the MLPM model as a neural network is suitable for handling high-dimensional data, while the SVM model is more suitable for handling low-dimensional data. Combining the MLPM model and the SVM model can take full advantage of their respective advantages. The MLPM model is used to extract the feature representation of the PEC time series, capturing the high-dimensional information in the data. Then, these features are input into the SVM model for classification to take advantage of its nonlinear classification ability and low-dimensional data classification advantage.

[0091] The combination of the MLPM model and the SVM model can complement each other's weaknesses. The MLPM model excels in extracting complex features of time series data, while the SVM model is robust to classification problems and small sample, nonlinear data sets. By combining the advantages of both, more accurate and reliable PEC time series classification results can be obtained. This is particularly suitable for high-purity gold detection scenarios with a purity of up to 99.9%, effectively distinguishing the subtle feature differences between high-purity gold and trace amounts of adulterated gold. Therefore, the MLPM model is combined with the SVM model, and the SVM model is used to replace the classification layer of the MLPM model.

[0092] The SVM model is used to finely classify the low-dimensional features output by the MLPM-P. The optimized P data is input into the SVM model for training, and the SVM model uses a radial basis function kernel with a penalty factor C set to 4.

[0093] S2.6: sequentially connecting the feature output end of the feature extraction module, the weighted fusion unit, and the feature input end of the trained classification module to obtain a trained hybrid model.

[0094] Figure 5 The architecture of the MLPM-weighted fusion unit-SVM model and its training schematic diagram in the embodiment of the application are shown. Among them, as shown in Figure 5 (a), the overall training process of the MLPM-weighted fusion unit-SVM model is: first using the trained MLPM model, then removing the classification layer of the MLPM model to obtain MLPM-P, replacing the classification layer of the MLPM model with the SVM model to obtain the final MLPM-weighted fusion unit-SVM model. Among them, Figure 5 (b) shows the training process of the SVM model, including using the trained MLPM model, then inputting all the training set data into the MLPM-P for calculation, extracting all P data, and performing weighted fusion optimization on the P data, and inputting the optimized P data into the SVM model for training. Figure 5 (c) shows the data analysis process of the MLPM-weighted fusion unit-SVM model, mainly including: the data is first processed by the MLPM-P, then the data is weighted and fused, and the optimized input is input into the SVM model again for analysis to obtain the final predicted sample label.

[0095] After obtaining the mixed model, the test data which has also been subjected to Z-score standardization is input into the mixed model for calculation to obtain the predicted gold label. Compared with the results of using the SVM model or the MLPM model alone, the sample amount of the training set used by the SVM model and the MLPM model used for the control experiment is twice that of the mixed model of the embodiment of the application, that is, the size dimension of the training set of the two models is (10400, 1004). In addition, the other parameters of the model are consistent with the mixed model.

[0096] The test results are shown in Table 2, and the confusion matrices of the prediction classification results of the three models are shown in Figure 6 、 Figure 7 、 Figure 8 Among them, the accuracy of the MLPM-weighted fusion unit-SVM model is the highest, which is 95.3%, and compared with the MLPM model alone, the accuracy is improved by about 3%.

[0097] Table 2 Classification results

[0098]

[0099] The main differences of the three models are compared by comparing the confusion matrix of SVM, MLPM and MLPM-weighted fusion unit-SVM. From the attribute table of 13 gold samples, it can be known that the purity of No. 12 gold sample is 99.90%, and the purity of No. 13 gold sample is 99.99%. The MLPM-weighted fusion unit-SVM model only misclassifies No. 13 gold sample, and the total misjudgment number is only 6. Therefore, compared with the MLPM model, the advantage of the MLPM-weighted fusion unit-SVM model mainly lies in more accurately classifying No. 12 gold sample and No. 13 gold sample. The purity of No. 12 gold sample and No. 13 gold sample is very similar. At the same time, from No. 8 gold sample with purity of 96% to No. 13 gold sample, the SVM model misclassifies only No. 13 gold sample by 16, which shows that in the subdivision of high-purity adulterated gold, the effect of the SVM model is better than that of the MLPM model, and for the classification of low-purity adulterated gold, the MLPM model is obviously better than the SVM model. Therefore, the MLPM-weighted fusion unit-SVM model integrates the advantages of the MLPM model and the SVM model, and makes up for each other's advantages, so as to achieve better prediction and classification effect. At the same time, the training set data amount of the MLPM-weighted fusion unit-SVM model is only half of that of the MLPM model and the SVM model, and the MLPM-weighted fusion unit-SVM model can still distinguish No. 12 and No. 13 more accurately. The MLPM-weighted fusion unit-SVM model can achieve higher prediction accuracy than the MLPM model and the SVM model by using only small sample training data, and the classification of high-purity adulterated gold is more detailed. In reducing the number of samples, the detection accuracy of the mixed classification model is improved, and the demand for samples in the detection scene such as gold detection is reduced.

[0100] Through the mixed architecture of "MLPM-P + weighted fusion unit + SVM", the organic combination of high-dimensional feature extraction and low-dimensional fine classification is realized; under the condition of small sample training, the classification precision and generalization ability are significantly improved through feature dimension reduction and model fusion; the above experiments prove that the MLPM-weighted fusion unit-SVM model can still obtain higher accuracy under the condition that the training set is only half of the control model, especially in the subdivision classification of high-purity gold.

[0101] S3: Collect the pulse eddy current signal and supplementary features corresponding to the gold to be tested.

[0102] The pulse eddy current detection equipment is used to collect the pulse eddy current signal and the supplementary feature corresponding to the gold to be detected. In the specific detection process, the pulse eddy current signal is collected by placing the gold to be detected above the sensor of the pulse eddy current detection equipment, and then the pulse eddy current detection equipment generates a rectangular wave, which is amplified by a power amplification module and then loaded into an excitation coil in the sensor. The excitation coil thus generates a primary magnetic field with a rectangular time-varying intensity. When the primary magnetic field jumps, an eddy current is induced in the gold plate. The induction and dissipation process of the eddy current is affected by the purity of the gold plate. During the induction and dissipation process of the eddy current, the eddy current generates a time-varying secondary magnetic field. The time-varying secondary magnetic field is converted into a voltage signal by a receiving coil in the sensor, amplified by a signal amplification module, and collected by a signal generation and collection dual-function module. The probe temperature of the pulse eddy current detection equipment is obtained by software monitoring the probe temperature. The collected data is preprocessed to obtain time series data.

[0103] S4: inputting the pulse eddy current signal and the supplementary feature into the trained hybrid model to output a prediction vector.

[0104] S5: converting the prediction vector of the gold to be detected into a one-hot code by using an argmax function, obtaining the corresponding number of the gold to be detected by using the one-hot code, and obtaining the corresponding purity value according to the number in Table 1 to evaluate the purity of the gold.

[0105] Embodiment 2

[0106] Based on the same concept, the embodiment of the present application also provides a hybrid model-based interlayer gold purity evaluation device, which comprises:

[0107] A first acquisition module is configured to acquire original data, wherein the original data comprises a first pulse eddy current signal and a first supplementary feature.

[0108] A hybrid model training module is configured to input the processed original data into a hybrid model for training to obtain a trained hybrid model.

[0109] A second acquisition module is configured to acquire test data of the gold to be detected, wherein the test data comprises a second pulse eddy current signal and a second supplementary feature corresponding to the gold to be detected.

[0110] A gold purity prediction module is configured to predict the purity of the gold to be detected and output a prediction vector.

[0111] A gold purity result identification module is configured to obtain the purity result of the gold to be detected according to the prediction vector.

[0112] Embodiment 3

[0113] Based on the same concept, the embodiment of the present application also provides an electronic device, which comprises:

[0114] a memory for storing a computer program;

[0115] a processor for implementing the mixed model based sandwich gold purity evaluation method of embodiment 1 when executing the computer program.

[0116] Embodiment 4

[0117] Based on the same concept, the embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the mixed model based sandwich gold purity evaluation method of embodiment 1.

[0118] The embodiment of the present application provides a mixed model based sandwich gold purity evaluation method, which collects PEC data of a gold block by a pulsed eddy current detection device, and evaluates the degree of gold adulteration by using deep learning technology. First, the relationship between the PEC signal collected during gold collection, the pulsed eddy current detection device and the environmental temperature data is analyzed to determine the training data set; second, the specific process of training the MLPM-weighted fusion unit-SVM model is given; finally, the MLPM-weighted fusion unit-SVM model is applied to classify the adulterated gold to obtain the final classification result. The mixed model is applied to the field of gold adulteration evaluation, and the experimental results show that compared with the MLPM model or the SVM model alone, the mixed model MLPM-weighted fusion unit-SVM has better classification effect in the gold adulteration evaluation task, and in the case of training with small sample data, it can still accurately classify gold samples with different adulteration degrees, and has sufficient generalization. And the feature data is bound with the physical attenuation stage of the gold pulsed eddy current signal, and the P data with physical meaning is designed, which solves the problem of local correlation loss when the traditional mixed model processes high-dimensional time series data. Therefore, the gold adulteration evaluation method provided by the present application has wide application prospect.

[0119] The content illustrated in the above embodiments should be understood as the embodiments only for more clearly illustrating the present application, and should not be used to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications of the present application, which all fall within the scope defined by the appended claims of the present application.

Claims

1. A method for evaluating the purity of a sandwiched gold based on a mixture model, characterized by, The method comprises the following steps: S1: collecting original data; wherein the original data comprises a first pulsed eddy current signal and a first supplementary feature; Wherein, the first pulsed eddy current signal is the signal returned by the test gold to the pulsed eddy current detection device after the pulsed eddy current detection device sends a pulsed waveform to the test gold and the test gold induces eddy current; S2: input the processed original data into the hybrid model for training to obtain the trained hybrid model; wherein the hybrid model comprises a feature extraction module, a weighted fusion unit and a classification module connected in sequence; wherein the specific steps of training the hybrid model comprise: S2.1: input the processed original data into the initial model corresponding to the feature extraction module for training to obtain the trained initial model, wherein the initial model comprises at least an input layer, a hidden layer and a classification layer; S2.2: remove the classification layer of the initial model to obtain the feature extraction module; S2.3: input the processed original data into the feature extraction module to obtain time series feature data; S2.4: input the time series feature data into the weighted fusion unit to obtain optimized time series feature data; Wherein, the weighted fusion unit divides the time series feature data into feature intervals corresponding to the decay stage based on the decay physical characteristics of the pulsed eddy current signal, and performs weighted fusion on the corresponding feature intervals; S2.5: input the optimized time series feature data into the classification module for training to obtain the trained classification module; S2.6: connect the feature output end of the feature extraction module, the weighted fusion unit and the feature input end of the trained classification module in sequence to obtain the trained hybrid model; S3: collecting a second pulsed eddy current signal and a second supplementary feature corresponding to the gold to be tested; S4: input the second pulsed eddy current signal and the second supplementary feature into the trained hybrid model to output a prediction vector; S5: convert the prediction vector into a one-hot code to obtain a corresponding number of the gold to be tested through the one-hot code, wherein the number is used to reflect the purity of the gold; Wherein, the first supplementary feature and the second supplementary feature both comprise the environmental temperature when collecting the pulsed eddy current signal, and the probe temperature, power supply temperature and heat sink temperature of the pulsed eddy current detection device.

2. The hybrid model based interlayer gold purity evaluation method according to claim 1, characterized in that, In the S2.4, the feature intervals corresponding to the decay stage are specifically: the rising edge segment, the fast decay segment and the slow decay segment.

3. The hybrid model based interlayer gold purity evaluation method of claim 1, wherein, In the S2.4, the weighted fusion adopts weighted splicing fusion, multiplies the feature intervals by corresponding weights respectively, and splices them in the original order to obtain the optimized time series feature data.

4. The hybrid model based interlayer gold purity evaluation method of claim 1, wherein, The initial model comprises an MLPM model, and the MLPM model has two data input channels, wherein the first channel is used to input the original data, and the second channel is used to input the processed original data, wherein the processed original data comprises standardized original data; The MLPM model is a multi-input channel multilayer perceptron model, which is a deformation model of the multilayer perceptron model. The MLPM model includes a two-channel input layer, a hidden layer for stacking feature extraction, and a classification layer for outputting a feature vector. One of the two data input channels of the two-channel input layer is used to input original data, and the other channel is used to input normalized data.

5. A sandwich gold purity evaluation device based on a mixture model, characterized by, The device comprises: A first acquisition module for acquiring original data, wherein the original data includes a first pulsed eddy current signal and first supplementary features; A hybrid model training module for training the processed original data in a hybrid model to obtain a trained hybrid model; The hybrid model comprises a feature extraction module, a weighted fusion unit, and a classification module connected in sequence. The specific steps for training the hybrid model include: S2.1: input the processed original data into the initial model corresponding to the feature extraction module for training to obtain a trained initial model, wherein the initial model includes at least an input layer, a hidden layer, and a classification layer; S2.2: remove the classification layer of the initial model to obtain the feature extraction module; S2.3: input the processed original data into the feature extraction module to obtain time series feature data; S2.4: input the time series feature data into the weighted fusion unit to obtain optimized time series feature data; The weighted fusion unit divides the time series feature data into feature intervals corresponding to the decay stage based on the decay physical properties of the pulsed eddy current signal, and performs weighted fusion on the corresponding feature intervals; S2.5: input the optimized time series feature data into the classification module for training to obtain a trained classification module; S2.6: connect the feature output end of the feature extraction module, the weighted fusion unit, and the feature input end of the trained classification module in sequence to obtain the trained hybrid model; A second acquisition module for acquiring test data of the test gold, wherein the test data includes a second pulsed eddy current signal corresponding to the test gold and second supplementary features; A gold purity prediction module for predicting the purity of the test gold and outputting a prediction vector; A gold purity result identification module for obtaining the purity result of the test gold according to the prediction vector; The first supplementary features and the second supplementary features both include the environmental temperature during the acquisition of the pulsed eddy current signal, and the probe temperature, power supply temperature, and heat sink temperature of the pulsed eddy current detection device.

6. An electronic device, characterized in that The device comprises: A memory for storing a computer program; A processor for executing the computer program to implement the hybrid model-based sandwich gold purity evaluation method of any one of claims 1-4.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is executed by a processor to implement the hybrid model-based sandwich gold purity evaluation method of any one of claims 1-4.

Images