Fluorite mine geological data edge node real-time processing system

Abstract

This application provides a real-time processing system for edge nodes of fluorite geological data, relating to the fields of intelligent power system inspection and non-contact fault diagnosis. First, it performs time alignment on heterogeneous data streams such as triaxial vibration acceleration, X-ray fluorescence pulse counting, and near-infrared spectral reflectance. Then, it adaptively adjusts the filtering intensity using the power spectral density of the vibration components, suppressing mechanical noise while preserving the narrow pulse characteristics of grade abrupt changes, thus solving the problems of detail loss and signal distortion caused by the overlap of noise bands and effective signals. Based on the energy spectrum characteristics of calcium and fluorine elements in fluorite, it performs significant weighted extraction, compressing the high-dimensional signal into a compact feature vector. Real-time grade inference is then completed through an edge-side quantization neural network. Finally, the compression ratio is dynamically adjusted according to the fluctuation amplitude of grade prediction: increasing the compression ratio to save bandwidth when the grade is stable, and decreasing the compression ratio to preserve key information when the grade changes abruptly, achieving an adaptive balance between transmission load and data fidelity.

Description

Technical Field

[0001] This application relates to the fields of edge computing and mining geological data processing technology, and more specifically, to a real-time processing system for edge nodes of fluorite geological data. Background Technology

[0002] As an important strategic non-metallic mineral resource, the real-time and accurate determination of ore grade during fluorite mining directly affects the optimization of mining plans, the improvement of mineral processing efficiency, and the control of production energy consumption. In underground or open-pit mining sites, various sensors such as X-ray fluorescence detectors, near-infrared spectrometers, and vibration accelerometers continuously generate high-frequency, multimodal raw data streams. However, the bandwidth of wireless communication links in mining areas is extremely limited, making it difficult to transmit all the raw data back to a remote cloud server for centralized processing and analysis in real time. Therefore, performing signal noise reduction, mineral feature extraction, grade prediction, and data compression and encapsulation on-site at edge computing nodes close to the data source has become a key technological approach to ensure the timeliness of mining decisions and the economy of data transmission.

[0003] However, existing edge-side geological data processing schemes have significant shortcomings in dealing with the complex working conditions of fluorite mines. The non-stationary vibration noise frequency generated by mechanical drilling highly overlaps with the effective signal frequency band of the geological detector. However, the computing power of edge nodes is limited, making it difficult to run highly complex time-frequency separation algorithms. Conventional low-pass filtering or simple averaging processes suppress noise but also smooth out the narrow pulse characteristics that represent grade changes, resulting in the loss of geological details and signal distortion, which directly affects the accuracy of subsequent grade prediction. In particular, the existing first-order recursive filtering mechanism passively adjusts the smoothing factor based solely on the background noise power spectral density, failing to consider the nonlinear physical coupling between mechanical drilling vibration and X-ray fluorescence pulse detection. Due to the brittle texture and uneven distribution of fluorite veins, the non-stationary harmonic vibrations induced when the drill bit cuts rock layers of different hardness will cause the detector probe to produce micron-level periodic displacements, forming a systematic intervention similar to amplitude modulation and frequency modulation on the pulse counting signal. This results in severe phase lag in the processed signal, accompanied by a large number of spurious peaks unrelated to the true grade, weakening the edge nodes' accuracy in identifying vein boundaries and making it difficult to meet the stringent requirements of real-time high-fidelity noise reduction in industrial settings.

[0004] Therefore, an optimized real-time processing system for edge nodes of fluorite geological data is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a real-time processing system for edge nodes of fluorite mineral geological data, comprising: The timing alignment module is used to perform timing alignment processing on the acquired raw sensor data stream to obtain the aligned raw matrix. The raw sensor data stream includes triaxial vibration acceleration components, X-ray fluorescence intensity pulse count, and near-infrared spectral reflectance. An adaptive noise suppression module is used to adaptively suppress environmental noise on the aligned original matrix to obtain a denoised geological signal; The feature saliency extraction module is used to extract fluorite feature saliency from the denoised geological signal to obtain salient mineral feature vectors; The edge inference module is used to input significant mineral feature vectors into a pre-deployed quantized neural network for forward inference to obtain edge grade prediction results; The dynamic compression and encapsulation module is used to determine the compression ratio coefficient based on the fluctuation range of the edge grade prediction results, and to perform dynamic compression and data encapsulation on the significant mineral feature vectors according to the compression ratio coefficient to obtain optimized transmission data packets.

[0006] Compared with existing technologies, the real-time processing system for edge nodes of fluorite ore geological data provided in this application solves the problem of geological data being unable to be transmitted back and centrally processed in real time under the condition of limited communication bandwidth in the mining area by constructing a complete processing link from multi-source sensor data alignment to grade prediction and adaptive compression transmission at the edge side. First, it performs time-series alignment on data streams from heterogeneous sensors such as triaxial vibration acceleration, X-ray fluorescence pulse counting, and near-infrared spectral reflectance to form a unified observation matrix. Then, it adaptively adjusts the filtering intensity using the power spectral density of vibration components to suppress mechanical construction noise while retaining narrow pulse characteristics representing grade changes, thus solving the problems of geological detail loss and signal distortion caused by the overlap of noise bands and effective signals. Next, it performs significant weighted extraction based on the energy spectrum characteristics of calcium and fluorine elements in fluorite, compressing the high-dimensional geological signal into a compact feature vector for fluorite grade determination. Then, it completes real-time grade inference through a quantized neural network deployed on the edge side. Finally, it dynamically adjusts the data compression ratio according to the fluctuation range of the grade prediction results, increasing the compression ratio during periods of stable grade to save bandwidth and decreasing the compression ratio during periods of grade changes to preserve key geological information, achieving an adaptive balance between transmission load and data fidelity. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1This is a system block diagram of a real-time processing system for edge nodes of fluorite ore geological data according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application; Figure 3 This is a block diagram of the adaptive noise suppression module in the real-time processing system for edge nodes of fluorite geological data according to an embodiment of this application; Figure 4 This is a block diagram of a time-series smoothing unit in a real-time processing system for edge nodes of fluorite geological data according to an embodiment of this application. Figure 5 This is a block diagram of the feature saliency extraction module in the real-time processing system for edge nodes of fluorite geological data according to an embodiment of this application; Figure 6 This is a block diagram of the edge inference module in the real-time processing system for edge nodes of fluorite geological data according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] This application uses system block diagrams and data flow diagrams to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing geological data processing schemes for the edge of fluorite mines face significant bottlenecks when dealing with complex mining conditions. The non-stationary vibration noise generated by mechanical drilling highly overlaps with the effective signal frequency band of the geological detector. Conventional filtering methods, while suppressing noise, also smooth out narrow pulse characteristics representing grade abrupt changes, leading to the loss of geological details and signal distortion. This severely restricts the real-time and accurate determination of vein grade by edge nodes. In particular, existing first-order recursive filtering mechanisms fail to consider the nonlinear physical coupling between mechanical vibration and X-ray fluorescence pulse detection. Harmonic vibrations induced when the drill bit cuts rock layers of varying hardness cause micrometer-level periodic displacements in the detector, systematically interfering with the pulse counting signal. This results in phase lag and numerous spurious peaks in the processed signal, making it difficult to meet the stringent real-time high-fidelity noise reduction requirements of industrial sites. Therefore, this application proposes a real-time processing system for edge nodes of fluorite mine geological data. Specifically, the system first performs time-series alignment on multiple sensor data streams, including triaxial vibration acceleration, X-ray fluorescence pulse count, and near-infrared spectral reflectance, to form a unified observation matrix. Then, it adaptively adjusts the filtering intensity using the power spectral density of the vibration components to suppress environmental noise while preserving effective geological features. Next, it performs significant weighted extraction based on the energy spectrum characteristics of calcium and fluorine elements in fluorite to generate a compact feature vector for grade determination. Real-time grade inference is then performed through a quantized neural network deployed at the edge. Finally, the data compression ratio is dynamically adjusted based on the fluctuation range of the grade prediction results to balance transmission load and data fidelity. In the noise suppression stage, a vibration-signal time-domain derivative coupling estimation and dynamic phase gain weight correction mechanism are further introduced. By quantifying the transient interference of mechanical vibration on the pulse signal and actively compressing the trust weight of the interfered signal accordingly, and combining this with a phase compensation term derived from the drilling rig's main frequency to perform phase alignment on historical reference signals, the system fundamentally eliminates geological feature distortion and phase lag caused by vibration phase interference, achieving a transient fidelity response to changes in fluorite vein grade.

[0014] Figure 1 This is a system block diagram of a real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in a real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application. Figure 1 and Figure 2As shown, the real-time processing system 100 for edge nodes of fluorite geological data according to an embodiment of this application includes: a time alignment module 110, used to perform time alignment processing on the acquired raw sensor data stream to obtain an aligned raw matrix, the raw sensor data stream including triaxial vibration acceleration components, X-ray fluorescence intensity pulse count, and near-infrared spectral reflectance; an adaptive noise suppression module 120, used to perform adaptive environmental noise suppression on the aligned raw matrix to obtain a denoised geological signal; a feature saliency extraction module 130, used to extract fluorite feature saliency from the denoised geological signal to obtain a significant mineral feature vector; an edge inference module 140, used to input the significant mineral feature vector into a pre-deployed quantized neural network for forward inference to obtain an edge grade prediction result; and a dynamic compression and encapsulation module 150, used to determine a compression ratio coefficient based on the fluctuation amplitude of the edge grade prediction result, and perform dynamic compression and data encapsulation on the significant mineral feature vector according to the compression ratio coefficient to obtain an optimized transmission data packet.

[0015] In the aforementioned real-time processing system 100 for edge nodes of fluorite mine geological data, the time alignment module 110 is used to perform time alignment processing on the acquired raw sensor data stream to obtain an aligned raw matrix. The raw sensor data stream includes triaxial vibration acceleration components, X-ray fluorescence intensity pulse counts, and near-infrared spectral reflectance. It should be noted that, since multiple heterogeneous sensors are deployed simultaneously at the fluorite mine mining site, the triaxial vibration accelerometer, X-ray fluorescence detector, and near-infrared spectrometer each have different sampling frequencies, analog-to-digital conversion accuracies, and signal output formats. Furthermore, the hardware clocks of each sensor have inherent drift deviations. If these asynchronously acquired raw data streams are directly sent to subsequent processing stages, the timestamp misalignment between different channels will cause the vibration information, fluorescence pulse information, and spectral reflectance information corresponding to the same geological sampling point to not strictly correspond in the time domain, thereby causing cross-time data aliasing in the subsequent noise suppression and feature extraction processes. Based on this, the technical solution of this application first performs time alignment processing on the acquired raw sensor data stream to obtain an aligned raw matrix. Through the above processing, multi-channel sensor data from different sampling rates and different hardware clocks can be unified to the same time base, ensuring that the vibration component, fluorescence pulse count and spectral reflectance at each time sampling point form a strict one-to-one correspondence, providing a time-domain consistent data foundation for subsequent adaptive noise suppression and fluorite feature extraction.

[0016] More specifically, in a concrete example of this application, the acquired raw sensor data streams are first subjected to concurrent conversion processing to obtain multiple raw digital signals. In fluorite mine drilling operations, a triaxial vibration accelerometer mounted on the drilling rig outputs analog voltage signals at a frequency of thousands of times per second, a X-ray fluorescence detector mounted at the drill pipe tip outputs discrete photon event sequences in a pulse counting manner, and a near-infrared spectrometer outputs a reflectance array with a fixed integration period. Edge nodes perform concurrent sampling of the above three types of analog or pulse signals through a multi-channel analog-to-digital converter, converting the raw physical quantities of each sensor into corresponding digital sequences, thereby obtaining multiple raw digital signals.

[0017] Subsequently, based on preset reference drift correction parameters, zero-point calibration is performed on multiple raw digital signals to obtain calibrated digital signals. Due to temperature variations in the fluorite mining environment and the aging of electronic components caused by long-term sensor operation, the zero-point reference of each sensor will slowly drift over time. For example, the output value of the vibration accelerometer deviates from the theoretical zero point in a static state, and the dark count rate of the fluorescence detector increases with increasing temperature. To eliminate such systematic deviations, the reference drift correction parameters, pre-calibrated and stored in the non-volatile memory of the edge nodes, are used to perform a correction operation by subtracting the corresponding zero-point offset from each of the multiple raw digital signals, so that the static reference of each channel signal returns to the nominal zero position, thereby obtaining the calibrated digital signal.

[0018] Subsequently, using the highest sampling frequency as the clock reference, interpolation and alignment encapsulation were performed on the calibrated digital signals to obtain the original alignment matrix. Since the original sampling frequencies of the three types of sensors differ—the sampling rate of the vibration accelerometer is much higher than that of the X-ray fluorescence detector and the near-infrared spectrometer—to achieve point-by-point alignment of the data from each channel in the time domain, the channel with the highest sampling frequency was selected as the unified clock reference. Linear interpolation or zero-order hold interpolation was performed on the fluorescence pulse counting channel and the spectral reflectance channel, which had lower sampling frequencies, to fill in their data point density to the same time resolution as the reference channel. After interpolation, the three calibrated and aligned digital signals were arranged column-by-column according to the time dimension and encapsulated into a two-dimensional matrix structure where rows represent sampling times and columns represent sensor channels, i.e., the original alignment matrix.

[0019] In the aforementioned real-time processing system 100 for edge nodes of fluorite ore geological data, the adaptive noise suppression module 120 is used to adaptively suppress environmental noise on the aligned original matrix to obtain a denoised geological signal. It should be noted that, given the high overlap between the non-stationary vibration noise generated by mechanical drilling at the fluorite mine site and the effective signal frequency band of the geological detector, and the drastic fluctuations in noise intensity with changes in drilling depth and rock hardness, fixed-parameter filtering methods, while suppressing noise, also smooth out the narrow pulse characteristics representing grade abrupt changes. Furthermore, the limited computing power of the edge nodes cannot support the real-time operation of highly complex time-frequency separation algorithms. Therefore, the technical solution of this application further performs adaptive environmental noise suppression on the aligned original matrix to obtain a denoised geological signal. Specifically, it extracts vibration components from the aligned original matrix and estimates the background noise power spectral density using fast Fourier transform. Then, it dynamically calculates a smoothing factor matching the current noise intensity using nonlinear mapping, and uses this smoothing factor to perform adaptive recursive filtering on the fluorescence pulse count components. Through the above processing, the filtering intensity can be synchronously adjusted with the real-time changes in the vibration environment of the mining site. When the noise increases, the smoothing intensity is increased to suppress the interference, and when the noise decreases, the smoothing intensity is reduced to retain the transient characteristics at the grade change point. Thus, real-time fidelity noise reduction of geological signals can be achieved under the limited computing power on the edge side.

[0020] Figure 3 This is a block diagram of the adaptive noise suppression module in the real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application. Figure 3 As shown, the adaptive noise suppression module 120 includes: a vibration feature analysis unit 121, used to perform vibration feature analysis and power spectrum estimation on the vibration components extracted from the aligned original matrix through fast Fourier transform to obtain the background noise power spectral density; a smoothing factor mapping unit 122, used to perform smoothing factor dynamic mapping estimation on the background noise power spectral density to obtain a smoothing factor; and a temporal smoothing unit 123, used to perform temporal smoothing processing on the fluorescence pulse count components in the aligned original matrix based on the smoothing factor to obtain a denoised geological signal.

[0021] In the aforementioned real-time processing system 100 for edge nodes of fluorite mine geological data, the vibration feature analysis unit 121 is used to perform vibration feature analysis and power spectrum estimation on the vibration components extracted from the aligned original matrix using Fast Fourier Transform to obtain the background noise power spectral density. It should be noted that, due to the non-stationary nature of mechanical vibration noise at fluorite mine mining sites, its frequency distribution and energy intensity continuously fluctuate with changes in the hardness of the rock layer being cut by the drill bit. The subsequent adaptive filtering stage requires a quantitative index that can reflect the current environmental interference level in real time to dynamically adjust the filtering intensity. Based on this, the technical solution of this application further uses Fast Fourier Transform to perform vibration feature analysis and power spectrum estimation on the vibration components extracted from the aligned original matrix to obtain the background noise power spectral density. Through the above processing, the time-domain vibration signal can be converted to the frequency domain and the energy distribution of mechanical noise in the current mining environment can be quantified, providing a real-time updated noise intensity reference benchmark for the subsequent dynamic calculation of the smoothing factor.

[0022] More specifically, in a concrete example of this application, the data columns corresponding to the triaxial vibration acceleration components are first extracted from the aligned original matrix by column index, obtaining vibration time series along the drilling direction, the radial horizontal direction, and the radial vertical direction, respectively. Then, a modulus synthesis operation is performed on the triaxial vibration time series, i.e., the arithmetic square root of the sum of the squares of the three axial acceleration values ​​at the same sampling time is calculated, merging the three-dimensional vibration information into a one-dimensional synthetic vibration amplitude sequence to eliminate the influence of vibration directionality on subsequent frequency domain analysis. Next, a Hanning window function is applied to the synthetic vibration amplitude sequence for windowing processing to suppress the spectral leakage effect caused by finite length truncation, making the frequency domain analysis results more accurately reflect the true frequency components of the vibration signal. After completing the windowing processing, a fast Fourier transform is performed on the windowed vibration sequence to convert it from the time domain to the frequency domain, obtaining the complex spectral coefficients corresponding to each frequency component. Then, the square of the modulus of the spectral coefficients is taken and divided by the sequence length to calculate the power value of the vibration signal at each discrete frequency point, forming the power spectrum estimation result. Finally, within the preset main frequency band of mechanical noise, the above power spectrum estimation results are integrated by frequency band integration, and the power values ​​at each frequency point in the frequency band are summed to obtain the background noise power spectral density that can characterize the overall vibration interference intensity of the current mining environment.

[0023] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the smoothing factor mapping unit 122 is used to dynamically map and estimate the smoothing factor of the background noise power spectral density to obtain the smoothing factor. It should be noted that since the background noise power spectral density is a continuous numerical quantity reflecting the intensity of current environmental vibration interference, while the control parameter required for the subsequent recursive filtering stage is a smoothing factor with a value between 0 and 1, a reasonable nonlinear correspondence needs to be established between the two so that the filtering intensity can automatically adjust with changes in noise level. Based on this, the technical solution of this application further performs dynamic mapping and estimation of the smoothing factor on the background noise power spectral density to obtain the smoothing factor. Through the above processing, changes in noise intensity can be converted into synchronous adjustments of the filter response weights in real time. When vibration interference increases, the confidence level of the current observation value is automatically reduced to enhance the smoothing effect; when vibration interference decreases, the confidence level of the current observation value is automatically increased to preserve transient changes in the geological signal.

[0024] More specifically, in a concrete example of this application, two pre-calibrated mapping control parameters are first read from the non-volatile memory of the edge node: a steepness adjustment coefficient and a noise power reference threshold. The steepness adjustment coefficient controls the response sensitivity of the smoothing factor to changes in noise power. The noise power reference threshold is a reference value for the average background noise power under normal drilling conditions in a fluorite mine, obtained through on-site no-load calibration tests before the edge node deployment. Then, the background noise power spectral density calculated at the current moment is subtracted from the noise power reference threshold to obtain the difference representing the current noise deviation from the normal level. This difference is then multiplied by the steepness adjustment coefficient and substituted into an S-shaped nonlinear mapping function for calculation. This involves exponentiation of the product with the natural constant as the base, adding 1, and then taking its reciprocal, thus mapping the noise power deviation to a continuous value between 0 and 1. When the background noise power spectral density is much higher than the reference threshold, the output of the mapping function approaches 0, meaning that the absorption of the current sampled value is minimized during subsequent filtering. The filter mainly relies on historical estimates for its output, thereby achieving sufficient smoothing of the observation data under strong noise conditions. When the background noise power spectral density is below or close to the reference threshold, the output of the mapping function approaches 1. The filter assigns a high confidence weight to the current sampled value, allowing for timely tracking of grade changes in the geological signal. After the above mapping operation, the output value is the smoothing factor at the current moment.

[0025] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the temporal smoothing unit 123 is used to perform temporal smoothing processing on the fluorescence pulse count components in the aligned original matrix based on a smoothing factor to obtain a denoised geological signal. It should be noted that, given that a smoothing factor capable of reflecting the current vibration interference level of the mining environment has been obtained through the aforementioned steps, and since the fluorescence pulse count components in the aligned original matrix, as the core observation data for fluorite grade determination, are still superimposed with broadband noise interference introduced by mechanical drilling, it is necessary to use this smoothing factor to perform dynamic weighted temporal filtering on the fluorescence pulse count components to suppress noise while preserving as much as possible the transient pulse characteristics reflecting spatial jumps in vein grade. Based on this, the technical solution of this application further performs temporal smoothing processing on the fluorescence pulse count components in the aligned original matrix based on a smoothing factor to obtain a denoised geological signal, that is, using the smoothing factor as a dynamic weighting coefficient between the current observation value and the historical estimate, and performing a first-order adaptive recursive filtering operation on the fluorescence pulse count sequence. Through the above processing, the reliance on historical estimates can be increased during periods of strong vibration and noise to fully suppress interference, while the response to current observations can be improved during periods of mild vibration and noise to preserve grade change information at the interface between the fluorite ore body and the surrounding rock. Thus, under the constraint of limited computing power at edge nodes, a denoised geological signal with improved signal-to-noise ratio and preserved geological features is output.

[0026] Figure 4 This is a block diagram of a time-series smoothing unit in a real-time processing system for edge nodes of fluorite mineral geological data according to a preferred embodiment of this application. Figure 4 As shown, the time-series smoothing unit 123 includes: a coupling estimation subunit 1231, used to perform vibration-signal time-domain derivative coupling estimation on the vibration components extracted from the aligned original matrix and the fluorescence pulse count to obtain a characteristic coupling factor; a gain weight correction subunit 1232, used to perform dynamic phase gain weight correction based on the characteristic coupling factor and combined with the smoothing factor to obtain a corrected gain weight; and a phase decoupling filtering subunit 1233, used to perform phase decoupling filtering on the fluorescence pulse count in the aligned original matrix and the historical denoised signal based on the corrected gain weight and by introducing a phase compensation term to obtain a denoised geological signal.

[0027] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the coupling estimation subunit 1231 is used to perform vibration-signal time-domain derivative coupling estimation on the vibration components extracted from the aligned original matrix and the fluorescence pulse count to obtain the characteristic coupling factor. It should be noted that, given that the aforementioned first-order recursive filtering mechanism passively adjusts the smoothing factor based solely on the background noise power spectral density, it fails to consider the nonlinear physical coupling relationship between mechanical drilling vibration and X-ray fluorescence pulse detection. Fluorite ore is brittle and unevenly distributed in the vein. When the drill bit cuts rock layers of different hardness, it induces non-stationary harmonic vibrations. Such vibrations cause micron-level periodic displacements between the detector probe and the ore surface, forming a systematic intervention similar to amplitude modulation and frequency modulation on the pulse count signal. The existing mechanism cannot identify from the underlying logic whether the signal fluctuations originate from the real spatial jump of the ore grade or from the interference modulation of the mechanical vibration phase. Based on this, the technical solution of this application further performs vibration-signal time-domain derivative coupling estimation on the vibration components extracted from the aligned original matrix and the fluorescence pulse count to obtain a characteristic coupling factor. By quantifying the transient correlation between vibration changes and pulse signal fluctuations, the physical homology of current geological signal fluctuations is identified. Through the above processing, nonlinear fluctuation characteristics caused by mechanical motion can be extracted from the observation sequence, providing quantitative decision support for subsequent accurate confidence weight allocation.

[0028] More specifically, in a concrete example of this application, the triaxial vibration acceleration components are first extracted from the aligned original matrix by column index. A magnitude synthesis operation is then performed on the three axial acceleration values ​​at the same sampling time, i.e., the arithmetic square root of the sum of the squares of the three axial accelerations is calculated to obtain the total vibration vector. Simultaneously, the X-ray fluorescence intensity pulse count at the current sampling time is extracted from the aligned original matrix. Subsequently, the first-order time-domain difference is calculated for the total vibration vector sequence and the pulse count sequence, i.e., the sampling value at the current time is subtracted from the sampling value at the previous time, to obtain the time-domain derivative of the X-ray fluorescence intensity and the time-domain derivative of the synthesized vibration vector. These two derivatives reflect the instantaneous rate of change of the fluorescence pulse signal and the mechanical vibration within adjacent sampling periods, respectively. Furthermore, a feature coupling factor is constructed using the cosine similarity between the two time-domain derivatives, with the specific calculation formula shown below: in, This represents the generated characteristic coupling factor, used to reflect the correlation strength between vibration changes and signal fluctuations. The time-domain derivative representing the intensity of X-ray fluorescence. The time-domain derivative of the composite vibration vector is represented. The modulus of the time-domain derivative of X-ray fluorescence intensity. This represents the magnitude of the time-domain derivative of the synthesized vibration vector. This represents a small positive constant used to ensure the stability of numerical calculations. When the characteristic coupling factor approaches 1, it indicates that the instantaneous fluctuation direction of the fluorescence pulse signal is highly consistent with the instantaneous change direction of the mechanical vibration, meaning that the current signal fluctuation is most likely due to physical interference from mechanical vibration rather than a real change in ore grade. When the characteristic coupling factor approaches 0, it indicates that there is no synchronous correlation between the two, and the current signal fluctuation is more likely to reflect the real spatial jump in the grade of the fluorite vein. In fluorite drilling operations, when the drill bit enters the fluorite-rich ore body from the surrounding rock layer, the fluorescence pulse count will increase stepwise due to the sudden increase in the calcium fluoride content in the ore. If the drilling speed and rock hardness do not change drastically at this time, and the rate of change of vibration acceleration remains stable, the time domain derivatives of the two are inconsistent, the characteristic coupling factor output is low, and the subsequent filtering stage will retain this grade change information. Conversely, when the drill bit encounters a hard interlayer causing a sudden increase in vibration acceleration, and the detector generates a synchronous jump in pulse counting due to probe displacement, the time-domain derivatives of the two change in the same direction, and the characteristic coupling factor output is high. The subsequent filtering stage will suppress this false fluctuation accordingly.

[0029] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the gain weight correction subunit 1232 is used to dynamically correct the phase gain weight based on the characteristic coupling factor and combined with the smoothing factor to obtain the corrected gain weight. It should be noted that, given that the aforementioned steps have already obtained the characteristic coupling factor capable of quantifying the transient correlation between mechanical vibration and fluorescence pulse signals, while the original smoothing factor only reflects the overall power level of environmental noise and has not yet incorporated the directional information of the physical coupling between vibration and signal, when vibration changes and signal fluctuations exhibit highly synchronous coupling, the observed signal is more likely to be subject to physical interference, and its trust weight should be rapidly compressed. Based on this, the technical solution of this application further uses the characteristic coupling factor and combined with the smoothing factor to dynamically correct the phase gain weight to obtain the corrected gain weight, introducing the characteristic coupling factor as an independent variable into the nonlinear mapping model to dynamically scale the original smoothing factor. Through the above processing, an active defense mechanism based on interferometric morphology can be established, avoiding false characteristic peaks caused by vibration by suppressing the gain intensity during periods of active vibration, making the processing logic sensitive to real grade changes while remaining robust to environmental phase interference.

[0030] More specifically, in a concrete example of this application, a pre-calibrated sensitivity adjustment coefficient is first read from the non-volatile memory of the edge node. This coefficient controls the response sensitivity of the feature coupling factor to the compression amplitude of the gain weight. Then, the absolute value of the feature coupling factor output from the previous step is taken to eliminate the positive and negative differences in the coupling direction, retaining only the amplitude information of the coupling strength. Next, this absolute value is multiplied by the sensitivity adjustment coefficient and then input into the hyperbolic tangent function for nonlinear mapping. The output range of the hyperbolic tangent function is between 0 and 1; the output approaches 1 when the input value is large and approaches 0 when the input value is small. Subtracting 1 from the mapping result yields a scaling factor that is inversely proportional to the coupling strength; that is, the higher the coupling strength, the smaller the scaling factor. Finally, this scaling factor is multiplied by the original smoothing factor to obtain the corrected gain weight. The specific calculation formula is shown below: in, The generated corrected gain weights determine the degree to which the filter absorbs the current sampled value. This represents the original smoothing factor. This represents the feature coupling factor of the input. Represents the absolute value of the characteristic coupling factor. This represents the preset sensitivity adjustment coefficient for the degree of phase interference. This represents the hyperbolic tangent function. When the absolute value of the characteristic coupling factor is high, the output of the hyperbolic tangent function approaches 1, the scaling factor approaches 0, and the correction gain weight is compressed to a level close to 0. The filter relies almost entirely on historical estimates for its output, thus excluding spurious fluctuations caused by vibration coupling from the filtering results. When the absolute value of the characteristic coupling factor is low, the output of the hyperbolic tangent function approaches zero, the scaling factor approaches 1, and the correction gain weight is approximately equal to the original smoothing factor. The filter responds normally to the grade change information in the current observation. In fluorite drilling operations, when the drill bit encounters a hard calcite interlayer, causing a sudden increase in vibration acceleration, and simultaneously the detector generates a synchronous jump in fluorescence pulse count due to periodic probe displacement, the characteristic coupling factor outputs a high value, the correction gain weight is rapidly compressed, and the filter automatically masks the observation data for this period to prevent spurious peaks from entering the subsequent grade prediction stage. When the drill bit smoothly transitions from the surrounding rock layer to the fluorite-rich ore body, the fluorescence pulse count increases stepwise due to the increased calcium fluoride content in the ore, but the vibration acceleration does not change synchronously. The characteristic coupling factor outputs a low value, the correction gain weight remains at a normal level, and the filter promptly tracks the grade change information and transmits it to the output signal.

[0031] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the phase decoupling filter subunit 1233 is used to perform phase decoupling filtering on the fluorescence pulse count and historical denoised signal in the aligned original matrix based on corrected gain weights and the introduction of a phase compensation term to obtain a denoised geological signal. It should be noted that, given that the aforementioned steps have already obtained corrected gain weights that incorporate vibration-signal coupling information, traditional first-order recursive filtering directly uses historical estimates as the prediction benchmark when performing residual feedback, without considering the influence of the periodic phase characteristics of mechanical drilling vibration on the time-domain offset of the historical reference signal. This results in an inherent phase lag in the filtered output, and minute geological feature details masked by environmental noise cannot be effectively restored. Based on this, the technical solution of this application further performs phase decoupling filtering on the fluorescence pulse count and historical denoised signal in the aligned original matrix based on corrected gain weights and the introduction of a phase compensation term to obtain a denoised geological signal. Through feedback compensation based on the laws of physical motion, a more realistic vein energy spectrum profile is recovered at the edge side. Through the above processing, real-time and high-fidelity restoration of geological signals can be achieved. By using residual compensation, minute geological feature details that are obscured by environmental noise can be recovered, and the inherent phase delay of traditional filtering algorithms can be offset.

[0032] More specifically, in a concrete example of this application, the X-ray fluorescence intensity pulse count at the current sampling moment is first extracted from the aligned original matrix as the current observation value. Simultaneously, the historical denoised signal output at the previous sampling moment is read from the circular buffer of the edge nodes as the prediction benchmark. Subsequently, the vibration phase angle at the current moment is derived in real time based on the rotation frequency of the drilling rig spindle, and a cosine operation is performed on this phase angle to generate a phase correction compensation term. The physical meaning of this compensation term is that the rotational motion of the drilling rig spindle has periodic characteristics, and the micron-level displacement between the detector probe and the ore surface varies with the spindle phase in a cosine manner. Therefore, the residual phase shift in the historical denoised signal can be aligned and corrected using this cosine term. Then, the historical denoised signal is multiplied by the phase correction compensation term to obtain the predicted value after phase-level alignment. This predicted value is then subtracted from the current observation value to obtain the residual signal reflecting the difference between the actual grade change and the residual noise. Finally, this residual signal is multiplied by the correction gain weight output in the aforementioned steps and superimposed on the historical denoised signal to complete one residual feedback recursive synthesis operation, outputting the denoised geological signal at the current moment. The specific calculation formula is as follows: in, This represents the final denoised geological signal. This represents the historical denoised signal stored in the edge node cache. This represents the input correction gain weight. This represents the original pulse count of X-ray fluorescence intensity. This represents the phase correction and compensation term generated based on the vibration phase law. This represents the current vibration phase angle derived in real time based on the drilling rig spindle rotation frequency. When the correction gain weight is compressed to near zero due to the high coupling between vibration and signal, the contribution of the residual term to the output is suppressed, and the denoised geological signal mainly follows the trend of historical estimates, thus shielding spurious fluctuations caused by vibration interference. When the correction gain weight remains at a normal level, the grade change information contained in the residual term is effectively absorbed into the output, while the phase correction compensation term aligns the historical reference signal at the phase level, eliminating signal contour distortion caused by phase delay in traditional recursive filtering. During fluorite drilling operations, when the drill bit traverses the boundary zone between the ore body and the surrounding rock at a stable rotation speed, the fluorescence pulse count exhibits a step change within several sampling cycles. At this time, the characteristic coupling factor outputs a low value, the correction gain weight remains normal, and the residual feedback mechanism incorporates the grade jump information into the output signal in a timely manner. Meanwhile, the phase compensation term synchronously corrects the phase shift of the historical signal caused by the rotation of the drilling rig, so that the output denoised geological signal accurately corresponds to the real spatial position of the vein boundary in the time domain. This avoids the problem of the vein boundary response lagging behind the actual position in traditional filtering methods, providing accurate and high-fidelity data input for subsequent fluorite feature saliency extraction and edge grade prediction.

[0033] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the feature saliency extraction module 130 is used to extract fluorite feature saliency from the denoised geological signal to obtain a salient mineral feature vector. It should be noted that, given that the denoised geological signal obtained after the aforementioned adaptive environmental noise suppression processing still contains mixed response information from various mineral components, including the energy spectrum characteristics of calcium and fluorine elements directly related to fluorite grade determination, as well as spectral interference components from associated minerals such as calcite and quartz, directly feeding all signal dimensions into the subsequent quantization neural network for inference would not only increase the computational burden under the limited computing power of the edge nodes but also reduce the accuracy of grade prediction due to the introduction of irrelevant features. Furthermore, existing data dimensionality reduction techniques are mostly based on mathematical statistical methods and do not consider the unique energy spectrum distribution characteristics of fluorite, resulting in the loss of key information related to grade during the compression process. Based on this, the technical solution of this application further extracts saliency of fluorite features from the denoised geological signal to obtain salient mineral feature vectors. Specifically, based on the energy spectrum characteristics of calcium and fluorine elements in fluorite minerals, signal slices corresponding to sensitive energy channels are extracted from the denoised geological signal. After saliency-weighted modeling, these slices are vectorized and mapped to the spectral components to construct a compact feature representation for fluorite grade determination. Through the above processing, high-dimensional geological signals can be compressed into low-dimensional feature vectors strongly correlated with fluorite grade in a domain knowledge-driven manner at the edge. This reduces the computational load of subsequent inference while retaining the mineralogical feature information most sensitive to grade changes, providing high signal-to-noise ratio and compact input data for grade prediction by the edge-end quantization neural network.

[0034] Figure 5 This is a block diagram of the feature saliency extraction module in the real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application. Figure 5 As shown, the feature saliency extraction module 130 includes: a sensitive energy channel slicing unit 131, used to extract signal slices corresponding to the sensitive energy channels of calcium and fluorine elements from the denoised geological signal based on the energy spectrum characteristics of fluorite minerals to obtain a feature frequency domain set; a saliency weighting unit 132, used to perform saliency weighting modeling on the feature frequency domain set to obtain a saliency index; and a vectorization mapping unit 133, used to perform vectorization mapping processing on the saliency index and the spectral components in the denoised geological signal to obtain a saliency mineral feature vector.

[0035] In the aforementioned real-time processing system 100 for edge nodes of fluorite mineral geological data, the sensitive energy channel slicing unit 131 is used to extract signal slices corresponding to the sensitive energy channels of calcium and fluorine elements from the denoised geological signal based on the energy spectrum characteristics of fluorite minerals to obtain a feature frequency domain set. It should be noted that since the denoised geological signal contains multiple elemental response information collected by the X-ray fluorescence detector across the full energy spectrum, and the chemical composition of fluorite is calcium fluoride, its grade determination is directly related only to the content of calcium and fluorine elements, a large amount of associated element energy channel data unrelated to fluorite grade in the full energy spectrum will interfere with subsequent feature modeling and increase the computational burden on edge nodes. Based on this, the technical solution of this application further extracts signal slices corresponding to the sensitive energy channels of calcium and fluorine elements from the denoised geological signal based on the energy spectrum characteristics of fluorite minerals to obtain a feature frequency domain set. Through the above processing, the data range for subsequent processing can be precisely limited to a specific energy spectrum interval strongly correlated with fluorite grade, eliminating the interference of associated mineral energy spectrum responses and reducing the dimensionality of the feature space.

[0036] More specifically, in a particular example of this application, a pre-configured fluorite mineral sensitive channel parameter table is first read from the non-volatile memory of the edge node. This parameter table records the energy ranges corresponding to the characteristic X-ray fluorescence peaks of calcium and the energy ranges corresponding to the characteristic responses of fluorine. These energy ranges are obtained before the edge node is deployed by performing energy spectrum calibration experiments on standard fluorite samples of known grades. Subsequently, the energy spectrum dimension of the X-ray fluorescence pulse count component is located from the denoised geological signal. According to the upper and lower limit indices of the calcium sensitive channel recorded in the parameter table, the pulse count values ​​corresponding to each discrete channel within this energy range are extracted to form a calcium channel signal slice. Simultaneously, according to the upper and lower limit indices of the fluorine sensitive channel, the pulse count values ​​within the corresponding energy range are extracted to form a fluorine channel signal slice. At fluorite mining sites, calcium-bearing minerals such as calcite are often associated with the ore body. The energy spectrum peaks of calcite partially overlap with those of fluorite. Therefore, when extracting calcium energy channel slices, the extraction range is precisely limited to the full width at half maximum (FWHM) of the characteristic calcium peaks in fluorite to reduce energy spectrum crosstalk from associated minerals like calcite. Furthermore, the calcium and fluorine energy channel signal slices are spliced ​​together according to their energy channel indexes, and energy accumulation is performed on the pulse count values ​​of each channel—that is, the count values ​​of adjacent channels within the same sensitive energy channel interval are summed to improve the statistical reliability of the signal. After the above slice extraction and energy accumulation, the output dataset is the characteristic frequency domain set, which contains only the calcium and fluorine sensitive energy channel response information directly corresponding to the chemical composition of fluorite.

[0037] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the significance weighting unit 132 is used to perform significance weighting modeling on the feature frequency domain set to obtain a significance index. It should be noted that, since the feature frequency domain set contains multiple discrete pulse count values ​​from the sensitive channels of calcium and fluorine elements, the contribution of each channel to the determination of fluorite grade is not the same, and the absolute magnitude of the original count values ​​lacks cross-time comparability due to factors such as detector sensitivity and ore surface distance. Therefore, it is necessary to establish a quantitative index that can comprehensively reflect the degree to which the current detection data deviates from the geological background level. Based on this, the technical solution of this application further performs significance weighting modeling on the feature frequency domain set to obtain a significance index. Through the above processing, the discrete response information of multiple sensitive channels can be fused into a scalarized significance metric, providing a unified criterion reflecting the intensity of mineral grade mutations for subsequent vectorized mapping.

[0038] More specifically, in a concrete example of this application, two pre-calibrated sets of reference parameters are first read from the non-volatile memory of the edge node: the background mean and the background standard deviation. The background mean is the average level of pulse counts for each sensitive channel under the geological background conditions of the surrounding rock in the mining area, and the background standard deviation is the fluctuation amplitude of the background signal. These reference parameters are obtained by performing multiple blank scans on the surrounding rock section of the mining area before the edge node is deployed. Simultaneously, a pre-configured feature contribution weight table is read from the memory. This weight table assigns different weighting coefficients to each sensitive channel for calcium and fluorine elements to correct for sensitivity deviations caused by differences in detector response efficiency among different channels. The fluorine element characteristic channel is given a relatively high weight value because it has stronger specificity in fluorite grade determination. Subsequently, a standardized deviation calculation is performed on each sensitive channel in the characteristic frequency domain set. This involves subtracting the corresponding background mean from the pulse count value of the current channel, then dividing by the sum of the background standard deviation and a small positive constant to obtain the normalized deviation of that channel relative to the geological background. The small positive constant is introduced to prevent numerical overflow when the background standard deviation approaches zero. Next, the normalized deviation of each channel is multiplied by its corresponding feature contribution weight to obtain a weighted deviation. Finally, the weighted deviations of all sensitive channels are summed to aggregate the response information of multiple channels into a single scalar value, namely the significance index. When the detection point is located inside a fluorite-rich ore body, the pulse count values ​​of the calcium and fluorine sensitive channels are higher than the background level of the surrounding rock. The normalized deviations of each channel are positive and have a large amplitude, resulting in a high positive significance index after weighted summation. When the detection point is located in the surrounding rock or low-grade ore area, the pulse count values ​​of each sensitive channel are close to the background mean, the normalized deviation is close to 0, and the significance index output is close to zero. When the detection point is located at the boundary between the fluorite ore body and the surrounding rock, the calcium channel may show a certain positive deviation due to the presence of associated calcite, but the deviation of the fluorine channel is low. Since the fluorine channel is given a higher characteristic contribution weight, the final significance index is still at a low to medium level, thus effectively distinguishing the differences in the energy spectrum response between the fluorite ore body and calcium-bearing associated minerals.

[0039] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the vectorization mapping unit 133 is used to perform vectorization mapping processing on the saliency index and the spectral components in the denoised geological signal to obtain a significant mineral feature vector. It should be noted that since the saliency index output by the aforementioned steps is a scalar value reflecting the degree of deviation of the fluorite mineral energy spectrum, while the denoised geological signal also contains mineral surface property information carried by the near-infrared spectral reflectance component, the two represent the characteristics of the fluorite ore body from two different physical dimensions: energy spectrum response and spectral reflectance. If grade prediction is based solely on information from a single dimension, misjudgments are likely to occur under conditions of complex ore body composition or strong interference from associated minerals. Therefore, it is necessary to fuse the two complementary feature information into a unified vectorized representation for subsequent quantization neural network inference. Based on this, the technical solution of this application further performs vectorization mapping processing on the saliency index and the spectral components in the denoised geological signal to obtain a significant mineral feature vector. Through the above processing, the grade mutation intensity information in the energy spectrum dimension and the mineral surface reflection characteristics in the spectral dimension can be fused into a feature vector with a compact dimension and complementary information, providing input data for the edge-end quantization neural network that takes into account both energy spectrum criteria and spectral criteria.

[0040] More specifically, in a particular example of this application, a nonlinear mapping operation is first applied to the significance index output from the aforementioned steps. This index is then substituted into the hyperbolic tangent function for range compression, ensuring the mapped value falls within the range of -1 to +1. This enhances the distinction between high-grade and low-grade features and suppresses extreme numerical fluctuations caused by occasional detector anomalies. Subsequently, near-infrared spectral reflectance components are extracted from the denoised geological signal by column index. These components contain reflectance values ​​of the ore surface in multiple near-infrared bands. Different minerals exhibit different absorption and reflection characteristics in the near-infrared bands due to differences in crystal structure and chemical bond vibration modes. The reflectance curve shape of fluorite in a specific near-infrared band shows identifiable differences from associated minerals such as calcite and quartz. Furthermore, the significance index, after nonlinear mapping, is used as a scalar multiplier to perform element-wise scaling on the values ​​of each band in the extracted near-infrared spectral reflectance component. This involves multiplying the reflectance value of each band by the mapped significance index, amplifying spectral features at sampling points with high fluorite grades and suppressing them at sampling points with low grades, thus achieving attention modulation of spectral features by the energy spectrum criterion. Finally, the scaled spectral components and the mapped significance index are concatenated in dimensional order to form a one-dimensional feature vector. This vector is then normalized using the L2 norm to scale its magnitude to a unit length, eliminating the scale inconsistency caused by differences in absolute signal intensity between different sampling points. The output is the significant mineral feature vector. In fluorite drilling operations, when the detection point is located inside a fluorite-rich ore body, the significance index outputs a high positive value, approaching 1 after nonlinear mapping. The unique near-infrared reflectance characteristics of fluorite in the spectral components are fully preserved and amplified, resulting in a significant mineral feature vector exhibiting a high numerical response in the corresponding dimension. When the detection point is located in the surrounding rock area, the significance index approaches 0, the spectral components are suppressed as a whole, and the values ​​of each dimension of the generated feature vector are at a low level, which enables the subsequent quantization neural network to quickly distinguish the ore body from the surrounding rock based on the overall amplitude distribution of the feature vector.

[0041] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the edge inference module 140 is used to input significant mineral feature vectors into a pre-deployed quantized neural network for forward inference to obtain the grade prediction result at the edge. It should be noted that, given that the aforementioned steps have compressed the high-dimensional geological signal into significant mineral feature vectors strongly correlated with fluorite grade, but these vectors are only numerical representations of mineralogical characteristics and have not yet been converted into grade values ​​that can directly guide mining decisions, and given the extremely high timeliness requirements for grade determination at fluorite mining sites, and the bandwidth limitations of the mining area's wireless communication links preventing the full feature data from being transmitted back to a remote cloud server for centralized inference, the mapping calculation from feature vectors to grade estimates needs to be completed locally at the edge node. Simultaneously, the computing resources of the edge node are limited by the microcontroller-level hardware platform, and the sensitivity of the deep learning model for identifying low-grade fluorite ore tends to decrease after compression to the edge. Therefore, a lightweight neural network with quantization processing is needed to achieve a balance between computational constraints and prediction accuracy. Based on this, the technical solution of this application further inputs significant mineral feature vectors into a pre-deployed quantization neural network for forward inference to obtain edge-end grade prediction results. Specifically, after normalizing the feature vectors, they are fed into the INT8 quantization model to complete interval probability deduction, and then the grade value is output through likelihood estimation. Through the above processing, real-time prediction of fluorite grade can be completed with millisecond-level response speed under the limited computing resources of edge nodes, providing grade fluctuation criteria for subsequent dynamic data compression strategies, and simultaneously providing on-site mining personnel with immediate grade reference information.

[0042] Specifically, the quantized neural network requires two preprocessing stages—offline training and quantization compression—before being deployed to edge nodes. In the offline training stage, a large amount of geological sample data with labeled grades is first collected from the historical exploration database of the fluorite mining area. Each sample data contains feature inputs with the same dimensional structure as the significant mineral feature vector and a true value label of calcium fluoride content determined by chemical analysis. These sample data are then divided into training and validation sets according to a preset ratio. Subsequently, a multi-layer fully connected neural network model is constructed on an offline server with floating-point computing capabilities. Using the training set samples as input and the cross-entropy loss function of grade as the optimization objective, the weight matrix and bias vector of each layer of the network are iteratively updated through the backpropagation algorithm until the grade classification accuracy on the validation set converges to a preset threshold. During the quantization and compression phase, post-training quantization is performed on the trained floating-point precision model. The numerical distribution range of activation values ​​for each layer is statistically analyzed using the calibration dataset. Based on this, the quantization scaling factor and zero-point offset parameter for each layer are calculated. The 32-bit floating-point format weight matrix and bias vector are converted to 8-bit integer format. This process compresses the model's storage volume and inference computation while keeping the loss of grade prediction accuracy within an acceptable range. Finally, the quantized model weight file, along with the quantization parameters of each layer, feature mean vector, and feature standard deviation vector, is written into the non-volatile memory of the edge nodes, completing the edge-side deployment of the model.

[0043] In a preferred embodiment, the quantization neural network adopts a fully connected network structure consisting of an input layer, three hidden layers, and an output layer cascaded sequentially. The number of neurons in the input layer is consistent with the dimension of the salient mineral feature vector, and it is used to receive the normalized feature vector. The first hidden layer contains 64 neurons, responsible for the initial nonlinear combination and abstract expression of the input features, and the activation function is ReLU. The second hidden layer contains 32 neurons, responsible for further feature compression and higher-order correlation modeling of the output of the first hidden layer, and the activation function is also ReLU. The third hidden layer contains 16 neurons, responsible for mapping higher-order features to a compact semantic space directly related to the grade determination, and the activation function is still ReLU. The number of neurons in the output layer is consistent with the preset total number of grade categories, and no activation function is set. It directly outputs the original score value corresponding to each grade, which is subsequently converted into a grade probability distribution through Softmax operation. The number of layers and neurons in each layer of the above network structure have been specifically optimized to address the computing power constraints of edge nodes. Under the INT8 quantization format, the storage volume of all weight parameters does not exceed tens of kilobytes, and the total number of integer multiplication and addition operations in a single forward inference is controlled to the order of thousands. It can complete a complete quality inference calculation with millisecond latency on a single-chip microcomputer-level edge hardware platform.

[0044] Figure 6This is a block diagram of the edge inference module in the real-time processing system for edge nodes of fluorite mineral geological data according to an embodiment of this application. Figure 6 As shown, the edge reasoning module 140 includes: a normalization unit 141, used to normalize the significant mineral feature vector to obtain a normalized feature vector; a probability inference unit 142, used to perform interval probability inference on the normalized feature vector using a quantized neural network engine to obtain a grade probability distribution; and a likelihood estimation unit 143, used to perform likelihood estimation on the grade probability distribution to obtain the edge grade prediction result.

[0045] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the normalization unit 141 is used to normalize the significant mineral feature vector to obtain a normalized feature vector. It should be noted that since the values ​​of each dimension in the significant mineral feature vector originate from the fusion mapping of different physical quantities, including the near-infrared spectral reflectance component scaled by the saliency index and the energy spectrum deviation scalar, there are differences in the numerical range and dimensions between each dimension. Furthermore, the input data used by the subsequent quantization neural network during the training phase has undergone specific standardization processing. If the distribution of the input data in the inference phase is inconsistent with that in the training phase, it will cause the activation values ​​of each layer of the network to deviate from the expected working range, reducing the accuracy of grade prediction. This is especially true in the INT8 quantization model, where the numerical range deviation of the input data will be further amplified by the quantization truncation effect. Based on this, the technical solution of this application further normalizes the significant mineral feature vector to obtain a normalized feature vector. Through the above processing, the numerical values ​​of each dimension of the feature vector can be uniformly mapped to a numerical range consistent with that of the training stage of the quantized neural network, ensuring that the input distribution in the inference stage is aligned with that in the training stage, and enabling the weight parameters of each layer of the quantized model to work normally within the expected numerical range.

[0046] More specifically, in a concrete example of this application, pre-stored calibration parameter pairs, including feature mean vectors and feature standard deviation vectors, are first read from the non-volatile memory of the edge nodes. These parameters are obtained during the model training phase by statistically calculating the salient mineral feature vectors of all samples in the training dataset, and are written to the edge node memory along with the weight file of the quantized neural network during model deployment. Subsequently, Z-Score standardization is performed on the salient mineral feature vectors, i.e., subtracting the feature mean at the corresponding position from each dimension of the vector, and then dividing by the sum of the feature standard deviation at the corresponding position and a small positive constant. The small positive constant is introduced to prevent numerical overflow when the feature standard deviation approaches 0. After the above standardization operation, the numerical distribution of each dimension is adjusted to a standard normal distribution with a mean of 0 and a standard deviation of 1. Furthermore, a numerical interval pruning operation is performed on the standardized vector, restricting the numerical values ​​of each dimension to the range of -3 to +3. Extreme values ​​exceeding this range are truncated to the boundary values ​​to avoid occasional outlier sampling points generating input values ​​that exceed the representation range of the INT8 quantization model. Finally, a linear scaling operation is performed on the clipped vector, mapping it from the interval of -3 to +3 to the interval of -1 to +1, so that the numerical range of the input data is strictly consistent with the input normalization interval used in the training stage of the quantized neural network. The output result is the normalized feature vector.

[0047] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the probability inference unit 142 is used to perform interval probability inference on the normalized feature vector using a quantized neural network engine to obtain the grade probability distribution. It should be noted that since the normalized feature vector is a standardized low-dimensional numerical sequence, the mapping relationship between its various dimensions and fluorite grade has highly nonlinear characteristics. It is impossible to accurately infer the grade level using simple linear regression or threshold determination methods. Furthermore, the computational resources of the edge node are limited by the microcontroller-level hardware platform, making it impossible to run floating-point precision deep learning models. Therefore, a lightweight neural network with INT8 quantization compression is needed to complete the inference calculation within the integer domain. Based on this, the technical solution of this application further utilizes a quantized neural network engine to perform interval probability inference on the normalized feature vector to obtain the grade probability distribution. Through the above processing, a nonlinear mapping from the feature space to the grade probability space can be completed under the limited integer computational capabilities of the edge node, outputting the confidence probability that the current detection point belongs to each preset grade level, providing a probabilistic decision-making basis for subsequent grade value estimation.

[0048] More specifically, in a concrete example of this application, the normalized feature vector is first loaded into the quantized neural network inference engine pre-deployed at the edge node. During the edge node's power-on initialization phase, this engine loads the INT8-quantized model weight file from non-volatile memory. The model structure is a multi-layer fully connected network, with the weight matrix and bias vector of each layer stored in 8-bit integer format, along with corresponding quantization scaling factors and zero-point offset parameters. Subsequently, the inference engine converts the normalized feature vector from the floating-point domain to the 8-bit integer domain according to the quantization scaling factor, obtaining the quantized input tensor. Then, the forward propagation operation of each hidden layer is sequentially performed on the quantized input tensor. The operation process for each layer includes performing integer matrix multiplication of the input tensor with the INT8 weight matrix of that layer, accumulating the multiplication result into a 32-bit integer accumulator to prevent intermediate result overflow, adding the quantized bias vector of that layer, then restoring the accumulated result to the floating-point domain through dequantization, applying the ReLU activation function to set negative values ​​to 0 to introduce non-linear expressive power, and finally requantizing the activated result into 8-bit integer format as the input to the next layer. After forward propagation through all hidden layers, the final output layer generates a raw score vector with the same dimension as the preset number of grade levels. Each element in this vector corresponds to the unnormalized score for one grade level interval. Finally, a Softmax normalization operation is performed on the raw score vector of the output layer. This involves taking the natural exponent of each element and dividing by the sum of the natural exponents of all elements, converting the raw scores for each grade level into probability values ​​between 0 and 1, with a total sum of 1. The output result is the grade probability distribution. In fluorite mining scenarios, grade levels are defined based on industrial indicators of the mining area. For example, calcium fluoride content is divided into three grade intervals: low grade, medium grade, and rich ore. The probability value corresponding to each grade in the grade probability distribution reflects the confidence level that the current detection point belongs to that grade interval.

[0049] In the aforementioned real-time processing system 100 for edge nodes of fluorite geological data, the likelihood estimation unit 143 is used to perform likelihood estimation on the grade probability distribution to obtain the grade prediction result at the edge. It should be noted that, since the grade probability distribution is a discrete probability vector, with each element corresponding to a confidence probability within a preset grade level interval, it has not yet been converted into a continuous grade value that can be directly used for mining decisions and subsequent data compression strategy determination. Therefore, a deterministic estimate that comprehensively reflects the fluorite content level at the current detection point needs to be extracted from the probability distribution. Based on this, the technical solution of this application further performs likelihood estimation on the grade probability distribution to obtain the grade prediction result at the edge. Through the above processing, the discrete grade level probability information can be converted into a continuous grade percentage value, providing a grade fluctuation criterion for subsequent dynamic data compression and encapsulation, while also providing on-site mining personnel with an instantly readable grade reference.

[0050] More specifically, in a concrete example of this application, a pre-configured table of grade level representative values ​​is first read from the non-volatile memory of the edge node. This table records the representative percentage value of calcium fluoride content corresponding to each grade level range. For example, the representative value for low grade is 30%, for medium grade is 55%, and for rich ore is 85%. These representative values ​​are configured before the edge node deployment based on the industrial grade indicators of the mining area. Subsequently, a weighted summation operation is performed on the probability of each grade level in the grade probability distribution and the corresponding representative value of the grade level. That is, the probability value of each grade level is multiplied by the representative percentage value of that grade, and then the product results of all grades are accumulated. The specific calculation formula is as follows: in, This indicates the generated edge grade prediction results. In the probability distribution of grade, the first... The probability value of each grade level. Indicates the first The representative percentage content value corresponding to each grade. This represents the total number of preset grade categories. After the above weighted summation operation, the obtained value is the expected grade value of the current detection point. Then, a temporal smoothing correction operation is performed on this expected grade value, that is, the expected grade value at the current moment is weighted and averaged with the grade prediction result stored in the edge node cache at the previous moment. This suppresses jumps in the grade prediction value caused by occasional probability fluctuations in a single inference. The weighting coefficient is controlled by a pre-calibrated temporal smoothing constant. After temporal smoothing correction, the output value is the edge-end grade prediction result. This result is synchronously written into the circular cache of the edge node for subsequent data dynamic compression and updated to the local display interface for real-time viewing by on-site operators.

[0051] In the aforementioned real-time processing system 100 for edge nodes of fluorite ore geological data, the dynamic compression and encapsulation module 150 is used to determine a compression ratio coefficient based on the fluctuation range of the edge-end grade prediction results, and to perform dynamic compression and data encapsulation on significant mineral feature vectors according to the compression ratio coefficient to obtain optimized transmission data packets. It should be noted that when fluorite ore edge nodes report data to a remote server via a mining wireless communication link, the available bandwidth is limited and the channel quality fluctuates with changes in the underground environment. If the significant mineral feature vectors generated in each sampling period are transmitted with their original precision, the limited communication resources will be quickly exhausted. Furthermore, the information value density of the data differs between the surrounding rock section with stable grades and the vein boundary section with drastic grade changes, requiring dynamic adjustment of the data compression intensity and transmission granularity based on the real-time trend of grade changes. Therefore, the technical solution of this application further determines a compression ratio coefficient based on the fluctuation range of the edge-end grade prediction results, and performs dynamic compression and data encapsulation on significant mineral feature vectors according to the compression ratio coefficient to obtain optimized transmission data packets. Through the above processing, the compression rate can be increased during periods of stable grade to save communication bandwidth, and the compression rate can be reduced during periods of sudden grade changes to preserve key geological information, thus achieving an adaptive balance between transmission load and data fidelity.

[0052] More specifically, in a concrete example of this application, the stability of the temporal volatility of the edge grade prediction results within a sliding window is first evaluated to obtain the compression ratio coefficient. Edge grade prediction results from the most recent consecutive sampling periods are read from the circular buffer of the edge nodes to form a grade time window sequence. The arithmetic mean of this window sequence is calculated as the window prediction mean. Then, the mean of the sum of squares of the differences between each grade prediction value within the window and the window prediction mean is calculated. The arithmetic square root of this mean is then taken to obtain the standard deviation of the grade prediction values ​​within the window. This standard deviation reflects the severity of grade changes in the current period. Furthermore, a pre-calibrated sensitivity adjustment factor is divided by the sum of the above standard deviation and a smoothing constant, and the calculation result is constrained between a preset minimum compression ratio and a maximum compression ratio using a limiting function. The specific calculation formula is as follows: in, This represents the generated compression ratio coefficient. Indicates the first [number]th ... Edge grade prediction results for each sampling period This represents the arithmetic mean of all predicted grades within the window. This indicates the length of the sliding window, which is the number of consecutive sampling periods involved in the evaluation. This indicates that a pre-calibrated sensitivity adjustment factor is used to control the driving force of volatility on the compression ratio. The smoothing constant is used to prevent numerical overflow when the denominator approaches 0. and These represent the minimum and maximum allowable compression ratios, respectively. When the grade prediction fluctuates drastically within the window, the standard deviation is large, and the compression ratio approaches the minimum compression ratio, allowing subsequent compression stages to retain more feature details. When the grade prediction remains stable within the window, the standard deviation approaches 0, and the compression ratio approaches the maximum compression ratio, allowing subsequent compression stages to reduce the data volume with a higher compression rate.

[0053] Subsequently, adaptive sparse compression processing is performed on the significant mineral feature vectors based on the compression ratio coefficient to obtain a compressed feature data stream. The number of feature vector dimensions to be retained in the current sampling period is determined according to the compression ratio coefficient, i.e., the total number of dimensions of the significant mineral feature vectors is divided by the compression ratio coefficient and rounded down to obtain the number of retained dimensions. The absolute values ​​of each dimension in the significant mineral feature vectors are sorted in descending order, and the top few dimensions with the highest absolute values ​​and their corresponding index numbers are selected and retained, while the values ​​of the remaining dimensions are set to 0, thus achieving sparsification based on amplitude sorting. The non-zero elements and their index numbers after sparsification are compressed and encoded using run-length encoding, replacing consecutive zero-value segments with run-length markers. The output encoded sequence is the compressed feature data stream. During periods of rapid grade change, the compression ratio coefficient is smaller, and a larger number of dimensions are retained, fully preserving the detailed information in the feature vectors. During periods of stable grade, the compression ratio coefficient is larger, retaining only a few key dimensions with the highest amplitudes, effectively reducing the data volume.

[0054] Next, the compressed feature data stream and the edge grade prediction results are encapsulated using a protocol to obtain an optimized transmission data packet. The edge grade prediction results of the current sampling period are used as key metadata fields and concatenated with the compressed feature data stream in binary format. The device identification code of the edge node, the timestamp of the current sampling time, the compression ratio coefficient used, and the compression algorithm identifier are sequentially written into the header of the data packet. A cyclic redundancy check code is appended to the tail of the data packet for the receiver to verify data integrity. Finally, an optimized transmission data packet conforming to the mining wireless communication protocol format is constructed.

[0055] In summary, the real-time processing system for edge nodes of fluorite mine geological data according to the embodiments of this application is explained. It solves the problem that geological data cannot be transmitted back in real time and analyzed on-site under the condition of limited communication bandwidth in the mining area by constructing a full-link processing flow from multi-source sensor data acquisition to compressed transmission at the edge side. The system first performs time alignment on the data streams from heterogeneous sensors such as triaxial vibration acceleration, X-ray fluorescence pulse counting, and near-infrared spectral reflectance to eliminate timescale drift of multi-source signals. Then, it adaptively adjusts the filtering intensity using the power spectral density of the vibration components, suppressing mechanical construction noise while preserving effective geological features representing grade changes, thus solving the problem of detail loss and signal distortion caused by the overlap of noise and signal frequency bands in conventional filtering methods. Next, it performs significant weighted extraction based on the energy spectrum characteristics of calcium and fluorine elements in fluorite, addressing the lack of specificity of general feature extraction methods for fluorite. The extracted feature vectors are then fed into an edge-side quantization neural network to complete real-time grade prediction. Finally, the compression ratio is dynamically adjusted based on the fluctuation range of the prediction results, performing adaptive compression and protocol encapsulation on the feature data. This effectively reduces the transmission load while ensuring the integrity of key geological information, enabling edge nodes to achieve real-time perception and efficient reporting of fluorite vein grade changes.

Claims

1. A real-time processing system for edge nodes of fluorite mineral geological data, characterized in that, include: The timing alignment module is used to perform timing alignment processing on the acquired raw sensor data stream to obtain the aligned raw matrix. The raw sensor data stream includes triaxial vibration acceleration components, X-ray fluorescence intensity pulse count, and near-infrared spectral reflectance. An adaptive noise suppression module is used to adaptively suppress environmental noise on the aligned original matrix to obtain a denoised geological signal; The feature saliency extraction module is used to extract fluorite feature saliency from the denoised geological signal to obtain salient mineral feature vectors; The edge inference module is used to input significant mineral feature vectors into a pre-deployed quantized neural network for forward inference to obtain edge grade prediction results; The dynamic compression and encapsulation module is used to determine the compression ratio coefficient based on the fluctuation range of the edge grade prediction results, and to perform dynamic compression and data encapsulation on the significant mineral feature vectors according to the compression ratio coefficient to obtain optimized transmission data packets.

2. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 1, characterized in that, The timing alignment module includes: The concurrent conversion unit is used to perform concurrent conversion processing on the acquired raw sensor data stream to obtain multiple raw digital signals; The zero-point calibration unit is used to perform zero-point calibration processing on multiple raw digital signals based on preset reference drift correction parameters to obtain calibrated digital signals. The interpolation alignment encapsulation unit is used to perform interpolation alignment encapsulation on the calibrated digital signal using the highest sampling frequency as a clock reference to obtain the aligned original matrix.

3. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 1, characterized in that, The adaptive noise suppression module includes: The vibration feature analysis unit is used to perform vibration feature analysis and power spectrum estimation on the vibration components extracted from the aligned original matrix through fast Fourier transform to obtain the background noise power spectral density. The smoothing factor mapping unit is used to perform dynamic mapping estimation of the background noise power spectral density to obtain the smoothing factor; The temporal smoothing unit is used to perform temporal smoothing on the fluorescence pulse count components in the aligned original matrix based on a smoothing factor to obtain a denoised geological signal.

4. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 1, characterized in that, The feature saliency extraction module includes: Sensitive energy channel slicing unit is used to extract signal slices corresponding to the sensitive energy channels of calcium and fluorine elements from denoised geological signals based on the energy spectrum characteristics of fluorite minerals to obtain a characteristic frequency domain set. The saliency weighting unit is used to perform saliency weighting modeling on the feature frequency domain set to obtain the saliency index; The vectorization mapping unit is used to perform vectorization mapping processing on the saliency index and the spectral components in the denoised geological signal to obtain saliency mineral feature vectors.

5. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 1, characterized in that, The edge reasoning module includes: The normalization unit is used to normalize the salient mineral feature vectors to obtain normalized feature vectors. The probability inference unit is used to perform interval probability inference on the normalized feature vector using a quantized neural network engine to obtain the grade probability distribution. The likelihood estimation unit is used to perform likelihood estimation on the grade probability distribution to obtain the grade prediction results at the edge.

6. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 1, characterized in that, The dynamic compression and encapsulation module includes: The stability assessment unit is used to assess the stability of the temporal volatility of the edge grade prediction results within the sliding window in order to obtain the compression ratio coefficient. A sparse compression unit is used to perform adaptive sparse compression processing on significant mineral feature vectors based on a compression ratio coefficient to obtain a compressed feature data stream. The protocol encapsulation unit is used to encapsulate the compressed feature data stream and the edge grade prediction results to obtain optimized transmission data packets.

7. The real-time processing system for edge nodes of fluorite mineral geological data according to claim 3, characterized in that, The timing smoothing unit includes: The coupling estimation subunit is used to perform vibration-signal time-domain derivative coupling estimation on the vibration components extracted from the aligned original matrix and the fluorescence pulse count to obtain the characteristic coupling factor; The gain weight correction subunit is used to perform dynamic phase gain weight correction based on the characteristic coupling factor and combined with the smoothing factor to obtain the corrected gain weight. The phase decoupling filter subunit is used to perform phase decoupling filtering on the fluorescence pulse count and historical denoised signal in the aligned original matrix based on the modified gain weight and the introduction of a phase compensation term to obtain the denoised geological signal.

Images