Interlayer gold purity evaluation method and device based on hybrid model, equipment and medium

Abstract

The invention discloses an interlayer gold purity evaluation method and device based on a hybrid model, equipment and a medium. The method comprises the following steps: collecting original data; inputting the processed original data into a 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 which are connected in sequence; collecting a second pulsed eddy current signal and a second supplementary feature corresponding to the gold to be detected; inputting the second pulse eddy current signal and the second supplementary feature into a trained hybrid model, and outputting a prediction vector; and converting the prediction vector into a one-hot code, and obtaining the number of the corresponding gold to be tested through the one-hot code. The problem that an existing classification mixing model cannot be suitable for a highly-adulterated gold purity detection scene is solved, the classification precision of adulterated gold purity detection is improved, and the application range of the classification mixing model in the field of scarce sample detection is widened.

Description

Technical Field

[0001] This invention relates to the technical field of pulsed eddy current metal detection using deep learning methods, and particularly to a method, apparatus, device, and medium for evaluating the purity of sandwiched gold based on a hybrid model. Background Technology

[0002] Currently, gold purity testing mainly relies on traditional methods such as chemical analysis, mass spectrometry, and spectral analysis. While these methods offer high accuracy, they generally suffer from drawbacks such as complex operation, long testing cycles, reliance on specialized equipment and personnel, and are mostly destructive testing methods, making them unsuitable for large-scale or rapid on-site testing scenarios. Pulsed eddy current testing (PECT), as a non-destructive testing technique, has been applied to the purity testing of gold in laminated gold plates.

[0003] For processing PEC signals, existing technologies mostly employ traditional machine learning methods. These methods rely on manual feature design and have limited ability to represent complex, high-dimensional PEC signals, especially when the sample size is small, resulting in insufficient classification accuracy and generalization ability.

[0004] In recent years, deep learning (DL) methods have demonstrated powerful capabilities in time series classification; for example, models such as multilayer perceptrons (MLPs) can automatically learn high-dimensional features. However, in gold purity assessment, pure deep learning models are prone to overfitting in small sample situations and lack the ability to distinguish subtle differences in high-purity adulterated gold.

[0005] In existing technologies, hybrid algorithms combining feedforward neural networks and support vector machines have been used for classification tasks, such as hybrid models of MLP and SVM. By using an MLP model to extract features and an SVM model to complete the classification, the recognition performance in some scenarios has been improved. However, due to the scarcity and high cost of gold samples, directly applying existing MLP-SVM hybrid models to gold detection scenarios presents several challenges. First, the low dimensionality and insufficient information density of the features generated by the MLP model's output layer make it difficult to capture the key features for determining the purity of gold samples. Furthermore, the limited amount of gold sample data means the SVM model cannot learn effective training information from low-quality features, ultimately leading to overfitting and a significant drop in classification accuracy. Second, existing models typically input the features directly into the SVM model for training after the MLP model outputs features, employing an indiscriminate feature processing strategy. This fails to highlight the core features directly related to purity in the gold pulse eddy current signal, ultimately making it difficult to meet the high-precision requirements of gold purity detection. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, device, and medium for evaluating the purity of sandwiched gold based on a hybrid model, in order to solve the problems that existing classification hybrid models are unsuitable for high-adulteration gold purity detection scenarios due to low feature information density, poor small sample adaptability, and indiscriminate feature processing. This will improve the classification accuracy of adulterated gold purity detection and broaden the applicability of classification hybrid models in the field of scarce sample detection.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for evaluating the purity of sandwiched gold based on a hybrid model, the method comprising the following steps:

[0008] S1: Acquire raw data; wherein the raw data includes a first pulse eddy current signal and a first supplementary feature;

[0009] The first pulsed eddy current signal is the signal returned to the pulsed eddy current detection device after the pulsed eddy current detection device sends a pulse waveform to the test gold and the test gold induces eddy currents.

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

[0011] S3: Acquire the second pulse eddy current signal and the second supplementary feature corresponding to the gold to be tested;

[0012] S4: Input the second pulse eddy current signal and the second supplementary feature into the trained hybrid model and output the prediction vector;

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

[0014] Furthermore, the specific steps for training the hybrid model include:

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

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

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

[0018] S2.4: Input 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 the trained classification module;

[0020] 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.

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

[0022] Furthermore, in S2.4, the characteristic intervals corresponding to the decay stage are specifically: the rising edge segment, the rapid decay segment, and the slow decay segment.

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

[0024] Furthermore, both the first supplementary feature and the second supplementary feature include the ambient temperature when acquiring the pulsed eddy current signal, as well as the probe temperature, power supply temperature, and heat sink temperature of the pulsed eddy current detection device.

[0025] Furthermore, the initial model includes an MLPM model, which has two data input channels. The first channel is used to input the raw data, and the second channel is used to input the processed raw data, wherein the processed raw data includes standardized raw data.

[0026] Based on the same concept, the present invention also provides a hybrid model-based device for evaluating the purity of sandwiched gold, the device comprising:

[0027] The first acquisition module is used to acquire raw data, wherein the raw data includes a first pulse eddy current signal and a first supplementary feature;

[0028] The hybrid model training module is used to input the processed raw data into the hybrid model for training, and obtain the trained hybrid model.

[0029] The second acquisition module is used to acquire the test data of the gold to be tested, wherein the test data includes the second pulse eddy current signal and the second supplementary feature corresponding to the gold to be tested.

[0030] The gold purity prediction module is used to predict the purity of the gold to be tested and output a prediction vector.

[0031] The gold purity determination module is used to obtain the purity result of the gold to be tested based on the prediction vector.

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

[0033] The device includes:

[0034] Memory, used to store computer programs;

[0035] A processor is configured to implement the hybrid model-based method for evaluating the purity of sandwiched gold as described above when executing the computer program.

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

[0037] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the hybrid model-based method for evaluating the purity of sandwiched gold as described above.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention proposes a hybrid model, the MLPM-Weighted Fusion Unit-SVM model, which specifically includes: removing the classification layer of the MLPM model; performing deep feature extraction on the input raw data based on the MLPM model without the classification layer; and inputting the extracted high-dimensional features into the SVM model for training, thereby constructing a hybrid model training method suitable for gold purity detection. This design expands the diversity of feature data output by the MLPM model, solving 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. Furthermore, based on the attenuation physical characteristics of pulsed eddy current signals, the data after feature extraction is weighted and fused, strengthening the features of the fast attenuation segment that is strongly correlated with purity in gold detection. This facilitates the SVM model to prioritize learning information directly related to gold purity, significantly improving the identification accuracy of high-precision adulterated gold. Attached Figure Description

[0040] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the sandwich gold purity evaluation method based on a hybrid model in an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of training a hybrid model according to an embodiment of the present invention;

[0043] Figure 3 This is a PEC signal image obtained by the pulsed eddy current device in this embodiment of the invention, when the probe is in both low temperature and high temperature states;

[0044] Figure 4 This is a typical image of the gold PEC signal detected by the device in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of the architecture and training of the MLPM-weighted fusion unit-SVM model in an embodiment of the present invention. Figure 5 (a) is a schematic diagram of the overall training process of the MLPM-weighted fusion unit-SVM model. Figure 5 (b) is a schematic diagram of the training process of the SVM model in an embodiment of the present invention. Figure 5 (c) is a schematic diagram of the data analysis process of the MLPM-weighted fusion unit-SVM model in an embodiment of the present invention;

[0046] Figure 6 This is a confusion matrix diagram of the prediction results of the SVM model in the embodiments of the present invention;

[0047] Figure 7 This is a confusion matrix diagram of the prediction results of the MLPM model in the embodiments of the present invention;

[0048] Figure 8 This is a confusion matrix diagram of the prediction results of the MLPM-weighted fusion unit-SVM model in the embodiments of the present invention. Detailed Implementation

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The technical solutions of this 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 may not be described again in some embodiments.

[0051] Example 1

[0052] Figure 1 A flowchart of a method for evaluating the purity of sandwiched gold based on a hybrid model, provided in an embodiment of the present invention, is shown. Figure 1As shown, the method for evaluating the purity of sandwiched gold includes the following steps:

[0053] S1: Collect raw data.

[0054] Pulsed eddy current (PECT) testing has been proven effective in detecting gold adulteration. Compared with traditional detection methods, PECT offers advantages such as speed, low cost, and non-contact operation, thus showing broad application prospects in the detection and evaluation of gold adulteration.

[0055] A large number of PEC signals and supplementary features from various gold samples were collected using a pulsed eddy current testing (PDC) device. These signals served as input data for a hybrid model. The PEC signal is the signal returned to the PDC device after the gold sample generates eddy currents following the transmission of a pulse waveform. Supplementary features include ambient temperature, and the probe temperature, power supply temperature, and heatsink temperature of the PDC device. The ambient temperature is the temperature at which the test was conducted, acquired using a sensor. The probe temperature is the temperature of the PDC device's probe itself. The power supply temperature and heatsink temperature are the temperatures of the hybrid power supply module of the PDC device. These three temperatures are monitored and controlled within a specific range by the temperature control module of the PDC device. The probe temperature of the PDC device can affect the results. Figure 3 The correlation between the amplitude of the pulsed eddy current and the probe temperature of the same gold sample is shown. It can be seen that the amplitude of the pulsed eddy current and the probe temperature are correlated during the acquisition of the same gold sample. Therefore, the probe temperature is also used as data for model training, and several other temperatures are also included in the training set as supplementary features. The PEC signal is a 1000-dimensional time series, and the ambient temperature, probe temperature, power supply temperature, and heat sink temperature are 1-dimensional time series. The input features are 1004-dimensional feature vectors.

[0056] The purity and other properties of each gold sample are shown in Table 1. PEC data of 13 gold samples with different purities and thicknesses were collected using a pulsed eddy current (PDC) detector. Both the training and test sets included data from the 13 gold samples. 200 data sets were collected for each gold sample, and data was acquired four times through rotation and flipping operations, resulting in 800 data points for each sample. The number of sample data points for each gold sample was 400 in both the training and test sets. Therefore, the dimensions of both the training and test sets were (5200, 1004).

[0057] Table 1 Gold Purity Properties

[0058]

[0059] S2: Input the processed raw data into the hybrid model for training to obtain the trained hybrid model, which includes a feature extraction module, a weighted fusion unit and a classification module connected in sequence.

[0060] Processing the raw data includes standardizing it using Z-score normalization, ensuring the mean is zero and the standard deviation is 1 across all dimensions. The resulting training data is then used for network training, and the test set is used for model performance evaluation.

[0061] The standardization method uses Z-score normalization, which is calculated as follows:

[0062] ,

[0063] in, The original data collected, where i is the number of dimensions, j is the amount of data in each dimension, and x is the number of dimensions. ji It is the j-th sample data in the i-th dimension, μ i and σ i Each dimension X is a separate dimension. i = (x 1i ,x 2i ,x 3i, …,x ji The mean and standard deviation of z on the z-axis. ji This represents the normalized data of the j-th sample across i dimensions. Where X... i = (x i1 ,x i2 ,x i3 …x ij ), is a tensor that flattens out when the dimension is i, where the number of dimensions i is 1004, and j is 5200 in both the training and test sets. μ i and σ i They are X j = (x 1j ,x 2j ,x 3j, …,x ij The mean and standard deviation of z ij This represents the normalized data of the j-th sample data in the i-th dimension.

[0064] Where, μ i and σ i The results are obtained using the following formulas:

[0065] ,

[0066] ,

[0067] Where N=5200.

[0068] in, Figure 4 The image shows the time variation of the amplitude of the induced pulse eddy current generated in the tested sample. Dimension i is obtained by dividing the first 100 milliseconds of the pulse eddy current decay signal into 1000 nodes with an interval of 0.1 milliseconds. The amplitude signal data corresponding to the 1000 nodes are used as the first 1000 dimensions. The ambient temperature, probe temperature, power supply temperature, and heat sink temperature data are used as the 1001st, 1002nd, 1003rd, and 1004th dimensions, respectively. j represents the sample data of the gold sample.

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

[0070] Due to the preciousness and scarcity of gold samples, obtaining a large amount of authentic, labeled gold sample data is difficult. Because gold is expensive and its supply is limited, collecting large amounts of PECT data with purity labels is costly and impractical. Therefore, introducing deep learning technology to classify gold into multiple categories (e.g., pure gold and adulterated gold), and then building a classification model based on the degree of adulteration by collecting and analyzing a small amount of labeled data, is of great significance for identifying gold purity.

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

[0072] In this embodiment of the invention, the initial model is a variant of the Multi-Input Channel Multi-Channel Perceptron (MLP) model, specifically the MLPM model. The MLPM model is suitable for learning feature representations from time-series data. It extracts key features from the data by stacking multiple network layers that fuse local features, enabling it to accurately capture subtle changes and evolutionary trends across different time periods. This characteristic gives the MLPM model a significant advantage in identifying complex patterns and anomalous events in PEC time series, demonstrating higher recognition accuracy and robustness.

[0073] The MLPM model includes a dual-channel input layer, a hidden layer for stacked feature extraction, and a classification layer for outputting feature vectors. One of the two input channels in the dual-channel input layer is used to input the raw data, and the other is used to input the standardized data. This dual-channel design eliminates the influence of probe temperature fluctuations. Of course, the initial model in this embodiment is not limited to a specific model; any deep learning model with classification capabilities can be applied to this embodiment, such as existing MLP models and LSTM models. In the single-channel input model described above, the standardized data is used for testing within the input channel.

[0074] In embodiments of the present invention, before training the MLPM model, the method further includes setting various hyperparameters of the MLPM model and configuring the training environment of the MLPM model. 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, this includes setting the number of blocks to 36 and initializing the learning rate to 1 × 10⁻⁶. −3 The training iterations consisted of 200 epochs. All weights and biases of the model were updated using the Adam optimizer, with an L2 regularization parameter of 1 × 10⁻⁶. −3 The cross-entropy loss function is used. Model parameters are updated using the gradient function. The batch size is 120. When the loss value stabilizes during training, the model is considered converged. To prevent overfitting, early stopping is performed at this point to obtain the trained MLPM model.

[0075] The training process was coded using the PyTorch machine learning library and performed on an Nvidia GeForce RTX 4070Ti GPU. The MLPM-36 training iteration (200 iterations) took 11 minutes and 3.6 seconds. As the number of blocks increases, the model's computation becomes more complex, and the training time also increases.

[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. MLPM-P is used to reduce the dimensionality of the data and amplify the differences between different feature data. MLPM-P is responsible for extracting discriminative features from high-dimensional PEC signals to achieve feature extraction.

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

[0079] The dataset used to train the MLPM model is then input back into MLPM-P. Using the weights of the MLPM model without backpropagation parameter updates, feature data is obtained through forward computation, resulting in the time-series feature data output by MLPM-P, i.e., the P data. The P data has an appropriate dimensionality, allowing the SVM model to better extract useful feature information. The reason for not using the data after the MLPM model's classification layer to train the SVM model is that the data type output by the classification layer is a feature vector corresponding to the one-hot code. Its data dimension is too low, almost directly representing the sample label, and it can represent too few features, so the subsequent classification model cannot obtain effective training information.

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

[0081] according to Figure 4 The graph showing the amplitude of the induced pulsed eddy current versus time reveals that the pulsed eddy current signal curve can be roughly divided into three stages: a rising edge segment of 0-5ms, a rapid decay segment of 5-50ms, and a slow decay segment of 50-100ms. In the rising edge segment, eddy currents are rapidly generated in the gold being tested after the excitation signal is applied. In the rapid decay segment, the eddy currents diffuse rapidly in the gold, and their energy dissipates quickly. During this stage, the signal amplitude decreases at the fastest rate and is strongly correlated with the conductivity of gold (which is directly related to gold purity). In the slow decay segment, most of the eddy current energy has dissipated, leaving only a small amount of eddy current that decays slowly. During this stage, the signal decay rate slows significantly.

[0082] This shows that among the 1000 input nodes, the 500 nodes in the rapid decay phase are core features, while the 500 nodes in the rising and slow decay phases are redundant data. Directly inputting the P data into the subsequent classification model for equivalent processing would dilute the core features with redundant data, reducing the model's classification accuracy in high-purity, small-sample gold detection scenarios. Therefore, this embodiment of the invention, based on the physical laws governing the time-varying pulsed eddy current signal, divides the P data into feature intervals corresponding to the decay phases and performs weighted fusion on the corresponding feature intervals.

[0083] The specific steps of weighted fusion include assigning different weights to the P data corresponding to different stages based on the physical prior knowledge of gold detection. For the fast decay segment, which is the core feature segment, the weight is set to 0.6, and for the rising edge segment and the slow decay segment, the weight is set to 0.2.

[0084] Multiply the P data corresponding to the feature intervals by their respective weights to obtain physically meaningful feature data. Concatenate the weighted P data in their original order to obtain optimized P data.

[0085] By multiplying each feature segment by a weight and then using a weighted concatenation and fusion method, the contribution of the core interval features is enhanced while the original feature dimensions and order of each interval are fully preserved, which facilitates 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 is well-suited for small sample sizes and non-linear data. SVM models can perform efficient classification with limited samples and can handle non-linear patterns through kernel function transformations. Therefore, SVM models excel in handling sparse labels and complex classification problems in PEC time series.

[0088] In high-dimensional data, neural networks generally perform better classification than SVM models. This is because: (1) Neural networks can learn non-linear relationships through multiple hidden layers and activation functions, thus handling more complex data distributions. In high-dimensional data, non-linear relationships are likely to exist, while SVM models can only introduce non-linearity through kernel functions, which is not applicable in some cases. (2) Neural networks can automatically learn feature representations through backpropagation, meaning they can extract the most discriminative features from the original data. In contrast, SVM models require manual feature selection and design, and feature selection can be more difficult for high-dimensional data. (3) Neural networks have advantages in handling large-scale datasets. By using deep learning frameworks and parallel computing techniques, neural networks can effectively handle large amounts of high-dimensional data, and classification performance can be improved by increasing the number of network layers and nodes.

[0089] It is evident that neural networks are suitable for processing high-dimensional data, but SVM models have an advantage when the data is linearly separable or has low feature dimensions. Furthermore, SVM models can provide interpretability for classification decisions. On low-dimensional data, SVM models outperform neural networks in classification: (1) For low-dimensional data, when the data is linearly separable, SVM models can quickly determine the optimal hyperplane to separate different categories of data. Neural networks, on the other hand, may learn unnecessarily complex models. (2) SVM models are highly robust to outliers. In low-dimensional data, scattered points or outliers are easily generated. Neural networks are susceptible to these disturbances, while SVM models can adjust the model's tolerance by introducing a penalty parameter C, thereby reducing the impact of these disturbances. Therefore, SVM models are more robust on low-dimensional data. (3) On low-dimensional data, neural networks may require more data for training and more computational resources and time. In contrast, SVM models train faster and perform well on datasets with fewer features.

[0090] In summary, the MLPM model, as a neural network, is well-suited for handling high-dimensional data, while the SVM model is better suited for handling low-dimensional data. Combining the MLPM and SVM models allows for full utilization of their respective advantages. The MLPM model is used to extract feature representations of PEC time series data, capturing high-dimensional information from the data. These features are then fed into the SVM model for classification, leveraging its non-linear classification capabilities and its advantage in classifying low-dimensional data.

[0091] Combining the MLPM and SVM models can complement each other's weaknesses. The MLPM model excels at extracting complex features from time-series data, while the SVM model is robust to classification problems and small, non-linear datasets. By leveraging the strengths of both, more accurate and reliable PEC time-series classification results can be obtained. This is particularly suitable for detecting high-purity gold (up to 99.9%), effectively distinguishing subtle differences between high-purity gold and trace adulterated gold. Therefore, combining the MLPM and SVM models, with the SVM model replacing the classification layer of the MLPM model, is recommended.

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

[0093] 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.

[0094] Figure 5 The diagram illustrates the architecture and training schematic of the MLPM-Weighted Fusion Unit-SVM model in an embodiment of the present invention. Wherein, as... Figure 5 As shown in (a), the overall training process of the MLPM-Weighted Fusion Unit-SVM model is as follows: First, the trained MLPM model is used; then, the classification layer of the MLPM model is removed to obtain MLPM-P; finally, the classification layer of the MLPM model is replaced with an SVM model to obtain the final MLPM-Weighted Fusion Unit-SVM model. Figure 5 (b) illustrates the training process of the SVM model, which includes using the trained MLPM model, then inputting all training set data into MLPM-P again for calculation, extracting all P data, 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, which mainly includes: the data is first processed by MLPM-P, then the data is weighted and fused for optimization, and the optimized input is input into the SVM model for analysis to obtain the final predicted sample label.

[0095] After obtaining the hybrid model, test data, which had also undergone Z-score normalization, was input into the hybrid model for calculation to obtain the predicted gold label. Compared with the results of using the SVM model or MLPM model alone, the training set size used by the SVM and MLPM models in the control experiment was twice that of the hybrid model using this embodiment of the invention, i.e., the size dimension of the training set for these two models was (10400, 1004). Apart from this, the configuration of other parameters of the model remained consistent with the hybrid model.

[0096] The test results are shown in Table 2. The confusion matrices of the prediction classification results of the three models are as follows: Figure 6 , Figure 7 , Figure 8 As shown, the MLPM-weighted fusion unit-SVM model achieved the highest accuracy of 95.3%, representing an improvement of approximately 3% compared to using the MLPM model alone.

[0097] Table 2 Classification Results

[0098]

[0099] By comparing the confusion matrices of three models—SVM, MLPM, and MLPM-Weighted Fusion Unit-SVM—the main differences between the three models can be identified. From the attribute table of the 13 gold samples, we know that the purity of gold sample 12 is 99.90%, and the purity of gold sample 13 is 99.99%. The MLPM-Weighted Fusion Unit-SVM model only misclassified gold sample 13, with a total of only 6 misclassifications. Therefore, compared to the MLPM model, the advantage of the MLPM-Weighted Fusion Unit-SVM model lies mainly in its more accurate classification of gold samples 12 and 13, whose purities are extremely similar. Meanwhile, from gold sample No. 8 to gold sample No. 13, which have a purity of over 96%, the SVM model only misclassified 16 samples of gold sample No. 13. This indicates that the SVM model performs better than the MLPM model in distinguishing high-purity adulterated gold. However, for differentiating low-purity adulterated gold, the MLPM model is significantly better than the SVM model. Therefore, the MLPM-weighted fusion unit-SVM model integrates the advantages of both the MLPM and SVM models, allowing them to complement each other and thus achieving better prediction and classification results. Meanwhile, the training set data size of the MLPM-Weighted Fusion Unit-SVM model is only half that of the MLPM model and the SVM model. The MLPM-Weighted Fusion Unit-SVM model can still distinguish between No. 12 and No. 13 more accurately. Compared with the MLPM model and the SVM model, the MLPM-Weighted Fusion Unit-SVM model can achieve a higher prediction accuracy than the latter two models with only a small sample training data. Moreover, it is more detailed in the classification of high-purity adulterated gold. While reducing the number of samples, it improves the detection accuracy of the hybrid classification model and reduces the sample requirements in detection scenarios such as gold detection where samples are difficult to obtain.

[0100] The hybrid architecture of "MLPM-P + weighted fusion unit + SVM" achieves an organic combination of high-dimensional feature extraction and low-dimensional fine classification. Under small sample training conditions, feature dimensionality reduction and model fusion significantly improve classification accuracy and generalization ability. The above experiments show that the MLPM-weighted fusion unit-SVM model can still achieve higher accuracy, especially in the fine classification of high-purity gold, even when the training set is only half that of the control model.

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

[0102] A pulsed eddy current testing device is used to collect pulsed eddy current signals and supplementary characteristics corresponding to the gold under test. In the specific testing process, the pulsed eddy current signal is collected by placing the gold under test above the sensor of the pulsed eddy current testing device. The device then generates a rectangular wave, which is amplified by a power amplifier module and applied to the excitation coil in the sensor. The excitation coil thus generates a primary magnetic field whose intensity varies with time, forming a rectangle. When the primary magnetic field undergoes a sudden change, eddy currents are induced in the gold plate. The induction and dissipation of eddy currents are affected by the purity of the gold plate. During the eddy current generation and dissipation process, the eddy currents generate a secondary magnetic field that varies with time. The receiving coil in the sensor converts the time-varying secondary magnetic field into a voltage signal, which is amplified by a signal amplification module. Data is acquired by a dual-function module for signal generation and acquisition. The probe temperature of the pulsed eddy current testing device is monitored by software. The collected data is preprocessed to obtain time-series data.

[0103] S4: Input the pulsed eddy current signal and supplementary features into the trained hybrid model and output the prediction vector.

[0104] S5: Convert the prediction vector of the gold to be tested into a one-hot code using the argmax function. Obtain the corresponding number of the gold to be tested using the one-hot code. Look up the corresponding purity value in Table 1 based on the number to assess the purity of the gold.

[0105] Example 2

[0106] Based on the same concept, embodiments of the present invention also provide a hybrid model-based device for evaluating the purity of sandwiched gold, the device comprising:

[0107] The first acquisition module is used to acquire raw data, wherein the raw data includes a first pulse eddy current signal and a first supplementary feature;

[0108] The hybrid model training module is used to input the processed raw data into the hybrid model for training, and obtain the trained hybrid model.

[0109] The second acquisition module is used to acquire the test data of the gold to be tested, wherein the test data includes the second pulse eddy current signal and the second supplementary feature corresponding to the gold to be tested.

[0110] The gold purity prediction module is used to predict the purity of the gold to be tested and output a prediction vector.

[0111] The gold purity determination module is used to obtain the purity result of the gold to be tested based on the prediction vector.

[0112] Example 3

[0113] Based on the same concept, embodiments of the present invention also provide an electronic device, the device comprising:

[0114] Memory, used to store computer programs;

[0115] A processor is used to implement the hybrid model-based method for evaluating the purity of sandwiched gold described in Embodiment 1 above when executing the computer program.

[0116] Example 4

[0117] Based on the same concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the sandwich gold purity evaluation method based on a hybrid model described in Embodiment 1 above.

[0118] This invention provides a hybrid model-based method for evaluating the purity of sandwiched gold. It involves collecting PEC data from gold blocks using a pulsed eddy current (PEC) detection device and utilizing deep learning technology to assess the degree of gold adulteration. First, the training dataset is determined by analyzing the relationship between the PEC signal during gold collection, the pulsed eddy current detection device data, and the ambient temperature data. 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 adulterated gold, yielding the final classification results. This invention successfully applies a hybrid model to the field of gold adulteration assessment. Experimental results show that compared to using the MLPM model or SVM model alone, the hybrid model MLPM-Weighted Fusion Unit-SVM exhibits better classification performance in gold adulteration assessment. Even with training on small sample data, it can still accurately classify gold samples with different degrees of adulteration and demonstrates sufficient generalization ability. Furthermore, by binding feature data to the physical attenuation stage of the gold pulsed eddy current signal, physically meaningful P-data is designed, solving the problem of local correlation loss when traditional hybrid models process high-dimensional time-series data. Therefore, the gold adulteration assessment method provided by this invention has broad application prospects.

[0119] The above embodiments should be understood as being used only to illustrate the present invention more clearly, and not to limit the scope of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. A method for evaluating the purity of sandwiched gold based on a hybrid model, characterized in that, The method includes the following steps: S1: Acquire raw data; wherein the raw data includes a first pulse eddy current signal and a first supplementary feature; The first pulsed eddy current signal is the signal returned to the pulsed eddy current detection device after the pulsed eddy current detection device sends a pulse waveform to the test gold and the test gold induces eddy currents. S2: Input the processed raw data into the hybrid model for training to obtain the trained hybrid model; wherein, the hybrid model includes a feature extraction module, a weighted fusion unit and a classification module connected in sequence; S3: Acquire the second pulse eddy current signal and the second supplementary feature corresponding to the gold to be tested; S4: Input the second pulse eddy current signal and the second supplementary feature into the trained hybrid model and output the prediction vector; S5: Convert the prediction vector into a one-hot code, and obtain the corresponding number of the gold to be tested through the one-hot code. The number is used to reflect the purity of the gold.

2. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 1, characterized in that, The specific steps for training a hybrid model include: S2.1: Input the processed raw data into the initial model corresponding to the feature extraction module for training to obtain the 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 raw 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; 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.

3. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 2, characterized in that, In the weighted fusion unit, based on the attenuation physical characteristics of the pulsed eddy current signal, the time series feature data is divided into feature intervals corresponding to the attenuation stage, and the corresponding feature intervals are weighted and fused.

4. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 3, characterized in that, In S2.4, the characteristic intervals corresponding to the decay stage are specifically: rising edge segment, fast decay segment, and slow decay segment.

5. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 4, characterized in that, In step S2.4, the weighted fusion adopts a weighted splicing fusion, which multiplies the feature intervals by their corresponding weights and splices them in the original order to obtain optimized time series feature data.

6. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 1, characterized in that, Both the first supplementary feature and the second supplementary feature include the ambient temperature when acquiring the pulsed eddy current signal, as well as the probe temperature, power supply temperature, and heat sink temperature of the pulsed eddy current detection device.

7. The method for evaluating the purity of sandwiched gold based on a hybrid model according to claim 2, characterized in that, The initial model includes an MLPM model, which has two data input channels. The first channel is used to input the raw data, and the second channel is used to input the processed raw data, which includes standardized raw data.

8. A device for evaluating the purity of sandwiched gold based on a hybrid model, characterized in that, The device includes: The first acquisition module is used to acquire raw data, wherein the raw data includes a first pulse eddy current signal and a first supplementary feature; The hybrid model training module is used to input the processed raw data into the hybrid model for training, and obtain the trained hybrid model. The second acquisition module is used to acquire the test data of the gold to be tested, wherein the test data includes the second pulse eddy current signal and the second supplementary feature corresponding to the gold to be tested. The gold purity prediction module is used to predict the purity of the gold to be tested and output a prediction vector. The gold purity determination module is used to obtain the purity result of the gold to be tested based on the prediction vector.

9. An electronic device, characterized in that, The device includes: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the sandwich gold purity assessment method based on a hybrid model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for evaluating the purity of sandwiched gold based on a hybrid model as described in any one of claims 1 to 7.

Images