Artificial intelligence (ai) rare earth detection method and system based on millimeter wave radar multi-metal reflection pattern

Abstract

This invention provides an AI-based rare earth detection method and system based on millimeter-wave radar multi-metal reflectance spectra, belonging to the field of data processing technology. The method includes: acquiring the raw millimeter-wave radar reflectance signal data of the sample to be tested; performing signal decomposition processing to separate the reflectance signal into multiple metal signal components based on metal reflectance characteristic parameters; calculating feature ratios, constructing feature ratio sequences between rare earth metal components and non-rare earth metal components based on rare earth metal components; performing local extremum extraction and segmented weighted superposition processing, assigning high weights to intervals with rare earth metal signal response patterns and low weights to intervals with non-rare earth metal signal response patterns in the feature ratios; extracting the signal response indicators of rare earth metals and performing quantitative analysis to generate rare earth metal content detection results. This invention improves the autonomy and accuracy of rare earth detection.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an AI rare earth detection method and system based on millimeter-wave radar multimetal reflectance spectra. Background Technology

[0002] Currently, rare earth element detection commonly employs methods based on traditional X-ray fluorescence analysis or inductively coupled plasma mass spectrometry. These methods typically involve sample pretreatment followed by quantitative analysis of the signal intensities generated by different elements using instruments. In recent years, some detection systems have incorporated millimeter-wave radar technology. By collecting the reflection characteristic curves of different metals to millimeter-wave signals and combining this with data processing algorithms, the types and amounts of metals contained in the sample can be identified, thereby achieving preliminary detection of rare earth elements.

[0003] However, in actual mine applications, existing millimeter-wave radar detection systems are susceptible to signal overlap interference in multi-metal mixed samples. For example, when a sample contains metals such as iron and nickel, the high reflectivity of these metals can mask the weak reflection signals of rare earth metals, leading to significant deviations in rare earth content detection results. Especially in scenarios involving continuous sampling and real-time monitoring, the system struggles to accurately separate and quantify the reflection information of rare earth elements, affecting detection reliability and ore grading decisions. Summary of the Invention

[0004] The purpose of this invention is to provide an AI rare earth detection method and system based on millimeter-wave radar multimetal reflectance spectra, aiming to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, an AI-based rare earth detection method based on millimeter-wave radar multimetal reflectance spectra, the method comprising:

[0007] Acquire raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to millimeter-wave signals;

[0008] The original reflected signal data is processed by signal decomposition, and the reflected signal is separated into multiple metal signal components based on the metal reflection characteristic parameters to obtain the dataset of each metal component.

[0009] The feature ratios of each metal component dataset are calculated. Based on the rare earth metal components, a feature ratio sequence between the rare earth metal components and the non-rare earth metal components is constructed to obtain the feature ratio data.

[0010] Based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. The intervals with rare earth metal signal response patterns in the feature ratios are given high weights, and the intervals with non-rare earth metal signal response patterns are given low weights, thus obtaining rare earth metal feature enhancement data.

[0011] Based on the characteristic enhancement data of rare earth metals, the signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate the rare earth metal content detection results.

[0012] Preferably, the original reflected signal data is subjected to signal decomposition processing, separating the reflected signal into multiple metal signal components based on metal reflection characteristic parameters, to obtain a dataset of each metal component, including:

[0013] Based on the pre-stored parameters of various metal reflection characteristics, the original reflected signal data is subjected to spectrum analysis, and the original signal is divided into multiple frequency bands according to the frequency response characteristics.

[0014] Based on the signal of each frequency band, the signal segments that match the characteristic parameters of different metals are extracted to obtain the preliminary response signal set of each metal;

[0015] The initial response signal set of each metal is denoised to filter out isolated high-frequency noise points and adaptively correct the signal baseline drift to obtain the denoised response signal set.

[0016] Based on the set of denoised response signals, the amplitude of each metal signal segment is normalized, and all segments are standardized to a uniform amplitude range to obtain the set of normalized response signals.

[0017] Each metal signal segment in the normalized response signal set is archived according to metal category to generate a dataset of each metal component.

[0018] Preferably, feature ratios are calculated for each metal component dataset. Using rare earth metal components as a benchmark, feature ratio sequences between rare earth metal components and non-rare earth metal components are constructed to obtain feature ratio data, including:

[0019] Iterate through all data points in each metal component dataset, using the signal intensity of rare earth metal components as reference data, and using the signal intensity of all non-rare earth metal components at the corresponding locations as comparison data.

[0020] Based on each set of reference data and comparison data, a ratio calculation is performed, dividing the reference data by the comparison data to obtain multiple ratio subsequences, each ratio subsequence corresponding to a non-rare earth metal component;

[0021] Apply a moving average to each ratio subsequence to remove abrupt outliers and smooth continuous data, resulting in a smoothed ratio subsequence.

[0022] The ratios of each ratio subsequence at each data point are combined sequentially to construct a feature vector. The feature vectors of all data points are then arranged in sequence according to the signal time sequence to generate a feature ratio dataset with a time sequence structure.

[0023] Preferably, based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. Intervals with rare earth metal signal response patterns are assigned high weights, while intervals with non-rare earth metal signal response patterns are assigned low weights, resulting in rare earth metal feature enhancement data, including:

[0024] The feature ratio data is segmented according to a preset interval length. Extreme value search is performed on the interval of each segment, and local maximum and minimum values ​​are extracted in sequence to form the feature sequence of the interval.

[0025] Calculate the similarity between the feature sequence of each interval and the preset rare earth metal signal template. If the similarity is greater than the preset similarity threshold, the interval is marked as a high-weight interval; otherwise, it is marked as a low-weight interval.

[0026] For each high-weight interval, multiply all ratio data within the interval by a preset high-weight factor; for low-weight intervals, multiply the ratio data within the interval by a preset low-weight factor, and satisfy the condition that the high-weight factor is greater than the low-weight factor.

[0027] All weighted segmented data are summed and superimposed to form a rare earth metal feature enhancement data sequence.

[0028] Preferably, based on the rare earth metal characteristic enhancement data, signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate rare earth metal content detection results, including:

[0029] Based on the characteristic enhancement data of rare earth metals, a main peak search is performed on the sequence to identify the main peak position and its signal intensity in the sequence, and the main peak signal response value is obtained.

[0030] Within a preset continuous interval on both sides of the main peak signal response value, the signal is integrated to obtain the area of ​​the interval reflecting the characteristic energy of rare earth.

[0031] Substitute the main peak signal response value and the interval area into the preset rare earth metal standard curve relationship to calculate the actual content of rare earth metals in the sample.

[0032] The actual content value is used as the rare earth metal content detection result, and this result is associated with the unique number of the sample to be tested.

[0033] Preferably, based on pre-stored various metal reflection characteristic parameters, spectral analysis is performed on the original reflected signal data, dividing the original signal into multiple frequency bands according to its frequency response characteristics, including:

[0034] Perform a time window truncation operation on the original reflected signal data to obtain signal segments within multiple time periods;

[0035] For each signal segment, the fast Fourier transform method is used to convert the time-domain signal into a frequency-domain signal, and the amplitude-frequency response curve of the signal segment is obtained.

[0036] Based on the typical frequency response range of various metal reflection characteristic parameters, the frequency domain signal is divided into different frequency intervals, and each frequency interval is generated as an independent frequency band.

[0037] The data from all frequency ranges are combined sequentially to form multiple frequency bands that cover the entire analysis range.

[0038] Preferably, based on the signal of each frequency band, signal segments matching the characteristic parameters of different metals are extracted to obtain a preliminary set of response signals for each metal, including:

[0039] Based on the signal in each frequency band, extract the amplitude sequence and phase sequence within that frequency band as features to be compared.

[0040] For each type of metal, the peak amplitude parameters and phase characteristic parameters at a specific preset frequency are called up and compared one by one with the characteristics to be compared in each frequency band to form a difference sequence for each frequency band.

[0041] Based on the difference sequence, a similarity calculation method is set, and the amplitude difference and phase difference of the frequency band are compared with the corresponding preset difference thresholds. If all differences are lower than their respective difference thresholds, the signal segment of the frequency band is determined to match the metal feature parameter, and a matching mark is generated.

[0042] All signal segments that are determined to be matched are archived into their corresponding metal categories to generate a preliminary set of response signals for each metal.

[0043] Preferably, the initial response signal set of each metal is denoised to filter isolated high-frequency noise points, and the signal baseline drift is adaptively corrected to obtain a denoised response signal set, including:

[0044] For each set of preliminary response signals, the signal curves are smoothed using a moving average method to generate preliminary smoothed signals.

[0045] For each data point of the preliminary smoothed signal, the difference between its amplitude and the amplitude of the adjacent data points is determined. If the amplitude change of the data point exceeds the preset amplitude change threshold, it is regarded as an isolated high-frequency noise point, and the average of the adjacent data points is used to replace the data point to generate a denoised signal.

[0046] Based on the denoised signal, its baseline trend curve is fitted, and the baseline trend curve is subtracted from the signal to form a set of denoised response signals.

[0047] Preferably, the similarity between the feature sequence of each interval and the preset rare earth metal signal template is calculated, including:

[0048] For the feature sequence of each interval, the sliding window method is used to segment it, dividing the feature sequence within the interval into several overlapping preliminary subsequences;

[0049] For each initial subsequence, perform a normalization operation to make the mean of the initial subsequence zero and the standard deviation one, thus obtaining a normalized subsequence;

[0050] Each normalized subsequence is compared point-to-point with a rare earth metal signal template fragment of the corresponding length. The local similarity score between the normalized subsequence and the template fragment is obtained by calculating the sum of the products of their respective components.

[0051] The local similarity scores of all normalized subsequences within the same interval are weighted and averaged to obtain the overall similarity score of that interval.

[0052] Secondly, an AI rare earth detection system based on millimeter-wave radar multimetal reflectance spectra, the system comprising:

[0053] The data acquisition module is used to acquire the raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to the millimeter-wave signal.

[0054] The signal decomposition module is used to perform signal decomposition processing on the original reflected signal data, separating the reflected signal into multiple metal signal components based on the metal reflection characteristic parameters, and obtaining the dataset of each metal component.

[0055] The feature ratio calculation module is used to calculate the feature ratio of each metal component dataset. Taking the rare earth metal component as the benchmark, it constructs the feature ratio sequence between the rare earth metal component and the non-rare earth metal component to obtain the feature ratio data.

[0056] The weighted enhancement module is used to perform local extremum extraction and segmented weighted superposition processing based on the feature ratio data. It assigns high weight to the intervals with rare earth metal signal response patterns in the feature ratios and low weight to the intervals with non-rare earth metal signal response patterns, thereby obtaining rare earth metal feature enhancement data.

[0057] The quantitative analysis module is used to extract the signal response indicators of rare earth metals based on the characteristic enhancement data of rare earth metals, perform quantitative analysis, and generate the rare earth metal content detection results.

[0058] The above-described solution of the present invention has at least the following beneficial effects:

[0059] This invention acquires the raw millimeter-wave radar reflection signal data of the sample under test, and decomposes and extracts features from the signal, effectively overcoming the problems of signal overlap and rare-earth signal masking caused by multi-metal contamination in existing technologies. By introducing signal decomposition processing based on metal reflection characteristic parameters, the response components of various metals in the raw reflection signal can be independently separated, significantly improving the distinguishability between rare-earth metal signals and high-reflectivity metal signals, and avoiding interference from main components such as iron and nickel on weak rare-earth signals.

[0060] Furthermore, by employing characteristic ratio calculation based on rare earth metal components and multi-step weighted enhancement processing, not only can the subtle changes in rare earth metal signals within the overall data be effectively amplified, but mechanisms such as extreme value extraction and segmented weighting can also accurately identify signal intervals with rare earth characteristics, significantly reducing the probability of false detections caused by impurity metals and noise. In continuous sampling and real-time on-site monitoring scenarios, this invention can stably separate and quantitatively analyze rare earth signals, achieving highly sensitive and interference-resistant detection of rare earth metal content in ores, improving detection accuracy and decision reliability, and contributing to the efficient development of rare earth resources and intelligent mine management. Attached Figure Description

[0061] Figure 1 This is a flowchart of an AI rare earth detection method based on millimeter-wave radar multimetal reflectance spectra provided by an embodiment of the present invention. Detailed Implementation

[0062] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0063] like Figure 1 As shown, embodiments of the present invention propose an AI rare earth detection method based on millimeter-wave radar multimetal reflectance spectra, the method comprising:

[0064] Acquire raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to millimeter-wave signals;

[0065] The original reflected signal data is processed by signal decomposition, and the reflected signal is separated into multiple metal signal components based on the metal reflection characteristic parameters to obtain the dataset of each metal component.

[0066] The feature ratios of each metal component dataset are calculated. Based on the rare earth metal components, a feature ratio sequence between the rare earth metal components and the non-rare earth metal components is constructed to obtain the feature ratio data.

[0067] Based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. The intervals with rare earth metal signal response patterns in the feature ratios are given high weights, and the intervals with non-rare earth metal signal response patterns are given low weights, thus obtaining rare earth metal feature enhancement data.

[0068] Based on the characteristic enhancement data of rare earth metals, the signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate the rare earth metal content detection results.

[0069] In this embodiment of the invention, high-precision and intelligent detection of rare earth metal components in the sample can be achieved by acquiring and processing the raw millimeter-wave radar reflection signal data of the sample in multiple stages. The method first acquires raw signals containing reflection response curves of various metals, which can completely cover the weak and strong signals of all metal components within the sample, ensuring the comprehensiveness of the detection basis. Subsequently, signal decomposition processing is used to effectively distinguish different metal signals, allowing the subtle response signals of rare earth metals to be accurately separated from the complex background and no longer masked by high-intensity main metal signals. The characteristic ratio calculation step further utilizes the relative changes between rare earth signals and other metal signals to enhance the recognizability of rare earth components in the overall data. After extreme value extraction and segmented weighting processing of the comparison value data, the system can perform interval recognition and enhancement of rare earth signal patterns, thereby suppressing noise and interference from irrelevant metals. Finally, the extracted rare earth signal response index can be accurately converted into rare earth metal content through quantitative analysis, achieving high sensitivity, anti-interference, and high accuracy detection of rare earth elements in complex mineral samples. For example, in mixed samples where the content of metals such as iron and nickel is much higher than that of rare earth elements, it can accurately identify rare earth signals and quantify their concentration, significantly improving the reliability and efficiency of on-site rare earth ore testing and reducing the need for laboratory analysis and the probability of misjudgment.

[0070] This includes acquiring the raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to the millimeter-wave signal, specifically including:

[0071] First, the mineral sample to be tested is placed within the effective detection area of ​​the millimeter-wave radar detection device, and the radar transmitting module is activated according to the preset parameter configuration, continuously transmitting millimeter-wave signals within a specific frequency range towards the sample. Simultaneously, the receiving module acquires the echo signals reflected from the sample surface and interior. The system performs preliminary preprocessing on the received raw echo signals, including analog signal amplification, filtering, and analog-to-digital conversion, generating digitized raw reflection signal data. Subsequently, the digitized signal is stored as a processable data stream, where each sampling point contains the instantaneous reflection response of each metal element to the millimeter-wave signal at a specific moment. This data stream comprehensively records the reflection characteristics of various metals within the sample to millimeter waves of different frequencies, providing fundamental data resources for subsequent decomposition, identification, and analysis of different metal components.

[0072] In a preferred embodiment of the present invention, the original reflected signal data is subjected to signal decomposition processing to separate the reflected signal into multiple metal signal components based on metal reflection characteristic parameters, thereby obtaining a dataset of each metal component, including:

[0073] Based on the pre-stored parameters of various metal reflection characteristics, the original reflected signal data is subjected to spectrum analysis, and the original signal is divided into multiple frequency bands according to the frequency response characteristics.

[0074] Based on the signal of each frequency band, the signal segments that match the characteristic parameters of different metals are extracted to obtain the preliminary response signal set of each metal;

[0075] The initial response signal set of each metal is denoised to filter out isolated high-frequency noise points and adaptively correct the signal baseline drift to obtain the denoised response signal set.

[0076] Based on the set of denoised response signals, the amplitude of each metal signal segment is normalized, and all segments are standardized to a uniform amplitude range to obtain the set of normalized response signals.

[0077] Each metal signal segment in the normalized response signal set is archived according to metal category to generate a dataset of each metal component.

[0078] In this embodiment of the invention, by performing signal decomposition based on metal reflection characteristic parameters on the original reflection signal data, signals of different metal categories can be extracted and processed in a targeted manner, greatly enhancing the detection capability of trace components such as rare earth metals. In the spectrum analysis stage, the original time-domain signal is first decomposed into its respective frequency response intervals, allowing for accurate division of the response frequency bands of different metals. For example, high-intensity reflection signals of metals such as iron and copper can be separated separately in their typical frequency bands, while rare earth metal signals, even with lower amplitudes, can be captured separately through parameter comparison. In the feature segment extraction stage, signals in each frequency band are compared with known metal parameters, retaining only the portions truly belonging to the specific metal response, significantly reducing the interference of impurity signals. Through denoising and baseline adaptive correction, high-frequency noise and signal drift are effectively eliminated, improving signal quality. Amplitude normalization processing eliminates systematic differences in signal amplitude between different samples or under different testing environments, making the final datasets of each metal component more uniform and comparable. Taking on-site testing as an example, after the above decomposition and normalization process, even in complex environments with large variations in impurity composition, the stability and traceability of each metal signal component can be guaranteed, which facilitates subsequent feature analysis and quantitative calculation.

[0079] In a preferred embodiment of the present invention, feature ratios are calculated for each metal component dataset. Using rare earth metal components as a benchmark, feature ratio sequences between rare earth metal components and non-rare earth metal components are constructed to obtain feature ratio data, including:

[0080] Iterate through all data points in each metal component dataset, using the signal intensity of rare earth metal components as reference data, and using the signal intensity of all non-rare earth metal components at the corresponding locations as comparison data.

[0081] Based on each set of reference data and comparison data, a ratio calculation is performed, dividing the reference data by the comparison data to obtain multiple ratio subsequences, each ratio subsequence corresponding to a non-rare earth metal component;

[0082] Apply a moving average to each ratio subsequence to remove abrupt outliers and smooth continuous data, resulting in a smoothed ratio subsequence.

[0083] The ratios of each ratio subsequence at each data point are combined sequentially to construct a feature vector. The feature vectors of all data points are then arranged in sequence according to the signal time sequence to generate a feature ratio dataset with a time sequence structure.

[0084] In this embodiment of the invention, by calculating the feature ratios of each metal component dataset, the differences between rare earth metal signals and other metal components can be highlighted in the signal environment of mixed metal samples. Specifically, each data point is traversed, and the signal intensity of the rare earth metal component is compared with the signals of other metals at the same time. For abnormal or abrupt values, a moving average is used for smoothing, making the result sequence more continuous and realistic, avoiding interference from single-point extreme values. The ratio results of all data points are combined in chronological order into a multi-dimensional feature vector sequence. This structure not only highlights the relative intensity of rare earth metals with other metals at each time point but also reflects the trend of signal changes with sample acquisition time. For example, during the continuous slice detection of a section of ore, if the rare earth metal content fluctuates or shows local enhancement, the aforementioned ratio features will form obvious peaks in the sequence, facilitating automatic detection system identification and alarm. This feature ratio dataset provides a solid data foundation for subsequent extreme value interval extraction, signal enhancement, and final quantitative content analysis, significantly improving the system's sensitivity and automation level, and contributing to high-throughput online monitoring and smart mine construction.

[0085] Specifically, for each set of reference data and comparison data, a ratio calculation is performed, dividing the reference data by the comparison data to obtain multiple ratio subsequences, including:

[0086] First, from the processed datasets of each metal component, data points at each time point are selected sequentially. The signal intensity of the rare earth metal component is used as the reference data for that time point, while the signal intensities of other metal components at the same time point are extracted as comparison data. For each set of reference and comparison data, a ratio calculation is performed: the value of the rare earth metal signal intensity at that time point is used as the numerator, and the signal intensity values ​​of each non-rare earth metal component at the same time point are used as the denominators, and the quotient is calculated. The result of each quotient calculation represents the relative amplitude relationship between the rare earth signal and a certain non-rare earth metal signal at the same sampling time. The above process is repeated for all sampling times and all non-rare earth metal components, resulting in multiple ratio subsequences arranged in chronological order. Each subsequence reflects the dynamic relationship between the rare earth signal and a specific non-rare earth metal signal.

[0087] Specifically, a moving average is applied to each ratio subsequence to remove abrupt outliers and smooth continuous data, resulting in a smoothed ratio subsequence, including:

[0088] For each ratio subsequence, a sliding window of length (e.g., a fixed number of consecutive data points) is set, and this window slides point by point across the entire ratio subsequence. At each step, the average value of all data points within the window is taken, and this average is used as the new data point at the current window center position, replacing the corresponding data points in the original ratio subsequence. In this way, the value of each data point is not only affected by itself but also by the combined effects of its preceding and following adjacent data points, effectively reducing the impact of sudden noise or outliers on the overall trend of the data sequence. After the moving average processing, the generated smooth ratio subsequence, while maintaining the main trend, greatly weakens data jumps caused by single-point anomalies or environmental disturbances, facilitating subsequent pattern recognition and interval analysis.

[0089] Specifically, the ratios of each ratio subsequence at each data point are sequentially combined to construct a feature vector. The feature vectors of all data points are then arranged according to the signal time sequence to generate a feature ratio dataset with a time-series structure, including:

[0090] For all smoothed ratio subsequences, at each identical sampling time, the ratio of each subsequence at that time is read and arranged in a fixed order to form a multidimensional feature vector. This feature vector fully reflects the relative relationship between the rare earth metal signal and all other non-rare earth metal signals at that sampling time. Arranging the feature vectors corresponding to each sampling time in chronological order of acquisition constitutes a multidimensional data sequence that varies over time, i.e., the feature ratio dataset. This dataset not only records the relative trends between the metal signals but also preserves the signal evolution details at each moment during the detection process, providing a rich information foundation for subsequent interval division, extreme value analysis, and rare earth signal identification.

[0091] In a preferred embodiment of the present invention, based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. High weights are assigned to intervals in the feature ratios that exhibit rare-earth metal signal response patterns, and low weights are assigned to intervals that exhibit non-rare-earth metal signal response patterns, thereby obtaining rare-earth metal feature enhancement data, including:

[0092] The feature ratio data is segmented according to a preset interval length. Extreme value search is performed on the interval of each segment, and local maximum and minimum values ​​are extracted in sequence to form the feature sequence of the interval.

[0093] Calculate the similarity between the feature sequence of each interval and the preset rare earth metal signal template. If the similarity is greater than the preset similarity threshold, the interval is marked as a high-weight interval; otherwise, it is marked as a low-weight interval.

[0094] For each high-weight interval, multiply all ratio data within the interval by a preset high-weight factor; for low-weight intervals, multiply the ratio data within the interval by a preset low-weight factor, and satisfy the condition that the high-weight factor is greater than the low-weight factor.

[0095] All weighted segmented data are summed and superimposed to form a rare earth metal feature enhancement data sequence.

[0096] In this embodiment of the invention, by extracting local extrema and performing segmented weighted superposition processing on the feature ratio data, the recognizability of rare earth metal signals in complex backgrounds can be significantly improved. The feature ratio data is segmented according to preset intervals, and extrema search and feature sequence extraction are performed on each segment interval. This effectively identifies abnormal peaks and valleys in the data; these extrema often correspond to local enhancement or weakening of rare earth metal signals. Subsequently, the similarity between the feature sequences of each interval and the rare earth metal signal template is calculated. Based on the calculation results, different weights are assigned to the intervals: highly matched intervals are weighted for enhancement, while mismatched intervals are weighted for weakening. Through this weighted superposition process, not only is the expression of rare earth metal feature signals in the global data strengthened, but the risk of false positives caused by impurity metals or noise is also greatly reduced. For example, when the signals of iron or copper in an ore sample fluctuate drastically, the system can automatically reduce the ratio weights of these intervals while weighting and amplifying the sensitive intervals where rare earth metal signals appear, thereby ensuring that the final extracted data more accurately reflects the actual changes in rare earth content. This processing method is applicable to a variety of practical detection scenarios. In particular, under conditions such as complex ores, coexistence of multiple metals, and strong background interference, it can still ensure the reliability and high accuracy of the detection results, thereby improving the feasibility of online automatic analysis and hierarchical decision-making.

[0097] The preset interval length specifically includes:

[0098] Before segmenting the feature ratio data, a fixed interval length is first set based on the actual application scenario and signal characteristics. This interval length can be set according to parameters such as historical detection data, target signal duration, and sampling rate. For example, the typical occurrence width of rare earth metal signal responses in samples can be statistically analyzed, and this width can be converted into the corresponding number of sampling points as the length benchmark for a single interval. After setting the interval length, the entire feature ratio dataset is sequentially divided into several continuous and equally long data intervals from the beginning. Each interval contains an equal number of sampling points, ensuring that subsequent extreme value search, feature sequence extraction, and weight allocation operations are completed within a standardized data window. This segmentation method with a preset length ensures that the statistical characteristics of the signal remain consistent within local intervals, which helps improve the accuracy of interval feature extraction and the stability of interval similarity comparison. In practice, if the detection task involves multiple signal durations, multiple candidate interval lengths can be set according to the actual target and compared. Finally, the interval length that best reflects the characteristics of the rare earth metal signal is selected for segmentation.

[0099] The preset high-weight factors and low-weight factors specifically include:

[0100] Before assigning weights to the segmented intervals of the feature ratio data, the values ​​of high-weight factors and low-weight factors need to be set according to the signal analysis task requirements and historical experimental data. High-weight factors are typically set to positive numbers greater than one, used to significantly amplify the influence of intervals highly similar to rare earth metal signal patterns in the final feature-enhanced data sequence. Low-weight factors are set to positive numbers no greater than one (e.g., one or a smaller fraction), used to weaken the contribution of data from intervals dissimilar to rare earth metal signal patterns. When setting these weights, the statistical distribution characteristics of rare earth signal and interference signal intervals in multiple samples can be analyzed to select weight combinations that effectively improve the recognition rate and reduce the false positive rate. During the assignment process, the system traverses all segmented intervals, multiplying all data points in high-similarity intervals by the high-weight factor and all data points in low-similarity intervals by the low-weight factor, ultimately outputting the weighted and accumulated feature-enhanced data. Through the distinguishing amplification and weakening effects of the weight factors, rare earth metal signal features can be more prominently expressed in complex backgrounds, effectively improving the robustness of signal recognition and the accuracy of automated system recognition.

[0101] In a preferred embodiment of the present invention, based on rare earth metal characteristic enhancement data, signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate rare earth metal content detection results, including:

[0102] Based on the characteristic enhancement data of rare earth metals, a main peak search is performed on the sequence to identify the main peak position and its signal intensity in the sequence, and the main peak signal response value is obtained.

[0103] Within a preset continuous interval on both sides of the main peak signal response value, the signal is integrated to obtain the area of ​​the interval reflecting the characteristic energy of rare earth.

[0104] Substitute the main peak signal response value and the interval area into the preset rare earth metal standard curve relationship to calculate the actual content of rare earth metals in the sample.

[0105] The actual content value is used as the rare earth metal content detection result, and this result is associated with the unique number of the sample to be tested.

[0106] In this embodiment of the invention, the original reflection signal data is subjected to spectral analysis based on pre-stored parameters of various metal reflection characteristics, and the signal is divided into multiple frequency bands, greatly improving the accuracy of signal decomposition and target metal identification. The use of time window truncation ensures detailed sampling of local signal changes, improving the resolution of subsequent frequency domain analysis. After each signal segment is accurately converted into a frequency domain signal through Fast Fourier Transform, the typical frequency responses corresponding to different metals are clearly separated. The system can accurately segment the original signal into several representative frequency intervals according to the parameter settings of each metal, and each interval reflects the typical response characteristics of the corresponding metal. For example, for composite mineral samples containing multiple metals, the system can automatically separate the frequency intervals of rare earth metals, iron, copper, etc., effectively avoiding cross-interference between different metal signals and laying a solid foundation for subsequent signal matching and feature extraction. This method significantly improves the accuracy of target metal detection in complex environments, shortens analysis time, and helps to achieve high-throughput continuous detection.

[0107] Specifically, based on rare earth metal characteristic enhancement data, a main peak search is performed on the sequence to identify the main peak position and its signal intensity, and the main peak signal response value is obtained, including:

[0108] First, starting from the beginning of the rare earth metal feature enhancement data sequence, each data point is traversed sequentially, and its signal intensity is compared with that of its neighboring data points. During the traversal, if the signal intensity of a data point is greater than that of its left and right neighbors, and this intensity value exceeds a preset background noise threshold, then this point is identified as a candidate peak. If multiple candidate peaks exist, the candidate point with the highest signal intensity is selected as the peak location, and its index in the sequence and its signal intensity are recorded. Finally, the specific location of the peak (i.e., the sampling number or corresponding time point) and the signal response value of the peak are output as the main feature parameters representing the significance of the rare earth metal signal.

[0109] Specifically, within a preset continuous interval to the left and right of the main peak signal response value, the signal is integrated to obtain the area of ​​the interval reflecting the characteristic energy of rare earth elements, which includes:

[0110] After determining the location of the main peak, based on the sequence index corresponding to the main peak, a predetermined number of sampling points are extended to the left and right to define a continuous interval symmetrical about the main peak. The length of this interval can be set based on experience or the peak width characteristics of the rare earth signal to ensure coverage of the main peak response and the main signal components around it. Within this continuous interval, the signal intensity of each sampling point is accumulated sequentially and multiplied by the sampling interval time or distance represented by each point. The accumulated result of all sampling points is used as the area value of the interval. This area value can comprehensively reflect the overall energy of the rare earth signal within the main peak and its neighborhood, eliminating random errors caused by single-point noise fluctuations, and is one of the key parameters for determining the actual content of rare earth metals.

[0111] Specifically, the main peak signal response value and the area of ​​the interval are substituted into the preset rare earth metal standard curve relationship to calculate the actual content of rare earth metals in the sample, including:

[0112] Based on the pre-calibrated rare earth metal standard curve of the detection system, the main peak signal response value and the area of ​​the interval are used as input parameters. The standard curve is obtained by fitting the actual content of a series of samples with known rare earth metal content to their corresponding main peak response value and area value, typically using linear fitting, piecewise linear fitting, or polynomial fitting methods. In practice, according to the relationship of the standard curve, the detected main peak signal response value and the area of ​​the interval are input, and the corresponding rare earth metal content is found or interpolated to achieve the quantitative conversion between signal intensity and actual content. The system finally outputs the actual rare earth metal content value in the detected sample for subsequent ore grading, resource assessment, or production control.

[0113] In a preferred embodiment of the present invention, based on pre-stored various metal reflection characteristic parameters, spectral analysis is performed on the original reflection signal data, and the original signal is divided into multiple frequency bands according to its frequency response characteristics, including:

[0114] Perform a time window truncation operation on the original reflected signal data to obtain signal segments within multiple time periods;

[0115] For each signal segment, the fast Fourier transform method is used to convert the time-domain signal into a frequency-domain signal, and the amplitude-frequency response curve of the signal segment is obtained.

[0116] Based on the typical frequency response range of various metal reflection characteristic parameters, the frequency domain signal is divided into different frequency intervals, and each frequency interval is generated as an independent frequency band.

[0117] The data from all frequency ranges are combined sequentially to form multiple frequency bands that cover the entire analysis range.

[0118] In this embodiment of the invention, by precisely matching each frequency band signal with different metal characteristic parameters and archiving them to obtain preliminary response signal sets for each metal, the accuracy of target metal identification can be effectively improved. Specifically, the system extracts amplitude and phase sequences from each frequency band signal, and then compares these features with standard parameters (such as peak amplitude, specific phase characteristics, etc.) for each metal. Using a set similarity algorithm and threshold judgment criteria, the system can filter out the signal segments that best match specific metals such as rare earth metals from all frequency bands. Thus, each archived response signal set has extremely high metal specificity, significantly reducing signal confusion caused by metal coexistence or noise. For example, when detecting samples containing multiple metals, even if some metal signals have weak amplitudes, as long as their characteristics are highly consistent with preset parameters, they can still be accurately identified and stored. This process lays the foundation for subsequent denoising and normalization operations, further improving the specificity, anti-interference capability, and detection reliability of the entire detection system.

[0119] Specifically, for each signal segment, a Fast Fourier Transform (FFT) is used to convert the time-domain signal into a frequency-domain signal, obtaining the amplitude-frequency response curve of that signal segment. This includes:

[0120] For each signal segment obtained after time window truncation, its raw time-domain data is first input to the signal processing unit. The system employs a Fast Fourier Transform (FFT) algorithm to perform spectral analysis on the time-domain signal, converting the amplitude information of continuous time sampling points into amplitude and phase information of different frequency components. After processing, the system outputs a sequence of data points with frequency as the independent variable and amplitude as the dependent variable, which is the amplitude-frequency characteristic curve of the signal segment. This curve visually reflects the energy distribution of the segment at various frequency components, providing a foundation for subsequent metal component identification and frequency band analysis.

[0121] Specifically, based on the typical frequency response range of various metal reflection characteristic parameters, the frequency domain signal is divided into different frequency intervals, with each frequency interval generated as an independent frequency band, including:

[0122] Based on pre-stored metal reflection characteristic parameters, the system can retrieve the typical response range of the target metal within the millimeter-wave frequency range. Subsequently, the system groups all data points on the amplitude-frequency response curve according to their frequency magnitude, dividing them into different frequency ranges. Each range corresponds to a specific response range for one or more metals, and a separate dataset for that range is generated, containing all amplitude and phase data within that frequency range. In this way, the original frequency domain signal can be divided into several frequency ranges covering the typical responses of all target metals, creating conditions for feature extraction and identification of each metal.

[0123] This involves combining data from all frequency ranges sequentially to form multiple frequency bands covering the entire analysis scope, specifically including:

[0124] The system sequentially arranges all individually generated frequency range data according to the sampling or detection order of the original signal, combining them into an ordered set of frequency band data. Each frequency band's data not only includes its amplitude, phase, and other characteristic information, but also retains the time information and frequency label corresponding to the original signal segment. The data set of all frequency bands collectively covers the response performance of the signal segment under test across the entire analysis frequency range. This combination process ensures that subsequent metal feature matching and comparison can be performed smoothly based on complete and unrestricted data, improving the overall accuracy of signal analysis and system compatibility.

[0125] In a preferred embodiment of the present invention, based on the signal of each frequency band, signal segments matching different metal characteristic parameters are extracted to obtain a preliminary response signal set for each metal, including:

[0126] Based on the signal in each frequency band, extract the amplitude sequence and phase sequence within that frequency band as features to be compared.

[0127] For each type of metal, the peak amplitude parameters and phase characteristic parameters at a specific preset frequency are called up and compared one by one with the characteristics to be compared in each frequency band to form a difference sequence for each frequency band.

[0128] Based on the difference sequence, a similarity calculation method is set, and the amplitude difference and phase difference of the frequency band are compared with the corresponding preset difference thresholds. If all differences are lower than their respective difference thresholds, the signal segment of the frequency band is determined to match the metal feature parameter, and a matching mark is generated.

[0129] All signal segments that are determined to be matched are archived into their corresponding metal categories to generate a preliminary set of response signals for each metal.

[0130] In this embodiment of the invention, by denoising the preliminary response signal sets of each metal, the purity of the signal and the accuracy of subsequent analysis can be significantly improved. Specifically, a moving average method is used to smooth each signal set, effectively reducing random high-frequency noise introduced by the measurement environment, equipment vibration, or instantaneous interference. For isolated high-amplitude abrupt changes in the signal, the difference in amplitude between the signal and the preceding and following data points is detected. After being identified as high-frequency noise points, the average of adjacent data is used to replace the noise points, thereby eliminating the distortion of the overall data by the outliers. In addition, the system automatically fits the baseline trend for each signal and subtracts the baseline drift, achieving adaptive correction for slow drift or low-frequency interference. After the above multi-level denoising and correction processing, the resulting denoised response signal set retains the effective components of the original signal while greatly reducing irrelevant interference. For example, when detecting complex mineral samples in the field, even if the signal has strong background noise and instrument baseline drift, the system can automatically recover high-quality metal response characteristics, providing a reliable data basis for further normalization and quantitative analysis, and effectively improving the accuracy and stability of the detection.

[0131] The preset specific frequencies include:

[0132] Before performing metal characteristic parameter matching, the system first pre-determines one or more typical response frequencies for each target metal based on historical detection data or publicly available literature describing the metal's physical properties. These preset frequencies typically correspond to the frequency points where the target metal exhibits significant reflection or absorption characteristics under millimeter-wave irradiation. During implementation, the system locates these specific frequencies in the signal spectrum and focuses on the amplitude, phase, and other parameter characteristics of the signal at those frequencies. For example, if rare earth metals exhibit a significant peak in reflectivity within a specific frequency range, the signal characteristics at that frequency point will be the focus of matching and analysis in subsequent data comparisons. The preset specific frequencies can be obtained through experimental calibration or flexibly adjusted according to different ore types to ensure the system's sensitive response and high-precision identification of various metal signals.

[0133] Specifically, based on the difference sequence, a similarity calculation method is set, and the amplitude difference and phase difference of the frequency band are compared with the corresponding preset difference thresholds, including:

[0134] First, for the characteristic parameters of the frequency band signal to be compared and the target metal, the corresponding amplitude and phase sequences are extracted respectively. The amplitude sequence of the frequency band signal is subtracted one-to-one from the amplitude sequence of the target metal parameters to form an amplitude difference sequence; the phase sequence is obtained in the same way to obtain a phase difference sequence. Then, according to a predetermined similarity calculation method, such as using the maximum absolute difference or root mean square difference, the system summarizes and calculates the amplitude difference sequence and phase difference sequence respectively to obtain amplitude difference index and phase difference index representing the overall degree of difference.

[0135] Each calculated difference index is compared with a pre-set difference threshold. If the amplitude difference index is less than or equal to the amplitude threshold, and the phase difference index is less than or equal to the phase threshold, the signal in that frequency band is considered to match the target metal's characteristic parameters; otherwise, it is considered a mismatch. The thresholds can be set based on historical comparison data for different metals and scenarios, or actual engineering requirements, thereby ensuring that the system can effectively suppress false positives and false negatives and improve the accuracy of metal matching.

[0136] In a preferred embodiment of the present invention, the preliminary response signal sets of each metal are denoised to filter isolated high-frequency noise points, and the signal baseline drift is adaptively corrected to obtain a denoised response signal set, including:

[0137] For each set of preliminary response signals, the signal curves are smoothed using a moving average method to generate preliminary smoothed signals.

[0138] For each data point of the preliminary smoothed signal, the difference between its amplitude and the amplitude of the adjacent data points is determined. If the amplitude change of the data point exceeds the preset amplitude change threshold, it is regarded as an isolated high-frequency noise point, and the average of the adjacent data points is used to replace the data point to generate a denoised signal.

[0139] Based on the denoised signal, its baseline trend curve is fitted, and the baseline trend curve is subtracted from the signal to form a set of denoised response signals.

[0140] In this embodiment of the invention, by segmenting the feature sequences of each interval using a sliding window method, combined with normalization and point-to-point comparison mechanisms, accurate similarity calculation between interval features and rare earth metal signal templates is achieved. The sliding window segmentation method can fully capture the detailed changes of the signal within the interval, improving the sensitivity of pattern recognition to local features. The normalization operation eliminates amplitude and scale differences between different sub-sequences, ensuring that the comparison results only reflect the consistency of morphology and change patterns. By mapping the normalized sub-sequences one-to-one with the signal template segments and calculating the multiplication and summation point by point to obtain the local similarity score, the system can accurately identify intervals that highly match the template. After weighted averaging of the local similarity scores of all normalized sub-sequences, the obtained overall interval similarity score reflects both the matching degree of the principal components of the signal and avoids the influence of individual outliers. For example, in complex scenarios of automatically identifying rare earth ore, the system can accurately determine whether there are rare earth metal response features in the signal interval based on this method, greatly enhancing the detection sensitivity and accuracy of rare earth signals, and providing data basis for subsequent weighted processing and final result output.

[0141] For each set of preliminary response signals, the signal curves are smoothed using a moving average method to generate preliminary smoothed signals, specifically including:

[0142] For the initial response signal set obtained under each metal category, a suitable sliding window length is first selected for the signal sequence. Then, starting from the beginning of the signal sequence, the window slides point by point, taking all data points within the window each time, calculating their arithmetic mean, and using this average as the new signal value at the center of the window. This process is repeated to cover the entire signal sequence, replacing each point in the original signal with the average of several points around it, resulting in a smoothed signal sequence. The moving average method effectively eliminates random high-frequency noise and sudden abnormal fluctuations during the measurement process, making the signal curve more continuous and accurately reflecting the changing trend of the metal response, laying a data foundation for subsequent noise point identification and baseline fitting.

[0143] The preset amplitude change threshold specifically includes:

[0144] When detecting anomalies in a smooth signal, the system needs to pre-set an amplitude change threshold. This threshold can be determined based on statistical analysis of historical sample signal amplitudes or actual testing experience, and is usually taken as the maximum expected change amplitude within the normal fluctuation range of the signal. In actual operation, when determining whether a sampling point is a noise point, the system calculates the amplitude difference between that point and its adjacent sampling points (such as one before and one after). If this difference is greater than the preset amplitude change threshold, the point is determined to be an abnormal abrupt change point or an isolated high-frequency noise point. In this case, the system replaces its value with the mean of the adjacent sampling points to ensure that the signal curve is continuous and smooth. Through this setting, sporadic abnormal signals can be effectively filtered out without damaging normal metallic response characteristics.

[0145] Specifically, based on the denoised signal, a baseline trend curve is fitted, and this baseline trend curve is subtracted from the signal to form a set of denoised response signals, including:

[0146] After noise removal and signal smoothing, the system employs a baseline fitting algorithm (such as least squares polynomial fitting) to perform global trend analysis on the remaining signal sequence. Using the amplitude and corresponding index of all sampling points as input, the system fits a baseline curve representing the overall trend of the signal. Subsequently, the system iterates through each sampling point of the signal sequence, subtracting the value of the fitted baseline curve at that point from the amplitude of the sampling point, thus automatically correcting for baseline drift. The resulting signal is the set of response signals after baseline drift removal. This processing eliminates the influence of slowly varying background factors such as instrument zero-point changes and ambient temperature drift, allowing the final signal to more accurately reflect the true instantaneous response of the metal composition and greatly improving the accuracy of subsequent feature analysis and quantitative detection.

[0147] In a preferred embodiment of the present invention, calculating the similarity between the feature sequence of each interval and the preset rare earth metal signal template includes:

[0148] For the feature sequence of each interval, the sliding window method is used to segment it, dividing the feature sequence within the interval into several overlapping preliminary subsequences;

[0149] For each initial subsequence, perform a normalization operation to make the mean of the initial subsequence zero and the standard deviation one, thus obtaining a normalized subsequence;

[0150] Each normalized subsequence is compared point-to-point with a rare earth metal signal template fragment of the corresponding length. The local similarity score between the normalized subsequence and the template fragment is obtained by calculating the sum of the products of their respective components.

[0151] The local similarity scores of all normalized subsequences within the same interval are weighted and averaged to obtain the overall similarity score of that interval.

[0152] In this embodiment of the invention, by segmenting the feature ratio data according to a preset interval length, performing extreme value search and feature sequence extraction on each segment interval, and combining similarity calculation and weighted processing, the expression and distinguishability of rare earth metal signals in the overall data can be further enhanced. Specifically, the interval segmentation and extreme value search process can effectively discover the main variation characteristics of the signal in different intervals, distinguishing rare earth signals from interference signals. For high similarity intervals, by assigning high weights and accumulating their data, the prominence of rare earth signals is significantly improved; while for low similarity intervals, by reducing their weights, background interference is effectively suppressed. Finally, the sum of all weighted segmented data forms a rare earth metal feature enhancement data sequence that can truly reflect the spatial distribution and content variation trend of rare earth metals in the sample. For example, during continuous ore sampling or real-time monitoring, this processing mechanism can dynamically capture the occurrence, enhancement, or weakening of rare earth signals, providing technical support for the efficient development and intelligent grading of rare earth mineral resources, and improving the automation level and adaptability of the system.

[0153] Embodiments of the present invention also provide an AI rare earth detection system based on millimeter-wave radar multimetal reflectance spectra, the system comprising:

[0154] The data acquisition module is used to acquire the raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to the millimeter-wave signal.

[0155] The signal decomposition module is used to perform signal decomposition processing on the original reflected signal data, separating the reflected signal into multiple metal signal components based on the metal reflection characteristic parameters, and obtaining the dataset of each metal component.

[0156] The feature ratio calculation module is used to calculate the feature ratio of each metal component dataset. Taking the rare earth metal component as the benchmark, it constructs the feature ratio sequence between the rare earth metal component and the non-rare earth metal component to obtain the feature ratio data.

[0157] The weighted enhancement module is used to perform local extremum extraction and segmented weighted superposition processing based on the feature ratio data. It assigns high weight to the intervals with rare earth metal signal response patterns in the feature ratios and low weight to the intervals with non-rare earth metal signal response patterns, thereby obtaining rare earth metal feature enhancement data.

[0158] The quantitative analysis module is used to extract the signal response indicators of rare earth metals based on the characteristic enhancement data of rare earth metals, perform quantitative analysis, and generate the rare earth metal content detection results.

[0159] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0160] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0161] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0162] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI-based rare earth detection method based on millimeter-wave radar multimetal reflectance spectra, characterized in that, The method includes: Acquire raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to millimeter-wave signals; The original reflected signal data is processed by signal decomposition, and the reflected signal is separated into multiple metal signal components based on the metal reflection characteristic parameters to obtain the dataset of each metal component. The feature ratios of each metal component dataset are calculated. Based on the rare earth metal components, a feature ratio sequence between the rare earth metal components and the non-rare earth metal components is constructed to obtain the feature ratio data. Based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. The intervals with rare earth metal signal response patterns in the feature ratios are given high weights, and the intervals with non-rare earth metal signal response patterns are given low weights, thus obtaining rare earth metal feature enhancement data. Based on the characteristic enhancement data of rare earth metals, the signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate the rare earth metal content detection results. The original reflected signal data is decomposed into multiple metallic signal components based on metallic reflection characteristic parameters, resulting in datasets for each metallic component, including: Based on the pre-stored parameters of various metal reflection characteristics, the original reflected signal data is subjected to spectrum analysis, and the original signal is divided into multiple frequency bands according to the frequency response characteristics. Based on the signal of each frequency band, the signal segments that match the characteristic parameters of different metals are extracted to obtain the preliminary response signal set of each metal; The initial response signal set of each metal is denoised to filter out isolated high-frequency noise points and adaptively correct the signal baseline drift to obtain the denoised response signal set. Based on the set of denoised response signals, the amplitude of each metal signal segment is normalized, and all segments are standardized to a uniform amplitude range to obtain the set of normalized response signals. Each metal signal segment in the normalized response signal set is archived according to metal category to generate a dataset of each metal component.

2. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, Feature ratios were calculated for each metal component dataset. Using rare earth metal components as a benchmark, feature ratio sequences were constructed between rare earth metal components and non-rare earth metal components, yielding feature ratio data, including: Iterate through all data points in each metal component dataset, using the signal intensity of rare earth metal components as reference data, and using the signal intensity of all non-rare earth metal components at the corresponding locations as comparison data. Based on each set of reference data and comparison data, a ratio calculation is performed, dividing the reference data by the comparison data to obtain multiple ratio subsequences, each ratio subsequence corresponding to a non-rare earth metal component; Apply a moving average to each ratio subsequence to remove abrupt outliers and smooth continuous data, resulting in a smoothed ratio subsequence. The ratios of each ratio subsequence at each data point are combined sequentially to construct a feature vector. The feature vectors of all data points are then arranged in sequence according to the signal time sequence to generate a feature ratio dataset with a time sequence structure.

3. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, Based on the feature ratio data, local extremum extraction and segmented weighted superposition processing are performed. Intervals exhibiting rare-earth metal signal response patterns are assigned high weights, while intervals exhibiting non-rare-earth metal signal response patterns are assigned low weights, resulting in rare-earth metal feature enhancement data, including: The feature ratio data is segmented according to a preset interval length. Extreme value search is performed on the interval of each segment, and local maximum and minimum values ​​are extracted in sequence to form the feature sequence of the interval. Calculate the similarity between the feature sequence of each interval and the preset rare earth metal signal template. If the similarity is greater than the preset similarity threshold, the interval is marked as a high-weight interval; otherwise, it is marked as a low-weight interval. For each high-weight interval, multiply all ratio data within the interval by a preset high-weight factor; for low-weight intervals, multiply the ratio data within the interval by a preset low-weight factor, and satisfy the condition that the high-weight factor is greater than the low-weight factor. All weighted segmented data are summed and superimposed to form a rare earth metal feature enhancement data sequence.

4. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, Based on the characteristic enhancement data of rare earth metals, signal response indicators of rare earth metals are extracted and quantitatively analyzed to generate rare earth metal content detection results, including: Based on the characteristic enhancement data of rare earth metals, a main peak search is performed on the sequence to identify the main peak position and its signal intensity in the sequence, and the main peak signal response value is obtained. Within a preset continuous interval on both sides of the main peak signal response value, the signal is integrated to obtain the area of ​​the interval reflecting the characteristic energy of rare earth. Substitute the main peak signal response value and the interval area into the preset rare earth metal standard curve relationship to calculate the actual content of rare earth metals in the sample. The actual content value is used as the rare earth metal content detection result, and this result is associated with the unique number of the sample to be tested.

5. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, Based on pre-stored parameters of various metal reflection characteristics, spectral analysis is performed on the original reflected signal data, dividing the original signal into multiple frequency bands according to its frequency response characteristics, including: Perform a time window truncation operation on the original reflected signal data to obtain signal segments within multiple time periods; For each signal segment, the fast Fourier transform method is used to convert the time-domain signal into a frequency-domain signal, and the amplitude-frequency response curve of the signal segment is obtained. Based on the typical frequency response range of various metal reflection characteristic parameters, the frequency domain signal is divided into different frequency intervals, and each frequency interval is generated as an independent frequency band. The data from all frequency ranges are combined sequentially to form multiple frequency bands that cover the entire analysis range.

6. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, Based on the signal in each frequency band, signal segments matching the characteristic parameters of different metals are extracted to obtain a preliminary set of response signals for each metal, including: Based on the signal in each frequency band, extract the amplitude sequence and phase sequence within that frequency band as features to be compared. For each type of metal, the peak amplitude parameters and phase characteristic parameters at a specific preset frequency are called up and compared one by one with the characteristics to be compared in each frequency band to form a difference sequence for each frequency band. Based on the difference sequence, a similarity calculation method is set, and the amplitude difference and phase difference of the frequency band are compared with the corresponding preset difference thresholds. If all differences are lower than their respective difference thresholds, the signal segment of the frequency band is determined to match the metal feature parameter, and a matching mark is generated. All signal segments that are determined to be matched are archived into their corresponding metal categories to generate a preliminary set of response signals for each metal.

7. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 1, characterized in that, The initial response signal sets of each metal are denoised to filter isolated high-frequency noise points, and the signal baseline drift is adaptively corrected to obtain a denoised response signal set, including: For each set of preliminary response signals, the signal curves are smoothed using a moving average method to generate preliminary smoothed signals. For each data point of the preliminary smoothed signal, the difference between its amplitude and the amplitude of the adjacent data points is determined. If the amplitude change of the data point exceeds the preset amplitude change threshold, it is regarded as an isolated high-frequency noise point, and the average of the adjacent data points is used to replace the data point to generate a denoised signal. Based on the denoised signal, its baseline trend curve is fitted, and the baseline trend curve is subtracted from the signal to form a set of denoised response signals.

8. The AI ​​rare earth detection method based on millimeter-wave radar multimetal reflectance spectra according to claim 3, characterized in that, Calculate the similarity between the feature sequences of each interval and the preset rare earth metal signal template, including: For the feature sequence of each interval, the sliding window method is used to segment it, dividing the feature sequence within the interval into several overlapping preliminary subsequences; For each initial subsequence, perform a normalization operation to make the mean of the initial subsequence zero and the standard deviation one, thus obtaining a normalized subsequence; Each normalized subsequence is compared point-to-point with a rare earth metal signal template fragment of the corresponding length. The local similarity score between the normalized subsequence and the template fragment is obtained by calculating the sum of the products of their respective components. The local similarity scores of all normalized subsequences within the same interval are weighted and averaged to obtain the overall similarity score of that interval.

9. An AI rare earth detection system based on millimeter-wave radar multimetal reflectance spectra, characterized in that, The system, used in any one of claims 1 to 8, comprises: The data acquisition module is used to acquire the raw millimeter-wave radar reflection signal data of the sample under test, including the reflection response curves of various metals in the sample to the millimeter-wave signal. The signal decomposition module is used to perform signal decomposition processing on the original reflected signal data, separating the reflected signal into multiple metal signal components based on the metal reflection characteristic parameters, and obtaining the dataset of each metal component. The feature ratio calculation module is used to calculate the feature ratio of each metal component dataset. Taking the rare earth metal component as the benchmark, it constructs the feature ratio sequence between the rare earth metal component and the non-rare earth metal component to obtain the feature ratio data. The weighted enhancement module is used to perform local extremum extraction and segmented weighted superposition processing based on the feature ratio data. It assigns high weight to the intervals with rare earth metal signal response patterns in the feature ratios and low weight to the intervals with non-rare earth metal signal response patterns, thereby obtaining rare earth metal feature enhancement data. The quantitative analysis module is used to extract the signal response indicators of rare earth metals based on the characteristic enhancement data of rare earth metals, perform quantitative analysis, and generate the rare earth metal content detection results.

Images